From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <M.Kretz@gsi.de>
Received: from lxmtout2.gsi.de (lxmtout2.gsi.de [140.181.3.112])
 by sourceware.org (Postfix) with ESMTPS id 7C184386103B;
 Fri, 18 Dec 2020 15:49:49 +0000 (GMT)
DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org 7C184386103B
Authentication-Results: sourceware.org;
 dmarc=none (p=none dis=none) header.from=gsi.de
Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=M.Kretz@gsi.de
Received: from localhost (localhost [127.0.0.1])
 by lxmtout2.gsi.de (Postfix) with ESMTP id 72226202AD6D;
 Fri, 18 Dec 2020 16:49:48 +0100 (CET)
X-Virus-Scanned: Debian amavisd-new at lxmtout2.gsi.de
Received: from lxmtout2.gsi.de ([127.0.0.1])
 by localhost (lxmtout2.gsi.de [127.0.0.1]) (amavisd-new, port 10024)
 with LMTP id V-ibOnya0daW; Fri, 18 Dec 2020 16:49:47 +0100 (CET)
Received: from srvex3.campus.gsi.de (srvex3.campus.gsi.de [10.10.4.16])
 (using TLSv1.2 with cipher ECDHE-RSA-AES128-SHA256 (128/128 bits))
 (No client certificate requested)
 by lxmtout2.gsi.de (Postfix) with ESMTPS id 88AF3202AD5D;
 Fri, 18 Dec 2020 16:49:45 +0100 (CET)
Received: from excalibur.localnet (140.181.3.12) by srvex3.campus.gsi.de
 (10.10.4.16) with Microsoft SMTP Server (version=TLS1_2,
 cipher=TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256_P256) id 15.1.2106.2; Fri, 18
 Dec 2020 16:49:43 +0100
From: Matthias Kretz <m.kretz@gsi.de>
To: <gcc-patches@gcc.gnu.org>, <libstdc++@gcc.gnu.org>
Subject: [PATCH 1/2] Add std::experimental::simd from the Parallelism TS 2
Date: Fri, 18 Dec 2020 16:49:43 +0100
Message-ID: <3133130.QemTDgPxUG@excalibur>
Organization: GSI Helmholtzzentrum =?UTF-8?B?ZsO8cg==?= Schwerionenforschung
MIME-Version: 1.0
Content-Type: multipart/mixed; boundary="nextPart3621190.fJnDbHsrEp"
Content-Transfer-Encoding: 7Bit
X-Originating-IP: [140.181.3.12]
X-ClientProxiedBy: SRVEX2.campus.gsi.de (10.10.4.15) To srvex3.campus.gsi.de
 (10.10.4.16)
X-Spam-Status: No, score=-11.4 required=5.0 tests=BAYES_00, BODY_8BITS,
 GIT_PATCH_0, KAM_DMARC_STATUS, KAM_NUMSUBJECT, KAM_SHORT, RCVD_IN_DNSWL_NONE,
 SCC_10_SHORT_WORD_LINES, SCC_20_SHORT_WORD_LINES, SCC_35_SHORT_WORD_LINES,
 SCC_5_SHORT_WORD_LINES, SPF_PASS, TXREP,
 T_SPF_HELO_PERMERROR autolearn=ham autolearn_force=no version=3.4.2
X-Spam-Checker-Version: SpamAssassin 3.4.2 (2018-09-13) on
 server2.sourceware.org
X-BeenThere: libstdc++@gcc.gnu.org
X-Mailman-Version: 2.1.29
Precedence: list
List-Id: Libstdc++ mailing list <libstdc++.gcc.gnu.org>
List-Unsubscribe: <https://gcc.gnu.org/mailman/options/libstdc++>,
 <mailto:libstdc++-request@gcc.gnu.org?subject=unsubscribe>
List-Archive: <https://gcc.gnu.org/pipermail/libstdc++/>
List-Post: <mailto:libstdc++@gcc.gnu.org>
List-Help: <mailto:libstdc++-request@gcc.gnu.org?subject=help>
List-Subscribe: <https://gcc.gnu.org/mailman/listinfo/libstdc++>,
 <mailto:libstdc++-request@gcc.gnu.org?subject=subscribe>
X-List-Received-Date: Fri, 18 Dec 2020 15:50:52 -0000

--nextPart3621190.fJnDbHsrEp
Content-Transfer-Encoding: quoted-printable
Content-Type: text/plain; charset="UTF-8"

Resending this patch with proper commit message and rebased on master.

=46rom: Matthias Kretz <kretz@kde.org>

Adds <experimental/simd>.

This implements the simd and simd_mask class templates via
[[gnu::vector_size(N)]] data members. It implements overloads for all of
<cmath> for simd. Explicit vectorization of the <cmath> functions is not
finished.
The majority of functions are marked as [[gnu::always_inline]] to enable
quasi-ODR-conforming linking of TUs with different -m flags.
Performance optimization was done for x86_64. ARM, Aarch64, and POWER
rely on the compiler to recognize reduction, conversion, and shuffle
patterns.
Besides verification using many different machine flages, the code was
also verified with different fast-math flags.

libstdc++-v3/ChangeLog:
	* doc/xml/manual/status_cxx2017.xml: Add implementation status
	of the Parallelism TS 2. Document implementation-defined types
	and behavior.
	* include/Makefile.am: Add new headers.
	* include/Makefile.in: Regenerate.
	* include/experimental/simd: New file. New header for
	Parallelism TS 2.
	* include/experimental/bits/numeric_traits.h: New file.
	Implementation of P1841R1 using internal naming. Addition of
	missing IEC559 functionality query.
	* include/experimental/bits/simd.h: New file. Definition of the
	public simd interfaces and general implementation helpers.
	* include/experimental/bits/simd_builtin.h: New file.
	Implementation of the _VecBuiltin simd_abi.
	* include/experimental/bits/simd_converter.h: New file. Generic
	simd conversions.
	* include/experimental/bits/simd_detail.h: New file. Internal
	macros for the simd implementation.
	* include/experimental/bits/simd_fixed_size.h: New file. Simd
	fixed_size ABI specific implementations.
	* include/experimental/bits/simd_math.h: New file. Math
	overloads for simd.
	* include/experimental/bits/simd_neon.h: New file. Simd NEON
	specific implementations.
	* include/experimental/bits/simd_ppc.h: New file. Implement bit
	shifts to avoid invalid results for integral types smaller than
	int.
	* include/experimental/bits/simd_scalar.h: New file. Simd scalar
	ABI specific implementations.
	* include/experimental/bits/simd_x86.h: New file. Simd x86
	specific implementations.
	* include/experimental/bits/simd_x86_conversions.h: New file.
	x86 specific conversion optimizations. The conversion patterns
	work around missing conversion patterns in the compiler and
	should be removed as soon as PR85048 is resolved.
	* testsuite/experimental/simd/standard_abi_usable.cc: New file.
	Test that all (not all fixed_size<N>, though) standard simd and
	simd_mask types are usable.
	* testsuite/experimental/simd/standard_abi_usable_2.cc: New
	file. As above but with -ffast-math.
	* testsuite/libstdc++-dg/conformance.exp: Don't build simd tests
	from the standard test loop. Instead use
	check_vect_support_and_set_flags to build simd tests with the
	relevant machine flags.
=2D--
 .../doc/xml/manual/status_cxx2017.xml         |  216 +
 libstdc++-v3/include/Makefile.am              |   13 +
 libstdc++-v3/include/Makefile.in              |   13 +
 .../experimental/bits/numeric_traits.h        |  567 ++
 libstdc++-v3/include/experimental/bits/simd.h | 5051 ++++++++++++++++
 .../include/experimental/bits/simd_builtin.h  | 2949 ++++++++++
 .../experimental/bits/simd_converter.h        |  354 ++
 .../include/experimental/bits/simd_detail.h   |  306 +
 .../experimental/bits/simd_fixed_size.h       | 2066 +++++++
 .../include/experimental/bits/simd_math.h     | 1500 +++++
 .../include/experimental/bits/simd_neon.h     |  519 ++
 .../include/experimental/bits/simd_ppc.h      |  123 +
 .../include/experimental/bits/simd_scalar.h   |  772 +++
 .../include/experimental/bits/simd_x86.h      | 5169 +++++++++++++++++
 .../experimental/bits/simd_x86_conversions.h  | 2029 +++++++
 libstdc++-v3/include/experimental/simd        |   70 +
 .../experimental/simd/standard_abi_usable.cc  |   64 +
 .../simd/standard_abi_usable_2.cc             |    4 +
 .../testsuite/libstdc++-dg/conformance.exp    |   18 +-
 19 files changed, 21802 insertions(+), 1 deletion(-)
 create mode 100644 libstdc++-v3/include/experimental/bits/numeric_traits.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_builtin.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_converter.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_detail.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_fixed_size.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_math.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_neon.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_ppc.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_scalar.h
 create mode 100644 libstdc++-v3/include/experimental/bits/simd_x86.h
 create mode 100644 libstdc++-v3/include/experimental/bits/
simd_x86_conversions.h
 create mode 100644 libstdc++-v3/include/experimental/simd
 create mode 100644 libstdc++-v3/testsuite/experimental/simd/
standard_abi_usable.cc
 create mode 100644 libstdc++-v3/testsuite/experimental/simd/
standard_abi_usable_2.cc


=2D-
=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=
=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=
=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=
=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=
=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=
=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=
=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=
=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=
=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80
 Dr. Matthias Kretz                           https://mattkretz.github.io
 GSI Helmholtz Centre for Heavy Ion Research               https://gsi.de
 std::experimental::simd              https://github.com/VcDevel/std-simd
=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=
=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=
=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=
=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=
=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=
=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=
=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=
=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=
=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80=E2=94=80
--nextPart3621190.fJnDbHsrEp
Content-Disposition: inline;
	filename="0001-Add-std-experimental-simd-from-the-Parallelism-TS-2.patch"
Content-Transfer-Encoding: quoted-printable
Content-Type: text/x-patch; charset="utf-8";
	name="0001-Add-std-experimental-simd-from-the-Parallelism-TS-2.patch"

diff --git a/libstdc++-v3/doc/xml/manual/status_cxx2017.xml b/libstdc++-v3/=
doc/xml/manual/status_cxx2017.xml
index e6834b3607a..bc740f8e1ba 100644
=2D-- a/libstdc++-v3/doc/xml/manual/status_cxx2017.xml
+++ b/libstdc++-v3/doc/xml/manual/status_cxx2017.xml
@@ -2869,6 +2869,17 @@ since C++14 and the implementation is complete.
       <entry>Library Fundamentals 2 TS</entry>
     </row>
=20
+    <row>
+      <entry>
+	<link xmlns:xlink=3D"http://www.w3.org/1999/xlink" xlink:href=3D"http://w=
ww.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0214r9.pdf">
+	  P0214R9
+	</link>
+      </entry>
+      <entry>Data-Parallel Types</entry>
+      <entry>Y</entry>
+      <entry>Parallelism 2 TS</entry>
+    </row>
+
   </tbody>
 </tgroup>
 </table>
@@ -3014,6 +3025,211 @@ since C++14 and the implementation is complete.
       If <code>!is_regular_file(p)</code>, an error is reported.
    </para>
=20
+   <section xml:id=3D"iso.2017.par2ts" xreflabel=3D"Implementation Specifi=
c Behavior of the Parallelism 2 TS"><info><title>Parallelism 2 TS</title></=
info>
+
+     <para>
+        <emphasis>9.3 [parallel.simd.abi]</emphasis>
+        <code>max_fixed_size&lt;T&gt;</code> is 32, except when targetting
+        AVX512BW and <code>sizeof(T)</code> is 1.
+     </para>
+
+     <para>
+        When targeting 32-bit x86,
+        <classname>simd_abi::compatible&lt;T&gt;</classname> is an alias f=
or
+        <classname>simd_abi::scalar</classname>.
+        When targeting 64-bit x86 (including x32) or Aarch64,
+        <classname>simd_abi::compatible&lt;T&gt;</classname> is an alias f=
or
+        <classname>simd_abi::_VecBuiltin&lt;16&gt;</classname>,
+        unless <code>T</code> is <code>long double</code>, in which case i=
t is
+        an alias for <classname>simd_abi::scalar</classname>.
+        When targeting ARM (but not Aarch64) with NEON support,
+        <classname>simd_abi::compatible&lt;T&gt;</classname> is an alias f=
or
+        <classname>simd_abi::_VecBuiltin&lt;16&gt;</classname>,
+        unless <code>sizeof(T) &gt; 4</code>, in which case it is
+        an alias for <classname>simd_abi::scalar</classname>. Additionally,
+        <classname>simd_abi::compatible&lt;float&gt;</classname> is an ali=
as for
+        <classname>simd_abi::scalar</classname> unless compiling with
+        -ffast-math.
+     </para>
+
+     <para>
+        When targeting x86 (both 32-bit and 64-bit),
+        <classname>simd_abi::native&lt;T&gt;</classname> is an alias for o=
ne of
+        <classname>simd_abi::scalar</classname>,
+        <classname>simd_abi::_VecBuiltin&lt;16&gt;</classname>,
+        <classname>simd_abi::_VecBuiltin&lt;32&gt;</classname>, or
+        <classname>simd_abi::_VecBltnBtmsk&lt;64&gt;</classname>, dependin=
g on
+        <code>T</code> and the machine options the compiler was invoked wi=
th.
+     </para>
+
+     <para>
+        When targeting ARM/Aarch64 or POWER,
+        <classname>simd_abi::native&lt;T&gt;</classname> is an alias for
+        <classname>simd_abi::scalar</classname> or
+        <classname>simd_abi::_VecBuiltin&lt;16&gt;</classname>, depending =
on
+        <code>T</code> and the machine options the compiler was invoked wi=
th.
+     </para>
+
+     <para>
+        For any other targeted machine
+        <classname>simd_abi::compatible&lt;T&gt;</classname> and
+        <classname>simd_abi::native&lt;T&gt;</classname> are aliases for
+        <classname>simd_abi::scalar</classname>. (subject to change)
+     </para>
+
+     <para>
+        The extended ABI tag types defined in the
+        <code>std::experimental::parallelism_v2::simd_abi</code> namespace=
 are:
+        <classname>simd_abi::_VecBuiltin&lt;Bytes&gt;</classname>, and
+        <classname>simd_abi::_VecBltnBtmsk&lt;Bytes&gt;</classname>.
+     </para>
+
+     <para>
+        <classname>simd_abi::deduce&lt;T, N, Abis...&gt;::type</classname>,
+        with <code>N &gt; 1</code> is an alias for an extended ABI tag, if=
 a
+        supported extended ABI tag exists. Otherwise it is an alias for
+        <classname>simd_abi::fixed_size&lt;N&gt;</classname>. The <classna=
me>
+        simd_abi::_VecBltnBtmsk</classname> ABI tag is preferred over
+        <classname>simd_abi::_VecBuiltin</classname>.
+     </para>
+
+     <para>
+        <emphasis>9.4 [parallel.simd.traits]</emphasis>
+        <classname>memory_alignment&lt;T, U&gt;::value</classname> is
+        <code>sizeof(U) * T::size()</code> rounded up to the next power-of=
=2Dtwo
+        value.
+     </para>
+
+     <para>
+        <emphasis>9.6.1 [parallel.simd.overview]</emphasis>
+        On ARM, <classname>simd&lt;T, _VecBuiltin&lt;Bytes&gt;&gt;</classn=
ame>
+        is supported if <code>__ARM_NEON</code> is defined and
+        <code>sizeof(T) &lt;=3D 4</code>. Additionally,
+        <code>sizeof(T) =3D=3D 8</code> with integral <code>T</code> is su=
pported if
+        <code>__ARM_ARCH &gt;=3D 8</code>, and <code>double</code> is supp=
orted if
+        <code>__aarch64__</code> is defined.
+
+        On POWER, <classname>simd&lt;T, _VecBuiltin&lt;Bytes&gt;&gt;</clas=
sname>
+        is supported if <code>__ALTIVEC__</code> is defined and <code>size=
of(T)
+        &lt; 8</code>. Additionally, <code>double</code> is supported if
+        <code>__VSX__</code> is defined, and any <code>T</code> with <code>
+        sizeof(T) &le; 8</code> is supported if <code>__POWER8_VECTOR__</c=
ode>
+        is defined.
+
+        On x86, given an extended ABI tag <code>Abi</code>,
+        <classname>simd&lt;T, Abi&gt;</classname> is supported according t=
o the
+        following table:
+        <table frame=3D"all" xml:id=3D"table.par2ts_simd_support">
+          <title>Support for Extended ABI Tags</title>
+
+          <tgroup cols=3D"4" align=3D"left" colsep=3D"0" rowsep=3D"1">
+          <colspec colname=3D"c1"/>
+          <colspec colname=3D"c2"/>
+          <colspec colname=3D"c3"/>
+          <colspec colname=3D"c4"/>
+            <thead>
+              <row>
+                <entry>ABI tag <code>Abi</code></entry>
+                <entry>value type <code>T</code></entry>
+                <entry>values for <code>Bytes</code></entry>
+                <entry>required machine option</entry>
+              </row>
+            </thead>
+
+            <tbody>
+              <row>
+                <entry morerows=3D"5">
+                  <classname>_VecBuiltin&lt;Bytes&gt;</classname>
+                </entry>
+                <entry morerows=3D"1"><code>float</code></entry>
+                <entry>8, 12, 16</entry>
+                <entry>"-msse"</entry>
+              </row>
+
+              <row>
+                <entry>20, 24, 28, 32</entry>
+                <entry>"-mavx"</entry>
+              </row>
+
+              <row>
+                <entry morerows=3D"1"><code>double</code></entry>
+                <entry>16</entry>
+                <entry>"-msse2"</entry>
+              </row>
+
+              <row>
+                <entry>24, 32</entry>
+                <entry>"-mavx"</entry>
+              </row>
+
+              <row>
+                <entry morerows=3D"1">
+                  integral types other than <code>bool</code>
+                </entry>
+                <entry>
+                  <code>Bytes</code> =E2=89=A4 16 and <code>Bytes</code> d=
ivisible by
+                  <code>sizeof(T)</code>
+                </entry>
+                <entry>"-msse2"</entry>
+              </row>
+
+              <row>
+                <entry>
+                  16 &lt; <code>Bytes</code> =E2=89=A4 32 and <code>Bytes<=
/code>
+                  divisible by <code>sizeof(T)</code>
+                </entry>
+                <entry>"-mavx2"</entry>
+              </row>
+
+              <row>
+                <entry morerows=3D"1">
+                  <classname>_VecBuiltin&lt;Bytes&gt;</classname> and
+                  <classname>_VecBltnBtmsk&lt;Bytes&gt;</classname>
+                </entry>
+                <entry>
+                  vectorizable types with <code>sizeof(T)</code> =E2=89=A5=
 4
+                </entry>
+                <entry morerows=3D"1">
+                  32 &lt; <code>Bytes</code> =E2=89=A4 64 and <code>Bytes<=
/code>
+                  divisible by <code>sizeof(T)</code>
+                </entry>
+                <entry>"-mavx512f"</entry>
+              </row>
+
+              <row>
+                <entry>
+                  vectorizable types with <code>sizeof(T)</code> &lt; 4
+                </entry>
+                <entry>"-mavx512bw"</entry>
+              </row>
+
+              <row>
+                <entry morerows=3D"1">
+                  <classname>_VecBltnBtmsk&lt;Bytes&gt;</classname>
+                </entry>
+                <entry>
+                  vectorizable types with <code>sizeof(T)</code> =E2=89=A5=
 4
+                </entry>
+                <entry morerows=3D"1">
+                  <code>Bytes</code> =E2=89=A4 32 and <code>Bytes</code> d=
ivisible by
+                  <code>sizeof(T)</code>
+                </entry>
+                <entry>"-mavx512vl"</entry>
+              </row>
+
+              <row>
+                <entry>
+                  vectorizable types with <code>sizeof(T)</code> &lt; 4
+                </entry>
+                <entry>"-mavx512bw" and "-mavx512vl"</entry>
+              </row>
+
+            </tbody>
+          </tgroup>
+        </table>
+     </para>
+
+   </section>
=20
 </section>
=20
diff --git a/libstdc++-v3/include/Makefile.am b/libstdc++-v3/include/Makefi=
le.am
index 958dfea5a98..3ad24267cfd 100644
=2D-- a/libstdc++-v3/include/Makefile.am
+++ b/libstdc++-v3/include/Makefile.am
@@ -746,6 +746,7 @@ experimental_headers =3D \
 	${experimental_srcdir}/ratio \
 	${experimental_srcdir}/regex \
 	${experimental_srcdir}/set \
+	${experimental_srcdir}/simd \
 	${experimental_srcdir}/socket \
 	${experimental_srcdir}/source_location \
 	${experimental_srcdir}/string \
@@ -765,7 +766,19 @@ experimental_bits_builddir =3D ./experimental/bits
 experimental_bits_headers =3D \
 	${experimental_bits_srcdir}/lfts_config.h \
 	${experimental_bits_srcdir}/net.h \
+	${experimental_bits_srcdir}/numeric_traits.h \
 	${experimental_bits_srcdir}/shared_ptr.h \
+	${experimental_bits_srcdir}/simd.h \
+	${experimental_bits_srcdir}/simd_builtin.h \
+	${experimental_bits_srcdir}/simd_converter.h \
+	${experimental_bits_srcdir}/simd_detail.h \
+	${experimental_bits_srcdir}/simd_fixed_size.h \
+	${experimental_bits_srcdir}/simd_math.h \
+	${experimental_bits_srcdir}/simd_neon.h \
+	${experimental_bits_srcdir}/simd_ppc.h \
+	${experimental_bits_srcdir}/simd_scalar.h \
+	${experimental_bits_srcdir}/simd_x86.h \
+	${experimental_bits_srcdir}/simd_x86_conversions.h \
 	${experimental_bits_srcdir}/string_view.tcc \
 	${experimental_bits_filesystem_headers}
=20
diff --git a/libstdc++-v3/include/Makefile.in b/libstdc++-v3/include/Makefi=
le.in
index b3256a7835e..2692b8352be 100644
=2D-- a/libstdc++-v3/include/Makefile.in
+++ b/libstdc++-v3/include/Makefile.in
@@ -1096,6 +1096,7 @@ experimental_headers =3D \
 	${experimental_srcdir}/ratio \
 	${experimental_srcdir}/regex \
 	${experimental_srcdir}/set \
+	${experimental_srcdir}/simd \
 	${experimental_srcdir}/socket \
 	${experimental_srcdir}/source_location \
 	${experimental_srcdir}/string \
@@ -1115,7 +1116,19 @@ experimental_bits_builddir =3D ./experimental/bits
 experimental_bits_headers =3D \
 	${experimental_bits_srcdir}/lfts_config.h \
 	${experimental_bits_srcdir}/net.h \
+	${experimental_bits_srcdir}/numeric_traits.h \
 	${experimental_bits_srcdir}/shared_ptr.h \
+	${experimental_bits_srcdir}/simd.h \
+	${experimental_bits_srcdir}/simd_builtin.h \
+	${experimental_bits_srcdir}/simd_converter.h \
+	${experimental_bits_srcdir}/simd_detail.h \
+	${experimental_bits_srcdir}/simd_fixed_size.h \
+	${experimental_bits_srcdir}/simd_math.h \
+	${experimental_bits_srcdir}/simd_neon.h \
+	${experimental_bits_srcdir}/simd_ppc.h \
+	${experimental_bits_srcdir}/simd_scalar.h \
+	${experimental_bits_srcdir}/simd_x86.h \
+	${experimental_bits_srcdir}/simd_x86_conversions.h \
 	${experimental_bits_srcdir}/string_view.tcc \
 	${experimental_bits_filesystem_headers}
=20
diff --git a/libstdc++-v3/include/experimental/bits/numeric_traits.h b/libs=
tdc++-v3/include/experimental/bits/numeric_traits.h
new file mode 100644
index 00000000000..1b60874b788
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/numeric_traits.h
@@ -0,0 +1,567 @@
+// Definition of numeric_limits replacement traits P1841R1 -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#include <type_traits>
+
+namespace std {
+
+template <template <typename> class _Trait, typename _Tp, typename =3D voi=
d>
+  struct __value_exists_impl : false_type {};
+
+template <template <typename> class _Trait, typename _Tp>
+  struct __value_exists_impl<_Trait, _Tp, void_t<decltype(_Trait<_Tp>::val=
ue)>>
+  : true_type {};
+
+template <typename _Tp, bool =3D is_arithmetic_v<_Tp>>
+  struct __digits_impl {};
+
+template <typename _Tp>
+  struct __digits_impl<_Tp, true>
+  {
+    static inline constexpr int value
+      =3D sizeof(_Tp) * __CHAR_BIT__ - is_signed_v<_Tp>;
+  };
+
+template <>
+  struct __digits_impl<float, true>
+  { static inline constexpr int value =3D __FLT_MANT_DIG__; };
+
+template <>
+  struct __digits_impl<double, true>
+  { static inline constexpr int value =3D __DBL_MANT_DIG__; };
+
+template <>
+  struct __digits_impl<long double, true>
+  { static inline constexpr int value =3D __LDBL_MANT_DIG__; };
+
+template <typename _Tp, bool =3D is_arithmetic_v<_Tp>>
+  struct __digits10_impl {};
+
+template <typename _Tp>
+  struct __digits10_impl<_Tp, true>
+  {
+    // The fraction 643/2136 approximates log10(2) to 7 significant digits.
+    static inline constexpr int value =3D __digits_impl<_Tp>::value * 643L=
 / 2136;
+  };
+
+template <>
+  struct __digits10_impl<float, true>
+  { static inline constexpr int value =3D __FLT_DIG__; };
+
+template <>
+  struct __digits10_impl<double, true>
+  { static inline constexpr int value =3D __DBL_DIG__; };
+
+template <>
+  struct __digits10_impl<long double, true>
+  { static inline constexpr int value =3D __LDBL_DIG__; };
+
+template <typename _Tp, bool =3D is_arithmetic_v<_Tp>>
+  struct __max_digits10_impl {};
+
+template <typename _Tp>
+  struct __max_digits10_impl<_Tp, true>
+  {
+    static inline constexpr int value
+      =3D is_floating_point_v<_Tp> ? 2 + __digits_impl<_Tp>::value * 643L =
/ 2136
+				 : __digits10_impl<_Tp>::value + 1;
+  };
+
+template <typename _Tp>
+  struct __max_exponent_impl {};
+
+template <>
+  struct __max_exponent_impl<float>
+  { static inline constexpr int value =3D __FLT_MAX_EXP__; };
+
+template <>
+  struct __max_exponent_impl<double>
+  { static inline constexpr int value =3D __DBL_MAX_EXP__; };
+
+template <>
+  struct __max_exponent_impl<long double>
+  { static inline constexpr int value =3D __LDBL_MAX_EXP__; };
+
+template <typename _Tp>
+  struct __max_exponent10_impl {};
+
+template <>
+  struct __max_exponent10_impl<float>
+  { static inline constexpr int value =3D __FLT_MAX_10_EXP__; };
+
+template <>
+  struct __max_exponent10_impl<double>
+  { static inline constexpr int value =3D __DBL_MAX_10_EXP__; };
+
+template <>
+  struct __max_exponent10_impl<long double>
+  { static inline constexpr int value =3D __LDBL_MAX_10_EXP__; };
+
+template <typename _Tp>
+  struct __min_exponent_impl {};
+
+template <>
+  struct __min_exponent_impl<float>
+  { static inline constexpr int value =3D __FLT_MIN_EXP__; };
+
+template <>
+  struct __min_exponent_impl<double>
+  { static inline constexpr int value =3D __DBL_MIN_EXP__; };
+
+template <>
+  struct __min_exponent_impl<long double>
+  { static inline constexpr int value =3D __LDBL_MIN_EXP__; };
+
+template <typename _Tp>
+  struct __min_exponent10_impl {};
+
+template <>
+  struct __min_exponent10_impl<float>
+  { static inline constexpr int value =3D __FLT_MIN_10_EXP__; };
+
+template <>
+  struct __min_exponent10_impl<double>
+  { static inline constexpr int value =3D __DBL_MIN_10_EXP__; };
+
+template <>
+  struct __min_exponent10_impl<long double>
+  { static inline constexpr int value =3D __LDBL_MIN_10_EXP__; };
+
+template <typename _Tp, bool =3D is_arithmetic_v<_Tp>>
+  struct __radix_impl {};
+
+template <typename _Tp>
+  struct __radix_impl<_Tp, true>
+  {
+    static inline constexpr int value
+      =3D is_floating_point_v<_Tp> ? __FLT_RADIX__ : 2;
+  };
+
+// [num.traits.util], numeric utility traits
+template <template <typename> class _Trait, typename _Tp>
+  struct __value_exists : __value_exists_impl<_Trait, _Tp> {};
+
+template <template <typename> class _Trait, typename _Tp>
+  inline constexpr bool __value_exists_v =3D __value_exists<_Trait, _Tp>::=
value;
+
+template <template <typename> class _Trait, typename _Tp, typename _Up =3D=
 _Tp>
+  inline constexpr _Up
+  __value_or(_Up __def =3D _Up()) noexcept
+  {
+    if constexpr (__value_exists_v<_Trait, _Tp>)
+      return static_cast<_Up>(_Trait<_Tp>::value);
+    else
+      return __def;
+  }
+
+template <typename _Tp, bool =3D is_arithmetic_v<_Tp>>
+  struct __norm_min_impl {};
+
+template <typename _Tp>
+  struct __norm_min_impl<_Tp, true>
+  { static inline constexpr _Tp value =3D 1; };
+
+template <>
+  struct __norm_min_impl<float, true>
+  { static inline constexpr float value =3D __FLT_MIN__; };
+
+template <>
+  struct __norm_min_impl<double, true>
+  { static inline constexpr double value =3D __DBL_MIN__; };
+
+template <>
+  struct __norm_min_impl<long double, true>
+  { static inline constexpr long double value =3D __LDBL_MIN__; };
+
+template <typename _Tp>
+  struct __denorm_min_impl : __norm_min_impl<_Tp> {};
+
+#if __FLT_HAS_DENORM__
+template <>
+  struct __denorm_min_impl<float>
+  { static inline constexpr float value =3D __FLT_DENORM_MIN__; };
+#endif
+
+#if __DBL_HAS_DENORM__
+template <>
+  struct __denorm_min_impl<double>
+  { static inline constexpr double value =3D __DBL_DENORM_MIN__; };
+#endif
+
+#if __LDBL_HAS_DENORM__
+template <>
+  struct __denorm_min_impl<long double>
+  { static inline constexpr long double value =3D __LDBL_DENORM_MIN__; };
+#endif
+
+template <typename _Tp>
+  struct __epsilon_impl {};
+
+template <>
+  struct __epsilon_impl<float>
+  { static inline constexpr float value =3D __FLT_EPSILON__; };
+
+template <>
+  struct __epsilon_impl<double>
+  { static inline constexpr double value =3D __DBL_EPSILON__; };
+
+template <>
+  struct __epsilon_impl<long double>
+  { static inline constexpr long double value =3D __LDBL_EPSILON__; };
+
+template <typename _Tp, bool =3D is_arithmetic_v<_Tp>>
+  struct __finite_min_impl {};
+
+template <typename _Tp>
+  struct __finite_min_impl<_Tp, true>
+  {
+    static inline constexpr _Tp value
+      =3D is_unsigned_v<_Tp> ? _Tp()
+			   : -2 * (_Tp(1) << __digits_impl<_Tp>::value - 1);
+  };
+
+template <>
+  struct __finite_min_impl<float, true>
+  { static inline constexpr float value =3D -__FLT_MAX__; };
+
+template <>
+  struct __finite_min_impl<double, true>
+  { static inline constexpr double value =3D -__DBL_MAX__; };
+
+template <>
+  struct __finite_min_impl<long double, true>
+  { static inline constexpr long double value =3D -__LDBL_MAX__; };
+
+template <typename _Tp, bool =3D is_arithmetic_v<_Tp>>
+  struct __finite_max_impl {};
+
+template <typename _Tp>
+  struct __finite_max_impl<_Tp, true>
+  { static inline constexpr _Tp value =3D ~__finite_min_impl<_Tp>::value; =
};
+
+template <>
+  struct __finite_max_impl<float, true>
+  { static inline constexpr float value =3D __FLT_MAX__; };
+
+template <>
+  struct __finite_max_impl<double, true>
+  { static inline constexpr double value =3D __DBL_MAX__; };
+
+template <>
+  struct __finite_max_impl<long double, true>
+  { static inline constexpr long double value =3D __LDBL_MAX__; };
+
+template <typename _Tp>
+  struct __infinity_impl {};
+
+#if __FLT_HAS_INFINITY__
+template <>
+  struct __infinity_impl<float>
+  { static inline constexpr float value =3D __builtin_inff(); };
+#endif
+
+#if __DBL_HAS_INFINITY__
+template <>
+  struct __infinity_impl<double>
+  { static inline constexpr double value =3D __builtin_inf(); };
+#endif
+
+#if __LDBL_HAS_INFINITY__
+template <>
+  struct __infinity_impl<long double>
+  { static inline constexpr long double value =3D __builtin_infl(); };
+#endif
+
+template <typename _Tp>
+  struct __quiet_NaN_impl {};
+
+#if __FLT_HAS_QUIET_NAN__
+template <>
+  struct __quiet_NaN_impl<float>
+  { static inline constexpr float value =3D __builtin_nanf(""); };
+#endif
+
+#if __DBL_HAS_QUIET_NAN__
+template <>
+  struct __quiet_NaN_impl<double>
+  { static inline constexpr double value =3D __builtin_nan(""); };
+#endif
+
+#if __LDBL_HAS_QUIET_NAN__
+template <>
+  struct __quiet_NaN_impl<long double>
+  { static inline constexpr long double value =3D __builtin_nanl(""); };
+#endif
+
+template <typename _Tp, bool =3D is_floating_point_v<_Tp>>
+  struct __reciprocal_overflow_threshold_impl {};
+
+template <typename _Tp>
+  struct __reciprocal_overflow_threshold_impl<_Tp, true>
+  {
+    // This typically yields a subnormal value. Is this incorrect for
+    // flush-to-zero configurations?
+    static constexpr _Tp _S_search(_Tp __ok, _Tp __overflows)
+    {
+      const _Tp __mid =3D (__ok + __overflows) / 2;
+      // 1/__mid without -ffast-math is not a constant expression if it
+      // overflows. Therefore divide 1 by the radix before division.
+      // Consequently finite_max (the threshold) must be scaled by the
+      // same value.
+      if (__mid =3D=3D __ok || __mid =3D=3D __overflows)
+	return __ok;
+      else if (_Tp(1) / (__radix_impl<_Tp>::value * __mid)
+	       <=3D __finite_max_impl<_Tp>::value / __radix_impl<_Tp>::value)
+	return _S_search(__mid, __overflows);
+      else
+	return _S_search(__ok, __mid);
+    }
+
+    static inline constexpr _Tp value
+      =3D _S_search(_Tp(1.01) / __finite_max_impl<_Tp>::value,
+		  _Tp(0.99) / __finite_max_impl<_Tp>::value);
+  };
+
+template <typename _Tp, bool =3D is_floating_point_v<_Tp>>
+  struct __round_error_impl {};
+
+template <typename _Tp>
+  struct __round_error_impl<_Tp, true>
+  { static inline constexpr _Tp value =3D 0.5; };
+
+template <typename _Tp>
+  struct __signaling_NaN_impl {};
+
+#if __FLT_HAS_QUIET_NAN__
+template <>
+  struct __signaling_NaN_impl<float>
+  { static inline constexpr float value =3D __builtin_nansf(""); };
+#endif
+
+#if __DBL_HAS_QUIET_NAN__
+template <>
+  struct __signaling_NaN_impl<double>
+  { static inline constexpr double value =3D __builtin_nans(""); };
+#endif
+
+#if __LDBL_HAS_QUIET_NAN__
+template <>
+  struct __signaling_NaN_impl<long double>
+  { static inline constexpr long double value =3D __builtin_nansl(""); };
+#endif
+
+// [num.traits.val], numeric distinguished value traits
+template <typename _Tp>
+  struct __denorm_min : __denorm_min_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __epsilon : __epsilon_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __finite_max : __finite_max_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __finite_min : __finite_min_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __infinity : __infinity_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __norm_min : __norm_min_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __quiet_NaN : __quiet_NaN_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __reciprocal_overflow_threshold
+  : __reciprocal_overflow_threshold_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __round_error : __round_error_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __signaling_NaN : __signaling_NaN_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  inline constexpr auto __denorm_min_v =3D __denorm_min<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __epsilon_v =3D __epsilon<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __finite_max_v =3D __finite_max<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __finite_min_v =3D __finite_min<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __infinity_v =3D __infinity<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __norm_min_v =3D __norm_min<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __quiet_NaN_v =3D __quiet_NaN<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __reciprocal_overflow_threshold_v
+    =3D __reciprocal_overflow_threshold<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __round_error_v =3D __round_error<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __signaling_NaN_v =3D __signaling_NaN<_Tp>::value;
+
+// [num.traits.char], numeric characteristics traits
+template <typename _Tp>
+  struct __digits : __digits_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __digits10 : __digits10_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __max_digits10 : __max_digits10_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __max_exponent : __max_exponent_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __max_exponent10 : __max_exponent10_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __min_exponent : __min_exponent_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __min_exponent10 : __min_exponent10_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  struct __radix : __radix_impl<remove_cv_t<_Tp>> {};
+
+template <typename _Tp>
+  inline constexpr auto __digits_v =3D __digits<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __digits10_v =3D __digits10<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __max_digits10_v =3D __max_digits10<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __max_exponent_v =3D __max_exponent<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __max_exponent10_v =3D __max_exponent10<_Tp>::valu=
e;
+
+template <typename _Tp>
+  inline constexpr auto __min_exponent_v =3D __min_exponent<_Tp>::value;
+
+template <typename _Tp>
+  inline constexpr auto __min_exponent10_v =3D __min_exponent10<_Tp>::valu=
e;
+
+template <typename _Tp>
+  inline constexpr auto __radix_v =3D __radix<_Tp>::value;
+
+// mkretz's extensions
+// TODO: does GCC tell me? __GCC_IEC_559 >=3D 2 is not the right answer
+template <typename _Tp>
+  struct __has_iec559_storage_format : true_type {};
+
+template <typename _Tp>
+  inline constexpr bool __has_iec559_storage_format_v
+    =3D __has_iec559_storage_format<_Tp>::value;
+
+/* To propose:
+   If __has_iec559_behavior<__quiet_NaN, T> is true the following holds:
+     - nan =3D=3D nan is false
+     - isnan(nan) is true
+     - isnan(nan + x) is true
+     - isnan(inf/inf) is true
+     - isnan(0/0) is true
+     - isunordered(nan, x) is true
+
+   If __has_iec559_behavior<__infinity, T> is true the following holds (x =
is
+   neither nan nor inf):
+     - isinf(inf) is true
+     - isinf(inf + x) is true
+     - isinf(1/0) is true
+ */
+template <template <typename> class _Trait, typename _Tp>
+  struct __has_iec559_behavior : false_type {};
+
+template <template <typename> class _Trait, typename _Tp>
+  inline constexpr bool __has_iec559_behavior_v
+    =3D __has_iec559_behavior<_Trait, _Tp>::value;
+
+#if !__FINITE_MATH_ONLY__
+#if __FLT_HAS_QUIET_NAN__
+template <>
+  struct __has_iec559_behavior<__quiet_NaN, float> : true_type {};
+#endif
+
+#if __DBL_HAS_QUIET_NAN__
+template <>
+  struct __has_iec559_behavior<__quiet_NaN, double> : true_type {};
+#endif
+
+#if __LDBL_HAS_QUIET_NAN__
+template <>
+  struct __has_iec559_behavior<__quiet_NaN, long double> : true_type {};
+#endif
+
+#if __FLT_HAS_INFINITY__
+template <>
+  struct __has_iec559_behavior<__infinity, float> : true_type {};
+#endif
+
+#if __DBL_HAS_INFINITY__
+template <>
+  struct __has_iec559_behavior<__infinity, double> : true_type {};
+#endif
+
+#if __LDBL_HAS_INFINITY__
+template <>
+  struct __has_iec559_behavior<__infinity, long double> : true_type {};
+#endif
+
+#ifdef __SUPPORT_SNAN__
+#if __FLT_HAS_QUIET_NAN__
+template <>
+  struct __has_iec559_behavior<__signaling_NaN, float> : true_type {};
+#endif
+
+#if __DBL_HAS_QUIET_NAN__
+template <>
+  struct __has_iec559_behavior<__signaling_NaN, double> : true_type {};
+#endif
+
+#if __LDBL_HAS_QUIET_NAN__
+template <>
+  struct __has_iec559_behavior<__signaling_NaN, long double> : true_type {=
};
+#endif
+
+#endif
+#endif // __FINITE_MATH_ONLY__
+
+} // namespace std
diff --git a/libstdc++-v3/include/experimental/bits/simd.h b/libstdc++-v3/i=
nclude/experimental/bits/simd.h
new file mode 100644
index 00000000000..00eec50d64f
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd.h
@@ -0,0 +1,5051 @@
+// Definition of the public simd interfaces -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_H
+#define _GLIBCXX_EXPERIMENTAL_SIMD_H
+
+#if __cplusplus >=3D 201703L
+
+#include "simd_detail.h"
+#include "numeric_traits.h"
+#include <bit>
+#include <bitset>
+#ifdef _GLIBCXX_DEBUG_UB
+#include <cstdio> // for stderr
+#endif
+#include <cstring>
+#include <functional>
+#include <iosfwd>
+#include <utility>
+
+#if _GLIBCXX_SIMD_X86INTRIN
+#include <x86intrin.h>
+#elif _GLIBCXX_SIMD_HAVE_NEON
+#include <arm_neon.h>
+#endif
+
+/* There are several closely related types, with the following naming
+ * convention:
+ * _Tp: vectorizable (arithmetic) type (or any type)
+ * _TV: __vector_type_t<_Tp, _Np>
+ * _TW: _SimdWrapper<_Tp, _Np>
+ * _TI: __intrinsic_type_t<_Tp, _Np>
+ * _TVT: _VectorTraits<_TV> or _VectorTraits<_TW>
+ * If one additional type is needed use _U instead of _T.
+ * Otherwise use _T\d, _TV\d, _TW\d, TI\d, _TVT\d.
+ *
+ * More naming conventions:
+ * _Ap or _Abi: An ABI tag from the simd_abi namespace
+ * _Ip: often used for integer types with sizeof(_Ip) =3D=3D sizeof(_Tp),
+ *      _IV, _IW as for _TV, _TW
+ * _Np: number of elements (not bytes)
+ * _Bytes: number of bytes
+ *
+ * Variable names:
+ * __k: mask object (vector- or bitmask)
+ */
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+
+#if !_GLIBCXX_SIMD_X86INTRIN
+using __m128  [[__gnu__::__vector_size__(16)]] =3D float;
+using __m128d [[__gnu__::__vector_size__(16)]] =3D double;
+using __m128i [[__gnu__::__vector_size__(16)]] =3D long long;
+using __m256  [[__gnu__::__vector_size__(32)]] =3D float;
+using __m256d [[__gnu__::__vector_size__(32)]] =3D double;
+using __m256i [[__gnu__::__vector_size__(32)]] =3D long long;
+using __m512  [[__gnu__::__vector_size__(64)]] =3D float;
+using __m512d [[__gnu__::__vector_size__(64)]] =3D double;
+using __m512i [[__gnu__::__vector_size__(64)]] =3D long long;
+#endif
+
+namespace simd_abi {
+// simd_abi forward declarations {{{
+// implementation details:
+struct _Scalar;
+
+template <int _Np>
+  struct _Fixed;
+
+// There are two major ABIs that appear on different architectures.
+// Both have non-boolean values packed into an N Byte register
+// -> #elements =3D N / sizeof(T)
+// Masks differ:
+// 1. Use value vector registers for masks (all 0 or all 1)
+// 2. Use bitmasks (mask registers) with one bit per value in the correspo=
nding
+//    value vector
+//
+// Both can be partially used, masking off the rest when doing horizontal
+// operations or operations that can trap (e.g. FP_INVALID or integer divi=
sion
+// by 0). This is encoded as the number of used bytes.
+template <int _UsedBytes>
+  struct _VecBuiltin;
+
+template <int _UsedBytes>
+  struct _VecBltnBtmsk;
+
+template <typename _Tp, int _Np>
+  using _VecN =3D _VecBuiltin<sizeof(_Tp) * _Np>;
+
+template <int _UsedBytes =3D 16>
+  using _Sse =3D _VecBuiltin<_UsedBytes>;
+
+template <int _UsedBytes =3D 32>
+  using _Avx =3D _VecBuiltin<_UsedBytes>;
+
+template <int _UsedBytes =3D 64>
+  using _Avx512 =3D _VecBltnBtmsk<_UsedBytes>;
+
+template <int _UsedBytes =3D 16>
+  using _Neon =3D _VecBuiltin<_UsedBytes>;
+
+// implementation-defined:
+using __sse =3D _Sse<>;
+using __avx =3D _Avx<>;
+using __avx512 =3D _Avx512<>;
+using __neon =3D _Neon<>;
+using __neon128 =3D _Neon<16>;
+using __neon64 =3D _Neon<8>;
+
+// standard:
+template <typename _Tp, size_t _Np, typename...>
+  struct deduce;
+
+template <int _Np>
+  using fixed_size =3D _Fixed<_Np>;
+
+using scalar =3D _Scalar;
+
+// }}}
+} // namespace simd_abi
+// forward declarations is_simd(_mask), simd(_mask), simd_size {{{
+template <typename _Tp>
+  struct is_simd;
+
+template <typename _Tp>
+  struct is_simd_mask;
+
+template <typename _Tp, typename _Abi>
+  class simd;
+
+template <typename _Tp, typename _Abi>
+  class simd_mask;
+
+template <typename _Tp, typename _Abi>
+  struct simd_size;
+
+// }}}
+// load/store flags {{{
+struct element_aligned_tag
+{
+  template <typename _Tp, typename _Up =3D typename _Tp::value_type>
+    static constexpr size_t _S_alignment =3D alignof(_Up);
+
+  template <typename _Tp, typename _Up>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _Up*
+    _S_apply(_Up* __ptr)
+    { return __ptr; }
+};
+
+struct vector_aligned_tag
+{
+  template <typename _Tp, typename _Up =3D typename _Tp::value_type>
+    static constexpr size_t _S_alignment
+      =3D std::__bit_ceil(sizeof(_Up) * _Tp::size());
+
+  template <typename _Tp, typename _Up>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _Up*
+    _S_apply(_Up* __ptr)
+    {
+      return static_cast<_Up*>(
+	__builtin_assume_aligned(__ptr, _S_alignment<_Tp, _Up>));
+    }
+};
+
+template <size_t _Np> struct overaligned_tag
+{
+  template <typename _Tp, typename _Up =3D typename _Tp::value_type>
+    static constexpr size_t _S_alignment =3D _Np;
+
+  template <typename _Tp, typename _Up>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _Up*
+    _S_apply(_Up* __ptr)
+    { return static_cast<_Up*>(__builtin_assume_aligned(__ptr, _Np)); }
+};
+
+inline constexpr element_aligned_tag element_aligned =3D {};
+
+inline constexpr vector_aligned_tag vector_aligned =3D {};
+
+template <size_t _Np>
+  inline constexpr overaligned_tag<_Np> overaligned =3D {};
+
+// }}}
+template <size_t _X>
+  using _SizeConstant =3D integral_constant<size_t, _X>;
+
+// unrolled/pack execution helpers
+// __execute_n_times{{{
+template <typename _Fp, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __execute_on_index_sequence(_Fp&& __f, index_sequence<_I...>)
+  { ((void)__f(_SizeConstant<_I>()), ...); }
+
+template <typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __execute_on_index_sequence(_Fp&&, index_sequence<>)
+  { }
+
+template <size_t _Np, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __execute_n_times(_Fp&& __f)
+  {
+    __execute_on_index_sequence(static_cast<_Fp&&>(__f),
+				make_index_sequence<_Np>{});
+  }
+
+// }}}
+// __generate_from_n_evaluations{{{
+template <typename _R, typename _Fp, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _R
+  __execute_on_index_sequence_with_return(_Fp&& __f, index_sequence<_I...>)
+  { return _R{__f(_SizeConstant<_I>())...}; }
+
+template <size_t _Np, typename _R, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _R
+  __generate_from_n_evaluations(_Fp&& __f)
+  {
+    return __execute_on_index_sequence_with_return<_R>(
+      static_cast<_Fp&&>(__f), make_index_sequence<_Np>{});
+  }
+
+// }}}
+// __call_with_n_evaluations{{{
+template <size_t... _I, typename _F0, typename _FArgs>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __call_with_n_evaluations(index_sequence<_I...>, _F0&& __f0, _FArgs&& __=
fargs)
+  { return __f0(__fargs(_SizeConstant<_I>())...); }
+
+template <size_t _Np, typename _F0, typename _FArgs>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __call_with_n_evaluations(_F0&& __f0, _FArgs&& __fargs)
+  {
+    return __call_with_n_evaluations(make_index_sequence<_Np>{},
+				     static_cast<_F0&&>(__f0),
+				     static_cast<_FArgs&&>(__fargs));
+  }
+
+// }}}
+// __call_with_subscripts{{{
+template <size_t _First =3D 0, size_t... _It, typename _Tp, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __call_with_subscripts(_Tp&& __x, index_sequence<_It...>, _Fp&& __fun)
+  { return __fun(__x[_First + _It]...); }
+
+template <size_t _Np, size_t _First =3D 0, typename _Tp, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __call_with_subscripts(_Tp&& __x, _Fp&& __fun)
+  {
+    return __call_with_subscripts<_First>(static_cast<_Tp&&>(__x),
+					  make_index_sequence<_Np>(),
+					  static_cast<_Fp&&>(__fun));
+  }
+
+// }}}
+
+// vvv ---- type traits ---- vvv
+// integer type aliases{{{
+using _UChar =3D unsigned char;
+using _SChar =3D signed char;
+using _UShort =3D unsigned short;
+using _UInt =3D unsigned int;
+using _ULong =3D unsigned long;
+using _ULLong =3D unsigned long long;
+using _LLong =3D long long;
+
+//}}}
+// __first_of_pack{{{
+template <typename _T0, typename...>
+  struct __first_of_pack
+  { using type =3D _T0; };
+
+template <typename... _Ts>
+  using __first_of_pack_t =3D typename __first_of_pack<_Ts...>::type;
+
+//}}}
+// __value_type_or_identity_t {{{
+template <typename _Tp>
+  typename _Tp::value_type
+  __value_type_or_identity_impl(int);
+
+template <typename _Tp>
+  _Tp
+  __value_type_or_identity_impl(float);
+
+template <typename _Tp>
+  using __value_type_or_identity_t
+    =3D decltype(__value_type_or_identity_impl<_Tp>(int()));
+
+// }}}
+// __is_vectorizable {{{
+template <typename _Tp>
+  struct __is_vectorizable : public is_arithmetic<_Tp> {};
+
+template <>
+  struct __is_vectorizable<bool> : public false_type {};
+
+template <typename _Tp>
+  inline constexpr bool __is_vectorizable_v =3D __is_vectorizable<_Tp>::va=
lue;
+
+// Deduces to a vectorizable type
+template <typename _Tp, typename =3D enable_if_t<__is_vectorizable_v<_Tp>>>
+  using _Vectorizable =3D _Tp;
+
+// }}}
+// _LoadStorePtr / __is_possible_loadstore_conversion {{{
+template <typename _Ptr, typename _ValueType>
+  struct __is_possible_loadstore_conversion
+  : conjunction<__is_vectorizable<_Ptr>, __is_vectorizable<_ValueType>> {};
+
+template <>
+  struct __is_possible_loadstore_conversion<bool, bool> : true_type {};
+
+// Deduces to a type allowed for load/store with the given value type.
+template <typename _Ptr, typename _ValueType,
+	  typename =3D enable_if_t<
+	    __is_possible_loadstore_conversion<_Ptr, _ValueType>::value>>
+  using _LoadStorePtr =3D _Ptr;
+
+// }}}
+// __is_bitmask{{{
+template <typename _Tp, typename =3D void_t<>>
+  struct __is_bitmask : false_type {};
+
+template <typename _Tp>
+  inline constexpr bool __is_bitmask_v =3D __is_bitmask<_Tp>::value;
+
+// the __mmaskXX case:
+template <typename _Tp>
+  struct __is_bitmask<_Tp,
+    void_t<decltype(declval<unsigned&>() =3D declval<_Tp>() & 1u)>>
+  : true_type {};
+
+// }}}
+// __int_for_sizeof{{{
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wpedantic"
+template <size_t _Bytes>
+  constexpr auto
+  __int_for_sizeof()
+  {
+    if constexpr (_Bytes =3D=3D sizeof(int))
+      return int();
+  #ifdef __clang__
+    else if constexpr (_Bytes =3D=3D sizeof(char))
+      return char();
+  #else
+    else if constexpr (_Bytes =3D=3D sizeof(_SChar))
+      return _SChar();
+  #endif
+    else if constexpr (_Bytes =3D=3D sizeof(short))
+      return short();
+  #ifndef __clang__
+    else if constexpr (_Bytes =3D=3D sizeof(long))
+      return long();
+  #endif
+    else if constexpr (_Bytes =3D=3D sizeof(_LLong))
+      return _LLong();
+  #ifdef __SIZEOF_INT128__
+    else if constexpr (_Bytes =3D=3D sizeof(__int128))
+      return __int128();
+  #endif // __SIZEOF_INT128__
+    else if constexpr (_Bytes % sizeof(int) =3D=3D 0)
+      {
+	constexpr size_t _Np =3D _Bytes / sizeof(int);
+	struct _Ip
+	{
+	  int _M_data[_Np];
+
+	  _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
+	  operator&(_Ip __rhs) const
+	  {
+	    return __generate_from_n_evaluations<_Np, _Ip>(
+	      [&](auto __i) { return __rhs._M_data[__i] & _M_data[__i]; });
+	  }
+
+	  _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
+	  operator|(_Ip __rhs) const
+	  {
+	    return __generate_from_n_evaluations<_Np, _Ip>(
+	      [&](auto __i) { return __rhs._M_data[__i] | _M_data[__i]; });
+	  }
+
+	  _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
+	  operator^(_Ip __rhs) const
+	  {
+	    return __generate_from_n_evaluations<_Np, _Ip>(
+	      [&](auto __i) { return __rhs._M_data[__i] ^ _M_data[__i]; });
+	  }
+
+	  _GLIBCXX_SIMD_INTRINSIC constexpr _Ip
+	  operator~() const
+	  {
+	    return __generate_from_n_evaluations<_Np, _Ip>(
+	      [&](auto __i) { return ~_M_data[__i]; });
+	  }
+	};
+	return _Ip{};
+      }
+    else
+      static_assert(_Bytes !=3D _Bytes, "this should be unreachable");
+  }
+#pragma GCC diagnostic pop
+
+template <typename _Tp>
+  using __int_for_sizeof_t =3D decltype(__int_for_sizeof<sizeof(_Tp)>());
+
+template <size_t _Np>
+  using __int_with_sizeof_t =3D decltype(__int_for_sizeof<_Np>());
+
+// }}}
+// __is_fixed_size_abi{{{
+template <typename _Tp>
+  struct __is_fixed_size_abi : false_type {};
+
+template <int _Np>
+  struct __is_fixed_size_abi<simd_abi::fixed_size<_Np>> : true_type {};
+
+template <typename _Tp>
+  inline constexpr bool __is_fixed_size_abi_v =3D __is_fixed_size_abi<_Tp>=
::value;
+
+// }}}
+// constexpr feature detection{{{
+constexpr inline bool __have_mmx =3D _GLIBCXX_SIMD_HAVE_MMX;
+constexpr inline bool __have_sse =3D _GLIBCXX_SIMD_HAVE_SSE;
+constexpr inline bool __have_sse2 =3D _GLIBCXX_SIMD_HAVE_SSE2;
+constexpr inline bool __have_sse3 =3D _GLIBCXX_SIMD_HAVE_SSE3;
+constexpr inline bool __have_ssse3 =3D _GLIBCXX_SIMD_HAVE_SSSE3;
+constexpr inline bool __have_sse4_1 =3D _GLIBCXX_SIMD_HAVE_SSE4_1;
+constexpr inline bool __have_sse4_2 =3D _GLIBCXX_SIMD_HAVE_SSE4_2;
+constexpr inline bool __have_xop =3D _GLIBCXX_SIMD_HAVE_XOP;
+constexpr inline bool __have_avx =3D _GLIBCXX_SIMD_HAVE_AVX;
+constexpr inline bool __have_avx2 =3D _GLIBCXX_SIMD_HAVE_AVX2;
+constexpr inline bool __have_bmi =3D _GLIBCXX_SIMD_HAVE_BMI1;
+constexpr inline bool __have_bmi2 =3D _GLIBCXX_SIMD_HAVE_BMI2;
+constexpr inline bool __have_lzcnt =3D _GLIBCXX_SIMD_HAVE_LZCNT;
+constexpr inline bool __have_sse4a =3D _GLIBCXX_SIMD_HAVE_SSE4A;
+constexpr inline bool __have_fma =3D _GLIBCXX_SIMD_HAVE_FMA;
+constexpr inline bool __have_fma4 =3D _GLIBCXX_SIMD_HAVE_FMA4;
+constexpr inline bool __have_f16c =3D _GLIBCXX_SIMD_HAVE_F16C;
+constexpr inline bool __have_popcnt =3D _GLIBCXX_SIMD_HAVE_POPCNT;
+constexpr inline bool __have_avx512f =3D _GLIBCXX_SIMD_HAVE_AVX512F;
+constexpr inline bool __have_avx512dq =3D _GLIBCXX_SIMD_HAVE_AVX512DQ;
+constexpr inline bool __have_avx512vl =3D _GLIBCXX_SIMD_HAVE_AVX512VL;
+constexpr inline bool __have_avx512bw =3D _GLIBCXX_SIMD_HAVE_AVX512BW;
+constexpr inline bool __have_avx512dq_vl =3D __have_avx512dq && __have_avx=
512vl;
+constexpr inline bool __have_avx512bw_vl =3D __have_avx512bw && __have_avx=
512vl;
+
+constexpr inline bool __have_neon =3D _GLIBCXX_SIMD_HAVE_NEON;
+constexpr inline bool __have_neon_a32 =3D _GLIBCXX_SIMD_HAVE_NEON_A32;
+constexpr inline bool __have_neon_a64 =3D _GLIBCXX_SIMD_HAVE_NEON_A64;
+constexpr inline bool __support_neon_float =3D
+#if defined __GCC_IEC_559
+  __GCC_IEC_559 =3D=3D 0;
+#elif defined __FAST_MATH__
+  true;
+#else
+  false;
+#endif
+
+#ifdef __POWER9_VECTOR__
+constexpr inline bool __have_power9vec =3D true;
+#else
+constexpr inline bool __have_power9vec =3D false;
+#endif
+#if defined __POWER8_VECTOR__
+constexpr inline bool __have_power8vec =3D true;
+#else
+constexpr inline bool __have_power8vec =3D __have_power9vec;
+#endif
+#if defined __VSX__
+constexpr inline bool __have_power_vsx =3D true;
+#else
+constexpr inline bool __have_power_vsx =3D __have_power8vec;
+#endif
+#if defined __ALTIVEC__
+constexpr inline bool __have_power_vmx =3D true;
+#else
+constexpr inline bool __have_power_vmx =3D __have_power_vsx;
+#endif
+
+// }}}
+// __is_scalar_abi {{{
+template <typename _Abi>
+  constexpr bool
+  __is_scalar_abi()
+  { return is_same_v<simd_abi::scalar, _Abi>; }
+
+// }}}
+// __abi_bytes_v {{{
+template <template <int> class _Abi, int _Bytes>
+  constexpr int
+  __abi_bytes_impl(_Abi<_Bytes>*)
+  { return _Bytes; }
+
+template <typename _Tp>
+  constexpr int
+  __abi_bytes_impl(_Tp*)
+  { return -1; }
+
+template <typename _Abi>
+  inline constexpr int __abi_bytes_v
+    =3D __abi_bytes_impl(static_cast<_Abi*>(nullptr));
+
+// }}}
+// __is_builtin_bitmask_abi {{{
+template <typename _Abi>
+  constexpr bool
+  __is_builtin_bitmask_abi()
+  { return is_same_v<simd_abi::_VecBltnBtmsk<__abi_bytes_v<_Abi>>, _Abi>; }
+
+// }}}
+// __is_sse_abi {{{
+template <typename _Abi>
+  constexpr bool
+  __is_sse_abi()
+  {
+    constexpr auto _Bytes =3D __abi_bytes_v<_Abi>;
+    return _Bytes <=3D 16 && is_same_v<simd_abi::_VecBuiltin<_Bytes>, _Abi=
>;
+  }
+
+// }}}
+// __is_avx_abi {{{
+template <typename _Abi>
+  constexpr bool
+  __is_avx_abi()
+  {
+    constexpr auto _Bytes =3D __abi_bytes_v<_Abi>;
+    return _Bytes > 16 && _Bytes <=3D 32
+	   && is_same_v<simd_abi::_VecBuiltin<_Bytes>, _Abi>;
+  }
+
+// }}}
+// __is_avx512_abi {{{
+template <typename _Abi>
+  constexpr bool
+  __is_avx512_abi()
+  {
+    constexpr auto _Bytes =3D __abi_bytes_v<_Abi>;
+    return _Bytes <=3D 64 && is_same_v<simd_abi::_Avx512<_Bytes>, _Abi>;
+  }
+
+// }}}
+// __is_neon_abi {{{
+template <typename _Abi>
+  constexpr bool
+  __is_neon_abi()
+  {
+    constexpr auto _Bytes =3D __abi_bytes_v<_Abi>;
+    return _Bytes <=3D 16 && is_same_v<simd_abi::_VecBuiltin<_Bytes>, _Abi=
>;
+  }
+
+// }}}
+// __make_dependent_t {{{
+template <typename, typename _Up>
+  struct __make_dependent
+  { using type =3D _Up; };
+
+template <typename _Tp, typename _Up>
+  using __make_dependent_t =3D typename __make_dependent<_Tp, _Up>::type;
+
+// }}}
+// ^^^ ---- type traits ---- ^^^
+
+// __invoke_ub{{{
+template <typename... _Args>
+  [[noreturn]] _GLIBCXX_SIMD_ALWAYS_INLINE void
+  __invoke_ub([[maybe_unused]] const char* __msg,
+	      [[maybe_unused]] const _Args&... __args)
+  {
+#ifdef _GLIBCXX_DEBUG_UB
+    __builtin_fprintf(stderr, __msg, __args...);
+    __builtin_trap();
+#else
+    __builtin_unreachable();
+#endif
+  }
+
+// }}}
+// __assert_unreachable{{{
+template <typename _Tp>
+  struct __assert_unreachable
+  { static_assert(!is_same_v<_Tp, _Tp>, "this should be unreachable"); };
+
+// }}}
+// __size_or_zero_v {{{
+template <typename _Tp, typename _Ap, size_t _Np =3D simd_size<_Tp, _Ap>::=
value>
+  constexpr size_t
+  __size_or_zero_dispatch(int)
+  { return _Np; }
+
+template <typename _Tp, typename _Ap>
+  constexpr size_t
+  __size_or_zero_dispatch(float)
+  { return 0; }
+
+template <typename _Tp, typename _Ap>
+  inline constexpr size_t __size_or_zero_v
+     =3D __size_or_zero_dispatch<_Tp, _Ap>(0);
+
+// }}}
+// __div_roundup {{{
+inline constexpr size_t
+__div_roundup(size_t __a, size_t __b)
+{ return (__a + __b - 1) / __b; }
+
+// }}}
+// _ExactBool{{{
+class _ExactBool
+{
+  const bool _M_data;
+
+public:
+  _GLIBCXX_SIMD_INTRINSIC constexpr _ExactBool(bool __b) : _M_data(__b) {}
+
+  _ExactBool(int) =3D delete;
+
+  _GLIBCXX_SIMD_INTRINSIC constexpr operator bool() const { return _M_data=
; }
+};
+
+// }}}
+// __may_alias{{{
+/**@internal
+ * Helper __may_alias<_Tp> that turns _Tp into the type to be used for an
+ * aliasing pointer. This adds the __may_alias attribute to _Tp (with comp=
ilers
+ * that support it).
+ */
+template <typename _Tp>
+  using __may_alias [[__gnu__::__may_alias__]] =3D _Tp;
+
+// }}}
+// _UnsupportedBase {{{
+// simd and simd_mask base for unsupported <_Tp, _Abi>
+struct _UnsupportedBase
+{
+  _UnsupportedBase() =3D delete;
+  _UnsupportedBase(const _UnsupportedBase&) =3D delete;
+  _UnsupportedBase& operator=3D(const _UnsupportedBase&) =3D delete;
+  ~_UnsupportedBase() =3D delete;
+};
+
+// }}}
+// _InvalidTraits {{{
+/**
+ * @internal
+ * Defines the implementation of __a given <_Tp, _Abi>.
+ *
+ * Implementations must ensure that only valid <_Tp, _Abi> instantiations =
are
+ * possible. Static assertions in the type definition do not suffice. It is
+ * important that SFINAE works.
+ */
+struct _InvalidTraits
+{
+  using _IsValid =3D false_type;
+  using _SimdBase =3D _UnsupportedBase;
+  using _MaskBase =3D _UnsupportedBase;
+
+  static constexpr size_t _S_full_size =3D 0;
+  static constexpr bool _S_is_partial =3D false;
+
+  static constexpr size_t _S_simd_align =3D 1;
+  struct _SimdImpl;
+  struct _SimdMember {};
+  struct _SimdCastType;
+
+  static constexpr size_t _S_mask_align =3D 1;
+  struct _MaskImpl;
+  struct _MaskMember {};
+  struct _MaskCastType;
+};
+
+// }}}
+// _SimdTraits {{{
+template <typename _Tp, typename _Abi, typename =3D void_t<>>
+  struct _SimdTraits : _InvalidTraits {};
+
+// }}}
+// __private_init, __bitset_init{{{
+/**
+ * @internal
+ * Tag used for private init constructor of simd and simd_mask
+ */
+inline constexpr struct _PrivateInit {} __private_init =3D {};
+
+inline constexpr struct _BitsetInit {} __bitset_init =3D {};
+
+// }}}
+// __is_narrowing_conversion<_From, _To>{{{
+template <typename _From, typename _To, bool =3D is_arithmetic_v<_From>,
+	  bool =3D is_arithmetic_v<_To>>
+  struct __is_narrowing_conversion;
+
+// ignore "signed/unsigned mismatch" in the following trait.
+// The implicit conversions will do the right thing here.
+template <typename _From, typename _To>
+  struct __is_narrowing_conversion<_From, _To, true, true>
+  : public __bool_constant<(
+      __digits_v<_From> > __digits_v<_To>
+      || __finite_max_v<_From> > __finite_max_v<_To>
+      || __finite_min_v<_From> < __finite_min_v<_To>
+      || (is_signed_v<_From> && is_unsigned_v<_To>))> {};
+
+template <typename _Tp>
+  struct __is_narrowing_conversion<_Tp, bool, true, true>
+  : public true_type {};
+
+template <>
+  struct __is_narrowing_conversion<bool, bool, true, true>
+  : public false_type {};
+
+template <typename _Tp>
+  struct __is_narrowing_conversion<_Tp, _Tp, true, true>
+  : public false_type {};
+
+template <typename _From, typename _To>
+  struct __is_narrowing_conversion<_From, _To, false, true>
+  : public negation<is_convertible<_From, _To>> {};
+
+// }}}
+// __converts_to_higher_integer_rank{{{
+template <typename _From, typename _To, bool =3D (sizeof(_From) < sizeof(_=
To))>
+  struct __converts_to_higher_integer_rank : public true_type {};
+
+// this may fail for char -> short if sizeof(char) =3D=3D sizeof(short)
+template <typename _From, typename _To>
+  struct __converts_to_higher_integer_rank<_From, _To, false>
+  : public is_same<decltype(declval<_From>() + declval<_To>()), _To> {};
+
+// }}}
+// __data(simd/simd_mask) {{{
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
+  __data(const simd<_Tp, _Ap>& __x);
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+  __data(simd<_Tp, _Ap>& __x);
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
+  __data(const simd_mask<_Tp, _Ap>& __x);
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+  __data(simd_mask<_Tp, _Ap>& __x);
+
+// }}}
+// _SimdConverter {{{
+template <typename _FromT, typename _FromA, typename _ToT, typename _ToA,
+	  typename =3D void>
+  struct _SimdConverter;
+
+template <typename _Tp, typename _Ap>
+  struct _SimdConverter<_Tp, _Ap, _Tp, _Ap, void>
+  {
+    template <typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC const _Up&
+      operator()(const _Up& __x)
+      { return __x; }
+  };
+
+// }}}
+// __to_value_type_or_member_type {{{
+template <typename _V>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __to_value_type_or_member_type(const _V& __x) -> decltype(__data(__x))
+  { return __data(__x); }
+
+template <typename _V>
+  _GLIBCXX_SIMD_INTRINSIC constexpr const typename _V::value_type&
+  __to_value_type_or_member_type(const typename _V::value_type& __x)
+  { return __x; }
+
+// }}}
+// __bool_storage_member_type{{{
+template <size_t _Size>
+  struct __bool_storage_member_type;
+
+template <size_t _Size>
+  using __bool_storage_member_type_t =3D
+    typename __bool_storage_member_type<_Size>::type;
+
+// }}}
+// _SimdTuple {{{
+// why not tuple?
+// 1. tuple gives no guarantee about the storage order, but I require
+// storage
+//    equivalent to array<_Tp, _Np>
+// 2. direct access to the element type (first template argument)
+// 3. enforces equal element type, only different _Abi types are allowed
+template <typename _Tp, typename... _Abis>
+  struct _SimdTuple;
+
+//}}}
+// __fixed_size_storage_t {{{
+template <typename _Tp, int _Np>
+  struct __fixed_size_storage;
+
+template <typename _Tp, int _Np>
+  using __fixed_size_storage_t =3D typename __fixed_size_storage<_Tp, _Np>=
::type;
+
+// }}}
+// _SimdWrapper fwd decl{{{
+template <typename _Tp, size_t _Size, typename =3D void_t<>>
+  struct _SimdWrapper;
+
+template <typename _Tp>
+  using _SimdWrapper8 =3D _SimdWrapper<_Tp, 8 / sizeof(_Tp)>;
+template <typename _Tp>
+  using _SimdWrapper16 =3D _SimdWrapper<_Tp, 16 / sizeof(_Tp)>;
+template <typename _Tp>
+  using _SimdWrapper32 =3D _SimdWrapper<_Tp, 32 / sizeof(_Tp)>;
+template <typename _Tp>
+  using _SimdWrapper64 =3D _SimdWrapper<_Tp, 64 / sizeof(_Tp)>;
+
+// }}}
+// __is_simd_wrapper {{{
+template <typename _Tp>
+  struct __is_simd_wrapper : false_type {};
+
+template <typename _Tp, size_t _Np>
+  struct __is_simd_wrapper<_SimdWrapper<_Tp, _Np>> : true_type {};
+
+template <typename _Tp>
+  inline constexpr bool __is_simd_wrapper_v =3D __is_simd_wrapper<_Tp>::va=
lue;
+
+// }}}
+// _BitOps {{{
+struct _BitOps
+{
+  // _S_bit_iteration {{{
+  template <typename _Tp, typename _Fp>
+    static void
+    _S_bit_iteration(_Tp __mask, _Fp&& __f)
+    {
+      static_assert(sizeof(_ULLong) >=3D sizeof(_Tp));
+      conditional_t<sizeof(_Tp) <=3D sizeof(_UInt), _UInt, _ULLong> __k;
+      if constexpr (is_convertible_v<_Tp, decltype(__k)>)
+	__k =3D __mask;
+      else
+	__k =3D __mask.to_ullong();
+      while(__k)
+	{
+	  __f(std::__countr_zero(__k));
+	  __k &=3D (__k - 1);
+	}
+    }
+
+  //}}}
+};
+
+//}}}
+// __increment, __decrement {{{
+template <typename _Tp =3D void>
+  struct __increment
+  { constexpr _Tp operator()(_Tp __a) const { return ++__a; } };
+
+template <>
+  struct __increment<void>
+  {
+    template <typename _Tp>
+      constexpr _Tp
+      operator()(_Tp __a) const
+      { return ++__a; }
+  };
+
+template <typename _Tp =3D void>
+  struct __decrement
+  { constexpr _Tp operator()(_Tp __a) const { return --__a; } };
+
+template <>
+  struct __decrement<void>
+  {
+    template <typename _Tp>
+      constexpr _Tp
+      operator()(_Tp __a) const
+      { return --__a; }
+  };
+
+// }}}
+// _ValuePreserving(OrInt) {{{
+template <typename _From, typename _To,
+	  typename =3D enable_if_t<negation<
+	    __is_narrowing_conversion<__remove_cvref_t<_From>, _To>>::value>>
+  using _ValuePreserving =3D _From;
+
+template <typename _From, typename _To,
+	  typename _DecayedFrom =3D __remove_cvref_t<_From>,
+	  typename =3D enable_if_t<conjunction<
+	    is_convertible<_From, _To>,
+	    disjunction<
+	      is_same<_DecayedFrom, _To>, is_same<_DecayedFrom, int>,
+	      conjunction<is_same<_DecayedFrom, _UInt>, is_unsigned<_To>>,
+	      negation<__is_narrowing_conversion<_DecayedFrom, _To>>>>::value>>
+  using _ValuePreservingOrInt =3D _From;
+
+// }}}
+// __intrinsic_type {{{
+template <typename _Tp, size_t _Bytes, typename =3D void_t<>>
+  struct __intrinsic_type;
+
+template <typename _Tp, size_t _Size>
+  using __intrinsic_type_t =3D
+    typename __intrinsic_type<_Tp, _Size * sizeof(_Tp)>::type;
+
+template <typename _Tp>
+  using __intrinsic_type2_t =3D typename __intrinsic_type<_Tp, 2>::type;
+template <typename _Tp>
+  using __intrinsic_type4_t =3D typename __intrinsic_type<_Tp, 4>::type;
+template <typename _Tp>
+  using __intrinsic_type8_t =3D typename __intrinsic_type<_Tp, 8>::type;
+template <typename _Tp>
+  using __intrinsic_type16_t =3D typename __intrinsic_type<_Tp, 16>::type;
+template <typename _Tp>
+  using __intrinsic_type32_t =3D typename __intrinsic_type<_Tp, 32>::type;
+template <typename _Tp>
+  using __intrinsic_type64_t =3D typename __intrinsic_type<_Tp, 64>::type;
+
+// }}}
+// _BitMask {{{
+template <size_t _Np, bool _Sanitized =3D false>
+  struct _BitMask;
+
+template <size_t _Np, bool _Sanitized>
+  struct __is_bitmask<_BitMask<_Np, _Sanitized>, void> : true_type {};
+
+template <size_t _Np>
+  using _SanitizedBitMask =3D _BitMask<_Np, true>;
+
+template <size_t _Np, bool _Sanitized>
+  struct _BitMask
+  {
+    static_assert(_Np > 0);
+
+    static constexpr size_t _NBytes =3D __div_roundup(_Np, __CHAR_BIT__);
+
+    using _Tp =3D conditional_t<_Np =3D=3D 1, bool,
+			      make_unsigned_t<__int_with_sizeof_t<std::min(
+				sizeof(_ULLong), std::__bit_ceil(_NBytes))>>>;
+
+    static constexpr int _S_array_size =3D __div_roundup(_NBytes, sizeof(_=
Tp));
+
+    _Tp _M_bits[_S_array_size];
+
+    static constexpr int _S_unused_bits
+      =3D _Np =3D=3D 1 ? 0 : _S_array_size * sizeof(_Tp) * __CHAR_BIT__ - =
_Np;
+
+    static constexpr _Tp _S_bitmask =3D +_Tp(~_Tp()) >> _S_unused_bits;
+
+    constexpr _BitMask() noexcept =3D default;
+
+    constexpr _BitMask(unsigned long long __x) noexcept
+      : _M_bits{static_cast<_Tp>(__x)} {}
+
+    _BitMask(bitset<_Np> __x) noexcept : _BitMask(__x.to_ullong()) {}
+
+    constexpr _BitMask(const _BitMask&) noexcept =3D default;
+
+    template <bool _RhsSanitized, typename =3D enable_if_t<_RhsSanitized =
=3D=3D false
+							 && _Sanitized =3D=3D true>>
+      constexpr _BitMask(const _BitMask<_Np, _RhsSanitized>& __rhs) noexce=
pt
+	: _BitMask(__rhs._M_sanitized()) {}
+
+    constexpr operator _SimdWrapper<bool, _Np>() const noexcept
+    {
+      static_assert(_S_array_size =3D=3D 1);
+      return _M_bits[0];
+    }
+
+    // precondition: is sanitized
+    constexpr _Tp
+    _M_to_bits() const noexcept
+    {
+      static_assert(_S_array_size =3D=3D 1);
+      return _M_bits[0];
+    }
+
+    // precondition: is sanitized
+    constexpr unsigned long long
+    to_ullong() const noexcept
+    {
+      static_assert(_S_array_size =3D=3D 1);
+      return _M_bits[0];
+    }
+
+    // precondition: is sanitized
+    constexpr unsigned long
+    to_ulong() const noexcept
+    {
+      static_assert(_S_array_size =3D=3D 1);
+      return _M_bits[0];
+    }
+
+    constexpr bitset<_Np>
+    _M_to_bitset() const noexcept
+    {
+      static_assert(_S_array_size =3D=3D 1);
+      return _M_bits[0];
+    }
+
+    constexpr decltype(auto)
+    _M_sanitized() const noexcept
+    {
+      if constexpr (_Sanitized)
+	return *this;
+      else if constexpr (_Np =3D=3D 1)
+	return _SanitizedBitMask<_Np>(_M_bits[0]);
+      else
+	{
+	  _SanitizedBitMask<_Np> __r =3D {};
+	  for (int __i =3D 0; __i < _S_array_size; ++__i)
+	    __r._M_bits[__i] =3D _M_bits[__i];
+	  if constexpr (_S_unused_bits > 0)
+	    __r._M_bits[_S_array_size - 1] &=3D _S_bitmask;
+	  return __r;
+	}
+    }
+
+    template <size_t _Mp, bool _LSanitized>
+      constexpr _BitMask<_Np + _Mp, _Sanitized>
+      _M_prepend(_BitMask<_Mp, _LSanitized> __lsb) const noexcept
+      {
+	constexpr size_t _RN =3D _Np + _Mp;
+	using _Rp =3D _BitMask<_RN, _Sanitized>;
+	if constexpr (_Rp::_S_array_size =3D=3D 1)
+	  {
+	    _Rp __r{{_M_bits[0]}};
+	    __r._M_bits[0] <<=3D _Mp;
+	    __r._M_bits[0] |=3D __lsb._M_sanitized()._M_bits[0];
+	    return __r;
+	  }
+	else
+	  __assert_unreachable<_Rp>();
+      }
+
+    // Return a new _BitMask with size _NewSize while dropping _DropLsb le=
ast
+    // significant bits. If the operation implicitly produces a sanitized =
bitmask,
+    // the result type will have _Sanitized set.
+    template <size_t _DropLsb, size_t _NewSize =3D _Np - _DropLsb>
+      constexpr auto
+      _M_extract() const noexcept
+      {
+	static_assert(_Np > _DropLsb);
+	static_assert(_DropLsb + _NewSize <=3D sizeof(_ULLong) * __CHAR_BIT__,
+		      "not implemented for bitmasks larger than one ullong");
+	if constexpr (_NewSize =3D=3D 1)
+	  // must sanitize because the return _Tp is bool
+	  return _SanitizedBitMask<1>(_M_bits[0] & (_Tp(1) << _DropLsb));
+	else
+	  return _BitMask<_NewSize,
+			  ((_NewSize + _DropLsb =3D=3D sizeof(_Tp) * __CHAR_BIT__
+			    && _NewSize + _DropLsb <=3D _Np)
+			   || ((_Sanitized || _Np =3D=3D sizeof(_Tp) * __CHAR_BIT__)
+			       && _NewSize + _DropLsb >=3D _Np))>(_M_bits[0]
+								>> _DropLsb);
+      }
+
+    // True if all bits are set. Implicitly sanitizes if _Sanitized =3D=3D=
 false.
+    constexpr bool
+    all() const noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	return _M_bits[0];
+      else if constexpr (!_Sanitized)
+	return _M_sanitized().all();
+      else
+	{
+	  constexpr _Tp __allbits =3D ~_Tp();
+	  for (int __i =3D 0; __i < _S_array_size - 1; ++__i)
+	    if (_M_bits[__i] !=3D __allbits)
+	      return false;
+	  return _M_bits[_S_array_size - 1] =3D=3D _S_bitmask;
+	}
+    }
+
+    // True if at least one bit is set. Implicitly sanitizes if _Sanitized=
 =3D=3D
+    // false.
+    constexpr bool
+    any() const noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	return _M_bits[0];
+      else if constexpr (!_Sanitized)
+	return _M_sanitized().any();
+      else
+	{
+	  for (int __i =3D 0; __i < _S_array_size - 1; ++__i)
+	    if (_M_bits[__i] !=3D 0)
+	      return true;
+	  return _M_bits[_S_array_size - 1] !=3D 0;
+	}
+    }
+
+    // True if no bit is set. Implicitly sanitizes if _Sanitized =3D=3D fa=
lse.
+    constexpr bool
+    none() const noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	return !_M_bits[0];
+      else if constexpr (!_Sanitized)
+	return _M_sanitized().none();
+      else
+	{
+	  for (int __i =3D 0; __i < _S_array_size - 1; ++__i)
+	    if (_M_bits[__i] !=3D 0)
+	      return false;
+	  return _M_bits[_S_array_size - 1] =3D=3D 0;
+	}
+    }
+
+    // Returns the number of set bits. Implicitly sanitizes if _Sanitized =
=3D=3D
+    // false.
+    constexpr int
+    count() const noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	return _M_bits[0];
+      else if constexpr (!_Sanitized)
+	return _M_sanitized().none();
+      else
+	{
+	  int __result =3D __builtin_popcountll(_M_bits[0]);
+	  for (int __i =3D 1; __i < _S_array_size; ++__i)
+	    __result +=3D __builtin_popcountll(_M_bits[__i]);
+	  return __result;
+	}
+    }
+
+    // Returns the bit at offset __i as bool.
+    constexpr bool
+    operator[](size_t __i) const noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	return _M_bits[0];
+      else if constexpr (_S_array_size =3D=3D 1)
+	return (_M_bits[0] >> __i) & 1;
+      else
+	{
+	  const size_t __j =3D __i / (sizeof(_Tp) * __CHAR_BIT__);
+	  const size_t __shift =3D __i % (sizeof(_Tp) * __CHAR_BIT__);
+	  return (_M_bits[__j] >> __shift) & 1;
+	}
+    }
+
+    template <size_t __i>
+      constexpr bool
+      operator[](_SizeConstant<__i>) const noexcept
+      {
+	static_assert(__i < _Np);
+	constexpr size_t __j =3D __i / (sizeof(_Tp) * __CHAR_BIT__);
+	constexpr size_t __shift =3D __i % (sizeof(_Tp) * __CHAR_BIT__);
+	return static_cast<bool>(_M_bits[__j] & (_Tp(1) << __shift));
+      }
+
+    // Set the bit at offset __i to __x.
+    constexpr void
+    set(size_t __i, bool __x) noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	_M_bits[0] =3D __x;
+      else if constexpr (_S_array_size =3D=3D 1)
+	{
+	  _M_bits[0] &=3D ~_Tp(_Tp(1) << __i);
+	  _M_bits[0] |=3D _Tp(_Tp(__x) << __i);
+	}
+      else
+	{
+	  const size_t __j =3D __i / (sizeof(_Tp) * __CHAR_BIT__);
+	  const size_t __shift =3D __i % (sizeof(_Tp) * __CHAR_BIT__);
+	  _M_bits[__j] &=3D ~_Tp(_Tp(1) << __shift);
+	  _M_bits[__j] |=3D _Tp(_Tp(__x) << __shift);
+	}
+    }
+
+    template <size_t __i>
+      constexpr void
+      set(_SizeConstant<__i>, bool __x) noexcept
+      {
+	static_assert(__i < _Np);
+	if constexpr (_Np =3D=3D 1)
+	  _M_bits[0] =3D __x;
+	else
+	  {
+	    constexpr size_t __j =3D __i / (sizeof(_Tp) * __CHAR_BIT__);
+	    constexpr size_t __shift =3D __i % (sizeof(_Tp) * __CHAR_BIT__);
+	    constexpr _Tp __mask =3D ~_Tp(_Tp(1) << __shift);
+	    _M_bits[__j] &=3D __mask;
+	    _M_bits[__j] |=3D _Tp(_Tp(__x) << __shift);
+	  }
+      }
+
+    // Inverts all bits. Sanitized input leads to sanitized output.
+    constexpr _BitMask
+    operator~() const noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	return !_M_bits[0];
+      else
+	{
+	  _BitMask __result{};
+	  for (int __i =3D 0; __i < _S_array_size - 1; ++__i)
+	    __result._M_bits[__i] =3D ~_M_bits[__i];
+	  if constexpr (_Sanitized)
+	    __result._M_bits[_S_array_size - 1]
+	      =3D _M_bits[_S_array_size - 1] ^ _S_bitmask;
+	  else
+	    __result._M_bits[_S_array_size - 1] =3D ~_M_bits[_S_array_size - 1];
+	  return __result;
+	}
+    }
+
+    constexpr _BitMask&
+    operator^=3D(const _BitMask& __b) & noexcept
+    {
+      __execute_n_times<_S_array_size>(
+	[&](auto __i) { _M_bits[__i] ^=3D __b._M_bits[__i]; });
+      return *this;
+    }
+
+    constexpr _BitMask&
+    operator|=3D(const _BitMask& __b) & noexcept
+    {
+      __execute_n_times<_S_array_size>(
+	[&](auto __i) { _M_bits[__i] |=3D __b._M_bits[__i]; });
+      return *this;
+    }
+
+    constexpr _BitMask&
+    operator&=3D(const _BitMask& __b) & noexcept
+    {
+      __execute_n_times<_S_array_size>(
+	[&](auto __i) { _M_bits[__i] &=3D __b._M_bits[__i]; });
+      return *this;
+    }
+
+    friend constexpr _BitMask
+    operator^(const _BitMask& __a, const _BitMask& __b) noexcept
+    {
+      _BitMask __r =3D __a;
+      __r ^=3D __b;
+      return __r;
+    }
+
+    friend constexpr _BitMask
+    operator|(const _BitMask& __a, const _BitMask& __b) noexcept
+    {
+      _BitMask __r =3D __a;
+      __r |=3D __b;
+      return __r;
+    }
+
+    friend constexpr _BitMask
+    operator&(const _BitMask& __a, const _BitMask& __b) noexcept
+    {
+      _BitMask __r =3D __a;
+      __r &=3D __b;
+      return __r;
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC
+    constexpr bool
+    _M_is_constprop() const
+    {
+      if constexpr (_S_array_size =3D=3D 0)
+	return __builtin_constant_p(_M_bits[0]);
+      else
+	{
+	  for (int __i =3D 0; __i < _S_array_size; ++__i)
+	    if (!__builtin_constant_p(_M_bits[__i]))
+	      return false;
+	  return true;
+	}
+    }
+  };
+
+// }}}
+
+// vvv ---- builtin vector types [[gnu::vector_size(N)]] and operations --=
=2D- vvv
+// __min_vector_size {{{
+template <typename _Tp =3D void>
+  static inline constexpr int __min_vector_size =3D 2 * sizeof(_Tp);
+
+#if _GLIBCXX_SIMD_HAVE_NEON
+template <>
+  inline constexpr int __min_vector_size<void> =3D 8;
+#else
+template <>
+  inline constexpr int __min_vector_size<void> =3D 16;
+#endif
+
+// }}}
+// __vector_type {{{
+template <typename _Tp, size_t _Np, typename =3D void>
+  struct __vector_type_n {};
+
+// substition failure for 0-element case
+template <typename _Tp>
+  struct __vector_type_n<_Tp, 0, void> {};
+
+// special case 1-element to be _Tp itself
+template <typename _Tp>
+  struct __vector_type_n<_Tp, 1, enable_if_t<__is_vectorizable_v<_Tp>>>
+  { using type =3D _Tp; };
+
+// else, use GNU-style builtin vector types
+template <typename _Tp, size_t _Np>
+  struct __vector_type_n<_Tp, _Np,
+			 enable_if_t<__is_vectorizable_v<_Tp> && _Np >=3D 2>>
+  {
+    static constexpr size_t _S_Np2 =3D std::__bit_ceil(_Np * sizeof(_Tp));
+
+    static constexpr size_t _S_Bytes =3D
+#ifdef __i386__
+      // Using [[gnu::vector_size(8)]] would wreak havoc on the FPU because
+      // those objects are passed via MMX registers and nothing ever calls=
 EMMS.
+      _S_Np2 =3D=3D 8 ? 16 :
+#endif
+      _S_Np2 < __min_vector_size<_Tp> ? __min_vector_size<_Tp>
+				      : _S_Np2;
+
+    using type [[__gnu__::__vector_size__(_S_Bytes)]] =3D _Tp;
+  };
+
+template <typename _Tp, size_t _Bytes, size_t =3D _Bytes % sizeof(_Tp)>
+  struct __vector_type;
+
+template <typename _Tp, size_t _Bytes>
+  struct __vector_type<_Tp, _Bytes, 0>
+  : __vector_type_n<_Tp, _Bytes / sizeof(_Tp)> {};
+
+template <typename _Tp, size_t _Size>
+  using __vector_type_t =3D typename __vector_type_n<_Tp, _Size>::type;
+
+template <typename _Tp>
+  using __vector_type2_t =3D typename __vector_type<_Tp, 2>::type;
+template <typename _Tp>
+  using __vector_type4_t =3D typename __vector_type<_Tp, 4>::type;
+template <typename _Tp>
+  using __vector_type8_t =3D typename __vector_type<_Tp, 8>::type;
+template <typename _Tp>
+  using __vector_type16_t =3D typename __vector_type<_Tp, 16>::type;
+template <typename _Tp>
+  using __vector_type32_t =3D typename __vector_type<_Tp, 32>::type;
+template <typename _Tp>
+  using __vector_type64_t =3D typename __vector_type<_Tp, 64>::type;
+
+// }}}
+// __is_vector_type {{{
+template <typename _Tp, typename =3D void_t<>>
+  struct __is_vector_type : false_type {};
+
+template <typename _Tp>
+  struct __is_vector_type<
+    _Tp, void_t<typename __vector_type<
+	   remove_reference_t<decltype(declval<_Tp>()[0])>, sizeof(_Tp)>::type>>
+    : is_same<_Tp, typename __vector_type<
+		     remove_reference_t<decltype(declval<_Tp>()[0])>,
+		     sizeof(_Tp)>::type> {};
+
+template <typename _Tp>
+  inline constexpr bool __is_vector_type_v =3D __is_vector_type<_Tp>::valu=
e;
+
+// }}}
+// _VectorTraits{{{
+template <typename _Tp, typename =3D void_t<>>
+  struct _VectorTraitsImpl;
+
+template <typename _Tp>
+  struct _VectorTraitsImpl<_Tp, enable_if_t<__is_vector_type_v<_Tp>>>
+  {
+    using type =3D _Tp;
+    using value_type =3D remove_reference_t<decltype(declval<_Tp>()[0])>;
+    static constexpr int _S_full_size =3D sizeof(_Tp) / sizeof(value_type);
+    using _Wrapper =3D _SimdWrapper<value_type, _S_full_size>;
+    template <typename _Up, int _W =3D _S_full_size>
+      static constexpr bool _S_is
+	=3D is_same_v<value_type, _Up> && _W =3D=3D _S_full_size;
+  };
+
+template <typename _Tp, size_t _Np>
+  struct _VectorTraitsImpl<_SimdWrapper<_Tp, _Np>,
+			   void_t<__vector_type_t<_Tp, _Np>>>
+  {
+    using type =3D __vector_type_t<_Tp, _Np>;
+    using value_type =3D _Tp;
+    static constexpr int _S_full_size =3D sizeof(type) / sizeof(value_type=
);
+    using _Wrapper =3D _SimdWrapper<_Tp, _Np>;
+    static constexpr bool _S_is_partial =3D (_Np =3D=3D _S_full_size);
+    static constexpr int _S_partial_width =3D _Np;
+    template <typename _Up, int _W =3D _S_full_size>
+      static constexpr bool _S_is
+	=3D is_same_v<value_type, _Up>&& _W =3D=3D _S_full_size;
+  };
+
+template <typename _Tp, typename =3D typename _VectorTraitsImpl<_Tp>::type>
+  using _VectorTraits =3D _VectorTraitsImpl<_Tp>;
+
+// }}}
+// __as_vector{{{
+template <typename _V>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __as_vector(_V __x)
+  {
+    if constexpr (__is_vector_type_v<_V>)
+      return __x;
+    else if constexpr (is_simd<_V>::value || is_simd_mask<_V>::value)
+      return __data(__x)._M_data;
+    else if constexpr (__is_vectorizable_v<_V>)
+      return __vector_type_t<_V, 2>{__x};
+    else
+      return __x._M_data;
+  }
+
+// }}}
+// __as_wrapper{{{
+template <size_t _Np =3D 0, typename _V>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __as_wrapper(_V __x)
+  {
+    if constexpr (__is_vector_type_v<_V>)
+      return _SimdWrapper<typename _VectorTraits<_V>::value_type,
+			  (_Np > 0 ? _Np : _VectorTraits<_V>::_S_full_size)>(__x);
+    else if constexpr (is_simd<_V>::value || is_simd_mask<_V>::value)
+      {
+	static_assert(_V::size() =3D=3D _Np);
+	return __data(__x);
+      }
+    else
+      {
+	static_assert(_V::_S_size =3D=3D _Np);
+	return __x;
+      }
+  }
+
+// }}}
+// __intrin_bitcast{{{
+template <typename _To, typename _From>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __intrin_bitcast(_From __v)
+  {
+    static_assert(__is_vector_type_v<_From> && __is_vector_type_v<_To>);
+    if constexpr (sizeof(_To) =3D=3D sizeof(_From))
+      return reinterpret_cast<_To>(__v);
+    else if constexpr (sizeof(_From) > sizeof(_To))
+      if constexpr (sizeof(_To) >=3D 16)
+	return reinterpret_cast<const __may_alias<_To>&>(__v);
+      else
+	{
+	  _To __r;
+	  __builtin_memcpy(&__r, &__v, sizeof(_To));
+	  return __r;
+	}
+#if _GLIBCXX_SIMD_X86INTRIN && !defined __clang__
+    else if constexpr (__have_avx && sizeof(_From) =3D=3D 16 && sizeof(_To=
) =3D=3D 32)
+      return reinterpret_cast<_To>(__builtin_ia32_ps256_ps(
+	reinterpret_cast<__vector_type_t<float, 4>>(__v)));
+    else if constexpr (__have_avx512f && sizeof(_From) =3D=3D 16
+		       && sizeof(_To) =3D=3D 64)
+      return reinterpret_cast<_To>(__builtin_ia32_ps512_ps(
+	reinterpret_cast<__vector_type_t<float, 4>>(__v)));
+    else if constexpr (__have_avx512f && sizeof(_From) =3D=3D 32
+		       && sizeof(_To) =3D=3D 64)
+      return reinterpret_cast<_To>(__builtin_ia32_ps512_256ps(
+	reinterpret_cast<__vector_type_t<float, 8>>(__v)));
+#endif // _GLIBCXX_SIMD_X86INTRIN
+    else if constexpr (sizeof(__v) <=3D 8)
+      return reinterpret_cast<_To>(
+	__vector_type_t<__int_for_sizeof_t<_From>, sizeof(_To) / sizeof(_From)>{
+	  reinterpret_cast<__int_for_sizeof_t<_From>>(__v)});
+    else
+      {
+	static_assert(sizeof(_To) > sizeof(_From));
+	_To __r =3D {};
+	__builtin_memcpy(&__r, &__v, sizeof(_From));
+	return __r;
+      }
+  }
+
+// }}}
+// __vector_bitcast{{{
+template <typename _To, size_t _NN =3D 0, typename _From,
+	  typename _FromVT =3D _VectorTraits<_From>,
+	  size_t _Np =3D _NN =3D=3D 0 ? sizeof(_From) / sizeof(_To) : _NN>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_To, _Np>
+  __vector_bitcast(_From __x)
+  {
+    using _R =3D __vector_type_t<_To, _Np>;
+    return __intrin_bitcast<_R>(__x);
+  }
+
+template <typename _To, size_t _NN =3D 0, typename _Tp, size_t _Nx,
+	  size_t _Np
+	  =3D _NN =3D=3D 0 ? sizeof(_SimdWrapper<_Tp, _Nx>) / sizeof(_To) : _NN>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_To, _Np>
+  __vector_bitcast(const _SimdWrapper<_Tp, _Nx>& __x)
+  {
+    static_assert(_Np > 1);
+    return __intrin_bitcast<__vector_type_t<_To, _Np>>(__x._M_data);
+  }
+
+// }}}
+// __convert_x86 declarations {{{
+#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
+template <typename _To, typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _To __convert_x86(_Tp);
+
+template <typename _To, typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _To __convert_x86(_Tp, _Tp);
+
+template <typename _To, typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _To __convert_x86(_Tp, _Tp, _Tp, _Tp);
+
+template <typename _To, typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _To __convert_x86(_Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp);
+
+template <typename _To, typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _To __convert_x86(_Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp, _Tp,=
 _Tp,
+		    _Tp, _Tp, _Tp, _Tp);
+#endif // _GLIBCXX_SIMD_WORKAROUND_PR85048
+
+//}}}
+// __bit_cast {{{
+template <typename _To, typename _From>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __bit_cast(const _From __x)
+  {
+    // TODO: implement with / replace by __builtin_bit_cast ASAP
+    static_assert(sizeof(_To) =3D=3D sizeof(_From));
+    constexpr bool __to_is_vectorizable
+      =3D is_arithmetic_v<_To> || is_enum_v<_To>;
+    constexpr bool __from_is_vectorizable
+      =3D is_arithmetic_v<_From> || is_enum_v<_From>;
+    if constexpr (__is_vector_type_v<_To> && __is_vector_type_v<_From>)
+      return reinterpret_cast<_To>(__x);
+    else if constexpr (__is_vector_type_v<_To> && __from_is_vectorizable)
+      {
+	using _FV [[gnu::vector_size(sizeof(_From))]] =3D _From;
+	return reinterpret_cast<_To>(_FV{__x});
+      }
+    else if constexpr (__to_is_vectorizable && __from_is_vectorizable)
+      {
+	using _TV [[gnu::vector_size(sizeof(_To))]] =3D _To;
+	using _FV [[gnu::vector_size(sizeof(_From))]] =3D _From;
+	return reinterpret_cast<_TV>(_FV{__x})[0];
+      }
+    else if constexpr (__to_is_vectorizable && __is_vector_type_v<_From>)
+      {
+	using _TV [[gnu::vector_size(sizeof(_To))]] =3D _To;
+	return reinterpret_cast<_TV>(__x)[0];
+      }
+    else
+      {
+	_To __r;
+	__builtin_memcpy(reinterpret_cast<char*>(&__r),
+			 reinterpret_cast<const char*>(&__x), sizeof(_To));
+	return __r;
+      }
+  }
+
+// }}}
+// __to_intrin {{{
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>,
+	  typename _R
+	  =3D __intrinsic_type_t<typename _TVT::value_type, _TVT::_S_full_size>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _R
+  __to_intrin(_Tp __x)
+  {
+    static_assert(sizeof(__x) <=3D sizeof(_R),
+		  "__to_intrin may never drop values off the end");
+    if constexpr (sizeof(__x) =3D=3D sizeof(_R))
+      return reinterpret_cast<_R>(__as_vector(__x));
+    else
+      {
+	using _Up =3D __int_for_sizeof_t<_Tp>;
+	return reinterpret_cast<_R>(
+	  __vector_type_t<_Up, sizeof(_R) / sizeof(_Up)>{__bit_cast<_Up>(__x)});
+      }
+  }
+
+// }}}
+// __make_vector{{{
+template <typename _Tp, typename... _Args>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, sizeof...(_Args)>
+  __make_vector(const _Args&... __args)
+  {
+    return __vector_type_t<_Tp, sizeof...(_Args)>{static_cast<_Tp>(__args)=
=2E..};
+  }
+
+// }}}
+// __vector_broadcast{{{
+template <size_t _Np, typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, _Np>
+  __vector_broadcast(_Tp __x)
+  {
+    return __call_with_n_evaluations<_Np>(
+      [](auto... __xx) { return __vector_type_t<_Tp, _Np>{__xx...}; },
+      [&__x](int) { return __x; });
+  }
+
+// }}}
+// __generate_vector{{{
+  template <typename _Tp, size_t _Np, typename _Gp, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, _Np>
+  __generate_vector_impl(_Gp&& __gen, index_sequence<_I...>)
+  {
+    return __vector_type_t<_Tp, _Np>{
+      static_cast<_Tp>(__gen(_SizeConstant<_I>()))...};
+  }
+
+template <typename _V, typename _VVT =3D _VectorTraits<_V>, typename _Gp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _V
+  __generate_vector(_Gp&& __gen)
+  {
+    if constexpr (__is_vector_type_v<_V>)
+      return __generate_vector_impl<typename _VVT::value_type,
+				    _VVT::_S_full_size>(
+	static_cast<_Gp&&>(__gen), make_index_sequence<_VVT::_S_full_size>());
+    else
+      return __generate_vector_impl<typename _VVT::value_type,
+				    _VVT::_S_partial_width>(
+	static_cast<_Gp&&>(__gen),
+	make_index_sequence<_VVT::_S_partial_width>());
+  }
+
+template <typename _Tp, size_t _Np, typename _Gp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Tp, _Np>
+  __generate_vector(_Gp&& __gen)
+  {
+    return __generate_vector_impl<_Tp, _Np>(static_cast<_Gp&&>(__gen),
+					    make_index_sequence<_Np>());
+  }
+
+// }}}
+// __xor{{{
+template <typename _TW>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _TW
+  __xor(_TW __a, _TW __b) noexcept
+  {
+    if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
+      {
+	using _Tp =3D typename conditional_t<__is_simd_wrapper_v<_TW>, _TW,
+					   _VectorTraitsImpl<_TW>>::value_type;
+	if constexpr (is_floating_point_v<_Tp>)
+	  {
+	    using _Ip =3D make_unsigned_t<__int_for_sizeof_t<_Tp>>;
+	    return __vector_bitcast<_Tp>(__vector_bitcast<_Ip>(__a)
+					 ^ __vector_bitcast<_Ip>(__b));
+	  }
+	else if constexpr (__is_vector_type_v<_TW>)
+	  return __a ^ __b;
+	else
+	  return __a._M_data ^ __b._M_data;
+      }
+    else
+      return __a ^ __b;
+  }
+
+// }}}
+// __or{{{
+template <typename _TW>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _TW
+  __or(_TW __a, _TW __b) noexcept
+  {
+    if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
+      {
+	using _Tp =3D typename conditional_t<__is_simd_wrapper_v<_TW>, _TW,
+					   _VectorTraitsImpl<_TW>>::value_type;
+	if constexpr (is_floating_point_v<_Tp>)
+	  {
+	    using _Ip =3D make_unsigned_t<__int_for_sizeof_t<_Tp>>;
+	    return __vector_bitcast<_Tp>(__vector_bitcast<_Ip>(__a)
+					 | __vector_bitcast<_Ip>(__b));
+	  }
+	else if constexpr (__is_vector_type_v<_TW>)
+	  return __a | __b;
+	else
+	  return __a._M_data | __b._M_data;
+      }
+    else
+      return __a | __b;
+  }
+
+// }}}
+// __and{{{
+template <typename _TW>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _TW
+  __and(_TW __a, _TW __b) noexcept
+  {
+    if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
+      {
+	using _Tp =3D typename conditional_t<__is_simd_wrapper_v<_TW>, _TW,
+					   _VectorTraitsImpl<_TW>>::value_type;
+	if constexpr (is_floating_point_v<_Tp>)
+	  {
+	    using _Ip =3D make_unsigned_t<__int_for_sizeof_t<_Tp>>;
+	    return __vector_bitcast<_Tp>(__vector_bitcast<_Ip>(__a)
+					 & __vector_bitcast<_Ip>(__b));
+	  }
+	else if constexpr (__is_vector_type_v<_TW>)
+	  return __a & __b;
+	else
+	  return __a._M_data & __b._M_data;
+      }
+    else
+      return __a & __b;
+  }
+
+// }}}
+// __andnot{{{
+#if _GLIBCXX_SIMD_X86INTRIN && !defined __clang__
+static constexpr struct
+{
+  _GLIBCXX_SIMD_INTRINSIC __v4sf
+  operator()(__v4sf __a, __v4sf __b) const noexcept
+  { return __builtin_ia32_andnps(__a, __b); }
+
+  _GLIBCXX_SIMD_INTRINSIC __v2df
+  operator()(__v2df __a, __v2df __b) const noexcept
+  { return __builtin_ia32_andnpd(__a, __b); }
+
+  _GLIBCXX_SIMD_INTRINSIC __v2di
+  operator()(__v2di __a, __v2di __b) const noexcept
+  { return __builtin_ia32_pandn128(__a, __b); }
+
+  _GLIBCXX_SIMD_INTRINSIC __v8sf
+  operator()(__v8sf __a, __v8sf __b) const noexcept
+  { return __builtin_ia32_andnps256(__a, __b); }
+
+  _GLIBCXX_SIMD_INTRINSIC __v4df
+  operator()(__v4df __a, __v4df __b) const noexcept
+  { return __builtin_ia32_andnpd256(__a, __b); }
+
+  _GLIBCXX_SIMD_INTRINSIC __v4di
+  operator()(__v4di __a, __v4di __b) const noexcept
+  {
+    if constexpr (__have_avx2)
+      return __builtin_ia32_andnotsi256(__a, __b);
+    else
+      return reinterpret_cast<__v4di>(
+	__builtin_ia32_andnpd256(reinterpret_cast<__v4df>(__a),
+				 reinterpret_cast<__v4df>(__b)));
+  }
+
+  _GLIBCXX_SIMD_INTRINSIC __v16sf
+  operator()(__v16sf __a, __v16sf __b) const noexcept
+  {
+    if constexpr (__have_avx512dq)
+      return _mm512_andnot_ps(__a, __b);
+    else
+      return reinterpret_cast<__v16sf>(
+	_mm512_andnot_si512(reinterpret_cast<__v8di>(__a),
+			    reinterpret_cast<__v8di>(__b)));
+  }
+
+  _GLIBCXX_SIMD_INTRINSIC __v8df
+  operator()(__v8df __a, __v8df __b) const noexcept
+  {
+    if constexpr (__have_avx512dq)
+      return _mm512_andnot_pd(__a, __b);
+    else
+      return reinterpret_cast<__v8df>(
+	_mm512_andnot_si512(reinterpret_cast<__v8di>(__a),
+			    reinterpret_cast<__v8di>(__b)));
+  }
+
+  _GLIBCXX_SIMD_INTRINSIC __v8di
+  operator()(__v8di __a, __v8di __b) const noexcept
+  { return _mm512_andnot_si512(__a, __b); }
+} _S_x86_andnot;
+#endif // _GLIBCXX_SIMD_X86INTRIN && !__clang__
+
+template <typename _TW>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _TW
+  __andnot(_TW __a, _TW __b) noexcept
+  {
+    if constexpr (__is_vector_type_v<_TW> || __is_simd_wrapper_v<_TW>)
+      {
+	using _TVT =3D conditional_t<__is_simd_wrapper_v<_TW>, _TW,
+				   _VectorTraitsImpl<_TW>>;
+	using _Tp =3D typename _TVT::value_type;
+#if _GLIBCXX_SIMD_X86INTRIN && !defined __clang__
+	if constexpr (sizeof(_TW) >=3D 16)
+	  {
+	    const auto __ai =3D __to_intrin(__a);
+	    const auto __bi =3D __to_intrin(__b);
+	    if (!__builtin_is_constant_evaluated()
+		&& !(__builtin_constant_p(__ai) && __builtin_constant_p(__bi)))
+	      {
+		const auto __r =3D _S_x86_andnot(__ai, __bi);
+		if constexpr (is_convertible_v<decltype(__r), _TW>)
+		  return __r;
+		else
+		  return reinterpret_cast<typename _TVT::type>(__r);
+	      }
+	  }
+#endif // _GLIBCXX_SIMD_X86INTRIN
+	using _Ip =3D make_unsigned_t<__int_for_sizeof_t<_Tp>>;
+	return __vector_bitcast<_Tp>(~__vector_bitcast<_Ip>(__a)
+				     & __vector_bitcast<_Ip>(__b));
+      }
+    else
+      return ~__a & __b;
+  }
+
+// }}}
+// __not{{{
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
+  __not(_Tp __a) noexcept
+  {
+    if constexpr (is_floating_point_v<typename _TVT::value_type>)
+      return reinterpret_cast<typename _TVT::type>(
+	~__vector_bitcast<unsigned>(__a));
+    else
+      return ~__a;
+  }
+
+// }}}
+// __concat{{{
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>,
+	  typename _R =3D __vector_type_t<typename _TVT::value_type,
+					_TVT::_S_full_size * 2>>
+  constexpr _R
+  __concat(_Tp a_, _Tp b_)
+  {
+#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_1
+    using _W
+      =3D conditional_t<is_floating_point_v<typename _TVT::value_type>, do=
uble,
+		      conditional_t<(sizeof(_Tp) >=3D 2 * sizeof(long long)),
+				    long long, typename _TVT::value_type>>;
+    constexpr int input_width =3D sizeof(_Tp) / sizeof(_W);
+    const auto __a =3D __vector_bitcast<_W>(a_);
+    const auto __b =3D __vector_bitcast<_W>(b_);
+    using _Up =3D __vector_type_t<_W, sizeof(_R) / sizeof(_W)>;
+#else
+    constexpr int input_width =3D _TVT::_S_full_size;
+    const _Tp& __a =3D a_;
+    const _Tp& __b =3D b_;
+    using _Up =3D _R;
+#endif
+    if constexpr (input_width =3D=3D 2)
+      return reinterpret_cast<_R>(_Up{__a[0], __a[1], __b[0], __b[1]});
+    else if constexpr (input_width =3D=3D 4)
+      return reinterpret_cast<_R>(
+	_Up{__a[0], __a[1], __a[2], __a[3], __b[0], __b[1], __b[2], __b[3]});
+    else if constexpr (input_width =3D=3D 8)
+      return reinterpret_cast<_R>(
+	_Up{__a[0], __a[1], __a[2], __a[3], __a[4], __a[5], __a[6], __a[7],
+	    __b[0], __b[1], __b[2], __b[3], __b[4], __b[5], __b[6], __b[7]});
+    else if constexpr (input_width =3D=3D 16)
+      return reinterpret_cast<_R>(
+	_Up{__a[0],  __a[1],  __a[2],  __a[3],  __a[4],  __a[5],  __a[6],
+	    __a[7],  __a[8],  __a[9],  __a[10], __a[11], __a[12], __a[13],
+	    __a[14], __a[15], __b[0],  __b[1],  __b[2],  __b[3],  __b[4],
+	    __b[5],  __b[6],  __b[7],  __b[8],  __b[9],  __b[10], __b[11],
+	    __b[12], __b[13], __b[14], __b[15]});
+    else if constexpr (input_width =3D=3D 32)
+      return reinterpret_cast<_R>(
+	_Up{__a[0],  __a[1],  __a[2],  __a[3],  __a[4],  __a[5],  __a[6],
+	    __a[7],  __a[8],  __a[9],  __a[10], __a[11], __a[12], __a[13],
+	    __a[14], __a[15], __a[16], __a[17], __a[18], __a[19], __a[20],
+	    __a[21], __a[22], __a[23], __a[24], __a[25], __a[26], __a[27],
+	    __a[28], __a[29], __a[30], __a[31], __b[0],  __b[1],  __b[2],
+	    __b[3],  __b[4],  __b[5],  __b[6],  __b[7],  __b[8],  __b[9],
+	    __b[10], __b[11], __b[12], __b[13], __b[14], __b[15], __b[16],
+	    __b[17], __b[18], __b[19], __b[20], __b[21], __b[22], __b[23],
+	    __b[24], __b[25], __b[26], __b[27], __b[28], __b[29], __b[30],
+	    __b[31]});
+  }
+
+// }}}
+// __zero_extend {{{
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  struct _ZeroExtendProxy
+  {
+    using value_type =3D typename _TVT::value_type;
+    static constexpr size_t _Np =3D _TVT::_S_full_size;
+    const _Tp __x;
+
+    template <typename _To, typename _ToVT =3D _VectorTraits<_To>,
+	      typename
+	      =3D enable_if_t<is_same_v<typename _ToVT::value_type, value_type>>>
+      _GLIBCXX_SIMD_INTRINSIC operator _To() const
+      {
+	constexpr size_t _ToN =3D _ToVT::_S_full_size;
+	if constexpr (_ToN =3D=3D _Np)
+	  return __x;
+	else if constexpr (_ToN =3D=3D 2 * _Np)
+	  {
+#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_3
+	    if constexpr (__have_avx && _TVT::template _S_is<float, 4>)
+	      return __vector_bitcast<value_type>(
+		_mm256_insertf128_ps(__m256(), __x, 0));
+	    else if constexpr (__have_avx && _TVT::template _S_is<double, 2>)
+	      return __vector_bitcast<value_type>(
+		_mm256_insertf128_pd(__m256d(), __x, 0));
+	    else if constexpr (__have_avx2 && _Np * sizeof(value_type) =3D=3D 16)
+	      return __vector_bitcast<value_type>(
+		_mm256_insertf128_si256(__m256i(), __to_intrin(__x), 0));
+	    else if constexpr (__have_avx512f && _TVT::template _S_is<float, 8>)
+	      {
+		if constexpr (__have_avx512dq)
+		  return __vector_bitcast<value_type>(
+		    _mm512_insertf32x8(__m512(), __x, 0));
+		else
+		  return reinterpret_cast<__m512>(
+		    _mm512_insertf64x4(__m512d(),
+				       reinterpret_cast<__m256d>(__x), 0));
+	      }
+	    else if constexpr (__have_avx512f
+			       && _TVT::template _S_is<double, 4>)
+	      return __vector_bitcast<value_type>(
+		_mm512_insertf64x4(__m512d(), __x, 0));
+	    else if constexpr (__have_avx512f && _Np * sizeof(value_type) =3D=3D =
32)
+	      return __vector_bitcast<value_type>(
+		_mm512_inserti64x4(__m512i(), __to_intrin(__x), 0));
+#endif
+	    return __concat(__x, _Tp());
+	  }
+	else if constexpr (_ToN =3D=3D 4 * _Np)
+	  {
+#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_3
+	    if constexpr (__have_avx512dq && _TVT::template _S_is<double, 2>)
+	      {
+		return __vector_bitcast<value_type>(
+		  _mm512_insertf64x2(__m512d(), __x, 0));
+	      }
+	    else if constexpr (__have_avx512f
+			       && is_floating_point_v<value_type>)
+	      {
+		return __vector_bitcast<value_type>(
+		  _mm512_insertf32x4(__m512(), reinterpret_cast<__m128>(__x),
+				     0));
+	      }
+	    else if constexpr (__have_avx512f && _Np * sizeof(value_type) =3D=3D =
16)
+	      {
+		return __vector_bitcast<value_type>(
+		  _mm512_inserti32x4(__m512i(), __to_intrin(__x), 0));
+	      }
+#endif
+	    return __concat(__concat(__x, _Tp()),
+			    __vector_type_t<value_type, _Np * 2>());
+	  }
+	else if constexpr (_ToN =3D=3D 8 * _Np)
+	  return __concat(operator __vector_type_t<value_type, _Np * 4>(),
+			  __vector_type_t<value_type, _Np * 4>());
+	else if constexpr (_ToN =3D=3D 16 * _Np)
+	  return __concat(operator __vector_type_t<value_type, _Np * 8>(),
+			  __vector_type_t<value_type, _Np * 8>());
+	else
+	  __assert_unreachable<_Tp>();
+      }
+  };
+
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC _ZeroExtendProxy<_Tp, _TVT>
+  __zero_extend(_Tp __x)
+  { return {__x}; }
+
+// }}}
+// __extract<_Np, By>{{{
+template <int _Offset,
+	  int _SplitBy,
+	  typename _Tp,
+	  typename _TVT =3D _VectorTraits<_Tp>,
+	  typename _R =3D __vector_type_t<typename _TVT::value_type,
+			  _TVT::_S_full_size / _SplitBy>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _R
+  __extract(_Tp __in)
+  {
+    using value_type =3D typename _TVT::value_type;
+#if _GLIBCXX_SIMD_X86INTRIN // {{{
+    if constexpr (sizeof(_Tp) =3D=3D 64 && _SplitBy =3D=3D 4 && _Offset > =
0)
+      {
+	if constexpr (__have_avx512dq && is_same_v<double, value_type>)
+	  return _mm512_extractf64x2_pd(__to_intrin(__in), _Offset);
+	else if constexpr (is_floating_point_v<value_type>)
+	  return __vector_bitcast<value_type>(
+	    _mm512_extractf32x4_ps(__intrin_bitcast<__m512>(__in), _Offset));
+	else
+	  return reinterpret_cast<_R>(
+	    _mm512_extracti32x4_epi32(__intrin_bitcast<__m512i>(__in),
+				      _Offset));
+      }
+    else
+#endif // _GLIBCXX_SIMD_X86INTRIN }}}
+      {
+#ifdef _GLIBCXX_SIMD_WORKAROUND_XXX_1
+	using _W =3D conditional_t<
+	  is_floating_point_v<value_type>, double,
+	  conditional_t<(sizeof(_R) >=3D 16), long long, value_type>>;
+	static_assert(sizeof(_R) % sizeof(_W) =3D=3D 0);
+	constexpr int __return_width =3D sizeof(_R) / sizeof(_W);
+	using _Up =3D __vector_type_t<_W, __return_width>;
+	const auto __x =3D __vector_bitcast<_W>(__in);
+#else
+      constexpr int __return_width =3D _TVT::_S_full_size / _SplitBy;
+      using _Up =3D _R;
+      const __vector_type_t<value_type, _TVT::_S_full_size>& __x
+	=3D __in; // only needed for _Tp =3D _SimdWrapper<value_type, _Np>
+#endif
+	constexpr int _O =3D _Offset * __return_width;
+	return __call_with_subscripts<__return_width, _O>(
+	  __x, [](auto... __entries) {
+	    return reinterpret_cast<_R>(_Up{__entries...});
+	  });
+      }
+  }
+
+// }}}
+// __lo/__hi64[z]{{{
+template <typename _Tp,
+	  typename _R
+	  =3D __vector_type8_t<typename _VectorTraits<_Tp>::value_type>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _R
+  __lo64(_Tp __x)
+  {
+    _R __r{};
+    __builtin_memcpy(&__r, &__x, 8);
+    return __r;
+  }
+
+template <typename _Tp,
+	  typename _R
+	  =3D __vector_type8_t<typename _VectorTraits<_Tp>::value_type>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _R
+  __hi64(_Tp __x)
+  {
+    static_assert(sizeof(_Tp) =3D=3D 16, "use __hi64z if you meant it");
+    _R __r{};
+    __builtin_memcpy(&__r, reinterpret_cast<const char*>(&__x) + 8, 8);
+    return __r;
+  }
+
+template <typename _Tp,
+	  typename _R
+	  =3D __vector_type8_t<typename _VectorTraits<_Tp>::value_type>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _R
+  __hi64z([[maybe_unused]] _Tp __x)
+  {
+    _R __r{};
+    if constexpr (sizeof(_Tp) =3D=3D 16)
+      __builtin_memcpy(&__r, reinterpret_cast<const char*>(&__x) + 8, 8);
+    return __r;
+  }
+
+// }}}
+// __lo/__hi128{{{
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __lo128(_Tp __x)
+  { return __extract<0, sizeof(_Tp) / 16>(__x); }
+
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __hi128(_Tp __x)
+  {
+    static_assert(sizeof(__x) =3D=3D 32);
+    return __extract<1, 2>(__x);
+  }
+
+// }}}
+// __lo/__hi256{{{
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __lo256(_Tp __x)
+  {
+    static_assert(sizeof(__x) =3D=3D 64);
+    return __extract<0, 2>(__x);
+  }
+
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __hi256(_Tp __x)
+  {
+    static_assert(sizeof(__x) =3D=3D 64);
+    return __extract<1, 2>(__x);
+  }
+
+// }}}
+// __auto_bitcast{{{
+template <typename _Tp>
+  struct _AutoCast
+  {
+    static_assert(__is_vector_type_v<_Tp>);
+
+    const _Tp __x;
+
+    template <typename _Up, typename _UVT =3D _VectorTraits<_Up>>
+      _GLIBCXX_SIMD_INTRINSIC constexpr operator _Up() const
+      { return __intrin_bitcast<typename _UVT::type>(__x); }
+  };
+
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _AutoCast<_Tp>
+  __auto_bitcast(const _Tp& __x)
+  { return {__x}; }
+
+template <typename _Tp, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC constexpr
+  _AutoCast<typename _SimdWrapper<_Tp, _Np>::_BuiltinType>
+  __auto_bitcast(const _SimdWrapper<_Tp, _Np>& __x)
+  { return {__x._M_data}; }
+
+// }}}
+// ^^^ ---- builtin vector types [[gnu::vector_size(N)]] and operations --=
=2D- ^^^
+
+#if _GLIBCXX_SIMD_HAVE_SSE_ABI
+// __bool_storage_member_type{{{
+#if _GLIBCXX_SIMD_HAVE_AVX512F && _GLIBCXX_SIMD_X86INTRIN
+template <size_t _Size>
+  struct __bool_storage_member_type
+  {
+    static_assert((_Size & (_Size - 1)) !=3D 0,
+		  "This trait may only be used for non-power-of-2 sizes. "
+		  "Power-of-2 sizes must be specialized.");
+    using type =3D
+      typename __bool_storage_member_type<std::__bit_ceil(_Size)>::type;
+  };
+
+template <>
+  struct __bool_storage_member_type<1> { using type =3D bool; };
+
+template <>
+  struct __bool_storage_member_type<2> { using type =3D __mmask8; };
+
+template <>
+  struct __bool_storage_member_type<4> { using type =3D __mmask8; };
+
+template <>
+  struct __bool_storage_member_type<8> { using type =3D __mmask8; };
+
+template <>
+  struct __bool_storage_member_type<16> { using type =3D __mmask16; };
+
+template <>
+  struct __bool_storage_member_type<32> { using type =3D __mmask32; };
+
+template <>
+  struct __bool_storage_member_type<64> { using type =3D __mmask64; };
+#endif // _GLIBCXX_SIMD_HAVE_AVX512F
+
+// }}}
+// __intrinsic_type (x86){{{
+// the following excludes bool via __is_vectorizable
+#if _GLIBCXX_SIMD_HAVE_SSE
+template <typename _Tp, size_t _Bytes>
+  struct __intrinsic_type<_Tp, _Bytes,
+			  enable_if_t<__is_vectorizable_v<_Tp> && _Bytes <=3D 64>>
+  {
+    static_assert(!is_same_v<_Tp, long double>,
+		  "no __intrinsic_type support for long double on x86");
+
+    static constexpr size_t _S_VBytes =3D _Bytes <=3D 16   ? 16
+					: _Bytes <=3D 32 ? 32
+						       : 64;
+
+    using type [[__gnu__::__vector_size__(_S_VBytes)]]
+    =3D conditional_t<is_integral_v<_Tp>, long long int, _Tp>;
+  };
+#endif // _GLIBCXX_SIMD_HAVE_SSE
+
+// }}}
+#endif // _GLIBCXX_SIMD_HAVE_SSE_ABI
+// __intrinsic_type (ARM){{{
+#if _GLIBCXX_SIMD_HAVE_NEON
+template <typename _Tp, size_t _Bytes>
+  struct __intrinsic_type<_Tp, _Bytes,
+			  enable_if_t<__is_vectorizable_v<_Tp> && _Bytes <=3D 16>>
+  {
+    static constexpr int _S_VBytes =3D _Bytes <=3D 8 ? 8 : 16;
+    using _Ip =3D __int_for_sizeof_t<_Tp>;
+    using _Up =3D conditional_t<
+      is_floating_point_v<_Tp>, _Tp,
+      conditional_t<is_unsigned_v<_Tp>, make_unsigned_t<_Ip>, _Ip>>;
+    using type [[__gnu__::__vector_size__(_S_VBytes)]] =3D _Up;
+  };
+#endif // _GLIBCXX_SIMD_HAVE_NEON
+
+// }}}
+// __intrinsic_type (PPC){{{
+#ifdef __ALTIVEC__
+template <typename _Tp>
+  struct __intrinsic_type_impl;
+
+#define _GLIBCXX_SIMD_PPC_INTRIN(_Tp)                                     =
     \
+  template <>                                                             =
     \
+    struct __intrinsic_type_impl<_Tp> { using type =3D __vector _Tp; }
+_GLIBCXX_SIMD_PPC_INTRIN(float);
+_GLIBCXX_SIMD_PPC_INTRIN(double);
+_GLIBCXX_SIMD_PPC_INTRIN(signed char);
+_GLIBCXX_SIMD_PPC_INTRIN(unsigned char);
+_GLIBCXX_SIMD_PPC_INTRIN(signed short);
+_GLIBCXX_SIMD_PPC_INTRIN(unsigned short);
+_GLIBCXX_SIMD_PPC_INTRIN(signed int);
+_GLIBCXX_SIMD_PPC_INTRIN(unsigned int);
+_GLIBCXX_SIMD_PPC_INTRIN(signed long);
+_GLIBCXX_SIMD_PPC_INTRIN(unsigned long);
+_GLIBCXX_SIMD_PPC_INTRIN(signed long long);
+_GLIBCXX_SIMD_PPC_INTRIN(unsigned long long);
+#undef _GLIBCXX_SIMD_PPC_INTRIN
+
+template <typename _Tp, size_t _Bytes>
+  struct __intrinsic_type<_Tp, _Bytes,
+			  enable_if_t<__is_vectorizable_v<_Tp> && _Bytes <=3D 16>>
+  {
+    static_assert(!is_same_v<_Tp, long double>,
+		  "no __intrinsic_type support for long double on PPC");
+#ifndef __VSX__
+    static_assert(!is_same_v<_Tp, double>,
+		  "no __intrinsic_type support for double on PPC w/o VSX");
+#endif
+#ifndef __POWER8_VECTOR__
+    static_assert(
+      !(is_integral_v<_Tp> && sizeof(_Tp) > 4),
+      "no __intrinsic_type support for integers larger than 4 Bytes "
+      "on PPC w/o POWER8 vectors");
+#endif
+    using type =3D typename __intrinsic_type_impl<conditional_t<
+      is_floating_point_v<_Tp>, _Tp, __int_for_sizeof_t<_Tp>>>::type;
+  };
+#endif // __ALTIVEC__
+
+// }}}
+// _SimdWrapper<bool>{{{1
+template <size_t _Width>
+  struct _SimdWrapper<bool, _Width,
+		      void_t<typename __bool_storage_member_type<_Width>::type>>
+  {
+    using _BuiltinType =3D typename __bool_storage_member_type<_Width>::ty=
pe;
+    using value_type =3D bool;
+
+    static constexpr size_t _S_full_size =3D sizeof(_BuiltinType) * __CHAR=
_BIT__;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<bool, _S_full_size>
+    __as_full_vector() const { return _M_data; }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper() =3D default;
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(_BuiltinType __k)
+      : _M_data(__k) {};
+
+    _GLIBCXX_SIMD_INTRINSIC operator const _BuiltinType&() const
+    { return _M_data; }
+
+    _GLIBCXX_SIMD_INTRINSIC operator _BuiltinType&()
+    { return _M_data; }
+
+    _GLIBCXX_SIMD_INTRINSIC _BuiltinType __intrin() const
+    { return _M_data; }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr value_type operator[](size_t __i) co=
nst
+    { return _M_data & (_BuiltinType(1) << __i); }
+
+    template <size_t __i>
+      _GLIBCXX_SIMD_INTRINSIC constexpr value_type
+      operator[](_SizeConstant<__i>) const
+      { return _M_data & (_BuiltinType(1) << __i); }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr void _M_set(size_t __i, value_type _=
_x)
+    {
+      if (__x)
+	_M_data |=3D (_BuiltinType(1) << __i);
+      else
+	_M_data &=3D ~(_BuiltinType(1) << __i);
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC
+    constexpr bool _M_is_constprop() const
+    { return __builtin_constant_p(_M_data); }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_none_of() const
+    {
+      if (__builtin_constant_p(_M_data))
+	{
+	  constexpr int __nbits =3D sizeof(_BuiltinType) * __CHAR_BIT__;
+	  constexpr _BuiltinType __active_mask
+	    =3D ~_BuiltinType() >> (__nbits - _Width);
+	  return (_M_data & __active_mask) =3D=3D 0;
+	}
+      return false;
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_all_of() const
+    {
+      if (__builtin_constant_p(_M_data))
+	{
+	  constexpr int __nbits =3D sizeof(_BuiltinType) * __CHAR_BIT__;
+	  constexpr _BuiltinType __active_mask
+	    =3D ~_BuiltinType() >> (__nbits - _Width);
+	  return (_M_data & __active_mask) =3D=3D __active_mask;
+	}
+      return false;
+    }
+
+    _BuiltinType _M_data;
+  };
+
+// _SimdWrapperBase{{{1
+template <bool _MustZeroInitPadding, typename _BuiltinType>
+  struct _SimdWrapperBase;
+
+template <typename _BuiltinType>
+  struct _SimdWrapperBase<false, _BuiltinType> // no padding or no SNaNs
+  {
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase() =3D default;
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase(_BuiltinType __init)
+      : _M_data(__init)
+    {}
+
+    _BuiltinType _M_data;
+  };
+
+template <typename _BuiltinType>
+  struct _SimdWrapperBase<true, _BuiltinType> // with padding that needs to
+					      // never become SNaN
+  {
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase() : _M_data() {}
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapperBase(_BuiltinType __init)
+      : _M_data(__init)
+    {}
+
+    _BuiltinType _M_data;
+  };
+
+// }}}
+// _SimdWrapper{{{
+template <typename _Tp, size_t _Width>
+  struct _SimdWrapper<
+    _Tp, _Width,
+    void_t<__vector_type_t<_Tp, _Width>, __intrinsic_type_t<_Tp, _Width>>>
+    : _SimdWrapperBase<__has_iec559_behavior<__signaling_NaN, _Tp>::value
+			 && sizeof(_Tp) * _Width
+			      =3D=3D sizeof(__vector_type_t<_Tp, _Width>),
+		       __vector_type_t<_Tp, _Width>>
+  {
+    using _Base
+      =3D _SimdWrapperBase<__has_iec559_behavior<__signaling_NaN, _Tp>::va=
lue
+			   && sizeof(_Tp) * _Width
+				=3D=3D sizeof(__vector_type_t<_Tp, _Width>),
+			 __vector_type_t<_Tp, _Width>>;
+
+    static_assert(__is_vectorizable_v<_Tp>);
+    static_assert(_Width >=3D 2); // 1 doesn't make sense, use _Tp directl=
y then
+
+    using _BuiltinType =3D __vector_type_t<_Tp, _Width>;
+    using value_type =3D _Tp;
+
+    static inline constexpr size_t _S_full_size
+      =3D sizeof(_BuiltinType) / sizeof(value_type);
+    static inline constexpr int _S_size =3D _Width;
+    static inline constexpr bool _S_is_partial =3D _S_full_size !=3D _S_si=
ze;
+
+    using _Base::_M_data;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<_Tp, _S_full_size>
+    __as_full_vector() const
+    { return _M_data; }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(initializer_list<_Tp> _=
_init)
+      : _Base(__generate_from_n_evaluations<_Width, _BuiltinType>(
+	[&](auto __i) { return __init.begin()[__i.value]; })) {}
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper() =3D default;
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(const _SimdWrapper&)
+      =3D default;
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(_SimdWrapper&&) =3D def=
ault;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper&
+    operator=3D(const _SimdWrapper&) =3D default;
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper&
+    operator=3D(_SimdWrapper&&) =3D default;
+
+    template <typename _V, typename =3D enable_if_t<disjunction_v<
+			     is_same<_V, __vector_type_t<_Tp, _Width>>,
+			     is_same<_V, __intrinsic_type_t<_Tp, _Width>>>>>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper(_V __x)
+      // __vector_bitcast can convert e.g. __m128 to __vector(2) float
+      : _Base(__vector_bitcast<_Tp, _Width>(__x)) {}
+
+    template <typename... _As,
+	      typename =3D enable_if_t<((is_same_v<simd_abi::scalar, _As> && ...)
+				      && sizeof...(_As) <=3D _Width)>>
+      _GLIBCXX_SIMD_INTRINSIC constexpr
+      operator _SimdTuple<_Tp, _As...>() const
+      {
+	const auto& dd =3D _M_data; // workaround for GCC7 ICE
+	return __generate_from_n_evaluations<sizeof...(_As),
+					     _SimdTuple<_Tp, _As...>>([&](
+	  auto __i) constexpr { return dd[int(__i)]; });
+      }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr operator const _BuiltinType&() const
+    { return _M_data; }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr operator _BuiltinType&()
+    { return _M_data; }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _Tp operator[](size_t __i) const
+    { return _M_data[__i]; }
+
+    template <size_t __i>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _Tp operator[](_SizeConstant<__i>)=
 const
+      { return _M_data[__i]; }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr void _M_set(size_t __i, _Tp __x)
+    { _M_data[__i] =3D __x; }
+
+    _GLIBCXX_SIMD_INTRINSIC
+    constexpr bool _M_is_constprop() const
+    { return __builtin_constant_p(_M_data); }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_none_of() const
+    {
+      if (__builtin_constant_p(_M_data))
+	{
+	  bool __r =3D true;
+	  if constexpr (is_floating_point_v<_Tp>)
+	    {
+	      using _Ip =3D __int_for_sizeof_t<_Tp>;
+	      const auto __intdata =3D __vector_bitcast<_Ip>(_M_data);
+	      __execute_n_times<_Width>(
+		[&](auto __i) { __r &=3D __intdata[__i.value] =3D=3D _Ip(); });
+	    }
+	  else
+	    __execute_n_times<_Width>(
+	      [&](auto __i) { __r &=3D _M_data[__i.value] =3D=3D _Tp(); });
+	  return __r;
+	}
+      return false;
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr bool _M_is_constprop_all_of() const
+    {
+      if (__builtin_constant_p(_M_data))
+	{
+	  bool __r =3D true;
+	  if constexpr (is_floating_point_v<_Tp>)
+	    {
+	      using _Ip =3D __int_for_sizeof_t<_Tp>;
+	      const auto __intdata =3D __vector_bitcast<_Ip>(_M_data);
+	      __execute_n_times<_Width>(
+		[&](auto __i) { __r &=3D __intdata[__i.value] =3D=3D ~_Ip(); });
+	    }
+	  else
+	    __execute_n_times<_Width>(
+	      [&](auto __i) { __r &=3D _M_data[__i.value] =3D=3D ~_Tp(); });
+	  return __r;
+	}
+      return false;
+    }
+  };
+
+// }}}
+
+// __vectorized_sizeof {{{
+template <typename _Tp>
+  constexpr size_t
+  __vectorized_sizeof()
+  {
+    if constexpr (!__is_vectorizable_v<_Tp>)
+      return 0;
+
+    if constexpr (sizeof(_Tp) <=3D 8)
+      {
+	// X86:
+	if constexpr (__have_avx512bw)
+	  return 64;
+	if constexpr (__have_avx512f && sizeof(_Tp) >=3D 4)
+	  return 64;
+	if constexpr (__have_avx2)
+	  return 32;
+	if constexpr (__have_avx && is_floating_point_v<_Tp>)
+	  return 32;
+	if constexpr (__have_sse2)
+	  return 16;
+	if constexpr (__have_sse && is_same_v<_Tp, float>)
+	  return 16;
+	/* The following is too much trouble because of mixed MMX and x87 code.
+	 * While nothing here explicitly calls MMX instructions of registers,
+	 * they are still emitted but no EMMS cleanup is done.
+	if constexpr (__have_mmx && sizeof(_Tp) <=3D 4 && is_integral_v<_Tp>)
+	  return 8;
+	 */
+
+	// PowerPC:
+	if constexpr (__have_power8vec
+		      || (__have_power_vmx && (sizeof(_Tp) < 8))
+		      || (__have_power_vsx && is_floating_point_v<_Tp>) )
+	  return 16;
+
+	// ARM:
+	if constexpr (__have_neon_a64
+		      || (__have_neon_a32 && !is_same_v<_Tp, double>) )
+	  return 16;
+	if constexpr (__have_neon
+		      && sizeof(_Tp) < 8
+		      // Only allow fp if the user allows non-ICE559 fp (e.g.
+		      // via -ffast-math). ARMv7 NEON fp is not conforming to
+		      // IEC559.
+		      && (__support_neon_float || !is_floating_point_v<_Tp>))
+	  return 16;
+      }
+
+    return sizeof(_Tp);
+  }
+
+// }}}
+namespace simd_abi {
+// most of simd_abi is defined in simd_detail.h
+template <typename _Tp>
+  inline constexpr int max_fixed_size
+    =3D (__have_avx512bw && sizeof(_Tp) =3D=3D 1) ? 64 : 32;
+
+// compatible {{{
+#if defined __x86_64__ || defined __aarch64__
+template <typename _Tp>
+  using compatible =3D conditional_t<(sizeof(_Tp) <=3D 8), _VecBuiltin<16>=
, scalar>;
+#elif defined __ARM_NEON
+// FIXME: not sure, probably needs to be scalar (or dependent on the hard-=
float
+// ABI?)
+template <typename _Tp>
+  using compatible
+    =3D conditional_t<(sizeof(_Tp) < 8
+		     && (__support_neon_float || !is_floating_point_v<_Tp>)),
+		    _VecBuiltin<16>, scalar>;
+#else
+template <typename>
+  using compatible =3D scalar;
+#endif
+
+// }}}
+// native {{{
+template <typename _Tp>
+  constexpr auto
+  __determine_native_abi()
+  {
+    constexpr size_t __bytes =3D __vectorized_sizeof<_Tp>();
+    if constexpr (__bytes =3D=3D sizeof(_Tp))
+      return static_cast<scalar*>(nullptr);
+    else if constexpr (__have_avx512vl || (__have_avx512f && __bytes =3D=
=3D 64))
+      return static_cast<_VecBltnBtmsk<__bytes>*>(nullptr);
+    else
+      return static_cast<_VecBuiltin<__bytes>*>(nullptr);
+  }
+
+template <typename _Tp, typename =3D enable_if_t<__is_vectorizable_v<_Tp>>>
+  using native =3D remove_pointer_t<decltype(__determine_native_abi<_Tp>()=
)>;
+
+// }}}
+// __default_abi {{{
+#if defined _GLIBCXX_SIMD_DEFAULT_ABI
+template <typename _Tp>
+  using __default_abi =3D _GLIBCXX_SIMD_DEFAULT_ABI<_Tp>;
+#else
+template <typename _Tp>
+  using __default_abi =3D compatible<_Tp>;
+#endif
+
+// }}}
+} // namespace simd_abi
+
+// traits {{{1
+// is_abi_tag {{{2
+template <typename _Tp, typename =3D void_t<>>
+  struct is_abi_tag : false_type {};
+
+template <typename _Tp>
+  struct is_abi_tag<_Tp, void_t<typename _Tp::_IsValidAbiTag>>
+  : public _Tp::_IsValidAbiTag {};
+
+template <typename _Tp>
+  inline constexpr bool is_abi_tag_v =3D is_abi_tag<_Tp>::value;
+
+// is_simd(_mask) {{{2
+template <typename _Tp>
+  struct is_simd : public false_type {};
+
+template <typename _Tp>
+  inline constexpr bool is_simd_v =3D is_simd<_Tp>::value;
+
+template <typename _Tp>
+  struct is_simd_mask : public false_type {};
+
+template <typename _Tp>
+inline constexpr bool is_simd_mask_v =3D is_simd_mask<_Tp>::value;
+
+// simd_size {{{2
+template <typename _Tp, typename _Abi, typename =3D void>
+  struct __simd_size_impl {};
+
+template <typename _Tp, typename _Abi>
+  struct __simd_size_impl<
+    _Tp, _Abi,
+    enable_if_t<conjunction_v<__is_vectorizable<_Tp>, is_abi_tag<_Abi>>>>
+    : _SizeConstant<_Abi::template _S_size<_Tp>> {};
+
+template <typename _Tp, typename _Abi =3D simd_abi::__default_abi<_Tp>>
+  struct simd_size : __simd_size_impl<_Tp, _Abi> {};
+
+template <typename _Tp, typename _Abi =3D simd_abi::__default_abi<_Tp>>
+  inline constexpr size_t simd_size_v =3D simd_size<_Tp, _Abi>::value;
+
+// simd_abi::deduce {{{2
+template <typename _Tp, size_t _Np, typename =3D void>
+  struct __deduce_impl;
+
+namespace simd_abi {
+/**
+ * @tparam _Tp   The requested `value_type` for the elements.
+ * @tparam _Np    The requested number of elements.
+ * @tparam _Abis This parameter is ignored, since this implementation cann=
ot
+ * make any use of it. Either __a good native ABI is matched and used as `=
type`
+ * alias, or the `fixed_size<_Np>` ABI is used, which internally is built =
from
+ * the best matching native ABIs.
+ */
+template <typename _Tp, size_t _Np, typename...>
+  struct deduce : __deduce_impl<_Tp, _Np> {};
+
+template <typename _Tp, size_t _Np, typename... _Abis>
+  using deduce_t =3D typename deduce<_Tp, _Np, _Abis...>::type;
+} // namespace simd_abi
+
+// }}}2
+// rebind_simd {{{2
+template <typename _Tp, typename _V, typename =3D void>
+  struct rebind_simd;
+
+template <typename _Tp, typename _Up, typename _Abi>
+  struct rebind_simd<
+    _Tp, simd<_Up, _Abi>,
+    void_t<simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _Abi>>>
+  {
+    using type
+      =3D simd<_Tp, simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _Abi>>;
+  };
+
+template <typename _Tp, typename _Up, typename _Abi>
+  struct rebind_simd<
+    _Tp, simd_mask<_Up, _Abi>,
+    void_t<simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _Abi>>>
+  {
+    using type
+      =3D simd_mask<_Tp, simd_abi::deduce_t<_Tp, simd_size_v<_Up, _Abi>, _=
Abi>>;
+  };
+
+template <typename _Tp, typename _V>
+  using rebind_simd_t =3D typename rebind_simd<_Tp, _V>::type;
+
+// resize_simd {{{2
+template <int _Np, typename _V, typename =3D void>
+  struct resize_simd;
+
+template <int _Np, typename _Tp, typename _Abi>
+  struct resize_simd<_Np, simd<_Tp, _Abi>,
+		     void_t<simd_abi::deduce_t<_Tp, _Np, _Abi>>>
+  { using type =3D simd<_Tp, simd_abi::deduce_t<_Tp, _Np, _Abi>>; };
+
+template <int _Np, typename _Tp, typename _Abi>
+  struct resize_simd<_Np, simd_mask<_Tp, _Abi>,
+		     void_t<simd_abi::deduce_t<_Tp, _Np, _Abi>>>
+  { using type =3D simd_mask<_Tp, simd_abi::deduce_t<_Tp, _Np, _Abi>>; };
+
+template <int _Np, typename _V>
+  using resize_simd_t =3D typename resize_simd<_Np, _V>::type;
+
+// }}}2
+// memory_alignment {{{2
+template <typename _Tp, typename _Up =3D typename _Tp::value_type>
+  struct memory_alignment
+  : public _SizeConstant<vector_aligned_tag::_S_alignment<_Tp, _Up>> {};
+
+template <typename _Tp, typename _Up =3D typename _Tp::value_type>
+  inline constexpr size_t memory_alignment_v =3D memory_alignment<_Tp, _Up=
>::value;
+
+// class template simd [simd] {{{1
+template <typename _Tp, typename _Abi =3D simd_abi::__default_abi<_Tp>>
+  class simd;
+
+template <typename _Tp, typename _Abi>
+  struct is_simd<simd<_Tp, _Abi>> : public true_type {};
+
+template <typename _Tp>
+  using native_simd =3D simd<_Tp, simd_abi::native<_Tp>>;
+
+template <typename _Tp, int _Np>
+  using fixed_size_simd =3D simd<_Tp, simd_abi::fixed_size<_Np>>;
+
+template <typename _Tp, size_t _Np>
+  using __deduced_simd =3D simd<_Tp, simd_abi::deduce_t<_Tp, _Np>>;
+
+// class template simd_mask [simd_mask] {{{1
+template <typename _Tp, typename _Abi =3D simd_abi::__default_abi<_Tp>>
+  class simd_mask;
+
+template <typename _Tp, typename _Abi>
+  struct is_simd_mask<simd_mask<_Tp, _Abi>> : public true_type {};
+
+template <typename _Tp>
+  using native_simd_mask =3D simd_mask<_Tp, simd_abi::native<_Tp>>;
+
+template <typename _Tp, int _Np>
+  using fixed_size_simd_mask =3D simd_mask<_Tp, simd_abi::fixed_size<_Np>>;
+
+template <typename _Tp, size_t _Np>
+  using __deduced_simd_mask =3D simd_mask<_Tp, simd_abi::deduce_t<_Tp, _Np=
>>;
+
+// casts [simd.casts] {{{1
+// static_simd_cast {{{2
+template <typename _Tp, typename _Up, typename _Ap, bool =3D is_simd_v<_Tp=
>,
+	  typename =3D void>
+  struct __static_simd_cast_return_type;
+
+template <typename _Tp, typename _A0, typename _Up, typename _Ap>
+  struct __static_simd_cast_return_type<simd_mask<_Tp, _A0>, _Up, _Ap, fal=
se,
+					void>
+  : __static_simd_cast_return_type<simd<_Tp, _A0>, _Up, _Ap> {};
+
+template <typename _Tp, typename _Up, typename _Ap>
+  struct __static_simd_cast_return_type<
+    _Tp, _Up, _Ap, true, enable_if_t<_Tp::size() =3D=3D simd_size_v<_Up, _=
Ap>>>
+  { using type =3D _Tp; };
+
+template <typename _Tp, typename _Ap>
+  struct __static_simd_cast_return_type<_Tp, _Tp, _Ap, false,
+#ifdef _GLIBCXX_SIMD_FIX_P2TS_ISSUE66
+					enable_if_t<__is_vectorizable_v<_Tp>>
+#else
+					void
+#endif
+					>
+  { using type =3D simd<_Tp, _Ap>; };
+
+template <typename _Tp, typename =3D void>
+  struct __safe_make_signed { using type =3D _Tp;};
+
+template <typename _Tp>
+  struct __safe_make_signed<_Tp, enable_if_t<is_integral_v<_Tp>>>
+  {
+    // the extra make_unsigned_t is because of PR85951
+    using type =3D make_signed_t<make_unsigned_t<_Tp>>;
+  };
+
+template <typename _Tp>
+  using safe_make_signed_t =3D typename __safe_make_signed<_Tp>::type;
+
+template <typename _Tp, typename _Up, typename _Ap>
+  struct __static_simd_cast_return_type<_Tp, _Up, _Ap, false,
+#ifdef _GLIBCXX_SIMD_FIX_P2TS_ISSUE66
+					enable_if_t<__is_vectorizable_v<_Tp>>
+#else
+					void
+#endif
+					>
+  {
+    using type =3D conditional_t<
+      (is_integral_v<_Up> && is_integral_v<_Tp> &&
+#ifndef _GLIBCXX_SIMD_FIX_P2TS_ISSUE65
+       is_signed_v<_Up> !=3D is_signed_v<_Tp> &&
+#endif
+       is_same_v<safe_make_signed_t<_Up>, safe_make_signed_t<_Tp>>),
+      simd<_Tp, _Ap>, fixed_size_simd<_Tp, simd_size_v<_Up, _Ap>>>;
+  };
+
+template <typename _Tp, typename _Up, typename _Ap,
+	  typename _R
+	  =3D typename __static_simd_cast_return_type<_Tp, _Up, _Ap>::type>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR _R
+  static_simd_cast(const simd<_Up, _Ap>& __x)
+  {
+    if constexpr (is_same<_R, simd<_Up, _Ap>>::value)
+      return __x;
+    else
+      {
+	_SimdConverter<_Up, _Ap, typename _R::value_type, typename _R::abi_type>
+	  __c;
+	return _R(__private_init, __c(__data(__x)));
+      }
+  }
+
+namespace __proposed {
+template <typename _Tp, typename _Up, typename _Ap,
+	  typename _R
+	  =3D typename __static_simd_cast_return_type<_Tp, _Up, _Ap>::type>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR typename _R::mask_type
+  static_simd_cast(const simd_mask<_Up, _Ap>& __x)
+  {
+    using _RM =3D typename _R::mask_type;
+    return {__private_init, _RM::abi_type::_MaskImpl::template _S_convert<
+			      typename _RM::simd_type::value_type>(__x)};
+  }
+} // namespace __proposed
+
+// simd_cast {{{2
+template <typename _Tp, typename _Up, typename _Ap,
+	  typename _To =3D __value_type_or_identity_t<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR auto
+  simd_cast(const simd<_ValuePreserving<_Up, _To>, _Ap>& __x)
+    -> decltype(static_simd_cast<_Tp>(__x))
+  { return static_simd_cast<_Tp>(__x); }
+
+namespace __proposed {
+template <typename _Tp, typename _Up, typename _Ap,
+	  typename _To =3D __value_type_or_identity_t<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR auto
+  simd_cast(const simd_mask<_ValuePreserving<_Up, _To>, _Ap>& __x)
+    -> decltype(static_simd_cast<_Tp>(__x))
+  { return static_simd_cast<_Tp>(__x); }
+} // namespace __proposed
+
+// }}}2
+// resizing_simd_cast {{{
+namespace __proposed {
+/* Proposed spec:
+
+template <class T, class U, class Abi>
+T resizing_simd_cast(const simd<U, Abi>& x)
+
+p1  Constraints:
+    - is_simd_v<T> is true and
+    - T::value_type is the same type as U
+
+p2  Returns:
+    A simd object with the i^th element initialized to x[i] for all i in t=
he
+    range of [0, min(T::size(), simd_size_v<U, Abi>)). If T::size() is lar=
ger
+    than simd_size_v<U, Abi>, the remaining elements are value-initialized.
+
+template <class T, class U, class Abi>
+T resizing_simd_cast(const simd_mask<U, Abi>& x)
+
+p1  Constraints: is_simd_mask_v<T> is true
+
+p2  Returns:
+    A simd_mask object with the i^th element initialized to x[i] for all i=
 in
+the range of [0, min(T::size(), simd_size_v<U, Abi>)). If T::size() is lar=
ger
+    than simd_size_v<U, Abi>, the remaining elements are initialized to fa=
lse.
+
+ */
+
+template <typename _Tp, typename _Up, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR enable_if_t<
+  conjunction_v<is_simd<_Tp>, is_same<typename _Tp::value_type, _Up>>, _Tp>
+  resizing_simd_cast(const simd<_Up, _Ap>& __x)
+  {
+    if constexpr (is_same_v<typename _Tp::abi_type, _Ap>)
+      return __x;
+    else if constexpr (simd_size_v<_Up, _Ap> =3D=3D 1)
+      {
+	_Tp __r{};
+	__r[0] =3D __x[0];
+	return __r;
+      }
+    else if constexpr (_Tp::size() =3D=3D 1)
+      return __x[0];
+    else if constexpr (sizeof(_Tp) =3D=3D sizeof(__x)
+		       && !__is_fixed_size_abi_v<_Ap>)
+      return {__private_init,
+	      __vector_bitcast<typename _Tp::value_type, _Tp::size()>(
+		_Ap::_S_masked(__data(__x))._M_data)};
+    else
+      {
+	_Tp __r{};
+	__builtin_memcpy(&__data(__r), &__data(__x),
+			 sizeof(_Up)
+			   * std::min(_Tp::size(), simd_size_v<_Up, _Ap>));
+	return __r;
+      }
+  }
+
+template <typename _Tp, typename _Up, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
+  enable_if_t<is_simd_mask_v<_Tp>, _Tp>
+  resizing_simd_cast(const simd_mask<_Up, _Ap>& __x)
+  {
+    return {__private_init, _Tp::abi_type::_MaskImpl::template _S_convert<
+			      typename _Tp::simd_type::value_type>(__x)};
+  }
+} // namespace __proposed
+
+// }}}
+// to_fixed_size {{{2
+template <typename _Tp, int _Np>
+  _GLIBCXX_SIMD_INTRINSIC fixed_size_simd<_Tp, _Np>
+  to_fixed_size(const fixed_size_simd<_Tp, _Np>& __x)
+  { return __x; }
+
+template <typename _Tp, int _Np>
+  _GLIBCXX_SIMD_INTRINSIC fixed_size_simd_mask<_Tp, _Np>
+  to_fixed_size(const fixed_size_simd_mask<_Tp, _Np>& __x)
+  { return __x; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  to_fixed_size(const simd<_Tp, _Ap>& __x)
+  {
+    return simd<_Tp, simd_abi::fixed_size<simd_size_v<_Tp, _Ap>>>([&__x](
+      auto __i) constexpr { return __x[__i]; });
+  }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  to_fixed_size(const simd_mask<_Tp, _Ap>& __x)
+  {
+    constexpr int _Np =3D simd_mask<_Tp, _Ap>::size();
+    fixed_size_simd_mask<_Tp, _Np> __r;
+    __execute_n_times<_Np>([&](auto __i) constexpr { __r[__i] =3D __x[__i]=
; });
+    return __r;
+  }
+
+// to_native {{{2
+template <typename _Tp, int _Np>
+  _GLIBCXX_SIMD_INTRINSIC
+  enable_if_t<(_Np =3D=3D native_simd<_Tp>::size()), native_simd<_Tp>>
+  to_native(const fixed_size_simd<_Tp, _Np>& __x)
+  {
+    alignas(memory_alignment_v<native_simd<_Tp>>) _Tp __mem[_Np];
+    __x.copy_to(__mem, vector_aligned);
+    return {__mem, vector_aligned};
+  }
+
+template <typename _Tp, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC
+  enable_if_t<(_Np =3D=3D native_simd_mask<_Tp>::size()), native_simd_mask=
<_Tp>>
+  to_native(const fixed_size_simd_mask<_Tp, _Np>& __x)
+  {
+    return native_simd_mask<_Tp>([&](auto __i) constexpr { return __x[__i]=
; });
+  }
+
+// to_compatible {{{2
+template <typename _Tp, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC enable_if_t<(_Np =3D=3D simd<_Tp>::size()), simd=
<_Tp>>
+  to_compatible(const simd<_Tp, simd_abi::fixed_size<_Np>>& __x)
+  {
+    alignas(memory_alignment_v<simd<_Tp>>) _Tp __mem[_Np];
+    __x.copy_to(__mem, vector_aligned);
+    return {__mem, vector_aligned};
+  }
+
+template <typename _Tp, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC
+  enable_if_t<(_Np =3D=3D simd_mask<_Tp>::size()), simd_mask<_Tp>>
+  to_compatible(const simd_mask<_Tp, simd_abi::fixed_size<_Np>>& __x)
+  { return simd_mask<_Tp>([&](auto __i) constexpr { return __x[__i]; }); }
+
+// masked assignment [simd_mask.where] {{{1
+
+// where_expression {{{1
+// const_where_expression<M, T> {{{2
+template <typename _M, typename _Tp>
+  class const_where_expression
+  {
+    using _V =3D _Tp;
+    static_assert(is_same_v<_V, __remove_cvref_t<_Tp>>);
+
+    struct _Wrapper { using value_type =3D _V; };
+
+  protected:
+    using _Impl =3D typename _V::_Impl;
+
+    using value_type =3D
+      typename conditional_t<is_arithmetic_v<_V>, _Wrapper, _V>::value_typ=
e;
+
+    _GLIBCXX_SIMD_INTRINSIC friend const _M&
+    __get_mask(const const_where_expression& __x)
+    { return __x._M_k; }
+
+    _GLIBCXX_SIMD_INTRINSIC friend const _Tp&
+    __get_lvalue(const const_where_expression& __x)
+    { return __x._M_value; }
+
+    const _M& _M_k;
+    _Tp& _M_value;
+
+  public:
+    const_where_expression(const const_where_expression&) =3D delete;
+    const_where_expression& operator=3D(const const_where_expression&) =3D=
 delete;
+
+    _GLIBCXX_SIMD_INTRINSIC const_where_expression(const _M& __kk, const _=
Tp& dd)
+      : _M_k(__kk), _M_value(const_cast<_Tp&>(dd)) {}
+
+    _GLIBCXX_SIMD_INTRINSIC _V
+    operator-() const&&
+    {
+      return {__private_init,
+	      _Impl::template _S_masked_unary<negate>(__data(_M_k),
+						      __data(_M_value))};
+    }
+
+    template <typename _Up, typename _Flags>
+      [[nodiscard]] _GLIBCXX_SIMD_INTRINSIC _V
+      copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) const=
&&
+      {
+	return {__private_init,
+		_Impl::_S_masked_load(__data(_M_value), __data(_M_k),
+				      _Flags::template _S_apply<_V>(__mem))};
+      }
+
+    template <typename _Up, typename _Flags>
+      _GLIBCXX_SIMD_INTRINSIC void
+      copy_to(_LoadStorePtr<_Up, value_type>* __mem, _Flags) const&&
+      {
+	_Impl::_S_masked_store(__data(_M_value),
+			       _Flags::template _S_apply<_V>(__mem),
+			       __data(_M_k));
+      }
+  };
+
+// const_where_expression<bool, T> {{{2
+template <typename _Tp>
+  class const_where_expression<bool, _Tp>
+  {
+    using _M =3D bool;
+    using _V =3D _Tp;
+
+    static_assert(is_same_v<_V, __remove_cvref_t<_Tp>>);
+
+    struct _Wrapper { using value_type =3D _V; };
+
+  protected:
+    using value_type =3D
+      typename conditional_t<is_arithmetic_v<_V>, _Wrapper, _V>::value_typ=
e;
+
+    _GLIBCXX_SIMD_INTRINSIC friend const _M&
+    __get_mask(const const_where_expression& __x)
+    { return __x._M_k; }
+
+    _GLIBCXX_SIMD_INTRINSIC friend const _Tp&
+    __get_lvalue(const const_where_expression& __x)
+    { return __x._M_value; }
+
+    const bool _M_k;
+    _Tp& _M_value;
+
+  public:
+    const_where_expression(const const_where_expression&) =3D delete;
+    const_where_expression& operator=3D(const const_where_expression&) =3D=
 delete;
+
+    _GLIBCXX_SIMD_INTRINSIC const_where_expression(const bool __kk, const =
_Tp& dd)
+      : _M_k(__kk), _M_value(const_cast<_Tp&>(dd)) {}
+
+    _GLIBCXX_SIMD_INTRINSIC _V operator-() const&&
+    { return _M_k ? -_M_value : _M_value; }
+
+    template <typename _Up, typename _Flags>
+      [[nodiscard]] _GLIBCXX_SIMD_INTRINSIC _V
+      copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) const=
&&
+      { return _M_k ? static_cast<_V>(__mem[0]) : _M_value; }
+
+    template <typename _Up, typename _Flags>
+      _GLIBCXX_SIMD_INTRINSIC void
+      copy_to(_LoadStorePtr<_Up, value_type>* __mem, _Flags) const&&
+      {
+	if (_M_k)
+	  __mem[0] =3D _M_value;
+      }
+  };
+
+// where_expression<M, T> {{{2
+template <typename _M, typename _Tp>
+  class where_expression : public const_where_expression<_M, _Tp>
+  {
+    using _Impl =3D typename const_where_expression<_M, _Tp>::_Impl;
+
+    static_assert(!is_const<_Tp>::value,
+		  "where_expression may only be instantiated with __a non-const "
+		  "_Tp parameter");
+
+    using typename const_where_expression<_M, _Tp>::value_type;
+    using const_where_expression<_M, _Tp>::_M_k;
+    using const_where_expression<_M, _Tp>::_M_value;
+
+    static_assert(
+      is_same<typename _M::abi_type, typename _Tp::abi_type>::value, "");
+    static_assert(_M::size() =3D=3D _Tp::size(), "");
+
+    _GLIBCXX_SIMD_INTRINSIC friend _Tp& __get_lvalue(where_expression& __x)
+    { return __x._M_value; }
+
+  public:
+    where_expression(const where_expression&) =3D delete;
+    where_expression& operator=3D(const where_expression&) =3D delete;
+
+    _GLIBCXX_SIMD_INTRINSIC where_expression(const _M& __kk, _Tp& dd)
+      : const_where_expression<_M, _Tp>(__kk, dd) {}
+
+    template <typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC void operator=3D(_Up&& __x) &&
+      {
+	_Impl::_S_masked_assign(__data(_M_k), __data(_M_value),
+				__to_value_type_or_member_type<_Tp>(
+				  static_cast<_Up&&>(__x)));
+      }
+
+#define _GLIBCXX_SIMD_OP_(__op, __name)                                   =
     \
+  template <typename _Up>                                                 =
     \
+    _GLIBCXX_SIMD_INTRINSIC void operator __op##=3D(_Up&& __x)&&          =
       \
+    {                                                                     =
     \
+      _Impl::template _S_masked_cassign(                                  =
     \
+	__data(_M_k), __data(_M_value),                                        \
+	__to_value_type_or_member_type<_Tp>(static_cast<_Up&&>(__x)),          \
+	[](auto __impl, auto __lhs, auto __rhs) constexpr {                    \
+	return __impl.__name(__lhs, __rhs);                                    \
+	});                                                                    \
+    }                                                                     =
     \
+  static_assert(true)
+    _GLIBCXX_SIMD_OP_(+, _S_plus);
+    _GLIBCXX_SIMD_OP_(-, _S_minus);
+    _GLIBCXX_SIMD_OP_(*, _S_multiplies);
+    _GLIBCXX_SIMD_OP_(/, _S_divides);
+    _GLIBCXX_SIMD_OP_(%, _S_modulus);
+    _GLIBCXX_SIMD_OP_(&, _S_bit_and);
+    _GLIBCXX_SIMD_OP_(|, _S_bit_or);
+    _GLIBCXX_SIMD_OP_(^, _S_bit_xor);
+    _GLIBCXX_SIMD_OP_(<<, _S_shift_left);
+    _GLIBCXX_SIMD_OP_(>>, _S_shift_right);
+#undef _GLIBCXX_SIMD_OP_
+
+    _GLIBCXX_SIMD_INTRINSIC void operator++() &&
+    {
+      __data(_M_value)
+	=3D _Impl::template _S_masked_unary<__increment>(__data(_M_k),
+						       __data(_M_value));
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC void operator++(int) &&
+    {
+      __data(_M_value)
+	=3D _Impl::template _S_masked_unary<__increment>(__data(_M_k),
+						       __data(_M_value));
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC void operator--() &&
+    {
+      __data(_M_value)
+	=3D _Impl::template _S_masked_unary<__decrement>(__data(_M_k),
+						       __data(_M_value));
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC void operator--(int) &&
+    {
+      __data(_M_value)
+	=3D _Impl::template _S_masked_unary<__decrement>(__data(_M_k),
+						       __data(_M_value));
+    }
+
+    // intentionally hides const_where_expression::copy_from
+    template <typename _Up, typename _Flags>
+      _GLIBCXX_SIMD_INTRINSIC void
+      copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) &&
+      {
+	__data(_M_value)
+	  =3D _Impl::_S_masked_load(__data(_M_value), __data(_M_k),
+				  _Flags::template _S_apply<_Tp>(__mem));
+      }
+  };
+
+// where_expression<bool, T> {{{2
+template <typename _Tp>
+  class where_expression<bool, _Tp> : public const_where_expression<bool, =
_Tp>
+  {
+    using _M =3D bool;
+    using typename const_where_expression<_M, _Tp>::value_type;
+    using const_where_expression<_M, _Tp>::_M_k;
+    using const_where_expression<_M, _Tp>::_M_value;
+
+  public:
+    where_expression(const where_expression&) =3D delete;
+    where_expression& operator=3D(const where_expression&) =3D delete;
+
+    _GLIBCXX_SIMD_INTRINSIC where_expression(const _M& __kk, _Tp& dd)
+      : const_where_expression<_M, _Tp>(__kk, dd) {}
+
+#define _GLIBCXX_SIMD_OP_(__op)                                           =
     \
+    template <typename _Up>                                               =
     \
+      _GLIBCXX_SIMD_INTRINSIC void operator __op(_Up&& __x)&&             =
     \
+      { if (_M_k) _M_value __op static_cast<_Up&&>(__x); }
+
+    _GLIBCXX_SIMD_OP_(=3D)
+    _GLIBCXX_SIMD_OP_(+=3D)
+    _GLIBCXX_SIMD_OP_(-=3D)
+    _GLIBCXX_SIMD_OP_(*=3D)
+    _GLIBCXX_SIMD_OP_(/=3D)
+    _GLIBCXX_SIMD_OP_(%=3D)
+    _GLIBCXX_SIMD_OP_(&=3D)
+    _GLIBCXX_SIMD_OP_(|=3D)
+    _GLIBCXX_SIMD_OP_(^=3D)
+    _GLIBCXX_SIMD_OP_(<<=3D)
+    _GLIBCXX_SIMD_OP_(>>=3D)
+  #undef _GLIBCXX_SIMD_OP_
+
+    _GLIBCXX_SIMD_INTRINSIC void operator++() &&
+    { if (_M_k) ++_M_value; }
+
+    _GLIBCXX_SIMD_INTRINSIC void operator++(int) &&
+    { if (_M_k) ++_M_value; }
+
+    _GLIBCXX_SIMD_INTRINSIC void operator--() &&
+    { if (_M_k) --_M_value; }
+
+    _GLIBCXX_SIMD_INTRINSIC void operator--(int) &&
+    { if (_M_k) --_M_value; }
+
+    // intentionally hides const_where_expression::copy_from
+    template <typename _Up, typename _Flags>
+      _GLIBCXX_SIMD_INTRINSIC void
+      copy_from(const _LoadStorePtr<_Up, value_type>* __mem, _Flags) &&
+      { if (_M_k) _M_value =3D __mem[0]; }
+  };
+
+// where {{{1
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC where_expression<simd_mask<_Tp, _Ap>, simd<_Tp, =
_Ap>>
+  where(const typename simd<_Tp, _Ap>::mask_type& __k, simd<_Tp, _Ap>& __v=
alue)
+  { return {__k, __value}; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC
+    const_where_expression<simd_mask<_Tp, _Ap>, simd<_Tp, _Ap>>
+    where(const typename simd<_Tp, _Ap>::mask_type& __k,
+	  const simd<_Tp, _Ap>& __value)
+  { return {__k, __value}; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC
+    where_expression<simd_mask<_Tp, _Ap>, simd_mask<_Tp, _Ap>>
+    where(const remove_const_t<simd_mask<_Tp, _Ap>>& __k,
+	  simd_mask<_Tp, _Ap>& __value)
+  { return {__k, __value}; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC
+    const_where_expression<simd_mask<_Tp, _Ap>, simd_mask<_Tp, _Ap>>
+    where(const remove_const_t<simd_mask<_Tp, _Ap>>& __k,
+	  const simd_mask<_Tp, _Ap>& __value)
+  { return {__k, __value}; }
+
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC where_expression<bool, _Tp>
+  where(_ExactBool __k, _Tp& __value)
+  { return {__k, __value}; }
+
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC const_where_expression<bool, _Tp>
+  where(_ExactBool __k, const _Tp& __value)
+  { return {__k, __value}; }
+
+  template <typename _Tp, typename _Ap>
+    void where(bool __k, simd<_Tp, _Ap>& __value) =3D delete;
+
+  template <typename _Tp, typename _Ap>
+    void where(bool __k, const simd<_Tp, _Ap>& __value) =3D delete;
+
+// proposed mask iterations {{{1
+namespace __proposed {
+template <size_t _Np>
+  class where_range
+  {
+    const bitset<_Np> __bits;
+
+  public:
+    where_range(bitset<_Np> __b) : __bits(__b) {}
+
+    class iterator
+    {
+      size_t __mask;
+      size_t __bit;
+
+      _GLIBCXX_SIMD_INTRINSIC void __next_bit()
+      { __bit =3D __builtin_ctzl(__mask); }
+
+      _GLIBCXX_SIMD_INTRINSIC void __reset_lsb()
+      {
+	// 01100100 - 1 =3D 01100011
+	__mask &=3D (__mask - 1);
+	// __asm__("btr %1,%0" : "+r"(__mask) : "r"(__bit));
+      }
+
+    public:
+      iterator(decltype(__mask) __m) : __mask(__m) { __next_bit(); }
+      iterator(const iterator&) =3D default;
+      iterator(iterator&&) =3D default;
+
+      _GLIBCXX_SIMD_ALWAYS_INLINE size_t operator->() const
+      { return __bit; }
+
+      _GLIBCXX_SIMD_ALWAYS_INLINE size_t operator*() const
+      { return __bit; }
+
+      _GLIBCXX_SIMD_ALWAYS_INLINE iterator& operator++()
+      {
+	__reset_lsb();
+	__next_bit();
+	return *this;
+      }
+
+      _GLIBCXX_SIMD_ALWAYS_INLINE iterator operator++(int)
+      {
+	iterator __tmp =3D *this;
+	__reset_lsb();
+	__next_bit();
+	return __tmp;
+      }
+
+      _GLIBCXX_SIMD_ALWAYS_INLINE bool operator=3D=3D(const iterator& __rh=
s) const
+      { return __mask =3D=3D __rhs.__mask; }
+
+      _GLIBCXX_SIMD_ALWAYS_INLINE bool operator!=3D(const iterator& __rhs)=
 const
+      { return __mask !=3D __rhs.__mask; }
+    };
+
+    iterator begin() const
+    { return __bits.to_ullong(); }
+
+    iterator end() const
+    { return 0; }
+  };
+
+template <typename _Tp, typename _Ap>
+  where_range<simd_size_v<_Tp, _Ap>>
+  where(const simd_mask<_Tp, _Ap>& __k)
+  { return __k.__to_bitset(); }
+
+} // namespace __proposed
+
+// }}}1
+// reductions [simd.reductions] {{{1
+  template <typename _Tp, typename _Abi, typename _BinaryOperation =3D plu=
s<>>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR _Tp
+  reduce(const simd<_Tp, _Abi>& __v,
+	 _BinaryOperation __binary_op =3D _BinaryOperation())
+  { return _Abi::_SimdImpl::_S_reduce(__v, __binary_op); }
+
+template <typename _M, typename _V, typename _BinaryOperation =3D plus<>>
+  _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
+  reduce(const const_where_expression<_M, _V>& __x,
+	 typename _V::value_type __identity_element,
+	 _BinaryOperation __binary_op)
+  {
+    if (__builtin_expect(none_of(__get_mask(__x)), false))
+      return __identity_element;
+
+    _V __tmp =3D __identity_element;
+    _V::_Impl::_S_masked_assign(__data(__get_mask(__x)), __data(__tmp),
+				__data(__get_lvalue(__x)));
+    return reduce(__tmp, __binary_op);
+  }
+
+template <typename _M, typename _V>
+  _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
+  reduce(const const_where_expression<_M, _V>& __x, plus<> __binary_op =3D=
 {})
+  { return reduce(__x, 0, __binary_op); }
+
+template <typename _M, typename _V>
+  _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
+  reduce(const const_where_expression<_M, _V>& __x, multiplies<> __binary_=
op)
+  { return reduce(__x, 1, __binary_op); }
+
+template <typename _M, typename _V>
+  _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
+  reduce(const const_where_expression<_M, _V>& __x, bit_and<> __binary_op)
+  { return reduce(__x, ~typename _V::value_type(), __binary_op); }
+
+template <typename _M, typename _V>
+  _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
+  reduce(const const_where_expression<_M, _V>& __x, bit_or<> __binary_op)
+  { return reduce(__x, 0, __binary_op); }
+
+template <typename _M, typename _V>
+  _GLIBCXX_SIMD_INTRINSIC typename _V::value_type
+  reduce(const const_where_expression<_M, _V>& __x, bit_xor<> __binary_op)
+  { return reduce(__x, 0, __binary_op); }
+
+// }}}1
+// algorithms [simd.alg] {{{
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
+  min(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
+  { return {__private_init, _Ap::_SimdImpl::_S_min(__data(__a), __data(__b=
))}; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
+  max(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
+  { return {__private_init, _Ap::_SimdImpl::_S_max(__data(__a), __data(__b=
))}; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
+  pair<simd<_Tp, _Ap>, simd<_Tp, _Ap>>
+  minmax(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
+  {
+    const auto pair_of_members
+      =3D _Ap::_SimdImpl::_S_minmax(__data(__a), __data(__b));
+    return {simd<_Tp, _Ap>(__private_init, pair_of_members.first),
+	    simd<_Tp, _Ap>(__private_init, pair_of_members.second)};
+  }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
+  clamp(const simd<_Tp, _Ap>& __v, const simd<_Tp, _Ap>& __lo,
+	const simd<_Tp, _Ap>& __hi)
+  {
+    using _Impl =3D typename _Ap::_SimdImpl;
+    return {__private_init,
+	    _Impl::_S_min(__data(__hi),
+			  _Impl::_S_max(__data(__lo), __data(__v)))};
+  }
+
+// }}}
+
+template <size_t... _Sizes, typename _Tp, typename _Ap,
+	  typename =3D enable_if_t<((_Sizes + ...) =3D=3D simd<_Tp, _Ap>::size())=
>>
+  inline tuple<simd<_Tp, simd_abi::deduce_t<_Tp, _Sizes>>...>
+  split(const simd<_Tp, _Ap>&);
+
+// __extract_part {{{
+template <int _Index, int _Total, int _Combine =3D 1, typename _Tp, size_t=
 _Np>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST
+  _SimdWrapper<_Tp, _Np / _Total * _Combine>
+  __extract_part(const _SimdWrapper<_Tp, _Np> __x);
+
+template <int Index, int Parts, int _Combine =3D 1, typename _Tp, typename=
 _A0,
+	  typename... _As>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __extract_part(const _SimdTuple<_Tp, _A0, _As...>& __x);
+
+// }}}
+// _SizeList {{{
+template <size_t _V0, size_t... _Values>
+  struct _SizeList
+  {
+    template <size_t _I>
+      static constexpr size_t _S_at(_SizeConstant<_I> =3D {})
+      {
+	if constexpr (_I =3D=3D 0)
+	  return _V0;
+	else
+	  return _SizeList<_Values...>::template _S_at<_I - 1>();
+      }
+
+    template <size_t _I>
+      static constexpr auto _S_before(_SizeConstant<_I> =3D {})
+      {
+	if constexpr (_I =3D=3D 0)
+	  return _SizeConstant<0>();
+	else
+	  return _SizeConstant<
+	    _V0 + _SizeList<_Values...>::template _S_before<_I - 1>()>();
+      }
+
+    template <size_t _Np>
+      static constexpr auto _S_pop_front(_SizeConstant<_Np> =3D {})
+      {
+	if constexpr (_Np =3D=3D 0)
+	  return _SizeList();
+	else
+	  return _SizeList<_Values...>::template _S_pop_front<_Np - 1>();
+      }
+  };
+
+// }}}
+// __extract_center {{{
+template <typename _Tp, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC _SimdWrapper<_Tp, _Np / 2>
+  __extract_center(_SimdWrapper<_Tp, _Np> __x)
+  {
+    static_assert(_Np >=3D 4);
+    static_assert(_Np % 4 =3D=3D 0); // x0 - x1 - x2 - x3 -> return {x1, x=
2}
+#if _GLIBCXX_SIMD_X86INTRIN    // {{{
+    if constexpr (__have_avx512f && sizeof(_Tp) * _Np =3D=3D 64)
+      {
+	const auto __intrin =3D __to_intrin(__x);
+	if constexpr (is_integral_v<_Tp>)
+	  return __vector_bitcast<_Tp>(_mm512_castsi512_si256(
+	    _mm512_shuffle_i32x4(__intrin, __intrin,
+				 1 + 2 * 0x4 + 2 * 0x10 + 3 * 0x40)));
+	else if constexpr (sizeof(_Tp) =3D=3D 4)
+	  return __vector_bitcast<_Tp>(_mm512_castps512_ps256(
+	    _mm512_shuffle_f32x4(__intrin, __intrin,
+				 1 + 2 * 0x4 + 2 * 0x10 + 3 * 0x40)));
+	else if constexpr (sizeof(_Tp) =3D=3D 8)
+	  return __vector_bitcast<_Tp>(_mm512_castpd512_pd256(
+	    _mm512_shuffle_f64x2(__intrin, __intrin,
+				 1 + 2 * 0x4 + 2 * 0x10 + 3 * 0x40)));
+	else
+	  __assert_unreachable<_Tp>();
+      }
+    else if constexpr (sizeof(_Tp) * _Np =3D=3D 32 && is_floating_point_v<=
_Tp>)
+      return __vector_bitcast<_Tp>(
+	_mm_shuffle_pd(__lo128(__vector_bitcast<double>(__x)),
+		       __hi128(__vector_bitcast<double>(__x)), 1));
+    else if constexpr (sizeof(__x) =3D=3D 32 && sizeof(_Tp) * _Np <=3D 32)
+      return __vector_bitcast<_Tp>(
+	_mm_alignr_epi8(__hi128(__vector_bitcast<_LLong>(__x)),
+			__lo128(__vector_bitcast<_LLong>(__x)),
+			sizeof(_Tp) * _Np / 4));
+    else
+#endif // _GLIBCXX_SIMD_X86INTRIN }}}
+      {
+	__vector_type_t<_Tp, _Np / 2> __r;
+	__builtin_memcpy(&__r,
+			 reinterpret_cast<const char*>(&__x)
+			   + sizeof(_Tp) * _Np / 4,
+			 sizeof(_Tp) * _Np / 2);
+	return __r;
+      }
+  }
+
+template <typename _Tp, typename _A0, typename... _As>
+  _GLIBCXX_SIMD_INTRINSIC
+  _SimdWrapper<_Tp, _SimdTuple<_Tp, _A0, _As...>::_S_size() / 2>
+  __extract_center(const _SimdTuple<_Tp, _A0, _As...>& __x)
+  {
+    if constexpr (sizeof...(_As) =3D=3D 0)
+      return __extract_center(__x.first);
+    else
+      return __extract_part<1, 4, 2>(__x);
+  }
+
+// }}}
+// __split_wrapper {{{
+template <size_t... _Sizes, typename _Tp, typename... _As>
+  auto
+  __split_wrapper(_SizeList<_Sizes...>, const _SimdTuple<_Tp, _As...>& __x)
+  {
+    return split<_Sizes...>(
+      fixed_size_simd<_Tp, _SimdTuple<_Tp, _As...>::_S_size()>(__private_i=
nit,
+							       __x));
+  }
+
+// }}}
+
+// split<simd>(simd) {{{
+template <typename _V, typename _Ap,
+	  size_t Parts =3D simd_size_v<typename _V::value_type, _Ap> / _V::size()>
+  enable_if_t<simd_size_v<typename _V::value_type, _Ap> =3D=3D Parts * _V:=
:size()
+	      && is_simd_v<_V>, array<_V, Parts>>
+  split(const simd<typename _V::value_type, _Ap>& __x)
+  {
+    using _Tp =3D typename _V::value_type;
+    if constexpr (Parts =3D=3D 1)
+      {
+	return {simd_cast<_V>(__x)};
+      }
+    else if (__x._M_is_constprop())
+      {
+	return __generate_from_n_evaluations<Parts, array<_V, Parts>>([&](
+	  auto __i) constexpr {
+	  return _V([&](auto __j) constexpr {
+	    return __x[__i * _V::size() + __j];
+	  });
+	});
+      }
+    else if constexpr (
+      __is_fixed_size_abi_v<_Ap>
+      && (is_same_v<typename _V::abi_type, simd_abi::scalar>
+	|| (__is_fixed_size_abi_v<typename _V::abi_type>
+	  && sizeof(_V) =3D=3D sizeof(_Tp) * _V::size() // _V doesn't have padding
+	  )))
+      {
+	// fixed_size -> fixed_size (w/o padding) or scalar
+#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
+      const __may_alias<_Tp>* const __element_ptr
+	=3D reinterpret_cast<const __may_alias<_Tp>*>(&__data(__x));
+      return __generate_from_n_evaluations<Parts, array<_V, Parts>>([&](
+	auto __i) constexpr {
+	return _V(__element_ptr + __i * _V::size(), vector_aligned);
+      });
+#else
+      const auto& __xx =3D __data(__x);
+      return __generate_from_n_evaluations<Parts, array<_V, Parts>>([&](
+	auto __i) constexpr {
+	[[maybe_unused]] constexpr size_t __offset
+	  =3D decltype(__i)::value * _V::size();
+	return _V([&](auto __j) constexpr {
+	  constexpr _SizeConstant<__j + __offset> __k;
+	  return __xx[__k];
+	});
+      });
+#endif
+    }
+  else if constexpr (is_same_v<typename _V::abi_type, simd_abi::scalar>)
+    {
+      // normally memcpy should work here as well
+      return __generate_from_n_evaluations<Parts, array<_V, Parts>>([&](
+	auto __i) constexpr { return __x[__i]; });
+    }
+  else
+    {
+      return __generate_from_n_evaluations<Parts, array<_V, Parts>>([&](
+	auto __i) constexpr {
+	if constexpr (__is_fixed_size_abi_v<typename _V::abi_type>)
+	  return _V([&](auto __j) constexpr {
+	    return __x[__i * _V::size() + __j];
+	  });
+	else
+	  return _V(__private_init,
+		    __extract_part<decltype(__i)::value, Parts>(__data(__x)));
+      });
+    }
+  }
+
+// }}}
+// split<simd_mask>(simd_mask) {{{
+template <typename _V, typename _Ap,
+	  size_t _Parts
+	  =3D simd_size_v<typename _V::simd_type::value_type, _Ap> / _V::size()>
+  enable_if_t<is_simd_mask_v<_V> && simd_size_v<typename
+    _V::simd_type::value_type, _Ap> =3D=3D _Parts * _V::size(), array<_V, =
_Parts>>
+  split(const simd_mask<typename _V::simd_type::value_type, _Ap>& __x)
+  {
+    if constexpr (is_same_v<_Ap, typename _V::abi_type>)
+      return {__x};
+    else if constexpr (_Parts =3D=3D 1)
+      return {__proposed::static_simd_cast<_V>(__x)};
+    else if constexpr (_Parts =3D=3D 2 && __is_sse_abi<typename _V::abi_ty=
pe>()
+		       && __is_avx_abi<_Ap>())
+      return {_V(__private_init, __lo128(__data(__x))),
+	      _V(__private_init, __hi128(__data(__x)))};
+    else if constexpr (_V::size() <=3D __CHAR_BIT__ * sizeof(_ULLong))
+      {
+	const bitset __bits =3D __x.__to_bitset();
+	return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
+	  auto __i) constexpr {
+	  constexpr size_t __offset =3D __i * _V::size();
+	  return _V(__bitset_init, (__bits >> __offset).to_ullong());
+	});
+      }
+    else
+      {
+	return __generate_from_n_evaluations<_Parts, array<_V, _Parts>>([&](
+	  auto __i) constexpr {
+	  constexpr size_t __offset =3D __i * _V::size();
+	  return _V(
+	    __private_init, [&](auto __j) constexpr {
+	      return __x[__j + __offset];
+	    });
+	});
+      }
+  }
+
+// }}}
+// split<_Sizes...>(simd) {{{
+template <size_t... _Sizes, typename _Tp, typename _Ap, typename>
+  _GLIBCXX_SIMD_ALWAYS_INLINE
+  tuple<simd<_Tp, simd_abi::deduce_t<_Tp, _Sizes>>...>
+  split(const simd<_Tp, _Ap>& __x)
+  {
+    using _SL =3D _SizeList<_Sizes...>;
+    using _Tuple =3D tuple<__deduced_simd<_Tp, _Sizes>...>;
+    constexpr size_t _Np =3D simd_size_v<_Tp, _Ap>;
+    constexpr size_t _N0 =3D _SL::template _S_at<0>();
+    using _V =3D __deduced_simd<_Tp, _N0>;
+
+    if (__x._M_is_constprop())
+      return __generate_from_n_evaluations<sizeof...(_Sizes), _Tuple>([&](
+	auto __i) constexpr {
+	using _Vi =3D __deduced_simd<_Tp, _SL::_S_at(__i)>;
+	constexpr size_t __offset =3D _SL::_S_before(__i);
+	return _Vi([&](auto __j) constexpr { return __x[__offset + __j]; });
+      });
+    else if constexpr (_Np =3D=3D _N0)
+      {
+	static_assert(sizeof...(_Sizes) =3D=3D 1);
+	return {simd_cast<_V>(__x)};
+      }
+    else if constexpr // split from fixed_size, such that __x::first.size =
=3D=3D _N0
+      (__is_fixed_size_abi_v<
+	 _Ap> && __fixed_size_storage_t<_Tp, _Np>::_S_first_size =3D=3D _N0)
+      {
+	static_assert(
+	  !__is_fixed_size_abi_v<typename _V::abi_type>,
+	  "How can <_Tp, _Np> be __a single _SimdTuple entry but __a "
+	  "fixed_size_simd "
+	  "when deduced?");
+	// extract first and recurse (__split_wrapper is needed to deduce a new
+	// _Sizes pack)
+	return tuple_cat(make_tuple(_V(__private_init, __data(__x).first)),
+			 __split_wrapper(_SL::template _S_pop_front<1>(),
+					 __data(__x).second));
+      }
+    else if constexpr ((!is_same_v<simd_abi::scalar,
+				   simd_abi::deduce_t<_Tp, _Sizes>> && ...)
+		       && (!__is_fixed_size_abi_v<
+			     simd_abi::deduce_t<_Tp, _Sizes>> && ...))
+      {
+	if constexpr (((_Sizes * 2 =3D=3D _Np) && ...))
+	  return {{__private_init, __extract_part<0, 2>(__data(__x))},
+		  {__private_init, __extract_part<1, 2>(__data(__x))}};
+	else if constexpr (is_same_v<_SizeList<_Sizes...>,
+				     _SizeList<_Np / 3, _Np / 3, _Np / 3>>)
+	  return {{__private_init, __extract_part<0, 3>(__data(__x))},
+		  {__private_init, __extract_part<1, 3>(__data(__x))},
+		  {__private_init, __extract_part<2, 3>(__data(__x))}};
+	else if constexpr (is_same_v<_SizeList<_Sizes...>,
+				     _SizeList<2 * _Np / 3, _Np / 3>>)
+	  return {{__private_init, __extract_part<0, 3, 2>(__data(__x))},
+		  {__private_init, __extract_part<2, 3>(__data(__x))}};
+	else if constexpr (is_same_v<_SizeList<_Sizes...>,
+				     _SizeList<_Np / 3, 2 * _Np / 3>>)
+	  return {{__private_init, __extract_part<0, 3>(__data(__x))},
+		  {__private_init, __extract_part<1, 3, 2>(__data(__x))}};
+	else if constexpr (is_same_v<_SizeList<_Sizes...>,
+				     _SizeList<_Np / 2, _Np / 4, _Np / 4>>)
+	  return {{__private_init, __extract_part<0, 2>(__data(__x))},
+		  {__private_init, __extract_part<2, 4>(__data(__x))},
+		  {__private_init, __extract_part<3, 4>(__data(__x))}};
+	else if constexpr (is_same_v<_SizeList<_Sizes...>,
+				     _SizeList<_Np / 4, _Np / 4, _Np / 2>>)
+	  return {{__private_init, __extract_part<0, 4>(__data(__x))},
+		  {__private_init, __extract_part<1, 4>(__data(__x))},
+		  {__private_init, __extract_part<1, 2>(__data(__x))}};
+	else if constexpr (is_same_v<_SizeList<_Sizes...>,
+				     _SizeList<_Np / 4, _Np / 2, _Np / 4>>)
+	  return {{__private_init, __extract_part<0, 4>(__data(__x))},
+		  {__private_init, __extract_center(__data(__x))},
+		  {__private_init, __extract_part<3, 4>(__data(__x))}};
+	else if constexpr (((_Sizes * 4 =3D=3D _Np) && ...))
+	  return {{__private_init, __extract_part<0, 4>(__data(__x))},
+		  {__private_init, __extract_part<1, 4>(__data(__x))},
+		  {__private_init, __extract_part<2, 4>(__data(__x))},
+		  {__private_init, __extract_part<3, 4>(__data(__x))}};
+	// else fall through
+      }
+#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
+    const __may_alias<_Tp>* const __element_ptr
+      =3D reinterpret_cast<const __may_alias<_Tp>*>(&__x);
+    return __generate_from_n_evaluations<sizeof...(_Sizes), _Tuple>([&](
+      auto __i) constexpr {
+      using _Vi =3D __deduced_simd<_Tp, _SL::_S_at(__i)>;
+      constexpr size_t __offset =3D _SL::_S_before(__i);
+      constexpr size_t __base_align =3D alignof(simd<_Tp, _Ap>);
+      constexpr size_t __a
+	=3D __base_align - ((__offset * sizeof(_Tp)) % __base_align);
+      constexpr size_t __b =3D ((__a - 1) & __a) ^ __a;
+      constexpr size_t __alignment =3D __b =3D=3D 0 ? __a : __b;
+      return _Vi(__element_ptr + __offset, overaligned<__alignment>);
+    });
+#else
+    return __generate_from_n_evaluations<sizeof...(_Sizes), _Tuple>([&](
+      auto __i) constexpr {
+      using _Vi =3D __deduced_simd<_Tp, _SL::_S_at(__i)>;
+      const auto& __xx =3D __data(__x);
+      using _Offset =3D decltype(_SL::_S_before(__i));
+      return _Vi([&](auto __j) constexpr {
+	constexpr _SizeConstant<_Offset::value + __j> __k;
+	return __xx[__k];
+      });
+    });
+#endif
+  }
+
+// }}}
+
+// __subscript_in_pack {{{
+template <size_t _I, typename _Tp, typename _Ap, typename... _As>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
+  __subscript_in_pack(const simd<_Tp, _Ap>& __x, const simd<_Tp, _As>&... =
__xs)
+  {
+    if constexpr (_I < simd_size_v<_Tp, _Ap>)
+      return __x[_I];
+    else
+      return __subscript_in_pack<_I - simd_size_v<_Tp, _Ap>>(__xs...);
+  }
+
+// }}}
+// __store_pack_of_simd {{{
+template <typename _Tp, typename _A0, typename... _As>
+  _GLIBCXX_SIMD_INTRINSIC void
+  __store_pack_of_simd(char* __mem, const simd<_Tp, _A0>& __x0,
+		       const simd<_Tp, _As>&... __xs)
+  {
+    constexpr size_t __n_bytes =3D sizeof(_Tp) * simd_size_v<_Tp, _A0>;
+    __builtin_memcpy(__mem, &__data(__x0), __n_bytes);
+    if constexpr (sizeof...(__xs) > 0)
+      __store_pack_of_simd(__mem + __n_bytes, __xs...);
+  }
+
+// }}}
+// concat(simd...) {{{
+template <typename _Tp, typename... _As>
+  inline _GLIBCXX_SIMD_CONSTEXPR
+  simd<_Tp, simd_abi::deduce_t<_Tp, (simd_size_v<_Tp, _As> + ...)>>
+  concat(const simd<_Tp, _As>&... __xs)
+  {
+    using _Rp =3D __deduced_simd<_Tp, (simd_size_v<_Tp, _As> + ...)>;
+    if constexpr (sizeof...(__xs) =3D=3D 1)
+      return simd_cast<_Rp>(__xs...);
+    else if ((... && __xs._M_is_constprop()))
+      return simd<_Tp,
+		  simd_abi::deduce_t<_Tp, (simd_size_v<_Tp, _As> + ...)>>([&](
+	auto __i) constexpr { return __subscript_in_pack<__i>(__xs...); });
+    else
+      {
+	_Rp __r{};
+	__store_pack_of_simd(reinterpret_cast<char*>(&__data(__r)), __xs...);
+	return __r;
+      }
+  }
+
+// }}}
+// concat(array<simd>) {{{
+template <typename _Tp, typename _Abi, size_t _Np>
+  _GLIBCXX_SIMD_ALWAYS_INLINE
+  _GLIBCXX_SIMD_CONSTEXPR __deduced_simd<_Tp, simd_size_v<_Tp, _Abi> * _Np>
+  concat(const array<simd<_Tp, _Abi>, _Np>& __x)
+  {
+    return __call_with_subscripts<_Np>(__x, [](const auto&... __xs) {
+      return concat(__xs...);
+    });
+  }
+
+// }}}
+
+// _SmartReference {{{
+template <typename _Up, typename _Accessor =3D _Up,
+	  typename _ValueType =3D typename _Up::value_type>
+  class _SmartReference
+  {
+    friend _Accessor;
+    int _M_index;
+    _Up& _M_obj;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _ValueType _M_read() const noexcept
+    {
+      if constexpr (is_arithmetic_v<_Up>)
+	return _M_obj;
+      else
+	return _M_obj[_M_index];
+    }
+
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC constexpr void _M_write(_Tp&& __x) const
+      { _Accessor::_S_set(_M_obj, _M_index, static_cast<_Tp&&>(__x)); }
+
+  public:
+    _GLIBCXX_SIMD_INTRINSIC constexpr
+    _SmartReference(_Up& __o, int __i) noexcept
+    : _M_index(__i), _M_obj(__o) {}
+
+    using value_type =3D _ValueType;
+
+    _GLIBCXX_SIMD_INTRINSIC _SmartReference(const _SmartReference&) =3D de=
lete;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr operator value_type() const noexcept
+    { return _M_read(); }
+
+    template <typename _Tp,
+	      typename
+	      =3D _ValuePreservingOrInt<__remove_cvref_t<_Tp>, value_type>>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference operator=3D(_Tp&& =
__x) &&
+      {
+	_M_write(static_cast<_Tp&&>(__x));
+	return {_M_obj, _M_index};
+      }
+
+#define _GLIBCXX_SIMD_OP_(__op)                                           =
     \
+    template <typename _Tp,                                               =
     \
+	      typename _TT                                                     \
+	      =3D decltype(declval<value_type>() __op declval<_Tp>()),           \
+	      typename =3D _ValuePreservingOrInt<__remove_cvref_t<_Tp>, _TT>,    \
+	      typename =3D _ValuePreservingOrInt<_TT, value_type>>               \
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference                   =
     \
+      operator __op##=3D(_Tp&& __x) &&                                    =
       \
+      {                                                                   =
     \
+	const value_type& __lhs =3D _M_read();                                   \
+	_M_write(__lhs __op __x);                                              \
+	return {_M_obj, _M_index};                                             \
+      }
+    _GLIBCXX_SIMD_ALL_ARITHMETICS(_GLIBCXX_SIMD_OP_);
+    _GLIBCXX_SIMD_ALL_SHIFTS(_GLIBCXX_SIMD_OP_);
+    _GLIBCXX_SIMD_ALL_BINARY(_GLIBCXX_SIMD_OP_);
+#undef _GLIBCXX_SIMD_OP_
+
+    template <typename _Tp =3D void,
+	      typename
+	      =3D decltype(++declval<conditional_t<true, value_type, _Tp>&>())>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference operator++() &&
+      {
+	value_type __x =3D _M_read();
+	_M_write(++__x);
+	return {_M_obj, _M_index};
+      }
+
+    template <typename _Tp =3D void,
+	      typename
+	      =3D decltype(declval<conditional_t<true, value_type, _Tp>&>()++)>
+      _GLIBCXX_SIMD_INTRINSIC constexpr value_type operator++(int) &&
+      {
+	const value_type __r =3D _M_read();
+	value_type __x =3D __r;
+	_M_write(++__x);
+	return __r;
+      }
+
+    template <typename _Tp =3D void,
+	      typename
+	      =3D decltype(--declval<conditional_t<true, value_type, _Tp>&>())>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SmartReference operator--() &&
+      {
+	value_type __x =3D _M_read();
+	_M_write(--__x);
+	return {_M_obj, _M_index};
+      }
+
+    template <typename _Tp =3D void,
+	      typename
+	      =3D decltype(declval<conditional_t<true, value_type, _Tp>&>()--)>
+      _GLIBCXX_SIMD_INTRINSIC constexpr value_type operator--(int) &&
+      {
+	const value_type __r =3D _M_read();
+	value_type __x =3D __r;
+	_M_write(--__x);
+	return __r;
+      }
+
+    _GLIBCXX_SIMD_INTRINSIC friend void
+    swap(_SmartReference&& __a, _SmartReference&& __b) noexcept(
+      conjunction<
+	is_nothrow_constructible<value_type, _SmartReference&&>,
+	is_nothrow_assignable<_SmartReference&&, value_type&&>>::value)
+    {
+      value_type __tmp =3D static_cast<_SmartReference&&>(__a);
+      static_cast<_SmartReference&&>(__a) =3D static_cast<value_type>(__b);
+      static_cast<_SmartReference&&>(__b) =3D std::move(__tmp);
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC friend void
+    swap(value_type& __a, _SmartReference&& __b) noexcept(
+      conjunction<
+	is_nothrow_constructible<value_type, value_type&&>,
+	is_nothrow_assignable<value_type&, value_type&&>,
+	is_nothrow_assignable<_SmartReference&&, value_type&&>>::value)
+    {
+      value_type __tmp(std::move(__a));
+      __a =3D static_cast<value_type>(__b);
+      static_cast<_SmartReference&&>(__b) =3D std::move(__tmp);
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC friend void
+    swap(_SmartReference&& __a, value_type& __b) noexcept(
+      conjunction<
+	is_nothrow_constructible<value_type, _SmartReference&&>,
+	is_nothrow_assignable<value_type&, value_type&&>,
+	is_nothrow_assignable<_SmartReference&&, value_type&&>>::value)
+    {
+      value_type __tmp(__a);
+      static_cast<_SmartReference&&>(__a) =3D std::move(__b);
+      __b =3D std::move(__tmp);
+    }
+  };
+
+// }}}
+// __scalar_abi_wrapper {{{
+template <int _Bytes>
+  struct __scalar_abi_wrapper
+  {
+    template <typename _Tp> static constexpr size_t _S_full_size =3D 1;
+    template <typename _Tp> static constexpr size_t _S_size =3D 1;
+    template <typename _Tp> static constexpr size_t _S_is_partial =3D fals=
e;
+
+    template <typename _Tp, typename _Abi =3D simd_abi::scalar>
+      static constexpr bool _S_is_valid_v
+	=3D _Abi::template _IsValid<_Tp>::value && sizeof(_Tp) =3D=3D _Bytes;
+  };
+
+// }}}
+// __decay_abi metafunction {{{
+template <typename _Tp>
+  struct __decay_abi { using type =3D _Tp; };
+
+template <int _Bytes>
+  struct __decay_abi<__scalar_abi_wrapper<_Bytes>>
+  { using type =3D simd_abi::scalar; };
+
+// }}}
+// __find_next_valid_abi metafunction {{{1
+// Given an ABI tag A<N>, find an N2 < N such that A<N2>::_S_is_valid_v<_T=
p> =3D=3D
+// true, N2 is a power-of-2, and A<N2>::_S_is_partial<_Tp> is false. Break
+// recursion at 2 elements in the resulting ABI tag. In this case
+// type::_S_is_valid_v<_Tp> may be false.
+template <template <int> class _Abi, int _Bytes, typename _Tp>
+  struct __find_next_valid_abi
+  {
+    static constexpr auto _S_choose()
+    {
+      constexpr int _NextBytes =3D std::__bit_ceil(_Bytes) / 2;
+      using _NextAbi =3D _Abi<_NextBytes>;
+      if constexpr (_NextBytes < sizeof(_Tp) * 2) // break recursion
+	return _Abi<_Bytes>();
+      else if constexpr (_NextAbi::template _S_is_partial<_Tp> =3D=3D false
+			 && _NextAbi::template _S_is_valid_v<_Tp>)
+	return _NextAbi();
+      else
+	return __find_next_valid_abi<_Abi, _NextBytes, _Tp>::_S_choose();
+    }
+
+    using type =3D decltype(_S_choose());
+  };
+
+template <int _Bytes, typename _Tp>
+  struct __find_next_valid_abi<__scalar_abi_wrapper, _Bytes, _Tp>
+  { using type =3D simd_abi::scalar; };
+
+// _AbiList {{{1
+template <template <int> class...>
+  struct _AbiList
+  {
+    template <typename, int> static constexpr bool _S_has_valid_abi =3D fa=
lse;
+    template <typename, int> using _FirstValidAbi =3D void;
+    template <typename, int> using _BestAbi =3D void;
+  };
+
+template <template <int> class _A0, template <int> class... _Rest>
+  struct _AbiList<_A0, _Rest...>
+  {
+    template <typename _Tp, int _Np>
+      static constexpr bool _S_has_valid_abi
+	=3D _A0<sizeof(_Tp) * _Np>::template _S_is_valid_v<
+	    _Tp> || _AbiList<_Rest...>::template _S_has_valid_abi<_Tp, _Np>;
+
+    template <typename _Tp, int _Np>
+      using _FirstValidAbi =3D conditional_t<
+	_A0<sizeof(_Tp) * _Np>::template _S_is_valid_v<_Tp>,
+	typename __decay_abi<_A0<sizeof(_Tp) * _Np>>::type,
+	typename _AbiList<_Rest...>::template _FirstValidAbi<_Tp, _Np>>;
+
+    template <typename _Tp, int _Np>
+      static constexpr auto _S_determine_best_abi()
+      {
+	static_assert(_Np >=3D 1);
+	constexpr int _Bytes =3D sizeof(_Tp) * _Np;
+	if constexpr (_Np =3D=3D 1)
+	  return __make_dependent_t<_Tp, simd_abi::scalar>{};
+	else
+	  {
+	    constexpr int __fullsize =3D _A0<_Bytes>::template _S_full_size<_Tp>;
+	    // _A0<_Bytes> is good if:
+	    // 1. The ABI tag is valid for _Tp
+	    // 2. The storage overhead is no more than padding to fill the next
+	    //    power-of-2 number of bytes
+	    if constexpr (_A0<_Bytes>::template _S_is_valid_v<
+			    _Tp> && __fullsize / 2 < _Np)
+	      return typename __decay_abi<_A0<_Bytes>>::type{};
+	    else
+	      {
+		using _B =3D
+		  typename __find_next_valid_abi<_A0, _Bytes, _Tp>::type;
+		if constexpr (_B::template _S_is_valid_v<
+				_Tp> && _B::template _S_size<_Tp> <=3D _Np)
+		  return _B{};
+		else
+		  return
+		    typename _AbiList<_Rest...>::template _BestAbi<_Tp, _Np>{};
+	      }
+	  }
+      }
+
+    template <typename _Tp, int _Np>
+      using _BestAbi =3D decltype(_S_determine_best_abi<_Tp, _Np>());
+  };
+
+// }}}1
+
+// the following lists all native ABIs, which makes them accessible to
+// simd_abi::deduce and select_best_vector_type_t (for fixed_size). Order
+// matters: Whatever comes first has higher priority.
+using _AllNativeAbis =3D _AbiList<simd_abi::_VecBltnBtmsk, simd_abi::_VecB=
uiltin,
+				__scalar_abi_wrapper>;
+
+// valid _SimdTraits specialization {{{1
+template <typename _Tp, typename _Abi>
+  struct _SimdTraits<_Tp, _Abi, void_t<typename _Abi::template _IsValid<_T=
p>>>
+  : _Abi::template __traits<_Tp> {};
+
+// __deduce_impl specializations {{{1
+// try all native ABIs (including scalar) first
+template <typename _Tp, size_t _Np>
+  struct __deduce_impl<
+    _Tp, _Np, enable_if_t<_AllNativeAbis::template _S_has_valid_abi<_Tp, _=
Np>>>
+  { using type =3D _AllNativeAbis::_FirstValidAbi<_Tp, _Np>; };
+
+// fall back to fixed_size only if scalar and native ABIs don't match
+template <typename _Tp, size_t _Np, typename =3D void>
+  struct __deduce_fixed_size_fallback {};
+
+template <typename _Tp, size_t _Np>
+  struct __deduce_fixed_size_fallback<_Tp, _Np,
+    enable_if_t<simd_abi::fixed_size<_Np>::template _S_is_valid_v<_Tp>>>
+  { using type =3D simd_abi::fixed_size<_Np>; };
+
+template <typename _Tp, size_t _Np, typename>
+  struct __deduce_impl : public __deduce_fixed_size_fallback<_Tp, _Np> {};
+
+//}}}1
+
+// simd_mask {{{
+template <typename _Tp, typename _Abi>
+  class simd_mask : public _SimdTraits<_Tp, _Abi>::_MaskBase
+  {
+    // types, tags, and friends {{{
+    using _Traits =3D _SimdTraits<_Tp, _Abi>;
+    using _MemberType =3D typename _Traits::_MaskMember;
+
+    // We map all masks with equal element sizeof to a single integer type=
, the
+    // one given by __int_for_sizeof_t<_Tp>. This is the approach
+    // [[gnu::vector_size(N)]] types take as well and it reduces the numbe=
r of
+    // template specializations in the implementation classes.
+    using _Ip =3D __int_for_sizeof_t<_Tp>;
+    static constexpr _Ip* _S_type_tag =3D nullptr;
+
+    friend typename _Traits::_MaskBase;
+    friend class simd<_Tp, _Abi>;       // to construct masks on return
+    friend typename _Traits::_SimdImpl; // to construct masks on return and
+					// inspect data on masked operations
+  public:
+    using _Impl =3D typename _Traits::_MaskImpl;
+    friend _Impl;
+
+    // }}}
+    // member types {{{
+    using value_type =3D bool;
+    using reference =3D _SmartReference<_MemberType, _Impl, value_type>;
+    using simd_type =3D simd<_Tp, _Abi>;
+    using abi_type =3D _Abi;
+
+    // }}}
+    static constexpr size_t size() // {{{
+    { return __size_or_zero_v<_Tp, _Abi>; }
+
+    // }}}
+    // constructors & assignment {{{
+    simd_mask() =3D default;
+    simd_mask(const simd_mask&) =3D default;
+    simd_mask(simd_mask&&) =3D default;
+    simd_mask& operator=3D(const simd_mask&) =3D default;
+    simd_mask& operator=3D(simd_mask&&) =3D default;
+
+    // }}}
+    // access to internal representation (optional feature) {{{
+    _GLIBCXX_SIMD_ALWAYS_INLINE explicit
+    simd_mask(typename _Traits::_MaskCastType __init)
+    : _M_data{__init} {}
+    // conversions to internal type is done in _MaskBase
+
+    // }}}
+    // bitset interface (extension to be proposed) {{{
+    // TS_FEEDBACK:
+    // Conversion of simd_mask to and from bitset makes it much easier to
+    // interface with other facilities. I suggest adding `static
+    // simd_mask::from_bitset` and `simd_mask::to_bitset`.
+    _GLIBCXX_SIMD_ALWAYS_INLINE static simd_mask
+    __from_bitset(bitset<size()> bs)
+    { return {__bitset_init, bs}; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE bitset<size()>
+    __to_bitset() const
+    { return _Impl::_S_to_bits(_M_data)._M_to_bitset(); }
+
+    // }}}
+    // explicit broadcast constructor {{{
+    _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
+    simd_mask(value_type __x)
+    : _M_data(_Impl::template _S_broadcast<_Ip>(__x)) {}
+
+    // }}}
+    // implicit type conversion constructor {{{
+  #ifdef _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
+    // proposed improvement
+    template <typename _Up, typename _A2,
+	      typename =3D enable_if_t<simd_size_v<_Up, _A2> =3D=3D size()>>
+      _GLIBCXX_SIMD_ALWAYS_INLINE explicit(sizeof(_MemberType)
+	  !=3D sizeof(typename _SimdTraits<_Up, _A2>::_MaskMember))
+      simd_mask(const simd_mask<_Up, _A2>& __x)
+      : simd_mask(__proposed::static_simd_cast<simd_mask>(__x)) {}
+  #else
+    // conforming to ISO/IEC 19570:2018
+    template <typename _Up, typename =3D enable_if_t<conjunction<
+			      is_same<abi_type, simd_abi::fixed_size<size()>>,
+			      is_same<_Up, _Up>>::value>>
+      _GLIBCXX_SIMD_ALWAYS_INLINE
+      simd_mask(const simd_mask<_Up, simd_abi::fixed_size<size()>>& __x)
+      : _M_data(_Impl::_S_from_bitmask(__data(__x), _S_type_tag)) {}
+  #endif
+
+    // }}}
+    // load constructor {{{
+    template <typename _Flags>
+      _GLIBCXX_SIMD_ALWAYS_INLINE
+      simd_mask(const value_type* __mem, _Flags)
+      : _M_data(_Impl::template _S_load<_Ip>(
+	_Flags::template _S_apply<simd_mask>(__mem))) {}
+
+    template <typename _Flags>
+      _GLIBCXX_SIMD_ALWAYS_INLINE
+      simd_mask(const value_type* __mem, simd_mask __k, _Flags)
+      : _M_data{}
+      {
+	_M_data
+	  =3D _Impl::_S_masked_load(_M_data, __k._M_data,
+				  _Flags::template _S_apply<simd_mask>(__mem));
+      }
+
+    // }}}
+    // loads [simd_mask.load] {{{
+    template <typename _Flags>
+      _GLIBCXX_SIMD_ALWAYS_INLINE void
+      copy_from(const value_type* __mem, _Flags)
+      {
+	_M_data =3D _Impl::template _S_load<_Ip>(
+	  _Flags::template _S_apply<simd_mask>(__mem));
+      }
+
+    // }}}
+    // stores [simd_mask.store] {{{
+    template <typename _Flags>
+      _GLIBCXX_SIMD_ALWAYS_INLINE void
+      copy_to(value_type* __mem, _Flags) const
+      { _Impl::_S_store(_M_data, _Flags::template _S_apply<simd_mask>(__me=
m)); }
+
+    // }}}
+    // scalar access {{{
+    _GLIBCXX_SIMD_ALWAYS_INLINE reference
+    operator[](size_t __i)
+    {
+      if (__i >=3D size())
+	__invoke_ub("Subscript %d is out of range [0, %d]", __i, size() - 1);
+      return {_M_data, int(__i)};
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE value_type
+    operator[](size_t __i) const
+    {
+      if (__i >=3D size())
+	__invoke_ub("Subscript %d is out of range [0, %d]", __i, size() - 1);
+      if constexpr (__is_scalar_abi<_Abi>())
+	return _M_data;
+      else
+	return static_cast<bool>(_M_data[__i]);
+    }
+
+    // }}}
+    // negation {{{
+    _GLIBCXX_SIMD_ALWAYS_INLINE simd_mask
+    operator!() const
+    { return {__private_init, _Impl::_S_bit_not(_M_data)}; }
+
+    // }}}
+    // simd_mask binary operators [simd_mask.binary] {{{
+  #ifdef _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
+    // simd_mask<int> && simd_mask<uint> needs disambiguation
+    template <typename _Up, typename _A2,
+	      typename
+	      =3D enable_if_t<is_convertible_v<simd_mask<_Up, _A2>, simd_mask>>>
+      _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
+      operator&&(const simd_mask& __x, const simd_mask<_Up, _A2>& __y)
+      {
+	return {__private_init,
+		_Impl::_S_logical_and(__x._M_data, simd_mask(__y)._M_data)};
+      }
+
+    template <typename _Up, typename _A2,
+	      typename
+	      =3D enable_if_t<is_convertible_v<simd_mask<_Up, _A2>, simd_mask>>>
+      _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
+      operator||(const simd_mask& __x, const simd_mask<_Up, _A2>& __y)
+      {
+	return {__private_init,
+		_Impl::_S_logical_or(__x._M_data, simd_mask(__y)._M_data)};
+      }
+  #endif // _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
+    operator&&(const simd_mask& __x, const simd_mask& __y)
+    {
+      return {__private_init, _Impl::_S_logical_and(__x._M_data, __y._M_da=
ta)};
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
+    operator||(const simd_mask& __x, const simd_mask& __y)
+    {
+      return {__private_init, _Impl::_S_logical_or(__x._M_data, __y._M_dat=
a)};
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
+    operator&(const simd_mask& __x, const simd_mask& __y)
+    { return {__private_init, _Impl::_S_bit_and(__x._M_data, __y._M_data)}=
; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
+    operator|(const simd_mask& __x, const simd_mask& __y)
+    { return {__private_init, _Impl::_S_bit_or(__x._M_data, __y._M_data)};=
 }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask
+    operator^(const simd_mask& __x, const simd_mask& __y)
+    { return {__private_init, _Impl::_S_bit_xor(__x._M_data, __y._M_data)}=
; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask&
+    operator&=3D(simd_mask& __x, const simd_mask& __y)
+    {
+      __x._M_data =3D _Impl::_S_bit_and(__x._M_data, __y._M_data);
+      return __x;
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask&
+    operator|=3D(simd_mask& __x, const simd_mask& __y)
+    {
+      __x._M_data =3D _Impl::_S_bit_or(__x._M_data, __y._M_data);
+      return __x;
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE friend simd_mask&
+    operator^=3D(simd_mask& __x, const simd_mask& __y)
+    {
+      __x._M_data =3D _Impl::_S_bit_xor(__x._M_data, __y._M_data);
+      return __x;
+    }
+
+    // }}}
+    // simd_mask compares [simd_mask.comparison] {{{
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
+    operator=3D=3D(const simd_mask& __x, const simd_mask& __y)
+    { return !operator!=3D(__x, __y); }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
+    operator!=3D(const simd_mask& __x, const simd_mask& __y)
+    { return {__private_init, _Impl::_S_bit_xor(__x._M_data, __y._M_data)}=
; }
+
+    // }}}
+    // private_init ctor {{{
+    _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
+    simd_mask(_PrivateInit, typename _Traits::_MaskMember __init)
+    : _M_data(__init) {}
+
+    // }}}
+    // private_init generator ctor {{{
+    template <typename _Fp, typename =3D decltype(bool(declval<_Fp>()(size=
_t())))>
+      _GLIBCXX_SIMD_INTRINSIC constexpr
+      simd_mask(_PrivateInit, _Fp&& __gen)
+      : _M_data()
+      {
+	__execute_n_times<size()>([&](auto __i) constexpr {
+	  _Impl::_S_set(_M_data, __i, __gen(__i));
+	});
+      }
+
+    // }}}
+    // bitset_init ctor {{{
+    _GLIBCXX_SIMD_INTRINSIC simd_mask(_BitsetInit, bitset<size()> __init)
+    : _M_data(
+	_Impl::_S_from_bitmask(_SanitizedBitMask<size()>(__init), _S_type_tag))
+    {}
+
+    // }}}
+    // __cvt {{{
+    // TS_FEEDBACK:
+    // The conversion operator this implements should be a ctor on simd_ma=
sk.
+    // Once you call .__cvt() on a simd_mask it converts conveniently.
+    // A useful variation: add `explicit(sizeof(_Tp) !=3D sizeof(_Up))`
+    struct _CvtProxy
+    {
+      template <typename _Up, typename _A2,
+		typename
+		=3D enable_if_t<simd_size_v<_Up, _A2> =3D=3D simd_size_v<_Tp, _Abi>>>
+	operator simd_mask<_Up, _A2>() &&
+	{
+	  using namespace std::experimental::__proposed;
+	  return static_simd_cast<simd_mask<_Up, _A2>>(_M_data);
+	}
+
+      const simd_mask<_Tp, _Abi>& _M_data;
+    };
+
+    _GLIBCXX_SIMD_INTRINSIC _CvtProxy
+    __cvt() const
+    { return {*this}; }
+
+    // }}}
+    // operator?: overloads (suggested extension) {{{
+  #ifdef __GXX_CONDITIONAL_IS_OVERLOADABLE__
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
+    operator?:(const simd_mask& __k, const simd_mask& __where_true,
+	       const simd_mask& __where_false)
+    {
+      auto __ret =3D __where_false;
+      _Impl::_S_masked_assign(__k._M_data, __ret._M_data, __where_true._M_=
data);
+      return __ret;
+    }
+
+    template <typename _U1, typename _U2,
+	      typename _Rp =3D simd<common_type_t<_U1, _U2>, _Abi>,
+	      typename =3D enable_if_t<conjunction_v<
+		is_convertible<_U1, _Rp>, is_convertible<_U2, _Rp>,
+		is_convertible<simd_mask, typename _Rp::mask_type>>>>
+      _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend _Rp
+      operator?:(const simd_mask& __k, const _U1& __where_true,
+		 const _U2& __where_false)
+      {
+	_Rp __ret =3D __where_false;
+	_Rp::_Impl::_S_masked_assign(
+	  __data(static_cast<typename _Rp::mask_type>(__k)), __data(__ret),
+	  __data(static_cast<_Rp>(__where_true)));
+	return __ret;
+      }
+
+  #ifdef _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
+    template <typename _Kp, typename _Ak, typename _Up, typename _Au,
+	      typename =3D enable_if_t<
+		conjunction_v<is_convertible<simd_mask<_Kp, _Ak>, simd_mask>,
+			      is_convertible<simd_mask<_Up, _Au>, simd_mask>>>>
+      _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd_mask
+      operator?:(const simd_mask<_Kp, _Ak>& __k, const simd_mask& __where_=
true,
+		 const simd_mask<_Up, _Au>& __where_false)
+      {
+	simd_mask __ret =3D __where_false;
+	_Impl::_S_masked_assign(simd_mask(__k)._M_data, __ret._M_data,
+				__where_true._M_data);
+	return __ret;
+      }
+  #endif // _GLIBCXX_SIMD_ENABLE_IMPLICIT_MASK_CAST
+  #endif // __GXX_CONDITIONAL_IS_OVERLOADABLE__
+
+    // }}}
+    // _M_is_constprop {{{
+    _GLIBCXX_SIMD_INTRINSIC constexpr bool
+    _M_is_constprop() const
+    {
+      if constexpr (__is_scalar_abi<_Abi>())
+	return __builtin_constant_p(_M_data);
+      else
+	return _M_data._M_is_constprop();
+    }
+
+    // }}}
+
+  private:
+    friend const auto& __data<_Tp, abi_type>(const simd_mask&);
+    friend auto& __data<_Tp, abi_type>(simd_mask&);
+    alignas(_Traits::_S_mask_align) _MemberType _M_data;
+  };
+
+// }}}
+
+// __data(simd_mask) {{{
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
+  __data(const simd_mask<_Tp, _Ap>& __x)
+  { return __x._M_data; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+  __data(simd_mask<_Tp, _Ap>& __x)
+  { return __x._M_data; }
+
+// }}}
+
+// simd_mask reductions [simd_mask.reductions] {{{
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+  all_of(const simd_mask<_Tp, _Abi>& __k) noexcept
+  {
+    if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
+      {
+	for (size_t __i =3D 0; __i < simd_size_v<_Tp, _Abi>; ++__i)
+	  if (!__k[__i])
+	    return false;
+	return true;
+      }
+    else
+      return _Abi::_MaskImpl::_S_all_of(__k);
+  }
+
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+  any_of(const simd_mask<_Tp, _Abi>& __k) noexcept
+  {
+    if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
+      {
+	for (size_t __i =3D 0; __i < simd_size_v<_Tp, _Abi>; ++__i)
+	  if (__k[__i])
+	    return true;
+	return false;
+      }
+    else
+      return _Abi::_MaskImpl::_S_any_of(__k);
+  }
+
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+  none_of(const simd_mask<_Tp, _Abi>& __k) noexcept
+  {
+    if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
+      {
+	for (size_t __i =3D 0; __i < simd_size_v<_Tp, _Abi>; ++__i)
+	  if (__k[__i])
+	    return false;
+	return true;
+      }
+    else
+      return _Abi::_MaskImpl::_S_none_of(__k);
+  }
+
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+  some_of(const simd_mask<_Tp, _Abi>& __k) noexcept
+  {
+    if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
+      {
+	for (size_t __i =3D 1; __i < simd_size_v<_Tp, _Abi>; ++__i)
+	  if (__k[__i] !=3D __k[__i - 1])
+	    return true;
+	return false;
+      }
+    else
+      return _Abi::_MaskImpl::_S_some_of(__k);
+  }
+
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
+  popcount(const simd_mask<_Tp, _Abi>& __k) noexcept
+  {
+    if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
+      {
+	const int __r =3D __call_with_subscripts<simd_size_v<_Tp, _Abi>>(
+	  __k, [](auto... __elements) { return ((__elements !=3D 0) + ...); });
+	if (__builtin_is_constant_evaluated() || __builtin_constant_p(__r))
+	  return __r;
+      }
+    return _Abi::_MaskImpl::_S_popcount(__k);
+  }
+
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
+  find_first_set(const simd_mask<_Tp, _Abi>& __k)
+  {
+    if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
+      {
+	constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	const size_t _Idx =3D __call_with_n_evaluations<_Np>(
+	  [](auto... __indexes) { return std::min({__indexes...}); },
+	  [&](auto __i) { return __k[__i] ? +__i : _Np; });
+	if (_Idx >=3D _Np)
+	  __invoke_ub("find_first_set(empty mask) is UB");
+	if (__builtin_constant_p(_Idx))
+	  return _Idx;
+      }
+    return _Abi::_MaskImpl::_S_find_first_set(__k);
+  }
+
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
+  find_last_set(const simd_mask<_Tp, _Abi>& __k)
+  {
+    if (__builtin_is_constant_evaluated() || __k._M_is_constprop())
+      {
+	constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	const int _Idx =3D __call_with_n_evaluations<_Np>(
+	  [](auto... __indexes) { return std::max({__indexes...}); },
+	  [&](auto __i) { return __k[__i] ? int(__i) : -1; });
+	if (_Idx < 0)
+	  __invoke_ub("find_first_set(empty mask) is UB");
+	if (__builtin_constant_p(_Idx))
+	  return _Idx;
+      }
+    return _Abi::_MaskImpl::_S_find_last_set(__k);
+  }
+
+_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+all_of(_ExactBool __x) noexcept
+{ return __x; }
+
+_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+any_of(_ExactBool __x) noexcept
+{ return __x; }
+
+_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+none_of(_ExactBool __x) noexcept
+{ return !__x; }
+
+_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR bool
+some_of(_ExactBool) noexcept
+{ return false; }
+
+_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
+popcount(_ExactBool __x) noexcept
+{ return __x; }
+
+_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
+find_first_set(_ExactBool)
+{ return 0; }
+
+_GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR int
+find_last_set(_ExactBool)
+{ return 0; }
+
+// }}}
+
+// _SimdIntOperators{{{1
+template <typename _V, typename _Impl, bool>
+  class _SimdIntOperators {};
+
+template <typename _V, typename _Impl>
+  class _SimdIntOperators<_V, _Impl, true>
+  {
+    _GLIBCXX_SIMD_INTRINSIC const _V& __derived() const
+    { return *static_cast<const _V*>(this); }
+
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static _GLIBCXX_SIMD_CONSTEXPR _V
+      _S_make_derived(_Tp&& __d)
+      { return {__private_init, static_cast<_Tp&&>(__d)}; }
+
+  public:
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator%=3D(_V& __lhs, const _V& _=
_x)
+    { return __lhs =3D __lhs % __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator&=3D(_V& __lhs, const _V& _=
_x)
+    { return __lhs =3D __lhs & __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator|=3D(_V& __lhs, const _V& _=
_x)
+    { return __lhs =3D __lhs | __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator^=3D(_V& __lhs, const _V& _=
_x)
+    { return __lhs =3D __lhs ^ __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator<<=3D(_V& __lhs, const _V& =
__x)
+    { return __lhs =3D __lhs << __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator>>=3D(_V& __lhs, const _V& =
__x)
+    { return __lhs =3D __lhs >> __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator<<=3D(_V& __lhs, int __x)
+    { return __lhs =3D __lhs << __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V& operator>>=3D(_V& __lhs, int __x)
+    { return __lhs =3D __lhs >> __x; }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V operator%(const _V& __x, const _V& _=
_y)
+    {
+      return _SimdIntOperators::_S_make_derived(
+	_Impl::_S_modulus(__data(__x), __data(__y)));
+    }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V operator&(const _V& __x, const _V& _=
_y)
+    {
+      return _SimdIntOperators::_S_make_derived(
+	_Impl::_S_bit_and(__data(__x), __data(__y)));
+    }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V operator|(const _V& __x, const _V& _=
_y)
+    {
+      return _SimdIntOperators::_S_make_derived(
+	_Impl::_S_bit_or(__data(__x), __data(__y)));
+    }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V operator^(const _V& __x, const _V& _=
_y)
+    {
+      return _SimdIntOperators::_S_make_derived(
+	_Impl::_S_bit_xor(__data(__x), __data(__y)));
+    }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V operator<<(const _V& __x, const _V& =
__y)
+    {
+      return _SimdIntOperators::_S_make_derived(
+	_Impl::_S_bit_shift_left(__data(__x), __data(__y)));
+    }
+
+    _GLIBCXX_SIMD_CONSTEXPR friend _V operator>>(const _V& __x, const _V& =
__y)
+    {
+      return _SimdIntOperators::_S_make_derived(
+	_Impl::_S_bit_shift_right(__data(__x), __data(__y)));
+    }
+
+    template <typename _VV =3D _V>
+      _GLIBCXX_SIMD_CONSTEXPR friend _V operator<<(const _V& __x, int __y)
+      {
+	using _Tp =3D typename _VV::value_type;
+	if (__y < 0)
+	  __invoke_ub("The behavior is undefined if the right operand of a "
+		      "shift operation is negative. [expr.shift]\nA shift by "
+		      "%d was requested",
+		      __y);
+	if (size_t(__y) >=3D sizeof(declval<_Tp>() << __y) * __CHAR_BIT__)
+	  __invoke_ub(
+	    "The behavior is undefined if the right operand of a "
+	    "shift operation is greater than or equal to the width of the "
+	    "promoted left operand. [expr.shift]\nA shift by %d was requested",
+	    __y);
+	return _SimdIntOperators::_S_make_derived(
+	  _Impl::_S_bit_shift_left(__data(__x), __y));
+      }
+
+    template <typename _VV =3D _V>
+      _GLIBCXX_SIMD_CONSTEXPR friend _V operator>>(const _V& __x, int __y)
+      {
+	using _Tp =3D typename _VV::value_type;
+	if (__y < 0)
+	  __invoke_ub(
+	    "The behavior is undefined if the right operand of a shift "
+	    "operation is negative. [expr.shift]\nA shift by %d was requested",
+	    __y);
+	if (size_t(__y) >=3D sizeof(declval<_Tp>() << __y) * __CHAR_BIT__)
+	  __invoke_ub(
+	    "The behavior is undefined if the right operand of a shift "
+	    "operation is greater than or equal to the width of the promoted "
+	    "left operand. [expr.shift]\nA shift by %d was requested",
+	    __y);
+	return _SimdIntOperators::_S_make_derived(
+	  _Impl::_S_bit_shift_right(__data(__x), __y));
+      }
+
+    // unary operators (for integral _Tp)
+    _GLIBCXX_SIMD_CONSTEXPR _V operator~() const
+    { return {__private_init, _Impl::_S_complement(__derived()._M_data)}; }
+  };
+
+//}}}1
+
+// simd {{{
+template <typename _Tp, typename _Abi>
+  class simd : public _SimdIntOperators<
+		 simd<_Tp, _Abi>, typename _SimdTraits<_Tp, _Abi>::_SimdImpl,
+		 conjunction<is_integral<_Tp>,
+			     typename _SimdTraits<_Tp, _Abi>::_IsValid>::value>,
+	       public _SimdTraits<_Tp, _Abi>::_SimdBase
+  {
+    using _Traits =3D _SimdTraits<_Tp, _Abi>;
+    using _MemberType =3D typename _Traits::_SimdMember;
+    using _CastType =3D typename _Traits::_SimdCastType;
+    static constexpr _Tp* _S_type_tag =3D nullptr;
+    friend typename _Traits::_SimdBase;
+
+  public:
+    using _Impl =3D typename _Traits::_SimdImpl;
+    friend _Impl;
+    friend _SimdIntOperators<simd, _Impl, true>;
+
+    using value_type =3D _Tp;
+    using reference =3D _SmartReference<_MemberType, _Impl, value_type>;
+    using mask_type =3D simd_mask<_Tp, _Abi>;
+    using abi_type =3D _Abi;
+
+    static constexpr size_t size()
+    { return __size_or_zero_v<_Tp, _Abi>; }
+
+    _GLIBCXX_SIMD_CONSTEXPR simd() =3D default;
+    _GLIBCXX_SIMD_CONSTEXPR simd(const simd&) =3D default;
+    _GLIBCXX_SIMD_CONSTEXPR simd(simd&&) noexcept =3D default;
+    _GLIBCXX_SIMD_CONSTEXPR simd& operator=3D(const simd&) =3D default;
+    _GLIBCXX_SIMD_CONSTEXPR simd& operator=3D(simd&&) noexcept =3D default;
+
+    // implicit broadcast constructor
+    template <typename _Up,
+	      typename =3D enable_if_t<!is_same_v<__remove_cvref_t<_Up>, bool>>>
+      _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR
+      simd(_ValuePreservingOrInt<_Up, value_type>&& __x)
+      : _M_data(
+	_Impl::_S_broadcast(static_cast<value_type>(static_cast<_Up&&>(__x))))
+      {}
+
+    // implicit type conversion constructor (convert from fixed_size to
+    // fixed_size)
+    template <typename _Up>
+      _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR
+      simd(const simd<_Up, simd_abi::fixed_size<size()>>& __x,
+	   enable_if_t<
+	     conjunction<
+	       is_same<simd_abi::fixed_size<size()>, abi_type>,
+	       negation<__is_narrowing_conversion<_Up, value_type>>,
+	       __converts_to_higher_integer_rank<_Up, value_type>>::value,
+	     void*> =3D nullptr)
+      : simd{static_cast<array<_Up, size()>>(__x).data(), vector_aligned} =
{}
+
+      // explicit type conversion constructor
+#ifdef _GLIBCXX_SIMD_ENABLE_STATIC_CAST
+    template <typename _Up, typename _A2,
+	      typename =3D decltype(static_simd_cast<simd>(
+		declval<const simd<_Up, _A2>&>()))>
+      _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
+      simd(const simd<_Up, _A2>& __x)
+      : simd(static_simd_cast<simd>(__x)) {}
+#endif // _GLIBCXX_SIMD_ENABLE_STATIC_CAST
+
+    // generator constructor
+    template <typename _Fp>
+      _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
+      simd(_Fp&& __gen, _ValuePreservingOrInt<decltype(declval<_Fp>()(
+						declval<_SizeConstant<0>&>())),
+					      value_type>* =3D nullptr)
+      : _M_data(_Impl::_S_generator(static_cast<_Fp&&>(__gen), _S_type_tag=
)) {}
+
+    // load constructor
+    template <typename _Up, typename _Flags>
+      _GLIBCXX_SIMD_ALWAYS_INLINE
+      simd(const _Up* __mem, _Flags)
+      : _M_data(
+	  _Impl::_S_load(_Flags::template _S_apply<simd>(__mem), _S_type_tag))
+      {}
+
+    // loads [simd.load]
+    template <typename _Up, typename _Flags>
+      _GLIBCXX_SIMD_ALWAYS_INLINE void
+      copy_from(const _Vectorizable<_Up>* __mem, _Flags)
+      {
+	_M_data =3D static_cast<decltype(_M_data)>(
+	  _Impl::_S_load(_Flags::template _S_apply<simd>(__mem), _S_type_tag));
+      }
+
+    // stores [simd.store]
+    template <typename _Up, typename _Flags>
+      _GLIBCXX_SIMD_ALWAYS_INLINE void
+      copy_to(_Vectorizable<_Up>* __mem, _Flags) const
+      {
+	_Impl::_S_store(_M_data, _Flags::template _S_apply<simd>(__mem),
+			_S_type_tag);
+      }
+
+    // scalar access
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR reference
+    operator[](size_t __i)
+    { return {_M_data, int(__i)}; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR value_type
+    operator[]([[maybe_unused]] size_t __i) const
+    {
+      if constexpr (__is_scalar_abi<_Abi>())
+	{
+	  _GLIBCXX_DEBUG_ASSERT(__i =3D=3D 0);
+	  return _M_data;
+	}
+      else
+	return _M_data[__i];
+    }
+
+    // increment and decrement:
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd&
+    operator++()
+    {
+      _Impl::_S_increment(_M_data);
+      return *this;
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
+    operator++(int)
+    {
+      simd __r =3D *this;
+      _Impl::_S_increment(_M_data);
+      return __r;
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd&
+    operator--()
+    {
+      _Impl::_S_decrement(_M_data);
+      return *this;
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
+    operator--(int)
+    {
+      simd __r =3D *this;
+      _Impl::_S_decrement(_M_data);
+      return __r;
+    }
+
+    // unary operators (for any _Tp)
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR mask_type
+    operator!() const
+    { return {__private_init, _Impl::_S_negate(_M_data)}; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
+    operator+() const
+    { return *this; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR simd
+    operator-() const
+    { return {__private_init, _Impl::_S_unary_minus(_M_data)}; }
+
+    // access to internal representation (suggested extension)
+    _GLIBCXX_SIMD_ALWAYS_INLINE explicit _GLIBCXX_SIMD_CONSTEXPR
+    simd(_CastType __init) : _M_data(__init) {}
+
+    // compound assignment [simd.cassign]
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
+    operator+=3D(simd& __lhs, const simd& __x)
+    { return __lhs =3D __lhs + __x; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
+    operator-=3D(simd& __lhs, const simd& __x)
+    { return __lhs =3D __lhs - __x; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
+    operator*=3D(simd& __lhs, const simd& __x)
+    { return __lhs =3D __lhs * __x; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd&
+    operator/=3D(simd& __lhs, const simd& __x)
+    { return __lhs =3D __lhs / __x; }
+
+    // binary operators [simd.binary]
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
+    operator+(const simd& __x, const simd& __y)
+    { return {__private_init, _Impl::_S_plus(__x._M_data, __y._M_data)}; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
+    operator-(const simd& __x, const simd& __y)
+    { return {__private_init, _Impl::_S_minus(__x._M_data, __y._M_data)}; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
+    operator*(const simd& __x, const simd& __y)
+    { return {__private_init, _Impl::_S_multiplies(__x._M_data, __y._M_dat=
a)}; }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
+    operator/(const simd& __x, const simd& __y)
+    { return {__private_init, _Impl::_S_divides(__x._M_data, __y._M_data)}=
; }
+
+    // compares [simd.comparison]
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
+    operator=3D=3D(const simd& __x, const simd& __y)
+    { return simd::_S_make_mask(_Impl::_S_equal_to(__x._M_data, __y._M_dat=
a)); }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
+    operator!=3D(const simd& __x, const simd& __y)
+    {
+      return simd::_S_make_mask(
+	_Impl::_S_not_equal_to(__x._M_data, __y._M_data));
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
+    operator<(const simd& __x, const simd& __y)
+    { return simd::_S_make_mask(_Impl::_S_less(__x._M_data, __y._M_data));=
 }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
+    operator<=3D(const simd& __x, const simd& __y)
+    {
+      return simd::_S_make_mask(_Impl::_S_less_equal(__x._M_data, __y._M_d=
ata));
+    }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
+    operator>(const simd& __x, const simd& __y)
+    { return simd::_S_make_mask(_Impl::_S_less(__y._M_data, __x._M_data));=
 }
+
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend mask_type
+    operator>=3D(const simd& __x, const simd& __y)
+    {
+      return simd::_S_make_mask(_Impl::_S_less_equal(__y._M_data, __x._M_d=
ata));
+    }
+
+    // operator?: overloads (suggested extension) {{{
+#ifdef __GXX_CONDITIONAL_IS_OVERLOADABLE__
+    _GLIBCXX_SIMD_ALWAYS_INLINE _GLIBCXX_SIMD_CONSTEXPR friend simd
+    operator?:(const mask_type& __k, const simd& __where_true,
+	const simd& __where_false)
+    {
+      auto __ret =3D __where_false;
+      _Impl::_S_masked_assign(__data(__k), __data(__ret), __data(__where_t=
rue));
+      return __ret;
+    }
+
+#endif // __GXX_CONDITIONAL_IS_OVERLOADABLE__
+    // }}}
+
+    // "private" because of the first arguments's namespace
+    _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR
+    simd(_PrivateInit, const _MemberType& __init)
+    : _M_data(__init) {}
+
+    // "private" because of the first arguments's namespace
+    _GLIBCXX_SIMD_INTRINSIC
+    simd(_BitsetInit, bitset<size()> __init) : _M_data()
+    { where(mask_type(__bitset_init, __init), *this) =3D ~*this; }
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr bool
+    _M_is_constprop() const
+    {
+      if constexpr (__is_scalar_abi<_Abi>())
+	return __builtin_constant_p(_M_data);
+      else
+	return _M_data._M_is_constprop();
+    }
+
+  private:
+    _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR static mask_type
+    _S_make_mask(typename mask_type::_MemberType __k)
+    { return {__private_init, __k}; }
+
+    friend const auto& __data<value_type, abi_type>(const simd&);
+    friend auto& __data<value_type, abi_type>(simd&);
+    alignas(_Traits::_S_simd_align) _MemberType _M_data;
+  };
+
+// }}}
+// __data {{{
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
+  __data(const simd<_Tp, _Ap>& __x)
+  { return __x._M_data; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+  __data(simd<_Tp, _Ap>& __x)
+  { return __x._M_data; }
+
+// }}}
+namespace __float_bitwise_operators { //{{{
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
+  operator^(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
+  {
+    return {__private_init,
+	    _Ap::_SimdImpl::_S_bit_xor(__data(__a), __data(__b))};
+  }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
+  operator|(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
+  {
+    return {__private_init,
+	    _Ap::_SimdImpl::_S_bit_or(__data(__a), __data(__b))};
+  }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_SIMD_CONSTEXPR simd<_Tp, _Ap>
+  operator&(const simd<_Tp, _Ap>& __a, const simd<_Tp, _Ap>& __b)
+  {
+    return {__private_init,
+	    _Ap::_SimdImpl::_S_bit_and(__data(__a), __data(__b))};
+  }
+} // namespace __float_bitwise_operators }}}
+
+_GLIBCXX_SIMD_END_NAMESPACE
+
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_H
+
+// vim: foldmethod=3Dmarker foldmarker=3D{{{,}}}
diff --git a/libstdc++-v3/include/experimental/bits/simd_builtin.h b/libstd=
c++-v3/include/experimental/bits/simd_builtin.h
new file mode 100644
index 00000000000..f2c99faa4ee
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_builtin.h
@@ -0,0 +1,2949 @@
+// Simd Abi specific implementations -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_ABIS_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_ABIS_H_
+
+#if __cplusplus >=3D 201703L
+
+#include <array>
+#include <cmath>
+#include <cstdlib>
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+// _S_allbits{{{
+template <typename _V>
+  static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_allbits
+    =3D reinterpret_cast<_V>(~__vector_type_t<char, sizeof(_V) / sizeof(ch=
ar)>());
+
+// }}}
+// _S_signmask, _S_absmask{{{
+template <typename _V, typename =3D _VectorTraits<_V>>
+  static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_signmask
+    =3D __xor(_V() + 1, _V() - 1);
+
+template <typename _V, typename =3D _VectorTraits<_V>>
+  static inline _GLIBCXX_SIMD_USE_CONSTEXPR _V _S_absmask
+    =3D __andnot(_S_signmask<_V>, _S_allbits<_V>);
+
+//}}}
+// __vector_permute<Indices...>{{{
+// Index =3D=3D -1 requests zeroing of the output element
+template <int... _Indices, typename _Tp, typename _TVT =3D _VectorTraits<_=
Tp>>
+  _Tp
+  __vector_permute(_Tp __x)
+  {
+    static_assert(sizeof...(_Indices) =3D=3D _TVT::_S_full_size);
+    return __make_vector<typename _TVT::value_type>(
+      (_Indices =3D=3D -1 ? 0 : __x[_Indices =3D=3D -1 ? 0 : _Indices])...=
);
+  }
+
+// }}}
+// __vector_shuffle<Indices...>{{{
+// Index =3D=3D -1 requests zeroing of the output element
+template <int... _Indices, typename _Tp, typename _TVT =3D _VectorTraits<_=
Tp>>
+  _Tp
+  __vector_shuffle(_Tp __x, _Tp __y)
+  {
+    return _Tp{(_Indices =3D=3D -1 ? 0
+		: _Indices < _TVT::_S_full_size
+		  ? __x[_Indices]
+		  : __y[_Indices - _TVT::_S_full_size])...};
+  }
+
+// }}}
+// __make_wrapper{{{
+template <typename _Tp, typename... _Args>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<_Tp, sizeof...(_Args)>
+  __make_wrapper(const _Args&... __args)
+  { return __make_vector<_Tp>(__args...); }
+
+// }}}
+// __wrapper_bitcast{{{
+template <typename _Tp, size_t _ToN =3D 0, typename _Up, size_t _M,
+	  size_t _Np =3D _ToN !=3D 0 ? _ToN : sizeof(_Up) * _M / sizeof(_Tp)>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<_Tp, _Np>
+  __wrapper_bitcast(_SimdWrapper<_Up, _M> __x)
+  {
+    static_assert(_Np > 1);
+    return __intrin_bitcast<__vector_type_t<_Tp, _Np>>(__x._M_data);
+  }
+
+// }}}
+// __shift_elements_right{{{
+// if (__shift % 2=E2=81=BF =3D=3D 0) =3D> the low n Bytes are correct
+template <unsigned __shift, typename _Tp, typename _TVT =3D _VectorTraits<=
_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC _Tp
+  __shift_elements_right(_Tp __v)
+  {
+    [[maybe_unused]] const auto __iv =3D __to_intrin(__v);
+    static_assert(__shift <=3D sizeof(_Tp));
+    if constexpr (__shift =3D=3D 0)
+      return __v;
+    else if constexpr (__shift =3D=3D sizeof(_Tp))
+      return _Tp();
+#if _GLIBCXX_SIMD_X86INTRIN // {{{
+    else if constexpr (__have_sse && __shift =3D=3D 8
+		       && _TVT::template _S_is<float, 4>)
+      return _mm_movehl_ps(__iv, __iv);
+    else if constexpr (__have_sse2 && __shift =3D=3D 8
+		       && _TVT::template _S_is<double, 2>)
+      return _mm_unpackhi_pd(__iv, __iv);
+    else if constexpr (__have_sse2 && sizeof(_Tp) =3D=3D 16)
+      return reinterpret_cast<typename _TVT::type>(
+	_mm_srli_si128(reinterpret_cast<__m128i>(__iv), __shift));
+    else if constexpr (__shift =3D=3D 16 && sizeof(_Tp) =3D=3D 32)
+      {
+	/*if constexpr (__have_avx && _TVT::template _S_is<double, 4>)
+	  return _mm256_permute2f128_pd(__iv, __iv, 0x81);
+	else if constexpr (__have_avx && _TVT::template _S_is<float, 8>)
+	  return _mm256_permute2f128_ps(__iv, __iv, 0x81);
+	else if constexpr (__have_avx)
+	  return reinterpret_cast<typename _TVT::type>(
+	    _mm256_permute2f128_si256(__iv, __iv, 0x81));
+	else*/
+	return __zero_extend(__hi128(__v));
+      }
+    else if constexpr (__have_avx2 && sizeof(_Tp) =3D=3D 32 && __shift < 1=
6)
+      {
+	const auto __vll =3D __vector_bitcast<_LLong>(__v);
+	return reinterpret_cast<typename _TVT::type>(
+	  _mm256_alignr_epi8(_mm256_permute2x128_si256(__vll, __vll, 0x81),
+			     __vll, __shift));
+      }
+    else if constexpr (__have_avx && sizeof(_Tp) =3D=3D 32 && __shift < 16)
+      {
+	const auto __vll =3D __vector_bitcast<_LLong>(__v);
+	return reinterpret_cast<typename _TVT::type>(
+	  __concat(_mm_alignr_epi8(__hi128(__vll), __lo128(__vll), __shift),
+		   _mm_srli_si128(__hi128(__vll), __shift)));
+      }
+    else if constexpr (sizeof(_Tp) =3D=3D 32 && __shift > 16)
+      return __zero_extend(__shift_elements_right<__shift - 16>(__hi128(__=
v)));
+    else if constexpr (sizeof(_Tp) =3D=3D 64 && __shift =3D=3D 32)
+      return __zero_extend(__hi256(__v));
+    else if constexpr (__have_avx512f && sizeof(_Tp) =3D=3D 64)
+      {
+	if constexpr (__shift >=3D 48)
+	  return __zero_extend(
+	    __shift_elements_right<__shift - 48>(__extract<3, 4>(__v)));
+	else if constexpr (__shift >=3D 32)
+	  return __zero_extend(
+	    __shift_elements_right<__shift - 32>(__hi256(__v)));
+	else if constexpr (__shift % 8 =3D=3D 0)
+	  return reinterpret_cast<typename _TVT::type>(
+	    _mm512_alignr_epi64(__m512i(), __intrin_bitcast<__m512i>(__v),
+				__shift / 8));
+	else if constexpr (__shift % 4 =3D=3D 0)
+	  return reinterpret_cast<typename _TVT::type>(
+	    _mm512_alignr_epi32(__m512i(), __intrin_bitcast<__m512i>(__v),
+				__shift / 4));
+	else if constexpr (__have_avx512bw && __shift < 16)
+	  {
+	    const auto __vll =3D __vector_bitcast<_LLong>(__v);
+	    return reinterpret_cast<typename _TVT::type>(
+	      _mm512_alignr_epi8(_mm512_shuffle_i32x4(__vll, __vll, 0xf9),
+				 __vll, __shift));
+	  }
+	else if constexpr (__have_avx512bw && __shift < 32)
+	  {
+	    const auto __vll =3D __vector_bitcast<_LLong>(__v);
+	    return reinterpret_cast<typename _TVT::type>(
+	      _mm512_alignr_epi8(_mm512_shuffle_i32x4(__vll, __m512i(), 0xee),
+				 _mm512_shuffle_i32x4(__vll, __vll, 0xf9),
+				 __shift - 16));
+	  }
+	else
+	  __assert_unreachable<_Tp>();
+      }
+  /*
+      } else if constexpr (__shift % 16 =3D=3D 0 && sizeof(_Tp) =3D=3D 64)
+	  return __auto_bitcast(__extract<__shift / 16, 4>(__v));
+  */
+#endif // _GLIBCXX_SIMD_X86INTRIN }}}
+    else
+      {
+	constexpr int __chunksize =3D __shift % 8 =3D=3D 0   ? 8
+				    : __shift % 4 =3D=3D 0 ? 4
+				    : __shift % 2 =3D=3D 0 ? 2
+						       : 1;
+	auto __w =3D __vector_bitcast<__int_with_sizeof_t<__chunksize>>(__v);
+	using _Up =3D decltype(__w);
+	return __intrin_bitcast<_Tp>(
+	  __call_with_n_evaluations<(sizeof(_Tp) - __shift) / __chunksize>(
+	    [](auto... __chunks) { return _Up{__chunks...}; },
+	    [&](auto __i) { return __w[__shift / __chunksize + __i]; }));
+      }
+  }
+
+// }}}
+// __extract_part(_SimdWrapper<_Tp, _Np>) {{{
+template <int _Index, int _Total, int _Combine, typename _Tp, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST
+  _SimdWrapper<_Tp, _Np / _Total * _Combine>
+  __extract_part(const _SimdWrapper<_Tp, _Np> __x)
+  {
+    if constexpr (_Index % 2 =3D=3D 0 && _Total % 2 =3D=3D 0 && _Combine %=
 2 =3D=3D 0)
+      return __extract_part<_Index / 2, _Total / 2, _Combine / 2>(__x);
+    else
+      {
+	constexpr size_t __values_per_part =3D _Np / _Total;
+	constexpr size_t __values_to_skip =3D _Index * __values_per_part;
+	constexpr size_t __return_size =3D __values_per_part * _Combine;
+	using _R =3D __vector_type_t<_Tp, __return_size>;
+	static_assert((_Index + _Combine) * __values_per_part * sizeof(_Tp)
+			<=3D sizeof(__x),
+		      "out of bounds __extract_part");
+	// the following assertion would ensure no "padding" to be read
+	// static_assert(_Total >=3D _Index + _Combine, "_Total must be greater
+	// than _Index");
+
+	// static_assert(__return_size * _Total =3D=3D _Np, "_Np must be divisible
+	// by _Total");
+	if (__x._M_is_constprop())
+	  return __generate_from_n_evaluations<__return_size, _R>(
+	    [&](auto __i) { return __x[__values_to_skip + __i]; });
+	if constexpr (_Index =3D=3D 0 && _Total =3D=3D 1)
+	  return __x;
+	else if constexpr (_Index =3D=3D 0)
+	  return __intrin_bitcast<_R>(__as_vector(__x));
+#if _GLIBCXX_SIMD_X86INTRIN // {{{
+	else if constexpr (sizeof(__x) =3D=3D 32
+			   && __return_size * sizeof(_Tp) <=3D 16)
+	  {
+	    constexpr size_t __bytes_to_skip =3D __values_to_skip * sizeof(_Tp);
+	    if constexpr (__bytes_to_skip =3D=3D 16)
+	      return __vector_bitcast<_Tp, __return_size>(
+		__hi128(__as_vector(__x)));
+	    else
+	      return __vector_bitcast<_Tp, __return_size>(
+		_mm_alignr_epi8(__hi128(__vector_bitcast<_LLong>(__x)),
+				__lo128(__vector_bitcast<_LLong>(__x)),
+				__bytes_to_skip));
+	  }
+#endif // _GLIBCXX_SIMD_X86INTRIN }}}
+	else if constexpr (_Index > 0
+			   && (__values_to_skip % __return_size !=3D 0
+			       || sizeof(_R) >=3D 8)
+			   && (__values_to_skip + __return_size) * sizeof(_Tp)
+				<=3D 64
+			   && sizeof(__x) >=3D 16)
+	  return __intrin_bitcast<_R>(
+	    __shift_elements_right<__values_to_skip * sizeof(_Tp)>(
+	      __as_vector(__x)));
+	else
+	  {
+	    _R __r =3D {};
+	    __builtin_memcpy(&__r,
+			     reinterpret_cast<const char*>(&__x)
+			       + sizeof(_Tp) * __values_to_skip,
+			     __return_size * sizeof(_Tp));
+	    return __r;
+	  }
+      }
+  }
+
+// }}}
+// __extract_part(_SimdWrapper<bool, _Np>) {{{
+template <int _Index, int _Total, int _Combine =3D 1, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<bool, _Np / _Total * _Com=
bine>
+  __extract_part(const _SimdWrapper<bool, _Np> __x)
+  {
+    static_assert(_Combine =3D=3D 1, "_Combine !=3D 1 not implemented");
+    static_assert(__have_avx512f && _Np =3D=3D _Np);
+    static_assert(_Total >=3D 2 && _Index + _Combine <=3D _Total && _Index=
 >=3D 0);
+    return __x._M_data >> (_Index * _Np / _Total);
+  }
+
+// }}}
+
+// __vector_convert {{{
+// implementation requires an index sequence
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d,
+		   index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i,
+		   index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i, _From __j,
+		   index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i, _From __j,
+		   _From __k, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
+	       static_cast<_Tp>(__k[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i, _From __j,
+		   _From __k, _From __l, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
+	       static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i, _From __j,
+		   _From __k, _From __l, _From __m, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
+	       static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
+	       static_cast<_Tp>(__m[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i, _From __j,
+		   _From __k, _From __l, _From __m, _From __n,
+		   index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
+	       static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
+	       static_cast<_Tp>(__m[_I])..., static_cast<_Tp>(__n[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i, _From __j,
+		   _From __k, _From __l, _From __m, _From __n, _From __o,
+		   index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
+	       static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
+	       static_cast<_Tp>(__m[_I])..., static_cast<_Tp>(__n[_I])...,
+	       static_cast<_Tp>(__o[_I])...};
+  }
+
+template <typename _To, typename _From, size_t... _I>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_From __a, _From __b, _From __c, _From __d, _From __e,
+		   _From __f, _From __g, _From __h, _From __i, _From __j,
+		   _From __k, _From __l, _From __m, _From __n, _From __o,
+		   _From __p, index_sequence<_I...>)
+  {
+    using _Tp =3D typename _VectorTraits<_To>::value_type;
+    return _To{static_cast<_Tp>(__a[_I])..., static_cast<_Tp>(__b[_I])...,
+	       static_cast<_Tp>(__c[_I])..., static_cast<_Tp>(__d[_I])...,
+	       static_cast<_Tp>(__e[_I])..., static_cast<_Tp>(__f[_I])...,
+	       static_cast<_Tp>(__g[_I])..., static_cast<_Tp>(__h[_I])...,
+	       static_cast<_Tp>(__i[_I])..., static_cast<_Tp>(__j[_I])...,
+	       static_cast<_Tp>(__k[_I])..., static_cast<_Tp>(__l[_I])...,
+	       static_cast<_Tp>(__m[_I])..., static_cast<_Tp>(__n[_I])...,
+	       static_cast<_Tp>(__o[_I])..., static_cast<_Tp>(__p[_I])...};
+  }
+
+// Defer actual conversion to the overload that takes an index sequence. N=
ote
+// that this function adds zeros or drops values off the end if you don't =
ensure
+// matching width.
+template <typename _To, typename... _From, size_t _FromSize>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _To
+  __vector_convert(_SimdWrapper<_From, _FromSize>... __xs)
+  {
+#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
+    using _From0 =3D __first_of_pack_t<_From...>;
+    using _FW =3D _SimdWrapper<_From0, _FromSize>;
+    if (!_FW::_S_is_partial && !(... && __xs._M_is_constprop()))
+      {
+	if constexpr ((sizeof...(_From) & (sizeof...(_From) - 1))
+		      =3D=3D 0) // power-of-two number of arguments
+	  return __convert_x86<_To>(__as_vector(__xs)...);
+	else // append zeros and recurse until the above branch is taken
+	  return __vector_convert<_To>(__xs..., _FW{});
+      }
+    else
+#endif
+      return __vector_convert<_To>(
+	__as_vector(__xs)...,
+	make_index_sequence<(sizeof...(__xs) =3D=3D 1 ? std::min(
+			       _VectorTraits<_To>::_S_full_size, int(_FromSize))
+						  : _FromSize)>());
+  }
+
+// }}}
+// __convert function{{{
+template <typename _To, typename _From, typename... _More>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __convert(_From __v0, _More... __vs)
+  {
+    static_assert((true && ... && is_same_v<_From, _More>) );
+    if constexpr (__is_vectorizable_v<_From>)
+      {
+	using _V =3D typename _VectorTraits<_To>::type;
+	using _Tp =3D typename _VectorTraits<_To>::value_type;
+	return _V{static_cast<_Tp>(__v0), static_cast<_Tp>(__vs)...};
+      }
+    else if constexpr (__is_vector_type_v<_From>)
+      return __convert<_To>(__as_wrapper(__v0), __as_wrapper(__vs)...);
+    else // _SimdWrapper arguments
+      {
+	constexpr size_t __input_size =3D _From::_S_size * (1 + sizeof...(_More));
+	if constexpr (__is_vectorizable_v<_To>)
+	  return __convert<__vector_type_t<_To, __input_size>>(__v0, __vs...);
+	else if constexpr (!__is_vector_type_v<_To>)
+	  return _To(__convert<typename _To::_BuiltinType>(__v0, __vs...));
+	else
+	  {
+	    static_assert(
+	      sizeof...(_More) =3D=3D 0
+		|| _VectorTraits<_To>::_S_full_size >=3D __input_size,
+	      "__convert(...) requires the input to fit into the output");
+	    return __vector_convert<_To>(__v0, __vs...);
+	  }
+      }
+  }
+
+// }}}
+// __convert_all{{{
+// Converts __v into array<_To, N>, where N is _NParts if non-zero or
+// otherwise deduced from _To such that N * #elements(_To) <=3D #elements(=
__v).
+// Note: this function may return less than all converted elements
+template <typename _To,
+	  size_t _NParts =3D 0, // allows to convert fewer or more (only last
+			      // _To, to be partially filled) than all
+	  size_t _Offset =3D 0, // where to start, # of elements (not Bytes or
+			      // Parts)
+	  typename _From, typename _FromVT =3D _VectorTraits<_From>>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __convert_all(_From __v)
+  {
+    if constexpr (is_arithmetic_v<_To> && _NParts !=3D 1)
+      {
+	static_assert(_Offset < _FromVT::_S_full_size);
+	constexpr auto _Np
+	  =3D _NParts =3D=3D 0 ? _FromVT::_S_partial_width - _Offset : _NParts;
+	return __generate_from_n_evaluations<_Np, array<_To, _Np>>(
+	  [&](auto __i) { return static_cast<_To>(__v[__i + _Offset]); });
+      }
+    else
+      {
+	static_assert(__is_vector_type_v<_To>);
+	using _ToVT =3D _VectorTraits<_To>;
+	if constexpr (__is_vector_type_v<_From>)
+	  return __convert_all<_To, _NParts>(__as_wrapper(__v));
+	else if constexpr (_NParts =3D=3D 1)
+	  {
+	    static_assert(_Offset % _ToVT::_S_full_size =3D=3D 0);
+	    return array<_To, 1>{__vector_convert<_To>(
+	      __extract_part<_Offset / _ToVT::_S_full_size,
+			     __div_roundup(_FromVT::_S_partial_width,
+					   _ToVT::_S_full_size)>(__v))};
+	  }
+#if _GLIBCXX_SIMD_X86INTRIN // {{{
+	else if constexpr (!__have_sse4_1 && _Offset =3D=3D 0
+	  && is_integral_v<typename _FromVT::value_type>
+	  && sizeof(typename _FromVT::value_type)
+	      < sizeof(typename _ToVT::value_type)
+	  && !(sizeof(typename _FromVT::value_type) =3D=3D 4
+	      && is_same_v<typename _ToVT::value_type, double>))
+	  {
+	    using _ToT =3D typename _ToVT::value_type;
+	    using _FromT =3D typename _FromVT::value_type;
+	    constexpr size_t _Np
+	      =3D _NParts !=3D 0
+		  ? _NParts
+		  : (_FromVT::_S_partial_width / _ToVT::_S_full_size);
+	    using _R =3D array<_To, _Np>;
+	    // __adjust modifies its input to have _Np (use _SizeConstant)
+	    // entries so that no unnecessary intermediate conversions are
+	    // requested and, more importantly, no intermediate conversions are
+	    // missing
+	    [[maybe_unused]] auto __adjust
+	      =3D [](auto __n,
+		   auto __vv) -> _SimdWrapper<_FromT, decltype(__n)::value> {
+	      return __vector_bitcast<_FromT, decltype(__n)::value>(__vv);
+	    };
+	    [[maybe_unused]] const auto __vi =3D __to_intrin(__v);
+	    auto&& __make_array =3D [](auto __x0, [[maybe_unused]] auto __x1) {
+	      if constexpr (_Np =3D=3D 1)
+		return _R{__intrin_bitcast<_To>(__x0)};
+	      else
+		return _R{__intrin_bitcast<_To>(__x0),
+			  __intrin_bitcast<_To>(__x1)};
+	    };
+
+	    if constexpr (_Np =3D=3D 0)
+	      return _R{};
+	    else if constexpr (sizeof(_FromT) =3D=3D 1 && sizeof(_ToT) =3D=3D 2)
+	      {
+		static_assert(is_integral_v<_FromT>);
+		static_assert(is_integral_v<_ToT>);
+		if constexpr (is_unsigned_v<_FromT>)
+		  return __make_array(_mm_unpacklo_epi8(__vi, __m128i()),
+				      _mm_unpackhi_epi8(__vi, __m128i()));
+		else
+		  return __make_array(
+		    _mm_srai_epi16(_mm_unpacklo_epi8(__vi, __vi), 8),
+		    _mm_srai_epi16(_mm_unpackhi_epi8(__vi, __vi), 8));
+	      }
+	    else if constexpr (sizeof(_FromT) =3D=3D 2 && sizeof(_ToT) =3D=3D 4)
+	      {
+		static_assert(is_integral_v<_FromT>);
+		if constexpr (is_floating_point_v<_ToT>)
+		  {
+		    const auto __ints
+		      =3D __convert_all<__vector_type16_t<int>, _Np>(
+			__adjust(_SizeConstant<_Np * 4>(), __v));
+		    return __generate_from_n_evaluations<_Np, _R>(
+		      [&](auto __i) {
+			return __vector_convert<_To>(__as_wrapper(__ints[__i]));
+		      });
+		  }
+		else if constexpr (is_unsigned_v<_FromT>)
+		  return __make_array(_mm_unpacklo_epi16(__vi, __m128i()),
+				      _mm_unpackhi_epi16(__vi, __m128i()));
+		else
+		  return __make_array(
+		    _mm_srai_epi32(_mm_unpacklo_epi16(__vi, __vi), 16),
+		    _mm_srai_epi32(_mm_unpackhi_epi16(__vi, __vi), 16));
+	      }
+	    else if constexpr (sizeof(_FromT) =3D=3D 4 && sizeof(_ToT) =3D=3D 8
+			       && is_integral_v<_FromT> && is_integral_v<_ToT>)
+	      {
+		if constexpr (is_unsigned_v<_FromT>)
+		  return __make_array(_mm_unpacklo_epi32(__vi, __m128i()),
+				      _mm_unpackhi_epi32(__vi, __m128i()));
+		else
+		  return __make_array(
+		    _mm_unpacklo_epi32(__vi, _mm_srai_epi32(__vi, 31)),
+		    _mm_unpackhi_epi32(__vi, _mm_srai_epi32(__vi, 31)));
+	      }
+	    else if constexpr (sizeof(_FromT) =3D=3D 4 && sizeof(_ToT) =3D=3D 8
+			       && is_integral_v<_FromT> && is_integral_v<_ToT>)
+	      {
+		if constexpr (is_unsigned_v<_FromT>)
+		  return __make_array(_mm_unpacklo_epi32(__vi, __m128i()),
+				      _mm_unpackhi_epi32(__vi, __m128i()));
+		else
+		  return __make_array(
+		    _mm_unpacklo_epi32(__vi, _mm_srai_epi32(__vi, 31)),
+		    _mm_unpackhi_epi32(__vi, _mm_srai_epi32(__vi, 31)));
+	      }
+	    else if constexpr (sizeof(_FromT) =3D=3D 1 && sizeof(_ToT) >=3D 4
+			       && is_signed_v<_FromT>)
+	      {
+		const __m128i __vv[2] =3D {_mm_unpacklo_epi8(__vi, __vi),
+					 _mm_unpackhi_epi8(__vi, __vi)};
+		const __vector_type_t<int, 4> __vvvv[4] =3D {
+		  __vector_bitcast<int>(_mm_unpacklo_epi16(__vv[0], __vv[0])),
+		  __vector_bitcast<int>(_mm_unpackhi_epi16(__vv[0], __vv[0])),
+		  __vector_bitcast<int>(_mm_unpacklo_epi16(__vv[1], __vv[1])),
+		  __vector_bitcast<int>(_mm_unpackhi_epi16(__vv[1], __vv[1]))};
+		if constexpr (sizeof(_ToT) =3D=3D 4)
+		  return __generate_from_n_evaluations<_Np, _R>([&](auto __i) {
+		    return __vector_convert<_To>(
+		      _SimdWrapper<int, 4>(__vvvv[__i] >> 24));
+		  });
+		else if constexpr (is_integral_v<_ToT>)
+		  return __generate_from_n_evaluations<_Np, _R>([&](auto __i) {
+		    const auto __signbits =3D __to_intrin(__vvvv[__i / 2] >> 31);
+		    const auto __sx32 =3D __to_intrin(__vvvv[__i / 2] >> 24);
+		    return __vector_bitcast<_ToT>(
+		      __i % 2 =3D=3D 0 ? _mm_unpacklo_epi32(__sx32, __signbits)
+				   : _mm_unpackhi_epi32(__sx32, __signbits));
+		  });
+		else
+		  return __generate_from_n_evaluations<_Np, _R>([&](auto __i) {
+		    const _SimdWrapper<int, 4> __int4 =3D __vvvv[__i / 2] >> 24;
+		    return __vector_convert<_To>(
+		      __i % 2 =3D=3D 0 ? __int4
+				   : _SimdWrapper<int, 4>(
+				     _mm_unpackhi_epi64(__to_intrin(__int4),
+							__to_intrin(__int4))));
+		  });
+	      }
+	    else if constexpr (sizeof(_FromT) =3D=3D 1 && sizeof(_ToT) =3D=3D 4)
+	      {
+		const auto __shorts =3D __convert_all<__vector_type16_t<
+		  conditional_t<is_signed_v<_FromT>, short, unsigned short>>>(
+		  __adjust(_SizeConstant<(_Np + 1) / 2 * 8>(), __v));
+		return __generate_from_n_evaluations<_Np, _R>([&](auto __i) {
+		  return __convert_all<_To>(__shorts[__i / 2])[__i % 2];
+		});
+	      }
+	    else if constexpr (sizeof(_FromT) =3D=3D 2 && sizeof(_ToT) =3D=3D 8
+			       && is_signed_v<_FromT> && is_integral_v<_ToT>)
+	      {
+		const __m128i __vv[2] =3D {_mm_unpacklo_epi16(__vi, __vi),
+					 _mm_unpackhi_epi16(__vi, __vi)};
+		const __vector_type16_t<int> __vvvv[4]
+		  =3D {__vector_bitcast<int>(
+		       _mm_unpacklo_epi32(_mm_srai_epi32(__vv[0], 16),
+					  _mm_srai_epi32(__vv[0], 31))),
+		     __vector_bitcast<int>(
+		       _mm_unpackhi_epi32(_mm_srai_epi32(__vv[0], 16),
+					  _mm_srai_epi32(__vv[0], 31))),
+		     __vector_bitcast<int>(
+		       _mm_unpacklo_epi32(_mm_srai_epi32(__vv[1], 16),
+					  _mm_srai_epi32(__vv[1], 31))),
+		     __vector_bitcast<int>(
+		       _mm_unpackhi_epi32(_mm_srai_epi32(__vv[1], 16),
+					  _mm_srai_epi32(__vv[1], 31)))};
+		return __generate_from_n_evaluations<_Np, _R>([&](auto __i) {
+		  return __vector_bitcast<_ToT>(__vvvv[__i]);
+		});
+	      }
+	    else if constexpr (sizeof(_FromT) <=3D 2 && sizeof(_ToT) =3D=3D 8)
+	      {
+		const auto __ints
+		  =3D __convert_all<__vector_type16_t<conditional_t<
+		    is_signed_v<_FromT> || is_floating_point_v<_ToT>, int,
+		    unsigned int>>>(
+		    __adjust(_SizeConstant<(_Np + 1) / 2 * 4>(), __v));
+		return __generate_from_n_evaluations<_Np, _R>([&](auto __i) {
+		  return __convert_all<_To>(__ints[__i / 2])[__i % 2];
+		});
+	      }
+	    else
+	      __assert_unreachable<_To>();
+	  }
+#endif // _GLIBCXX_SIMD_X86INTRIN }}}
+	else if constexpr ((_FromVT::_S_partial_width - _Offset)
+			   > _ToVT::_S_full_size)
+	  {
+	    /*
+	    static_assert(
+	      (_FromVT::_S_partial_width & (_FromVT::_S_partial_width - 1)) =3D=3D
+	    0,
+	      "__convert_all only supports power-of-2 number of elements.
+	    Otherwise " "the return type cannot be array<_To, N>.");
+	      */
+	    constexpr size_t _NTotal
+	      =3D (_FromVT::_S_partial_width - _Offset) / _ToVT::_S_full_size;
+	    constexpr size_t _Np =3D _NParts =3D=3D 0 ? _NTotal : _NParts;
+	    static_assert(
+	      _Np <=3D _NTotal
+	      || (_Np =3D=3D _NTotal + 1
+		  && (_FromVT::_S_partial_width - _Offset) % _ToVT::_S_full_size
+		       > 0));
+	    using _R =3D array<_To, _Np>;
+	    if constexpr (_Np =3D=3D 1)
+	      return _R{__vector_convert<_To>(
+		__extract_part<_Offset, _FromVT::_S_partial_width,
+			       _ToVT::_S_full_size>(__v))};
+	    else
+	      return __generate_from_n_evaluations<_Np, _R>([&](
+		auto __i) constexpr {
+		auto __part
+		  =3D __extract_part<__i * _ToVT::_S_full_size + _Offset,
+				   _FromVT::_S_partial_width,
+				   _ToVT::_S_full_size>(__v);
+		return __vector_convert<_To>(__part);
+	      });
+	  }
+	else if constexpr (_Offset =3D=3D 0)
+	  return array<_To, 1>{__vector_convert<_To>(__v)};
+	else
+	  return array<_To, 1>{__vector_convert<_To>(
+	    __extract_part<_Offset, _FromVT::_S_partial_width,
+			   _FromVT::_S_partial_width - _Offset>(__v))};
+      }
+  }
+
+// }}}
+
+// _GnuTraits {{{
+template <typename _Tp, typename _Mp, typename _Abi, size_t _Np>
+  struct _GnuTraits
+  {
+    using _IsValid =3D true_type;
+    using _SimdImpl =3D typename _Abi::_SimdImpl;
+    using _MaskImpl =3D typename _Abi::_MaskImpl;
+
+    // simd and simd_mask member types {{{
+    using _SimdMember =3D _SimdWrapper<_Tp, _Np>;
+    using _MaskMember =3D _SimdWrapper<_Mp, _Np>;
+    static constexpr size_t _S_simd_align =3D alignof(_SimdMember);
+    static constexpr size_t _S_mask_align =3D alignof(_MaskMember);
+
+    // }}}
+    // size metadata {{{
+    static constexpr size_t _S_full_size =3D _SimdMember::_S_full_size;
+    static constexpr bool _S_is_partial =3D _SimdMember::_S_is_partial;
+
+    // }}}
+    // _SimdBase / base class for simd, providing extra conversions {{{
+    struct _SimdBase2
+    {
+      explicit operator __intrinsic_type_t<_Tp, _Np>() const
+      {
+	return __to_intrin(static_cast<const simd<_Tp, _Abi>*>(this)->_M_data);
+      }
+      explicit operator __vector_type_t<_Tp, _Np>() const
+      {
+	return static_cast<const simd<_Tp, _Abi>*>(this)->_M_data.__builtin();
+      }
+    };
+
+    struct _SimdBase1
+    {
+      explicit operator __intrinsic_type_t<_Tp, _Np>() const
+      { return __data(*static_cast<const simd<_Tp, _Abi>*>(this)); }
+    };
+
+    using _SimdBase =3D conditional_t<
+      is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::va=
lue,
+      _SimdBase1, _SimdBase2>;
+
+    // }}}
+    // _MaskBase {{{
+    struct _MaskBase2
+    {
+      explicit operator __intrinsic_type_t<_Tp, _Np>() const
+      {
+	return static_cast<const simd_mask<_Tp, _Abi>*>(this)
+	  ->_M_data.__intrin();
+      }
+      explicit operator __vector_type_t<_Tp, _Np>() const
+      {
+	return static_cast<const simd_mask<_Tp, _Abi>*>(this)->_M_data._M_data;
+      }
+    };
+
+    struct _MaskBase1
+    {
+      explicit operator __intrinsic_type_t<_Tp, _Np>() const
+      { return __data(*static_cast<const simd_mask<_Tp, _Abi>*>(this)); }
+    };
+
+    using _MaskBase =3D conditional_t<
+      is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::va=
lue,
+      _MaskBase1, _MaskBase2>;
+
+    // }}}
+    // _MaskCastType {{{
+    // parameter type of one explicit simd_mask constructor
+    class _MaskCastType
+    {
+      using _Up =3D __intrinsic_type_t<_Tp, _Np>;
+      _Up _M_data;
+
+    public:
+      _MaskCastType(_Up __x) : _M_data(__x) {}
+      operator _MaskMember() const { return _M_data; }
+    };
+
+    // }}}
+    // _SimdCastType {{{
+    // parameter type of one explicit simd constructor
+    class _SimdCastType1
+    {
+      using _Ap =3D __intrinsic_type_t<_Tp, _Np>;
+      _SimdMember _M_data;
+
+    public:
+      _SimdCastType1(_Ap __a) : _M_data(__vector_bitcast<_Tp>(__a)) {}
+      operator _SimdMember() const { return _M_data; }
+    };
+
+    class _SimdCastType2
+    {
+      using _Ap =3D __intrinsic_type_t<_Tp, _Np>;
+      using _B =3D __vector_type_t<_Tp, _Np>;
+      _SimdMember _M_data;
+
+    public:
+      _SimdCastType2(_Ap __a) : _M_data(__vector_bitcast<_Tp>(__a)) {}
+      _SimdCastType2(_B __b) : _M_data(__b) {}
+      operator _SimdMember() const { return _M_data; }
+    };
+
+    using _SimdCastType =3D conditional_t<
+      is_same<__intrinsic_type_t<_Tp, _Np>, __vector_type_t<_Tp, _Np>>::va=
lue,
+      _SimdCastType1, _SimdCastType2>;
+    //}}}
+  };
+
+// }}}
+struct _CommonImplX86;
+struct _CommonImplNeon;
+struct _CommonImplBuiltin;
+template <typename _Abi> struct _SimdImplBuiltin;
+template <typename _Abi> struct _MaskImplBuiltin;
+template <typename _Abi> struct _SimdImplX86;
+template <typename _Abi> struct _MaskImplX86;
+template <typename _Abi> struct _SimdImplNeon;
+template <typename _Abi> struct _MaskImplNeon;
+template <typename _Abi> struct _SimdImplPpc;
+
+// simd_abi::_VecBuiltin {{{
+template <int _UsedBytes>
+  struct simd_abi::_VecBuiltin
+  {
+    template <typename _Tp>
+      static constexpr size_t _S_size =3D _UsedBytes / sizeof(_Tp);
+
+    // validity traits {{{
+    struct _IsValidAbiTag : __bool_constant<(_UsedBytes > 1)> {};
+
+    template <typename _Tp>
+      struct _IsValidSizeFor
+	: __bool_constant<(_UsedBytes / sizeof(_Tp) > 1
+			   && _UsedBytes % sizeof(_Tp) =3D=3D 0
+			   && _UsedBytes <=3D __vectorized_sizeof<_Tp>()
+			   && (!__have_avx512f || _UsedBytes <=3D 32))> {};
+
+    template <typename _Tp>
+      struct _IsValid : conjunction<_IsValidAbiTag, __is_vectorizable<_Tp>,
+				    _IsValidSizeFor<_Tp>> {};
+
+    template <typename _Tp>
+      static constexpr bool _S_is_valid_v =3D _IsValid<_Tp>::value;
+
+    // }}}
+    // _SimdImpl/_MaskImpl {{{
+#if _GLIBCXX_SIMD_X86INTRIN
+    using _CommonImpl =3D _CommonImplX86;
+    using _SimdImpl =3D _SimdImplX86<_VecBuiltin<_UsedBytes>>;
+    using _MaskImpl =3D _MaskImplX86<_VecBuiltin<_UsedBytes>>;
+#elif _GLIBCXX_SIMD_HAVE_NEON
+    using _CommonImpl =3D _CommonImplNeon;
+    using _SimdImpl =3D _SimdImplNeon<_VecBuiltin<_UsedBytes>>;
+    using _MaskImpl =3D _MaskImplNeon<_VecBuiltin<_UsedBytes>>;
+#else
+    using _CommonImpl =3D _CommonImplBuiltin;
+#ifdef __ALTIVEC__
+    using _SimdImpl =3D _SimdImplPpc<_VecBuiltin<_UsedBytes>>;
+#else
+    using _SimdImpl =3D _SimdImplBuiltin<_VecBuiltin<_UsedBytes>>;
+#endif
+    using _MaskImpl =3D _MaskImplBuiltin<_VecBuiltin<_UsedBytes>>;
+#endif
+
+    // }}}
+    // __traits {{{
+    template <typename _Tp>
+      using _MaskValueType =3D __int_for_sizeof_t<_Tp>;
+
+    template <typename _Tp>
+      using __traits
+	=3D conditional_t<_S_is_valid_v<_Tp>,
+			_GnuTraits<_Tp, _MaskValueType<_Tp>,
+				   _VecBuiltin<_UsedBytes>, _S_size<_Tp>>,
+			_InvalidTraits>;
+
+    //}}}
+    // size metadata {{{
+    template <typename _Tp>
+      static constexpr size_t _S_full_size =3D __traits<_Tp>::_S_full_size;
+
+    template <typename _Tp>
+      static constexpr bool _S_is_partial =3D __traits<_Tp>::_S_is_partial;
+
+    // }}}
+    // implicit masks {{{
+    template <typename _Tp>
+      using _MaskMember =3D _SimdWrapper<_MaskValueType<_Tp>, _S_size<_Tp>=
>;
+
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_implicit_mask()
+      {
+	using _UV =3D typename _MaskMember<_Tp>::_BuiltinType;
+	if constexpr (!_MaskMember<_Tp>::_S_is_partial)
+	  return ~_UV();
+	else
+	  {
+	    constexpr auto __size =3D _S_size<_Tp>;
+	    _GLIBCXX_SIMD_USE_CONSTEXPR auto __r =3D __generate_vector<_UV>(
+	      [](auto __i) constexpr { return __i < __size ? -1 : 0; });
+	    return __r;
+	  }
+      }
+
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr __intrinsic_type_t<_Tp,
+								  _S_size<_Tp>>
+      _S_implicit_mask_intrin()
+      {
+	return __to_intrin(
+	  __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>()._M_data));
+      }
+
+    template <typename _TW, typename _TVT =3D _VectorTraits<_TW>>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _TW _S_masked(_TW __x)
+      {
+	using _Tp =3D typename _TVT::value_type;
+	if constexpr (!_MaskMember<_Tp>::_S_is_partial)
+	  return __x;
+	else
+	  return __and(__as_vector(__x),
+		       __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>()));
+      }
+
+    template <typename _TW, typename _TVT =3D _VectorTraits<_TW>>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr auto
+      __make_padding_nonzero(_TW __x)
+      {
+	using _Tp =3D typename _TVT::value_type;
+	if constexpr (!_S_is_partial<_Tp>)
+	  return __x;
+	else
+	  {
+	    _GLIBCXX_SIMD_USE_CONSTEXPR auto __implicit_mask
+	      =3D __vector_bitcast<_Tp>(_S_implicit_mask<_Tp>());
+	    if constexpr (is_integral_v<_Tp>)
+	      return __or(__x, ~__implicit_mask);
+	    else
+	      {
+		_GLIBCXX_SIMD_USE_CONSTEXPR auto __one
+		  =3D __andnot(__implicit_mask,
+			     __vector_broadcast<_S_full_size<_Tp>>(_Tp(1)));
+		// it's not enough to return `x | 1_in_padding` because the
+		// padding in x might be inf or nan (independent of
+		// __FINITE_MATH_ONLY__, because it's about padding bits)
+		return __or(__and(__x, __implicit_mask), __one);
+	      }
+	  }
+      }
+    // }}}
+  };
+
+// }}}
+// simd_abi::_VecBltnBtmsk {{{
+template <int _UsedBytes>
+  struct simd_abi::_VecBltnBtmsk
+  {
+    template <typename _Tp>
+      static constexpr size_t _S_size =3D _UsedBytes / sizeof(_Tp);
+
+    // validity traits {{{
+    struct _IsValidAbiTag : __bool_constant<(_UsedBytes > 1)> {};
+
+    template <typename _Tp>
+      struct _IsValidSizeFor
+	: __bool_constant<(_UsedBytes / sizeof(_Tp) > 1
+			   && _UsedBytes % sizeof(_Tp) =3D=3D 0 && _UsedBytes <=3D 64
+			   && (_UsedBytes > 32 || __have_avx512vl))> {};
+
+    // Bitmasks require at least AVX512F. If sizeof(_Tp) < 4 the AVX512BW =
is also
+    // required.
+    template <typename _Tp>
+      struct _IsValid
+	: conjunction<
+	    _IsValidAbiTag, __bool_constant<__have_avx512f>,
+	    __bool_constant<__have_avx512bw || (sizeof(_Tp) >=3D 4)>,
+	    __bool_constant<(__vectorized_sizeof<_Tp>() > sizeof(_Tp))>,
+	    _IsValidSizeFor<_Tp>> {};
+
+    template <typename _Tp>
+      static constexpr bool _S_is_valid_v =3D _IsValid<_Tp>::value;
+
+    // }}}
+    // simd/_MaskImpl {{{
+  #if _GLIBCXX_SIMD_X86INTRIN
+    using _CommonImpl =3D _CommonImplX86;
+    using _SimdImpl =3D _SimdImplX86<_VecBltnBtmsk<_UsedBytes>>;
+    using _MaskImpl =3D _MaskImplX86<_VecBltnBtmsk<_UsedBytes>>;
+  #else
+    template <int>
+      struct _MissingImpl;
+
+    using _CommonImpl =3D _MissingImpl<_UsedBytes>;
+    using _SimdImpl =3D _MissingImpl<_UsedBytes>;
+    using _MaskImpl =3D _MissingImpl<_UsedBytes>;
+  #endif
+
+    // }}}
+    // __traits {{{
+    template <typename _Tp>
+      using _MaskMember =3D _SimdWrapper<bool, _S_size<_Tp>>;
+
+    template <typename _Tp>
+      using __traits =3D conditional_t<
+	_S_is_valid_v<_Tp>,
+	_GnuTraits<_Tp, bool, _VecBltnBtmsk<_UsedBytes>, _S_size<_Tp>>,
+	_InvalidTraits>;
+
+    //}}}
+    // size metadata {{{
+    template <typename _Tp>
+      static constexpr size_t _S_full_size =3D __traits<_Tp>::_S_full_size;
+    template <typename _Tp>
+      static constexpr bool _S_is_partial =3D __traits<_Tp>::_S_is_partial;
+
+    // }}}
+    // implicit mask {{{
+  private:
+    template <typename _Tp>
+      using _ImplicitMask =3D _SimdWrapper<bool, _S_size<_Tp>>;
+
+  public:
+    template <size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr __bool_storage_member_type_=
t<_Np>
+      __implicit_mask_n()
+      {
+	using _Tp =3D __bool_storage_member_type_t<_Np>;
+	return _Np < sizeof(_Tp) * __CHAR_BIT__ ? _Tp((1ULL << _Np) - 1) : ~_Tp();
+      }
+
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _ImplicitMask<_Tp>
+      _S_implicit_mask()
+      { return __implicit_mask_n<_S_size<_Tp>>(); }
+
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr __bool_storage_member_type_=
t<
+	_S_size<_Tp>>
+      _S_implicit_mask_intrin()
+      { return __implicit_mask_n<_S_size<_Tp>>(); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_masked(_SimdWrapper<_Tp, _Np> __x)
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  if constexpr (_Np < 8 || (_Np & (_Np - 1)) !=3D 0)
+	    return _MaskImpl::_S_bit_and(
+	      __x, _SimdWrapper<_Tp, _Np>(
+		     __bool_storage_member_type_t<_Np>((1ULL << _Np) - 1)));
+	  else
+	    return __x;
+	else
+	  return _S_masked(__x._M_data);
+      }
+
+    template <typename _TV>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _TV
+      _S_masked(_TV __x)
+      {
+	using _Tp =3D typename _VectorTraits<_TV>::value_type;
+	static_assert(
+	  !__is_bitmask_v<_TV>,
+	  "_VecBltnBtmsk::_S_masked cannot work on bitmasks, since it doesn't "
+	  "know the number of elements. Use _SimdWrapper<bool, N> instead.");
+	if constexpr (_S_is_partial<_Tp>)
+	  {
+	    constexpr size_t _Np =3D _S_size<_Tp>;
+	    return __make_dependent_t<_TV, _CommonImpl>::_S_blend(
+	      _S_implicit_mask<_Tp>(), _SimdWrapper<_Tp, _Np>(),
+	      _SimdWrapper<_Tp, _Np>(__x));
+	  }
+	else
+	  return __x;
+      }
+
+    template <typename _TV, typename _TVT =3D _VectorTraits<_TV>>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr auto
+      __make_padding_nonzero(_TV __x)
+      {
+	using _Tp =3D typename _TVT::value_type;
+	if constexpr (!_S_is_partial<_Tp>)
+	  return __x;
+	else
+	  {
+	    constexpr size_t _Np =3D _S_size<_Tp>;
+	    if constexpr (is_integral_v<typename _TVT::value_type>)
+	      return __x
+		     | __generate_vector<_Tp, _S_full_size<_Tp>>(
+		       [](auto __i) -> _Tp {
+			 if (__i < _Np)
+			   return 0;
+			 else
+			   return 1;
+		       });
+	    else
+	      return __make_dependent_t<_TV, _CommonImpl>::_S_blend(
+		       _S_implicit_mask<_Tp>(),
+		       _SimdWrapper<_Tp, _Np>(
+			 __vector_broadcast<_S_full_size<_Tp>>(_Tp(1))),
+		       _SimdWrapper<_Tp, _Np>(__x))
+		._M_data;
+	  }
+      }
+
+    // }}}
+  };
+
+//}}}
+// _CommonImplBuiltin {{{
+struct _CommonImplBuiltin
+{
+  // _S_converts_via_decomposition{{{
+  // This lists all cases where a __vector_convert needs to fall back to
+  // conversion of individual scalars (i.e. decompose the input vector into
+  // scalars, convert, compose output vector). In those cases, _S_masked_l=
oad &
+  // _S_masked_store prefer to use the _S_bit_iteration implementation.
+  template <typename _From, typename _To, size_t _ToSize>
+    static inline constexpr bool __converts_via_decomposition_v
+      =3D sizeof(_From) !=3D sizeof(_To);
+
+  // }}}
+  // _S_load{{{
+  template <typename _Tp, size_t _Np, size_t _Bytes =3D _Np * sizeof(_Tp)>
+    _GLIBCXX_SIMD_INTRINSIC static __vector_type_t<_Tp, _Np>
+    _S_load(const void* __p)
+    {
+      static_assert(_Np > 1);
+      static_assert(_Bytes % sizeof(_Tp) =3D=3D 0);
+      using _Rp =3D __vector_type_t<_Tp, _Np>;
+      if constexpr (sizeof(_Rp) =3D=3D _Bytes)
+	{
+	  _Rp __r;
+	  __builtin_memcpy(&__r, __p, _Bytes);
+	  return __r;
+	}
+      else
+	{
+#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90424
+	  using _Up =3D conditional_t<
+	    is_integral_v<_Tp>,
+	    conditional_t<_Bytes % 4 =3D=3D 0,
+			  conditional_t<_Bytes % 8 =3D=3D 0, long long, int>,
+			  conditional_t<_Bytes % 2 =3D=3D 0, short, signed char>>,
+	    conditional_t<(_Bytes < 8 || _Np % 2 =3D=3D 1 || _Np =3D=3D 2), _Tp,
+			  double>>;
+	  using _V =3D __vector_type_t<_Up, _Np * sizeof(_Tp) / sizeof(_Up)>;
+	  if constexpr (sizeof(_V) !=3D sizeof(_Rp))
+	    { // on i386 with 4 < _Bytes <=3D 8
+	      _Rp __r{};
+	      __builtin_memcpy(&__r, __p, _Bytes);
+	      return __r;
+	    }
+	  else
+#else // _GLIBCXX_SIMD_WORKAROUND_PR90424
+	  using _V =3D _Rp;
+#endif // _GLIBCXX_SIMD_WORKAROUND_PR90424
+	    {
+	      _V __r{};
+	      static_assert(_Bytes <=3D sizeof(_V));
+	      __builtin_memcpy(&__r, __p, _Bytes);
+	      return reinterpret_cast<_Rp>(__r);
+	    }
+	}
+    }
+
+  // }}}
+  // _S_store {{{
+  template <size_t _ReqBytes =3D 0, typename _TV>
+    _GLIBCXX_SIMD_INTRINSIC static void _S_store(_TV __x, void* __addr)
+    {
+      constexpr size_t _Bytes =3D _ReqBytes =3D=3D 0 ? sizeof(__x) : _ReqB=
ytes;
+      static_assert(sizeof(__x) >=3D _Bytes);
+
+      if constexpr (__is_vector_type_v<_TV>)
+	{
+	  using _Tp =3D typename _VectorTraits<_TV>::value_type;
+	  constexpr size_t _Np =3D _Bytes / sizeof(_Tp);
+	  static_assert(_Np * sizeof(_Tp) =3D=3D _Bytes);
+
+#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90424
+	  using _Up =3D conditional_t<
+	    (is_integral_v<_Tp> || _Bytes < 4),
+	    conditional_t<(sizeof(__x) > sizeof(long long)), long long, _Tp>,
+	    float>;
+	  const auto __v =3D __vector_bitcast<_Up>(__x);
+#else // _GLIBCXX_SIMD_WORKAROUND_PR90424
+	  const __vector_type_t<_Tp, _Np> __v =3D __x;
+#endif // _GLIBCXX_SIMD_WORKAROUND_PR90424
+
+	  if constexpr ((_Bytes & (_Bytes - 1)) !=3D 0)
+	    {
+	      constexpr size_t _MoreBytes =3D std::__bit_ceil(_Bytes);
+	      alignas(decltype(__v)) char __tmp[_MoreBytes];
+	      __builtin_memcpy(__tmp, &__v, _MoreBytes);
+	      __builtin_memcpy(__addr, __tmp, _Bytes);
+	    }
+	  else
+	    __builtin_memcpy(__addr, &__v, _Bytes);
+	}
+      else
+	__builtin_memcpy(__addr, &__x, _Bytes);
+    }
+
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> __=
x,
+						 void* __addr)
+    { _S_store<_Np * sizeof(_Tp)>(__x._M_data, __addr); }
+
+  // }}}
+  // _S_store_bool_array(_BitMask) {{{
+  template <size_t _Np, bool _Sanitized>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr void
+    _S_store_bool_array(_BitMask<_Np, _Sanitized> __x, bool* __mem)
+    {
+      if constexpr (_Np =3D=3D 1)
+	__mem[0] =3D __x[0];
+      else if constexpr (_Np =3D=3D 2)
+	{
+	  short __bool2 =3D (__x._M_to_bits() * 0x81) & 0x0101;
+	  _S_store<_Np>(__bool2, __mem);
+	}
+      else if constexpr (_Np =3D=3D 3)
+	{
+	  int __bool3 =3D (__x._M_to_bits() * 0x4081) & 0x010101;
+	  _S_store<_Np>(__bool3, __mem);
+	}
+      else
+	{
+	  __execute_n_times<__div_roundup(_Np, 4)>([&](auto __i) {
+	    constexpr int __offset =3D __i * 4;
+	    constexpr int __remaining =3D _Np - __offset;
+	    if constexpr (__remaining > 4 && __remaining <=3D 7)
+	      {
+		const _ULLong __bool7
+		  =3D (__x.template _M_extract<__offset>()._M_to_bits()
+		     * 0x40810204081ULL)
+		    & 0x0101010101010101ULL;
+		_S_store<__remaining>(__bool7, __mem + __offset);
+	      }
+	    else if constexpr (__remaining >=3D 4)
+	      {
+		int __bits =3D __x.template _M_extract<__offset>()._M_to_bits();
+		if constexpr (__remaining > 7)
+		  __bits &=3D 0xf;
+		const int __bool4 =3D (__bits * 0x204081) & 0x01010101;
+		_S_store<4>(__bool4, __mem + __offset);
+	      }
+	  });
+	}
+    }
+
+  // }}}
+  // _S_blend{{{
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr auto
+    _S_blend(_SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k,
+	     _SimdWrapper<_Tp, _Np> __at0, _SimdWrapper<_Tp, _Np> __at1)
+    { return __k._M_data ? __at1._M_data : __at0._M_data; }
+
+  // }}}
+};
+
+// }}}
+// _SimdImplBuiltin {{{1
+template <typename _Abi>
+  struct _SimdImplBuiltin
+  {
+    // member types {{{2
+    template <typename _Tp>
+      static constexpr size_t _S_max_store_size =3D 16;
+
+    using abi_type =3D _Abi;
+
+    template <typename _Tp>
+      using _TypeTag =3D _Tp*;
+
+    template <typename _Tp>
+      using _SimdMember =3D typename _Abi::template __traits<_Tp>::_SimdMe=
mber;
+
+    template <typename _Tp>
+      using _MaskMember =3D typename _Abi::template _MaskMember<_Tp>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_size =3D _Abi::template _S_size<_Tp>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_full_size =3D _Abi::template _S_full_size=
<_Tp>;
+
+    using _CommonImpl =3D typename _Abi::_CommonImpl;
+    using _SuperImpl =3D typename _Abi::_SimdImpl;
+    using _MaskImpl =3D typename _Abi::_MaskImpl;
+
+    // _M_make_simd(_SimdWrapper/__intrinsic_type_t) {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static simd<_Tp, _Abi>
+      _M_make_simd(_SimdWrapper<_Tp, _Np> __x)
+      { return {__private_init, __x}; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static simd<_Tp, _Abi>
+      _M_make_simd(__intrinsic_type_t<_Tp, _Np> __x)
+      { return {__private_init, __vector_bitcast<_Tp>(__x)}; }
+
+    // _S_broadcast {{{2
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdMember<_Tp>
+      _S_broadcast(_Tp __x) noexcept
+      { return __vector_broadcast<_S_full_size<_Tp>>(__x); }
+
+    // _S_generator {{{2
+    template <typename _Fp, typename _Tp>
+      inline static constexpr _SimdMember<_Tp> _S_generator(_Fp&& __gen,
+							    _TypeTag<_Tp>)
+      {
+	return __generate_vector<_Tp, _S_full_size<_Tp>>([&](
+	  auto __i) constexpr {
+	  if constexpr (__i < _S_size<_Tp>)
+	    return __gen(__i);
+	  else
+	    return 0;
+	});
+      }
+
+    // _S_load {{{2
+    template <typename _Tp, typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdMember<_Tp>
+      _S_load(const _Up* __mem, _TypeTag<_Tp>) noexcept
+      {
+	constexpr size_t _Np =3D _S_size<_Tp>;
+	constexpr size_t __max_load_size
+	  =3D (sizeof(_Up) >=3D 4 && __have_avx512f) || __have_avx512bw   ? 64
+	    : (is_floating_point_v<_Up> && __have_avx) || __have_avx2 ? 32
+								      : 16;
+	constexpr size_t __bytes_to_load =3D sizeof(_Up) * _Np;
+	if constexpr (sizeof(_Up) > 8)
+	  return __generate_vector<_Tp, _SimdMember<_Tp>::_S_full_size>([&](
+	    auto __i) constexpr {
+	    return static_cast<_Tp>(__i < _Np ? __mem[__i] : 0);
+	  });
+	else if constexpr (is_same_v<_Up, _Tp>)
+	  return _CommonImpl::template _S_load<_Tp, _S_full_size<_Tp>,
+					       _Np * sizeof(_Tp)>(__mem);
+	else if constexpr (__bytes_to_load <=3D __max_load_size)
+	  return __convert<_SimdMember<_Tp>>(
+	    _CommonImpl::template _S_load<_Up, _Np>(__mem));
+	else if constexpr (__bytes_to_load % __max_load_size =3D=3D 0)
+	  {
+	    constexpr size_t __n_loads =3D __bytes_to_load / __max_load_size;
+	    constexpr size_t __elements_per_load =3D _Np / __n_loads;
+	    return __call_with_n_evaluations<__n_loads>(
+	      [](auto... __uncvted) {
+		return __convert<_SimdMember<_Tp>>(__uncvted...);
+	      },
+	      [&](auto __i) {
+		return _CommonImpl::template _S_load<_Up, __elements_per_load>(
+		  __mem + __i * __elements_per_load);
+	      });
+	  }
+	else if constexpr (__bytes_to_load % (__max_load_size / 2) =3D=3D 0
+			   && __max_load_size > 16)
+	  { // e.g. int[] -> <char, 12> with AVX2
+	    constexpr size_t __n_loads
+	      =3D __bytes_to_load / (__max_load_size / 2);
+	    constexpr size_t __elements_per_load =3D _Np / __n_loads;
+	    return __call_with_n_evaluations<__n_loads>(
+	      [](auto... __uncvted) {
+		return __convert<_SimdMember<_Tp>>(__uncvted...);
+	      },
+	      [&](auto __i) {
+		return _CommonImpl::template _S_load<_Up, __elements_per_load>(
+		  __mem + __i * __elements_per_load);
+	      });
+	  }
+	else // e.g. int[] -> <char, 9>
+	  return __call_with_subscripts(
+	    __mem, make_index_sequence<_Np>(), [](auto... __args) {
+	      return __vector_type_t<_Tp, _S_full_size<_Tp>>{
+		static_cast<_Tp>(__args)...};
+	    });
+      }
+
+    // _S_masked_load {{{2
+    template <typename _Tp, size_t _Np, typename _Up>
+      static inline _SimdWrapper<_Tp, _Np>
+      _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
+		     const _Up* __mem) noexcept
+      {
+	_BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k), [&](auto __i) {
+	  __merge._M_set(__i, static_cast<_Tp>(__mem[__i]));
+	});
+	return __merge;
+      }
+
+    // _S_store {{{2
+    template <typename _Tp, typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_store(_SimdMember<_Tp> __v, _Up* __mem, _TypeTag<_Tp>) noexcept
+      {
+	// TODO: converting int -> "smaller int" can be optimized with AVX512
+	constexpr size_t _Np =3D _S_size<_Tp>;
+	constexpr size_t __max_store_size
+	  =3D _SuperImpl::template _S_max_store_size<_Up>;
+	if constexpr (sizeof(_Up) > 8)
+	  __execute_n_times<_Np>([&](auto __i) constexpr {
+	    __mem[__i] =3D __v[__i];
+	  });
+	else if constexpr (is_same_v<_Up, _Tp>)
+	  _CommonImpl::_S_store(__v, __mem);
+	else if constexpr (sizeof(_Up) * _Np <=3D __max_store_size)
+	  _CommonImpl::_S_store(_SimdWrapper<_Up, _Np>(__convert<_Up>(__v)),
+				__mem);
+	else
+	  {
+	    constexpr size_t __vsize =3D __max_store_size / sizeof(_Up);
+	    // round up to convert the last partial vector as well:
+	    constexpr size_t __stores =3D __div_roundup(_Np, __vsize);
+	    constexpr size_t __full_stores =3D _Np / __vsize;
+	    using _V =3D __vector_type_t<_Up, __vsize>;
+	    const array<_V, __stores> __converted
+	      =3D __convert_all<_V, __stores>(__v);
+	    __execute_n_times<__full_stores>([&](auto __i) constexpr {
+	      _CommonImpl::_S_store(__converted[__i], __mem + __i * __vsize);
+	    });
+	    if constexpr (__full_stores < __stores)
+	      _CommonImpl::template _S_store<(_Np - __full_stores * __vsize)
+					     * sizeof(_Up)>(
+		__converted[__full_stores], __mem + __full_stores * __vsize);
+	  }
+      }
+
+    // _S_masked_store_nocvt {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
+			    _MaskMember<_Tp> __k)
+      {
+	_BitOps::_S_bit_iteration(
+	  _MaskImpl::_S_to_bits(__k), [&](auto __i) constexpr {
+	    __mem[__i] =3D __v[__i];
+	  });
+      }
+
+    // _S_masked_store {{{2
+    template <typename _TW, typename _TVT =3D _VectorTraits<_TW>,
+	      typename _Tp =3D typename _TVT::value_type, typename _Up>
+      static inline void
+      _S_masked_store(const _TW __v, _Up* __mem, const _MaskMember<_Tp> __=
k)
+	noexcept
+      {
+	constexpr size_t _TV_size =3D _S_size<_Tp>;
+	[[maybe_unused]] const auto __vi =3D __to_intrin(__v);
+	constexpr size_t __max_store_size
+	  =3D _SuperImpl::template _S_max_store_size<_Up>;
+	if constexpr (
+	  is_same_v<
+	    _Tp,
+	    _Up> || (is_integral_v<_Tp> && is_integral_v<_Up> && sizeof(_Tp) =3D=
=3D sizeof(_Up)))
+	  {
+	    // bitwise or no conversion, reinterpret:
+	    const _MaskMember<_Up> __kk =3D [&]() {
+	      if constexpr (__is_bitmask_v<decltype(__k)>)
+		return _MaskMember<_Up>(__k._M_data);
+	      else
+		return __wrapper_bitcast<__int_for_sizeof_t<_Up>>(__k);
+	    }();
+	    _SuperImpl::_S_masked_store_nocvt(__wrapper_bitcast<_Up>(__v),
+					      __mem, __kk);
+	  }
+	else if constexpr (__vectorized_sizeof<_Up>() > sizeof(_Up)
+			   && !_CommonImpl::
+				template __converts_via_decomposition_v<
+				  _Tp, _Up, __max_store_size>)
+	  { // conversion via decomposition is better handled via the
+	    // bit_iteration
+	    // fallback below
+	    constexpr size_t _UW_size
+	      =3D std::min(_TV_size, __max_store_size / sizeof(_Up));
+	    static_assert(_UW_size <=3D _TV_size);
+	    using _UW =3D _SimdWrapper<_Up, _UW_size>;
+	    using _UV =3D __vector_type_t<_Up, _UW_size>;
+	    using _UAbi =3D simd_abi::deduce_t<_Up, _UW_size>;
+	    if constexpr (_UW_size =3D=3D _TV_size) // one convert+store
+	      {
+		const _UW __converted =3D __convert<_UW>(__v);
+		_SuperImpl::_S_masked_store_nocvt(
+		  __converted, __mem,
+		  _UAbi::_MaskImpl::template _S_convert<
+		    __int_for_sizeof_t<_Up>>(__k));
+	      }
+	    else
+	      {
+		static_assert(_UW_size * sizeof(_Up) =3D=3D __max_store_size);
+		constexpr size_t _NFullStores =3D _TV_size / _UW_size;
+		constexpr size_t _NAllStores
+		  =3D __div_roundup(_TV_size, _UW_size);
+		constexpr size_t _NParts =3D _S_full_size<_Tp> / _UW_size;
+		const array<_UV, _NAllStores> __converted
+		  =3D __convert_all<_UV, _NAllStores>(__v);
+		__execute_n_times<_NFullStores>([&](auto __i) {
+		  _SuperImpl::_S_masked_store_nocvt(
+		    _UW(__converted[__i]), __mem + __i * _UW_size,
+		    _UAbi::_MaskImpl::template _S_convert<
+		      __int_for_sizeof_t<_Up>>(
+		      __extract_part<__i, _NParts>(__k.__as_full_vector())));
+		});
+		if constexpr (_NAllStores
+			      > _NFullStores) // one partial at the end
+		  _SuperImpl::_S_masked_store_nocvt(
+		    _UW(__converted[_NFullStores]),
+		    __mem + _NFullStores * _UW_size,
+		    _UAbi::_MaskImpl::template _S_convert<
+		      __int_for_sizeof_t<_Up>>(
+		      __extract_part<_NFullStores, _NParts>(
+			__k.__as_full_vector())));
+	      }
+	  }
+	else
+	  _BitOps::_S_bit_iteration(
+	    _MaskImpl::_S_to_bits(__k), [&](auto __i) constexpr {
+	      __mem[__i] =3D static_cast<_Up>(__v[__i]);
+	    });
+      }
+
+    // _S_complement {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_complement(_SimdWrapper<_Tp, _Np> __x) noexcept
+      { return ~__x._M_data; }
+
+    // _S_unary_minus {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_unary_minus(_SimdWrapper<_Tp, _Np> __x) noexcept
+      {
+	// GCC doesn't use the psign instructions, but pxor & psub seem to be
+	// just as good a choice as pcmpeqd & psign. So meh.
+	return -__x._M_data;
+      }
+
+    // arithmetic operators {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_plus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __x._M_data + __y._M_data; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_minus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __x._M_data - __y._M_data; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_multiplies(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __x._M_data * __y._M_data; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_divides(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      {
+	// Note that division by 0 is always UB, so we must ensure we avoid the
+	// case for partial registers
+	if constexpr (!_Abi::template _S_is_partial<_Tp>)
+	  return __x._M_data / __y._M_data;
+	else
+	  return __x._M_data / _Abi::__make_padding_nonzero(__y._M_data);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_modulus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      {
+	if constexpr (!_Abi::template _S_is_partial<_Tp>)
+	  return __x._M_data % __y._M_data;
+	else
+	  return __as_vector(__x)
+		 % _Abi::__make_padding_nonzero(__as_vector(__y));
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_and(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __and(__x, __y); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_or(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __or(__x, __y); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_xor(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __xor(__x, __y); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np>=
 __y)
+      { return __x._M_data << __y._M_data; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np=
> __y)
+      { return __x._M_data >> __y._M_data; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x, int __y)
+      { return __x._M_data << __y; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x, int __y)
+      { return __x._M_data >> __y; }
+
+    // compares {{{2
+    // _S_equal_to {{{3
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __x._M_data =3D=3D __y._M_data; }
+
+    // _S_not_equal_to {{{3
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_not_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> _=
_y)
+      { return __x._M_data !=3D __y._M_data; }
+
+    // _S_less {{{3
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_less(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __x._M_data < __y._M_data; }
+
+    // _S_less_equal {{{3
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_less_equal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      { return __x._M_data <=3D __y._M_data; }
+
+    // _S_negate {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_negate(_SimdWrapper<_Tp, _Np> __x) noexcept
+      { return !__x._M_data; }
+
+    // _S_min, _S_max, _S_minmax {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC static constexpr
+      _SimdWrapper<_Tp, _Np>
+      _S_min(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
+      { return __a._M_data < __b._M_data ? __a._M_data : __b._M_data; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC static constexpr
+      _SimdWrapper<_Tp, _Np>
+      _S_max(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
+      { return __a._M_data > __b._M_data ? __a._M_data : __b._M_data; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_NORMAL_MATH _GLIBCXX_SIMD_INTRINSIC static constexpr
+      pair<_SimdWrapper<_Tp, _Np>, _SimdWrapper<_Tp, _Np>>
+      _S_minmax(_SimdWrapper<_Tp, _Np> __a, _SimdWrapper<_Tp, _Np> __b)
+      {
+	return {__a._M_data < __b._M_data ? __a._M_data : __b._M_data,
+		__a._M_data < __b._M_data ? __b._M_data : __a._M_data};
+      }
+
+    // reductions {{{2
+    template <size_t _Np, size_t... _Is, size_t... _Zeros, typename _Tp,
+	      typename _BinaryOperation>
+      _GLIBCXX_SIMD_INTRINSIC static _Tp
+      _S_reduce_partial(index_sequence<_Is...>, index_sequence<_Zeros...>,
+			simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
+      {
+	using _V =3D __vector_type_t<_Tp, _Np / 2>;
+	static_assert(sizeof(_V) <=3D sizeof(__x));
+	// _S_full_size is the size of the smallest native SIMD register that
+	// can store _Np/2 elements:
+	using _FullSimd =3D __deduced_simd<_Tp, _VectorTraits<_V>::_S_full_size>;
+	using _HalfSimd =3D __deduced_simd<_Tp, _Np / 2>;
+	const auto __xx =3D __as_vector(__x);
+	return _HalfSimd::abi_type::_SimdImpl::_S_reduce(
+	  static_cast<_HalfSimd>(__as_vector(__binary_op(
+	    static_cast<_FullSimd>(__intrin_bitcast<_V>(__xx)),
+	    static_cast<_FullSimd>(__intrin_bitcast<_V>(
+	      __vector_permute<(_Np / 2 + _Is)..., (int(_Zeros * 0) - 1)...>(
+		__xx)))))),
+	  __binary_op);
+      }
+
+    template <typename _Tp, typename _BinaryOperation>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _Tp
+      _S_reduce(simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
+      {
+	constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	if constexpr (_Np =3D=3D 1)
+	  return __x[0];
+	else if constexpr (_Np =3D=3D 2)
+	  return __binary_op(simd<_Tp, simd_abi::scalar>(__x[0]),
+			     simd<_Tp, simd_abi::scalar>(__x[1]))[0];
+	else if constexpr (_Abi::template _S_is_partial<_Tp>) //{{{
+	  {
+	    [[maybe_unused]] constexpr auto __full_size
+	      =3D _Abi::template _S_full_size<_Tp>;
+	    if constexpr (_Np =3D=3D 3)
+	      return __binary_op(
+		__binary_op(simd<_Tp, simd_abi::scalar>(__x[0]),
+			    simd<_Tp, simd_abi::scalar>(__x[1])),
+		simd<_Tp, simd_abi::scalar>(__x[2]))[0];
+	    else if constexpr (is_same_v<__remove_cvref_t<_BinaryOperation>,
+					 plus<>>)
+	      {
+		using _Ap =3D simd_abi::deduce_t<_Tp, __full_size>;
+		return _Ap::_SimdImpl::_S_reduce(
+		  simd<_Tp, _Ap>(__private_init,
+				 _Abi::_S_masked(__as_vector(__x))),
+		  __binary_op);
+	      }
+	    else if constexpr (is_same_v<__remove_cvref_t<_BinaryOperation>,
+					 multiplies<>>)
+	      {
+		using _Ap =3D simd_abi::deduce_t<_Tp, __full_size>;
+		using _TW =3D _SimdWrapper<_Tp, __full_size>;
+		_GLIBCXX_SIMD_USE_CONSTEXPR auto __implicit_mask_full
+		  =3D _Abi::template _S_implicit_mask<_Tp>().__as_full_vector();
+		_GLIBCXX_SIMD_USE_CONSTEXPR _TW __one
+		  =3D __vector_broadcast<__full_size>(_Tp(1));
+		const _TW __x_full =3D __data(__x).__as_full_vector();
+		const _TW __x_padded_with_ones
+		  =3D _Ap::_CommonImpl::_S_blend(__implicit_mask_full, __one,
+					       __x_full);
+		return _Ap::_SimdImpl::_S_reduce(
+		  simd<_Tp, _Ap>(__private_init, __x_padded_with_ones),
+		  __binary_op);
+	      }
+	    else if constexpr (_Np & 1)
+	      {
+		using _Ap =3D simd_abi::deduce_t<_Tp, _Np - 1>;
+		return __binary_op(
+		  simd<_Tp, simd_abi::scalar>(_Ap::_SimdImpl::_S_reduce(
+		    simd<_Tp, _Ap>(
+		      __intrin_bitcast<__vector_type_t<_Tp, _Np - 1>>(
+			__as_vector(__x))),
+		    __binary_op)),
+		  simd<_Tp, simd_abi::scalar>(__x[_Np - 1]))[0];
+	      }
+	    else
+	      return _S_reduce_partial<_Np>(
+		make_index_sequence<_Np / 2>(),
+		make_index_sequence<__full_size - _Np / 2>(), __x, __binary_op);
+	  }                                   //}}}
+	else if constexpr (sizeof(__x) =3D=3D 16) //{{{
+	  {
+	    if constexpr (_Np =3D=3D 16)
+	      {
+		const auto __y =3D __data(__x);
+		__x =3D __binary_op(
+		  _M_make_simd<_Tp, _Np>(
+		    __vector_permute<0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6,
+				     7, 7>(__y)),
+		  _M_make_simd<_Tp, _Np>(
+		    __vector_permute<8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13,
+				     14, 14, 15, 15>(__y)));
+	      }
+	    if constexpr (_Np >=3D 8)
+	      {
+		const auto __y =3D __vector_bitcast<short>(__data(__x));
+		__x =3D __binary_op(
+		  _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
+		    __vector_permute<0, 0, 1, 1, 2, 2, 3, 3>(__y))),
+		  _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
+		    __vector_permute<4, 4, 5, 5, 6, 6, 7, 7>(__y))));
+	      }
+	    if constexpr (_Np >=3D 4)
+	      {
+		using _Up =3D conditional_t<is_floating_point_v<_Tp>, float, int>;
+		const auto __y =3D __vector_bitcast<_Up>(__data(__x));
+		__x =3D __binary_op(__x,
+				  _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
+				    __vector_permute<3, 2, 1, 0>(__y))));
+	      }
+	    using _Up =3D conditional_t<is_floating_point_v<_Tp>, double, _LLong>;
+	    const auto __y =3D __vector_bitcast<_Up>(__data(__x));
+	    __x =3D __binary_op(__x, _M_make_simd<_Tp, _Np>(__vector_bitcast<_Tp>(
+				     __vector_permute<1, 1>(__y))));
+	    return __x[0];
+	  } //}}}
+	else
+	  {
+	    static_assert(sizeof(__x) > __min_vector_size<_Tp>);
+	    static_assert((_Np & (_Np - 1)) =3D=3D 0); // _Np must be a power of 2
+	    using _Ap =3D simd_abi::deduce_t<_Tp, _Np / 2>;
+	    using _V =3D simd<_Tp, _Ap>;
+	    return _Ap::_SimdImpl::_S_reduce(
+	      __binary_op(_V(__private_init, __extract<0, 2>(__as_vector(__x))),
+			  _V(__private_init,
+			     __extract<1, 2>(__as_vector(__x)))),
+	      static_cast<_BinaryOperation&&>(__binary_op));
+	  }
+      }
+
+    // math {{{2
+    // frexp, modf and copysign implemented in simd_math.h
+#define _GLIBCXX_SIMD_MATH_FALLBACK(__name)                               =
     \
+    template <typename _Tp, typename... _More>                            =
     \
+      static _Tp _S_##__name(const _Tp& __x, const _More&... __more)      =
     \
+      {                                                                   =
     \
+	return __generate_vector<_Tp>(                                         \
+	  [&](auto __i) { return __name(__x[__i], __more[__i]...); });         \
+      }
+
+#define _GLIBCXX_SIMD_MATH_FALLBACK_MASKRET(__name)                       =
     \
+    template <typename _Tp, typename... _More>                            =
     \
+      static typename _Tp::mask_type _S_##__name(const _Tp& __x,          =
     \
+						 const _More&... __more)       \
+      {                                                                   =
     \
+	return __generate_vector<_Tp>(                                         \
+	  [&](auto __i) { return __name(__x[__i], __more[__i]...); });         \
+      }
+
+#define _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(_RetTp, __name)              =
     \
+    template <typename _Tp, typename... _More>                            =
     \
+      static auto _S_##__name(const _Tp& __x, const _More&... __more)     =
     \
+      {                                                                   =
     \
+	return __fixed_size_storage_t<_RetTp,                                  \
+				      _VectorTraits<_Tp>::_S_partial_width>::  \
+	  _S_generate([&](auto __meta) constexpr {                             \
+	    return __meta._S_generator(                                        \
+	      [&](auto __i) {                                                  \
+		return __name(__x[__meta._S_offset + __i],                     \
+			      __more[__meta._S_offset + __i]...);              \
+	      },                                                               \
+	      static_cast<_RetTp*>(nullptr));                                  \
+	  });                                                                  \
+      }
+
+    _GLIBCXX_SIMD_MATH_FALLBACK(acos)
+    _GLIBCXX_SIMD_MATH_FALLBACK(asin)
+    _GLIBCXX_SIMD_MATH_FALLBACK(atan)
+    _GLIBCXX_SIMD_MATH_FALLBACK(atan2)
+    _GLIBCXX_SIMD_MATH_FALLBACK(cos)
+    _GLIBCXX_SIMD_MATH_FALLBACK(sin)
+    _GLIBCXX_SIMD_MATH_FALLBACK(tan)
+    _GLIBCXX_SIMD_MATH_FALLBACK(acosh)
+    _GLIBCXX_SIMD_MATH_FALLBACK(asinh)
+    _GLIBCXX_SIMD_MATH_FALLBACK(atanh)
+    _GLIBCXX_SIMD_MATH_FALLBACK(cosh)
+    _GLIBCXX_SIMD_MATH_FALLBACK(sinh)
+    _GLIBCXX_SIMD_MATH_FALLBACK(tanh)
+    _GLIBCXX_SIMD_MATH_FALLBACK(exp)
+    _GLIBCXX_SIMD_MATH_FALLBACK(exp2)
+    _GLIBCXX_SIMD_MATH_FALLBACK(expm1)
+    _GLIBCXX_SIMD_MATH_FALLBACK(ldexp)
+    _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(int, ilogb)
+    _GLIBCXX_SIMD_MATH_FALLBACK(log)
+    _GLIBCXX_SIMD_MATH_FALLBACK(log10)
+    _GLIBCXX_SIMD_MATH_FALLBACK(log1p)
+    _GLIBCXX_SIMD_MATH_FALLBACK(log2)
+    _GLIBCXX_SIMD_MATH_FALLBACK(logb)
+
+    // modf implemented in simd_math.h
+    _GLIBCXX_SIMD_MATH_FALLBACK(scalbn)
+    _GLIBCXX_SIMD_MATH_FALLBACK(scalbln)
+    _GLIBCXX_SIMD_MATH_FALLBACK(cbrt)
+    _GLIBCXX_SIMD_MATH_FALLBACK(fabs)
+    _GLIBCXX_SIMD_MATH_FALLBACK(pow)
+    _GLIBCXX_SIMD_MATH_FALLBACK(sqrt)
+    _GLIBCXX_SIMD_MATH_FALLBACK(erf)
+    _GLIBCXX_SIMD_MATH_FALLBACK(erfc)
+    _GLIBCXX_SIMD_MATH_FALLBACK(lgamma)
+    _GLIBCXX_SIMD_MATH_FALLBACK(tgamma)
+
+    _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long, lrint)
+    _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long long, llrint)
+
+    _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long, lround)
+    _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET(long long, llround)
+
+    _GLIBCXX_SIMD_MATH_FALLBACK(fmod)
+    _GLIBCXX_SIMD_MATH_FALLBACK(remainder)
+
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      static _Tp
+      _S_remquo(const _Tp __x, const _Tp __y,
+		__fixed_size_storage_t<int, _TVT::_S_partial_width>* __z)
+      {
+	return __generate_vector<_Tp>([&](auto __i) {
+	  int __tmp;
+	  auto __r =3D remquo(__x[__i], __y[__i], &__tmp);
+	  __z->_M_set(__i, __tmp);
+	  return __r;
+	});
+      }
+
+    // copysign in simd_math.h
+    _GLIBCXX_SIMD_MATH_FALLBACK(nextafter)
+    _GLIBCXX_SIMD_MATH_FALLBACK(fdim)
+    _GLIBCXX_SIMD_MATH_FALLBACK(fmax)
+    _GLIBCXX_SIMD_MATH_FALLBACK(fmin)
+    _GLIBCXX_SIMD_MATH_FALLBACK(fma)
+
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_isgreater(_SimdWrapper<_Tp, _Np> __x,
+		   _SimdWrapper<_Tp, _Np> __y) noexcept
+      {
+	using _Ip =3D __int_for_sizeof_t<_Tp>;
+	const auto __xn =3D __vector_bitcast<_Ip>(__x);
+	const auto __yn =3D __vector_bitcast<_Ip>(__y);
+	const auto __xp =3D __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
+	const auto __yp =3D __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
+	return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
+			__xp > __yp);
+      }
+
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_isgreaterequal(_SimdWrapper<_Tp, _Np> __x,
+			_SimdWrapper<_Tp, _Np> __y) noexcept
+      {
+	using _Ip =3D __int_for_sizeof_t<_Tp>;
+	const auto __xn =3D __vector_bitcast<_Ip>(__x);
+	const auto __yn =3D __vector_bitcast<_Ip>(__y);
+	const auto __xp =3D __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
+	const auto __yp =3D __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
+	return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
+			__xp >=3D __yp);
+      }
+
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_isless(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y) no=
except
+      {
+	using _Ip =3D __int_for_sizeof_t<_Tp>;
+	const auto __xn =3D __vector_bitcast<_Ip>(__x);
+	const auto __yn =3D __vector_bitcast<_Ip>(__y);
+	const auto __xp =3D __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
+	const auto __yp =3D __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
+	return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
+			__xp < __yp);
+      }
+
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_islessequal(_SimdWrapper<_Tp, _Np> __x,
+		     _SimdWrapper<_Tp, _Np> __y) noexcept
+      {
+	using _Ip =3D __int_for_sizeof_t<_Tp>;
+	const auto __xn =3D __vector_bitcast<_Ip>(__x);
+	const auto __yn =3D __vector_bitcast<_Ip>(__y);
+	const auto __xp =3D __xn < 0 ? -(__xn & __finite_max_v<_Ip>) : __xn;
+	const auto __yp =3D __yn < 0 ? -(__yn & __finite_max_v<_Ip>) : __yn;
+	return __andnot(_SuperImpl::_S_isunordered(__x, __y)._M_data,
+			__xp <=3D __yp);
+      }
+
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_islessgreater(_SimdWrapper<_Tp, _Np> __x,
+		       _SimdWrapper<_Tp, _Np> __y) noexcept
+      {
+	return __andnot(_SuperImpl::_S_isunordered(__x, __y),
+			_SuperImpl::_S_not_equal_to(__x, __y));
+      }
+
+#undef _GLIBCXX_SIMD_MATH_FALLBACK
+#undef _GLIBCXX_SIMD_MATH_FALLBACK_MASKRET
+#undef _GLIBCXX_SIMD_MATH_FALLBACK_FIXEDRET
+    // _S_abs {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+    _S_abs(_SimdWrapper<_Tp, _Np> __x) noexcept
+    {
+      // if (__builtin_is_constant_evaluated())
+      //  {
+      //    return __x._M_data < 0 ? -__x._M_data : __x._M_data;
+      //  }
+      if constexpr (is_floating_point_v<_Tp>)
+	// `v < 0 ? -v : v` cannot compile to the efficient implementation of
+	// masking the signbit off because it must consider v =3D=3D -0
+
+	// ~(-0.) & v would be easy, but breaks with fno-signed-zeros
+	return __and(_S_absmask<__vector_type_t<_Tp, _Np>>, __x._M_data);
+      else
+	return __x._M_data < 0 ? -__x._M_data : __x._M_data;
+    }
+
+    // }}}3
+    // _S_plus_minus {{{
+    // Returns __x + __y - __y without -fassociative-math optimizing to __=
x.
+    // - _TV must be __vector_type_t<floating-point type, N>.
+    // - _UV must be _TV or floating-point type.
+    template <typename _TV, typename _UV>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _TV _S_plus_minus(_TV __x,
+							       _UV __y) noexcept
+    {
+  #if defined __i386__ && !defined __SSE_MATH__
+      if constexpr (sizeof(__x) =3D=3D 8)
+	{ // operations on __x would use the FPU
+	  static_assert(is_same_v<_TV, __vector_type_t<float, 2>>);
+	  const auto __x4 =3D __vector_bitcast<float, 4>(__x);
+	  if constexpr (is_same_v<_TV, _UV>)
+	    return __vector_bitcast<float, 2>(
+	      _S_plus_minus(__x4, __vector_bitcast<float, 4>(__y)));
+	  else
+	    return __vector_bitcast<float, 2>(_S_plus_minus(__x4, __y));
+	}
+  #endif
+  #if !defined __clang__ && __GCC_IEC_559 =3D=3D 0
+      if (__builtin_is_constant_evaluated()
+	  || (__builtin_constant_p(__x) && __builtin_constant_p(__y)))
+	return (__x + __y) - __y;
+      else
+	return [&] {
+	  __x +=3D __y;
+	  if constexpr(__have_sse)
+	    {
+	      if constexpr (sizeof(__x) >=3D 16)
+		asm("" : "+x"(__x));
+	      else if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
+		asm("" : "+x"(__x[0]), "+x"(__x[1]));
+	      else
+		__assert_unreachable<_TV>();
+	    }
+	  else if constexpr(__have_neon)
+	    asm("" : "+w"(__x));
+	  else if constexpr (__have_power_vmx)
+	    {
+	      if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
+		asm("" : "+fgr"(__x[0]), "+fgr"(__x[1]));
+	      else
+		asm("" : "+v"(__x));
+	    }
+	  else
+	    asm("" : "+g"(__x));
+	  return __x - __y;
+	}();
+  #else
+      return (__x + __y) - __y;
+  #endif
+    }
+
+    // }}}
+    // _S_nearbyint {{{3
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_nearbyint(_Tp __x_) noexcept
+    {
+      using value_type =3D typename _TVT::value_type;
+      using _V =3D typename _TVT::type;
+      const _V __x =3D __x_;
+      const _V __absx =3D __and(__x, _S_absmask<_V>);
+      static_assert(__CHAR_BIT__ * sizeof(1ull) >=3D __digits_v<value_type=
>);
+      _GLIBCXX_SIMD_USE_CONSTEXPR _V __shifter_abs
+	=3D _V() + (1ull << (__digits_v<value_type> - 1));
+      const _V __shifter =3D __or(__and(_S_signmask<_V>, __x), __shifter_a=
bs);
+      const _V __shifted =3D _S_plus_minus(__x, __shifter);
+      return __absx < __shifter_abs ? __shifted : __x;
+    }
+
+    // _S_rint {{{3
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_rint(_Tp __x) noexcept
+    {
+      return _SuperImpl::_S_nearbyint(__x);
+    }
+
+    // _S_trunc {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+    _S_trunc(_SimdWrapper<_Tp, _Np> __x)
+    {
+      using _V =3D __vector_type_t<_Tp, _Np>;
+      const _V __absx =3D __and(__x._M_data, _S_absmask<_V>);
+      static_assert(__CHAR_BIT__ * sizeof(1ull) >=3D __digits_v<_Tp>);
+      constexpr _Tp __shifter =3D 1ull << (__digits_v<_Tp> - 1);
+      _V __truncated =3D _S_plus_minus(__absx, __shifter);
+      __truncated -=3D __truncated > __absx ? _V() + 1 : _V();
+      return __absx < __shifter ? __or(__xor(__absx, __x._M_data), __trunc=
ated)
+				: __x._M_data;
+    }
+
+    // _S_round {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+    _S_round(_SimdWrapper<_Tp, _Np> __x)
+    {
+      const auto __abs_x =3D _SuperImpl::_S_abs(__x);
+      const auto __t_abs =3D _SuperImpl::_S_trunc(__abs_x)._M_data;
+      const auto __r_abs // round(abs(x)) =3D
+	=3D __t_abs + (__abs_x._M_data - __t_abs >=3D _Tp(.5) ? _Tp(1) : 0);
+      return __or(__xor(__abs_x._M_data, __x._M_data), __r_abs);
+    }
+
+    // _S_floor {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+    _S_floor(_SimdWrapper<_Tp, _Np> __x)
+    {
+      const auto __y =3D _SuperImpl::_S_trunc(__x)._M_data;
+      const auto __negative_input
+	=3D __vector_bitcast<_Tp>(__x._M_data < __vector_broadcast<_Np, _Tp>(0));
+      const auto __mask
+	=3D __andnot(__vector_bitcast<_Tp>(__y =3D=3D __x._M_data), __negative_in=
put);
+      return __or(__andnot(__mask, __y),
+		  __and(__mask, __y - __vector_broadcast<_Np, _Tp>(1)));
+    }
+
+    // _S_ceil {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+    _S_ceil(_SimdWrapper<_Tp, _Np> __x)
+    {
+      const auto __y =3D _SuperImpl::_S_trunc(__x)._M_data;
+      const auto __negative_input
+	=3D __vector_bitcast<_Tp>(__x._M_data < __vector_broadcast<_Np, _Tp>(0));
+      const auto __inv_mask
+	=3D __or(__vector_bitcast<_Tp>(__y =3D=3D __x._M_data), __negative_input);
+      return __or(__and(__inv_mask, __y),
+		  __andnot(__inv_mask, __y + __vector_broadcast<_Np, _Tp>(1)));
+    }
+
+    // _S_isnan {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+    _S_isnan([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
+    {
+  #if __FINITE_MATH_ONLY__
+      return {}; // false
+  #elif !defined __SUPPORT_SNAN__
+      return ~(__x._M_data =3D=3D __x._M_data);
+  #elif defined __STDC_IEC_559__
+      using _Ip =3D __int_for_sizeof_t<_Tp>;
+      const auto __absn =3D __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
+      const auto __infn
+	=3D __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__infinity_v<_Tp>));
+      return __infn < __absn;
+  #else
+  #error "Not implemented: how to support SNaN but non-IEC559 floating-poi=
nt?"
+  #endif
+    }
+
+    // _S_isfinite {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+    _S_isfinite([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
+    {
+  #if __FINITE_MATH_ONLY__
+      using _UV =3D typename _MaskMember<_Tp>::_BuiltinType;
+      _GLIBCXX_SIMD_USE_CONSTEXPR _UV __alltrue =3D ~_UV();
+      return __alltrue;
+  #else
+      // if all exponent bits are set, __x is either inf or NaN
+      using _Ip =3D __int_for_sizeof_t<_Tp>;
+      const auto __absn =3D __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
+      const auto __maxn
+	=3D __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__finite_max_v<_Tp>));
+      return __absn <=3D __maxn;
+  #endif
+    }
+
+    // _S_isunordered {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+    _S_isunordered(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+    {
+      return __or(_S_isnan(__x), _S_isnan(__y));
+    }
+
+    // _S_signbit {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+    _S_signbit(_SimdWrapper<_Tp, _Np> __x)
+    {
+      using _Ip =3D __int_for_sizeof_t<_Tp>;
+      return __vector_bitcast<_Ip>(__x) < 0;
+      // Arithmetic right shift (SRA) would also work (instead of compare)=
, but
+      // 64-bit SRA isn't available on x86 before AVX512. And in general,
+      // compares are more likely to be efficient than SRA.
+    }
+
+    // _S_isinf {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+    _S_isinf([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x)
+    {
+  #if __FINITE_MATH_ONLY__
+      return {}; // false
+  #else
+      return _SuperImpl::template _S_equal_to<_Tp, _Np>(_SuperImpl::_S_abs=
(__x),
+							__vector_broadcast<_Np>(
+							  __infinity_v<_Tp>));
+      // alternative:
+      // compare to inf using the corresponding integer type
+      /*
+	 return
+	 __vector_bitcast<_Tp>(__vector_bitcast<__int_for_sizeof_t<_Tp>>(
+			       _S_abs(__x)._M_data)
+	 =3D=3D
+	 __vector_bitcast<__int_for_sizeof_t<_Tp>>(__vector_broadcast<_Np>(
+	 __infinity_v<_Tp>)));
+	 */
+  #endif
+    }
+
+    // _S_isnormal {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+    _S_isnormal(_SimdWrapper<_Tp, _Np> __x)
+    {
+      using _Ip =3D __int_for_sizeof_t<_Tp>;
+      const auto __absn =3D __vector_bitcast<_Ip>(_SuperImpl::_S_abs(__x));
+      const auto __minn
+	=3D __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__norm_min_v<_Tp>));
+  #if __FINITE_MATH_ONLY__
+      return __absn >=3D __minn;
+  #else
+      const auto __maxn
+	=3D __vector_bitcast<_Ip>(__vector_broadcast<_Np>(__finite_max_v<_Tp>));
+      return __minn <=3D __absn && __absn <=3D __maxn;
+  #endif
+    }
+
+    // _S_fpclassify {{{3
+    template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static __fixed_size_storage_t<int, _Np>
+    _S_fpclassify(_SimdWrapper<_Tp, _Np> __x)
+    {
+      using _I =3D __int_for_sizeof_t<_Tp>;
+      const auto __xn
+	=3D __vector_bitcast<_I>(__to_intrin(_SuperImpl::_S_abs(__x)));
+      constexpr size_t _NI =3D sizeof(__xn) / sizeof(_I);
+      _GLIBCXX_SIMD_USE_CONSTEXPR auto __minn
+	=3D __vector_bitcast<_I>(__vector_broadcast<_NI>(__norm_min_v<_Tp>));
+      _GLIBCXX_SIMD_USE_CONSTEXPR auto __infn
+	=3D __vector_bitcast<_I>(__vector_broadcast<_NI>(__infinity_v<_Tp>));
+
+      _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_normal
+	=3D __vector_broadcast<_NI, _I>(FP_NORMAL);
+  #if !__FINITE_MATH_ONLY__
+      _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_nan
+	=3D __vector_broadcast<_NI, _I>(FP_NAN);
+      _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_infinite
+	=3D __vector_broadcast<_NI, _I>(FP_INFINITE);
+  #endif
+  #ifndef __FAST_MATH__
+      _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_subnormal
+	=3D __vector_broadcast<_NI, _I>(FP_SUBNORMAL);
+  #endif
+      _GLIBCXX_SIMD_USE_CONSTEXPR auto __fp_zero
+	=3D __vector_broadcast<_NI, _I>(FP_ZERO);
+
+      __vector_type_t<_I, _NI>
+	__tmp =3D __xn < __minn
+  #ifdef __FAST_MATH__
+		  ? __fp_zero
+  #else
+		  ? (__xn =3D=3D 0 ? __fp_zero : __fp_subnormal)
+  #endif
+  #if __FINITE_MATH_ONLY__
+		  : __fp_normal;
+  #else
+		  : (__xn < __infn ? __fp_normal
+				   : (__xn =3D=3D __infn ? __fp_infinite : __fp_nan));
+  #endif
+
+      if constexpr (sizeof(_I) =3D=3D sizeof(int))
+	{
+	  using _FixedInt =3D __fixed_size_storage_t<int, _Np>;
+	  const auto __as_int =3D __vector_bitcast<int, _Np>(__tmp);
+	  if constexpr (_FixedInt::_S_tuple_size =3D=3D 1)
+	    return {__as_int};
+	  else if constexpr (_FixedInt::_S_tuple_size =3D=3D 2
+			     && is_same_v<
+			       typename _FixedInt::_SecondType::_FirstAbi,
+			       simd_abi::scalar>)
+	    return {__extract<0, 2>(__as_int), __as_int[_Np - 1]};
+	  else if constexpr (_FixedInt::_S_tuple_size =3D=3D 2)
+	    return {__extract<0, 2>(__as_int),
+		    __auto_bitcast(__extract<1, 2>(__as_int))};
+	  else
+	    __assert_unreachable<_Tp>();
+	}
+      else if constexpr (_Np =3D=3D 2 && sizeof(_I) =3D=3D 8
+			 && __fixed_size_storage_t<int, _Np>::_S_tuple_size =3D=3D 2)
+	{
+	  const auto __aslong =3D __vector_bitcast<_LLong>(__tmp);
+	  return {int(__aslong[0]), {int(__aslong[1])}};
+	}
+  #if _GLIBCXX_SIMD_X86INTRIN
+      else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(__tmp) =3D=3D 32
+			 && __fixed_size_storage_t<int, _Np>::_S_tuple_size =3D=3D 1)
+	return {_mm_packs_epi32(__to_intrin(__lo128(__tmp)),
+				__to_intrin(__hi128(__tmp)))};
+      else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(__tmp) =3D=3D 64
+			 && __fixed_size_storage_t<int, _Np>::_S_tuple_size =3D=3D 1)
+	return {_mm512_cvtepi64_epi32(__to_intrin(__tmp))};
+  #endif // _GLIBCXX_SIMD_X86INTRIN
+      else if constexpr (__fixed_size_storage_t<int, _Np>::_S_tuple_size =
=3D=3D 1)
+	return {__call_with_subscripts<_Np>(__vector_bitcast<_LLong>(__tmp),
+					    [](auto... __l) {
+					      return __make_wrapper<int>(__l...);
+					    })};
+      else
+	__assert_unreachable<_Tp>();
+    }
+
+    // _S_increment & _S_decrement{{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_increment(_SimdWrapper<_Tp, _Np>& __x)
+      { __x =3D __x._M_data + 1; }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_decrement(_SimdWrapper<_Tp, _Np>& __x)
+      { __x =3D __x._M_data - 1; }
+
+    // smart_reference access {{{2
+    template <typename _Tp, size_t _Np, typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC constexpr static void
+      _S_set(_SimdWrapper<_Tp, _Np>& __v, int __i, _Up&& __x) noexcept
+      { __v._M_set(__i, static_cast<_Up&&>(__x)); }
+
+    // _S_masked_assign{{{2
+    template <typename _Tp, typename _K, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(_SimdWrapper<_K, _Np> __k, _SimdWrapper<_Tp, _Np>& =
__lhs,
+		       __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs)
+      {
+	if (__k._M_is_constprop_none_of())
+	  return;
+	else if (__k._M_is_constprop_all_of())
+	  __lhs =3D __rhs;
+	else
+	  __lhs =3D _CommonImpl::_S_blend(__k, __lhs, __rhs);
+      }
+
+    template <typename _Tp, typename _K, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(_SimdWrapper<_K, _Np> __k, _SimdWrapper<_Tp, _Np>& =
__lhs,
+		       __type_identity_t<_Tp> __rhs)
+      {
+	if (__k._M_is_constprop_none_of())
+	  return;
+	else if (__k._M_is_constprop_all_of())
+	  __lhs =3D __vector_broadcast<_Np>(__rhs);
+	else if (__builtin_constant_p(__rhs) && __rhs =3D=3D 0)
+	  {
+	    if constexpr (!is_same_v<bool, _K>)
+	      // the __andnot optimization only makes sense if __k._M_data is a
+	      // vector register
+	      __lhs._M_data
+		=3D __andnot(__vector_bitcast<_Tp>(__k), __lhs._M_data);
+	    else
+	      // for AVX512/__mmask, a _mm512_maskz_mov is best
+	      __lhs
+		=3D _CommonImpl::_S_blend(__k, __lhs, _SimdWrapper<_Tp, _Np>());
+	  }
+	else
+	  __lhs =3D _CommonImpl::_S_blend(__k, __lhs,
+					_SimdWrapper<_Tp, _Np>(
+					  __vector_broadcast<_Np>(__rhs)));
+      }
+
+    // _S_masked_cassign {{{2
+    template <typename _Op, typename _Tp, typename _K, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_cassign(const _SimdWrapper<_K, _Np> __k,
+			_SimdWrapper<_Tp, _Np>& __lhs,
+			const __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs,
+			_Op __op)
+      {
+	if (__k._M_is_constprop_none_of())
+	  return;
+	else if (__k._M_is_constprop_all_of())
+	  __lhs =3D __op(_SuperImpl{}, __lhs, __rhs);
+	else
+	  __lhs =3D _CommonImpl::_S_blend(__k, __lhs,
+					__op(_SuperImpl{}, __lhs, __rhs));
+      }
+
+    template <typename _Op, typename _Tp, typename _K, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_cassign(const _SimdWrapper<_K, _Np> __k,
+			_SimdWrapper<_Tp, _Np>& __lhs,
+			const __type_identity_t<_Tp> __rhs, _Op __op)
+      { _S_masked_cassign(__k, __lhs, __vector_broadcast<_Np>(__rhs), __op=
); }
+
+    // _S_masked_unary {{{2
+    template <template <typename> class _Op, typename _Tp, typename _K,
+	      size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_masked_unary(const _SimdWrapper<_K, _Np> __k,
+		      const _SimdWrapper<_Tp, _Np> __v)
+      {
+	if (__k._M_is_constprop_none_of())
+	  return __v;
+	auto __vv =3D _M_make_simd(__v);
+	_Op<decltype(__vv)> __op;
+	if (__k._M_is_constprop_all_of())
+	  return __data(__op(__vv));
+	else
+	  return _CommonImpl::_S_blend(__k, __v, __data(__op(__vv)));
+      }
+
+    //}}}2
+  };
+
+// _MaskImplBuiltinMixin {{{1
+struct _MaskImplBuiltinMixin
+{
+  template <typename _Tp>
+    using _TypeTag =3D _Tp*;
+
+  // _S_to_maskvector {{{
+  template <typename _Up, size_t _ToN =3D 1>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
+    _S_to_maskvector(bool __x)
+    {
+      static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
+      return __x ? __vector_type_t<_Up, _ToN>{~_Up()}
+		 : __vector_type_t<_Up, _ToN>{};
+    }
+
+  template <typename _Up, size_t _UpN =3D 0, size_t _Np, bool _Sanitized,
+	    size_t _ToN =3D _UpN =3D=3D 0 ? _Np : _UpN>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
+    _S_to_maskvector(_BitMask<_Np, _Sanitized> __x)
+    {
+      static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
+      return __generate_vector<__vector_type_t<_Up, _ToN>>([&](
+	auto __i) constexpr {
+	if constexpr (__i < _Np)
+	  return __x[__i] ? ~_Up() : _Up();
+	else
+	  return _Up();
+      });
+    }
+
+  template <typename _Up, size_t _UpN =3D 0, typename _Tp, size_t _Np,
+	    size_t _ToN =3D _UpN =3D=3D 0 ? _Np : _UpN>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
+    _S_to_maskvector(_SimdWrapper<_Tp, _Np> __x)
+    {
+      static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
+      using _TW =3D _SimdWrapper<_Tp, _Np>;
+      using _UW =3D _SimdWrapper<_Up, _ToN>;
+      if constexpr (sizeof(_Up) =3D=3D sizeof(_Tp) && sizeof(_TW) =3D=3D s=
izeof(_UW))
+	return __wrapper_bitcast<_Up, _ToN>(__x);
+      else if constexpr (is_same_v<_Tp, bool>) // bits -> vector
+	return _S_to_maskvector<_Up, _ToN>(_BitMask<_Np>(__x._M_data));
+      else
+	{ // vector -> vector
+	  /*
+	  [[maybe_unused]] const auto __y =3D __vector_bitcast<_Up>(__x._M_data);
+	  if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 4 && sizeof(__=
y) =3D=3D
+	  16) return __vector_permute<1, 3, -1, -1>(__y); else if constexpr
+	  (sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 2
+			     && sizeof(__y) =3D=3D 16)
+	    return __vector_permute<1, 3, 5, 7, -1, -1, -1, -1>(__y);
+	  else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 2
+			     && sizeof(__y) =3D=3D 16)
+	    return __vector_permute<3, 7, -1, -1, -1, -1, -1, -1>(__y);
+	  else if constexpr (sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 1
+			     && sizeof(__y) =3D=3D 16)
+	    return __vector_permute<1, 3, 5, 7, 9, 11, 13, 15, -1, -1, -1, -1,
+	  -1, -1, -1, -1>(__y); else if constexpr (sizeof(_Tp) =3D=3D 4 &&
+	  sizeof(_Up) =3D=3D 1
+			     && sizeof(__y) =3D=3D 16)
+	    return __vector_permute<3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1,
+	  -1, -1, -1, -1, -1>(__y); else if constexpr (sizeof(_Tp) =3D=3D 8 &&
+	  sizeof(_Up) =3D=3D 1
+			     && sizeof(__y) =3D=3D 16)
+	    return __vector_permute<7, 15, -1, -1, -1, -1, -1, -1, -1, -1, -1,
+	  -1, -1, -1, -1, -1>(__y); else
+	  */
+	  {
+	    return __generate_vector<__vector_type_t<_Up, _ToN>>([&](
+	      auto __i) constexpr {
+	      if constexpr (__i < _Np)
+		return _Up(__x[__i.value]);
+	      else
+		return _Up();
+	    });
+	  }
+	}
+    }
+
+  // }}}
+  // _S_to_bits {{{
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
+    _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
+    {
+      static_assert(!is_same_v<_Tp, bool>);
+      static_assert(_Np <=3D __CHAR_BIT__ * sizeof(_ULLong));
+      using _Up =3D make_unsigned_t<__int_for_sizeof_t<_Tp>>;
+      const auto __bools
+	=3D __vector_bitcast<_Up>(__x) >> (sizeof(_Up) * __CHAR_BIT__ - 1);
+      _ULLong __r =3D 0;
+      __execute_n_times<_Np>(
+	[&](auto __i) { __r |=3D _ULLong(__bools[__i.value]) << __i; });
+      return __r;
+    }
+
+  // }}}
+};
+
+// _MaskImplBuiltin {{{1
+template <typename _Abi>
+  struct _MaskImplBuiltin : _MaskImplBuiltinMixin
+  {
+    using _MaskImplBuiltinMixin::_S_to_bits;
+    using _MaskImplBuiltinMixin::_S_to_maskvector;
+
+    // member types {{{
+    template <typename _Tp>
+      using _SimdMember =3D typename _Abi::template __traits<_Tp>::_SimdMe=
mber;
+
+    template <typename _Tp>
+      using _MaskMember =3D typename _Abi::template _MaskMember<_Tp>;
+
+    using _SuperImpl =3D typename _Abi::_MaskImpl;
+    using _CommonImpl =3D typename _Abi::_CommonImpl;
+
+    template <typename _Tp>
+      static constexpr size_t _S_size =3D simd_size_v<_Tp, _Abi>;
+
+    // }}}
+    // _S_broadcast {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_broadcast(bool __x)
+      {
+	return __x ? _Abi::template _S_implicit_mask<_Tp>()
+		   : _MaskMember<_Tp>();
+      }
+
+    // }}}
+    // _S_load {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_load(const bool* __mem)
+      {
+	using _I =3D __int_for_sizeof_t<_Tp>;
+	if constexpr (sizeof(_Tp) =3D=3D sizeof(bool))
+	  {
+	    const auto __bools
+	      =3D _CommonImpl::template _S_load<_I, _S_size<_Tp>>(__mem);
+	    // bool is {0, 1}, everything else is UB
+	    return __bools > 0;
+	  }
+	else
+	  return __generate_vector<_I, _S_size<_Tp>>([&](auto __i) constexpr {
+	    return __mem[__i] ? ~_I() : _I();
+	  });
+      }
+
+    // }}}
+    // _S_convert {{{
+    template <typename _Tp, size_t _Np, bool _Sanitized>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr auto
+      _S_convert(_BitMask<_Np, _Sanitized> __x)
+      {
+	if constexpr (__is_builtin_bitmask_abi<_Abi>())
+	  return _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>(__x._M_to_bits());
+	else
+	  return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
+						       _S_size<_Tp>>(
+	    __x._M_sanitized());
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr auto
+      _S_convert(_SimdWrapper<bool, _Np> __x)
+      {
+	if constexpr (__is_builtin_bitmask_abi<_Abi>())
+	  return _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>(__x._M_data);
+	else
+	  return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
+						       _S_size<_Tp>>(
+	    _BitMask<_Np>(__x._M_data)._M_sanitized());
+      }
+
+    template <typename _Tp, typename _Up, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr auto
+      _S_convert(_SimdWrapper<_Up, _Np> __x)
+      {
+	if constexpr (__is_builtin_bitmask_abi<_Abi>())
+	  return _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>(
+	    _SuperImpl::_S_to_bits(__x));
+	else
+	  return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
+						       _S_size<_Tp>>(__x);
+      }
+
+    template <typename _Tp, typename _Up, typename _UAbi>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr auto
+      _S_convert(simd_mask<_Up, _UAbi> __x)
+      {
+	if constexpr (__is_builtin_bitmask_abi<_Abi>())
+	  {
+	    using _R =3D _SimdWrapper<bool, simd_size_v<_Tp, _Abi>>;
+	    if constexpr (__is_builtin_bitmask_abi<_UAbi>()) // bits -> bits
+	      return _R(__data(__x));
+	    else if constexpr (__is_scalar_abi<_UAbi>()) // bool -> bits
+	      return _R(__data(__x));
+	    else if constexpr (__is_fixed_size_abi_v<_UAbi>) // bitset -> bits
+	      return _R(__data(__x)._M_to_bits());
+	    else // vector -> bits
+	      return _R(_UAbi::_MaskImpl::_S_to_bits(__data(__x))._M_to_bits());
+	  }
+	else
+	  return _SuperImpl::template _S_to_maskvector<__int_for_sizeof_t<_Tp>,
+						       _S_size<_Tp>>(
+	    __data(__x));
+      }
+
+    // }}}
+    // _S_masked_load {{{2
+    template <typename _Tp, size_t _Np>
+      static inline _SimdWrapper<_Tp, _Np>
+      _S_masked_load(_SimdWrapper<_Tp, _Np> __merge,
+		     _SimdWrapper<_Tp, _Np> __mask, const bool* __mem) noexcept
+      {
+	// AVX(2) has 32/64 bit maskload, but nothing at 8 bit granularity
+	auto __tmp =3D __wrapper_bitcast<__int_for_sizeof_t<_Tp>>(__merge);
+	_BitOps::_S_bit_iteration(_SuperImpl::_S_to_bits(__mask),
+				  [&](auto __i) {
+				    __tmp._M_set(__i, -__mem[__i]);
+				  });
+	__merge =3D __wrapper_bitcast<_Tp>(__tmp);
+	return __merge;
+      }
+
+    // _S_store {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> =
__v,
+						   bool* __mem) noexcept
+      {
+	__execute_n_times<_Np>([&](auto __i) constexpr {
+	  __mem[__i] =3D __v[__i];
+	});
+      }
+
+    // _S_masked_store {{{2
+    template <typename _Tp, size_t _Np>
+      static inline void
+      _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, bool* __mem,
+		      const _SimdWrapper<_Tp, _Np> __k) noexcept
+      {
+	_BitOps::_S_bit_iteration(
+	  _SuperImpl::_S_to_bits(__k), [&](auto __i) constexpr {
+	    __mem[__i] =3D __v[__i];
+	  });
+      }
+
+    // _S_from_bitmask{{{2
+    template <size_t _Np, typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_from_bitmask(_SanitizedBitMask<_Np> __bits, _TypeTag<_Tp>)
+      {
+	return _SuperImpl::template _S_to_maskvector<_Tp, _S_size<_Tp>>(__bits);
+      }
+
+    // logical and bitwise operators {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_logical_and(const _SimdWrapper<_Tp, _Np>& __x,
+		     const _SimdWrapper<_Tp, _Np>& __y)
+      { return __and(__x._M_data, __y._M_data); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_logical_or(const _SimdWrapper<_Tp, _Np>& __x,
+		    const _SimdWrapper<_Tp, _Np>& __y)
+      { return __or(__x._M_data, __y._M_data); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_not(const _SimdWrapper<_Tp, _Np>& __x)
+      {
+	if constexpr (_Abi::template _S_is_partial<_Tp>)
+	  return __andnot(__x, __wrapper_bitcast<_Tp>(
+				 _Abi::template _S_implicit_mask<_Tp>()));
+	else
+	  return __not(__x._M_data);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_and(const _SimdWrapper<_Tp, _Np>& __x,
+		 const _SimdWrapper<_Tp, _Np>& __y)
+      { return __and(__x._M_data, __y._M_data); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_or(const _SimdWrapper<_Tp, _Np>& __x,
+		const _SimdWrapper<_Tp, _Np>& __y)
+      { return __or(__x._M_data, __y._M_data); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_xor(const _SimdWrapper<_Tp, _Np>& __x,
+		 const _SimdWrapper<_Tp, _Np>& __y)
+      { return __xor(__x._M_data, __y._M_data); }
+
+    // smart_reference access {{{2
+    template <typename _Tp, size_t _Np>
+      static constexpr void _S_set(_SimdWrapper<_Tp, _Np>& __k, int __i,
+				   bool __x) noexcept
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  __k._M_set(__i, __x);
+	else
+	  {
+	    static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
+	    if (__builtin_is_constant_evaluated())
+	      {
+		__k =3D __generate_from_n_evaluations<_Np,
+						    __vector_type_t<_Tp, _Np>>(
+		  [&](auto __j) {
+		    if (__i =3D=3D __j)
+		      return _Tp(-__x);
+		    else
+		      return __k[+__j];
+		  });
+	      }
+	    else
+	      __k._M_data[__i] =3D -__x;
+	  }
+      }
+
+    // _S_masked_assign{{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(_SimdWrapper<_Tp, _Np> __k,
+		       _SimdWrapper<_Tp, _Np>& __lhs,
+		       __type_identity_t<_SimdWrapper<_Tp, _Np>> __rhs)
+      { __lhs =3D _CommonImpl::_S_blend(__k, __lhs, __rhs); }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(_SimdWrapper<_Tp, _Np> __k,
+		       _SimdWrapper<_Tp, _Np>& __lhs, bool __rhs)
+      {
+	if (__builtin_constant_p(__rhs))
+	  {
+	    if (__rhs =3D=3D false)
+	      __lhs =3D __andnot(__k, __lhs);
+	    else
+	      __lhs =3D __or(__k, __lhs);
+	    return;
+	  }
+	__lhs =3D _CommonImpl::_S_blend(__k, __lhs,
+				      __data(simd_mask<_Tp, _Abi>(__rhs)));
+      }
+
+    //}}}2
+    // _S_all_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool
+      _S_all_of(simd_mask<_Tp, _Abi> __k)
+      {
+	return __call_with_subscripts(
+	  __data(__k), make_index_sequence<_S_size<_Tp>>(),
+	  [](const auto... __ent) constexpr { return (... && !(__ent =3D=3D 0)); =
});
+      }
+
+    // }}}
+    // _S_any_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool
+      _S_any_of(simd_mask<_Tp, _Abi> __k)
+      {
+	return __call_with_subscripts(
+	  __data(__k), make_index_sequence<_S_size<_Tp>>(),
+	  [](const auto... __ent) constexpr { return (... || !(__ent =3D=3D 0)); =
});
+      }
+
+    // }}}
+    // _S_none_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool
+      _S_none_of(simd_mask<_Tp, _Abi> __k)
+      {
+	return __call_with_subscripts(
+	  __data(__k), make_index_sequence<_S_size<_Tp>>(),
+	  [](const auto... __ent) constexpr { return (... && (__ent =3D=3D 0)); }=
);
+      }
+
+    // }}}
+    // _S_some_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool
+      _S_some_of(simd_mask<_Tp, _Abi> __k)
+      {
+	const int __n_true =3D _S_popcount(__k);
+	return __n_true > 0 && __n_true < int(_S_size<_Tp>);
+      }
+
+    // }}}
+    // _S_popcount {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_popcount(simd_mask<_Tp, _Abi> __k)
+      {
+	using _I =3D __int_for_sizeof_t<_Tp>;
+	if constexpr (is_default_constructible_v<simd<_I, _Abi>>)
+	  return -reduce(
+	    simd<_I, _Abi>(__private_init, __wrapper_bitcast<_I>(__data(__k))));
+	else
+	  return -reduce(__bit_cast<rebind_simd_t<_I, simd<_Tp, _Abi>>>(
+	    simd<_Tp, _Abi>(__private_init, __data(__k))));
+      }
+
+    // }}}
+    // _S_find_first_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_first_set(simd_mask<_Tp, _Abi> __k)
+      {
+	return std::__countr_zero(
+	  _SuperImpl::_S_to_bits(__data(__k))._M_to_bits());
+      }
+
+    // }}}
+    // _S_find_last_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_last_set(simd_mask<_Tp, _Abi> __k)
+      {
+	return std::__bit_width(
+	  _SuperImpl::_S_to_bits(__data(__k))._M_to_bits()) - 1;
+      }
+
+    // }}}
+  };
+
+//}}}1
+_GLIBCXX_SIMD_END_NAMESPACE
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_ABIS_H_
+
+// vim: foldmethod=3Dmarker foldmarker=3D{{{,}}} sw=3D2 noet ts=3D8 sts=3D=
2 tw=3D80
diff --git a/libstdc++-v3/include/experimental/bits/simd_converter.h b/libs=
tdc++-v3/include/experimental/bits/simd_converter.h
new file mode 100644
index 00000000000..dc4598743f9
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_converter.h
@@ -0,0 +1,354 @@
+// Generic simd conversions -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_CONVERTER_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_CONVERTER_H_
+
+#if __cplusplus >=3D 201703L
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+// _SimdConverter scalar -> scalar {{{
+template <typename _From, typename _To>
+  struct _SimdConverter<_From, simd_abi::scalar, _To, simd_abi::scalar,
+			enable_if_t<!is_same_v<_From, _To>>>
+  {
+    _GLIBCXX_SIMD_INTRINSIC constexpr _To operator()(_From __a) const noex=
cept
+    { return static_cast<_To>(__a); }
+  };
+
+// }}}
+// _SimdConverter scalar -> "native" {{{
+template <typename _From, typename _To, typename _Abi>
+  struct _SimdConverter<_From, simd_abi::scalar, _To, _Abi,
+			enable_if_t<!is_same_v<_Abi, simd_abi::scalar>>>
+  {
+    using _Ret =3D typename _Abi::template __traits<_To>::_SimdMember;
+
+    template <typename... _More>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _Ret
+      operator()(_From __a, _More... __more) const noexcept
+      {
+	static_assert(sizeof...(_More) + 1 =3D=3D _Abi::template _S_size<_To>);
+	static_assert(conjunction_v<is_same<_From, _More>...>);
+	return __make_vector<_To>(__a, __more...);
+      }
+  };
+
+// }}}
+// _SimdConverter "native 1" -> "native 2" {{{
+template <typename _From, typename _To, typename _AFrom, typename _ATo>
+  struct _SimdConverter<
+    _From, _AFrom, _To, _ATo,
+    enable_if_t<!disjunction_v<
+      __is_fixed_size_abi<_AFrom>, __is_fixed_size_abi<_ATo>,
+      is_same<_AFrom, simd_abi::scalar>, is_same<_ATo, simd_abi::scalar>,
+      conjunction<is_same<_From, _To>, is_same<_AFrom, _ATo>>>>>
+  {
+    using _Arg =3D typename _AFrom::template __traits<_From>::_SimdMember;
+    using _Ret =3D typename _ATo::template __traits<_To>::_SimdMember;
+    using _V =3D __vector_type_t<_To, simd_size_v<_To, _ATo>>;
+
+    template <typename... _More>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _Ret
+      operator()(_Arg __a, _More... __more) const noexcept
+      { return __vector_convert<_V>(__a, __more...); }
+  };
+
+// }}}
+// _SimdConverter scalar -> fixed_size<1> {{{1
+template <typename _From, typename _To>
+  struct _SimdConverter<_From, simd_abi::scalar, _To, simd_abi::fixed_size=
<1>,
+			void>
+  {
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple<_To, simd_abi::scalar>
+    operator()(_From __x) const noexcept
+    { return {static_cast<_To>(__x)}; }
+  };
+
+// _SimdConverter fixed_size<1> -> scalar {{{1
+template <typename _From, typename _To>
+  struct _SimdConverter<_From, simd_abi::fixed_size<1>, _To, simd_abi::sca=
lar,
+			void>
+  {
+    _GLIBCXX_SIMD_INTRINSIC constexpr _To
+    operator()(_SimdTuple<_From, simd_abi::scalar> __x) const noexcept
+    { return {static_cast<_To>(__x.first)}; }
+  };
+
+// _SimdConverter fixed_size<_Np> -> fixed_size<_Np> {{{1
+template <typename _From, typename _To, int _Np>
+  struct _SimdConverter<_From, simd_abi::fixed_size<_Np>, _To,
+			simd_abi::fixed_size<_Np>,
+			enable_if_t<!is_same_v<_From, _To>>>
+  {
+    using _Ret =3D __fixed_size_storage_t<_To, _Np>;
+    using _Arg =3D __fixed_size_storage_t<_From, _Np>;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _Ret
+    operator()(const _Arg& __x) const noexcept
+    {
+      if constexpr (is_same_v<_From, _To>)
+	return __x;
+
+      // special case (optimize) int signedness casts
+      else if constexpr (sizeof(_From) =3D=3D sizeof(_To)
+			 && is_integral_v<_From> && is_integral_v<_To>)
+	return __bit_cast<_Ret>(__x);
+
+      // special case if all ABI tags in _Ret are scalar
+      else if constexpr (__is_scalar_abi<typename _Ret::_FirstAbi>())
+	{
+	  return __call_with_subscripts(
+	    __x, make_index_sequence<_Np>(),
+	    [](auto... __values) constexpr->_Ret {
+	      return __make_simd_tuple<_To, decltype((void) __values,
+						     simd_abi::scalar())...>(
+		static_cast<_To>(__values)...);
+	    });
+	}
+
+      // from one vector to one vector
+      else if constexpr (_Arg::_S_first_size =3D=3D _Ret::_S_first_size)
+	{
+	  _SimdConverter<_From, typename _Arg::_FirstAbi, _To,
+			 typename _Ret::_FirstAbi>
+	    __native_cvt;
+	  if constexpr (_Arg::_S_tuple_size =3D=3D 1)
+	    return {__native_cvt(__x.first)};
+	  else
+	    {
+	      constexpr size_t _NRemain =3D _Np - _Arg::_S_first_size;
+	      _SimdConverter<_From, simd_abi::fixed_size<_NRemain>, _To,
+			     simd_abi::fixed_size<_NRemain>>
+		__remainder_cvt;
+	      return {__native_cvt(__x.first), __remainder_cvt(__x.second)};
+	    }
+	}
+
+      // from one vector to multiple vectors
+      else if constexpr (_Arg::_S_first_size > _Ret::_S_first_size)
+	{
+	  const auto __multiple_return_chunks
+	    =3D __convert_all<__vector_type_t<_To, _Ret::_S_first_size>>(
+	      __x.first);
+	  constexpr auto __converted =3D __multiple_return_chunks.size()
+				       * _Ret::_FirstAbi::template _S_size<_To>;
+	  constexpr auto __remaining =3D _Np - __converted;
+	  if constexpr (_Arg::_S_tuple_size =3D=3D 1 && __remaining =3D=3D 0)
+	    return __to_simd_tuple<_To, _Np>(__multiple_return_chunks);
+	  else if constexpr (_Arg::_S_tuple_size =3D=3D 1)
+	    { // e.g. <int, 3> -> <double, 2, 1> or <short, 7> -> <double, 4, 2,
+	      // 1>
+	      using _RetRem
+		=3D __remove_cvref_t<decltype(__simd_tuple_pop_front<__converted>(
+		  _Ret()))>;
+	      const auto __return_chunks2
+		=3D __convert_all<__vector_type_t<_To, _RetRem::_S_first_size>, 0,
+				__converted>(__x.first);
+	      constexpr auto __converted2
+		=3D __converted
+		  + __return_chunks2.size() * _RetRem::_S_first_size;
+	      if constexpr (__converted2 =3D=3D _Np)
+		return __to_simd_tuple<_To, _Np>(__multiple_return_chunks,
+						 __return_chunks2);
+	      else
+		{
+		  using _RetRem2 =3D __remove_cvref_t<
+		    decltype(__simd_tuple_pop_front<__return_chunks2.size()
+						    * _RetRem::_S_first_size>(
+		      _RetRem()))>;
+		  const auto __return_chunks3 =3D __convert_all<
+		    __vector_type_t<_To, _RetRem2::_S_first_size>, 0,
+		    __converted2>(__x.first);
+		  constexpr auto __converted3
+		    =3D __converted2
+		      + __return_chunks3.size() * _RetRem2::_S_first_size;
+		  if constexpr (__converted3 =3D=3D _Np)
+		    return __to_simd_tuple<_To, _Np>(__multiple_return_chunks,
+						     __return_chunks2,
+						     __return_chunks3);
+		  else
+		    {
+		      using _RetRem3
+			=3D __remove_cvref_t<decltype(__simd_tuple_pop_front<
+						    __return_chunks3.size()
+						    * _RetRem2::_S_first_size>(
+			  _RetRem2()))>;
+		      const auto __return_chunks4 =3D __convert_all<
+			__vector_type_t<_To, _RetRem3::_S_first_size>, 0,
+			__converted3>(__x.first);
+		      constexpr auto __converted4
+			=3D __converted3
+			  + __return_chunks4.size() * _RetRem3::_S_first_size;
+		      if constexpr (__converted4 =3D=3D _Np)
+			return __to_simd_tuple<_To, _Np>(
+			  __multiple_return_chunks, __return_chunks2,
+			  __return_chunks3, __return_chunks4);
+		      else
+			__assert_unreachable<_To>();
+		    }
+		}
+	    }
+	  else
+	    {
+	      constexpr size_t _NRemain =3D _Np - _Arg::_S_first_size;
+	      _SimdConverter<_From, simd_abi::fixed_size<_NRemain>, _To,
+			     simd_abi::fixed_size<_NRemain>>
+		__remainder_cvt;
+	      return __simd_tuple_concat(
+		__to_simd_tuple<_To, _Arg::_S_first_size>(
+		  __multiple_return_chunks),
+		__remainder_cvt(__x.second));
+	    }
+	}
+
+      // from multiple vectors to one vector
+      // _Arg::_S_first_size < _Ret::_S_first_size
+      // a) heterogeneous input at the end of the tuple (possible with par=
tial
+      //    native registers in _Ret)
+      else if constexpr (_Ret::_S_tuple_size =3D=3D 1
+			 && _Np % _Arg::_S_first_size !=3D 0)
+	{
+	  static_assert(_Ret::_FirstAbi::template _S_is_partial<_To>);
+	  return _Ret{__generate_from_n_evaluations<
+	    _Np, typename _VectorTraits<typename _Ret::_FirstType>::type>(
+	    [&](auto __i) { return static_cast<_To>(__x[__i]); })};
+	}
+      else
+	{
+	  static_assert(_Arg::_S_tuple_size > 1);
+	  constexpr auto __n
+	    =3D __div_roundup(_Ret::_S_first_size, _Arg::_S_first_size);
+	  return __call_with_n_evaluations<__n>(
+	    [&__x](auto... __uncvted) {
+	      // assuming _Arg Abi tags for all __i are _Arg::_FirstAbi
+	      _SimdConverter<_From, typename _Arg::_FirstAbi, _To,
+			     typename _Ret::_FirstAbi>
+		__native_cvt;
+	      if constexpr (_Ret::_S_tuple_size =3D=3D 1)
+		return _Ret{__native_cvt(__uncvted...)};
+	      else
+		return _Ret{
+		  __native_cvt(__uncvted...),
+		  _SimdConverter<
+		    _From, simd_abi::fixed_size<_Np - _Ret::_S_first_size>, _To,
+		    simd_abi::fixed_size<_Np - _Ret::_S_first_size>>()(
+		    __simd_tuple_pop_front<_Ret::_S_first_size>(__x))};
+	    },
+	    [&__x](auto __i) { return __get_tuple_at<__i>(__x); });
+	}
+    }
+  };
+
+// _SimdConverter "native" -> fixed_size<_Np> {{{1
+// i.e. 1 register to ? registers
+template <typename _From, typename _Ap, typename _To, int _Np>
+  struct _SimdConverter<_From, _Ap, _To, simd_abi::fixed_size<_Np>,
+			enable_if_t<!__is_fixed_size_abi_v<_Ap>>>
+  {
+    static_assert(
+      _Np =3D=3D simd_size_v<_From, _Ap>,
+      "_SimdConverter to fixed_size only works for equal element counts");
+
+    using _Ret =3D __fixed_size_storage_t<_To, _Np>;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _Ret
+    operator()(typename _SimdTraits<_From, _Ap>::_SimdMember __x) const no=
except
+    {
+      if constexpr (_Ret::_S_tuple_size =3D=3D 1)
+	return {__vector_convert<typename _Ret::_FirstType::_BuiltinType>(__x)};
+      else
+	{
+	  using _FixedNp =3D simd_abi::fixed_size<_Np>;
+	  _SimdConverter<_From, _FixedNp, _To, _FixedNp> __fixed_cvt;
+	  using _FromFixedStorage =3D __fixed_size_storage_t<_From, _Np>;
+	  if constexpr (_FromFixedStorage::_S_tuple_size =3D=3D 1)
+	    return __fixed_cvt(_FromFixedStorage{__x});
+	  else if constexpr (_FromFixedStorage::_S_tuple_size =3D=3D 2)
+	    {
+	      _FromFixedStorage __tmp;
+	      static_assert(sizeof(__tmp) <=3D sizeof(__x));
+	      __builtin_memcpy(&__tmp.first, &__x, sizeof(__tmp.first));
+	      __builtin_memcpy(&__tmp.second.first,
+			       reinterpret_cast<const char*>(&__x)
+				 + sizeof(__tmp.first),
+			       sizeof(__tmp.second.first));
+	      return __fixed_cvt(__tmp);
+	    }
+	  else
+	    __assert_unreachable<_From>();
+	}
+    }
+  };
+
+// _SimdConverter fixed_size<_Np> -> "native" {{{1
+// i.e. ? register to 1 registers
+template <typename _From, int _Np, typename _To, typename _Ap>
+  struct _SimdConverter<_From, simd_abi::fixed_size<_Np>, _To, _Ap,
+			enable_if_t<!__is_fixed_size_abi_v<_Ap>>>
+  {
+    static_assert(
+      _Np =3D=3D simd_size_v<_To, _Ap>,
+      "_SimdConverter to fixed_size only works for equal element counts");
+
+    using _Arg =3D __fixed_size_storage_t<_From, _Np>;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr
+      typename _SimdTraits<_To, _Ap>::_SimdMember
+      operator()(_Arg __x) const noexcept
+    {
+      if constexpr (_Arg::_S_tuple_size =3D=3D 1)
+	return __vector_convert<__vector_type_t<_To, _Np>>(__x.first);
+      else if constexpr (_Arg::_S_is_homogeneous)
+	return __call_with_n_evaluations<_Arg::_S_tuple_size>(
+	  [](auto... __members) {
+	    if constexpr ((is_convertible_v<decltype(__members), _To> && ...))
+	      return __vector_type_t<_To, _Np>{static_cast<_To>(__members)...};
+	    else
+	      return __vector_convert<__vector_type_t<_To, _Np>>(__members...);
+	  },
+	  [&](auto __i) { return __get_tuple_at<__i>(__x); });
+      else if constexpr (__fixed_size_storage_t<_To, _Np>::_S_tuple_size =
=3D=3D 1)
+	{
+	  _SimdConverter<_From, simd_abi::fixed_size<_Np>, _To,
+			 simd_abi::fixed_size<_Np>>
+	    __fixed_cvt;
+	  return __fixed_cvt(__x).first;
+	}
+      else
+	{
+	  const _SimdWrapper<_From, _Np> __xv
+	    =3D __generate_from_n_evaluations<_Np, __vector_type_t<_From, _Np>>(
+	      [&](auto __i) { return __x[__i]; });
+	  return __vector_convert<__vector_type_t<_To, _Np>>(__xv);
+	}
+    }
+  };
+
+// }}}1
+_GLIBCXX_SIMD_END_NAMESPACE
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_CONVERTER_H_
+
+// vim: foldmethod=3Dmarker sw=3D2 noet ts=3D8 sts=3D2 tw=3D80
diff --git a/libstdc++-v3/include/experimental/bits/simd_detail.h b/libstdc=
++-v3/include/experimental/bits/simd_detail.h
new file mode 100644
index 00000000000..a49a9d88b7f
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_detail.h
@@ -0,0 +1,306 @@
+// Internal macros for the simd implementation -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_DETAIL_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_DETAIL_H_
+
+#if __cplusplus >=3D 201703L
+
+#include <cstddef>
+#include <cstdint>
+
+
+#define _GLIBCXX_SIMD_BEGIN_NAMESPACE                                     =
     \
+  namespace std _GLIBCXX_VISIBILITY(default)                              =
     \
+  {                                                                       =
     \
+    _GLIBCXX_BEGIN_NAMESPACE_VERSION                                      =
     \
+      namespace experimental {                                            =
     \
+      inline namespace parallelism_v2 {
+#define _GLIBCXX_SIMD_END_NAMESPACE                                       =
     \
+  }                                                                       =
     \
+  }                                                                       =
     \
+  _GLIBCXX_END_NAMESPACE_VERSION                                          =
     \
+  }
+
+// ISA extension detection. The following defines all the _GLIBCXX_SIMD_HA=
VE_XXX
+// macros ARM{{{
+#if defined __ARM_NEON
+#define _GLIBCXX_SIMD_HAVE_NEON 1
+#else
+#define _GLIBCXX_SIMD_HAVE_NEON 0
+#endif
+#if defined __ARM_NEON && (__ARM_ARCH >=3D 8 || defined __aarch64__)
+#define _GLIBCXX_SIMD_HAVE_NEON_A32 1
+#else
+#define _GLIBCXX_SIMD_HAVE_NEON_A32 0
+#endif
+#if defined __ARM_NEON && defined __aarch64__
+#define _GLIBCXX_SIMD_HAVE_NEON_A64 1
+#else
+#define _GLIBCXX_SIMD_HAVE_NEON_A64 0
+#endif
+//}}}
+// x86{{{
+#ifdef __MMX__
+#define _GLIBCXX_SIMD_HAVE_MMX 1
+#else
+#define _GLIBCXX_SIMD_HAVE_MMX 0
+#endif
+#if defined __SSE__ || defined __x86_64__
+#define _GLIBCXX_SIMD_HAVE_SSE 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSE 0
+#endif
+#if defined __SSE2__ || defined __x86_64__
+#define _GLIBCXX_SIMD_HAVE_SSE2 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSE2 0
+#endif
+#ifdef __SSE3__
+#define _GLIBCXX_SIMD_HAVE_SSE3 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSE3 0
+#endif
+#ifdef __SSSE3__
+#define _GLIBCXX_SIMD_HAVE_SSSE3 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSSE3 0
+#endif
+#ifdef __SSE4_1__
+#define _GLIBCXX_SIMD_HAVE_SSE4_1 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSE4_1 0
+#endif
+#ifdef __SSE4_2__
+#define _GLIBCXX_SIMD_HAVE_SSE4_2 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSE4_2 0
+#endif
+#ifdef __XOP__
+#define _GLIBCXX_SIMD_HAVE_XOP 1
+#else
+#define _GLIBCXX_SIMD_HAVE_XOP 0
+#endif
+#ifdef __AVX__
+#define _GLIBCXX_SIMD_HAVE_AVX 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX 0
+#endif
+#ifdef __AVX2__
+#define _GLIBCXX_SIMD_HAVE_AVX2 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX2 0
+#endif
+#ifdef __BMI__
+#define _GLIBCXX_SIMD_HAVE_BMI1 1
+#else
+#define _GLIBCXX_SIMD_HAVE_BMI1 0
+#endif
+#ifdef __BMI2__
+#define _GLIBCXX_SIMD_HAVE_BMI2 1
+#else
+#define _GLIBCXX_SIMD_HAVE_BMI2 0
+#endif
+#ifdef __LZCNT__
+#define _GLIBCXX_SIMD_HAVE_LZCNT 1
+#else
+#define _GLIBCXX_SIMD_HAVE_LZCNT 0
+#endif
+#ifdef __SSE4A__
+#define _GLIBCXX_SIMD_HAVE_SSE4A 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSE4A 0
+#endif
+#ifdef __FMA__
+#define _GLIBCXX_SIMD_HAVE_FMA 1
+#else
+#define _GLIBCXX_SIMD_HAVE_FMA 0
+#endif
+#ifdef __FMA4__
+#define _GLIBCXX_SIMD_HAVE_FMA4 1
+#else
+#define _GLIBCXX_SIMD_HAVE_FMA4 0
+#endif
+#ifdef __F16C__
+#define _GLIBCXX_SIMD_HAVE_F16C 1
+#else
+#define _GLIBCXX_SIMD_HAVE_F16C 0
+#endif
+#ifdef __POPCNT__
+#define _GLIBCXX_SIMD_HAVE_POPCNT 1
+#else
+#define _GLIBCXX_SIMD_HAVE_POPCNT 0
+#endif
+#ifdef __AVX512F__
+#define _GLIBCXX_SIMD_HAVE_AVX512F 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX512F 0
+#endif
+#ifdef __AVX512DQ__
+#define _GLIBCXX_SIMD_HAVE_AVX512DQ 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX512DQ 0
+#endif
+#ifdef __AVX512VL__
+#define _GLIBCXX_SIMD_HAVE_AVX512VL 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX512VL 0
+#endif
+#ifdef __AVX512BW__
+#define _GLIBCXX_SIMD_HAVE_AVX512BW 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX512BW 0
+#endif
+
+#if _GLIBCXX_SIMD_HAVE_SSE
+#define _GLIBCXX_SIMD_HAVE_SSE_ABI 1
+#else
+#define _GLIBCXX_SIMD_HAVE_SSE_ABI 0
+#endif
+#if _GLIBCXX_SIMD_HAVE_SSE2
+#define _GLIBCXX_SIMD_HAVE_FULL_SSE_ABI 1
+#else
+#define _GLIBCXX_SIMD_HAVE_FULL_SSE_ABI 0
+#endif
+
+#if _GLIBCXX_SIMD_HAVE_AVX
+#define _GLIBCXX_SIMD_HAVE_AVX_ABI 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX_ABI 0
+#endif
+#if _GLIBCXX_SIMD_HAVE_AVX2
+#define _GLIBCXX_SIMD_HAVE_FULL_AVX_ABI 1
+#else
+#define _GLIBCXX_SIMD_HAVE_FULL_AVX_ABI 0
+#endif
+
+#if _GLIBCXX_SIMD_HAVE_AVX512F
+#define _GLIBCXX_SIMD_HAVE_AVX512_ABI 1
+#else
+#define _GLIBCXX_SIMD_HAVE_AVX512_ABI 0
+#endif
+#if _GLIBCXX_SIMD_HAVE_AVX512BW
+#define _GLIBCXX_SIMD_HAVE_FULL_AVX512_ABI 1
+#else
+#define _GLIBCXX_SIMD_HAVE_FULL_AVX512_ABI 0
+#endif
+
+#if defined __x86_64__ && !_GLIBCXX_SIMD_HAVE_SSE2
+#error "Use of SSE2 is required on AMD64"
+#endif
+//}}}
+
+#ifdef __clang__
+#define _GLIBCXX_SIMD_NORMAL_MATH
+#else
+#define _GLIBCXX_SIMD_NORMAL_MATH                                         =
     \
+  [[__gnu__::__optimize__("finite-math-only,no-signed-zeros")]]
+#endif
+#define _GLIBCXX_SIMD_NEVER_INLINE [[__gnu__::__noinline__]]
+#define _GLIBCXX_SIMD_INTRINSIC                                           =
     \
+  [[__gnu__::__always_inline__, __gnu__::__artificial__]] inline
+#define _GLIBCXX_SIMD_ALWAYS_INLINE [[__gnu__::__always_inline__]] inline
+#define _GLIBCXX_SIMD_IS_UNLIKELY(__x) __builtin_expect(__x, 0)
+#define _GLIBCXX_SIMD_IS_LIKELY(__x) __builtin_expect(__x, 1)
+
+#if defined __STRICT_ANSI__ && __STRICT_ANSI__
+#define _GLIBCXX_SIMD_CONSTEXPR
+#define _GLIBCXX_SIMD_USE_CONSTEXPR_API const
+#else
+#define _GLIBCXX_SIMD_CONSTEXPR constexpr
+#define _GLIBCXX_SIMD_USE_CONSTEXPR_API constexpr
+#endif
+
+#if defined __clang__
+#define _GLIBCXX_SIMD_USE_CONSTEXPR const
+#else
+#define _GLIBCXX_SIMD_USE_CONSTEXPR constexpr
+#endif
+
+#define _GLIBCXX_SIMD_LIST_BINARY(__macro) __macro(|) __macro(&) __macro(^)
+#define _GLIBCXX_SIMD_LIST_SHIFTS(__macro) __macro(<<) __macro(>>)
+#define _GLIBCXX_SIMD_LIST_ARITHMETICS(__macro)                           =
     \
+  __macro(+) __macro(-) __macro(*) __macro(/) __macro(%)
+
+#define _GLIBCXX_SIMD_ALL_BINARY(__macro)                                 =
     \
+  _GLIBCXX_SIMD_LIST_BINARY(__macro) static_assert(true)
+#define _GLIBCXX_SIMD_ALL_SHIFTS(__macro)                                 =
     \
+  _GLIBCXX_SIMD_LIST_SHIFTS(__macro) static_assert(true)
+#define _GLIBCXX_SIMD_ALL_ARITHMETICS(__macro)                            =
     \
+  _GLIBCXX_SIMD_LIST_ARITHMETICS(__macro) static_assert(true)
+
+#ifdef _GLIBCXX_SIMD_NO_ALWAYS_INLINE
+#undef _GLIBCXX_SIMD_ALWAYS_INLINE
+#define _GLIBCXX_SIMD_ALWAYS_INLINE inline
+#undef _GLIBCXX_SIMD_INTRINSIC
+#define _GLIBCXX_SIMD_INTRINSIC inline
+#endif
+
+#if _GLIBCXX_SIMD_HAVE_SSE || _GLIBCXX_SIMD_HAVE_MMX
+#define _GLIBCXX_SIMD_X86INTRIN 1
+#else
+#define _GLIBCXX_SIMD_X86INTRIN 0
+#endif
+
+// workaround macros {{{
+// use aliasing loads to help GCC understand the data accesses better
+// This also seems to hide a miscompilation on swap(x[i], x[i + 1]) with
+// fixed_size_simd<float, 16> x.
+#define _GLIBCXX_SIMD_USE_ALIASING_LOADS 1
+
+// vector conversions on x86 not optimized:
+#if _GLIBCXX_SIMD_X86INTRIN
+#define _GLIBCXX_SIMD_WORKAROUND_PR85048 1
+#endif
+
+// integer division not optimized
+#define _GLIBCXX_SIMD_WORKAROUND_PR90993 1
+
+// very bad codegen for extraction and concatenation of 128/256 "subregist=
ers"
+// with sizeof(element type) < 8: https://godbolt.org/g/mqUsgM
+#if _GLIBCXX_SIMD_X86INTRIN
+#define _GLIBCXX_SIMD_WORKAROUND_XXX_1 1
+#endif
+
+// bad codegen for 8 Byte memcpy to __vector_type_t<char, 16>
+#define _GLIBCXX_SIMD_WORKAROUND_PR90424 1
+
+// bad codegen for zero-extend using simple concat(__x, 0)
+#if _GLIBCXX_SIMD_X86INTRIN
+#define _GLIBCXX_SIMD_WORKAROUND_XXX_3 1
+#endif
+
+// https://github.com/cplusplus/parallelism-ts/issues/65 (incorrect return=
 type
+// of static_simd_cast)
+#define _GLIBCXX_SIMD_FIX_P2TS_ISSUE65 1
+
+// https://github.com/cplusplus/parallelism-ts/issues/66 (incorrect SFINAE
+// constraint on (static)_simd_cast)
+#define _GLIBCXX_SIMD_FIX_P2TS_ISSUE66 1
+// }}}
+
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_DETAIL_H_
+
+// vim: foldmethod=3Dmarker
diff --git a/libstdc++-v3/include/experimental/bits/simd_fixed_size.h b/lib=
stdc++-v3/include/experimental/bits/simd_fixed_size.h
new file mode 100644
index 00000000000..fba8c7e466e
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_fixed_size.h
@@ -0,0 +1,2066 @@
+// Simd fixed_size ABI specific implementations -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+/*
+ * The fixed_size ABI gives the following guarantees:
+ *  - simd objects are passed via the stack
+ *  - memory layout of `simd<_Tp, _Np>` is equivalent to `array<_Tp, _Np>`
+ *  - alignment of `simd<_Tp, _Np>` is `_Np * sizeof(_Tp)` if _Np is __a
+ *    power-of-2 value, otherwise `std::__bit_ceil(_Np * sizeof(_Tp))` (No=
te:
+ *    if the alignment were to exceed the system/compiler maximum, it is b=
ounded
+ *    to that maximum)
+ *  - simd_mask objects are passed like bitset<_Np>
+ *  - memory layout of `simd_mask<_Tp, _Np>` is equivalent to `bitset<_Np>`
+ *  - alignment of `simd_mask<_Tp, _Np>` is equal to the alignment of
+ *    `bitset<_Np>`
+ */
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_FIXED_SIZE_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_FIXED_SIZE_H_
+
+#if __cplusplus >=3D 201703L
+
+#include <array>
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+
+// __simd_tuple_element {{{
+template <size_t _I, typename _Tp>
+  struct __simd_tuple_element;
+
+template <typename _Tp, typename _A0, typename... _As>
+  struct __simd_tuple_element<0, _SimdTuple<_Tp, _A0, _As...>>
+  { using type =3D simd<_Tp, _A0>; };
+
+template <size_t _I, typename _Tp, typename _A0, typename... _As>
+  struct __simd_tuple_element<_I, _SimdTuple<_Tp, _A0, _As...>>
+  {
+    using type =3D
+      typename __simd_tuple_element<_I - 1, _SimdTuple<_Tp, _As...>>::type;
+  };
+
+template <size_t _I, typename _Tp>
+  using __simd_tuple_element_t =3D typename __simd_tuple_element<_I, _Tp>:=
:type;
+
+// }}}
+// __simd_tuple_concat {{{
+
+template <typename _Tp, typename... _A0s, typename... _A1s>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple<_Tp, _A0s..., _A1s...>
+  __simd_tuple_concat(const _SimdTuple<_Tp, _A0s...>& __left,
+		      const _SimdTuple<_Tp, _A1s...>& __right)
+  {
+    if constexpr (sizeof...(_A0s) =3D=3D 0)
+      return __right;
+    else if constexpr (sizeof...(_A1s) =3D=3D 0)
+      return __left;
+    else
+      return {__left.first, __simd_tuple_concat(__left.second, __right)};
+  }
+
+template <typename _Tp, typename _A10, typename... _A1s>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple<_Tp, simd_abi::scalar, _A10,
+					       _A1s...>
+  __simd_tuple_concat(const _Tp& __left,
+		      const _SimdTuple<_Tp, _A10, _A1s...>& __right)
+  { return {__left, __right}; }
+
+// }}}
+// __simd_tuple_pop_front {{{
+// Returns the next _SimdTuple in __x that has _Np elements less.
+// Precondition: _Np must match the number of elements in __first (recursi=
vely)
+template <size_t _Np, typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr decltype(auto)
+  __simd_tuple_pop_front(_Tp&& __x)
+  {
+    if constexpr (_Np =3D=3D 0)
+      return static_cast<_Tp&&>(__x);
+    else
+      {
+	using _Up =3D __remove_cvref_t<_Tp>;
+	static_assert(_Np >=3D _Up::_S_first_size);
+	return __simd_tuple_pop_front<_Np - _Up::_S_first_size>(__x.second);
+      }
+  }
+
+// }}}
+// __get_simd_at<_Np> {{{1
+struct __as_simd {};
+
+struct __as_simd_tuple {};
+
+template <typename _Tp, typename _A0, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr simd<_Tp, _A0>
+  __simd_tuple_get_impl(__as_simd, const _SimdTuple<_Tp, _A0, _Abis...>& _=
_t,
+			_SizeConstant<0>)
+  { return {__private_init, __t.first}; }
+
+template <typename _Tp, typename _A0, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr const auto&
+  __simd_tuple_get_impl(__as_simd_tuple,
+			const _SimdTuple<_Tp, _A0, _Abis...>& __t,
+			_SizeConstant<0>)
+  { return __t.first; }
+
+template <typename _Tp, typename _A0, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+  __simd_tuple_get_impl(__as_simd_tuple, _SimdTuple<_Tp, _A0, _Abis...>& _=
_t,
+			_SizeConstant<0>)
+  { return __t.first; }
+
+template <typename _R, size_t _Np, typename _Tp, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __simd_tuple_get_impl(_R, const _SimdTuple<_Tp, _Abis...>& __t,
+			_SizeConstant<_Np>)
+  { return __simd_tuple_get_impl(_R(), __t.second, _SizeConstant<_Np - 1>(=
)); }
+
+template <size_t _Np, typename _Tp, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+  __simd_tuple_get_impl(__as_simd_tuple, _SimdTuple<_Tp, _Abis...>& __t,
+			_SizeConstant<_Np>)
+  {
+    return __simd_tuple_get_impl(__as_simd_tuple(), __t.second,
+				 _SizeConstant<_Np - 1>());
+  }
+
+template <size_t _Np, typename _Tp, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __get_simd_at(const _SimdTuple<_Tp, _Abis...>& __t)
+  { return __simd_tuple_get_impl(__as_simd(), __t, _SizeConstant<_Np>()); }
+
+// }}}
+// __get_tuple_at<_Np> {{{
+template <size_t _Np, typename _Tp, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto
+  __get_tuple_at(const _SimdTuple<_Tp, _Abis...>& __t)
+  {
+    return __simd_tuple_get_impl(__as_simd_tuple(), __t, _SizeConstant<_Np=
>());
+  }
+
+template <size_t _Np, typename _Tp, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+  __get_tuple_at(_SimdTuple<_Tp, _Abis...>& __t)
+  {
+    return __simd_tuple_get_impl(__as_simd_tuple(), __t, _SizeConstant<_Np=
>());
+  }
+
+// __tuple_element_meta {{{1
+template <typename _Tp, typename _Abi, size_t _Offset>
+  struct __tuple_element_meta : public _Abi::_SimdImpl
+  {
+    static_assert(is_same_v<typename _Abi::_SimdImpl::abi_type,
+			    _Abi>); // this fails e.g. when _SimdImpl is an
+				    // alias for _SimdImplBuiltin<_DifferentAbi>
+    using value_type =3D _Tp;
+    using abi_type =3D _Abi;
+    using _Traits =3D _SimdTraits<_Tp, _Abi>;
+    using _MaskImpl =3D typename _Abi::_MaskImpl;
+    using _MaskMember =3D typename _Traits::_MaskMember;
+    using simd_type =3D simd<_Tp, _Abi>;
+    static constexpr size_t _S_offset =3D _Offset;
+    static constexpr size_t _S_size() { return simd_size<_Tp, _Abi>::value=
; }
+    static constexpr _MaskImpl _S_mask_impl =3D {};
+
+    template <size_t _Np, bool _Sanitized>
+      _GLIBCXX_SIMD_INTRINSIC static auto
+      _S_submask(_BitMask<_Np, _Sanitized> __bits)
+      { return __bits.template _M_extract<_Offset, _S_size()>(); }
+
+    template <size_t _Np, bool _Sanitized>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember
+      _S_make_mask(_BitMask<_Np, _Sanitized> __bits)
+      {
+	return _MaskImpl::template _S_convert<_Tp>(
+	  __bits.template _M_extract<_Offset, _S_size()>()._M_sanitized());
+      }
+
+    _GLIBCXX_SIMD_INTRINSIC static _ULLong
+    _S_mask_to_shifted_ullong(_MaskMember __k)
+    { return _MaskImpl::_S_to_bits(__k).to_ullong() << _Offset; }
+  };
+
+template <size_t _Offset, typename _Tp, typename _Abi, typename... _As>
+  __tuple_element_meta<_Tp, _Abi, _Offset>
+  __make_meta(const _SimdTuple<_Tp, _Abi, _As...>&)
+  { return {}; }
+
+// }}}1
+// _WithOffset wrapper class {{{
+template <size_t _Offset, typename _Base>
+  struct _WithOffset : public _Base
+  {
+    static inline constexpr size_t _S_offset =3D _Offset;
+
+    _GLIBCXX_SIMD_INTRINSIC char* _M_as_charptr()
+    {
+      return reinterpret_cast<char*>(this)
+	     + _S_offset * sizeof(typename _Base::value_type);
+    }
+
+    _GLIBCXX_SIMD_INTRINSIC const char* _M_as_charptr() const
+    {
+      return reinterpret_cast<const char*>(this)
+	     + _S_offset * sizeof(typename _Base::value_type);
+    }
+  };
+
+// make _WithOffset<_WithOffset> ill-formed to use:
+template <size_t _O0, size_t _O1, typename _Base>
+  struct _WithOffset<_O0, _WithOffset<_O1, _Base>> {};
+
+template <size_t _Offset, typename _Tp>
+  decltype(auto)
+  __add_offset(_Tp& __base)
+  { return static_cast<_WithOffset<_Offset, __remove_cvref_t<_Tp>>&>(__bas=
e); }
+
+template <size_t _Offset, typename _Tp>
+  decltype(auto)
+  __add_offset(const _Tp& __base)
+  {
+    return static_cast<const _WithOffset<_Offset, __remove_cvref_t<_Tp>>&>(
+      __base);
+  }
+
+template <size_t _Offset, size_t _ExistingOffset, typename _Tp>
+  decltype(auto)
+  __add_offset(_WithOffset<_ExistingOffset, _Tp>& __base)
+  {
+    return static_cast<_WithOffset<_Offset + _ExistingOffset, _Tp>&>(
+      static_cast<_Tp&>(__base));
+  }
+
+template <size_t _Offset, size_t _ExistingOffset, typename _Tp>
+  decltype(auto)
+  __add_offset(const _WithOffset<_ExistingOffset, _Tp>& __base)
+  {
+    return static_cast<const _WithOffset<_Offset + _ExistingOffset, _Tp>&>(
+      static_cast<const _Tp&>(__base));
+  }
+
+template <typename _Tp>
+  constexpr inline size_t __offset =3D 0;
+
+template <size_t _Offset, typename _Tp>
+  constexpr inline size_t __offset<_WithOffset<_Offset, _Tp>>
+    =3D _WithOffset<_Offset, _Tp>::_S_offset;
+
+template <typename _Tp>
+  constexpr inline size_t __offset<const _Tp> =3D __offset<_Tp>;
+
+template <typename _Tp>
+  constexpr inline size_t __offset<_Tp&> =3D __offset<_Tp>;
+
+template <typename _Tp>
+  constexpr inline size_t __offset<_Tp&&> =3D __offset<_Tp>;
+
+// }}}
+// _SimdTuple specializations {{{1
+// empty {{{2
+template <typename _Tp>
+  struct _SimdTuple<_Tp>
+  {
+    using value_type =3D _Tp;
+    static constexpr size_t _S_tuple_size =3D 0;
+    static constexpr size_t _S_size() { return 0; }
+  };
+
+// _SimdTupleData {{{2
+template <typename _FirstType, typename _SecondType>
+  struct _SimdTupleData
+  {
+    _FirstType first;
+    _SecondType second;
+
+    _GLIBCXX_SIMD_INTRINSIC
+    constexpr bool _M_is_constprop() const
+    {
+      if constexpr (is_class_v<_FirstType>)
+	return first._M_is_constprop() && second._M_is_constprop();
+      else
+	return __builtin_constant_p(first) && second._M_is_constprop();
+    }
+  };
+
+template <typename _FirstType, typename _Tp>
+  struct _SimdTupleData<_FirstType, _SimdTuple<_Tp>>
+  {
+    _FirstType first;
+    static constexpr _SimdTuple<_Tp> second =3D {};
+
+    _GLIBCXX_SIMD_INTRINSIC
+    constexpr bool _M_is_constprop() const
+    {
+      if constexpr (is_class_v<_FirstType>)
+	return first._M_is_constprop();
+      else
+	return __builtin_constant_p(first);
+    }
+  };
+
+// 1 or more {{{2
+template <typename _Tp, typename _Abi0, typename... _Abis>
+  struct _SimdTuple<_Tp, _Abi0, _Abis...>
+    : _SimdTupleData<typename _SimdTraits<_Tp, _Abi0>::_SimdMember,
+		     _SimdTuple<_Tp, _Abis...>>
+  {
+    static_assert(!__is_fixed_size_abi_v<_Abi0>);
+    using value_type =3D _Tp;
+    using _FirstType =3D typename _SimdTraits<_Tp, _Abi0>::_SimdMember;
+    using _FirstAbi =3D _Abi0;
+    using _SecondType =3D _SimdTuple<_Tp, _Abis...>;
+    static constexpr size_t _S_tuple_size =3D sizeof...(_Abis) + 1;
+
+    static constexpr size_t _S_size()
+    { return simd_size_v<_Tp, _Abi0> + _SecondType::_S_size(); }
+
+    static constexpr size_t _S_first_size =3D simd_size_v<_Tp, _Abi0>;
+    static constexpr bool _S_is_homogeneous =3D (is_same_v<_Abi0, _Abis> &=
& ...);
+
+    using _Base =3D _SimdTupleData<typename _SimdTraits<_Tp, _Abi0>::_Simd=
Member,
+				 _SimdTuple<_Tp, _Abis...>>;
+    using _Base::first;
+    using _Base::second;
+
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple() =3D default;
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple(const _SimdTuple&) =3D de=
fault;
+    _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple& operator=3D(const _SimdT=
uple&)
+      =3D default;
+
+    template <typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple(_Up&& __x)
+      : _Base{static_cast<_Up&&>(__x)} {}
+
+    template <typename _Up, typename _Up2>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple(_Up&& __x, _Up2&& __y)
+      : _Base{static_cast<_Up&&>(__x), static_cast<_Up2&&>(__y)} {}
+
+    template <typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple(_Up&& __x, _SimdTuple<_=
Tp>)
+      : _Base{static_cast<_Up&&>(__x)} {}
+
+    _GLIBCXX_SIMD_INTRINSIC char* _M_as_charptr()
+    { return reinterpret_cast<char*>(this); }
+
+    _GLIBCXX_SIMD_INTRINSIC const char* _M_as_charptr() const
+    { return reinterpret_cast<const char*>(this); }
+
+    template <size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC constexpr auto& _M_at()
+      {
+	if constexpr (_Np =3D=3D 0)
+	  return first;
+	else
+	  return second.template _M_at<_Np - 1>();
+      }
+
+    template <size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC constexpr const auto& _M_at() const
+      {
+	if constexpr (_Np =3D=3D 0)
+	  return first;
+	else
+	  return second.template _M_at<_Np - 1>();
+      }
+
+    template <size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC constexpr auto _M_simd_at() const
+      {
+	if constexpr (_Np =3D=3D 0)
+	  return simd<_Tp, _Abi0>(__private_init, first);
+	else
+	  return second.template _M_simd_at<_Np - 1>();
+      }
+
+    template <size_t _Offset =3D 0, typename _Fp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdTuple
+      _S_generate(_Fp&& __gen, _SizeConstant<_Offset> =3D {})
+      {
+	auto&& __first =3D __gen(__tuple_element_meta<_Tp, _Abi0, _Offset>());
+	if constexpr (_S_tuple_size =3D=3D 1)
+	  return {__first};
+	else
+	  return {__first,
+		  _SecondType::_S_generate(
+		    static_cast<_Fp&&>(__gen),
+		    _SizeConstant<_Offset + simd_size_v<_Tp, _Abi0>>())};
+      }
+
+    template <size_t _Offset =3D 0, typename _Fp, typename... _More>
+      _GLIBCXX_SIMD_INTRINSIC _SimdTuple
+      _M_apply_wrapped(_Fp&& __fun, const _More&... __more) const
+      {
+	auto&& __first
+	  =3D __fun(__make_meta<_Offset>(*this), first, __more.first...);
+	if constexpr (_S_tuple_size =3D=3D 1)
+	  return {__first};
+	else
+	  return {
+	    __first,
+	    second.template _M_apply_wrapped<_Offset + simd_size_v<_Tp, _Abi0>>(
+	      static_cast<_Fp&&>(__fun), __more.second...)};
+      }
+
+    template <typename _Tup>
+      _GLIBCXX_SIMD_INTRINSIC constexpr decltype(auto)
+      _M_extract_argument(_Tup&& __tup) const
+      {
+	using _TupT =3D typename __remove_cvref_t<_Tup>::value_type;
+	if constexpr (is_same_v<_SimdTuple, __remove_cvref_t<_Tup>>)
+	  return __tup.first;
+	else if (__builtin_is_constant_evaluated())
+	  return __fixed_size_storage_t<_TupT, _S_first_size>::_S_generate([&](
+	    auto __meta) constexpr {
+	    return __meta._S_generator(
+	      [&](auto __i) constexpr { return __tup[__i]; },
+	      static_cast<_TupT*>(nullptr));
+	  });
+	else
+	  return [&]() {
+	    __fixed_size_storage_t<_TupT, _S_first_size> __r;
+	    __builtin_memcpy(__r._M_as_charptr(), __tup._M_as_charptr(),
+			     sizeof(__r));
+	    return __r;
+	  }();
+      }
+
+    template <typename _Tup>
+      _GLIBCXX_SIMD_INTRINSIC constexpr auto&
+      _M_skip_argument(_Tup&& __tup) const
+      {
+	static_assert(_S_tuple_size > 1);
+	using _Up =3D __remove_cvref_t<_Tup>;
+	constexpr size_t __off =3D __offset<_Up>;
+	if constexpr (_S_first_size =3D=3D _Up::_S_first_size && __off =3D=3D 0)
+	  return __tup.second;
+	else if constexpr (_S_first_size > _Up::_S_first_size
+			   && _S_first_size % _Up::_S_first_size =3D=3D 0
+			   && __off =3D=3D 0)
+	  return __simd_tuple_pop_front<_S_first_size>(__tup);
+	else if constexpr (_S_first_size + __off < _Up::_S_first_size)
+	  return __add_offset<_S_first_size>(__tup);
+	else if constexpr (_S_first_size + __off =3D=3D _Up::_S_first_size)
+	  return __tup.second;
+	else
+	  __assert_unreachable<_Tup>();
+      }
+
+    template <size_t _Offset, typename... _More>
+      _GLIBCXX_SIMD_INTRINSIC constexpr void
+      _M_assign_front(const _SimdTuple<_Tp, _Abi0, _More...>& __x) &
+      {
+	static_assert(_Offset =3D=3D 0);
+	first =3D __x.first;
+	if constexpr (sizeof...(_More) > 0)
+	  {
+	    static_assert(sizeof...(_Abis) >=3D sizeof...(_More));
+	    second.template _M_assign_front<0>(__x.second);
+	  }
+      }
+
+    template <size_t _Offset>
+      _GLIBCXX_SIMD_INTRINSIC constexpr void
+      _M_assign_front(const _FirstType& __x) &
+      {
+	static_assert(_Offset =3D=3D 0);
+	first =3D __x;
+      }
+
+    template <size_t _Offset, typename... _As>
+      _GLIBCXX_SIMD_INTRINSIC constexpr void
+      _M_assign_front(const _SimdTuple<_Tp, _As...>& __x) &
+      {
+	__builtin_memcpy(_M_as_charptr() + _Offset * sizeof(value_type),
+			 __x._M_as_charptr(),
+			 sizeof(_Tp) * _SimdTuple<_Tp, _As...>::_S_size());
+      }
+
+    /*
+     * Iterate over the first objects in this _SimdTuple and call __fun fo=
r each
+     * of them. If additional arguments are passed via __more, chunk them =
into
+     * _SimdTuple or __vector_type_t objects of the same number of values.
+     */
+    template <typename _Fp, typename... _More>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _SimdTuple
+      _M_apply_per_chunk(_Fp&& __fun, _More&&... __more) const
+      {
+	if constexpr ((...
+		       || conjunction_v<
+			 is_lvalue_reference<_More>,
+			 negation<is_const<remove_reference_t<_More>>>>) )
+	  {
+	    // need to write back at least one of __more after calling __fun
+	    auto&& __first =3D [&](auto... __args) constexpr
+	    {
+	      auto __r =3D __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), first,
+			       __args...);
+	      [[maybe_unused]] auto&& __ignore_me =3D {(
+		[](auto&& __dst, const auto& __src) {
+		  if constexpr (is_assignable_v<decltype(__dst),
+						decltype(__dst)>)
+		    {
+		      __dst.template _M_assign_front<__offset<decltype(__dst)>>(
+			__src);
+		    }
+		}(static_cast<_More&&>(__more), __args),
+		0)...};
+	      return __r;
+	    }
+	    (_M_extract_argument(__more)...);
+	    if constexpr (_S_tuple_size =3D=3D 1)
+	      return {__first};
+	    else
+	      return {__first,
+		      second._M_apply_per_chunk(static_cast<_Fp&&>(__fun),
+						_M_skip_argument(__more)...)};
+	  }
+	else
+	  {
+	    auto&& __first =3D __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), first,
+				   _M_extract_argument(__more)...);
+	    if constexpr (_S_tuple_size =3D=3D 1)
+	      return {__first};
+	    else
+	      return {__first,
+		      second._M_apply_per_chunk(static_cast<_Fp&&>(__fun),
+						_M_skip_argument(__more)...)};
+	  }
+      }
+
+    template <typename _R =3D _Tp, typename _Fp, typename... _More>
+      _GLIBCXX_SIMD_INTRINSIC auto _M_apply_r(_Fp&& __fun,
+					      const _More&... __more) const
+      {
+	auto&& __first =3D __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), first,
+			       __more.first...);
+	if constexpr (_S_tuple_size =3D=3D 1)
+	  return __first;
+	else
+	  return __simd_tuple_concat<_R>(
+	    __first, second.template _M_apply_r<_R>(static_cast<_Fp&&>(__fun),
+						    __more.second...));
+      }
+
+    template <typename _Fp, typename... _More>
+      _GLIBCXX_SIMD_INTRINSIC constexpr friend _SanitizedBitMask<_S_size()>
+      _M_test(const _Fp& __fun, const _SimdTuple& __x, const _More&... __m=
ore)
+      {
+	const _SanitizedBitMask<_S_first_size> __first
+	  =3D _Abi0::_MaskImpl::_S_to_bits(
+	    __fun(__tuple_element_meta<_Tp, _Abi0, 0>(), __x.first,
+		  __more.first...));
+	if constexpr (_S_tuple_size =3D=3D 1)
+	  return __first;
+	else
+	  return _M_test(__fun, __x.second, __more.second...)
+	    ._M_prepend(__first);
+      }
+
+    template <typename _Up, _Up _I>
+      _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
+      operator[](integral_constant<_Up, _I>) const noexcept
+      {
+	if constexpr (_I < simd_size_v<_Tp, _Abi0>)
+	  return _M_subscript_read(_I);
+	else
+	  return second[integral_constant<_Up, _I - simd_size_v<_Tp, _Abi0>>()];
+      }
+
+    _Tp operator[](size_t __i) const noexcept
+    {
+      if constexpr (_S_tuple_size =3D=3D 1)
+	return _M_subscript_read(__i);
+      else
+	{
+#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
+	  return reinterpret_cast<const __may_alias<_Tp>*>(this)[__i];
+#else
+	  if constexpr (__is_scalar_abi<_Abi0>())
+	    {
+	      const _Tp* ptr =3D &first;
+	      return ptr[__i];
+	    }
+	  else
+	    return __i < simd_size_v<_Tp, _Abi0>
+		     ? _M_subscript_read(__i)
+		     : second[__i - simd_size_v<_Tp, _Abi0>];
+#endif
+	}
+    }
+
+    void _M_set(size_t __i, _Tp __val) noexcept
+    {
+      if constexpr (_S_tuple_size =3D=3D 1)
+	return _M_subscript_write(__i, __val);
+      else
+	{
+#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
+	  reinterpret_cast<__may_alias<_Tp>*>(this)[__i] =3D __val;
+#else
+	  if (__i < simd_size_v<_Tp, _Abi0>)
+	    _M_subscript_write(__i, __val);
+	  else
+	    second._M_set(__i - simd_size_v<_Tp, _Abi0>, __val);
+#endif
+	}
+    }
+
+  private:
+    // _M_subscript_read/_write {{{
+    _Tp _M_subscript_read([[maybe_unused]] size_t __i) const noexcept
+    {
+      if constexpr (__is_vectorizable_v<_FirstType>)
+	return first;
+      else
+	return first[__i];
+    }
+
+    void _M_subscript_write([[maybe_unused]] size_t __i, _Tp __y) noexcept
+    {
+      if constexpr (__is_vectorizable_v<_FirstType>)
+	first =3D __y;
+      else
+	first._M_set(__i, __y);
+    }
+
+    // }}}
+  };
+
+// __make_simd_tuple {{{1
+template <typename _Tp, typename _A0>
+  _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0>
+  __make_simd_tuple(simd<_Tp, _A0> __x0)
+  { return {__data(__x0)}; }
+
+template <typename _Tp, typename _A0, typename... _As>
+  _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0, _As...>
+  __make_simd_tuple(const simd<_Tp, _A0>& __x0, const simd<_Tp, _As>&... _=
_xs)
+  { return {__data(__x0), __make_simd_tuple(__xs...)}; }
+
+template <typename _Tp, typename _A0>
+  _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0>
+  __make_simd_tuple(const typename _SimdTraits<_Tp, _A0>::_SimdMember& __a=
rg0)
+  { return {__arg0}; }
+
+template <typename _Tp, typename _A0, typename _A1, typename... _Abis>
+  _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp, _A0, _A1, _Abis...>
+  __make_simd_tuple(
+    const typename _SimdTraits<_Tp, _A0>::_SimdMember& __arg0,
+    const typename _SimdTraits<_Tp, _A1>::_SimdMember& __arg1,
+    const typename _SimdTraits<_Tp, _Abis>::_SimdMember&... __args)
+  { return {__arg0, __make_simd_tuple<_Tp, _A1, _Abis...>(__arg1, __args..=
=2E)}; }
+
+// __to_simd_tuple {{{1
+template <typename _Tp, size_t _Np, typename _V, size_t _NV, typename... _=
VX>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __fixed_size_storage_t<_Tp, _Np>
+  __to_simd_tuple(const array<_V, _NV>& __from, const _VX... __fromX);
+
+template <typename _Tp, size_t _Np,
+	  size_t _Offset =3D 0, // skip this many elements in __from0
+	  typename _R =3D __fixed_size_storage_t<_Tp, _Np>, typename _V0,
+	  typename _V0VT =3D _VectorTraits<_V0>, typename... _VX>
+  _GLIBCXX_SIMD_INTRINSIC _R constexpr __to_simd_tuple(const _V0 __from0,
+						       const _VX... __fromX)
+  {
+    static_assert(is_same_v<typename _V0VT::value_type, _Tp>);
+    static_assert(_Offset < _V0VT::_S_full_size);
+    using _R0 =3D __vector_type_t<_Tp, _R::_S_first_size>;
+    if constexpr (_R::_S_tuple_size =3D=3D 1)
+      {
+	if constexpr (_Np =3D=3D 1)
+	  return _R{__from0[_Offset]};
+	else if constexpr (_Offset =3D=3D 0 && _V0VT::_S_full_size >=3D _Np)
+	  return _R{__intrin_bitcast<_R0>(__from0)};
+	else if constexpr (_Offset * 2 =3D=3D _V0VT::_S_full_size
+			   && _V0VT::_S_full_size / 2 >=3D _Np)
+	  return _R{__intrin_bitcast<_R0>(__extract_part<1, 2>(__from0))};
+	else if constexpr (_Offset * 4 =3D=3D _V0VT::_S_full_size
+			   && _V0VT::_S_full_size / 4 >=3D _Np)
+	  return _R{__intrin_bitcast<_R0>(__extract_part<1, 4>(__from0))};
+	else
+	  __assert_unreachable<_Tp>();
+      }
+    else
+      {
+	if constexpr (1 =3D=3D _R::_S_first_size)
+	  { // extract one scalar and recurse
+	    if constexpr (_Offset + 1 < _V0VT::_S_full_size)
+	      return _R{__from0[_Offset],
+			__to_simd_tuple<_Tp, _Np - 1, _Offset + 1>(__from0,
+								   __fromX...)};
+	    else
+	      return _R{__from0[_Offset],
+			__to_simd_tuple<_Tp, _Np - 1, 0>(__fromX...)};
+	  }
+
+	// place __from0 into _R::first and recurse for __fromX -> _R::second
+	else if constexpr (_V0VT::_S_full_size =3D=3D _R::_S_first_size
+			   && _Offset =3D=3D 0)
+	  return _R{__from0,
+		    __to_simd_tuple<_Tp, _Np - _R::_S_first_size>(__fromX...)};
+
+	// place lower part of __from0 into _R::first and recurse with _Offset
+	else if constexpr (_V0VT::_S_full_size > _R::_S_first_size
+			   && _Offset =3D=3D 0)
+	  return _R{__intrin_bitcast<_R0>(__from0),
+		    __to_simd_tuple<_Tp, _Np - _R::_S_first_size,
+				    _R::_S_first_size>(__from0, __fromX...)};
+
+	// place lower part of second quarter of __from0 into _R::first and
+	// recurse with _Offset
+	else if constexpr (_Offset * 4 =3D=3D _V0VT::_S_full_size
+			   && _V0VT::_S_full_size >=3D 4 * _R::_S_first_size)
+	  return _R{__intrin_bitcast<_R0>(__extract_part<2, 4>(__from0)),
+		    __to_simd_tuple<_Tp, _Np - _R::_S_first_size,
+				    _Offset + _R::_S_first_size>(__from0,
+								 __fromX...)};
+
+	// place lower half of high half of __from0 into _R::first and recurse
+	// with _Offset
+	else if constexpr (_Offset * 2 =3D=3D _V0VT::_S_full_size
+			   && _V0VT::_S_full_size >=3D 4 * _R::_S_first_size)
+	  return _R{__intrin_bitcast<_R0>(__extract_part<2, 4>(__from0)),
+		    __to_simd_tuple<_Tp, _Np - _R::_S_first_size,
+				    _Offset + _R::_S_first_size>(__from0,
+								 __fromX...)};
+
+	// place high half of __from0 into _R::first and recurse with __fromX
+	else if constexpr (_Offset * 2 =3D=3D _V0VT::_S_full_size
+			   && _V0VT::_S_full_size / 2 >=3D _R::_S_first_size)
+	  return _R{__intrin_bitcast<_R0>(__extract_part<1, 2>(__from0)),
+		    __to_simd_tuple<_Tp, _Np - _R::_S_first_size, 0>(
+		      __fromX...)};
+
+	// ill-formed if some unforseen pattern is needed
+	else
+	  __assert_unreachable<_Tp>();
+      }
+  }
+
+template <typename _Tp, size_t _Np, typename _V, size_t _NV, typename... _=
VX>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __fixed_size_storage_t<_Tp, _Np>
+  __to_simd_tuple(const array<_V, _NV>& __from, const _VX... __fromX)
+  {
+    if constexpr (is_same_v<_Tp, _V>)
+      {
+	static_assert(
+	  sizeof...(_VX) =3D=3D 0,
+	  "An array of scalars must be the last argument to __to_simd_tuple");
+	return __call_with_subscripts(
+	  __from,
+	  make_index_sequence<_NV>(), [&](const auto... __args) constexpr {
+	    return __simd_tuple_concat(
+	      _SimdTuple<_Tp, simd_abi::scalar>{__args}..., _SimdTuple<_Tp>());
+	  });
+      }
+    else
+      return __call_with_subscripts(
+	__from,
+	make_index_sequence<_NV>(), [&](const auto... __args) constexpr {
+	  return __to_simd_tuple<_Tp, _Np>(__args..., __fromX...);
+	});
+  }
+
+template <size_t, typename _Tp>
+  using __to_tuple_helper =3D _Tp;
+
+template <typename _Tp, typename _A0, size_t _NOut, size_t _Np,
+	  size_t... _Indexes>
+  _GLIBCXX_SIMD_INTRINSIC __fixed_size_storage_t<_Tp, _NOut>
+  __to_simd_tuple_impl(index_sequence<_Indexes...>,
+      const array<__vector_type_t<_Tp, simd_size_v<_Tp, _A0>>, _Np>& __arg=
s)
+  {
+    return __make_simd_tuple<_Tp, __to_tuple_helper<_Indexes, _A0>...>(
+      __args[_Indexes]...);
+  }
+
+template <typename _Tp, typename _A0, size_t _NOut, size_t _Np,
+	  typename _R =3D __fixed_size_storage_t<_Tp, _NOut>>
+  _GLIBCXX_SIMD_INTRINSIC _R
+  __to_simd_tuple_sized(
+    const array<__vector_type_t<_Tp, simd_size_v<_Tp, _A0>>, _Np>& __args)
+  {
+    static_assert(_Np * simd_size_v<_Tp, _A0> >=3D _NOut);
+    return __to_simd_tuple_impl<_Tp, _A0, _NOut>(
+      make_index_sequence<_R::_S_tuple_size>(), __args);
+  }
+
+// __optimize_simd_tuple {{{1
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC _SimdTuple<_Tp>
+  __optimize_simd_tuple(const _SimdTuple<_Tp>)
+  { return {}; }
+
+template <typename _Tp, typename _Ap>
+  _GLIBCXX_SIMD_INTRINSIC const _SimdTuple<_Tp, _Ap>&
+  __optimize_simd_tuple(const _SimdTuple<_Tp, _Ap>& __x)
+  { return __x; }
+
+template <typename _Tp, typename _A0, typename _A1, typename... _Abis,
+	  typename _R =3D __fixed_size_storage_t<
+	    _Tp, _SimdTuple<_Tp, _A0, _A1, _Abis...>::_S_size()>>
+  _GLIBCXX_SIMD_INTRINSIC _R
+  __optimize_simd_tuple(const _SimdTuple<_Tp, _A0, _A1, _Abis...>& __x)
+  {
+    using _Tup =3D _SimdTuple<_Tp, _A0, _A1, _Abis...>;
+    if constexpr (is_same_v<_R, _Tup>)
+      return __x;
+    else if constexpr (is_same_v<typename _R::_FirstType,
+				 typename _Tup::_FirstType>)
+      return {__x.first, __optimize_simd_tuple(__x.second)};
+    else if constexpr (__is_scalar_abi<_A0>()
+		       || _A0::template _S_is_partial<_Tp>)
+      return {__generate_from_n_evaluations<_R::_S_first_size,
+					    typename _R::_FirstType>(
+		[&](auto __i) { return __x[__i]; }),
+	      __optimize_simd_tuple(
+		__simd_tuple_pop_front<_R::_S_first_size>(__x))};
+    else if constexpr (is_same_v<_A0, _A1>
+	&& _R::_S_first_size =3D=3D simd_size_v<_Tp, _A0> + simd_size_v<_Tp, _A1>)
+      return {__concat(__x.template _M_at<0>(), __x.template _M_at<1>()),
+	      __optimize_simd_tuple(__x.second.second)};
+    else if constexpr (sizeof...(_Abis) >=3D 2
+	&& _R::_S_first_size =3D=3D (4 * simd_size_v<_Tp, _A0>)
+	&& simd_size_v<_Tp, _A0> =3D=3D __simd_tuple_element_t<
+	    (sizeof...(_Abis) >=3D 2 ? 3 : 0), _Tup>::size())
+      return {
+	__concat(__concat(__x.template _M_at<0>(), __x.template _M_at<1>()),
+		 __concat(__x.template _M_at<2>(), __x.template _M_at<3>())),
+	__optimize_simd_tuple(__x.second.second.second.second)};
+    else
+      {
+	static_assert(sizeof(_R) =3D=3D sizeof(__x));
+	_R __r;
+	__builtin_memcpy(__r._M_as_charptr(), __x._M_as_charptr(),
+			 sizeof(_Tp) * _R::_S_size());
+	return __r;
+      }
+  }
+
+// __for_each(const _SimdTuple &, Fun) {{{1
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(const _SimdTuple<_Tp, _A0>& __t, _Fp&& __fun)
+  { static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__t), __t.first); }
+
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _A1,
+	  typename... _As, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(const _SimdTuple<_Tp, _A0, _A1, _As...>& __t, _Fp&& __fun)
+  {
+    __fun(__make_meta<_Offset>(__t), __t.first);
+    __for_each<_Offset + simd_size<_Tp, _A0>::value>(__t.second,
+						     static_cast<_Fp&&>(__fun));
+  }
+
+// __for_each(_SimdTuple &, Fun) {{{1
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(_SimdTuple<_Tp, _A0>& __t, _Fp&& __fun)
+  { static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__t), __t.first); }
+
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _A1,
+	  typename... _As, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(_SimdTuple<_Tp, _A0, _A1, _As...>& __t, _Fp&& __fun)
+  {
+    __fun(__make_meta<_Offset>(__t), __t.first);
+    __for_each<_Offset + simd_size<_Tp, _A0>::value>(__t.second,
+						     static_cast<_Fp&&>(__fun));
+  }
+
+// __for_each(_SimdTuple &, const _SimdTuple &, Fun) {{{1
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(_SimdTuple<_Tp, _A0>& __a, const _SimdTuple<_Tp, _A0>& __b,
+	     _Fp&& __fun)
+  {
+    static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__a), __a.first, __b.fi=
rst);
+  }
+
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _A1,
+	  typename... _As, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(_SimdTuple<_Tp, _A0, _A1, _As...>& __a,
+	     const _SimdTuple<_Tp, _A0, _A1, _As...>& __b, _Fp&& __fun)
+  {
+    __fun(__make_meta<_Offset>(__a), __a.first, __b.first);
+    __for_each<_Offset + simd_size<_Tp, _A0>::value>(__a.second, __b.secon=
d,
+						     static_cast<_Fp&&>(__fun));
+  }
+
+// __for_each(const _SimdTuple &, const _SimdTuple &, Fun) {{{1
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(const _SimdTuple<_Tp, _A0>& __a, const _SimdTuple<_Tp, _A0>& =
__b,
+	     _Fp&& __fun)
+  {
+    static_cast<_Fp&&>(__fun)(__make_meta<_Offset>(__a), __a.first, __b.fi=
rst);
+  }
+
+template <size_t _Offset =3D 0, typename _Tp, typename _A0, typename _A1,
+	  typename... _As, typename _Fp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr void
+  __for_each(const _SimdTuple<_Tp, _A0, _A1, _As...>& __a,
+	     const _SimdTuple<_Tp, _A0, _A1, _As...>& __b, _Fp&& __fun)
+  {
+    __fun(__make_meta<_Offset>(__a), __a.first, __b.first);
+    __for_each<_Offset + simd_size<_Tp, _A0>::value>(__a.second, __b.secon=
d,
+						     static_cast<_Fp&&>(__fun));
+  }
+
+// }}}1
+// __extract_part(_SimdTuple) {{{
+template <int _Index, int _Total, int _Combine, typename _Tp, typename _A0,
+	  typename... _As>
+  _GLIBCXX_SIMD_INTRINSIC auto // __vector_type_t or _SimdTuple
+  __extract_part(const _SimdTuple<_Tp, _A0, _As...>& __x)
+  {
+    // worst cases:
+    // (a) 4, 4, 4 =3D> 3, 3, 3, 3 (_Total =3D 4)
+    // (b) 2, 2, 2 =3D> 3, 3       (_Total =3D 2)
+    // (c) 4, 2 =3D> 2, 2, 2       (_Total =3D 3)
+    using _Tuple =3D _SimdTuple<_Tp, _A0, _As...>;
+    static_assert(_Index + _Combine <=3D _Total && _Index >=3D 0 && _Total=
 >=3D 1);
+    constexpr size_t _Np =3D _Tuple::_S_size();
+    static_assert(_Np >=3D _Total && _Np % _Total =3D=3D 0);
+    constexpr size_t __values_per_part =3D _Np / _Total;
+    [[maybe_unused]] constexpr size_t __values_to_skip
+      =3D _Index * __values_per_part;
+    constexpr size_t __return_size =3D __values_per_part * _Combine;
+    using _RetAbi =3D simd_abi::deduce_t<_Tp, __return_size>;
+
+    // handle (optimize) the simple cases
+    if constexpr (_Index =3D=3D 0 && _Tuple::_S_first_size =3D=3D __return=
_size)
+      return __x.first._M_data;
+    else if constexpr (_Index =3D=3D 0 && _Total =3D=3D _Combine)
+      return __x;
+    else if constexpr (_Index =3D=3D 0 && _Tuple::_S_first_size >=3D __ret=
urn_size)
+      return __intrin_bitcast<__vector_type_t<_Tp, __return_size>>(
+	__as_vector(__x.first));
+
+    // recurse to skip unused data members at the beginning of _SimdTuple
+    else if constexpr (__values_to_skip >=3D _Tuple::_S_first_size)
+      { // recurse
+	if constexpr (_Tuple::_S_first_size % __values_per_part =3D=3D 0)
+	  {
+	    constexpr int __parts_in_first
+	      =3D _Tuple::_S_first_size / __values_per_part;
+	    return __extract_part<_Index - __parts_in_first,
+				  _Total - __parts_in_first, _Combine>(
+	      __x.second);
+	  }
+	else
+	  return __extract_part<__values_to_skip - _Tuple::_S_first_size,
+				_Np - _Tuple::_S_first_size, __return_size>(
+	    __x.second);
+      }
+
+    // extract from multiple _SimdTuple data members
+    else if constexpr (__return_size > _Tuple::_S_first_size - __values_to=
_skip)
+      {
+#ifdef _GLIBCXX_SIMD_USE_ALIASING_LOADS
+	const __may_alias<_Tp>* const element_ptr
+	  =3D reinterpret_cast<const __may_alias<_Tp>*>(&__x) + __values_to_skip;
+	return __as_vector(simd<_Tp, _RetAbi>(element_ptr, element_aligned));
+#else
+	[[maybe_unused]] constexpr size_t __offset =3D __values_to_skip;
+	return __as_vector(simd<_Tp, _RetAbi>([&](auto __i) constexpr {
+	  constexpr _SizeConstant<__i + __offset> __k;
+	  return __x[__k];
+	}));
+#endif
+      }
+
+    // all of the return values are in __x.first
+    else if constexpr (_Tuple::_S_first_size % __values_per_part =3D=3D 0)
+      return __extract_part<_Index, _Tuple::_S_first_size / __values_per_p=
art,
+			    _Combine>(__x.first);
+    else
+      return __extract_part<__values_to_skip, _Tuple::_S_first_size,
+			    _Combine * __values_per_part>(__x.first);
+  }
+
+// }}}
+// __fixed_size_storage_t<_Tp, _Np>{{{
+template <typename _Tp, int _Np, typename _Tuple,
+	  typename _Next =3D simd<_Tp, _AllNativeAbis::_BestAbi<_Tp, _Np>>,
+	  int _Remain =3D _Np - int(_Next::size())>
+  struct __fixed_size_storage_builder;
+
+template <typename _Tp, int _Np>
+  struct __fixed_size_storage
+  : public __fixed_size_storage_builder<_Tp, _Np, _SimdTuple<_Tp>> {};
+
+template <typename _Tp, int _Np, typename... _As, typename _Next>
+  struct __fixed_size_storage_builder<_Tp, _Np, _SimdTuple<_Tp, _As...>, _=
Next,
+				      0>
+  { using type =3D _SimdTuple<_Tp, _As..., typename _Next::abi_type>; };
+
+template <typename _Tp, int _Np, typename... _As, typename _Next, int _Rem=
ain>
+  struct __fixed_size_storage_builder<_Tp, _Np, _SimdTuple<_Tp, _As...>, _=
Next,
+				      _Remain>
+  {
+    using type =3D typename __fixed_size_storage_builder<
+      _Tp, _Remain, _SimdTuple<_Tp, _As..., typename _Next::abi_type>>::ty=
pe;
+  };
+
+// }}}
+// _AbisInSimdTuple {{{
+template <typename _Tp>
+  struct _SeqOp;
+
+template <size_t _I0, size_t... _Is>
+  struct _SeqOp<index_sequence<_I0, _Is...>>
+  {
+    using _FirstPlusOne =3D index_sequence<_I0 + 1, _Is...>;
+    using _NotFirstPlusOne =3D index_sequence<_I0, (_Is + 1)...>;
+    template <size_t _First, size_t _Add>
+    using _Prepend =3D index_sequence<_First, _I0 + _Add, (_Is + _Add)...>;
+  };
+
+template <typename _Tp>
+  struct _AbisInSimdTuple;
+
+template <typename _Tp>
+  struct _AbisInSimdTuple<_SimdTuple<_Tp>>
+  {
+    using _Counts =3D index_sequence<0>;
+    using _Begins =3D index_sequence<0>;
+  };
+
+template <typename _Tp, typename _Ap>
+  struct _AbisInSimdTuple<_SimdTuple<_Tp, _Ap>>
+  {
+    using _Counts =3D index_sequence<1>;
+    using _Begins =3D index_sequence<0>;
+  };
+
+template <typename _Tp, typename _A0, typename... _As>
+  struct _AbisInSimdTuple<_SimdTuple<_Tp, _A0, _A0, _As...>>
+  {
+    using _Counts =3D typename _SeqOp<typename _AbisInSimdTuple<
+      _SimdTuple<_Tp, _A0, _As...>>::_Counts>::_FirstPlusOne;
+    using _Begins =3D typename _SeqOp<typename _AbisInSimdTuple<
+      _SimdTuple<_Tp, _A0, _As...>>::_Begins>::_NotFirstPlusOne;
+  };
+
+template <typename _Tp, typename _A0, typename _A1, typename... _As>
+  struct _AbisInSimdTuple<_SimdTuple<_Tp, _A0, _A1, _As...>>
+  {
+    using _Counts =3D typename _SeqOp<typename _AbisInSimdTuple<
+      _SimdTuple<_Tp, _A1, _As...>>::_Counts>::template _Prepend<1, 0>;
+    using _Begins =3D typename _SeqOp<typename _AbisInSimdTuple<
+      _SimdTuple<_Tp, _A1, _As...>>::_Begins>::template _Prepend<0, 1>;
+  };
+
+// }}}
+// __autocvt_to_simd {{{
+template <typename _Tp, bool =3D is_arithmetic_v<__remove_cvref_t<_Tp>>>
+  struct __autocvt_to_simd
+  {
+    _Tp _M_data;
+    using _TT =3D __remove_cvref_t<_Tp>;
+
+    operator _TT()
+    { return _M_data; }
+
+    operator _TT&()
+    {
+      static_assert(is_lvalue_reference<_Tp>::value, "");
+      static_assert(!is_const<_Tp>::value, "");
+      return _M_data;
+    }
+
+    operator _TT*()
+    {
+      static_assert(is_lvalue_reference<_Tp>::value, "");
+      static_assert(!is_const<_Tp>::value, "");
+      return &_M_data;
+    }
+
+    constexpr inline __autocvt_to_simd(_Tp dd) : _M_data(dd) {}
+
+    template <typename _Abi>
+      operator simd<typename _TT::value_type, _Abi>()
+      { return {__private_init, _M_data}; }
+
+    template <typename _Abi>
+      operator simd<typename _TT::value_type, _Abi>&()
+      {
+	return *reinterpret_cast<simd<typename _TT::value_type, _Abi>*>(
+	  &_M_data);
+      }
+
+    template <typename _Abi>
+      operator simd<typename _TT::value_type, _Abi>*()
+      {
+	return reinterpret_cast<simd<typename _TT::value_type, _Abi>*>(
+	  &_M_data);
+      }
+  };
+
+template <typename _Tp>
+  __autocvt_to_simd(_Tp &&) -> __autocvt_to_simd<_Tp>;
+
+template <typename _Tp>
+  struct __autocvt_to_simd<_Tp, true>
+  {
+    using _TT =3D __remove_cvref_t<_Tp>;
+    _Tp _M_data;
+    fixed_size_simd<_TT, 1> _M_fd;
+
+    constexpr inline __autocvt_to_simd(_Tp dd) : _M_data(dd), _M_fd(_M_dat=
a) {}
+
+    ~__autocvt_to_simd()
+    { _M_data =3D __data(_M_fd).first; }
+
+    operator fixed_size_simd<_TT, 1>()
+    { return _M_fd; }
+
+    operator fixed_size_simd<_TT, 1> &()
+    {
+      static_assert(is_lvalue_reference<_Tp>::value, "");
+      static_assert(!is_const<_Tp>::value, "");
+      return _M_fd;
+    }
+
+    operator fixed_size_simd<_TT, 1> *()
+    {
+      static_assert(is_lvalue_reference<_Tp>::value, "");
+      static_assert(!is_const<_Tp>::value, "");
+      return &_M_fd;
+    }
+  };
+
+// }}}
+
+struct _CommonImplFixedSize;
+template <int _Np> struct _SimdImplFixedSize;
+template <int _Np> struct _MaskImplFixedSize;
+// simd_abi::_Fixed {{{
+template <int _Np>
+  struct simd_abi::_Fixed
+  {
+    template <typename _Tp> static constexpr size_t _S_size =3D _Np;
+    template <typename _Tp> static constexpr size_t _S_full_size =3D _Np;
+    // validity traits {{{
+    struct _IsValidAbiTag : public __bool_constant<(_Np > 0)> {};
+
+    template <typename _Tp>
+      struct _IsValidSizeFor
+      : __bool_constant<(_Np <=3D simd_abi::max_fixed_size<_Tp>)> {};
+
+    template <typename _Tp>
+      struct _IsValid : conjunction<_IsValidAbiTag, __is_vectorizable<_Tp>,
+				    _IsValidSizeFor<_Tp>> {};
+
+    template <typename _Tp>
+      static constexpr bool _S_is_valid_v =3D _IsValid<_Tp>::value;
+
+    // }}}
+    // _S_masked {{{
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
+    _S_masked(_BitMask<_Np> __x)
+    { return __x._M_sanitized(); }
+
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
+    _S_masked(_SanitizedBitMask<_Np> __x)
+    { return __x; }
+
+    // }}}
+    // _*Impl {{{
+    using _CommonImpl =3D _CommonImplFixedSize;
+    using _SimdImpl =3D _SimdImplFixedSize<_Np>;
+    using _MaskImpl =3D _MaskImplFixedSize<_Np>;
+
+    // }}}
+    // __traits {{{
+    template <typename _Tp, bool =3D _S_is_valid_v<_Tp>>
+      struct __traits : _InvalidTraits {};
+
+    template <typename _Tp>
+      struct __traits<_Tp, true>
+      {
+	using _IsValid =3D true_type;
+	using _SimdImpl =3D _SimdImplFixedSize<_Np>;
+	using _MaskImpl =3D _MaskImplFixedSize<_Np>;
+
+	// simd and simd_mask member types {{{
+	using _SimdMember =3D __fixed_size_storage_t<_Tp, _Np>;
+	using _MaskMember =3D _SanitizedBitMask<_Np>;
+
+	static constexpr size_t _S_simd_align
+	  =3D std::__bit_ceil(_Np * sizeof(_Tp));
+
+	static constexpr size_t _S_mask_align =3D alignof(_MaskMember);
+
+	// }}}
+	// _SimdBase / base class for simd, providing extra conversions {{{
+	struct _SimdBase
+	{
+	  // The following ensures, function arguments are passed via the stack.
+	  // This is important for ABI compatibility across TU boundaries
+	  _SimdBase(const _SimdBase&) {}
+	  _SimdBase() =3D default;
+
+	  explicit operator const _SimdMember &() const
+	  { return static_cast<const simd<_Tp, _Fixed>*>(this)->_M_data; }
+
+	  explicit operator array<_Tp, _Np>() const
+	  {
+	    array<_Tp, _Np> __r;
+	    // _SimdMember can be larger because of higher alignment
+	    static_assert(sizeof(__r) <=3D sizeof(_SimdMember), "");
+	    __builtin_memcpy(__r.data(), &static_cast<const _SimdMember&>(*this),
+			     sizeof(__r));
+	    return __r;
+	  }
+	};
+
+	// }}}
+	// _MaskBase {{{
+	// empty. The bitset interface suffices
+	struct _MaskBase {};
+
+	// }}}
+	// _SimdCastType {{{
+	struct _SimdCastType
+	{
+	  _SimdCastType(const array<_Tp, _Np>&);
+	  _SimdCastType(const _SimdMember& dd) : _M_data(dd) {}
+	  explicit operator const _SimdMember &() const { return _M_data; }
+
+	private:
+	  const _SimdMember& _M_data;
+	};
+
+	// }}}
+	// _MaskCastType {{{
+	class _MaskCastType
+	{
+	  _MaskCastType() =3D delete;
+	};
+	// }}}
+      };
+    // }}}
+  };
+
+// }}}
+// _CommonImplFixedSize {{{
+struct _CommonImplFixedSize
+{
+  // _S_store {{{
+  template <typename _Tp, typename... _As>
+    _GLIBCXX_SIMD_INTRINSIC static void
+    _S_store(const _SimdTuple<_Tp, _As...>& __x, void* __addr)
+    {
+      constexpr size_t _Np =3D _SimdTuple<_Tp, _As...>::_S_size();
+      __builtin_memcpy(__addr, &__x, _Np * sizeof(_Tp));
+    }
+
+  // }}}
+};
+
+// }}}
+// _SimdImplFixedSize {{{1
+// fixed_size should not inherit from _SimdMathFallback in order for
+// specializations in the used _SimdTuple Abis to get used
+template <int _Np>
+  struct _SimdImplFixedSize
+  {
+    // member types {{{2
+    using _MaskMember =3D _SanitizedBitMask<_Np>;
+
+    template <typename _Tp>
+      using _SimdMember =3D __fixed_size_storage_t<_Tp, _Np>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_tuple_size =3D _SimdMember<_Tp>::_S_tuple=
_size;
+
+    template <typename _Tp>
+      using _Simd =3D simd<_Tp, simd_abi::fixed_size<_Np>>;
+
+    template <typename _Tp>
+      using _TypeTag =3D _Tp*;
+
+    // broadcast {{{2
+    template <typename _Tp>
+      static constexpr inline _SimdMember<_Tp> _S_broadcast(_Tp __x) noexc=
ept
+      {
+	return _SimdMember<_Tp>::_S_generate([&](auto __meta) constexpr {
+	  return __meta._S_broadcast(__x);
+	});
+      }
+
+    // _S_generator {{{2
+    template <typename _Fp, typename _Tp>
+      static constexpr inline _SimdMember<_Tp> _S_generator(_Fp&& __gen,
+							    _TypeTag<_Tp>)
+      {
+	return _SimdMember<_Tp>::_S_generate([&__gen](auto __meta) constexpr {
+	  return __meta._S_generator(
+	    [&](auto __i) constexpr {
+	      return __i < _Np ? __gen(_SizeConstant<__meta._S_offset + __i>())
+			       : 0;
+	    },
+	    _TypeTag<_Tp>());
+	});
+      }
+
+    // _S_load {{{2
+    template <typename _Tp, typename _Up>
+      static inline _SimdMember<_Tp> _S_load(const _Up* __mem,
+					     _TypeTag<_Tp>) noexcept
+      {
+	return _SimdMember<_Tp>::_S_generate([&](auto __meta) {
+	  return __meta._S_load(&__mem[__meta._S_offset], _TypeTag<_Tp>());
+	});
+      }
+
+    // _S_masked_load {{{2
+    template <typename _Tp, typename... _As, typename _Up>
+      static inline _SimdTuple<_Tp, _As...>
+      _S_masked_load(const _SimdTuple<_Tp, _As...>& __old,
+		     const _MaskMember __bits, const _Up* __mem) noexcept
+      {
+	auto __merge =3D __old;
+	__for_each(__merge, [&](auto __meta, auto& __native) {
+	  if (__meta._S_submask(__bits).any())
+#pragma GCC diagnostic push
+	  // __mem + __mem._S_offset could be UB ([expr.add]/4.3, but it punts
+	  // the responsibility for avoiding UB to the caller of the masked load
+	  // via the mask. Consequently, the compiler may assume this branch is
+	  // unreachable, if the pointer arithmetic is UB.
+#pragma GCC diagnostic ignored "-Warray-bounds"
+	    __native
+	      =3D __meta._S_masked_load(__native, __meta._S_make_mask(__bits),
+				      __mem + __meta._S_offset);
+#pragma GCC diagnostic pop
+	});
+	return __merge;
+      }
+
+    // _S_store {{{2
+    template <typename _Tp, typename _Up>
+      static inline void _S_store(const _SimdMember<_Tp>& __v, _Up* __mem,
+				  _TypeTag<_Tp>) noexcept
+      {
+	__for_each(__v, [&](auto __meta, auto __native) {
+	  __meta._S_store(__native, &__mem[__meta._S_offset], _TypeTag<_Tp>());
+	});
+      }
+
+    // _S_masked_store {{{2
+    template <typename _Tp, typename... _As, typename _Up>
+      static inline void _S_masked_store(const _SimdTuple<_Tp, _As...>& __=
v,
+					 _Up* __mem,
+					 const _MaskMember __bits) noexcept
+      {
+	__for_each(__v, [&](auto __meta, auto __native) {
+	  if (__meta._S_submask(__bits).any())
+#pragma GCC diagnostic push
+	  // __mem + __mem._S_offset could be UB ([expr.add]/4.3, but it punts
+	  // the responsibility for avoiding UB to the caller of the masked
+	  // store via the mask. Consequently, the compiler may assume this
+	  // branch is unreachable, if the pointer arithmetic is UB.
+#pragma GCC diagnostic ignored "-Warray-bounds"
+	    __meta._S_masked_store(__native, __mem + __meta._S_offset,
+				   __meta._S_make_mask(__bits));
+#pragma GCC diagnostic pop
+	});
+      }
+
+    // negation {{{2
+    template <typename _Tp, typename... _As>
+      static inline _MaskMember
+      _S_negate(const _SimdTuple<_Tp, _As...>& __x) noexcept
+      {
+	_MaskMember __bits =3D 0;
+	__for_each(
+	  __x, [&__bits](auto __meta, auto __native) constexpr {
+	    __bits
+	      |=3D __meta._S_mask_to_shifted_ullong(__meta._S_negate(__native));
+	  });
+	return __bits;
+      }
+
+    // reductions {{{2
+    template <typename _Tp, typename _BinaryOperation>
+      static constexpr inline _Tp _S_reduce(const _Simd<_Tp>& __x,
+					    const _BinaryOperation& __binary_op)
+      {
+	using _Tup =3D _SimdMember<_Tp>;
+	const _Tup& __tup =3D __data(__x);
+	if constexpr (_Tup::_S_tuple_size =3D=3D 1)
+	  return _Tup::_FirstAbi::_SimdImpl::_S_reduce(
+	    __tup.template _M_simd_at<0>(), __binary_op);
+	else if constexpr (_Tup::_S_tuple_size =3D=3D 2 && _Tup::_S_size() > 2
+			   && _Tup::_SecondType::_S_size() =3D=3D 1)
+	  {
+	    return __binary_op(simd<_Tp, simd_abi::scalar>(
+				 reduce(__tup.template _M_simd_at<0>(),
+					__binary_op)),
+			       __tup.template _M_simd_at<1>())[0];
+	  }
+	else if constexpr (_Tup::_S_tuple_size =3D=3D 2 && _Tup::_S_size() > 4
+			   && _Tup::_SecondType::_S_size() =3D=3D 2)
+	  {
+	    return __binary_op(
+	      simd<_Tp, simd_abi::scalar>(
+		reduce(__tup.template _M_simd_at<0>(), __binary_op)),
+	      simd<_Tp, simd_abi::scalar>(
+		reduce(__tup.template _M_simd_at<1>(), __binary_op)))[0];
+	  }
+	else
+	  {
+	    const auto& __x2 =3D __call_with_n_evaluations<
+	      __div_roundup(_Tup::_S_tuple_size, 2)>(
+	      [](auto __first_simd, auto... __remaining) {
+		if constexpr (sizeof...(__remaining) =3D=3D 0)
+		  return __first_simd;
+		else
+		  {
+		    using _Tup2
+		      =3D _SimdTuple<_Tp,
+				   typename decltype(__first_simd)::abi_type,
+				   typename decltype(__remaining)::abi_type...>;
+		    return fixed_size_simd<_Tp, _Tup2::_S_size()>(
+		      __private_init,
+		      __make_simd_tuple(__first_simd, __remaining...));
+		  }
+	      },
+	      [&](auto __i) {
+		auto __left =3D __tup.template _M_simd_at<2 * __i>();
+		if constexpr (2 * __i + 1 =3D=3D _Tup::_S_tuple_size)
+		  return __left;
+		else
+		  {
+		    auto __right =3D __tup.template _M_simd_at<2 * __i + 1>();
+		    using _LT =3D decltype(__left);
+		    using _RT =3D decltype(__right);
+		    if constexpr (_LT::size() =3D=3D _RT::size())
+		      return __binary_op(__left, __right);
+		    else
+		      {
+			_GLIBCXX_SIMD_USE_CONSTEXPR_API
+			typename _LT::mask_type __k(
+			  __private_init,
+			  [](auto __j) constexpr { return __j < _RT::size(); });
+			_LT __ext_right =3D __left;
+			where(__k, __ext_right)
+			  =3D __proposed::resizing_simd_cast<_LT>(__right);
+			where(__k, __left) =3D __binary_op(__left, __ext_right);
+			return __left;
+		      }
+		  }
+	      });
+	    return reduce(__x2, __binary_op);
+	  }
+      }
+
+    // _S_min, _S_max {{{2
+    template <typename _Tp, typename... _As>
+      static inline constexpr _SimdTuple<_Tp, _As...>
+      _S_min(const _SimdTuple<_Tp, _As...>& __a,
+	     const _SimdTuple<_Tp, _As...>& __b)
+      {
+	return __a._M_apply_per_chunk(
+	  [](auto __impl, auto __aa, auto __bb) constexpr {
+	    return __impl._S_min(__aa, __bb);
+	  },
+	  __b);
+      }
+
+    template <typename _Tp, typename... _As>
+      static inline constexpr _SimdTuple<_Tp, _As...>
+      _S_max(const _SimdTuple<_Tp, _As...>& __a,
+	     const _SimdTuple<_Tp, _As...>& __b)
+      {
+	return __a._M_apply_per_chunk(
+	  [](auto __impl, auto __aa, auto __bb) constexpr {
+	    return __impl._S_max(__aa, __bb);
+	  },
+	  __b);
+      }
+
+    // _S_complement {{{2
+    template <typename _Tp, typename... _As>
+      static inline constexpr _SimdTuple<_Tp, _As...>
+      _S_complement(const _SimdTuple<_Tp, _As...>& __x) noexcept
+      {
+	return __x._M_apply_per_chunk([](auto __impl, auto __xx) constexpr {
+	  return __impl._S_complement(__xx);
+	});
+      }
+
+    // _S_unary_minus {{{2
+    template <typename _Tp, typename... _As>
+      static inline constexpr _SimdTuple<_Tp, _As...>
+      _S_unary_minus(const _SimdTuple<_Tp, _As...>& __x) noexcept
+      {
+	return __x._M_apply_per_chunk([](auto __impl, auto __xx) constexpr {
+	  return __impl._S_unary_minus(__xx);
+	});
+      }
+
+    // arithmetic operators {{{2
+
+#define _GLIBCXX_SIMD_FIXED_OP(name_, op_)                                =
     \
+    template <typename _Tp, typename... _As>                              =
     \
+      static inline constexpr _SimdTuple<_Tp, _As...> name_(              =
     \
+	const _SimdTuple<_Tp, _As...> __x, const _SimdTuple<_Tp, _As...> __y)  \
+      {                                                                   =
     \
+	return __x._M_apply_per_chunk(                                         \
+	  [](auto __impl, auto __xx, auto __yy) constexpr {                    \
+	    return __impl.name_(__xx, __yy);                                   \
+	  },                                                                   \
+	  __y);                                                                \
+      }
+
+    _GLIBCXX_SIMD_FIXED_OP(_S_plus, +)
+    _GLIBCXX_SIMD_FIXED_OP(_S_minus, -)
+    _GLIBCXX_SIMD_FIXED_OP(_S_multiplies, *)
+    _GLIBCXX_SIMD_FIXED_OP(_S_divides, /)
+    _GLIBCXX_SIMD_FIXED_OP(_S_modulus, %)
+    _GLIBCXX_SIMD_FIXED_OP(_S_bit_and, &)
+    _GLIBCXX_SIMD_FIXED_OP(_S_bit_or, |)
+    _GLIBCXX_SIMD_FIXED_OP(_S_bit_xor, ^)
+    _GLIBCXX_SIMD_FIXED_OP(_S_bit_shift_left, <<)
+    _GLIBCXX_SIMD_FIXED_OP(_S_bit_shift_right, >>)
+#undef _GLIBCXX_SIMD_FIXED_OP
+
+    template <typename _Tp, typename... _As>
+      static inline constexpr _SimdTuple<_Tp, _As...>
+      _S_bit_shift_left(const _SimdTuple<_Tp, _As...>& __x, int __y)
+      {
+	return __x._M_apply_per_chunk([__y](auto __impl, auto __xx) constexpr {
+	  return __impl._S_bit_shift_left(__xx, __y);
+	});
+      }
+
+    template <typename _Tp, typename... _As>
+      static inline constexpr _SimdTuple<_Tp, _As...>
+      _S_bit_shift_right(const _SimdTuple<_Tp, _As...>& __x, int __y)
+      {
+	return __x._M_apply_per_chunk([__y](auto __impl, auto __xx) constexpr {
+	  return __impl._S_bit_shift_right(__xx, __y);
+	});
+      }
+
+  // math {{{2
+#define _GLIBCXX_SIMD_APPLY_ON_TUPLE(_RetTp, __name)                      =
     \
+    template <typename _Tp, typename... _As, typename... _More>           =
     \
+      static inline __fixed_size_storage_t<_RetTp, _Np>                   =
     \
+	_S_##__name(const _SimdTuple<_Tp, _As...>& __x,                        \
+		    const _More&... __more)                                    \
+      {                                                                   =
     \
+	if constexpr (sizeof...(_More) =3D=3D 0)                                 =
  \
+	  {                                                                    \
+	    if constexpr (is_same_v<_Tp, _RetTp>)                              \
+	      return __x._M_apply_per_chunk(                                   \
+		[](auto __impl, auto __xx) constexpr {                         \
+		  using _V =3D typename decltype(__impl)::simd_type;             \
+		  return __data(__name(_V(__private_init, __xx)));             \
+		});                                                            \
+	    else                                                               \
+	      return __optimize_simd_tuple(                                    \
+		__x.template _M_apply_r<_RetTp>([](auto __impl, auto __xx) {   \
+		  return __impl._S_##__name(__xx);                             \
+		}));                                                           \
+	  }                                                                    \
+	else if constexpr (                                                    \
+	  is_same_v<                                                           \
+	    _Tp,                                                               \
+	    _RetTp> && (... && is_same_v<_SimdTuple<_Tp, _As...>, _More>) )    \
+	  return __x._M_apply_per_chunk(                                       \
+	    [](auto __impl, auto __xx, auto... __pack) constexpr {             \
+	      using _V =3D typename decltype(__impl)::simd_type;                 \
+	      return __data(__name(_V(__private_init, __xx),                   \
+				   _V(__private_init, __pack)...));            \
+	    },                                                                 \
+	    __more...);                                                        \
+	else if constexpr (is_same_v<_Tp, _RetTp>)                             \
+	  return __x._M_apply_per_chunk(                                       \
+	    [](auto __impl, auto __xx, auto... __pack) constexpr {             \
+	      using _V =3D typename decltype(__impl)::simd_type;                 \
+	      return __data(__name(_V(__private_init, __xx),                   \
+				   __autocvt_to_simd(__pack)...));             \
+	    },                                                                 \
+	    __more...);                                                        \
+	else                                                                   \
+	  __assert_unreachable<_Tp>();                                         \
+      }
+
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, acos)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, asin)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, atan)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, atan2)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, cos)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, sin)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, tan)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, acosh)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, asinh)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, atanh)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, cosh)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, sinh)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, tanh)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, exp)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, exp2)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, expm1)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(int, ilogb)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log10)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log1p)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, log2)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, logb)
+    // modf implemented in simd_math.h
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp,
+				 scalbn) // double scalbn(double x, int exp);
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, scalbln)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, cbrt)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, abs)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fabs)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, pow)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, sqrt)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, erf)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, erfc)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, lgamma)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, tgamma)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, trunc)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, ceil)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, floor)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, nearbyint)
+
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, rint)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(long, lrint)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(long long, llrint)
+
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, round)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(long, lround)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(long long, llround)
+
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, ldexp)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fmod)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, remainder)
+    // copysign in simd_math.h
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, nextafter)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fdim)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fmax)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fmin)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(_Tp, fma)
+    _GLIBCXX_SIMD_APPLY_ON_TUPLE(int, fpclassify)
+#undef _GLIBCXX_SIMD_APPLY_ON_TUPLE
+
+    template <typename _Tp, typename... _Abis>
+      static _SimdTuple<_Tp, _Abis...> _S_remquo(
+	const _SimdTuple<_Tp, _Abis...>& __x,
+	const _SimdTuple<_Tp, _Abis...>& __y,
+	__fixed_size_storage_t<int, _SimdTuple<_Tp, _Abis...>::_S_size()>* __z)
+      {
+	return __x._M_apply_per_chunk(
+	  [](auto __impl, const auto __xx, const auto __yy, auto& __zz) {
+	    return __impl._S_remquo(__xx, __yy, &__zz);
+	  },
+	  __y, *__z);
+      }
+
+    template <typename _Tp, typename... _As>
+      static inline _SimdTuple<_Tp, _As...>
+      _S_frexp(const _SimdTuple<_Tp, _As...>& __x,
+	       __fixed_size_storage_t<int, _Np>& __exp) noexcept
+      {
+	return __x._M_apply_per_chunk(
+	  [](auto __impl, const auto& __a, auto& __b) {
+	    return __data(
+	      frexp(typename decltype(__impl)::simd_type(__private_init, __a),
+		    __autocvt_to_simd(__b)));
+	  },
+	  __exp);
+      }
+
+#define _GLIBCXX_SIMD_TEST_ON_TUPLE_(name_)                               =
     \
+    template <typename _Tp, typename... _As>                              =
     \
+      static inline _MaskMember                                           =
     \
+	_S_##name_(const _SimdTuple<_Tp, _As...>& __x) noexcept                \
+      {                                                                   =
     \
+	return _M_test([](auto __impl,                                         \
+			  auto __xx) { return __impl._S_##name_(__xx); },      \
+		       __x);                                                   \
+      }
+
+    _GLIBCXX_SIMD_TEST_ON_TUPLE_(isinf)
+    _GLIBCXX_SIMD_TEST_ON_TUPLE_(isfinite)
+    _GLIBCXX_SIMD_TEST_ON_TUPLE_(isnan)
+    _GLIBCXX_SIMD_TEST_ON_TUPLE_(isnormal)
+    _GLIBCXX_SIMD_TEST_ON_TUPLE_(signbit)
+#undef _GLIBCXX_SIMD_TEST_ON_TUPLE_
+
+    // _S_increment & _S_decrement{{{2
+    template <typename... _Ts>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr void
+      _S_increment(_SimdTuple<_Ts...>& __x)
+      {
+	__for_each(
+	  __x, [](auto __meta, auto& native) constexpr {
+	    __meta._S_increment(native);
+	  });
+      }
+
+    template <typename... _Ts>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr void
+      _S_decrement(_SimdTuple<_Ts...>& __x)
+      {
+	__for_each(
+	  __x, [](auto __meta, auto& native) constexpr {
+	    __meta._S_decrement(native);
+	  });
+      }
+
+    // compares {{{2
+#define _GLIBCXX_SIMD_CMP_OPERATIONS(__cmp)                               =
     \
+    template <typename _Tp, typename... _As>                              =
     \
+      _GLIBCXX_SIMD_INTRINSIC constexpr static _MaskMember                =
     \
+      __cmp(const _SimdTuple<_Tp, _As...>& __x,                           =
     \
+	    const _SimdTuple<_Tp, _As...>& __y)                                \
+      {                                                                   =
     \
+	return _M_test(                                                        \
+	  [](auto __impl, auto __xx, auto __yy) constexpr {                    \
+	    return __impl.__cmp(__xx, __yy);                                   \
+	  },                                                                   \
+	  __x, __y);                                                           \
+      }
+
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_equal_to)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_not_equal_to)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_less)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_less_equal)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isless)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_islessequal)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isgreater)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isgreaterequal)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_islessgreater)
+    _GLIBCXX_SIMD_CMP_OPERATIONS(_S_isunordered)
+#undef _GLIBCXX_SIMD_CMP_OPERATIONS
+
+    // smart_reference access {{{2
+    template <typename _Tp, typename... _As, typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC static void _S_set(_SimdTuple<_Tp, _As...>& =
__v,
+						 int __i, _Up&& __x) noexcept
+      { __v._M_set(__i, static_cast<_Up&&>(__x)); }
+
+    // _S_masked_assign {{{2
+    template <typename _Tp, typename... _As>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(const _MaskMember __bits, _SimdTuple<_Tp, _As...>& =
__lhs,
+		       const __type_identity_t<_SimdTuple<_Tp, _As...>>& __rhs)
+      {
+	__for_each(
+	  __lhs, __rhs,
+	  [&](auto __meta, auto& __native_lhs, auto __native_rhs) constexpr {
+	    __meta._S_masked_assign(__meta._S_make_mask(__bits), __native_lhs,
+				    __native_rhs);
+	  });
+      }
+
+    // Optimization for the case where the RHS is a scalar. No need to bro=
adcast
+    // the scalar to a simd first.
+    template <typename _Tp, typename... _As>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(const _MaskMember __bits, _SimdTuple<_Tp, _As...>& =
__lhs,
+		       const __type_identity_t<_Tp> __rhs)
+      {
+	__for_each(
+	  __lhs, [&](auto __meta, auto& __native_lhs) constexpr {
+	    __meta._S_masked_assign(__meta._S_make_mask(__bits), __native_lhs,
+				    __rhs);
+	  });
+      }
+
+    // _S_masked_cassign {{{2
+    template <typename _Op, typename _Tp, typename... _As>
+      static inline void _S_masked_cassign(const _MaskMember __bits,
+					   _SimdTuple<_Tp, _As...>& __lhs,
+					   const _SimdTuple<_Tp, _As...>& __rhs,
+					   _Op __op)
+      {
+	__for_each(
+	  __lhs, __rhs,
+	  [&](auto __meta, auto& __native_lhs, auto __native_rhs) constexpr {
+	    __meta.template _S_masked_cassign(__meta._S_make_mask(__bits),
+					      __native_lhs, __native_rhs, __op);
+	  });
+      }
+
+    // Optimization for the case where the RHS is a scalar. No need to bro=
adcast
+    // the scalar to a simd first.
+    template <typename _Op, typename _Tp, typename... _As>
+      static inline void _S_masked_cassign(const _MaskMember __bits,
+					   _SimdTuple<_Tp, _As...>& __lhs,
+					   const _Tp& __rhs, _Op __op)
+      {
+	__for_each(
+	  __lhs, [&](auto __meta, auto& __native_lhs) constexpr {
+	    __meta.template _S_masked_cassign(__meta._S_make_mask(__bits),
+					      __native_lhs, __rhs, __op);
+	  });
+      }
+
+    // _S_masked_unary {{{2
+    template <template <typename> class _Op, typename _Tp, typename... _As>
+      static inline _SimdTuple<_Tp, _As...>
+      _S_masked_unary(const _MaskMember __bits,
+		      const _SimdTuple<_Tp, _As...> __v) // TODO: const-ref __v?
+      {
+	return __v._M_apply_wrapped([&__bits](auto __meta,
+					      auto __native) constexpr {
+	  return __meta.template _S_masked_unary<_Op>(__meta._S_make_mask(
+							__bits),
+						      __native);
+	});
+      }
+
+    // }}}2
+  };
+
+// _MaskImplFixedSize {{{1
+template <int _Np>
+  struct _MaskImplFixedSize
+  {
+    static_assert(
+      sizeof(_ULLong) * __CHAR_BIT__ >=3D _Np,
+      "The fixed_size implementation relies on one _ULLong being able to s=
tore "
+      "all boolean elements."); // required in load & store
+
+    // member types {{{
+    using _Abi =3D simd_abi::fixed_size<_Np>;
+
+    using _MaskMember =3D _SanitizedBitMask<_Np>;
+
+    template <typename _Tp>
+      using _FirstAbi =3D typename __fixed_size_storage_t<_Tp, _Np>::_Firs=
tAbi;
+
+    template <typename _Tp>
+      using _TypeTag =3D _Tp*;
+
+    // }}}
+    // _S_broadcast {{{
+    template <typename>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
+      _S_broadcast(bool __x)
+      { return __x ? ~_MaskMember() : _MaskMember(); }
+
+    // }}}
+    // _S_load {{{
+    template <typename>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
+      _S_load(const bool* __mem)
+      {
+	using _Ip =3D __int_for_sizeof_t<bool>;
+	// the following load uses element_aligned and relies on __mem already
+	// carrying alignment information from when this load function was
+	// called.
+	const simd<_Ip, _Abi> __bools(reinterpret_cast<const __may_alias<_Ip>*>(
+					__mem),
+				      element_aligned);
+	return __data(__bools !=3D 0);
+      }
+
+    // }}}
+    // _S_to_bits {{{
+    template <bool _Sanitized>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
+      _S_to_bits(_BitMask<_Np, _Sanitized> __x)
+      {
+	if constexpr (_Sanitized)
+	  return __x;
+	else
+	  return __x._M_sanitized();
+      }
+
+    // }}}
+    // _S_convert {{{
+    template <typename _Tp, typename _Up, typename _UAbi>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
+      _S_convert(simd_mask<_Up, _UAbi> __x)
+      {
+	return _UAbi::_MaskImpl::_S_to_bits(__data(__x))
+	  .template _M_extract<0, _Np>();
+      }
+
+    // }}}
+    // _S_from_bitmask {{{2
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember
+      _S_from_bitmask(_MaskMember __bits, _TypeTag<_Tp>) noexcept
+      { return __bits; }
+
+    // _S_load {{{2
+    static inline _MaskMember _S_load(const bool* __mem) noexcept
+    {
+      // TODO: _UChar is not necessarily the best type to use here. For sm=
aller
+      // _Np _UShort, _UInt, _ULLong, float, and double can be more effici=
ent.
+      _ULLong __r =3D 0;
+      using _Vs =3D __fixed_size_storage_t<_UChar, _Np>;
+      __for_each(_Vs{}, [&](auto __meta, auto) {
+	__r |=3D __meta._S_mask_to_shifted_ullong(
+	  __meta._S_mask_impl._S_load(&__mem[__meta._S_offset],
+				      _SizeConstant<__meta._S_size()>()));
+      });
+      return __r;
+    }
+
+    // _S_masked_load {{{2
+    static inline _MaskMember _S_masked_load(_MaskMember __merge,
+					     _MaskMember __mask,
+					     const bool* __mem) noexcept
+    {
+      _BitOps::_S_bit_iteration(__mask.to_ullong(), [&](auto __i) {
+	__merge.set(__i, __mem[__i]);
+      });
+      return __merge;
+    }
+
+    // _S_store {{{2
+    static inline void _S_store(const _MaskMember __bitmask,
+				bool* __mem) noexcept
+    {
+      if constexpr (_Np =3D=3D 1)
+	__mem[0] =3D __bitmask[0];
+      else
+	_FirstAbi<_UChar>::_CommonImpl::_S_store_bool_array(__bitmask, __mem);
+    }
+
+    // _S_masked_store {{{2
+    static inline void _S_masked_store(const _MaskMember __v, bool* __mem,
+				       const _MaskMember __k) noexcept
+    {
+      _BitOps::_S_bit_iteration(__k, [&](auto __i) { __mem[__i] =3D __v[__=
i]; });
+    }
+
+    // logical and bitwise operators {{{2
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember
+    _S_logical_and(const _MaskMember& __x, const _MaskMember& __y) noexcept
+    { return __x & __y; }
+
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember
+    _S_logical_or(const _MaskMember& __x, const _MaskMember& __y) noexcept
+    { return __x | __y; }
+
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember
+    _S_bit_not(const _MaskMember& __x) noexcept
+    { return ~__x; }
+
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember
+    _S_bit_and(const _MaskMember& __x, const _MaskMember& __y) noexcept
+    { return __x & __y; }
+
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember
+    _S_bit_or(const _MaskMember& __x, const _MaskMember& __y) noexcept
+    { return __x | __y; }
+
+    _GLIBCXX_SIMD_INTRINSIC static _MaskMember
+    _S_bit_xor(const _MaskMember& __x, const _MaskMember& __y) noexcept
+    { return __x ^ __y; }
+
+    // smart_reference access {{{2
+    _GLIBCXX_SIMD_INTRINSIC static void _S_set(_MaskMember& __k, int __i,
+					       bool __x) noexcept
+    { __k.set(__i, __x); }
+
+    // _S_masked_assign {{{2
+    _GLIBCXX_SIMD_INTRINSIC static void
+    _S_masked_assign(const _MaskMember __k, _MaskMember& __lhs,
+		     const _MaskMember __rhs)
+    { __lhs =3D (__lhs & ~__k) | (__rhs & __k); }
+
+    // Optimization for the case where the RHS is a scalar.
+    _GLIBCXX_SIMD_INTRINSIC static void _S_masked_assign(const _MaskMember=
 __k,
+							 _MaskMember& __lhs,
+							 const bool __rhs)
+    {
+      if (__rhs)
+	__lhs |=3D __k;
+      else
+	__lhs &=3D ~__k;
+    }
+
+    // }}}2
+    // _S_all_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_all_of(simd_mask<_Tp, _Abi> _=
_k)
+      { return __data(__k).all(); }
+
+    // }}}
+    // _S_any_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_any_of(simd_mask<_Tp, _Abi> _=
_k)
+      { return __data(__k).any(); }
+
+    // }}}
+    // _S_none_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_none_of(simd_mask<_Tp, _Abi> =
__k)
+      { return __data(__k).none(); }
+
+    // }}}
+    // _S_some_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool
+      _S_some_of([[maybe_unused]] simd_mask<_Tp, _Abi> __k)
+      {
+	if constexpr (_Np =3D=3D 1)
+	  return false;
+	else
+	  return __data(__k).any() && !__data(__k).all();
+      }
+
+    // }}}
+    // _S_popcount {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int _S_popcount(simd_mask<_Tp, _Abi> =
__k)
+      { return __data(__k).count(); }
+
+    // }}}
+    // _S_find_first_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_first_set(simd_mask<_Tp, _Abi> __k)
+      { return std::__countr_zero(__data(__k).to_ullong()); }
+
+    // }}}
+    // _S_find_last_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_last_set(simd_mask<_Tp, _Abi> __k)
+      { return std::__bit_width(__data(__k).to_ullong()) - 1; }
+
+    // }}}
+  };
+// }}}1
+
+_GLIBCXX_SIMD_END_NAMESPACE
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_FIXED_SIZE_H_
+
+// vim: foldmethod=3Dmarker sw=3D2 noet ts=3D8 sts=3D2 tw=3D80
diff --git a/libstdc++-v3/include/experimental/bits/simd_math.h b/libstdc++=
=2Dv3/include/experimental/bits/simd_math.h
new file mode 100644
index 00000000000..bbaa899faa2
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_math.h
@@ -0,0 +1,1500 @@
+// Math overloads for simd -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_MATH_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_MATH_H_
+
+#if __cplusplus >=3D 201703L
+
+#include <utility>
+#include <iomanip>
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+template <typename _Tp, typename _V>
+  using _Samesize =3D fixed_size_simd<_Tp, _V::size()>;
+
+// _Math_return_type {{{
+template <typename _DoubleR, typename _Tp, typename _Abi>
+  struct _Math_return_type;
+
+template <typename _DoubleR, typename _Tp, typename _Abi>
+  using _Math_return_type_t =3D
+    typename _Math_return_type<_DoubleR, _Tp, _Abi>::type;
+
+template <typename _Tp, typename _Abi>
+  struct _Math_return_type<double, _Tp, _Abi>
+  { using type =3D simd<_Tp, _Abi>; };
+
+template <typename _Tp, typename _Abi>
+  struct _Math_return_type<bool, _Tp, _Abi>
+  { using type =3D simd_mask<_Tp, _Abi>; };
+
+template <typename _DoubleR, typename _Tp, typename _Abi>
+  struct _Math_return_type
+  { using type =3D fixed_size_simd<_DoubleR, simd_size_v<_Tp, _Abi>>; };
+
+//}}}
+// _GLIBCXX_SIMD_MATH_CALL_ {{{
+#define _GLIBCXX_SIMD_MATH_CALL_(__name)                                  =
     \
+template <typename _Tp, typename _Abi, typename...,                       =
     \
+	  typename _R =3D _Math_return_type_t<                                   \
+	    decltype(std::__name(declval<double>())), _Tp, _Abi>>              \
+  enable_if_t<is_floating_point_v<_Tp>, _R>                               =
     \
+  __name(simd<_Tp, _Abi> __x)                                             =
     \
+  { return {__private_init, _Abi::_SimdImpl::_S_##__name(__data(__x))}; }
+
+// }}}
+//_Extra_argument_type{{{
+template <typename _Up, typename _Tp, typename _Abi>
+  struct _Extra_argument_type;
+
+template <typename _Tp, typename _Abi>
+  struct _Extra_argument_type<_Tp*, _Tp, _Abi>
+  {
+    using type =3D simd<_Tp, _Abi>*;
+    static constexpr double* declval();
+    static constexpr bool __needs_temporary_scalar =3D true;
+
+    _GLIBCXX_SIMD_INTRINSIC static constexpr auto _S_data(type __x)
+    { return &__data(*__x); }
+  };
+
+template <typename _Up, typename _Tp, typename _Abi>
+  struct _Extra_argument_type<_Up*, _Tp, _Abi>
+  {
+    static_assert(is_integral_v<_Up>);
+    using type =3D fixed_size_simd<_Up, simd_size_v<_Tp, _Abi>>*;
+    static constexpr _Up* declval();
+    static constexpr bool __needs_temporary_scalar =3D true;
+
+    _GLIBCXX_SIMD_INTRINSIC static constexpr auto _S_data(type __x)
+    { return &__data(*__x); }
+  };
+
+template <typename _Tp, typename _Abi>
+  struct _Extra_argument_type<_Tp, _Tp, _Abi>
+  {
+    using type =3D simd<_Tp, _Abi>;
+    static constexpr double declval();
+    static constexpr bool __needs_temporary_scalar =3D false;
+
+    _GLIBCXX_SIMD_INTRINSIC static constexpr decltype(auto)
+    _S_data(const type& __x)
+    { return __data(__x); }
+  };
+
+template <typename _Up, typename _Tp, typename _Abi>
+  struct _Extra_argument_type
+  {
+    static_assert(is_integral_v<_Up>);
+    using type =3D fixed_size_simd<_Up, simd_size_v<_Tp, _Abi>>;
+    static constexpr _Up declval();
+    static constexpr bool __needs_temporary_scalar =3D false;
+
+    _GLIBCXX_SIMD_INTRINSIC static constexpr decltype(auto)
+    _S_data(const type& __x)
+    { return __data(__x); }
+  };
+
+//}}}
+// _GLIBCXX_SIMD_MATH_CALL2_ {{{
+#define _GLIBCXX_SIMD_MATH_CALL2_(__name, arg2_)                          =
     \
+template <                                                                =
     \
+  typename _Tp, typename _Abi, typename...,                               =
     \
+  typename _Arg2 =3D _Extra_argument_type<arg2_, _Tp, _Abi>,              =
       \
+  typename _R =3D _Math_return_type_t<                                    =
       \
+    decltype(std::__name(declval<double>(), _Arg2::declval())), _Tp, _Abi>=
>    \
+  enable_if_t<is_floating_point_v<_Tp>, _R>                               =
     \
+  __name(const simd<_Tp, _Abi>& __x, const typename _Arg2::type& __y)     =
     \
+  {                                                                       =
     \
+    return {__private_init,                                               =
     \
+	    _Abi::_SimdImpl::_S_##__name(__data(__x), _Arg2::_S_data(__y))};   \
+  }                                                                       =
     \
+template <typename _Up, typename _Tp, typename _Abi>                      =
     \
+  _GLIBCXX_SIMD_INTRINSIC _Math_return_type_t<                            =
     \
+    decltype(std::__name(                                                 =
     \
+      declval<double>(),                                                  =
     \
+      declval<enable_if_t<                                                =
     \
+	conjunction_v<                                                         \
+	  is_same<arg2_, _Tp>,                                                 \
+	  negation<is_same<__remove_cvref_t<_Up>, simd<_Tp, _Abi>>>,           \
+	  is_convertible<_Up, simd<_Tp, _Abi>>, is_floating_point<_Tp>>,       \
+	double>>())),                                                          \
+    _Tp, _Abi>                                                            =
     \
+  __name(_Up&& __xx, const simd<_Tp, _Abi>& __yy)                         =
     \
+  { return __name(simd<_Tp, _Abi>(static_cast<_Up&&>(__xx)), __yy); }
+
+// }}}
+// _GLIBCXX_SIMD_MATH_CALL3_ {{{
+#define _GLIBCXX_SIMD_MATH_CALL3_(__name, arg2_, arg3_)                   =
     \
+template <typename _Tp, typename _Abi, typename...,                       =
     \
+	  typename _Arg2 =3D _Extra_argument_type<arg2_, _Tp, _Abi>,             \
+	  typename _Arg3 =3D _Extra_argument_type<arg3_, _Tp, _Abi>,             \
+	  typename _R =3D _Math_return_type_t<                                   \
+	    decltype(std::__name(declval<double>(), _Arg2::declval(),          \
+				 _Arg3::declval())),                           \
+	    _Tp, _Abi>>                                                        \
+  enable_if_t<is_floating_point_v<_Tp>, _R>                               =
     \
+  __name(const simd<_Tp, _Abi>& __x, const typename _Arg2::type& __y,     =
     \
+	 const typename _Arg3::type& __z)                                      \
+  {                                                                       =
     \
+    return {__private_init,                                               =
     \
+	    _Abi::_SimdImpl::_S_##__name(__data(__x), _Arg2::_S_data(__y),     \
+					 _Arg3::_S_data(__z))};                \
+  }                                                                       =
     \
+template <                                                                =
     \
+  typename _T0, typename _T1, typename _T2, typename...,                  =
     \
+  typename _U0 =3D __remove_cvref_t<_T0>,                                 =
       \
+  typename _U1 =3D __remove_cvref_t<_T1>,                                 =
       \
+  typename _U2 =3D __remove_cvref_t<_T2>,                                 =
       \
+  typename _Simd =3D conditional_t<is_simd_v<_U1>, _U1, _U2>,             =
       \
+  typename =3D enable_if_t<conjunction_v<                                 =
       \
+    is_simd<_Simd>, is_convertible<_T0&&, _Simd>,                         =
     \
+    is_convertible<_T1&&, _Simd>, is_convertible<_T2&&, _Simd>,           =
     \
+    negation<conjunction<                                                 =
     \
+      is_simd<_U0>, is_floating_point<__value_type_or_identity_t<_U0>>>>>>=
>    \
+  _GLIBCXX_SIMD_INTRINSIC decltype(__name(declval<const _Simd&>(),        =
     \
+					  declval<const _Simd&>(),             \
+					  declval<const _Simd&>()))            \
+  __name(_T0&& __xx, _T1&& __yy, _T2&& __zz)                              =
     \
+  {                                                                       =
     \
+    return __name(_Simd(static_cast<_T0&&>(__xx)),                        =
     \
+		  _Simd(static_cast<_T1&&>(__yy)),                             \
+		  _Simd(static_cast<_T2&&>(__zz)));                            \
+  }
+
+// }}}
+// __cosSeries {{{
+template <typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE static simd<float, _Abi>
+  __cosSeries(const simd<float, _Abi>& __x)
+  {
+    const simd<float, _Abi> __x2 =3D __x * __x;
+    simd<float, _Abi> __y;
+    __y =3D 0x1.ap-16f;                  //  1/8!
+    __y =3D __y * __x2 - 0x1.6c1p-10f;   // -1/6!
+    __y =3D __y * __x2 + 0x1.555556p-5f; //  1/4!
+    return __y * (__x2 * __x2) - .5f * __x2 + 1.f;
+  }
+
+template <typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE static simd<double, _Abi>
+  __cosSeries(const simd<double, _Abi>& __x)
+  {
+    const simd<double, _Abi> __x2 =3D __x * __x;
+    simd<double, _Abi> __y;
+    __y =3D 0x1.AC00000000000p-45;              //  1/16!
+    __y =3D __y * __x2 - 0x1.9394000000000p-37; // -1/14!
+    __y =3D __y * __x2 + 0x1.1EED8C0000000p-29; //  1/12!
+    __y =3D __y * __x2 - 0x1.27E4FB7400000p-22; // -1/10!
+    __y =3D __y * __x2 + 0x1.A01A01A018000p-16; //  1/8!
+    __y =3D __y * __x2 - 0x1.6C16C16C16C00p-10; // -1/6!
+    __y =3D __y * __x2 + 0x1.5555555555554p-5;  //  1/4!
+    return (__y * __x2 - .5f) * __x2 + 1.f;
+  }
+
+// }}}
+// __sinSeries {{{
+template <typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE static simd<float, _Abi>
+  __sinSeries(const simd<float, _Abi>& __x)
+  {
+    const simd<float, _Abi> __x2 =3D __x * __x;
+    simd<float, _Abi> __y;
+    __y =3D -0x1.9CC000p-13f;            // -1/7!
+    __y =3D __y * __x2 + 0x1.111100p-7f; //  1/5!
+    __y =3D __y * __x2 - 0x1.555556p-3f; // -1/3!
+    return __y * (__x2 * __x) + __x;
+  }
+
+template <typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE static simd<double, _Abi>
+  __sinSeries(const simd<double, _Abi>& __x)
+  {
+    // __x  =3D [0, 0.7854 =3D pi/4]
+    // __x=C2=B2 =3D [0, 0.6169 =3D pi=C2=B2/8]
+    const simd<double, _Abi> __x2 =3D __x * __x;
+    simd<double, _Abi> __y;
+    __y =3D -0x1.ACF0000000000p-41;             // -1/15!
+    __y =3D __y * __x2 + 0x1.6124400000000p-33; //  1/13!
+    __y =3D __y * __x2 - 0x1.AE64567000000p-26; // -1/11!
+    __y =3D __y * __x2 + 0x1.71DE3A5540000p-19; //  1/9!
+    __y =3D __y * __x2 - 0x1.A01A01A01A000p-13; // -1/7!
+    __y =3D __y * __x2 + 0x1.1111111111110p-7;  //  1/5!
+    __y =3D __y * __x2 - 0x1.5555555555555p-3;  // -1/3!
+    return __y * (__x2 * __x) + __x;
+  }
+
+// }}}
+// __zero_low_bits {{{
+template <int _Bits, typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi>
+  __zero_low_bits(simd<_Tp, _Abi> __x)
+  {
+    const simd<_Tp, _Abi> __bitmask
+      =3D __bit_cast<_Tp>(~make_unsigned_t<__int_for_sizeof_t<_Tp>>() << _=
Bits);
+    return {__private_init,
+	    _Abi::_SimdImpl::_S_bit_and(__data(__x), __data(__bitmask))};
+  }
+
+// }}}
+// __fold_input {{{
+
+/**@internal
+ * Fold @p x into [-=C2=BC=CF=80, =C2=BC=CF=80] and remember the quadrant =
it came from:
+ * quadrant 0: [-=C2=BC=CF=80,  =C2=BC=CF=80]
+ * quadrant 1: [ =C2=BC=CF=80,  =C2=BE=CF=80]
+ * quadrant 2: [ =C2=BE=CF=80, 1=C2=BC=CF=80]
+ * quadrant 3: [1=C2=BC=CF=80, 1=C2=BE=CF=80]
+ *
+ * The algorithm determines `y` as the multiple `x - y * =C2=BC=CF=80 =3D =
[-=C2=BC=CF=80, =C2=BC=CF=80]`. Using
+ * a bitmask, `y` is reduced to `quadrant`. `y` can be calculated as
+ * ```
+ * y =3D trunc(x / =C2=BC=CF=80);
+ * y +=3D fmod(y, 2);
+ * ```
+ * This can be simplified by moving the (implicit) division by 2 into the
+ * truncation expression. The `+=3D fmod` effect can the be achieved by us=
ing
+ * rounding instead of truncation: `y =3D round(x / =C2=BD=CF=80) * 2`. If=
 precision allows,
+ * `2/=CF=80 * x` is better (faster).
+ */
+template <typename _Tp, typename _Abi>
+  struct _Folded
+  {
+    simd<_Tp, _Abi> _M_x;
+    rebind_simd_t<int, simd<_Tp, _Abi>> _M_quadrant;
+  };
+
+namespace __math_float {
+inline constexpr float __pi_over_4 =3D 0x1.921FB6p-1f; // =CF=80/4
+inline constexpr float __2_over_pi =3D 0x1.45F306p-1f; // 2/=CF=80
+inline constexpr float __pi_2_5bits0
+  =3D 0x1.921fc0p0f; // =CF=80/2, 5 0-bits (least significant)
+inline constexpr float __pi_2_5bits0_rem
+  =3D -0x1.5777a6p-21f; // =CF=80/2 - __pi_2_5bits0
+} // namespace __math_float
+namespace __math_double {
+inline constexpr double __pi_over_4 =3D 0x1.921fb54442d18p-1; // =CF=80/4
+inline constexpr double __2_over_pi =3D 0x1.45F306DC9C883p-1; // 2/=CF=80
+inline constexpr double __pi_2 =3D 0x1.921fb54442d18p0;       // =CF=80/2
+} // namespace __math_double
+
+template <typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _Folded<float, _Abi>
+  __fold_input(const simd<float, _Abi>& __x)
+  {
+    using _V =3D simd<float, _Abi>;
+    using _IV =3D rebind_simd_t<int, _V>;
+    using namespace __math_float;
+    _Folded<float, _Abi> __r;
+    __r._M_x =3D abs(__x);
+#if 0
+    // zero most mantissa bits:
+    constexpr float __1_over_pi =3D 0x1.45F306p-2f; // 1/=CF=80
+    const auto __y =3D (__r._M_x * __1_over_pi + 0x1.8p23f) - 0x1.8p23f;
+    // split =CF=80 into 4 parts, the first three with 13 trailing zeros (=
to make the
+    // following multiplications precise):
+    constexpr float __pi0 =3D 0x1.920000p1f;
+    constexpr float __pi1 =3D 0x1.fb4000p-11f;
+    constexpr float __pi2 =3D 0x1.444000p-23f;
+    constexpr float __pi3 =3D 0x1.68c234p-38f;
+    __r._M_x - __y*__pi0 - __y*__pi1 - __y*__pi2 - __y*__pi3
+#else
+    if (_GLIBCXX_SIMD_IS_UNLIKELY(all_of(__r._M_x < __pi_over_4)))
+      __r._M_quadrant =3D 0;
+    else if (_GLIBCXX_SIMD_IS_LIKELY(all_of(__r._M_x < 6 * __pi_over_4)))
+      {
+	const _V __y =3D nearbyint(__r._M_x * __2_over_pi);
+	__r._M_quadrant =3D static_simd_cast<_IV>(__y) & 3; // __y mod 4
+	__r._M_x -=3D __y * __pi_2_5bits0;
+	__r._M_x -=3D __y * __pi_2_5bits0_rem;
+      }
+    else
+      {
+	using __math_double::__2_over_pi;
+	using __math_double::__pi_2;
+	using _VD =3D rebind_simd_t<double, _V>;
+	_VD __xd =3D static_simd_cast<_VD>(__r._M_x);
+	_VD __y =3D nearbyint(__xd * __2_over_pi);
+	__r._M_quadrant =3D static_simd_cast<_IV>(__y) & 3; // =3D __y mod 4
+	__r._M_x =3D static_simd_cast<_V>(__xd - __y * __pi_2);
+      }
+#endif
+    return __r;
+  }
+
+template <typename _Abi>
+  _GLIBCXX_SIMD_ALWAYS_INLINE _Folded<double, _Abi>
+  __fold_input(const simd<double, _Abi>& __x)
+  {
+    using _V =3D simd<double, _Abi>;
+    using _IV =3D rebind_simd_t<int, _V>;
+    using namespace __math_double;
+
+    _Folded<double, _Abi> __r;
+    __r._M_x =3D abs(__x);
+    if (_GLIBCXX_SIMD_IS_UNLIKELY(all_of(__r._M_x < __pi_over_4)))
+      {
+	__r._M_quadrant =3D 0;
+	return __r;
+      }
+    const _V __y =3D nearbyint(__r._M_x / (2 * __pi_over_4));
+    __r._M_quadrant =3D static_simd_cast<_IV>(__y) & 3;
+
+    if (_GLIBCXX_SIMD_IS_LIKELY(all_of(__r._M_x < 1025 * __pi_over_4)))
+      {
+	// x - y * pi/2, y uses no more than 11 mantissa bits
+	__r._M_x -=3D __y * 0x1.921FB54443000p0;
+	__r._M_x -=3D __y * -0x1.73DCB3B39A000p-43;
+	__r._M_x -=3D __y * 0x1.45C06E0E68948p-86;
+      }
+    else if (_GLIBCXX_SIMD_IS_LIKELY(all_of(__y <=3D 0x1.0p30)))
+      {
+	// x - y * pi/2, y uses no more than 29 mantissa bits
+	__r._M_x -=3D __y * 0x1.921FB40000000p0;
+	__r._M_x -=3D __y * 0x1.4442D00000000p-24;
+	__r._M_x -=3D __y * 0x1.8469898CC5170p-48;
+      }
+    else
+      {
+	// x - y * pi/2, y may require all mantissa bits
+	const _V __y_hi =3D __zero_low_bits<26>(__y);
+	const _V __y_lo =3D __y - __y_hi;
+	const auto __pi_2_1 =3D 0x1.921FB50000000p0;
+	const auto __pi_2_2 =3D 0x1.110B460000000p-26;
+	const auto __pi_2_3 =3D 0x1.1A62630000000p-54;
+	const auto __pi_2_4 =3D 0x1.8A2E03707344Ap-81;
+	__r._M_x =3D __r._M_x - __y_hi * __pi_2_1
+		   - max(__y_hi * __pi_2_2, __y_lo * __pi_2_1)
+		   - min(__y_hi * __pi_2_2, __y_lo * __pi_2_1)
+		   - max(__y_hi * __pi_2_3, __y_lo * __pi_2_2)
+		   - min(__y_hi * __pi_2_3, __y_lo * __pi_2_2)
+		   - max(__y * __pi_2_4, __y_lo * __pi_2_3)
+		   - min(__y * __pi_2_4, __y_lo * __pi_2_3);
+      }
+    return __r;
+  }
+
+// }}}
+// __extract_exponent_as_int {{{
+template <typename _Tp, typename _Abi>
+  rebind_simd_t<int, simd<_Tp, _Abi>>
+  __extract_exponent_as_int(const simd<_Tp, _Abi>& __v)
+  {
+    using _Vp =3D simd<_Tp, _Abi>;
+    using _Up =3D make_unsigned_t<__int_for_sizeof_t<_Tp>>;
+    using namespace std::experimental::__float_bitwise_operators;
+    const _Vp __exponent_mask
+      =3D __infinity_v<_Tp>; // 0x7f800000 or 0x7ff0000000000000
+    return static_simd_cast<rebind_simd_t<int, _Vp>>(
+      __bit_cast<rebind_simd_t<_Up, _Vp>>(__v & __exponent_mask)
+      >> (__digits_v<_Tp> - 1));
+  }
+
+// }}}
+// __impl_or_fallback {{{
+template <typename ImplFun, typename FallbackFun, typename... _Args>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __impl_or_fallback_dispatch(int, ImplFun&& __impl_fun, FallbackFun&&,
+			      _Args&&... __args)
+    -> decltype(__impl_fun(static_cast<_Args&&>(__args)...))
+  { return __impl_fun(static_cast<_Args&&>(__args)...); }
+
+template <typename ImplFun, typename FallbackFun, typename... _Args>
+  inline auto
+  __impl_or_fallback_dispatch(float, ImplFun&&, FallbackFun&& __fallback_f=
un,
+			      _Args&&... __args)
+    -> decltype(__fallback_fun(static_cast<_Args&&>(__args)...))
+  { return __fallback_fun(static_cast<_Args&&>(__args)...); }
+
+template <typename... _Args>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __impl_or_fallback(_Args&&... __args)
+  {
+    return __impl_or_fallback_dispatch(int(), static_cast<_Args&&>(__args)=
=2E..);
+  }
+//}}}
+
+// trigonometric functions {{{
+_GLIBCXX_SIMD_MATH_CALL_(acos)
+_GLIBCXX_SIMD_MATH_CALL_(asin)
+_GLIBCXX_SIMD_MATH_CALL_(atan)
+_GLIBCXX_SIMD_MATH_CALL2_(atan2, _Tp)
+
+/*
+ * algorithm for sine and cosine:
+ *
+ * The result can be calculated with sine or cosine depending on the =CF=
=80/4 section
+ * the input is in. sine   =E2=89=88 __x + __x=C2=B3 cosine =E2=89=88 1 - =
__x=C2=B2
+ *
+ * sine:
+ * Map -__x to __x and invert the output
+ * Extend precision of __x - n * =CF=80/4 by calculating
+ * ((__x - n * p1) - n * p2) - n * p3 (p1 + p2 + p3 =3D =CF=80/4)
+ *
+ * Calculate Taylor series with tuned coefficients.
+ * Fix sign.
+ */
+// cos{{{
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  cos(const simd<_Tp, _Abi>& __x)
+  {
+    using _V =3D simd<_Tp, _Abi>;
+    if constexpr (__is_scalar_abi<_Abi>() || __is_fixed_size_abi_v<_Abi>)
+      return {__private_init, _Abi::_SimdImpl::_S_cos(__data(__x))};
+    else
+      {
+	if constexpr (is_same_v<_Tp, float>)
+	  if (_GLIBCXX_SIMD_IS_UNLIKELY(any_of(abs(__x) >=3D 393382)))
+	    return static_simd_cast<_V>(
+	      cos(static_simd_cast<rebind_simd_t<double, _V>>(__x)));
+
+	const auto __f =3D __fold_input(__x);
+	// quadrant | effect
+	//        0 | cosSeries, +
+	//        1 | sinSeries, -
+	//        2 | cosSeries, -
+	//        3 | sinSeries, +
+	using namespace std::experimental::__float_bitwise_operators;
+	const _V __sign_flip
+	  =3D _V(-0.f) & static_simd_cast<_V>((1 + __f._M_quadrant) << 30);
+
+	const auto __need_cos =3D (__f._M_quadrant & 1) =3D=3D 0;
+	if (_GLIBCXX_SIMD_IS_UNLIKELY(all_of(__need_cos)))
+	  return __sign_flip ^ __cosSeries(__f._M_x);
+	else if (_GLIBCXX_SIMD_IS_UNLIKELY(none_of(__need_cos)))
+	  return __sign_flip ^ __sinSeries(__f._M_x);
+	else // some_of(__need_cos)
+	  {
+	    _V __r =3D __sinSeries(__f._M_x);
+	    where(__need_cos.__cvt(), __r) =3D __cosSeries(__f._M_x);
+	    return __r ^ __sign_flip;
+	  }
+      }
+  }
+
+template <typename _Tp>
+  _GLIBCXX_SIMD_ALWAYS_INLINE
+  enable_if_t<is_floating_point<_Tp>::value, simd<_Tp, simd_abi::scalar>>
+  cos(simd<_Tp, simd_abi::scalar> __x)
+  { return std::cos(__data(__x)); }
+
+//}}}
+// sin{{{
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  sin(const simd<_Tp, _Abi>& __x)
+  {
+    using _V =3D simd<_Tp, _Abi>;
+    if constexpr (__is_scalar_abi<_Abi>() || __is_fixed_size_abi_v<_Abi>)
+      return {__private_init, _Abi::_SimdImpl::_S_sin(__data(__x))};
+    else
+      {
+	if constexpr (is_same_v<_Tp, float>)
+	  if (_GLIBCXX_SIMD_IS_UNLIKELY(any_of(abs(__x) >=3D 527449)))
+	    return static_simd_cast<_V>(
+	      sin(static_simd_cast<rebind_simd_t<double, _V>>(__x)));
+
+	const auto __f =3D __fold_input(__x);
+	// quadrant | effect
+	//        0 | sinSeries
+	//        1 | cosSeries
+	//        2 | sinSeries, sign flip
+	//        3 | cosSeries, sign flip
+	using namespace std::experimental::__float_bitwise_operators;
+	const auto __sign_flip
+	  =3D (__x ^ static_simd_cast<_V>(1 - __f._M_quadrant)) & _V(_Tp(-0.));
+
+	const auto __need_sin =3D (__f._M_quadrant & 1) =3D=3D 0;
+	if (_GLIBCXX_SIMD_IS_UNLIKELY(all_of(__need_sin)))
+	  return __sign_flip ^ __sinSeries(__f._M_x);
+	else if (_GLIBCXX_SIMD_IS_UNLIKELY(none_of(__need_sin)))
+	  return __sign_flip ^ __cosSeries(__f._M_x);
+	else // some_of(__need_sin)
+	  {
+	    _V __r =3D __cosSeries(__f._M_x);
+	    where(__need_sin.__cvt(), __r) =3D __sinSeries(__f._M_x);
+	    return __sign_flip ^ __r;
+	  }
+      }
+  }
+
+template <typename _Tp>
+  _GLIBCXX_SIMD_ALWAYS_INLINE
+  enable_if_t<is_floating_point<_Tp>::value, simd<_Tp, simd_abi::scalar>>
+  sin(simd<_Tp, simd_abi::scalar> __x)
+  { return std::sin(__data(__x)); }
+
+//}}}
+_GLIBCXX_SIMD_MATH_CALL_(tan)
+_GLIBCXX_SIMD_MATH_CALL_(acosh)
+_GLIBCXX_SIMD_MATH_CALL_(asinh)
+_GLIBCXX_SIMD_MATH_CALL_(atanh)
+_GLIBCXX_SIMD_MATH_CALL_(cosh)
+_GLIBCXX_SIMD_MATH_CALL_(sinh)
+_GLIBCXX_SIMD_MATH_CALL_(tanh)
+// }}}
+// exponential functions {{{
+_GLIBCXX_SIMD_MATH_CALL_(exp)
+_GLIBCXX_SIMD_MATH_CALL_(exp2)
+_GLIBCXX_SIMD_MATH_CALL_(expm1)
+
+// }}}
+// frexp {{{
+#if _GLIBCXX_SIMD_X86INTRIN
+template <typename _Tp, size_t _Np>
+  _SimdWrapper<_Tp, _Np>
+  __getexp(_SimdWrapper<_Tp, _Np> __x)
+  {
+    if constexpr (__have_avx512vl && __is_sse_ps<_Tp, _Np>())
+      return __auto_bitcast(_mm_getexp_ps(__to_intrin(__x)));
+    else if constexpr (__have_avx512f && __is_sse_ps<_Tp, _Np>())
+      return __auto_bitcast(_mm512_getexp_ps(__auto_bitcast(__to_intrin(__=
x))));
+    else if constexpr (__have_avx512vl && __is_sse_pd<_Tp, _Np>())
+      return _mm_getexp_pd(__x);
+    else if constexpr (__have_avx512f && __is_sse_pd<_Tp, _Np>())
+      return __lo128(_mm512_getexp_pd(__auto_bitcast(__x)));
+    else if constexpr (__have_avx512vl && __is_avx_ps<_Tp, _Np>())
+      return _mm256_getexp_ps(__x);
+    else if constexpr (__have_avx512f && __is_avx_ps<_Tp, _Np>())
+      return __lo256(_mm512_getexp_ps(__auto_bitcast(__x)));
+    else if constexpr (__have_avx512vl && __is_avx_pd<_Tp, _Np>())
+      return _mm256_getexp_pd(__x);
+    else if constexpr (__have_avx512f && __is_avx_pd<_Tp, _Np>())
+      return __lo256(_mm512_getexp_pd(__auto_bitcast(__x)));
+    else if constexpr (__is_avx512_ps<_Tp, _Np>())
+      return _mm512_getexp_ps(__x);
+    else if constexpr (__is_avx512_pd<_Tp, _Np>())
+      return _mm512_getexp_pd(__x);
+    else
+      __assert_unreachable<_Tp>();
+  }
+
+template <typename _Tp, size_t _Np>
+  _SimdWrapper<_Tp, _Np>
+  __getmant_avx512(_SimdWrapper<_Tp, _Np> __x)
+  {
+    if constexpr (__have_avx512vl && __is_sse_ps<_Tp, _Np>())
+      return __auto_bitcast(_mm_getmant_ps(__to_intrin(__x), _MM_MANT_NORM=
_p5_1,
+					   _MM_MANT_SIGN_src));
+    else if constexpr (__have_avx512f && __is_sse_ps<_Tp, _Np>())
+      return __auto_bitcast(_mm512_getmant_ps(__auto_bitcast(__to_intrin(_=
_x)),
+					      _MM_MANT_NORM_p5_1,
+					      _MM_MANT_SIGN_src));
+    else if constexpr (__have_avx512vl && __is_sse_pd<_Tp, _Np>())
+      return _mm_getmant_pd(__x, _MM_MANT_NORM_p5_1, _MM_MANT_SIGN_src);
+    else if constexpr (__have_avx512f && __is_sse_pd<_Tp, _Np>())
+      return __lo128(_mm512_getmant_pd(__auto_bitcast(__x), _MM_MANT_NORM_=
p5_1,
+				       _MM_MANT_SIGN_src));
+    else if constexpr (__have_avx512vl && __is_avx_ps<_Tp, _Np>())
+      return _mm256_getmant_ps(__x, _MM_MANT_NORM_p5_1, _MM_MANT_SIGN_src);
+    else if constexpr (__have_avx512f && __is_avx_ps<_Tp, _Np>())
+      return __lo256(_mm512_getmant_ps(__auto_bitcast(__x), _MM_MANT_NORM_=
p5_1,
+				       _MM_MANT_SIGN_src));
+    else if constexpr (__have_avx512vl && __is_avx_pd<_Tp, _Np>())
+      return _mm256_getmant_pd(__x, _MM_MANT_NORM_p5_1, _MM_MANT_SIGN_src);
+    else if constexpr (__have_avx512f && __is_avx_pd<_Tp, _Np>())
+      return __lo256(_mm512_getmant_pd(__auto_bitcast(__x), _MM_MANT_NORM_=
p5_1,
+				       _MM_MANT_SIGN_src));
+    else if constexpr (__is_avx512_ps<_Tp, _Np>())
+      return _mm512_getmant_ps(__x, _MM_MANT_NORM_p5_1, _MM_MANT_SIGN_src);
+    else if constexpr (__is_avx512_pd<_Tp, _Np>())
+      return _mm512_getmant_pd(__x, _MM_MANT_NORM_p5_1, _MM_MANT_SIGN_src);
+    else
+      __assert_unreachable<_Tp>();
+  }
+#endif // _GLIBCXX_SIMD_X86INTRIN
+
+/**
+ * splits @p __v into exponent and mantissa, the sign is kept with the man=
tissa
+ *
+ * The return value will be in the range [0.5, 1.0[
+ * The @p __e value will be an integer defining the power-of-two exponent
+ */
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  frexp(const simd<_Tp, _Abi>& __x, _Samesize<int, simd<_Tp, _Abi>>* __exp)
+  {
+    if constexpr (simd_size_v<_Tp, _Abi> =3D=3D 1)
+      {
+	int __tmp;
+	const auto __r =3D std::frexp(__x[0], &__tmp);
+	(*__exp)[0] =3D __tmp;
+	return __r;
+      }
+    else if constexpr (__is_fixed_size_abi_v<_Abi>)
+      {
+	return {__private_init,
+		_Abi::_SimdImpl::_S_frexp(__data(__x), __data(*__exp))};
+#if _GLIBCXX_SIMD_X86INTRIN
+      }
+    else if constexpr (__have_avx512f)
+      {
+	constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	constexpr size_t _NI =3D _Np < 4 ? 4 : _Np;
+	const auto __v =3D __data(__x);
+	const auto __isnonzero
+	  =3D _Abi::_SimdImpl::_S_isnonzerovalue_mask(__v._M_data);
+	const _SimdWrapper<int, _NI> __exp_plus1
+	  =3D 1 + __convert<_SimdWrapper<int, _NI>>(__getexp(__v))._M_data;
+	const _SimdWrapper<int, _Np> __e =3D __wrapper_bitcast<int, _Np>(
+	  _Abi::_CommonImpl::_S_blend(_SimdWrapper<bool, _NI>(__isnonzero),
+				      _SimdWrapper<int, _NI>(), __exp_plus1));
+	simd_abi::deduce_t<int, _Np>::_CommonImpl::_S_store(__e, __exp);
+	return {__private_init,
+		_Abi::_CommonImpl::_S_blend(_SimdWrapper<bool, _Np>(
+					      __isnonzero),
+					    __v, __getmant_avx512(__v))};
+#endif // _GLIBCXX_SIMD_X86INTRIN
+      }
+    else
+      {
+	// fallback implementation
+	static_assert(sizeof(_Tp) =3D=3D 4 || sizeof(_Tp) =3D=3D 8);
+	using _V =3D simd<_Tp, _Abi>;
+	using _IV =3D rebind_simd_t<int, _V>;
+	using namespace std::experimental::__proposed;
+	using namespace std::experimental::__float_bitwise_operators;
+
+	constexpr int __exp_adjust =3D sizeof(_Tp) =3D=3D 4 ? 0x7e : 0x3fe;
+	constexpr int __exp_offset =3D sizeof(_Tp) =3D=3D 4 ? 0x70 : 0x200;
+	constexpr _Tp __subnorm_scale =3D sizeof(_Tp) =3D=3D 4 ? 0x1p112 : 0x1p51=
2;
+	_GLIBCXX_SIMD_USE_CONSTEXPR_API _V __exponent_mask
+	  =3D __infinity_v<_Tp>; // 0x7f800000 or 0x7ff0000000000000
+	_GLIBCXX_SIMD_USE_CONSTEXPR_API _V __p5_1_exponent
+	  =3D -(2 - __epsilon_v<_Tp>) / 2; // 0xbf7fffff or 0xbfefffffffffffff
+
+	_V __mant =3D __p5_1_exponent & (__exponent_mask | __x); // +/-[.5, 1)
+	const _IV __exponent_bits =3D __extract_exponent_as_int(__x);
+	if (_GLIBCXX_SIMD_IS_LIKELY(all_of(isnormal(__x))))
+	  {
+	    *__exp
+	      =3D simd_cast<_Samesize<int, _V>>(__exponent_bits - __exp_adjust);
+	    return __mant;
+	  }
+
+#if __FINITE_MATH_ONLY__
+	// at least one element of __x is 0 or subnormal, the rest is normal
+	// (inf and NaN are excluded by -ffinite-math-only)
+	const auto __iszero_inf_nan =3D __x =3D=3D 0;
+#else
+	const auto __as_int
+	  =3D __bit_cast<rebind_simd_t<__int_for_sizeof_t<_Tp>, _V>>(abs(__x));
+	const auto __inf
+	  =3D __bit_cast<rebind_simd_t<__int_for_sizeof_t<_Tp>, _V>>(
+	    _V(__infinity_v<_Tp>));
+	const auto __iszero_inf_nan =3D static_simd_cast<typename _V::mask_type>(
+	  __as_int =3D=3D 0 || __as_int >=3D __inf);
+#endif
+
+	const _V __scaled_subnormal =3D __x * __subnorm_scale;
+	const _V __mant_subnormal
+	  =3D __p5_1_exponent & (__exponent_mask | __scaled_subnormal);
+	where(!isnormal(__x), __mant) =3D __mant_subnormal;
+	where(__iszero_inf_nan, __mant) =3D __x;
+	_IV __e =3D __extract_exponent_as_int(__scaled_subnormal);
+	using _MaskType =3D
+	  typename conditional_t<sizeof(typename _V::value_type) =3D=3D sizeof(in=
t),
+				 _V, _IV>::mask_type;
+	const _MaskType __value_isnormal =3D isnormal(__x).__cvt();
+	where(__value_isnormal.__cvt(), __e) =3D __exponent_bits;
+	static_assert(sizeof(_IV) =3D=3D sizeof(__value_isnormal));
+	const _IV __offset
+	  =3D (__bit_cast<_IV>(__value_isnormal) & _IV(__exp_adjust))
+	    | (__bit_cast<_IV>(static_simd_cast<_MaskType>(__exponent_bits =3D=3D=
 0)
+			       & static_simd_cast<_MaskType>(__x !=3D 0))
+	       & _IV(__exp_adjust + __exp_offset));
+	*__exp =3D simd_cast<_Samesize<int, _V>>(__e - __offset);
+	return __mant;
+      }
+  }
+
+// }}}
+_GLIBCXX_SIMD_MATH_CALL2_(ldexp, int)
+_GLIBCXX_SIMD_MATH_CALL_(ilogb)
+
+// logarithms {{{
+_GLIBCXX_SIMD_MATH_CALL_(log)
+_GLIBCXX_SIMD_MATH_CALL_(log10)
+_GLIBCXX_SIMD_MATH_CALL_(log1p)
+_GLIBCXX_SIMD_MATH_CALL_(log2)
+
+//}}}
+// logb{{{
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point<_Tp>::value, simd<_Tp, _Abi>>
+  logb(const simd<_Tp, _Abi>& __x)
+  {
+    constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+    if constexpr (_Np =3D=3D 1)
+      return std::logb(__x[0]);
+    else if constexpr (__is_fixed_size_abi_v<_Abi>)
+      {
+	return {__private_init,
+		__data(__x)._M_apply_per_chunk([](auto __impl, auto __xx) {
+		  using _V =3D typename decltype(__impl)::simd_type;
+		  return __data(
+		    std::experimental::logb(_V(__private_init, __xx)));
+		})};
+      }
+#if _GLIBCXX_SIMD_X86INTRIN // {{{
+    else if constexpr (__have_avx512vl && __is_sse_ps<_Tp, _Np>())
+      return {__private_init,
+	      __auto_bitcast(_mm_getexp_ps(__to_intrin(__as_vector(__x))))};
+    else if constexpr (__have_avx512vl && __is_sse_pd<_Tp, _Np>())
+      return {__private_init, _mm_getexp_pd(__data(__x))};
+    else if constexpr (__have_avx512vl && __is_avx_ps<_Tp, _Np>())
+      return {__private_init, _mm256_getexp_ps(__data(__x))};
+    else if constexpr (__have_avx512vl && __is_avx_pd<_Tp, _Np>())
+      return {__private_init, _mm256_getexp_pd(__data(__x))};
+    else if constexpr (__have_avx512f && __is_avx_ps<_Tp, _Np>())
+      return {__private_init,
+	      __lo256(_mm512_getexp_ps(__auto_bitcast(__data(__x))))};
+    else if constexpr (__have_avx512f && __is_avx_pd<_Tp, _Np>())
+      return {__private_init,
+	      __lo256(_mm512_getexp_pd(__auto_bitcast(__data(__x))))};
+    else if constexpr (__is_avx512_ps<_Tp, _Np>())
+      return {__private_init, _mm512_getexp_ps(__data(__x))};
+    else if constexpr (__is_avx512_pd<_Tp, _Np>())
+      return {__private_init, _mm512_getexp_pd(__data(__x))};
+#endif // _GLIBCXX_SIMD_X86INTRIN }}}
+    else
+      {
+	using _V =3D simd<_Tp, _Abi>;
+	using namespace std::experimental::__proposed;
+	auto __is_normal =3D isnormal(__x);
+
+	// work on abs(__x) to reflect the return value on Linux for negative
+	// inputs (domain-error =3D> implementation-defined value is returned)
+	const _V abs_x =3D abs(__x);
+
+	// __exponent(__x) returns the exponent value (bias removed) as
+	// simd<_Up> with integral _Up
+	auto&& __exponent =3D [](const _V& __v) {
+	  using namespace std::experimental::__proposed;
+	  using _IV =3D rebind_simd_t<
+	    conditional_t<sizeof(_Tp) =3D=3D sizeof(_LLong), _LLong, int>, _V>;
+	  return (__bit_cast<_IV>(__v) >> (__digits_v<_Tp> - 1))
+		 - (__max_exponent_v<_Tp> - 1);
+	};
+	_V __r =3D static_simd_cast<_V>(__exponent(abs_x));
+	if (_GLIBCXX_SIMD_IS_LIKELY(all_of(__is_normal)))
+	  // without corner cases (nan, inf, subnormal, zero) we have our
+	  // answer:
+	  return __r;
+	const auto __is_zero =3D __x =3D=3D 0;
+	const auto __is_nan =3D isnan(__x);
+	const auto __is_inf =3D isinf(__x);
+	where(__is_zero, __r) =3D -__infinity_v<_Tp>;
+	where(__is_nan, __r) =3D __x;
+	where(__is_inf, __r) =3D __infinity_v<_Tp>;
+	__is_normal |=3D __is_zero || __is_nan || __is_inf;
+	if (all_of(__is_normal))
+	  // at this point everything but subnormals is handled
+	  return __r;
+	// subnormals repeat the exponent extraction after multiplication of the
+	// input with __a floating point value that has 112 (0x70) in its exponent
+	// (not too big for sp and large enough for dp)
+	const _V __scaled =3D abs_x * _Tp(0x1p112);
+	_V __scaled_exp =3D static_simd_cast<_V>(__exponent(__scaled) - 112);
+	where(__is_normal, __scaled_exp) =3D __r;
+	return __scaled_exp;
+      }
+  }
+
+//}}}
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  modf(const simd<_Tp, _Abi>& __x, simd<_Tp, _Abi>* __iptr)
+  {
+    if constexpr (__is_scalar_abi<_Abi>()
+		  || (__is_fixed_size_abi_v<
+			_Abi> && simd_size_v<_Tp, _Abi> =3D=3D 1))
+      {
+	_Tp __tmp;
+	_Tp __r =3D std::modf(__x[0], &__tmp);
+	__iptr[0] =3D __tmp;
+	return __r;
+      }
+    else
+      {
+	const auto __integral =3D trunc(__x);
+	*__iptr =3D __integral;
+	auto __r =3D __x - __integral;
+#if !__FINITE_MATH_ONLY__
+	where(isinf(__x), __r) =3D _Tp();
+#endif
+	return copysign(__r, __x);
+      }
+  }
+
+_GLIBCXX_SIMD_MATH_CALL2_(scalbn, int)
+_GLIBCXX_SIMD_MATH_CALL2_(scalbln, long)
+
+_GLIBCXX_SIMD_MATH_CALL_(cbrt)
+
+_GLIBCXX_SIMD_MATH_CALL_(abs)
+_GLIBCXX_SIMD_MATH_CALL_(fabs)
+
+// [parallel.simd.math] only asks for is_floating_point_v<_Tp> and forgot =
to
+// allow signed integral _Tp
+template <typename _Tp, typename _Abi>
+  enable_if_t<!is_floating_point_v<_Tp> && is_signed_v<_Tp>, simd<_Tp, _Ab=
i>>
+  abs(const simd<_Tp, _Abi>& __x)
+  { return {__private_init, _Abi::_SimdImpl::_S_abs(__data(__x))}; }
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<!is_floating_point_v<_Tp> && is_signed_v<_Tp>, simd<_Tp, _Ab=
i>>
+  fabs(const simd<_Tp, _Abi>& __x)
+  { return {__private_init, _Abi::_SimdImpl::_S_abs(__data(__x))}; }
+
+// the following are overloads for functions in <cstdlib> and not covered =
by
+// [parallel.simd.math]. I don't see much value in making them work, though
+/*
+template <typename _Abi> simd<long, _Abi> labs(const simd<long, _Abi> &__x)
+{ return {__private_init, _Abi::_SimdImpl::abs(__data(__x))}; }
+
+template <typename _Abi> simd<long long, _Abi> llabs(const simd<long long,=
 _Abi>
+&__x)
+{ return {__private_init, _Abi::_SimdImpl::abs(__data(__x))}; }
+*/
+
+#define _GLIBCXX_SIMD_CVTING2(_NAME)                                      =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const simd<_Tp, _Abi>& __x, const __type_identity_t<simd<_Tp, _Abi>>& =
__y) \
+  {                                                                       =
     \
+    return _NAME(__x, __y);                                               =
     \
+  }                                                                       =
     \
+                                                                          =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const __type_identity_t<simd<_Tp, _Abi>>& __x, const simd<_Tp, _Abi>& =
__y) \
+  {                                                                       =
     \
+    return _NAME(__x, __y);                                               =
     \
+  }
+
+#define _GLIBCXX_SIMD_CVTING3(_NAME)                                      =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const __type_identity_t<simd<_Tp, _Abi>>& __x, const simd<_Tp, _Abi>& =
__y, \
+    const simd<_Tp, _Abi>& __z)                                           =
     \
+  {                                                                       =
     \
+    return _NAME(__x, __y, __z);                                          =
     \
+  }                                                                       =
     \
+                                                                          =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const simd<_Tp, _Abi>& __x, const __type_identity_t<simd<_Tp, _Abi>>& =
__y, \
+    const simd<_Tp, _Abi>& __z)                                           =
     \
+  {                                                                       =
     \
+    return _NAME(__x, __y, __z);                                          =
     \
+  }                                                                       =
     \
+                                                                          =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const simd<_Tp, _Abi>& __x, const simd<_Tp, _Abi>& __y,               =
     \
+    const __type_identity_t<simd<_Tp, _Abi>>& __z)                        =
     \
+  {                                                                       =
     \
+    return _NAME(__x, __y, __z);                                          =
     \
+  }                                                                       =
     \
+                                                                          =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const simd<_Tp, _Abi>& __x, const __type_identity_t<simd<_Tp, _Abi>>& =
__y, \
+    const __type_identity_t<simd<_Tp, _Abi>>& __z)                        =
     \
+  {                                                                       =
     \
+    return _NAME(__x, __y, __z);                                          =
     \
+  }                                                                       =
     \
+                                                                          =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const __type_identity_t<simd<_Tp, _Abi>>& __x, const simd<_Tp, _Abi>& =
__y, \
+    const __type_identity_t<simd<_Tp, _Abi>>& __z)                        =
     \
+  {                                                                       =
     \
+    return _NAME(__x, __y, __z);                                          =
     \
+  }                                                                       =
     \
+                                                                          =
     \
+template <typename _Tp, typename _Abi>                                    =
     \
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi> _NAME(                          =
     \
+    const __type_identity_t<simd<_Tp, _Abi>>& __x,                        =
     \
+    const __type_identity_t<simd<_Tp, _Abi>>& __y, const simd<_Tp, _Abi>& =
__z) \
+  {                                                                       =
     \
+    return _NAME(__x, __y, __z);                                          =
     \
+  }
+
+template <typename _R, typename _ToApply, typename _Tp, typename... _Tps>
+  _GLIBCXX_SIMD_INTRINSIC _R
+  __fixed_size_apply(_ToApply&& __apply, const _Tp& __arg0,
+		     const _Tps&... __args)
+  {
+    return {__private_init,
+	    __data(__arg0)._M_apply_per_chunk(
+	      [&](auto __impl, const auto&... __inner) {
+		using _V =3D typename decltype(__impl)::simd_type;
+		return __data(__apply(_V(__private_init, __inner)...));
+	      },
+	      __data(__args)...)};
+  }
+
+template <typename _VV>
+  __remove_cvref_t<_VV>
+  __hypot(_VV __x, _VV __y)
+  {
+    using _V =3D __remove_cvref_t<_VV>;
+    using _Tp =3D typename _V::value_type;
+    if constexpr (_V::size() =3D=3D 1)
+      return std::hypot(_Tp(__x[0]), _Tp(__y[0]));
+    else if constexpr (__is_fixed_size_abi_v<typename _V::abi_type>)
+      {
+	return __fixed_size_apply<_V>([](auto __a,
+					 auto __b) { return hypot(__a, __b); },
+				      __x, __y);
+      }
+    else
+      {
+	// A simple solution for _Tp =3D=3D float would be to cast to double and
+	// simply calculate sqrt(x=C2=B2+y=C2=B2) as it can't over-/underflow any=
more with
+	// dp. It still needs the Annex F fixups though and isn't faster on
+	// Skylake-AVX512 (not even for SSE and AVX vectors, and really bad for
+	// AVX-512).
+	using namespace __float_bitwise_operators;
+	_V __absx =3D abs(__x);          // no error
+	_V __absy =3D abs(__y);          // no error
+	_V __hi =3D max(__absx, __absy); // no error
+	_V __lo =3D min(__absy, __absx); // no error
+
+	// round __hi down to the next power-of-2:
+	_GLIBCXX_SIMD_USE_CONSTEXPR_API _V __inf(__infinity_v<_Tp>);
+
+#ifndef __FAST_MATH__
+	if constexpr (__have_neon && !__have_neon_a32)
+	  { // With ARMv7 NEON, we have no subnormals and must use slightly
+	    // different strategy
+	    const _V __hi_exp =3D __hi & __inf;
+	    _V __scale_back =3D __hi_exp;
+	    // For large exponents (max & max/2) the inversion comes too close
+	    // to subnormals. Subtract 3 from the exponent:
+	    where(__hi_exp > 1, __scale_back) =3D __hi_exp * _Tp(0.125);
+	    // Invert and adjust for the off-by-one error of inversion via xor:
+	    const _V __scale =3D (__scale_back ^ __inf) * _Tp(.5);
+	    const _V __h1 =3D __hi * __scale;
+	    const _V __l1 =3D __lo * __scale;
+	    _V __r =3D __scale_back * sqrt(__h1 * __h1 + __l1 * __l1);
+	    // Fix up hypot(0, 0) to not be NaN:
+	    where(__hi =3D=3D 0, __r) =3D 0;
+	    return __r;
+	  }
+#endif
+
+#ifdef __FAST_MATH__
+	// With fast-math, ignore precision of subnormals and inputs from
+	// __finite_max_v/2 to __finite_max_v. This removes all
+	// branching/masking.
+	if constexpr (true)
+#else
+	if (_GLIBCXX_SIMD_IS_LIKELY(all_of(isnormal(__x))
+				    && all_of(isnormal(__y))))
+#endif
+	  {
+	    const _V __hi_exp =3D __hi & __inf;
+	    //((__hi + __hi) & __inf) ^ __inf almost works for computing
+	    //__scale,
+	    // except when (__hi + __hi) & __inf =3D=3D __inf, in which case __sc=
ale
+	    // becomes 0 (should be min/2 instead) and thus loses the
+	    // information from __lo.
+#ifdef __FAST_MATH__
+	    using _Ip =3D __int_for_sizeof_t<_Tp>;
+	    using _IV =3D rebind_simd_t<_Ip, _V>;
+	    const auto __as_int =3D __bit_cast<_IV>(__hi_exp);
+	    const _V __scale
+	      =3D __bit_cast<_V>(2 * __bit_cast<_Ip>(_Tp(1)) - __as_int);
+#else
+	    const _V __scale =3D (__hi_exp ^ __inf) * _Tp(.5);
+#endif
+	    _GLIBCXX_SIMD_USE_CONSTEXPR_API _V __mant_mask
+	      =3D __norm_min_v<_Tp> - __denorm_min_v<_Tp>;
+	    const _V __h1 =3D (__hi & __mant_mask) | _V(1);
+	    const _V __l1 =3D __lo * __scale;
+	    return __hi_exp * sqrt(__h1 * __h1 + __l1 * __l1);
+	  }
+	else
+	  {
+	    // slower path to support subnormals
+	    // if __hi is subnormal, avoid scaling by inf & final mul by 0
+	    // (which yields NaN) by using min()
+	    _V __scale =3D _V(1 / __norm_min_v<_Tp>);
+	    // invert exponent w/o error and w/o using the slow divider unit:
+	    // xor inverts the exponent but off by 1. Multiplication with .5
+	    // adjusts for the discrepancy.
+	    where(__hi >=3D __norm_min_v<_Tp>, __scale)
+	      =3D ((__hi & __inf) ^ __inf) * _Tp(.5);
+	    // adjust final exponent for subnormal inputs
+	    _V __hi_exp =3D __norm_min_v<_Tp>;
+	    where(__hi >=3D __norm_min_v<_Tp>, __hi_exp)
+	      =3D __hi & __inf;         // no error
+	    _V __h1 =3D __hi * __scale; // no error
+	    _V __l1 =3D __lo * __scale; // no error
+
+	    // sqrt(x=C2=B2+y=C2=B2) =3D e*sqrt((x/e)=C2=B2+(y/e)=C2=B2):
+	    // this ensures no overflow in the argument to sqrt
+	    _V __r =3D __hi_exp * sqrt(__h1 * __h1 + __l1 * __l1);
+#ifdef __STDC_IEC_559__
+	    // fixup for Annex F requirements
+	    // the naive fixup goes like this:
+	    //
+	    // where(__l1 =3D=3D 0, __r)                      =3D __hi;
+	    // where(isunordered(__x, __y), __r)          =3D __quiet_NaN_v<_Tp>;
+	    // where(isinf(__absx) || isinf(__absy), __r) =3D __inf;
+	    //
+	    // The fixup can be prepared in parallel with the sqrt, requiring a
+	    // single blend step after hi_exp * sqrt, reducing latency and
+	    // throughput:
+	    _V __fixup =3D __hi; // __lo =3D=3D 0
+	    where(isunordered(__x, __y), __fixup) =3D __quiet_NaN_v<_Tp>;
+	    where(isinf(__absx) || isinf(__absy), __fixup) =3D __inf;
+	    where(!(__lo =3D=3D 0 || isunordered(__x, __y)
+		    || (isinf(__absx) || isinf(__absy))),
+		  __fixup)
+	      =3D __r;
+	    __r =3D __fixup;
+#endif
+	    return __r;
+	  }
+      }
+  }
+
+template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi>
+  hypot(const simd<_Tp, _Abi>& __x, const simd<_Tp, _Abi>& __y)
+  {
+    return __hypot<conditional_t<__is_fixed_size_abi_v<_Abi>,
+				 const simd<_Tp, _Abi>&, simd<_Tp, _Abi>>>(__x,
+									   __y);
+  }
+
+_GLIBCXX_SIMD_CVTING2(hypot)
+
+  template <typename _VV>
+  __remove_cvref_t<_VV>
+  __hypot(_VV __x, _VV __y, _VV __z)
+  {
+    using _V =3D __remove_cvref_t<_VV>;
+    using _Abi =3D typename _V::abi_type;
+    using _Tp =3D typename _V::value_type;
+    /* FIXME: enable after PR77776 is resolved
+    if constexpr (_V::size() =3D=3D 1)
+      return std::hypot(_Tp(__x[0]), _Tp(__y[0]), _Tp(__z[0]));
+    else
+    */
+    if constexpr (__is_fixed_size_abi_v<_Abi> && _V::size() > 1)
+      {
+	return __fixed_size_apply<simd<_Tp, _Abi>>(
+	  [](auto __a, auto __b, auto __c) { return hypot(__a, __b, __c); },
+	  __x, __y, __z);
+      }
+    else
+      {
+	using namespace __float_bitwise_operators;
+	const _V __absx =3D abs(__x);                 // no error
+	const _V __absy =3D abs(__y);                 // no error
+	const _V __absz =3D abs(__z);                 // no error
+	_V __hi =3D max(max(__absx, __absy), __absz); // no error
+	_V __l0 =3D min(__absz, max(__absx, __absy)); // no error
+	_V __l1 =3D min(__absy, __absx);              // no error
+	if constexpr (__digits_v<_Tp> =3D=3D 64 && __max_exponent_v<_Tp> =3D=3D 0=
x4000
+		      && __min_exponent_v<_Tp> =3D=3D -0x3FFD && _V::size() =3D=3D 1)
+	  { // Seems like x87 fp80, where bit 63 is always 1 unless subnormal or
+	    // NaN. In this case the bit-tricks don't work, they require IEC559
+	    // binary32 or binary64 format.
+#ifdef __STDC_IEC_559__
+	    // fixup for Annex F requirements
+	    if (isinf(__absx[0]) || isinf(__absy[0]) || isinf(__absz[0]))
+	      return __infinity_v<_Tp>;
+	    else if (isunordered(__absx[0], __absy[0] + __absz[0]))
+	      return __quiet_NaN_v<_Tp>;
+	    else if (__l0[0] =3D=3D 0 && __l1[0] =3D=3D 0)
+	      return __hi;
+#endif
+	    _V __hi_exp =3D __hi;
+	    const _ULLong __tmp =3D 0x8000'0000'0000'0000ull;
+	    __builtin_memcpy(&__data(__hi_exp), &__tmp, 8);
+	    const _V __scale =3D 1 / __hi_exp;
+	    __hi *=3D __scale;
+	    __l0 *=3D __scale;
+	    __l1 *=3D __scale;
+	    return __hi_exp * sqrt((__l0 * __l0 + __l1 * __l1) + __hi * __hi);
+	  }
+	else
+	  {
+	    // round __hi down to the next power-of-2:
+	    _GLIBCXX_SIMD_USE_CONSTEXPR_API _V __inf(__infinity_v<_Tp>);
+
+#ifndef __FAST_MATH__
+	    if constexpr (_V::size() > 1 && __have_neon && !__have_neon_a32)
+	      { // With ARMv7 NEON, we have no subnormals and must use slightly
+		// different strategy
+		const _V __hi_exp =3D __hi & __inf;
+		_V __scale_back =3D __hi_exp;
+		// For large exponents (max & max/2) the inversion comes too
+		// close to subnormals. Subtract 3 from the exponent:
+		where(__hi_exp > 1, __scale_back) =3D __hi_exp * _Tp(0.125);
+		// Invert and adjust for the off-by-one error of inversion via
+		// xor:
+		const _V __scale =3D (__scale_back ^ __inf) * _Tp(.5);
+		const _V __h1 =3D __hi * __scale;
+		__l0 *=3D __scale;
+		__l1 *=3D __scale;
+		_V __lo =3D __l0 * __l0
+			  + __l1 * __l1; // add the two smaller values first
+		asm("" : "+m"(__lo));
+		_V __r =3D __scale_back * sqrt(__h1 * __h1 + __lo);
+		// Fix up hypot(0, 0, 0) to not be NaN:
+		where(__hi =3D=3D 0, __r) =3D 0;
+		return __r;
+	      }
+#endif
+
+#ifdef __FAST_MATH__
+	    // With fast-math, ignore precision of subnormals and inputs from
+	    // __finite_max_v/2 to __finite_max_v. This removes all
+	    // branching/masking.
+	    if constexpr (true)
+#else
+	    if (_GLIBCXX_SIMD_IS_LIKELY(all_of(isnormal(__x))
+					&& all_of(isnormal(__y))
+					&& all_of(isnormal(__z))))
+#endif
+	      {
+		const _V __hi_exp =3D __hi & __inf;
+		//((__hi + __hi) & __inf) ^ __inf almost works for computing
+		//__scale, except when (__hi + __hi) & __inf =3D=3D __inf, in which
+		// case __scale
+		// becomes 0 (should be min/2 instead) and thus loses the
+		// information from __lo.
+#ifdef __FAST_MATH__
+		using _Ip =3D __int_for_sizeof_t<_Tp>;
+		using _IV =3D rebind_simd_t<_Ip, _V>;
+		const auto __as_int =3D __bit_cast<_IV>(__hi_exp);
+		const _V __scale
+		  =3D __bit_cast<_V>(2 * __bit_cast<_Ip>(_Tp(1)) - __as_int);
+#else
+		const _V __scale =3D (__hi_exp ^ __inf) * _Tp(.5);
+#endif
+		constexpr _Tp __mant_mask
+		  =3D __norm_min_v<_Tp> - __denorm_min_v<_Tp>;
+		const _V __h1 =3D (__hi & _V(__mant_mask)) | _V(1);
+		__l0 *=3D __scale;
+		__l1 *=3D __scale;
+		const _V __lo
+		  =3D __l0 * __l0
+		    + __l1 * __l1; // add the two smaller values first
+		return __hi_exp * sqrt(__lo + __h1 * __h1);
+	      }
+	    else
+	      {
+		// slower path to support subnormals
+		// if __hi is subnormal, avoid scaling by inf & final mul by 0
+		// (which yields NaN) by using min()
+		_V __scale =3D _V(1 / __norm_min_v<_Tp>);
+		// invert exponent w/o error and w/o using the slow divider
+		// unit: xor inverts the exponent but off by 1. Multiplication
+		// with .5 adjusts for the discrepancy.
+		where(__hi >=3D __norm_min_v<_Tp>, __scale)
+		  =3D ((__hi & __inf) ^ __inf) * _Tp(.5);
+		// adjust final exponent for subnormal inputs
+		_V __hi_exp =3D __norm_min_v<_Tp>;
+		where(__hi >=3D __norm_min_v<_Tp>, __hi_exp)
+		  =3D __hi & __inf;         // no error
+		_V __h1 =3D __hi * __scale; // no error
+		__l0 *=3D __scale;          // no error
+		__l1 *=3D __scale;          // no error
+		_V __lo =3D __l0 * __l0
+			  + __l1 * __l1; // add the two smaller values first
+		_V __r =3D __hi_exp * sqrt(__lo + __h1 * __h1);
+#ifdef __STDC_IEC_559__
+		// fixup for Annex F requirements
+		_V __fixup =3D __hi; // __lo =3D=3D 0
+		// where(__lo =3D=3D 0, __fixup)                   =3D __hi;
+		where(isunordered(__x, __y + __z), __fixup)
+		  =3D __quiet_NaN_v<_Tp>;
+		where(isinf(__absx) || isinf(__absy) || isinf(__absz), __fixup)
+		  =3D __inf;
+		// Instead of __lo =3D=3D 0, the following could depend on __h1=C2=B2 =
=3D=3D
+		// __h1=C2=B2 + __lo (i.e. __hi is so much larger than the other two
+		// inputs that the result is exactly __hi). While this may
+		// improve precision, it is likely to reduce efficiency if the
+		// ISA has FMAs (because __h1=C2=B2 + __lo is an FMA, but the
+		// intermediate
+		// __h1=C2=B2 must be kept)
+		where(!(__lo =3D=3D 0 || isunordered(__x, __y + __z)
+			|| isinf(__absx) || isinf(__absy) || isinf(__absz)),
+		      __fixup)
+		  =3D __r;
+		__r =3D __fixup;
+#endif
+		return __r;
+	      }
+	  }
+      }
+  }
+
+  template <typename _Tp, typename _Abi>
+  _GLIBCXX_SIMD_INTRINSIC simd<_Tp, _Abi>
+  hypot(const simd<_Tp, _Abi>& __x, const simd<_Tp, _Abi>& __y,
+	const simd<_Tp, _Abi>& __z)
+  {
+    return __hypot<conditional_t<__is_fixed_size_abi_v<_Abi>,
+				 const simd<_Tp, _Abi>&, simd<_Tp, _Abi>>>(__x,
+									   __y,
+									   __z);
+  }
+
+_GLIBCXX_SIMD_CVTING3(hypot)
+
+_GLIBCXX_SIMD_MATH_CALL2_(pow, _Tp)
+
+_GLIBCXX_SIMD_MATH_CALL_(sqrt)
+_GLIBCXX_SIMD_MATH_CALL_(erf)
+_GLIBCXX_SIMD_MATH_CALL_(erfc)
+_GLIBCXX_SIMD_MATH_CALL_(lgamma)
+_GLIBCXX_SIMD_MATH_CALL_(tgamma)
+_GLIBCXX_SIMD_MATH_CALL_(ceil)
+_GLIBCXX_SIMD_MATH_CALL_(floor)
+_GLIBCXX_SIMD_MATH_CALL_(nearbyint)
+_GLIBCXX_SIMD_MATH_CALL_(rint)
+_GLIBCXX_SIMD_MATH_CALL_(lrint)
+_GLIBCXX_SIMD_MATH_CALL_(llrint)
+
+_GLIBCXX_SIMD_MATH_CALL_(round)
+_GLIBCXX_SIMD_MATH_CALL_(lround)
+_GLIBCXX_SIMD_MATH_CALL_(llround)
+
+_GLIBCXX_SIMD_MATH_CALL_(trunc)
+
+_GLIBCXX_SIMD_MATH_CALL2_(fmod, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(remainder, _Tp)
+_GLIBCXX_SIMD_MATH_CALL3_(remquo, _Tp, int*)
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  copysign(const simd<_Tp, _Abi>& __x, const simd<_Tp, _Abi>& __y)
+  {
+    if constexpr (simd_size_v<_Tp, _Abi> =3D=3D 1)
+      return std::copysign(__x[0], __y[0]);
+    else if constexpr (is_same_v<_Tp, long double> && sizeof(_Tp) =3D=3D 1=
2)
+      // Remove this case once __bit_cast is implemented via __builtin_bit=
_cast.
+      // It is necessary, because __signmask below cannot be computed at c=
ompile
+      // time.
+      return simd<_Tp, _Abi>(
+	[&](auto __i) { return std::copysign(__x[__i], __y[__i]); });
+    else
+      {
+	using _V =3D simd<_Tp, _Abi>;
+	using namespace std::experimental::__float_bitwise_operators;
+	_GLIBCXX_SIMD_USE_CONSTEXPR_API auto __signmask =3D _V(1) ^ _V(-1);
+	return (__x & (__x ^ __signmask)) | (__y & __signmask);
+      }
+  }
+
+_GLIBCXX_SIMD_MATH_CALL2_(nextafter, _Tp)
+// not covered in [parallel.simd.math]:
+// _GLIBCXX_SIMD_MATH_CALL2_(nexttoward, long double)
+_GLIBCXX_SIMD_MATH_CALL2_(fdim, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(fmax, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(fmin, _Tp)
+
+_GLIBCXX_SIMD_MATH_CALL3_(fma, _Tp, _Tp)
+_GLIBCXX_SIMD_MATH_CALL_(fpclassify)
+_GLIBCXX_SIMD_MATH_CALL_(isfinite)
+
+// isnan and isinf require special treatment because old glibc may declare
+// `int isinf(double)`.
+template <typename _Tp, typename _Abi, typename...,
+	  typename _R =3D _Math_return_type_t<bool, _Tp, _Abi>>
+  enable_if_t<is_floating_point_v<_Tp>, _R>
+  isinf(simd<_Tp, _Abi> __x)
+  { return {__private_init, _Abi::_SimdImpl::_S_isinf(__data(__x))}; }
+
+template <typename _Tp, typename _Abi, typename...,
+	  typename _R =3D _Math_return_type_t<bool, _Tp, _Abi>>
+  enable_if_t<is_floating_point_v<_Tp>, _R>
+  isnan(simd<_Tp, _Abi> __x)
+  { return {__private_init, _Abi::_SimdImpl::_S_isnan(__data(__x))}; }
+
+_GLIBCXX_SIMD_MATH_CALL_(isnormal)
+
+template <typename..., typename _Tp, typename _Abi>
+  simd_mask<_Tp, _Abi>
+  signbit(simd<_Tp, _Abi> __x)
+  {
+    if constexpr (is_integral_v<_Tp>)
+      {
+	if constexpr (is_unsigned_v<_Tp>)
+	  return simd_mask<_Tp, _Abi>{}; // false
+	else
+	  return __x < 0;
+      }
+    else
+      return {__private_init, _Abi::_SimdImpl::_S_signbit(__data(__x))};
+  }
+
+_GLIBCXX_SIMD_MATH_CALL2_(isgreater, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(isgreaterequal, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(isless, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(islessequal, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(islessgreater, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(isunordered, _Tp)
+
+/* not covered in [parallel.simd.math]
+template <typename _Abi> __doublev<_Abi> nan(const char* tagp);
+template <typename _Abi> __floatv<_Abi> nanf(const char* tagp);
+template <typename _Abi> __ldoublev<_Abi> nanl(const char* tagp);
+
+template <typename _V> struct simd_div_t {
+    _V quot, rem;
+};
+
+template <typename _Abi>
+simd_div_t<_SCharv<_Abi>> div(_SCharv<_Abi> numer,
+					 _SCharv<_Abi> denom);
+template <typename _Abi>
+simd_div_t<__shortv<_Abi>> div(__shortv<_Abi> numer,
+					 __shortv<_Abi> denom);
+template <typename _Abi>
+simd_div_t<__intv<_Abi>> div(__intv<_Abi> numer, __intv<_Abi> denom);
+template <typename _Abi>
+simd_div_t<__longv<_Abi>> div(__longv<_Abi> numer,
+					__longv<_Abi> denom);
+template <typename _Abi>
+simd_div_t<__llongv<_Abi>> div(__llongv<_Abi> numer,
+					 __llongv<_Abi> denom);
+*/
+
+// special math {{{
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  assoc_laguerre(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& =
__n,
+		 const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __m,
+		 const simd<_Tp, _Abi>& __x)
+  {
+    return simd<_Tp, _Abi>([&](auto __i) {
+      return std::assoc_laguerre(__n[__i], __m[__i], __x[__i]);
+    });
+  }
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  assoc_legendre(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& =
__n,
+		 const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __m,
+		 const simd<_Tp, _Abi>& __x)
+  {
+    return simd<_Tp, _Abi>([&](auto __i) {
+      return std::assoc_legendre(__n[__i], __m[__i], __x[__i]);
+    });
+  }
+
+_GLIBCXX_SIMD_MATH_CALL2_(beta, _Tp)
+_GLIBCXX_SIMD_MATH_CALL_(comp_ellint_1)
+_GLIBCXX_SIMD_MATH_CALL_(comp_ellint_2)
+_GLIBCXX_SIMD_MATH_CALL2_(comp_ellint_3, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(cyl_bessel_i, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(cyl_bessel_j, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(cyl_bessel_k, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(cyl_neumann, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(ellint_1, _Tp)
+_GLIBCXX_SIMD_MATH_CALL2_(ellint_2, _Tp)
+_GLIBCXX_SIMD_MATH_CALL3_(ellint_3, _Tp, _Tp)
+_GLIBCXX_SIMD_MATH_CALL_(expint)
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  hermite(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __n,
+	  const simd<_Tp, _Abi>& __x)
+  {
+    return simd<_Tp, _Abi>(
+      [&](auto __i) { return std::hermite(__n[__i], __x[__i]); });
+  }
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  laguerre(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __n,
+	   const simd<_Tp, _Abi>& __x)
+  {
+    return simd<_Tp, _Abi>(
+      [&](auto __i) { return std::laguerre(__n[__i], __x[__i]); });
+  }
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  legendre(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __n,
+	   const simd<_Tp, _Abi>& __x)
+  {
+    return simd<_Tp, _Abi>(
+      [&](auto __i) { return std::legendre(__n[__i], __x[__i]); });
+  }
+
+_GLIBCXX_SIMD_MATH_CALL_(riemann_zeta)
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  sph_bessel(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __n,
+	     const simd<_Tp, _Abi>& __x)
+  {
+    return simd<_Tp, _Abi>(
+      [&](auto __i) { return std::sph_bessel(__n[__i], __x[__i]); });
+  }
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  sph_legendre(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __=
l,
+	       const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __m,
+	       const simd<_Tp, _Abi>& theta)
+  {
+    return simd<_Tp, _Abi>([&](auto __i) {
+      return std::assoc_legendre(__l[__i], __m[__i], theta[__i]);
+    });
+  }
+
+template <typename _Tp, typename _Abi>
+  enable_if_t<is_floating_point_v<_Tp>, simd<_Tp, _Abi>>
+  sph_neumann(const fixed_size_simd<unsigned, simd_size_v<_Tp, _Abi>>& __n,
+	      const simd<_Tp, _Abi>& __x)
+  {
+    return simd<_Tp, _Abi>(
+      [&](auto __i) { return std::sph_neumann(__n[__i], __x[__i]); });
+  }
+// }}}
+
+#undef _GLIBCXX_SIMD_MATH_CALL_
+#undef _GLIBCXX_SIMD_MATH_CALL2_
+#undef _GLIBCXX_SIMD_MATH_CALL3_
+
+_GLIBCXX_SIMD_END_NAMESPACE
+
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_MATH_H_
+
+// vim: foldmethod=3Dmarker sw=3D2 ts=3D8 noet sts=3D2
diff --git a/libstdc++-v3/include/experimental/bits/simd_neon.h b/libstdc++=
=2Dv3/include/experimental/bits/simd_neon.h
new file mode 100644
index 00000000000..a3a8ffe165f
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_neon.h
@@ -0,0 +1,519 @@
+// Simd NEON specific implementations -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_NEON_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_NEON_H_
+
+#if __cplusplus >=3D 201703L
+
+#if !_GLIBCXX_SIMD_HAVE_NEON
+#error "simd_neon.h may only be included when NEON on ARM is available"
+#endif
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+
+// _CommonImplNeon {{{
+struct _CommonImplNeon : _CommonImplBuiltin
+{
+  // _S_store {{{
+  using _CommonImplBuiltin::_S_store;
+
+  // }}}
+};
+
+// }}}
+// _SimdImplNeon {{{
+template <typename _Abi>
+  struct _SimdImplNeon : _SimdImplBuiltin<_Abi>
+  {
+    using _Base =3D _SimdImplBuiltin<_Abi>;
+
+    template <typename _Tp>
+      using _MaskMember =3D typename _Base::template _MaskMember<_Tp>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_max_store_size =3D 16;
+
+    // _S_masked_load {{{
+    template <typename _Tp, size_t _Np, typename _Up>
+      static inline _SimdWrapper<_Tp, _Np>
+      _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
+		     const _Up* __mem) noexcept
+      {
+	__execute_n_times<_Np>([&](auto __i) {
+	  if (__k[__i] !=3D 0)
+	    __merge._M_set(__i, static_cast<_Tp>(__mem[__i]));
+	});
+	return __merge;
+      }
+
+    // }}}
+    // _S_masked_store_nocvt {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
+			    _MaskMember<_Tp> __k)
+      {
+	__execute_n_times<_Np>([&](auto __i) {
+	  if (__k[__i] !=3D 0)
+	    __mem[__i] =3D __v[__i];
+	});
+      }
+
+    // }}}
+    // _S_reduce {{{
+    template <typename _Tp, typename _BinaryOperation>
+      _GLIBCXX_SIMD_INTRINSIC static _Tp
+      _S_reduce(simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
+      {
+	constexpr size_t _Np =3D __x.size();
+	if constexpr (sizeof(__x) =3D=3D 16 && _Np >=3D 4
+		      && !_Abi::template _S_is_partial<_Tp>)
+	  {
+	    const auto __halves =3D split<simd<_Tp, simd_abi::_Neon<8>>>(__x);
+	    const auto __y =3D __binary_op(__halves[0], __halves[1]);
+	    return _SimdImplNeon<simd_abi::_Neon<8>>::_S_reduce(
+	      __y, static_cast<_BinaryOperation&&>(__binary_op));
+	  }
+	else if constexpr (_Np =3D=3D 8)
+	  {
+	    __x =3D __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
+				     __vector_permute<1, 0, 3, 2, 5, 4, 7, 6>(
+				       __x._M_data)));
+	    __x =3D __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
+				     __vector_permute<3, 2, 1, 0, 7, 6, 5, 4>(
+				       __x._M_data)));
+	    __x =3D __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
+				     __vector_permute<7, 6, 5, 4, 3, 2, 1, 0>(
+				       __x._M_data)));
+	    return __x[0];
+	  }
+	else if constexpr (_Np =3D=3D 4)
+	  {
+	    __x
+	      =3D __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
+				   __vector_permute<1, 0, 3, 2>(__x._M_data)));
+	    __x
+	      =3D __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
+				   __vector_permute<3, 2, 1, 0>(__x._M_data)));
+	    return __x[0];
+	  }
+	else if constexpr (_Np =3D=3D 2)
+	  {
+	    __x =3D __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
+				     __vector_permute<1, 0>(__x._M_data)));
+	    return __x[0];
+	  }
+	else
+	  return _Base::_S_reduce(__x,
+				  static_cast<_BinaryOperation&&>(__binary_op));
+      }
+
+    // }}}
+    // math {{{
+    // _S_sqrt {{{
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      _GLIBCXX_SIMD_INTRINSIC static _Tp _S_sqrt(_Tp __x)
+      {
+	if constexpr (__have_neon_a64)
+	  {
+	    const auto __intrin =3D __to_intrin(__x);
+	    if constexpr (_TVT::template _S_is<float, 2>)
+	      return vsqrt_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<float, 4>)
+	      return vsqrtq_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 1>)
+	      return vsqrt_f64(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 2>)
+	      return vsqrtq_f64(__intrin);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_sqrt(__x);
+      }
+
+    // }}}
+    // _S_trunc {{{
+    template <typename _TW, typename _TVT =3D _VectorTraits<_TW>>
+      _GLIBCXX_SIMD_INTRINSIC static _TW _S_trunc(_TW __x)
+      {
+	using _Tp =3D typename _TVT::value_type;
+	if constexpr (__have_neon_a32)
+	  {
+	    const auto __intrin =3D __to_intrin(__x);
+	    if constexpr (_TVT::template _S_is<float, 2>)
+	      return vrnd_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<float, 4>)
+	      return vrndq_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 1>)
+	      return vrnd_f64(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 2>)
+	      return vrndq_f64(__intrin);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (is_same_v<_Tp, float>)
+	  {
+	    auto __intrin =3D __to_intrin(__x);
+	    if constexpr (sizeof(__x) =3D=3D 16)
+	      __intrin =3D vcvtq_f32_s32(vcvtq_s32_f32(__intrin));
+	    else
+	      __intrin =3D vcvt_f32_s32(vcvt_s32_f32(__intrin));
+	    return _Base::_S_abs(__x)._M_data < 0x1p23f
+		     ? __vector_bitcast<float>(__intrin)
+		     : __x._M_data;
+	  }
+	else
+	  return _Base::_S_trunc(__x);
+      }
+
+    // }}}
+    // _S_round {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_round(_SimdWrapper<_Tp, _Np> __x)
+      {
+	if constexpr (__have_neon_a32)
+	  {
+	    const auto __intrin =3D __to_intrin(__x);
+	    if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(__x) =3D=3D 8)
+	      return vrnda_f32(__intrin);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(__x) =3D=3D 16)
+	      return vrndaq_f32(__intrin);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(__x) =3D=3D 8)
+	      return vrnda_f64(__intrin);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(__x) =3D=3D 16)
+	      return vrndaq_f64(__intrin);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_round(__x);
+      }
+
+    // }}}
+    // _S_floor {{{
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      _GLIBCXX_SIMD_INTRINSIC static _Tp _S_floor(_Tp __x)
+      {
+	if constexpr (__have_neon_a32)
+	  {
+	    const auto __intrin =3D __to_intrin(__x);
+	    if constexpr (_TVT::template _S_is<float, 2>)
+	      return vrndm_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<float, 4>)
+	      return vrndmq_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 1>)
+	      return vrndm_f64(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 2>)
+	      return vrndmq_f64(__intrin);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_floor(__x);
+      }
+
+    // }}}
+    // _S_ceil {{{
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      _GLIBCXX_SIMD_INTRINSIC static _Tp _S_ceil(_Tp __x)
+      {
+	if constexpr (__have_neon_a32)
+	  {
+	    const auto __intrin =3D __to_intrin(__x);
+	    if constexpr (_TVT::template _S_is<float, 2>)
+	      return vrndp_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<float, 4>)
+	      return vrndpq_f32(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 1>)
+	      return vrndp_f64(__intrin);
+	    else if constexpr (_TVT::template _S_is<double, 2>)
+	      return vrndpq_f64(__intrin);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_ceil(__x);
+      }
+
+    //}}} }}}
+  }; // }}}
+// _MaskImplNeonMixin {{{
+struct _MaskImplNeonMixin
+{
+  using _Base =3D _MaskImplBuiltinMixin;
+
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
+    _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
+    {
+      if (__builtin_is_constant_evaluated())
+	return _Base::_S_to_bits(__x);
+
+      using _I =3D __int_for_sizeof_t<_Tp>;
+      if constexpr (sizeof(__x) =3D=3D 16)
+	{
+	  auto __asint =3D __vector_bitcast<_I>(__x);
+#ifdef __aarch64__
+	  [[maybe_unused]] constexpr auto __zero =3D decltype(__asint)();
+#else
+	  [[maybe_unused]] constexpr auto __zero =3D decltype(__lo64(__asint))();
+#endif
+	  if constexpr (sizeof(_Tp) =3D=3D 1)
+	    {
+	      constexpr auto __bitsel
+		=3D __generate_from_n_evaluations<16, __vector_type_t<_I, 16>>(
+		  [&](auto __i) {
+		    return static_cast<_I>(
+		      __i < _Np ? (__i < 8 ? 1 << __i : 1 << (__i - 8)) : 0);
+		  });
+	      __asint &=3D __bitsel;
+#ifdef __aarch64__
+	      return __vector_bitcast<_UShort>(
+		vpaddq_s8(vpaddq_s8(vpaddq_s8(__asint, __zero), __zero),
+			  __zero))[0];
+#else
+	      return __vector_bitcast<_UShort>(
+		vpadd_s8(vpadd_s8(vpadd_s8(__lo64(__asint), __hi64(__asint)),
+				  __zero),
+			 __zero))[0];
+#endif
+	    }
+	  else if constexpr (sizeof(_Tp) =3D=3D 2)
+	    {
+	      constexpr auto __bitsel
+		=3D __generate_from_n_evaluations<8, __vector_type_t<_I, 8>>(
+		  [&](auto __i) {
+		    return static_cast<_I>(__i < _Np ? 1 << __i : 0);
+		  });
+	      __asint &=3D __bitsel;
+#ifdef __aarch64__
+	      return vpaddq_s16(vpaddq_s16(vpaddq_s16(__asint, __zero), __zero),
+				__zero)[0];
+#else
+	      return vpadd_s16(
+		vpadd_s16(vpadd_s16(__lo64(__asint), __hi64(__asint)), __zero),
+		__zero)[0];
+#endif
+	    }
+	  else if constexpr (sizeof(_Tp) =3D=3D 4)
+	    {
+	      constexpr auto __bitsel
+		=3D __generate_from_n_evaluations<4, __vector_type_t<_I, 4>>(
+		  [&](auto __i) {
+		    return static_cast<_I>(__i < _Np ? 1 << __i : 0);
+		  });
+	      __asint &=3D __bitsel;
+#ifdef __aarch64__
+	      return vpaddq_s32(vpaddq_s32(__asint, __zero), __zero)[0];
+#else
+	      return vpadd_s32(vpadd_s32(__lo64(__asint), __hi64(__asint)),
+			       __zero)[0];
+#endif
+	    }
+	  else if constexpr (sizeof(_Tp) =3D=3D 8)
+	    return (__asint[0] & 1) | (__asint[1] & 2);
+	  else
+	    __assert_unreachable<_Tp>();
+	}
+      else if constexpr (sizeof(__x) =3D=3D 8)
+	{
+	  auto __asint =3D __vector_bitcast<_I>(__x);
+	  [[maybe_unused]] constexpr auto __zero =3D decltype(__asint)();
+	  if constexpr (sizeof(_Tp) =3D=3D 1)
+	    {
+	      constexpr auto __bitsel
+		=3D __generate_from_n_evaluations<8, __vector_type_t<_I, 8>>(
+		  [&](auto __i) {
+		    return static_cast<_I>(__i < _Np ? 1 << __i : 0);
+		  });
+	      __asint &=3D __bitsel;
+	      return vpadd_s8(vpadd_s8(vpadd_s8(__asint, __zero), __zero),
+			      __zero)[0];
+	    }
+	  else if constexpr (sizeof(_Tp) =3D=3D 2)
+	    {
+	      constexpr auto __bitsel
+		=3D __generate_from_n_evaluations<4, __vector_type_t<_I, 4>>(
+		  [&](auto __i) {
+		    return static_cast<_I>(__i < _Np ? 1 << __i : 0);
+		  });
+	      __asint &=3D __bitsel;
+	      return vpadd_s16(vpadd_s16(__asint, __zero), __zero)[0];
+	    }
+	  else if constexpr (sizeof(_Tp) =3D=3D 4)
+	    {
+	      __asint &=3D __make_vector<_I>(0x1, 0x2);
+	      return vpadd_s32(__asint, __zero)[0];
+	    }
+	  else
+	    __assert_unreachable<_Tp>();
+	}
+      else
+	return _Base::_S_to_bits(__x);
+    }
+};
+
+// }}}
+// _MaskImplNeon {{{
+template <typename _Abi>
+  struct _MaskImplNeon : _MaskImplNeonMixin, _MaskImplBuiltin<_Abi>
+  {
+    using _MaskImplBuiltinMixin::_S_to_maskvector;
+    using _MaskImplNeonMixin::_S_to_bits;
+    using _Base =3D _MaskImplBuiltin<_Abi>;
+    using _Base::_S_convert;
+
+    // _S_all_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_all_of(simd_mask<_Tp, _Abi> _=
_k)
+      {
+	const auto __kk
+	  =3D __vector_bitcast<char>(__k._M_data)
+	    | ~__vector_bitcast<char>(_Abi::template _S_implicit_mask<_Tp>());
+	if constexpr (sizeof(__k) =3D=3D 16)
+	  {
+	    const auto __x =3D __vector_bitcast<long long>(__kk);
+	    return __x[0] + __x[1] =3D=3D -2;
+	  }
+	else if constexpr (sizeof(__k) <=3D 8)
+	  return __bit_cast<__int_for_sizeof_t<decltype(__kk)>>(__kk) =3D=3D -1;
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    // }}}
+    // _S_any_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_any_of(simd_mask<_Tp, _Abi> _=
_k)
+      {
+	const auto __kk
+	  =3D __vector_bitcast<char>(__k._M_data)
+	    | ~__vector_bitcast<char>(_Abi::template _S_implicit_mask<_Tp>());
+	if constexpr (sizeof(__k) =3D=3D 16)
+	  {
+	    const auto __x =3D __vector_bitcast<long long>(__kk);
+	    return (__x[0] | __x[1]) !=3D 0;
+	  }
+	else if constexpr (sizeof(__k) <=3D 8)
+	  return __bit_cast<__int_for_sizeof_t<decltype(__kk)>>(__kk) !=3D 0;
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    // }}}
+    // _S_none_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_none_of(simd_mask<_Tp, _Abi> =
__k)
+      {
+	const auto __kk =3D _Abi::_S_masked(__k._M_data);
+	if constexpr (sizeof(__k) =3D=3D 16)
+	  {
+	    const auto __x =3D __vector_bitcast<long long>(__kk);
+	    return (__x[0] | __x[1]) =3D=3D 0;
+	  }
+	else if constexpr (sizeof(__k) <=3D 8)
+	  return __bit_cast<__int_for_sizeof_t<decltype(__kk)>>(__kk) =3D=3D 0;
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    // }}}
+    // _S_some_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_some_of(simd_mask<_Tp, _Abi> =
__k)
+      {
+	if constexpr (sizeof(__k) <=3D 8)
+	  {
+	    const auto __kk =3D __vector_bitcast<char>(__k._M_data)
+			      | ~__vector_bitcast<char>(
+				_Abi::template _S_implicit_mask<_Tp>());
+	    using _Up =3D make_unsigned_t<__int_for_sizeof_t<decltype(__kk)>>;
+	    return __bit_cast<_Up>(__kk) + 1 > 1;
+	  }
+	else
+	  return _Base::_S_some_of(__k);
+      }
+
+    // }}}
+    // _S_popcount {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int _S_popcount(simd_mask<_Tp, _Abi> =
__k)
+      {
+	if constexpr (sizeof(_Tp) =3D=3D 1)
+	  {
+	    const auto __s8 =3D __vector_bitcast<_SChar>(__k._M_data);
+	    int8x8_t __tmp =3D __lo64(__s8) + __hi64z(__s8);
+	    return -vpadd_s8(vpadd_s8(vpadd_s8(__tmp, int8x8_t()), int8x8_t()),
+			     int8x8_t())[0];
+	  }
+	else if constexpr (sizeof(_Tp) =3D=3D 2)
+	  {
+	    const auto __s16 =3D __vector_bitcast<short>(__k._M_data);
+	    int16x4_t __tmp =3D __lo64(__s16) + __hi64z(__s16);
+	    return -vpadd_s16(vpadd_s16(__tmp, int16x4_t()), int16x4_t())[0];
+	  }
+	else if constexpr (sizeof(_Tp) =3D=3D 4)
+	  {
+	    const auto __s32 =3D __vector_bitcast<int>(__k._M_data);
+	    int32x2_t __tmp =3D __lo64(__s32) + __hi64z(__s32);
+	    return -vpadd_s32(__tmp, int32x2_t())[0];
+	  }
+	else if constexpr (sizeof(_Tp) =3D=3D 8)
+	  {
+	    static_assert(sizeof(__k) =3D=3D 16);
+	    const auto __s64 =3D __vector_bitcast<long>(__k._M_data);
+	    return -(__s64[0] + __s64[1]);
+	  }
+      }
+
+    // }}}
+    // _S_find_first_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_first_set(simd_mask<_Tp, _Abi> __k)
+      {
+	// TODO: the _Base implementation is not optimal for NEON
+	return _Base::_S_find_first_set(__k);
+      }
+
+    // }}}
+    // _S_find_last_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_last_set(simd_mask<_Tp, _Abi> __k)
+      {
+	// TODO: the _Base implementation is not optimal for NEON
+	return _Base::_S_find_last_set(__k);
+      }
+
+    // }}}
+  }; // }}}
+
+_GLIBCXX_SIMD_END_NAMESPACE
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_NEON_H_
+// vim: foldmethod=3Dmarker sw=3D2 noet ts=3D8 sts=3D2 tw=3D80
diff --git a/libstdc++-v3/include/experimental/bits/simd_ppc.h b/libstdc++-=
v3/include/experimental/bits/simd_ppc.h
new file mode 100644
index 00000000000..c00d2323ac6
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_ppc.h
@@ -0,0 +1,123 @@
+// Simd PowerPC specific implementations -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_PPC_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_PPC_H_
+
+#if __cplusplus >=3D 201703L
+
+#ifndef __ALTIVEC__
+#error "simd_ppc.h may only be included when AltiVec/VMX is available"
+#endif
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+
+// _SimdImplPpc {{{
+template <typename _Abi>
+  struct _SimdImplPpc : _SimdImplBuiltin<_Abi>
+  {
+    using _Base =3D _SimdImplBuiltin<_Abi>;
+
+    // Byte and halfword shift instructions on PPC only consider the low 3=
 or 4
+    // bits of the RHS. Consequently, shifting by sizeof(_Tp)*CHAR_BIT (or=
 more)
+    // is UB without extra measures. To match scalar behavior, byte and ha=
lfword
+    // shifts need an extra fixup step.
+
+    // _S_bit_shift_left {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np>=
 __y)
+      {
+	__x =3D _Base::_S_bit_shift_left(__x, __y);
+	if constexpr (sizeof(_Tp) < sizeof(int))
+	  __x._M_data
+	    =3D (__y._M_data < sizeof(_Tp) * __CHAR_BIT__) & __x._M_data;
+	return __x;
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_left(_SimdWrapper<_Tp, _Np> __x, int __y)
+      {
+	__x =3D _Base::_S_bit_shift_left(__x, __y);
+	if constexpr (sizeof(_Tp) < sizeof(int))
+	  {
+	    if (__y >=3D sizeof(_Tp) * __CHAR_BIT__)
+	      return {};
+	  }
+	return __x;
+      }
+
+    // }}}
+    // _S_bit_shift_right {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np=
> __y)
+      {
+	if constexpr (sizeof(_Tp) < sizeof(int))
+	  {
+	    constexpr int __nbits =3D sizeof(_Tp) * __CHAR_BIT__;
+	    if constexpr (is_unsigned_v<_Tp>)
+	      return (__y._M_data < __nbits)
+		     & _Base::_S_bit_shift_right(__x, __y)._M_data;
+	    else
+	      {
+		_Base::_S_masked_assign(_SimdWrapper<_Tp, _Np>(__y._M_data
+							       >=3D __nbits),
+					__y, __nbits - 1);
+		return _Base::_S_bit_shift_right(__x, __y);
+	      }
+	  }
+	else
+	  return _Base::_S_bit_shift_right(__x, __y);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_shift_right(_SimdWrapper<_Tp, _Np> __x, int __y)
+      {
+	if constexpr (sizeof(_Tp) < sizeof(int))
+	  {
+	    constexpr int __nbits =3D sizeof(_Tp) * __CHAR_BIT__;
+	    if (__y >=3D __nbits)
+	      {
+		if constexpr (is_unsigned_v<_Tp>)
+		  return {};
+		else
+		  return _Base::_S_bit_shift_right(__x, __nbits - 1);
+	      }
+	  }
+	return _Base::_S_bit_shift_right(__x, __y);
+      }
+
+    // }}}
+  };
+
+// }}}
+
+_GLIBCXX_SIMD_END_NAMESPACE
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_PPC_H_
+
+// vim: foldmethod=3Dmarker sw=3D2 noet ts=3D8 sts=3D2 tw=3D80
diff --git a/libstdc++-v3/include/experimental/bits/simd_scalar.h b/libstdc=
++-v3/include/experimental/bits/simd_scalar.h
new file mode 100644
index 00000000000..7680bc39c30
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_scalar.h
@@ -0,0 +1,772 @@
+// Simd scalar ABI specific implementations -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_SCALAR_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_SCALAR_H_
+#if __cplusplus >=3D 201703L
+
+#include <cmath>
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+
+// __promote_preserving_unsigned{{{
+// work around crazy semantics of unsigned integers of lower rank than int:
+// Before applying an operator the operands are promoted to int. In which =
case
+// over- or underflow is UB, even though the operand types were unsigned.
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC constexpr decltype(auto)
+  __promote_preserving_unsigned(const _Tp& __x)
+  {
+    if constexpr (is_signed_v<decltype(+__x)> && is_unsigned_v<_Tp>)
+      return static_cast<unsigned int>(__x);
+    else
+      return __x;
+  }
+
+// }}}
+
+struct _CommonImplScalar;
+struct _CommonImplBuiltin;
+struct _SimdImplScalar;
+struct _MaskImplScalar;
+
+// simd_abi::_Scalar {{{
+struct simd_abi::_Scalar
+{
+  template <typename _Tp>
+    static constexpr size_t _S_size =3D 1;
+
+  template <typename _Tp>
+    static constexpr size_t _S_full_size =3D 1;
+
+  template <typename _Tp>
+    static constexpr bool _S_is_partial =3D false;
+
+  struct _IsValidAbiTag : true_type {};
+
+  template <typename _Tp>
+    struct _IsValidSizeFor : true_type {};
+
+  template <typename _Tp>
+    struct _IsValid : __is_vectorizable<_Tp> {};
+
+  template <typename _Tp>
+    static constexpr bool _S_is_valid_v =3D _IsValid<_Tp>::value;
+
+  _GLIBCXX_SIMD_INTRINSIC static constexpr bool _S_masked(bool __x)
+  { return __x; }
+
+  using _CommonImpl =3D _CommonImplScalar;
+  using _SimdImpl =3D _SimdImplScalar;
+  using _MaskImpl =3D _MaskImplScalar;
+
+  template <typename _Tp, bool =3D _S_is_valid_v<_Tp>>
+    struct __traits : _InvalidTraits {};
+
+  template <typename _Tp>
+    struct __traits<_Tp, true>
+    {
+      using _IsValid =3D true_type;
+      using _SimdImpl =3D _SimdImplScalar;
+      using _MaskImpl =3D _MaskImplScalar;
+      using _SimdMember =3D _Tp;
+      using _MaskMember =3D bool;
+
+      static constexpr size_t _S_simd_align =3D alignof(_SimdMember);
+      static constexpr size_t _S_mask_align =3D alignof(_MaskMember);
+
+      // nothing the user can spell converts to/from simd/simd_mask
+      struct _SimdCastType { _SimdCastType() =3D delete; };
+      struct _MaskCastType { _MaskCastType() =3D delete; };
+      struct _SimdBase {};
+      struct _MaskBase {};
+    };
+};
+
+// }}}
+// _CommonImplScalar {{{
+struct _CommonImplScalar
+{
+  // _S_store {{{
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static void _S_store(_Tp __x, void* __addr)
+    { __builtin_memcpy(__addr, &__x, sizeof(_Tp)); }
+
+  // }}}
+  // _S_store_bool_array(_BitMask) {{{
+  template <size_t _Np, bool _Sanitized>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr void
+    _S_store_bool_array(_BitMask<_Np, _Sanitized> __x, bool* __mem)
+    {
+      __make_dependent_t<decltype(__x), _CommonImplBuiltin>::_S_store_bool=
_array(
+	__x, __mem);
+    }
+
+  // }}}
+};
+
+// }}}
+// _SimdImplScalar {{{
+struct _SimdImplScalar
+{
+  // member types {{{2
+  using abi_type =3D simd_abi::scalar;
+
+  template <typename _Tp>
+    using _TypeTag =3D _Tp*;
+
+  // _S_broadcast {{{2
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _Tp _S_broadcast(_Tp __x) noe=
xcept
+    { return __x; }
+
+  // _S_generator {{{2
+  template <typename _Fp, typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _Tp _S_generator(_Fp&& __gen,
+							      _TypeTag<_Tp>)
+    { return __gen(_SizeConstant<0>()); }
+
+  // _S_load {{{2
+  template <typename _Tp, typename _Up>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_load(const _Up* __mem,
+					       _TypeTag<_Tp>) noexcept
+    { return static_cast<_Tp>(__mem[0]); }
+
+  // _S_masked_load {{{2
+  template <typename _Tp, typename _Up>
+    static inline _Tp _S_masked_load(_Tp __merge, bool __k,
+				     const _Up* __mem) noexcept
+    {
+      if (__k)
+	__merge =3D static_cast<_Tp>(__mem[0]);
+      return __merge;
+    }
+
+  // _S_store {{{2
+  template <typename _Tp, typename _Up>
+    static inline void _S_store(_Tp __v, _Up* __mem, _TypeTag<_Tp>) noexce=
pt
+    { __mem[0] =3D static_cast<_Up>(__v); }
+
+  // _S_masked_store {{{2
+  template <typename _Tp, typename _Up>
+    static inline void _S_masked_store(const _Tp __v, _Up* __mem,
+				       const bool __k) noexcept
+    { if (__k) __mem[0] =3D __v; }
+
+  // _S_negate {{{2
+  template <typename _Tp>
+    static constexpr inline bool _S_negate(_Tp __x) noexcept
+    { return !__x; }
+
+  // _S_reduce {{{2
+  template <typename _Tp, typename _BinaryOperation>
+    static constexpr inline _Tp
+    _S_reduce(const simd<_Tp, simd_abi::scalar>& __x, _BinaryOperation&)
+    { return __x._M_data; }
+
+  // _S_min, _S_max {{{2
+  template <typename _Tp>
+    static constexpr inline _Tp _S_min(const _Tp __a, const _Tp __b)
+    { return std::min(__a, __b); }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_max(const _Tp __a, const _Tp __b)
+    { return std::max(__a, __b); }
+
+  // _S_complement {{{2
+  template <typename _Tp>
+    static constexpr inline _Tp _S_complement(_Tp __x) noexcept
+    { return static_cast<_Tp>(~__x); }
+
+  // _S_unary_minus {{{2
+  template <typename _Tp>
+    static constexpr inline _Tp _S_unary_minus(_Tp __x) noexcept
+    { return static_cast<_Tp>(-__x); }
+
+  // arithmetic operators {{{2
+  template <typename _Tp>
+    static constexpr inline _Tp _S_plus(_Tp __x, _Tp __y)
+    {
+      return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+			      + __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_minus(_Tp __x, _Tp __y)
+    {
+      return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+			      - __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_multiplies(_Tp __x, _Tp __y)
+    {
+      return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+			      * __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_divides(_Tp __x, _Tp __y)
+    {
+      return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+			      / __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_modulus(_Tp __x, _Tp __y)
+    {
+      return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+			      % __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_bit_and(_Tp __x, _Tp __y)
+    {
+      if constexpr (is_floating_point_v<_Tp>)
+	{
+	  using _Ip =3D __int_for_sizeof_t<_Tp>;
+	  return __bit_cast<_Tp>(__bit_cast<_Ip>(__x) & __bit_cast<_Ip>(__y));
+	}
+      else
+	return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+				& __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_bit_or(_Tp __x, _Tp __y)
+    {
+      if constexpr (is_floating_point_v<_Tp>)
+	{
+	  using _Ip =3D __int_for_sizeof_t<_Tp>;
+	  return __bit_cast<_Tp>(__bit_cast<_Ip>(__x) | __bit_cast<_Ip>(__y));
+	}
+      else
+	return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+				| __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_bit_xor(_Tp __x, _Tp __y)
+    {
+      if constexpr (is_floating_point_v<_Tp>)
+	{
+	  using _Ip =3D __int_for_sizeof_t<_Tp>;
+	  return __bit_cast<_Tp>(__bit_cast<_Ip>(__x) ^ __bit_cast<_Ip>(__y));
+	}
+      else
+	return static_cast<_Tp>(__promote_preserving_unsigned(__x)
+				^ __promote_preserving_unsigned(__y));
+    }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_bit_shift_left(_Tp __x, int __y)
+    { return static_cast<_Tp>(__promote_preserving_unsigned(__x) << __y); }
+
+  template <typename _Tp>
+    static constexpr inline _Tp _S_bit_shift_right(_Tp __x, int __y)
+    { return static_cast<_Tp>(__promote_preserving_unsigned(__x) >> __y); }
+
+  // math {{{2
+  // frexp, modf and copysign implemented in simd_math.h
+  template <typename _Tp>
+    using _ST =3D _SimdTuple<_Tp, simd_abi::scalar>;
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_acos(_Tp __x)
+    { return std::acos(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_asin(_Tp __x)
+    { return std::asin(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_atan(_Tp __x)
+    { return std::atan(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_cos(_Tp __x)
+    { return std::cos(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_sin(_Tp __x)
+    { return std::sin(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_tan(_Tp __x)
+    { return std::tan(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_acosh(_Tp __x)
+    { return std::acosh(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_asinh(_Tp __x)
+    { return std::asinh(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_atanh(_Tp __x)
+    { return std::atanh(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_cosh(_Tp __x)
+    { return std::cosh(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_sinh(_Tp __x)
+    { return std::sinh(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_tanh(_Tp __x)
+    { return std::tanh(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_atan2(_Tp __x, _Tp __y)
+    { return std::atan2(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_exp(_Tp __x)
+    { return std::exp(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_exp2(_Tp __x)
+    { return std::exp2(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_expm1(_Tp __x)
+    { return std::expm1(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_log(_Tp __x)
+    { return std::log(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_log10(_Tp __x)
+    { return std::log10(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_log1p(_Tp __x)
+    { return std::log1p(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_log2(_Tp __x)
+    { return std::log2(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_logb(_Tp __x)
+    { return std::logb(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _ST<int> _S_ilogb(_Tp __x)
+    { return {std::ilogb(__x)}; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_pow(_Tp __x, _Tp __y)
+    { return std::pow(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_abs(_Tp __x)
+    { return std::abs(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_fabs(_Tp __x)
+    { return std::fabs(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_sqrt(_Tp __x)
+    { return std::sqrt(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_cbrt(_Tp __x)
+    { return std::cbrt(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_erf(_Tp __x)
+    { return std::erf(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_erfc(_Tp __x)
+    { return std::erfc(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_lgamma(_Tp __x)
+    { return std::lgamma(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_tgamma(_Tp __x)
+    { return std::tgamma(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_trunc(_Tp __x)
+    { return std::trunc(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_floor(_Tp __x)
+    { return std::floor(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_ceil(_Tp __x)
+    { return std::ceil(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_nearbyint(_Tp __x)
+    { return std::nearbyint(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_rint(_Tp __x)
+    { return std::rint(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _ST<long> _S_lrint(_Tp __x)
+    { return {std::lrint(__x)}; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _ST<long long> _S_llrint(_Tp __x)
+    { return {std::llrint(__x)}; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_round(_Tp __x)
+    { return std::round(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _ST<long> _S_lround(_Tp __x)
+    { return {std::lround(__x)}; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _ST<long long> _S_llround(_Tp __x)
+    { return {std::llround(__x)}; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_ldexp(_Tp __x, _ST<int> __y)
+    { return std::ldexp(__x, __y.first); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_scalbn(_Tp __x, _ST<int> __y)
+    { return std::scalbn(__x, __y.first); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_scalbln(_Tp __x, _ST<long> __y)
+    { return std::scalbln(__x, __y.first); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_fmod(_Tp __x, _Tp __y)
+    { return std::fmod(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_remainder(_Tp __x, _Tp __y)
+    { return std::remainder(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_nextafter(_Tp __x, _Tp __y)
+    { return std::nextafter(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_fdim(_Tp __x, _Tp __y)
+    { return std::fdim(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_fmax(_Tp __x, _Tp __y)
+    { return std::fmax(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_fmin(_Tp __x, _Tp __y)
+    { return std::fmin(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_fma(_Tp __x, _Tp __y, _Tp __z)
+    { return std::fma(__x, __y, __z); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_remquo(_Tp __x, _Tp __y, _ST<int=
>* __z)
+    { return std::remquo(__x, __y, &__z->first); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static _ST<int> _S_fpclassify(_Tp __=
x)
+    { return {std::fpclassify(__x)}; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isfinite(_Tp __x)
+    { return std::isfinite(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isinf(_Tp __x)
+    { return std::isinf(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isnan(_Tp __x)
+    { return std::isnan(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isnormal(_Tp __x)
+    { return std::isnormal(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_signbit(_Tp __x)
+    { return std::signbit(__x); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isgreater(_Tp __x, _T=
p __y)
+    { return std::isgreater(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isgreaterequal(_Tp __=
x,
+								    _Tp __y)
+    { return std::isgreaterequal(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isless(_Tp __x, _Tp _=
_y)
+    { return std::isless(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_islessequal(_Tp __x, =
_Tp __y)
+    { return std::islessequal(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_islessgreater(_Tp __x,
+								   _Tp __y)
+    { return std::islessgreater(__x, __y); }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_isunordered(_Tp __x,
+								 _Tp __y)
+    { return std::isunordered(__x, __y); }
+
+  // _S_increment & _S_decrement{{{2
+  template <typename _Tp>
+    constexpr static inline void _S_increment(_Tp& __x)
+    { ++__x; }
+
+  template <typename _Tp>
+    constexpr static inline void _S_decrement(_Tp& __x)
+    { --__x; }
+
+
+  // compares {{{2
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_equal_to(_Tp __x, _Tp=
 __y)
+    { return __x =3D=3D __y; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_not_equal_to(_Tp __x,
+								  _Tp __y)
+    { return __x !=3D __y; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_less(_Tp __x, _Tp __y)
+    { return __x < __y; }
+
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool _S_less_equal(_Tp __x,
+								_Tp __y)
+    { return __x <=3D __y; }
+
+  // smart_reference access {{{2
+  template <typename _Tp, typename _Up>
+    constexpr static void _S_set(_Tp& __v, [[maybe_unused]] int __i,
+				 _Up&& __x) noexcept
+    {
+      _GLIBCXX_DEBUG_ASSERT(__i =3D=3D 0);
+      __v =3D static_cast<_Up&&>(__x);
+    }
+
+  // _S_masked_assign {{{2
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static void
+    _S_masked_assign(bool __k, _Tp& __lhs, _Tp __rhs)
+    { if (__k) __lhs =3D __rhs; }
+
+  // _S_masked_cassign {{{2
+  template <typename _Op, typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static void
+    _S_masked_cassign(const bool __k, _Tp& __lhs, const _Tp __rhs, _Op __o=
p)
+    { if (__k) __lhs =3D __op(_SimdImplScalar{}, __lhs, __rhs); }
+
+  // _S_masked_unary {{{2
+  template <template <typename> class _Op, typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static _Tp _S_masked_unary(const boo=
l __k,
+								 const _Tp __v)
+    { return static_cast<_Tp>(__k ? _Op<_Tp>{}(__v) : __v); }
+
+  // }}}2
+};
+
+// }}}
+// _MaskImplScalar {{{
+struct _MaskImplScalar
+{
+  // member types {{{
+  template <typename _Tp>
+    using _TypeTag =3D _Tp*;
+
+  // }}}
+  // _S_broadcast {{{
+  template <typename>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr bool _S_broadcast(bool __x)
+    { return __x; }
+
+  // }}}
+  // _S_load {{{
+  template <typename>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr bool _S_load(const bool* __me=
m)
+    { return __mem[0]; }
+
+  // }}}
+  // _S_to_bits {{{
+  _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<1>
+  _S_to_bits(bool __x)
+  { return __x; }
+
+  // }}}
+  // _S_convert {{{
+  template <typename, bool _Sanitized>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr bool
+    _S_convert(_BitMask<1, _Sanitized> __x)
+    { return __x[0]; }
+
+  template <typename, typename _Up, typename _UAbi>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr bool
+    _S_convert(simd_mask<_Up, _UAbi> __x)
+    { return __x[0]; }
+
+  // }}}
+  // _S_from_bitmask {{{2
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool
+    _S_from_bitmask(_SanitizedBitMask<1> __bits, _TypeTag<_Tp>) noexcept
+    { return __bits[0]; }
+
+  // _S_masked_load {{{2
+  _GLIBCXX_SIMD_INTRINSIC constexpr static bool
+  _S_masked_load(bool __merge, bool __mask, const bool* __mem) noexcept
+  {
+    if (__mask)
+      __merge =3D __mem[0];
+    return __merge;
+  }
+
+  // _S_store {{{2
+  _GLIBCXX_SIMD_INTRINSIC static void _S_store(bool __v, bool* __mem) noex=
cept
+  { __mem[0] =3D __v; }
+
+  // _S_masked_store {{{2
+  _GLIBCXX_SIMD_INTRINSIC static void
+  _S_masked_store(const bool __v, bool* __mem, const bool __k) noexcept
+  {
+    if (__k)
+      __mem[0] =3D __v;
+  }
+
+  // logical and bitwise operators {{{2
+  static constexpr bool _S_logical_and(bool __x, bool __y)
+  { return __x && __y; }
+
+  static constexpr bool _S_logical_or(bool __x, bool __y)
+  { return __x || __y; }
+
+  static constexpr bool _S_bit_not(bool __x)
+  { return !__x; }
+
+  static constexpr bool _S_bit_and(bool __x, bool __y)
+  { return __x && __y; }
+
+  static constexpr bool _S_bit_or(bool __x, bool __y)
+  { return __x || __y; }
+
+  static constexpr bool _S_bit_xor(bool __x, bool __y)
+  { return __x !=3D __y; }
+
+  // smart_reference access {{{2
+  constexpr static void _S_set(bool& __k, [[maybe_unused]] int __i,
+			       bool __x) noexcept
+  {
+    _GLIBCXX_DEBUG_ASSERT(__i =3D=3D 0);
+    __k =3D __x;
+  }
+
+  // _S_masked_assign {{{2
+  _GLIBCXX_SIMD_INTRINSIC static void _S_masked_assign(bool __k, bool& __l=
hs,
+						       bool __rhs)
+  {
+    if (__k)
+      __lhs =3D __rhs;
+  }
+
+  // }}}2
+  // _S_all_of {{{
+  template <typename _Tp, typename _Abi>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool
+    _S_all_of(simd_mask<_Tp, _Abi> __k)
+    { return __k._M_data; }
+
+  // }}}
+  // _S_any_of {{{
+  template <typename _Tp, typename _Abi>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool
+    _S_any_of(simd_mask<_Tp, _Abi> __k)
+    { return __k._M_data; }
+
+  // }}}
+  // _S_none_of {{{
+  template <typename _Tp, typename _Abi>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool
+    _S_none_of(simd_mask<_Tp, _Abi> __k)
+    { return !__k._M_data; }
+
+  // }}}
+  // _S_some_of {{{
+  template <typename _Tp, typename _Abi>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static bool
+    _S_some_of(simd_mask<_Tp, _Abi>)
+    { return false; }
+
+  // }}}
+  // _S_popcount {{{
+  template <typename _Tp, typename _Abi>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static int
+    _S_popcount(simd_mask<_Tp, _Abi> __k)
+    { return __k._M_data; }
+
+  // }}}
+  // _S_find_first_set {{{
+  template <typename _Tp, typename _Abi>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static int
+    _S_find_first_set(simd_mask<_Tp, _Abi>)
+    { return 0; }
+
+  // }}}
+  // _S_find_last_set {{{
+  template <typename _Tp, typename _Abi>
+    _GLIBCXX_SIMD_INTRINSIC constexpr static int
+    _S_find_last_set(simd_mask<_Tp, _Abi>)
+    { return 0; }
+
+  // }}}
+};
+
+// }}}
+
+_GLIBCXX_SIMD_END_NAMESPACE
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_SCALAR_H_
+
+// vim: foldmethod=3Dmarker sw=3D2 noet ts=3D8 sts=3D2 tw=3D80
diff --git a/libstdc++-v3/include/experimental/bits/simd_x86.h b/libstdc++-=
v3/include/experimental/bits/simd_x86.h
new file mode 100644
index 00000000000..d1d7b9d4bf3
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_x86.h
@@ -0,0 +1,5169 @@
+// Simd x86 specific implementations -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_X86_H_
+#define _GLIBCXX_EXPERIMENTAL_SIMD_X86_H_
+
+#if __cplusplus >=3D 201703L
+
+#if !_GLIBCXX_SIMD_X86INTRIN
+#error                                                                    =
     \
+  "simd_x86.h may only be included when MMX or SSE on x86(_64) are availab=
le"
+#endif
+
+_GLIBCXX_SIMD_BEGIN_NAMESPACE
+
+// __to_masktype {{{
+// Given <T, N> return <__int_for_sizeof_t<T>, N>. For _SimdWrapper and
+// __vector_type_t.
+template <typename _Tp, size_t _Np>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _SimdWrapper<__int_for_sizeof_t<_Tp>, =
_Np>
+  __to_masktype(_SimdWrapper<_Tp, _Np> __x)
+  {
+    return reinterpret_cast<__vector_type_t<__int_for_sizeof_t<_Tp>, _Np>>(
+      __x._M_data);
+  }
+
+template <typename _TV,
+	  typename _TVT
+	  =3D enable_if_t<__is_vector_type_v<_TV>, _VectorTraits<_TV>>,
+	  typename _Up =3D __int_for_sizeof_t<typename _TVT::value_type>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr __vector_type_t<_Up, _TVT::_S_full_siz=
e>
+  __to_masktype(_TV __x)
+  { return reinterpret_cast<__vector_type_t<_Up, _TVT::_S_full_size>>(__x)=
; }
+
+// }}}
+// __interleave128_lo {{{
+template <typename _Ap, typename _B, typename _Tp =3D common_type_t<_Ap, _=
B>,
+	  typename _Trait =3D _VectorTraits<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr _Tp
+  __interleave128_lo(const _Ap& __av, const _B& __bv)
+  {
+    const _Tp __a(__av);
+    const _Tp __b(__bv);
+    if constexpr (sizeof(_Tp) =3D=3D 16 && _Trait::_S_full_size =3D=3D 2)
+      return _Tp{__a[0], __b[0]};
+    else if constexpr (sizeof(_Tp) =3D=3D 16 && _Trait::_S_full_size =3D=
=3D 4)
+      return _Tp{__a[0], __b[0], __a[1], __b[1]};
+    else if constexpr (sizeof(_Tp) =3D=3D 16 && _Trait::_S_full_size =3D=
=3D 8)
+      return _Tp{__a[0], __b[0], __a[1], __b[1],
+		 __a[2], __b[2], __a[3], __b[3]};
+    else if constexpr (sizeof(_Tp) =3D=3D 16 && _Trait::_S_full_size =3D=
=3D 16)
+      return _Tp{__a[0], __b[0], __a[1], __b[1], __a[2], __b[2],
+		 __a[3], __b[3], __a[4], __b[4], __a[5], __b[5],
+		 __a[6], __b[6], __a[7], __b[7]};
+    else if constexpr (sizeof(_Tp) =3D=3D 32 && _Trait::_S_full_size =3D=
=3D 4)
+      return _Tp{__a[0], __b[0], __a[2], __b[2]};
+    else if constexpr (sizeof(_Tp) =3D=3D 32 && _Trait::_S_full_size =3D=
=3D 8)
+      return _Tp{__a[0], __b[0], __a[1], __b[1],
+		 __a[4], __b[4], __a[5], __b[5]};
+    else if constexpr (sizeof(_Tp) =3D=3D 32 && _Trait::_S_full_size =3D=
=3D 16)
+      return _Tp{__a[0],  __b[0],  __a[1],  __b[1], __a[2], __b[2],
+		 __a[3],  __b[3],  __a[8],  __b[8], __a[9], __b[9],
+		 __a[10], __b[10], __a[11], __b[11]};
+    else if constexpr (sizeof(_Tp) =3D=3D 32 && _Trait::_S_full_size =3D=
=3D 32)
+      return _Tp{__a[0],  __b[0],  __a[1],  __b[1],  __a[2],  __b[2],  __a=
[3],
+		 __b[3],  __a[4],  __b[4],  __a[5],  __b[5],  __a[6],  __b[6],
+		 __a[7],  __b[7],  __a[16], __b[16], __a[17], __b[17], __a[18],
+		 __b[18], __a[19], __b[19], __a[20], __b[20], __a[21], __b[21],
+		 __a[22], __b[22], __a[23], __b[23]};
+    else if constexpr (sizeof(_Tp) =3D=3D 64 && _Trait::_S_full_size =3D=
=3D 8)
+      return _Tp{__a[0], __b[0], __a[2], __b[2],
+		 __a[4], __b[4], __a[6], __b[6]};
+    else if constexpr (sizeof(_Tp) =3D=3D 64 && _Trait::_S_full_size =3D=
=3D 16)
+      return _Tp{__a[0],  __b[0],  __a[1],  __b[1], __a[4], __b[4],
+		 __a[5],  __b[5],  __a[8],  __b[8], __a[9], __b[9],
+		 __a[12], __b[12], __a[13], __b[13]};
+    else if constexpr (sizeof(_Tp) =3D=3D 64 && _Trait::_S_full_size =3D=
=3D 32)
+      return _Tp{__a[0],  __b[0],  __a[1],  __b[1],  __a[2],  __b[2],  __a=
[3],
+		 __b[3],  __a[8],  __b[8],  __a[9],  __b[9],  __a[10], __b[10],
+		 __a[11], __b[11], __a[16], __b[16], __a[17], __b[17], __a[18],
+		 __b[18], __a[19], __b[19], __a[24], __b[24], __a[25], __b[25],
+		 __a[26], __b[26], __a[27], __b[27]};
+    else if constexpr (sizeof(_Tp) =3D=3D 64 && _Trait::_S_full_size =3D=
=3D 64)
+      return _Tp{__a[0],  __b[0],  __a[1],  __b[1],  __a[2],  __b[2],  __a=
[3],
+		 __b[3],  __a[4],  __b[4],  __a[5],  __b[5],  __a[6],  __b[6],
+		 __a[7],  __b[7],  __a[16], __b[16], __a[17], __b[17], __a[18],
+		 __b[18], __a[19], __b[19], __a[20], __b[20], __a[21], __b[21],
+		 __a[22], __b[22], __a[23], __b[23], __a[32], __b[32], __a[33],
+		 __b[33], __a[34], __b[34], __a[35], __b[35], __a[36], __b[36],
+		 __a[37], __b[37], __a[38], __b[38], __a[39], __b[39], __a[48],
+		 __b[48], __a[49], __b[49], __a[50], __b[50], __a[51], __b[51],
+		 __a[52], __b[52], __a[53], __b[53], __a[54], __b[54], __a[55],
+		 __b[55]};
+    else
+      __assert_unreachable<_Tp>();
+  }
+
+// }}}
+// __is_zero{{{
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC constexpr bool
+  __is_zero(_Tp __a)
+  {
+    if (!__builtin_is_constant_evaluated())
+      {
+	if constexpr (__have_avx)
+	  {
+	    if constexpr (_TVT::template _S_is<float, 8>)
+	      return _mm256_testz_ps(__a, __a);
+	    else if constexpr (_TVT::template _S_is<double, 4>)
+	      return _mm256_testz_pd(__a, __a);
+	    else if constexpr (sizeof(_Tp) =3D=3D 32)
+	      return _mm256_testz_si256(__to_intrin(__a), __to_intrin(__a));
+	    else if constexpr (_TVT::template _S_is<float>)
+	      return _mm_testz_ps(__to_intrin(__a), __to_intrin(__a));
+	    else if constexpr (_TVT::template _S_is<double, 2>)
+	      return _mm_testz_pd(__a, __a);
+	    else
+	      return _mm_testz_si128(__to_intrin(__a), __to_intrin(__a));
+	  }
+	else if constexpr (__have_sse4_1)
+	  return _mm_testz_si128(__intrin_bitcast<__m128i>(__a),
+				 __intrin_bitcast<__m128i>(__a));
+      }
+    else if constexpr (sizeof(_Tp) <=3D 8)
+      return reinterpret_cast<__int_for_sizeof_t<_Tp>>(__a) =3D=3D 0;
+    else
+      {
+	const auto __b =3D __vector_bitcast<_LLong>(__a);
+	if constexpr (sizeof(__b) =3D=3D 16)
+	  return (__b[0] | __b[1]) =3D=3D 0;
+	else if constexpr (sizeof(__b) =3D=3D 32)
+	  return __is_zero(__lo128(__b) | __hi128(__b));
+	else if constexpr (sizeof(__b) =3D=3D 64)
+	  return __is_zero(__lo256(__b) | __hi256(__b));
+	else
+	  __assert_unreachable<_Tp>();
+      }
+  }
+
+// }}}
+// __movemask{{{
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST int
+  __movemask(_Tp __a)
+  {
+    if constexpr (sizeof(_Tp) =3D=3D 32)
+      {
+	if constexpr (_TVT::template _S_is<float>)
+	  return _mm256_movemask_ps(__to_intrin(__a));
+	else if constexpr (_TVT::template _S_is<double>)
+	  return _mm256_movemask_pd(__to_intrin(__a));
+	else
+	  return _mm256_movemask_epi8(__to_intrin(__a));
+      }
+    else if constexpr (_TVT::template _S_is<float>)
+      return _mm_movemask_ps(__to_intrin(__a));
+    else if constexpr (_TVT::template _S_is<double>)
+      return _mm_movemask_pd(__to_intrin(__a));
+    else
+      return _mm_movemask_epi8(__to_intrin(__a));
+  }
+
+// }}}
+// __testz{{{
+template <typename _TI, typename _TVT =3D _VectorTraits<_TI>>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
+  __testz(_TI __a, _TI __b)
+  {
+    static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_t=
ype,
+						    _TVT::_S_full_size>>);
+    if (!__builtin_is_constant_evaluated())
+      {
+	if constexpr (sizeof(_TI) =3D=3D 32)
+	  {
+	    if constexpr (_TVT::template _S_is<float>)
+	      return _mm256_testz_ps(__to_intrin(__a), __to_intrin(__b));
+	    else if constexpr (_TVT::template _S_is<double>)
+	      return _mm256_testz_pd(__to_intrin(__a), __to_intrin(__b));
+	    else
+	      return _mm256_testz_si256(__to_intrin(__a), __to_intrin(__b));
+	  }
+	else if constexpr (_TVT::template _S_is<float> && __have_avx)
+	  return _mm_testz_ps(__to_intrin(__a), __to_intrin(__b));
+	else if constexpr (_TVT::template _S_is<double> && __have_avx)
+	  return _mm_testz_pd(__to_intrin(__a), __to_intrin(__b));
+	else if constexpr (__have_sse4_1)
+	  return _mm_testz_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
+				 __intrin_bitcast<__m128i>(__to_intrin(__b)));
+	else
+	  return __movemask(0 =3D=3D __and(__a, __b)) !=3D 0;
+      }
+    else
+      return __is_zero(__and(__a, __b));
+  }
+
+// }}}
+// __testc{{{
+// requires SSE4.1 or above
+template <typename _TI, typename _TVT =3D _VectorTraits<_TI>>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
+  __testc(_TI __a, _TI __b)
+  {
+    static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_t=
ype,
+						    _TVT::_S_full_size>>);
+    if (__builtin_is_constant_evaluated())
+      return __is_zero(__andnot(__a, __b));
+
+    if constexpr (sizeof(_TI) =3D=3D 32)
+      {
+	if constexpr (_TVT::template _S_is<float>)
+	  return _mm256_testc_ps(__a, __b);
+	else if constexpr (_TVT::template _S_is<double>)
+	  return _mm256_testc_pd(__a, __b);
+	else
+	  return _mm256_testc_si256(__to_intrin(__a), __to_intrin(__b));
+      }
+    else if constexpr (_TVT::template _S_is<float> && __have_avx)
+      return _mm_testc_ps(__to_intrin(__a), __to_intrin(__b));
+    else if constexpr (_TVT::template _S_is<double> && __have_avx)
+      return _mm_testc_pd(__to_intrin(__a), __to_intrin(__b));
+    else
+      {
+	static_assert(is_same_v<_TI, _TI> && __have_sse4_1);
+	return _mm_testc_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
+			       __intrin_bitcast<__m128i>(__to_intrin(__b)));
+      }
+  }
+
+// }}}
+// __testnzc{{{
+template <typename _TI, typename _TVT =3D _VectorTraits<_TI>>
+  _GLIBCXX_SIMD_INTRINSIC _GLIBCXX_CONST constexpr int
+  __testnzc(_TI __a, _TI __b)
+  {
+    static_assert(is_same_v<_TI, __intrinsic_type_t<typename _TVT::value_t=
ype,
+						    _TVT::_S_full_size>>);
+    if (!__builtin_is_constant_evaluated())
+      {
+	if constexpr (sizeof(_TI) =3D=3D 32)
+	  {
+	    if constexpr (_TVT::template _S_is<float>)
+	      return _mm256_testnzc_ps(__a, __b);
+	    else if constexpr (_TVT::template _S_is<double>)
+	      return _mm256_testnzc_pd(__a, __b);
+	    else
+	      return _mm256_testnzc_si256(__to_intrin(__a), __to_intrin(__b));
+	  }
+	else if constexpr (_TVT::template _S_is<float> && __have_avx)
+	  return _mm_testnzc_ps(__to_intrin(__a), __to_intrin(__b));
+	else if constexpr (_TVT::template _S_is<double> && __have_avx)
+	  return _mm_testnzc_pd(__to_intrin(__a), __to_intrin(__b));
+	else if constexpr (__have_sse4_1)
+	  return _mm_testnzc_si128(__intrin_bitcast<__m128i>(__to_intrin(__a)),
+				   __intrin_bitcast<__m128i>(__to_intrin(__b)));
+	else
+	  return __movemask(0 =3D=3D __and(__a, __b)) =3D=3D 0
+		 && __movemask(0 =3D=3D __andnot(__a, __b)) =3D=3D 0;
+      }
+    else
+      return !(__is_zero(__and(__a, __b)) || __is_zero(__andnot(__a, __b))=
);
+  }
+
+// }}}
+// __xzyw{{{
+// shuffles the complete vector, swapping the inner two quarters. Often us=
eful
+// for AVX for fixing up a shuffle result.
+template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+  _GLIBCXX_SIMD_INTRINSIC _Tp
+  __xzyw(_Tp __a)
+  {
+    if constexpr (sizeof(_Tp) =3D=3D 16)
+      {
+	const auto __x =3D __vector_bitcast<conditional_t<
+	  is_floating_point_v<typename _TVT::value_type>, float, int>>(__a);
+	return reinterpret_cast<_Tp>(
+	  decltype(__x){__x[0], __x[2], __x[1], __x[3]});
+      }
+    else if constexpr (sizeof(_Tp) =3D=3D 32)
+      {
+	const auto __x =3D __vector_bitcast<conditional_t<
+	  is_floating_point_v<typename _TVT::value_type>, double, _LLong>>(__a);
+	return reinterpret_cast<_Tp>(
+	  decltype(__x){__x[0], __x[2], __x[1], __x[3]});
+      }
+    else if constexpr (sizeof(_Tp) =3D=3D 64)
+      {
+	const auto __x =3D __vector_bitcast<conditional_t<
+	  is_floating_point_v<typename _TVT::value_type>, double, _LLong>>(__a);
+	return reinterpret_cast<_Tp>(decltype(__x){__x[0], __x[1], __x[4],
+						   __x[5], __x[2], __x[3],
+						   __x[6], __x[7]});
+      }
+    else
+      __assert_unreachable<_Tp>();
+  }
+
+// }}}
+// __maskload_epi32{{{
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __maskload_epi32(const int* __ptr, _Tp __k)
+  {
+    if constexpr (sizeof(__k) =3D=3D 16)
+      return _mm_maskload_epi32(__ptr, __k);
+    else
+      return _mm256_maskload_epi32(__ptr, __k);
+  }
+
+// }}}
+// __maskload_epi64{{{
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __maskload_epi64(const _LLong* __ptr, _Tp __k)
+  {
+    if constexpr (sizeof(__k) =3D=3D 16)
+      return _mm_maskload_epi64(__ptr, __k);
+    else
+      return _mm256_maskload_epi64(__ptr, __k);
+  }
+
+// }}}
+// __maskload_ps{{{
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __maskload_ps(const float* __ptr, _Tp __k)
+  {
+    if constexpr (sizeof(__k) =3D=3D 16)
+      return _mm_maskload_ps(__ptr, __k);
+    else
+      return _mm256_maskload_ps(__ptr, __k);
+  }
+
+// }}}
+// __maskload_pd{{{
+template <typename _Tp>
+  _GLIBCXX_SIMD_INTRINSIC auto
+  __maskload_pd(const double* __ptr, _Tp __k)
+  {
+    if constexpr (sizeof(__k) =3D=3D 16)
+      return _mm_maskload_pd(__ptr, __k);
+    else
+      return _mm256_maskload_pd(__ptr, __k);
+  }
+
+// }}}
+
+#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
+#include "simd_x86_conversions.h"
+#endif
+
+// ISA & type detection {{{
+template <typename _Tp, size_t _Np>
+  constexpr bool
+  __is_sse_ps()
+  {
+    return __have_sse
+	   && is_same_v<_Tp,
+			float> && sizeof(__intrinsic_type_t<_Tp, _Np>) =3D=3D 16;
+  }
+
+template <typename _Tp, size_t _Np>
+  constexpr bool
+  __is_sse_pd()
+  {
+    return __have_sse2
+	   && is_same_v<_Tp,
+			double> && sizeof(__intrinsic_type_t<_Tp, _Np>) =3D=3D 16;
+  }
+
+template <typename _Tp, size_t _Np>
+  constexpr bool
+  __is_avx_ps()
+  {
+    return __have_avx
+	   && is_same_v<_Tp,
+			float> && sizeof(__intrinsic_type_t<_Tp, _Np>) =3D=3D 32;
+  }
+
+template <typename _Tp, size_t _Np>
+  constexpr bool
+  __is_avx_pd()
+  {
+    return __have_avx
+	   && is_same_v<_Tp,
+			double> && sizeof(__intrinsic_type_t<_Tp, _Np>) =3D=3D 32;
+  }
+
+template <typename _Tp, size_t _Np>
+  constexpr bool
+  __is_avx512_ps()
+  {
+    return __have_avx512f
+	   && is_same_v<_Tp,
+			float> && sizeof(__intrinsic_type_t<_Tp, _Np>) =3D=3D 64;
+  }
+
+template <typename _Tp, size_t _Np>
+  constexpr bool
+  __is_avx512_pd()
+  {
+    return __have_avx512f
+	   && is_same_v<_Tp,
+			double> && sizeof(__intrinsic_type_t<_Tp, _Np>) =3D=3D 64;
+  }
+
+// }}}
+struct _MaskImplX86Mixin;
+
+// _CommonImplX86 {{{
+struct _CommonImplX86 : _CommonImplBuiltin
+{
+#ifdef _GLIBCXX_SIMD_WORKAROUND_PR85048
+  // _S_converts_via_decomposition {{{
+  template <typename _From, typename _To, size_t _ToSize>
+    static constexpr bool _S_converts_via_decomposition()
+    {
+      if constexpr (is_integral_v<
+		      _From> && is_integral_v<_To> && sizeof(_From) =3D=3D 8
+		    && _ToSize =3D=3D 16)
+	return (sizeof(_To) =3D=3D 2 && !__have_ssse3)
+	       || (sizeof(_To) =3D=3D 1 && !__have_avx512f);
+      else if constexpr (is_floating_point_v<_From> && is_integral_v<_To>)
+	return ((sizeof(_From) =3D=3D 4 || sizeof(_From) =3D=3D 8) && sizeof(_To)=
 =3D=3D 8
+		&& !__have_avx512dq)
+	       || (sizeof(_From) =3D=3D 8 && sizeof(_To) =3D=3D 4 && !__have_sse4=
_1
+		   && _ToSize =3D=3D 16);
+      else if constexpr (
+	is_integral_v<_From> && is_floating_point_v<_To> && sizeof(_From) =3D=3D 8
+	&& !__have_avx512dq)
+	return (sizeof(_To) =3D=3D 4 && _ToSize =3D=3D 16)
+	       || (sizeof(_To) =3D=3D 8 && _ToSize < 64);
+      else
+	return false;
+    }
+
+  template <typename _From, typename _To, size_t _ToSize>
+    static inline constexpr bool __converts_via_decomposition_v
+      =3D _S_converts_via_decomposition<_From, _To, _ToSize>();
+
+  // }}}
+#endif
+  // _S_store {{{
+  using _CommonImplBuiltin::_S_store;
+
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> __=
x,
+						 void* __addr)
+    {
+      constexpr size_t _Bytes =3D _Np * sizeof(_Tp);
+
+      if constexpr ((_Bytes & (_Bytes - 1)) !=3D 0 && __have_avx512bw_vl)
+	{
+	  const auto __v =3D __to_intrin(__x);
+
+	  if constexpr (_Bytes & 1)
+	    {
+	      if constexpr (_Bytes < 16)
+		_mm_mask_storeu_epi8(__addr, 0xffffu >> (16 - _Bytes),
+				     __intrin_bitcast<__m128i>(__v));
+	      else if constexpr (_Bytes < 32)
+		_mm256_mask_storeu_epi8(__addr, 0xffffffffu >> (32 - _Bytes),
+					__intrin_bitcast<__m256i>(__v));
+	      else
+		_mm512_mask_storeu_epi8(__addr,
+					0xffffffffffffffffull >> (64 - _Bytes),
+					__intrin_bitcast<__m512i>(__v));
+	    }
+	  else if constexpr (_Bytes & 2)
+	    {
+	      if constexpr (_Bytes < 16)
+		_mm_mask_storeu_epi16(__addr, 0xffu >> (8 - _Bytes / 2),
+				      __intrin_bitcast<__m128i>(__v));
+	      else if constexpr (_Bytes < 32)
+		_mm256_mask_storeu_epi16(__addr, 0xffffu >> (16 - _Bytes / 2),
+					 __intrin_bitcast<__m256i>(__v));
+	      else
+		_mm512_mask_storeu_epi16(__addr,
+					 0xffffffffull >> (32 - _Bytes / 2),
+					 __intrin_bitcast<__m512i>(__v));
+	    }
+	  else if constexpr (_Bytes & 4)
+	    {
+	      if constexpr (_Bytes < 16)
+		_mm_mask_storeu_epi32(__addr, 0xfu >> (4 - _Bytes / 4),
+				      __intrin_bitcast<__m128i>(__v));
+	      else if constexpr (_Bytes < 32)
+		_mm256_mask_storeu_epi32(__addr, 0xffu >> (8 - _Bytes / 4),
+					 __intrin_bitcast<__m256i>(__v));
+	      else
+		_mm512_mask_storeu_epi32(__addr, 0xffffull >> (16 - _Bytes / 4),
+					 __intrin_bitcast<__m512i>(__v));
+	    }
+	  else
+	    {
+	      static_assert(
+		_Bytes > 16,
+		"_Bytes < 16 && (_Bytes & 7) =3D=3D 0 && (_Bytes & (_Bytes "
+		"- 1)) !=3D 0 is impossible");
+	      if constexpr (_Bytes < 32)
+		_mm256_mask_storeu_epi64(__addr, 0xfu >> (4 - _Bytes / 8),
+					 __intrin_bitcast<__m256i>(__v));
+	      else
+		_mm512_mask_storeu_epi64(__addr, 0xffull >> (8 - _Bytes / 8),
+					 __intrin_bitcast<__m512i>(__v));
+	    }
+	}
+      else
+	_CommonImplBuiltin::_S_store(__x, __addr);
+    }
+
+  // }}}
+  // _S_store_bool_array(_BitMask) {{{
+  template <size_t _Np, bool _Sanitized>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr void
+    _S_store_bool_array(const _BitMask<_Np, _Sanitized> __x, bool* __mem)
+    {
+      if constexpr (__have_avx512bw_vl) // don't care for BW w/o VL
+	_S_store<_Np>(1 & __vector_bitcast<_UChar, _Np>([=3D]() constexpr {
+			if constexpr (_Np <=3D 16)
+			  return _mm_movm_epi8(__x._M_to_bits());
+			else if constexpr (_Np <=3D 32)
+			  return _mm256_movm_epi8(__x._M_to_bits());
+			else if constexpr (_Np <=3D 64)
+			  return _mm512_movm_epi8(__x._M_to_bits());
+			else
+			  __assert_unreachable<_SizeConstant<_Np>>();
+		      }()),
+		      __mem);
+      else if constexpr (__have_bmi2)
+	{
+	  if constexpr (_Np <=3D 4)
+	    _S_store<_Np>(_pdep_u32(__x._M_to_bits(), 0x01010101U), __mem);
+	  else
+	    __execute_n_times<__div_roundup(_Np, sizeof(size_t))>(
+	      [&](auto __i) {
+		constexpr size_t __offset =3D __i * sizeof(size_t);
+		constexpr int __todo =3D std::min(sizeof(size_t), _Np - __offset);
+		if constexpr (__todo =3D=3D 1)
+		  __mem[__offset] =3D __x[__offset];
+		else
+		  {
+		    const auto __bools =3D
+#ifdef __x86_64__
+		      _pdep_u64(__x.template _M_extract<__offset>().to_ullong(),
+				0x0101010101010101ULL);
+#else // __x86_64__
+		      _pdep_u32(
+			__x.template _M_extract<__offset>()._M_to_bits(),
+			0x01010101U);
+#endif // __x86_64__
+		    _S_store<__todo>(__bools, __mem + __offset);
+		  }
+	      });
+	}
+      else if constexpr (__have_sse2 && _Np > 7)
+	__execute_n_times<__div_roundup(_Np, 16)>([&](auto __i) {
+	  constexpr int __offset =3D __i * 16;
+	  constexpr int __todo =3D std::min(16, int(_Np) - __offset);
+	  const int __bits =3D __x.template _M_extract<__offset>()._M_to_bits();
+	  __vector_type16_t<_UChar> __bools;
+	  if constexpr (__have_avx512f)
+	    {
+	      auto __as32bits
+		=3D _mm512_maskz_mov_epi32(__bits, __to_intrin(
+						   __vector_broadcast<16>(1)));
+	      auto __as16bits
+		=3D __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
+					    __todo > 8 ? __hi256(__as32bits)
+						       : __m256i()));
+	      __bools =3D __vector_bitcast<_UChar>(
+		_mm_packs_epi16(__lo128(__as16bits), __hi128(__as16bits)));
+	    }
+	  else
+	    {
+	      using _V =3D __vector_type_t<_UChar, 16>;
+	      auto __tmp =3D _mm_cvtsi32_si128(__bits);
+	      __tmp =3D _mm_unpacklo_epi8(__tmp, __tmp);
+	      __tmp =3D _mm_unpacklo_epi16(__tmp, __tmp);
+	      __tmp =3D _mm_unpacklo_epi32(__tmp, __tmp);
+	      _V __tmp2 =3D reinterpret_cast<_V>(__tmp);
+	      __tmp2 &=3D _V{1, 2, 4, 8, 16, 32, 64, 128,
+			   1, 2, 4, 8, 16, 32, 64, 128}; // mask bit index
+	      __bools =3D (__tmp2 =3D=3D 0) + 1; // 0xff -> 0x00 | 0x00 -> 0x01
+	    }
+	  _S_store<__todo>(__bools, __mem + __offset);
+	});
+      else
+	_CommonImplBuiltin::_S_store_bool_array(__x, __mem);
+    }
+
+  // }}}
+  // _S_blend_avx512 {{{
+  // Returns: __k ? __b : __a
+  // TODO: reverse __a and __b to match COND_EXPR
+  // Requires: _TV to be a __vector_type_t matching valuetype for the bitm=
ask
+  //           __k
+  template <typename _Kp, typename _TV>
+    _GLIBCXX_SIMD_INTRINSIC static _TV
+    _S_blend_avx512(const _Kp __k, const _TV __a, const _TV __b) noexcept
+    {
+#ifdef __clang__
+      // FIXME: this does a boolean choice, not a blend
+      return __k ? __a : __b;
+#else
+      static_assert(__is_vector_type_v<_TV>);
+      using _Tp =3D typename _VectorTraits<_TV>::value_type;
+      static_assert(sizeof(_TV) >=3D 16);
+      static_assert(sizeof(_Tp) <=3D 8);
+      using _IntT
+	=3D conditional_t<(sizeof(_Tp) > 2),
+			conditional_t<sizeof(_Tp) =3D=3D 4, int, long long>,
+			conditional_t<sizeof(_Tp) =3D=3D 1, char, short>>;
+      [[maybe_unused]] const auto __aa =3D __vector_bitcast<_IntT>(__a);
+      [[maybe_unused]] const auto __bb =3D __vector_bitcast<_IntT>(__b);
+      if constexpr (sizeof(_TV) =3D=3D 64)
+	{
+	  if constexpr (sizeof(_Tp) =3D=3D 1)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmb_512_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 2)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmw_512_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 4 && is_floating_point_v<_Tp>)
+	    return __builtin_ia32_blendmps_512_mask(__a, __b, __k);
+	  else if constexpr (sizeof(_Tp) =3D=3D 4)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmd_512_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 8 && is_floating_point_v<_Tp>)
+	    return __builtin_ia32_blendmpd_512_mask(__a, __b, __k);
+	  else if constexpr (sizeof(_Tp) =3D=3D 8)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmq_512_mask(__aa, __bb, __k));
+	}
+      else if constexpr (sizeof(_TV) =3D=3D 32)
+	{
+	  if constexpr (sizeof(_Tp) =3D=3D 1)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmb_256_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 2)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmw_256_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 4 && is_floating_point_v<_Tp>)
+	    return __builtin_ia32_blendmps_256_mask(__a, __b, __k);
+	  else if constexpr (sizeof(_Tp) =3D=3D 4)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmd_256_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 8 && is_floating_point_v<_Tp>)
+	    return __builtin_ia32_blendmpd_256_mask(__a, __b, __k);
+	  else if constexpr (sizeof(_Tp) =3D=3D 8)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmq_256_mask(__aa, __bb, __k));
+	}
+      else if constexpr (sizeof(_TV) =3D=3D 16)
+	{
+	  if constexpr (sizeof(_Tp) =3D=3D 1)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmb_128_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 2)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmw_128_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 4 && is_floating_point_v<_Tp>)
+	    return __builtin_ia32_blendmps_128_mask(__a, __b, __k);
+	  else if constexpr (sizeof(_Tp) =3D=3D 4)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmd_128_mask(__aa, __bb, __k));
+	  else if constexpr (sizeof(_Tp) =3D=3D 8 && is_floating_point_v<_Tp>)
+	    return __builtin_ia32_blendmpd_128_mask(__a, __b, __k);
+	  else if constexpr (sizeof(_Tp) =3D=3D 8)
+	    return reinterpret_cast<_TV>(
+	      __builtin_ia32_blendmq_128_mask(__aa, __bb, __k));
+	}
+#endif
+    }
+
+  // }}}
+  // _S_blend_intrin {{{
+  // Returns: __k ? __b : __a
+  // TODO: reverse __a and __b to match COND_EXPR
+  // Requires: _Tp to be an intrinsic type (integers blend per byte) and 1=
6/32
+  //           Bytes wide
+  template <typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static _Tp _S_blend_intrin(_Tp __k, _Tp __a,
+						       _Tp __b) noexcept
+    {
+      static_assert(is_same_v<decltype(__to_intrin(__a)), _Tp>);
+      constexpr struct
+      {
+	_GLIBCXX_SIMD_INTRINSIC __m128 operator()(__m128 __a, __m128 __b,
+						  __m128 __k) const noexcept
+	{
+	  return __builtin_ia32_blendvps(__a, __b, __k);
+	}
+	_GLIBCXX_SIMD_INTRINSIC __m128d operator()(__m128d __a, __m128d __b,
+						   __m128d __k) const noexcept
+	{
+	  return __builtin_ia32_blendvpd(__a, __b, __k);
+	}
+	_GLIBCXX_SIMD_INTRINSIC __m128i operator()(__m128i __a, __m128i __b,
+						   __m128i __k) const noexcept
+	{
+	  return reinterpret_cast<__m128i>(
+	    __builtin_ia32_pblendvb128(reinterpret_cast<__v16qi>(__a),
+				       reinterpret_cast<__v16qi>(__b),
+				       reinterpret_cast<__v16qi>(__k)));
+	}
+	_GLIBCXX_SIMD_INTRINSIC __m256 operator()(__m256 __a, __m256 __b,
+						  __m256 __k) const noexcept
+	{
+	  return __builtin_ia32_blendvps256(__a, __b, __k);
+	}
+	_GLIBCXX_SIMD_INTRINSIC __m256d operator()(__m256d __a, __m256d __b,
+						   __m256d __k) const noexcept
+	{
+	  return __builtin_ia32_blendvpd256(__a, __b, __k);
+	}
+	_GLIBCXX_SIMD_INTRINSIC __m256i operator()(__m256i __a, __m256i __b,
+						   __m256i __k) const noexcept
+	{
+	  if constexpr (__have_avx2)
+	    return reinterpret_cast<__m256i>(
+	      __builtin_ia32_pblendvb256(reinterpret_cast<__v32qi>(__a),
+					 reinterpret_cast<__v32qi>(__b),
+					 reinterpret_cast<__v32qi>(__k)));
+	  else
+	    return reinterpret_cast<__m256i>(
+	      __builtin_ia32_blendvps256(reinterpret_cast<__v8sf>(__a),
+					 reinterpret_cast<__v8sf>(__b),
+					 reinterpret_cast<__v8sf>(__k)));
+	}
+      } __eval;
+      return __eval(__a, __b, __k);
+    }
+
+  // }}}
+  // _S_blend {{{
+  // Returns: __k ? __at1 : __at0
+  // TODO: reverse __at0 and __at1 to match COND_EXPR
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+    _S_blend(_SimdWrapper<bool, _Np> __k, _SimdWrapper<_Tp, _Np> __at0,
+	     _SimdWrapper<_Tp, _Np> __at1)
+    {
+      static_assert(is_same_v<_Tp, _Tp> && __have_avx512f);
+      if (__k._M_is_constprop() && __at0._M_is_constprop()
+	  && __at1._M_is_constprop())
+	return __generate_from_n_evaluations<_Np,
+					     __vector_type_t<_Tp, _Np>>([&](
+	  auto __i) constexpr { return __k[__i] ? __at1[__i] : __at0[__i]; });
+      else if constexpr (sizeof(__at0) =3D=3D 64
+			 || (__have_avx512vl && sizeof(__at0) >=3D 16))
+	return _S_blend_avx512(__k._M_data, __at0._M_data, __at1._M_data);
+      else
+	{
+	  static_assert((__have_avx512vl && sizeof(__at0) < 16)
+			|| !__have_avx512vl);
+	  constexpr size_t __size =3D (__have_avx512vl ? 16 : 64) / sizeof(_Tp);
+	  return __vector_bitcast<_Tp, _Np>(
+	    _S_blend_avx512(__k._M_data, __vector_bitcast<_Tp, __size>(__at0),
+			    __vector_bitcast<_Tp, __size>(__at1)));
+	}
+    }
+
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+    _S_blend(_SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k,
+	     _SimdWrapper<_Tp, _Np> __at0, _SimdWrapper<_Tp, _Np> __at1)
+    {
+      const auto __kk =3D __wrapper_bitcast<_Tp>(__k);
+      if (__builtin_is_constant_evaluated()
+	  || (__kk._M_is_constprop() && __at0._M_is_constprop()
+	      && __at1._M_is_constprop()))
+	{
+	  auto __r =3D __or(__andnot(__kk, __at0), __and(__kk, __at1));
+	  if (__r._M_is_constprop())
+	    return __r;
+	}
+      if constexpr (((__have_avx512f && sizeof(__at0) =3D=3D 64) || __have=
_avx512vl)
+		    && (sizeof(_Tp) >=3D 4 || __have_avx512bw))
+	// convert to bitmask and call overload above
+	return _S_blend(
+	  _SimdWrapper<bool, _Np>(
+	    __make_dependent_t<_Tp, _MaskImplX86Mixin>::_S_to_bits(__k)
+	      ._M_to_bits()),
+	  __at0, __at1);
+      else
+	{
+	  // Since GCC does not assume __k to be a mask, using the builtin
+	  // conditional operator introduces an extra compare against 0 before
+	  // blending. So we rather call the intrinsic here.
+	  if constexpr (__have_sse4_1)
+	    return _S_blend_intrin(__to_intrin(__kk), __to_intrin(__at0),
+				   __to_intrin(__at1));
+	  else
+	    return __or(__andnot(__kk, __at0), __and(__kk, __at1));
+	}
+    }
+
+  // }}}
+};
+
+// }}}
+// _SimdImplX86 {{{
+template <typename _Abi>
+  struct _SimdImplX86 : _SimdImplBuiltin<_Abi>
+  {
+    using _Base =3D _SimdImplBuiltin<_Abi>;
+
+    template <typename _Tp>
+      using _MaskMember =3D typename _Base::template _MaskMember<_Tp>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_full_size =3D _Abi::template _S_full_size=
<_Tp>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_size =3D _Abi::template _S_size<_Tp>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_max_store_size
+	=3D (sizeof(_Tp) >=3D 4 && __have_avx512f) || __have_avx512bw  ? 64
+	  : (is_floating_point_v<_Tp>&& __have_avx) || __have_avx2 ? 32
+								   : 16;
+    using _MaskImpl =3D typename _Abi::_MaskImpl;
+
+    // _S_masked_load {{{
+    template <typename _Tp, size_t _Np, typename _Up>
+      static inline _SimdWrapper<_Tp, _Np>
+      _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
+		     const _Up* __mem) noexcept
+      {
+	static_assert(_Np =3D=3D _S_size<_Tp>);
+	if constexpr (is_same_v<_Tp, _Up> || // no conversion
+		      (sizeof(_Tp) =3D=3D sizeof(_Up)
+		       && is_integral_v<
+			    _Tp> =3D=3D is_integral_v<_Up>) // conversion via bit
+							// reinterpretation
+	)
+	  {
+	    [[maybe_unused]] const auto __intrin =3D __to_intrin(__merge);
+	    if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512bw_vl)
+			  && sizeof(_Tp) =3D=3D 1)
+	      {
+		const auto __kk =3D _MaskImpl::_S_to_bits(__k)._M_to_bits();
+		if constexpr (sizeof(__intrin) =3D=3D 16)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm_mask_loadu_epi8(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__merge) =3D=3D 32)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm256_mask_loadu_epi8(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__merge) =3D=3D 64)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm512_mask_loadu_epi8(__intrin, __kk, __mem));
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512bw_vl)
+			       && sizeof(_Tp) =3D=3D 2)
+	      {
+		const auto __kk =3D _MaskImpl::_S_to_bits(__k)._M_to_bits();
+		if constexpr (sizeof(__intrin) =3D=3D 16)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm_mask_loadu_epi16(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 32)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm256_mask_loadu_epi16(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 64)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm512_mask_loadu_epi16(__intrin, __kk, __mem));
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
+			       && sizeof(_Tp) =3D=3D 4 && is_integral_v<_Up>)
+	      {
+		const auto __kk =3D _MaskImpl::_S_to_bits(__k)._M_to_bits();
+		if constexpr (sizeof(__intrin) =3D=3D 16)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm_mask_loadu_epi32(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 32)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm256_mask_loadu_epi32(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 64)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm512_mask_loadu_epi32(__intrin, __kk, __mem));
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
+			       && sizeof(_Tp) =3D=3D 4 && is_floating_point_v<_Up>)
+	      {
+		const auto __kk =3D _MaskImpl::_S_to_bits(__k)._M_to_bits();
+		if constexpr (sizeof(__intrin) =3D=3D 16)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm_mask_loadu_ps(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 32)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm256_mask_loadu_ps(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 64)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm512_mask_loadu_ps(__intrin, __kk, __mem));
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (__have_avx2 && sizeof(_Tp) =3D=3D 4
+			       && is_integral_v<_Up>)
+	      {
+		static_assert(sizeof(__intrin) =3D=3D 16 || sizeof(__intrin) =3D=3D 32);
+		__merge
+		  =3D __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
+			 __vector_bitcast<_Tp, _Np>(
+			   __maskload_epi32(reinterpret_cast<const int*>(__mem),
+					    __to_intrin(__k))));
+	      }
+	    else if constexpr (__have_avx && sizeof(_Tp) =3D=3D 4)
+	      {
+		static_assert(sizeof(__intrin) =3D=3D 16 || sizeof(__intrin) =3D=3D 32);
+		__merge
+		  =3D __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
+			 __vector_bitcast<_Tp, _Np>(
+			   __maskload_ps(reinterpret_cast<const float*>(__mem),
+					 __to_intrin(__k))));
+	      }
+	    else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
+			       && sizeof(_Tp) =3D=3D 8 && is_integral_v<_Up>)
+	      {
+		const auto __kk =3D _MaskImpl::_S_to_bits(__k)._M_to_bits();
+		if constexpr (sizeof(__intrin) =3D=3D 16)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm_mask_loadu_epi64(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 32)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm256_mask_loadu_epi64(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 64)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm512_mask_loadu_epi64(__intrin, __kk, __mem));
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr ((__is_avx512_abi<_Abi>() || __have_avx512vl)
+			       && sizeof(_Tp) =3D=3D 8 && is_floating_point_v<_Up>)
+	      {
+		const auto __kk =3D _MaskImpl::_S_to_bits(__k)._M_to_bits();
+		if constexpr (sizeof(__intrin) =3D=3D 16)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm_mask_loadu_pd(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 32)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm256_mask_loadu_pd(__intrin, __kk, __mem));
+		else if constexpr (sizeof(__intrin) =3D=3D 64)
+		  __merge =3D __vector_bitcast<_Tp, _Np>(
+		    _mm512_mask_loadu_pd(__intrin, __kk, __mem));
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (__have_avx2 && sizeof(_Tp) =3D=3D 8
+			       && is_integral_v<_Up>)
+	      {
+		static_assert(sizeof(__intrin) =3D=3D 16 || sizeof(__intrin) =3D=3D 32);
+		__merge
+		  =3D __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
+			 __vector_bitcast<_Tp, _Np>(__maskload_epi64(
+			   reinterpret_cast<const _LLong*>(__mem),
+			   __to_intrin(__k))));
+	      }
+	    else if constexpr (__have_avx && sizeof(_Tp) =3D=3D 8)
+	      {
+		static_assert(sizeof(__intrin) =3D=3D 16 || sizeof(__intrin) =3D=3D 32);
+		__merge
+		  =3D __or(__andnot(__vector_bitcast<_Tp>(__k), __merge._M_data),
+			 __vector_bitcast<_Tp, _Np>(
+			   __maskload_pd(reinterpret_cast<const double*>(__mem),
+					 __to_intrin(__k))));
+	      }
+	    else
+	      _BitOps::_S_bit_iteration(_MaskImpl::_S_to_bits(__k),
+					[&](auto __i) {
+					  __merge._M_set(__i, static_cast<_Tp>(
+								__mem[__i]));
+					});
+	  }
+	/* Very uncertain, that the following improves anything. Needs
+	benchmarking
+	 * before it's activated.
+	else if constexpr (sizeof(_Up) <=3D 8 && // no long double
+			   !__converts_via_decomposition_v<
+			     _Up, _Tp,
+			     sizeof(__merge)> // conversion via decomposition
+					      // is better handled via the
+					      // bit_iteration fallback below
+	)
+	  {
+	    // TODO: copy pattern from _S_masked_store, which doesn't resort to
+	    // fixed_size
+	    using _Ap       =3D simd_abi::deduce_t<_Up, _Np>;
+	    using _ATraits =3D _SimdTraits<_Up, _Ap>;
+	    using _AImpl   =3D typename _ATraits::_SimdImpl;
+	    typename _ATraits::_SimdMember __uncvted{};
+	    typename _ATraits::_MaskMember __kk =3D _Ap::_MaskImpl::template
+	_S_convert<_Up>(__k);
+	    __uncvted =3D _AImpl::_S_masked_load(__uncvted, __kk, __mem);
+	    _SimdConverter<_Up, _Ap, _Tp, _Abi> __converter;
+	    _Base::_S_masked_assign(__k, __merge, __converter(__uncvted));
+	  }
+	  */
+	else
+	  __merge =3D _Base::_S_masked_load(__merge, __k, __mem);
+	return __merge;
+      }
+
+    // }}}
+    // _S_masked_store_nocvt {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
+			    _SimdWrapper<bool, _Np> __k)
+      {
+	[[maybe_unused]] const auto __vi =3D __to_intrin(__v);
+	if constexpr (sizeof(__vi) =3D=3D 64)
+	  {
+	    static_assert(sizeof(__v) =3D=3D 64 && __have_avx512f);
+	    if constexpr (__have_avx512bw && sizeof(_Tp) =3D=3D 1)
+	      _mm512_mask_storeu_epi8(__mem, __k, __vi);
+	    else if constexpr (__have_avx512bw && sizeof(_Tp) =3D=3D 2)
+	      _mm512_mask_storeu_epi16(__mem, __k, __vi);
+	    else if constexpr (__have_avx512f && sizeof(_Tp) =3D=3D 4)
+	      {
+		if constexpr (is_integral_v<_Tp>)
+		  _mm512_mask_storeu_epi32(__mem, __k, __vi);
+		else
+		  _mm512_mask_storeu_ps(__mem, __k, __vi);
+	      }
+	    else if constexpr (__have_avx512f && sizeof(_Tp) =3D=3D 8)
+	      {
+		if constexpr (is_integral_v<_Tp>)
+		  _mm512_mask_storeu_epi64(__mem, __k, __vi);
+		else
+		  _mm512_mask_storeu_pd(__mem, __k, __vi);
+	      }
+#if 0 // with KNL either sizeof(_Tp) >=3D 4 or sizeof(_vi) <=3D 32
+      // with Skylake-AVX512, __have_avx512bw is true
+	  else if constexpr (__have_sse2)
+	    {
+	      using _M   =3D __vector_type_t<_Tp, _Np>;
+	      using _MVT =3D _VectorTraits<_M>;
+	      _mm_maskmoveu_si128(__auto_bitcast(__extract<0, 4>(__v._M_data)),
+				  __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(__k._M_data)=
),
+				  reinterpret_cast<char*>(__mem));
+	      _mm_maskmoveu_si128(__auto_bitcast(__extract<1, 4>(__v._M_data)),
+				  __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(
+				    __k._M_data >> 1 * _MVT::_S_full_size)),
+				  reinterpret_cast<char*>(__mem) + 1 * 16);
+	      _mm_maskmoveu_si128(__auto_bitcast(__extract<2, 4>(__v._M_data)),
+				  __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(
+				    __k._M_data >> 2 * _MVT::_S_full_size)),
+				  reinterpret_cast<char*>(__mem) + 2 * 16);
+	      if constexpr (_Np > 48 / sizeof(_Tp))
+		_mm_maskmoveu_si128(
+		  __auto_bitcast(__extract<3, 4>(__v._M_data)),
+		  __auto_bitcast(_MaskImpl::template _S_convert<_Tp, _Np>(
+		    __k._M_data >> 3 * _MVT::_S_full_size)),
+		  reinterpret_cast<char*>(__mem) + 3 * 16);
+	    }
+#endif
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (sizeof(__vi) =3D=3D 32)
+	  {
+	    if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 1)
+	      _mm256_mask_storeu_epi8(__mem, __k, __vi);
+	    else if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 2)
+	      _mm256_mask_storeu_epi16(__mem, __k, __vi);
+	    else if constexpr (__have_avx512vl && sizeof(_Tp) =3D=3D 4)
+	      {
+		if constexpr (is_integral_v<_Tp>)
+		  _mm256_mask_storeu_epi32(__mem, __k, __vi);
+		else
+		  _mm256_mask_storeu_ps(__mem, __k, __vi);
+	      }
+	    else if constexpr (__have_avx512vl && sizeof(_Tp) =3D=3D 8)
+	      {
+		if constexpr (is_integral_v<_Tp>)
+		  _mm256_mask_storeu_epi64(__mem, __k, __vi);
+		else
+		  _mm256_mask_storeu_pd(__mem, __k, __vi);
+	      }
+	    else if constexpr (__have_avx512f
+			       && (sizeof(_Tp) >=3D 4 || __have_avx512bw))
+	      {
+		// use a 512-bit maskstore, using zero-extension of the bitmask
+		_S_masked_store_nocvt(
+		  _SimdWrapper64<_Tp>(
+		    __intrin_bitcast<__vector_type64_t<_Tp>>(__v._M_data)),
+		  __mem, _SimdWrapper<bool, 64 / sizeof(_Tp)>(__k._M_data));
+	      }
+	    else
+	      _S_masked_store_nocvt(__v, __mem,
+				    _MaskImpl::template _S_to_maskvector<
+				      __int_for_sizeof_t<_Tp>, _Np>(__k));
+	  }
+	else if constexpr (sizeof(__vi) =3D=3D 16)
+	  {
+	    if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 1)
+	      _mm_mask_storeu_epi8(__mem, __k, __vi);
+	    else if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 2)
+	      _mm_mask_storeu_epi16(__mem, __k, __vi);
+	    else if constexpr (__have_avx512vl && sizeof(_Tp) =3D=3D 4)
+	      {
+		if constexpr (is_integral_v<_Tp>)
+		  _mm_mask_storeu_epi32(__mem, __k, __vi);
+		else
+		  _mm_mask_storeu_ps(__mem, __k, __vi);
+	      }
+	    else if constexpr (__have_avx512vl && sizeof(_Tp) =3D=3D 8)
+	      {
+		if constexpr (is_integral_v<_Tp>)
+		  _mm_mask_storeu_epi64(__mem, __k, __vi);
+		else
+		  _mm_mask_storeu_pd(__mem, __k, __vi);
+	      }
+	    else if constexpr (__have_avx512f
+			       && (sizeof(_Tp) >=3D 4 || __have_avx512bw))
+	      {
+		// use a 512-bit maskstore, using zero-extension of the bitmask
+		_S_masked_store_nocvt(
+		  _SimdWrapper64<_Tp>(
+		    __intrin_bitcast<__intrinsic_type64_t<_Tp>>(__v._M_data)),
+		  __mem, _SimdWrapper<bool, 64 / sizeof(_Tp)>(__k._M_data));
+	      }
+	    else
+	      _S_masked_store_nocvt(__v, __mem,
+				    _MaskImpl::template _S_to_maskvector<
+				      __int_for_sizeof_t<_Tp>, _Np>(__k));
+	  }
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
+			    _SimdWrapper<__int_for_sizeof_t<_Tp>, _Np> __k)
+      {
+	if constexpr (sizeof(__v) <=3D 16)
+	  {
+	    [[maybe_unused]] const auto __vi
+	      =3D __intrin_bitcast<__m128i>(__as_vector(__v));
+	    [[maybe_unused]] const auto __ki
+	      =3D __intrin_bitcast<__m128i>(__as_vector(__k));
+	    if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 1)
+	      _mm_mask_storeu_epi8(__mem, _mm_movepi8_mask(__ki), __vi);
+	    else if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 2)
+	      _mm_mask_storeu_epi16(__mem, _mm_movepi16_mask(__ki), __vi);
+	    else if constexpr (__have_avx2 && sizeof(_Tp) =3D=3D 4
+			       && is_integral_v<_Tp>)
+	      _mm_maskstore_epi32(reinterpret_cast<int*>(__mem), __ki, __vi);
+	    else if constexpr (__have_avx && sizeof(_Tp) =3D=3D 4)
+	      _mm_maskstore_ps(reinterpret_cast<float*>(__mem), __ki,
+			       __vector_bitcast<float>(__vi));
+	    else if constexpr (__have_avx2 && sizeof(_Tp) =3D=3D 8
+			       && is_integral_v<_Tp>)
+	      _mm_maskstore_epi64(reinterpret_cast<_LLong*>(__mem), __ki, __vi);
+	    else if constexpr (__have_avx && sizeof(_Tp) =3D=3D 8)
+	      _mm_maskstore_pd(reinterpret_cast<double*>(__mem), __ki,
+			       __vector_bitcast<double>(__vi));
+	    else if constexpr (__have_sse2)
+	      _mm_maskmoveu_si128(__vi, __ki, reinterpret_cast<char*>(__mem));
+	  }
+	else if constexpr (sizeof(__v) =3D=3D 32)
+	  {
+	    [[maybe_unused]] const auto __vi
+	      =3D __intrin_bitcast<__m256i>(__as_vector(__v));
+	    [[maybe_unused]] const auto __ki
+	      =3D __intrin_bitcast<__m256i>(__as_vector(__k));
+	    if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 1)
+	      _mm256_mask_storeu_epi8(__mem, _mm256_movepi8_mask(__ki), __vi);
+	    else if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 2)
+	      _mm256_mask_storeu_epi16(__mem, _mm256_movepi16_mask(__ki), __vi);
+	    else if constexpr (__have_avx2 && sizeof(_Tp) =3D=3D 4
+			       && is_integral_v<_Tp>)
+	      _mm256_maskstore_epi32(reinterpret_cast<int*>(__mem), __ki, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4)
+	      _mm256_maskstore_ps(reinterpret_cast<float*>(__mem), __ki,
+				  __vector_bitcast<float>(__v));
+	    else if constexpr (__have_avx2 && sizeof(_Tp) =3D=3D 8
+			       && is_integral_v<_Tp>)
+	      _mm256_maskstore_epi64(reinterpret_cast<_LLong*>(__mem), __ki,
+				     __vi);
+	    else if constexpr (__have_avx && sizeof(_Tp) =3D=3D 8)
+	      _mm256_maskstore_pd(reinterpret_cast<double*>(__mem), __ki,
+				  __vector_bitcast<double>(__v));
+	    else if constexpr (__have_sse2)
+	      {
+		_mm_maskmoveu_si128(__lo128(__vi), __lo128(__ki),
+				    reinterpret_cast<char*>(__mem));
+		_mm_maskmoveu_si128(__hi128(__vi), __hi128(__ki),
+				    reinterpret_cast<char*>(__mem) + 16);
+	      }
+	  }
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    // }}}
+    // _S_masked_store {{{
+    template <typename _Tp, size_t _Np, typename _Up>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, _Up* __mem,
+		      const _MaskMember<_Tp> __k) noexcept
+      {
+	if constexpr (is_integral_v<
+			_Tp> && is_integral_v<_Up> && sizeof(_Tp) > sizeof(_Up)
+		      && __have_avx512f && (sizeof(_Tp) >=3D 4 || __have_avx512bw)
+		      && (sizeof(__v) =3D=3D 64 || __have_avx512vl))
+	  { // truncating store
+	    const auto __vi =3D __to_intrin(__v);
+	    const auto __kk =3D _MaskImpl::_S_to_bits(__k)._M_to_bits();
+	    if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 4
+			  && sizeof(__vi) =3D=3D 64)
+	      _mm512_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 4
+			       && sizeof(__vi) =3D=3D 32)
+	      _mm256_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 4
+			       && sizeof(__vi) =3D=3D 16)
+	      _mm_mask_cvtepi64_storeu_epi32(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 2
+			       && sizeof(__vi) =3D=3D 64)
+	      _mm512_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 2
+			       && sizeof(__vi) =3D=3D 32)
+	      _mm256_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 2
+			       && sizeof(__vi) =3D=3D 16)
+	      _mm_mask_cvtepi64_storeu_epi16(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 64)
+	      _mm512_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 32)
+	      _mm256_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 16)
+	      _mm_mask_cvtepi64_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 2
+			       && sizeof(__vi) =3D=3D 64)
+	      _mm512_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 2
+			       && sizeof(__vi) =3D=3D 32)
+	      _mm256_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 2
+			       && sizeof(__vi) =3D=3D 16)
+	      _mm_mask_cvtepi32_storeu_epi16(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 64)
+	      _mm512_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 32)
+	      _mm256_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 16)
+	      _mm_mask_cvtepi32_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 64)
+	      _mm512_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 32)
+	      _mm256_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
+	    else if constexpr (sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 1
+			       && sizeof(__vi) =3D=3D 16)
+	      _mm_mask_cvtepi16_storeu_epi8(__mem, __kk, __vi);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  _Base::_S_masked_store(__v, __mem, __k);
+      }
+
+    // }}}
+    // _S_multiplies {{{
+    template <typename _V, typename _VVT =3D _VectorTraits<_V>>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _V _S_multiplies(_V __x, _V=
 __y)
+      {
+	using _Tp =3D typename _VVT::value_type;
+	if (__builtin_is_constant_evaluated() || __x._M_is_constprop()
+	    || __y._M_is_constprop())
+	  return __as_vector(__x) * __as_vector(__y);
+	else if constexpr (sizeof(_Tp) =3D=3D 1)
+	  {
+	    if constexpr (sizeof(_V) =3D=3D 2)
+	      {
+		const auto __xs =3D reinterpret_cast<short>(__x._M_data);
+		const auto __ys =3D reinterpret_cast<short>(__y._M_data);
+		return reinterpret_cast<__vector_type_t<_Tp, 2>>(short(
+		  ((__xs * __ys) & 0xff) | ((__xs >> 8) * (__ys & 0xff00))));
+	      }
+	    else if constexpr (sizeof(_V) =3D=3D 4 && _VVT::_S_partial_width =3D=
=3D 3)
+	      {
+		const auto __xi =3D reinterpret_cast<int>(__x._M_data);
+		const auto __yi =3D reinterpret_cast<int>(__y._M_data);
+		return reinterpret_cast<__vector_type_t<_Tp, 3>>(
+		  ((__xi * __yi) & 0xff)
+		  | (((__xi >> 8) * (__yi & 0xff00)) & 0xff00)
+		  | ((__xi >> 16) * (__yi & 0xff0000)));
+	      }
+	    else if constexpr (sizeof(_V) =3D=3D 4)
+	      {
+		const auto __xi =3D reinterpret_cast<int>(__x._M_data);
+		const auto __yi =3D reinterpret_cast<int>(__y._M_data);
+		return reinterpret_cast<__vector_type_t<_Tp, 4>>(
+		  ((__xi * __yi) & 0xff)
+		  | (((__xi >> 8) * (__yi & 0xff00)) & 0xff00)
+		  | (((__xi >> 16) * (__yi & 0xff0000)) & 0xff0000)
+		  | ((__xi >> 24) * (__yi & 0xff000000u)));
+	      }
+	    else if constexpr (sizeof(_V) =3D=3D 8 && __have_avx2
+			       && is_signed_v<_Tp>)
+	      return __convert<typename _VVT::type>(
+		__vector_bitcast<short>(_mm_cvtepi8_epi16(__to_intrin(__x)))
+		* __vector_bitcast<short>(_mm_cvtepi8_epi16(__to_intrin(__y))));
+	    else if constexpr (sizeof(_V) =3D=3D 8 && __have_avx2
+			       && is_unsigned_v<_Tp>)
+	      return __convert<typename _VVT::type>(
+		__vector_bitcast<short>(_mm_cvtepu8_epi16(__to_intrin(__x)))
+		* __vector_bitcast<short>(_mm_cvtepu8_epi16(__to_intrin(__y))));
+	    else
+	      {
+		// codegen of `x*y` is suboptimal (as of GCC 9.0.1)
+		constexpr size_t __full_size =3D _VVT::_S_full_size;
+		constexpr int _Np =3D sizeof(_V) >=3D 16 ? __full_size / 2 : 8;
+		using _ShortW =3D _SimdWrapper<short, _Np>;
+		const _ShortW __even =3D __vector_bitcast<short, _Np>(__x)
+				       * __vector_bitcast<short, _Np>(__y);
+		_ShortW __high_byte =3D _ShortW()._M_data - 256;
+		//[&]() { asm("" : "+x"(__high_byte._M_data)); }();
+		const _ShortW __odd
+		  =3D (__vector_bitcast<short, _Np>(__x) >> 8)
+		    * (__vector_bitcast<short, _Np>(__y) & __high_byte._M_data);
+		if constexpr (__have_avx512bw && sizeof(_V) > 2)
+		  return _CommonImplX86::_S_blend_avx512(
+		    0xaaaa'aaaa'aaaa'aaaaLL, __vector_bitcast<_Tp>(__even),
+		    __vector_bitcast<_Tp>(__odd));
+		else if constexpr (__have_sse4_1 && sizeof(_V) > 2)
+		  return _CommonImplX86::_S_blend_intrin(__to_intrin(
+							   __high_byte),
+							 __to_intrin(__even),
+							 __to_intrin(__odd));
+		else
+		  return __to_intrin(
+		    __or(__andnot(__high_byte, __even), __odd));
+	      }
+	  }
+	else
+	  return _Base::_S_multiplies(__x, __y);
+      }
+
+    // }}}
+    // _S_divides {{{
+#ifdef _GLIBCXX_SIMD_WORKAROUND_PR90993
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_divides(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      {
+	if (!__builtin_is_constant_evaluated()
+	    && !__builtin_constant_p(__y._M_data))
+	  if constexpr (is_integral_v<_Tp> && sizeof(_Tp) <=3D 4)
+	    { // use divps - codegen of `x/y` is suboptimal (as of GCC 9.0.1)
+	      // Note that using floating-point division is likely to raise the
+	      // *Inexact* exception flag and thus appears like an invalid
+	      // "as-if" transformation. However, C++ doesn't specify how the
+	      // fpenv can be observed and points to C. C says that function
+	      // calls are assumed to potentially raise fp exceptions, unless
+	      // documented otherwise. Consequently, operator/, which is a
+	      // function call, may raise fp exceptions.
+	      /*const struct _CsrGuard
+	      {
+		const unsigned _M_data =3D _mm_getcsr();
+		_CsrGuard()
+		{
+		  _mm_setcsr(0x9f80); // turn off FP exceptions and
+	      flush-to-zero
+		}
+		~_CsrGuard() { _mm_setcsr(_M_data); }
+	      } __csr;*/
+	      using _Float =3D conditional_t<sizeof(_Tp) =3D=3D 4, double, float>;
+	      constexpr size_t __n_intermediate
+		=3D std::min(_Np, (__have_avx512f ? 64
+				 : __have_avx   ? 32
+						: 16)
+				  / sizeof(_Float));
+	      using _FloatV =3D __vector_type_t<_Float, __n_intermediate>;
+	      constexpr size_t __n_floatv
+		=3D __div_roundup(_Np, __n_intermediate);
+	      using _R =3D __vector_type_t<_Tp, _Np>;
+	      const auto __xf =3D __convert_all<_FloatV, __n_floatv>(__x);
+	      const auto __yf =3D __convert_all<_FloatV, __n_floatv>(
+		_Abi::__make_padding_nonzero(__as_vector(__y)));
+	      return __call_with_n_evaluations<__n_floatv>(
+		[](auto... __quotients) {
+		  return __vector_convert<_R>(__quotients...);
+		},
+		[&__xf,
+		 &__yf](auto __i) -> _SimdWrapper<_Float, __n_intermediate> {
+#if !defined __clang__ && __GCC_IEC_559 =3D=3D 0
+		  // If -freciprocal-math is active, using the `/` operator is
+		  // incorrect because it may be translated to an imprecise
+		  // multiplication with reciprocal. We need to use inline
+		  // assembly to force a real division.
+		  _FloatV __r;
+		  if constexpr (__have_avx) // -mno-sse2avx is irrelevant
+					    // because once -mavx is given, GCC
+					    // emits VEX encoded vdivp[sd]
+		    {
+		      if constexpr (sizeof(_Tp) =3D=3D 4)
+			asm("vdivpd\t{%2, %1, %0|%0, %1, %2}"
+			    : "=3Dx"(__r)
+			    : "x"(__xf[__i]), "x"(__yf[__i]));
+		      else
+			asm("vdivps\t{%2, %1, %0|%0, %1, %2}"
+			    : "=3Dx"(__r)
+			    : "x"(__xf[__i]), "x"(__yf[__i]));
+		    }
+		  else
+		    {
+		      __r =3D __xf[__i];
+		      if constexpr (sizeof(_Tp) =3D=3D 4)
+			asm("divpd\t{%1, %0|%0, %1}"
+			    : "=3Dx"(__r)
+			    : "x"(__yf[__i]));
+		      else
+			asm("divps\t{%1, %0|%0, %1}"
+			    : "=3Dx"(__r)
+			    : "x"(__yf[__i]));
+		    }
+		  return __r;
+#else
+		  return __xf[__i] / __yf[__i];
+#endif
+		});
+	    }
+	/* 64-bit int division is potentially optimizable via double division if
+	 * the value in __x is small enough and the conversion between
+	 * int<->double is efficient enough:
+	else if constexpr (is_integral_v<_Tp> && is_unsigned_v<_Tp> &&
+			   sizeof(_Tp) =3D=3D 8)
+	  {
+	    if constexpr (__have_sse4_1 && sizeof(__x) =3D=3D 16)
+	      {
+		if (_mm_test_all_zeros(__x, __m128i{0xffe0'0000'0000'0000ull,
+						    0xffe0'0000'0000'0000ull}))
+		  {
+		    __x._M_data | 0x __vector_convert<__m128d>(__x._M_data)
+		  }
+	      }
+	  }
+	  */
+	return _Base::_S_divides(__x, __y);
+      }
+  #endif // _GLIBCXX_SIMD_WORKAROUND_PR90993
+
+    // }}}
+    // _S_modulus {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_modulus(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      {
+	if (__builtin_is_constant_evaluated()
+	    || __builtin_constant_p(__y._M_data) || sizeof(_Tp) >=3D 8)
+	  return _Base::_S_modulus(__x, __y);
+	else
+	  return _Base::_S_minus(__x, _S_multiplies(__y, _S_divides(__x, __y)));
+      }
+
+    // }}}
+    // _S_bit_shift_left {{{
+    // Notes on UB. C++2a [expr.shift] says:
+    // -1- [...] The operands shall be of integral or unscoped enumeration=
 type
+    //     and integral promotions are performed. The type of the result i=
s that
+    //     of the promoted left operand. The behavior is undefined if the =
right
+    //     operand is negative, or greater than or equal to the width of t=
he
+    //     promoted left operand.
+    // -2- The value of E1 << E2 is the unique value congruent to E1=C3=97=
2^E2 modulo
+    //     2^N, where N is the width of the type of the result.
+    //
+    // C++17 [expr.shift] says:
+    // -2- The value of E1 << E2 is E1 left-shifted E2 bit positions; vaca=
ted
+    //     bits are zero-filled. If E1 has an unsigned type, the value of =
the
+    //     result is E1 =C3=97 2^E2 , reduced modulo one more than the max=
imum value
+    //     representable in the result type. Otherwise, if E1 has a signed=
 type
+    //     and non-negative value, and E1 =C3=97 2^E2 is representable in =
the
+    //     corresponding unsigned type of the result type, then that value,
+    //     converted to the result type, is the resulting value; otherwise=
, the
+    //     behavior is undefined.
+    //
+    // Consequences:
+    // With C++2a signed and unsigned types have the same UB
+    // characteristics:
+    // - left shift is not UB for 0 <=3D RHS < max(32, #bits(T))
+    //
+    // With C++17 there's little room for optimizations because the standa=
rd
+    // requires all shifts to happen on promoted integrals (i.e. int). Thu=
s,
+    // short and char shifts must assume shifts affect bits of neighboring
+    // values.
+  #ifndef _GLIBCXX_SIMD_NO_SHIFT_OPT
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      inline _GLIBCXX_CONST static typename _TVT::type
+      _S_bit_shift_left(_Tp __xx, int __y)
+      {
+	using _V =3D typename _TVT::type;
+	using _Up =3D typename _TVT::value_type;
+	_V __x =3D __xx;
+	[[maybe_unused]] const auto __ix =3D __to_intrin(__x);
+	if (__builtin_is_constant_evaluated())
+	  return __x << __y;
+#if __cplusplus > 201703
+	// after C++17, signed shifts have no UB, and behave just like unsigned
+	// shifts
+	else if constexpr (sizeof(_Up) =3D=3D 1 && is_signed_v<_Up>)
+	  return __vector_bitcast<_Up>(
+	    _S_bit_shift_left(__vector_bitcast<make_unsigned_t<_Up>>(__x),
+			      __y));
+#endif
+	else if constexpr (sizeof(_Up) =3D=3D 1)
+	  {
+	    // (cf. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D83894)
+	    if (__builtin_constant_p(__y))
+	      {
+		if (__y =3D=3D 0)
+		  return __x;
+		else if (__y =3D=3D 1)
+		  return __x + __x;
+		else if (__y =3D=3D 2)
+		  {
+		    __x =3D __x + __x;
+		    return __x + __x;
+		  }
+		else if (__y > 2 && __y < 8)
+		  {
+		    if constexpr (sizeof(__x) > sizeof(unsigned))
+		      {
+			const _UChar __mask =3D 0xff << __y; // precomputed vector
+			return __vector_bitcast<_Up>(
+			  __vector_bitcast<_UChar>(
+			    __vector_bitcast<unsigned>(__x) << __y)
+			  & __mask);
+		      }
+		    else
+		      {
+			const unsigned __mask
+			  =3D (0xff & (0xff << __y)) * 0x01010101u;
+			return reinterpret_cast<_V>(
+			  static_cast<__int_for_sizeof_t<_V>>(
+			    unsigned(
+			      reinterpret_cast<__int_for_sizeof_t<_V>>(__x)
+			      << __y)
+			    & __mask));
+		      }
+		  }
+		else if (__y >=3D 8 && __y < 32)
+		  return _V();
+		else
+		  __builtin_unreachable();
+	      }
+	    // general strategy in the following: use an sllv instead of sll
+	    // instruction, because it's 2 to 4 times faster:
+	    else if constexpr (__have_avx512bw_vl && sizeof(__x) =3D=3D 16)
+	      return __vector_bitcast<_Up>(_mm256_cvtepi16_epi8(
+		_mm256_sllv_epi16(_mm256_cvtepi8_epi16(__ix),
+				  _mm256_set1_epi16(__y))));
+	    else if constexpr (__have_avx512bw && sizeof(__x) =3D=3D 32)
+	      return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
+		_mm512_sllv_epi16(_mm512_cvtepi8_epi16(__ix),
+				  _mm512_set1_epi16(__y))));
+	    else if constexpr (__have_avx512bw && sizeof(__x) =3D=3D 64)
+	      {
+		const auto __shift =3D _mm512_set1_epi16(__y);
+		return __vector_bitcast<_Up>(
+		  __concat(_mm512_cvtepi16_epi8(_mm512_sllv_epi16(
+			     _mm512_cvtepi8_epi16(__lo256(__ix)), __shift)),
+			   _mm512_cvtepi16_epi8(_mm512_sllv_epi16(
+			     _mm512_cvtepi8_epi16(__hi256(__ix)), __shift))));
+	      }
+	    else if constexpr (__have_avx2 && sizeof(__x) =3D=3D 32)
+	      {
+#if 1
+		const auto __shift =3D _mm_cvtsi32_si128(__y);
+		auto __k
+		  =3D _mm256_sll_epi16(_mm256_slli_epi16(~__m256i(), 8), __shift);
+		__k |=3D _mm256_srli_epi16(__k, 8);
+		return __vector_bitcast<_Up>(_mm256_sll_epi32(__ix, __shift)
+					     & __k);
+#else
+		const _Up __k =3D 0xff << __y;
+		return __vector_bitcast<_Up>(__vector_bitcast<int>(__x) << __y)
+		       & __k;
+#endif
+	      }
+	    else
+	      {
+		const auto __shift =3D _mm_cvtsi32_si128(__y);
+		auto __k
+		  =3D _mm_sll_epi16(_mm_slli_epi16(~__m128i(), 8), __shift);
+		__k |=3D _mm_srli_epi16(__k, 8);
+		return __intrin_bitcast<_V>(_mm_sll_epi16(__ix, __shift) & __k);
+	      }
+	  }
+	return __x << __y;
+      }
+
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      inline _GLIBCXX_CONST static typename _TVT::type
+      _S_bit_shift_left(_Tp __xx, typename _TVT::type __y)
+      {
+	using _V =3D typename _TVT::type;
+	using _Up =3D typename _TVT::value_type;
+	_V __x =3D __xx;
+	[[maybe_unused]] const auto __ix =3D __to_intrin(__x);
+	[[maybe_unused]] const auto __iy =3D __to_intrin(__y);
+	if (__builtin_is_constant_evaluated())
+	  return __x << __y;
+#if __cplusplus > 201703
+	// after C++17, signed shifts have no UB, and behave just like unsigned
+	// shifts
+	else if constexpr (is_signed_v<_Up>)
+	  return __vector_bitcast<_Up>(
+	    _S_bit_shift_left(__vector_bitcast<make_unsigned_t<_Up>>(__x),
+			      __vector_bitcast<make_unsigned_t<_Up>>(__y)));
+#endif
+	else if constexpr (sizeof(_Up) =3D=3D 1)
+	  {
+	    if constexpr (sizeof __ix =3D=3D 64 && __have_avx512bw)
+	      return __vector_bitcast<_Up>(__concat(
+		_mm512_cvtepi16_epi8(
+		  _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__lo256(__ix)),
+				    _mm512_cvtepu8_epi16(__lo256(__iy)))),
+		_mm512_cvtepi16_epi8(
+		  _mm512_sllv_epi16(_mm512_cvtepu8_epi16(__hi256(__ix)),
+				    _mm512_cvtepu8_epi16(__hi256(__iy))))));
+	    else if constexpr (sizeof __ix =3D=3D 32 && __have_avx512bw)
+	      return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
+		_mm512_sllv_epi16(_mm512_cvtepu8_epi16(__ix),
+				  _mm512_cvtepu8_epi16(__iy))));
+	    else if constexpr (sizeof __x <=3D 8 && __have_avx512bw_vl)
+	      return __intrin_bitcast<_V>(
+		_mm_cvtepi16_epi8(_mm_sllv_epi16(_mm_cvtepu8_epi16(__ix),
+						 _mm_cvtepu8_epi16(__iy))));
+	    else if constexpr (sizeof __ix =3D=3D 16 && __have_avx512bw_vl)
+	      return __intrin_bitcast<_V>(_mm256_cvtepi16_epi8(
+		_mm256_sllv_epi16(_mm256_cvtepu8_epi16(__ix),
+				  _mm256_cvtepu8_epi16(__iy))));
+	    else if constexpr (sizeof __ix =3D=3D 16 && __have_avx512bw)
+	      return __intrin_bitcast<_V>(
+		__lo128(_mm512_cvtepi16_epi8(_mm512_sllv_epi16(
+		  _mm512_cvtepu8_epi16(_mm256_castsi128_si256(__ix)),
+		  _mm512_cvtepu8_epi16(_mm256_castsi128_si256(__iy))))));
+	    else if constexpr (__have_sse4_1 && sizeof(__x) =3D=3D 16)
+	      {
+		auto __mask
+		  =3D __vector_bitcast<_Up>(__vector_bitcast<short>(__y) << 5);
+		auto __x4
+		  =3D __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 4);
+		__x4 &=3D char(0xf0);
+		__x =3D reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x4)));
+		__mask +=3D __mask;
+		auto __x2
+		  =3D __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 2);
+		__x2 &=3D char(0xfc);
+		__x =3D reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x2)));
+		__mask +=3D __mask;
+		auto __x1 =3D __x + __x;
+		__x =3D reinterpret_cast<_V>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__x), __to_intrin(__x1)));
+		return __x
+		       & ((__y & char(0xf8)) =3D=3D 0); // y > 7 nulls the result
+	      }
+	    else if constexpr (sizeof(__x) =3D=3D 16)
+	      {
+		auto __mask
+		  =3D __vector_bitcast<_UChar>(__vector_bitcast<short>(__y) << 5);
+		auto __x4
+		  =3D __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 4);
+		__x4 &=3D char(0xf0);
+		__x =3D __vector_bitcast<_SChar>(__mask) < 0 ? __x4 : __x;
+		__mask +=3D __mask;
+		auto __x2
+		  =3D __vector_bitcast<_Up>(__vector_bitcast<short>(__x) << 2);
+		__x2 &=3D char(0xfc);
+		__x =3D __vector_bitcast<_SChar>(__mask) < 0 ? __x2 : __x;
+		__mask +=3D __mask;
+		auto __x1 =3D __x + __x;
+		__x =3D __vector_bitcast<_SChar>(__mask) < 0 ? __x1 : __x;
+		return __x
+		       & ((__y & char(0xf8)) =3D=3D 0); // y > 7 nulls the result
+	      }
+	    else
+	      return __x << __y;
+	  }
+	else if constexpr (sizeof(_Up) =3D=3D 2)
+	  {
+	    if constexpr (sizeof __ix =3D=3D 64 && __have_avx512bw)
+	      return __vector_bitcast<_Up>(_mm512_sllv_epi16(__ix, __iy));
+	    else if constexpr (sizeof __ix =3D=3D 32 && __have_avx512bw_vl)
+	      return __vector_bitcast<_Up>(_mm256_sllv_epi16(__ix, __iy));
+	    else if constexpr (sizeof __ix =3D=3D 32 && __have_avx512bw)
+	      return __vector_bitcast<_Up>(
+		__lo256(_mm512_sllv_epi16(_mm512_castsi256_si512(__ix),
+					  _mm512_castsi256_si512(__iy))));
+	    else if constexpr (sizeof __ix =3D=3D 32 && __have_avx2)
+	      {
+		const auto __ux =3D __vector_bitcast<unsigned>(__x);
+		const auto __uy =3D __vector_bitcast<unsigned>(__y);
+		return __vector_bitcast<_Up>(_mm256_blend_epi16(
+		  __auto_bitcast(__ux << (__uy & 0x0000ffffu)),
+		  __auto_bitcast((__ux & 0xffff0000u) << (__uy >> 16)), 0xaa));
+	      }
+	    else if constexpr (sizeof __ix =3D=3D 16 && __have_avx512bw_vl)
+	      return __intrin_bitcast<_V>(_mm_sllv_epi16(__ix, __iy));
+	    else if constexpr (sizeof __ix =3D=3D 16 && __have_avx512bw)
+	      return __intrin_bitcast<_V>(
+		__lo128(_mm512_sllv_epi16(_mm512_castsi128_si512(__ix),
+					  _mm512_castsi128_si512(__iy))));
+	    else if constexpr (sizeof __ix =3D=3D 16 && __have_avx2)
+	      {
+		const auto __ux =3D __vector_bitcast<unsigned>(__ix);
+		const auto __uy =3D __vector_bitcast<unsigned>(__iy);
+		return __intrin_bitcast<_V>(_mm_blend_epi16(
+		  __auto_bitcast(__ux << (__uy & 0x0000ffffu)),
+		  __auto_bitcast((__ux & 0xffff0000u) << (__uy >> 16)), 0xaa));
+	      }
+	    else if constexpr (sizeof __ix =3D=3D 16)
+	      {
+		using _Float4 =3D __vector_type_t<float, 4>;
+		using _Int4 =3D __vector_type_t<int, 4>;
+		using _UInt4 =3D __vector_type_t<unsigned, 4>;
+		const _UInt4 __yu
+		  =3D reinterpret_cast<_UInt4>(__to_intrin(__y + (0x3f8 >> 3)));
+		return __x
+		       * __intrin_bitcast<_V>(
+			 __vector_convert<_Int4>(_SimdWrapper<float, 4>(
+			   reinterpret_cast<_Float4>(__yu << 23)))
+			 | (__vector_convert<_Int4>(_SimdWrapper<float, 4>(
+			      reinterpret_cast<_Float4>((__yu >> 16) << 23)))
+			    << 16));
+	      }
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (sizeof(_Up) =3D=3D 4 && sizeof __ix =3D=3D 16
+			   && !__have_avx2)
+	  // latency is suboptimal, but throughput is at full speedup
+	  return __intrin_bitcast<_V>(
+	    __vector_bitcast<unsigned>(__ix)
+	    * __vector_convert<__vector_type16_t<int>>(
+	      _SimdWrapper<float, 4>(__vector_bitcast<float>(
+		(__vector_bitcast<unsigned, 4>(__y) << 23) + 0x3f80'0000))));
+	else if constexpr (sizeof(_Up) =3D=3D 8 && sizeof __ix =3D=3D 16
+			   && !__have_avx2)
+	  {
+	    const auto __lo =3D _mm_sll_epi64(__ix, __iy);
+	    const auto __hi
+	      =3D _mm_sll_epi64(__ix, _mm_unpackhi_epi64(__iy, __iy));
+	    if constexpr (__have_sse4_1)
+	      return __vector_bitcast<_Up>(_mm_blend_epi16(__lo, __hi, 0xf0));
+	    else
+	      return __vector_bitcast<_Up>(
+		_mm_move_sd(__vector_bitcast<double>(__hi),
+			    __vector_bitcast<double>(__lo)));
+	  }
+	else
+	  return __x << __y;
+      }
+#endif // _GLIBCXX_SIMD_NO_SHIFT_OPT
+
+    // }}}
+    // _S_bit_shift_right {{{
+#ifndef _GLIBCXX_SIMD_NO_SHIFT_OPT
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      inline _GLIBCXX_CONST static typename _TVT::type
+      _S_bit_shift_right(_Tp __xx, int __y)
+      {
+	using _V =3D typename _TVT::type;
+	using _Up =3D typename _TVT::value_type;
+	_V __x =3D __xx;
+	[[maybe_unused]] const auto __ix =3D __to_intrin(__x);
+	if (__builtin_is_constant_evaluated())
+	  return __x >> __y;
+	else if (__builtin_constant_p(__y)
+		 && is_unsigned_v<
+		   _Up> && __y >=3D int(sizeof(_Up) * __CHAR_BIT__))
+	  return _V();
+	else if constexpr (sizeof(_Up) =3D=3D 1 && is_unsigned_v<_Up>) //{{{
+	  return __intrin_bitcast<_V>(__vector_bitcast<_UShort>(__ix) >> __y)
+		 & _Up(0xff >> __y);
+	//}}}
+	else if constexpr (sizeof(_Up) =3D=3D 1 && is_signed_v<_Up>) //{{{
+	  return __intrin_bitcast<_V>(
+	    (__vector_bitcast<_UShort>(__vector_bitcast<short>(__ix)
+				       >> (__y + 8))
+	     << 8)
+	    | (__vector_bitcast<_UShort>(
+		 __vector_bitcast<short>(__vector_bitcast<_UShort>(__ix) << 8)
+		 >> __y)
+	       >> 8));
+	//}}}
+	// GCC optimizes sizeof =3D=3D 2, 4, and unsigned 8 as expected
+	else if constexpr (sizeof(_Up) =3D=3D 8 && is_signed_v<_Up>) //{{{
+	  {
+	    if (__y > 32)
+	      return (__intrin_bitcast<_V>(__vector_bitcast<int>(__ix) >> 32)
+		      & _Up(0xffff'ffff'0000'0000ull))
+		     | __vector_bitcast<_Up>(
+		       __vector_bitcast<int>(__vector_bitcast<_ULLong>(__ix)
+					     >> 32)
+		       >> (__y - 32));
+	    else
+	      return __intrin_bitcast<_V>(__vector_bitcast<_ULLong>(__ix)
+					  >> __y)
+		     | __vector_bitcast<_Up>(
+		       __vector_bitcast<int>(__ix & -0x8000'0000'0000'0000ll)
+		       >> __y);
+	  }
+	//}}}
+	else
+	  return __x >> __y;
+      }
+
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      inline _GLIBCXX_CONST static typename _TVT::type
+      _S_bit_shift_right(_Tp __xx, typename _TVT::type __y)
+      {
+	using _V =3D typename _TVT::type;
+	using _Up =3D typename _TVT::value_type;
+	_V __x =3D __xx;
+	[[maybe_unused]] const auto __ix =3D __to_intrin(__x);
+	[[maybe_unused]] const auto __iy =3D __to_intrin(__y);
+	if (__builtin_is_constant_evaluated()
+	    || (__builtin_constant_p(__x) && __builtin_constant_p(__y)))
+	  return __x >> __y;
+	else if constexpr (sizeof(_Up) =3D=3D 1) //{{{
+	  {
+	    if constexpr (sizeof(__x) <=3D 8 && __have_avx512bw_vl)
+	      return __intrin_bitcast<_V>(_mm_cvtepi16_epi8(
+		is_signed_v<_Up> ? _mm_srav_epi16(_mm_cvtepi8_epi16(__ix),
+						  _mm_cvtepi8_epi16(__iy))
+				 : _mm_srlv_epi16(_mm_cvtepu8_epi16(__ix),
+						  _mm_cvtepu8_epi16(__iy))));
+	    if constexpr (sizeof(__x) =3D=3D 16 && __have_avx512bw_vl)
+	      return __intrin_bitcast<_V>(_mm256_cvtepi16_epi8(
+		is_signed_v<_Up>
+		  ? _mm256_srav_epi16(_mm256_cvtepi8_epi16(__ix),
+				      _mm256_cvtepi8_epi16(__iy))
+		  : _mm256_srlv_epi16(_mm256_cvtepu8_epi16(__ix),
+				      _mm256_cvtepu8_epi16(__iy))));
+	    else if constexpr (sizeof(__x) =3D=3D 32 && __have_avx512bw)
+	      return __vector_bitcast<_Up>(_mm512_cvtepi16_epi8(
+		is_signed_v<_Up>
+		  ? _mm512_srav_epi16(_mm512_cvtepi8_epi16(__ix),
+				      _mm512_cvtepi8_epi16(__iy))
+		  : _mm512_srlv_epi16(_mm512_cvtepu8_epi16(__ix),
+				      _mm512_cvtepu8_epi16(__iy))));
+	    else if constexpr (sizeof(__x) =3D=3D 64 && is_signed_v<_Up>)
+	      return __vector_bitcast<_Up>(_mm512_mask_mov_epi8(
+		_mm512_srav_epi16(__ix, _mm512_srli_epi16(__iy, 8)),
+		0x5555'5555'5555'5555ull,
+		_mm512_srav_epi16(
+		  _mm512_slli_epi16(__ix, 8),
+		  _mm512_maskz_add_epi8(0x5555'5555'5555'5555ull, __iy,
+					_mm512_set1_epi16(8)))));
+	    else if constexpr (sizeof(__x) =3D=3D 64 && is_unsigned_v<_Up>)
+	      return __vector_bitcast<_Up>(_mm512_mask_mov_epi8(
+		_mm512_srlv_epi16(__ix, _mm512_srli_epi16(__iy, 8)),
+		0x5555'5555'5555'5555ull,
+		_mm512_srlv_epi16(
+		  _mm512_maskz_mov_epi8(0x5555'5555'5555'5555ull, __ix),
+		  _mm512_maskz_mov_epi8(0x5555'5555'5555'5555ull, __iy))));
+	    /* This has better throughput but higher latency than the impl below
+	    else if constexpr (__have_avx2 && sizeof(__x) =3D=3D 16 &&
+			       is_unsigned_v<_Up>)
+	      {
+		const auto __shorts =3D __to_intrin(_S_bit_shift_right(
+		  __vector_bitcast<_UShort>(_mm256_cvtepu8_epi16(__ix)),
+		  __vector_bitcast<_UShort>(_mm256_cvtepu8_epi16(__iy))));
+		return __vector_bitcast<_Up>(
+		  _mm_packus_epi16(__lo128(__shorts), __hi128(__shorts)));
+	      }
+	      */
+	    else if constexpr (__have_avx2 && sizeof(__x) > 8)
+	      // the following uses vpsr[al]vd, which requires AVX2
+	      if constexpr (is_signed_v<_Up>)
+		{
+		  const auto r3 =3D __vector_bitcast<_UInt>(
+				    (__vector_bitcast<int>(__x)
+				     >> (__vector_bitcast<_UInt>(__y) >> 24)))
+				  & 0xff000000u;
+		  const auto r2
+		    =3D __vector_bitcast<_UInt>(
+			((__vector_bitcast<int>(__x) << 8)
+			 >> ((__vector_bitcast<_UInt>(__y) << 8) >> 24)))
+		      & 0xff000000u;
+		  const auto r1
+		    =3D __vector_bitcast<_UInt>(
+			((__vector_bitcast<int>(__x) << 16)
+			 >> ((__vector_bitcast<_UInt>(__y) << 16) >> 24)))
+		      & 0xff000000u;
+		  const auto r0 =3D __vector_bitcast<_UInt>(
+		    (__vector_bitcast<int>(__x) << 24)
+		    >> ((__vector_bitcast<_UInt>(__y) << 24) >> 24));
+		  return __vector_bitcast<_Up>(r3 | (r2 >> 8) | (r1 >> 16)
+					       | (r0 >> 24));
+		}
+	      else
+		{
+		  const auto r3 =3D (__vector_bitcast<_UInt>(__x)
+				   >> (__vector_bitcast<_UInt>(__y) >> 24))
+				  & 0xff000000u;
+		  const auto r2
+		    =3D ((__vector_bitcast<_UInt>(__x) << 8)
+		       >> ((__vector_bitcast<_UInt>(__y) << 8) >> 24))
+		      & 0xff000000u;
+		  const auto r1
+		    =3D ((__vector_bitcast<_UInt>(__x) << 16)
+		       >> ((__vector_bitcast<_UInt>(__y) << 16) >> 24))
+		      & 0xff000000u;
+		  const auto r0
+		    =3D (__vector_bitcast<_UInt>(__x) << 24)
+		      >> ((__vector_bitcast<_UInt>(__y) << 24) >> 24);
+		  return __vector_bitcast<_Up>(r3 | (r2 >> 8) | (r1 >> 16)
+					       | (r0 >> 24));
+		}
+	    else if constexpr (__have_sse4_1
+			       && is_unsigned_v<_Up> && sizeof(__x) > 2)
+	      {
+		auto __x128 =3D __vector_bitcast<_Up>(__ix);
+		auto __mask
+		  =3D __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__iy) << 5);
+		auto __x4 =3D __vector_bitcast<_Up>(
+		  (__vector_bitcast<_UShort>(__x128) >> 4) & _UShort(0xff0f));
+		__x128 =3D __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x4)));
+		__mask +=3D __mask;
+		auto __x2 =3D __vector_bitcast<_Up>(
+		  (__vector_bitcast<_UShort>(__x128) >> 2) & _UShort(0xff3f));
+		__x128 =3D __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x2)));
+		__mask +=3D __mask;
+		auto __x1 =3D __vector_bitcast<_Up>(
+		  (__vector_bitcast<_UShort>(__x128) >> 1) & _UShort(0xff7f));
+		__x128 =3D __vector_bitcast<_Up>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__x128), __to_intrin(__x1)));
+		return __intrin_bitcast<_V>(
+		  __x128
+		  & ((__vector_bitcast<_Up>(__iy) & char(0xf8))
+		     =3D=3D 0)); // y > 7 nulls the result
+	      }
+	    else if constexpr (__have_sse4_1
+			       && is_signed_v<_Up> && sizeof(__x) > 2)
+	      {
+		auto __mask =3D __vector_bitcast<_UChar>(
+		  __vector_bitcast<_UShort>(__iy) << 5);
+		auto __maskl =3D [&]() {
+		  return __to_intrin(__vector_bitcast<_UShort>(__mask) << 8);
+		};
+		auto __xh =3D __vector_bitcast<short>(__ix);
+		auto __xl =3D __vector_bitcast<short>(__ix) << 8;
+		auto __xh4 =3D __xh >> 4;
+		auto __xl4 =3D __xl >> 4;
+		__xh =3D __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh4)));
+		__xl =3D __vector_bitcast<short>(
+		  _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
+						  __to_intrin(__xl4)));
+		__mask +=3D __mask;
+		auto __xh2 =3D __xh >> 2;
+		auto __xl2 =3D __xl >> 2;
+		__xh =3D __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh2)));
+		__xl =3D __vector_bitcast<short>(
+		  _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
+						  __to_intrin(__xl2)));
+		__mask +=3D __mask;
+		auto __xh1 =3D __xh >> 1;
+		auto __xl1 =3D __xl >> 1;
+		__xh =3D __vector_bitcast<short>(_CommonImplX86::_S_blend_intrin(
+		  __to_intrin(__mask), __to_intrin(__xh), __to_intrin(__xh1)));
+		__xl =3D __vector_bitcast<short>(
+		  _CommonImplX86::_S_blend_intrin(__maskl(), __to_intrin(__xl),
+						  __to_intrin(__xl1)));
+		return __intrin_bitcast<_V>(
+		  (__vector_bitcast<_Up>((__xh & short(0xff00)))
+		   | __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__xl)
+					   >> 8))
+		  & ((__vector_bitcast<_Up>(__iy) & char(0xf8))
+		     =3D=3D 0)); // y > 7 nulls the result
+	      }
+	    else if constexpr (is_unsigned_v<_Up> && sizeof(__x) > 2) // SSE2
+	      {
+		auto __mask
+		  =3D __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__y) << 5);
+		auto __x4 =3D __vector_bitcast<_Up>(
+		  (__vector_bitcast<_UShort>(__x) >> 4) & _UShort(0xff0f));
+		__x =3D __mask > 0x7f ? __x4 : __x;
+		__mask +=3D __mask;
+		auto __x2 =3D __vector_bitcast<_Up>(
+		  (__vector_bitcast<_UShort>(__x) >> 2) & _UShort(0xff3f));
+		__x =3D __mask > 0x7f ? __x2 : __x;
+		__mask +=3D __mask;
+		auto __x1 =3D __vector_bitcast<_Up>(
+		  (__vector_bitcast<_UShort>(__x) >> 1) & _UShort(0xff7f));
+		__x =3D __mask > 0x7f ? __x1 : __x;
+		return __x
+		       & ((__y & char(0xf8)) =3D=3D 0); // y > 7 nulls the result
+	      }
+	    else if constexpr (sizeof(__x) > 2) // signed SSE2
+	      {
+		static_assert(is_signed_v<_Up>);
+		auto __maskh =3D __vector_bitcast<_UShort>(__y) << 5;
+		auto __maskl =3D __vector_bitcast<_UShort>(__y) << (5 + 8);
+		auto __xh =3D __vector_bitcast<short>(__x);
+		auto __xl =3D __vector_bitcast<short>(__x) << 8;
+		auto __xh4 =3D __xh >> 4;
+		auto __xl4 =3D __xl >> 4;
+		__xh =3D __maskh > 0x7fff ? __xh4 : __xh;
+		__xl =3D __maskl > 0x7fff ? __xl4 : __xl;
+		__maskh +=3D __maskh;
+		__maskl +=3D __maskl;
+		auto __xh2 =3D __xh >> 2;
+		auto __xl2 =3D __xl >> 2;
+		__xh =3D __maskh > 0x7fff ? __xh2 : __xh;
+		__xl =3D __maskl > 0x7fff ? __xl2 : __xl;
+		__maskh +=3D __maskh;
+		__maskl +=3D __maskl;
+		auto __xh1 =3D __xh >> 1;
+		auto __xl1 =3D __xl >> 1;
+		__xh =3D __maskh > 0x7fff ? __xh1 : __xh;
+		__xl =3D __maskl > 0x7fff ? __xl1 : __xl;
+		__x =3D __vector_bitcast<_Up>((__xh & short(0xff00)))
+		      | __vector_bitcast<_Up>(__vector_bitcast<_UShort>(__xl)
+					      >> 8);
+		return __x
+		       & ((__y & char(0xf8)) =3D=3D 0); // y > 7 nulls the result
+	      }
+	    else
+	      return __x >> __y;
+	  }                                                      //}}}
+	else if constexpr (sizeof(_Up) =3D=3D 2 && sizeof(__x) >=3D 4) //{{{
+	  {
+	    [[maybe_unused]] auto __blend_0xaa =3D [](auto __a, auto __b) {
+	      if constexpr (sizeof(__a) =3D=3D 16)
+		return _mm_blend_epi16(__to_intrin(__a), __to_intrin(__b),
+				       0xaa);
+	      else if constexpr (sizeof(__a) =3D=3D 32)
+		return _mm256_blend_epi16(__to_intrin(__a), __to_intrin(__b),
+					  0xaa);
+	      else if constexpr (sizeof(__a) =3D=3D 64)
+		return _mm512_mask_blend_epi16(0xaaaa'aaaaU, __to_intrin(__a),
+					       __to_intrin(__b));
+	      else
+		__assert_unreachable<decltype(__a)>();
+	    };
+	    if constexpr (__have_avx512bw_vl && sizeof(_Tp) <=3D 16)
+	      return __intrin_bitcast<_V>(is_signed_v<_Up>
+					    ? _mm_srav_epi16(__ix, __iy)
+					    : _mm_srlv_epi16(__ix, __iy));
+	    else if constexpr (__have_avx512bw_vl && sizeof(_Tp) =3D=3D 32)
+	      return __vector_bitcast<_Up>(is_signed_v<_Up>
+					     ? _mm256_srav_epi16(__ix, __iy)
+					     : _mm256_srlv_epi16(__ix, __iy));
+	    else if constexpr (__have_avx512bw && sizeof(_Tp) =3D=3D 64)
+	      return __vector_bitcast<_Up>(is_signed_v<_Up>
+					     ? _mm512_srav_epi16(__ix, __iy)
+					     : _mm512_srlv_epi16(__ix, __iy));
+	    else if constexpr (__have_avx2 && is_signed_v<_Up>)
+	      return __intrin_bitcast<_V>(
+		__blend_0xaa(((__vector_bitcast<int>(__ix) << 16)
+			      >> (__vector_bitcast<int>(__iy) & 0xffffu))
+			       >> 16,
+			     __vector_bitcast<int>(__ix)
+			       >> (__vector_bitcast<int>(__iy) >> 16)));
+	    else if constexpr (__have_avx2 && is_unsigned_v<_Up>)
+	      return __intrin_bitcast<_V>(
+		__blend_0xaa((__vector_bitcast<_UInt>(__ix) & 0xffffu)
+			       >> (__vector_bitcast<_UInt>(__iy) & 0xffffu),
+			     __vector_bitcast<_UInt>(__ix)
+			       >> (__vector_bitcast<_UInt>(__iy) >> 16)));
+	    else if constexpr (__have_sse4_1)
+	      {
+		auto __mask =3D __vector_bitcast<_UShort>(__iy);
+		auto __x128 =3D __vector_bitcast<_Up>(__ix);
+		//__mask *=3D 0x0808;
+		__mask =3D (__mask << 3) | (__mask << 11);
+		// do __x128 =3D 0 where __y[4] is set
+		__x128 =3D __vector_bitcast<_Up>(
+		  _mm_blendv_epi8(__to_intrin(__x128), __m128i(),
+				  __to_intrin(__mask)));
+		// do __x128 =3D>> 8 where __y[3] is set
+		__x128 =3D __vector_bitcast<_Up>(
+		  _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 8),
+				  __to_intrin(__mask +=3D __mask)));
+		// do __x128 =3D>> 4 where __y[2] is set
+		__x128 =3D __vector_bitcast<_Up>(
+		  _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 4),
+				  __to_intrin(__mask +=3D __mask)));
+		// do __x128 =3D>> 2 where __y[1] is set
+		__x128 =3D __vector_bitcast<_Up>(
+		  _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 2),
+				  __to_intrin(__mask +=3D __mask)));
+		// do __x128 =3D>> 1 where __y[0] is set
+		return __intrin_bitcast<_V>(
+		  _mm_blendv_epi8(__to_intrin(__x128), __to_intrin(__x128 >> 1),
+				  __to_intrin(__mask + __mask)));
+	      }
+	    else
+	      {
+		auto __k =3D __vector_bitcast<_UShort>(__iy) << 11;
+		auto __x128 =3D __vector_bitcast<_Up>(__ix);
+		auto __mask =3D [](__vector_type16_t<_UShort> __kk) {
+		  return __vector_bitcast<short>(__kk) < 0;
+		};
+		// do __x128 =3D 0 where __y[4] is set
+		__x128 =3D __mask(__k) ? decltype(__x128)() : __x128;
+		// do __x128 =3D>> 8 where __y[3] is set
+		__x128 =3D __mask(__k +=3D __k) ? __x128 >> 8 : __x128;
+		// do __x128 =3D>> 4 where __y[2] is set
+		__x128 =3D __mask(__k +=3D __k) ? __x128 >> 4 : __x128;
+		// do __x128 =3D>> 2 where __y[1] is set
+		__x128 =3D __mask(__k +=3D __k) ? __x128 >> 2 : __x128;
+		// do __x128 =3D>> 1 where __y[0] is set
+		return __intrin_bitcast<_V>(__mask(__k + __k) ? __x128 >> 1
+							      : __x128);
+	      }
+	  }                                                  //}}}
+	else if constexpr (sizeof(_Up) =3D=3D 4 && !__have_avx2) //{{{
+	  {
+	    if constexpr (is_unsigned_v<_Up>)
+	      {
+		// x >> y =3D=3D x * 2^-y =3D=3D (x * 2^(31-y)) >> 31
+		const __m128 __factor_f =3D reinterpret_cast<__m128>(
+		  0x4f00'0000u - (__vector_bitcast<unsigned, 4>(__y) << 23));
+		const __m128i __factor
+		  =3D __builtin_constant_p(__factor_f)
+		      ? __to_intrin(
+			__make_vector<unsigned>(__factor_f[0], __factor_f[1],
+						__factor_f[2], __factor_f[3]))
+		      : _mm_cvttps_epi32(__factor_f);
+		const auto __r02
+		  =3D _mm_srli_epi64(_mm_mul_epu32(__ix, __factor), 31);
+		const auto __r13 =3D _mm_mul_epu32(_mm_srli_si128(__ix, 4),
+						 _mm_srli_si128(__factor, 4));
+		if constexpr (__have_sse4_1)
+		  return __intrin_bitcast<_V>(
+		    _mm_blend_epi16(_mm_slli_epi64(__r13, 1), __r02, 0x33));
+		else
+		  return __intrin_bitcast<_V>(
+		    __r02 | _mm_slli_si128(_mm_srli_epi64(__r13, 31), 4));
+	      }
+	    else
+	      {
+		auto __shift =3D [](auto __a, auto __b) {
+		  if constexpr (is_signed_v<_Up>)
+		    return _mm_sra_epi32(__a, __b);
+		  else
+		    return _mm_srl_epi32(__a, __b);
+		};
+		const auto __r0
+		  =3D __shift(__ix, _mm_unpacklo_epi32(__iy, __m128i()));
+		const auto __r1 =3D __shift(__ix, _mm_srli_epi64(__iy, 32));
+		const auto __r2
+		  =3D __shift(__ix, _mm_unpackhi_epi32(__iy, __m128i()));
+		const auto __r3 =3D __shift(__ix, _mm_srli_si128(__iy, 12));
+		if constexpr (__have_sse4_1)
+		  return __intrin_bitcast<_V>(
+		    _mm_blend_epi16(_mm_blend_epi16(__r1, __r0, 0x3),
+				    _mm_blend_epi16(__r3, __r2, 0x30), 0xf0));
+		else
+		  return __intrin_bitcast<_V>(_mm_unpacklo_epi64(
+		    _mm_unpacklo_epi32(__r0, _mm_srli_si128(__r1, 4)),
+		    _mm_unpackhi_epi32(__r2, _mm_srli_si128(__r3, 4))));
+	      }
+	  } //}}}
+	else
+	  return __x >> __y;
+      }
+#endif // _GLIBCXX_SIMD_NO_SHIFT_OPT
+
+    // }}}
+    // compares {{{
+    // _S_equal_to {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      {
+	if constexpr (__is_avx512_abi<_Abi>()) // {{{
+	  {
+	    if (__builtin_is_constant_evaluated()
+		|| (__x._M_is_constprop() && __y._M_is_constprop()))
+	      return _MaskImpl::_S_to_bits(
+		__as_wrapper<_Np>(__x._M_data =3D=3D __y._M_data));
+
+	    constexpr auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    [[maybe_unused]] const auto __xi =3D __to_intrin(__x);
+	    [[maybe_unused]] const auto __yi =3D __to_intrin(__y);
+	    if constexpr (is_floating_point_v<_Tp>)
+	      {
+		if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+		  return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
+		else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+		  return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
+		else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+		  return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
+		else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+		  return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
+		else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+		  return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
+		else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+		  return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_EQ_OQ);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 2)
+	      return _mm512_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 1)
+	      return _mm512_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 2)
+	      return _mm256_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 1)
+	      return _mm256_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 2)
+	      return _mm_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 1)
+	      return _mm_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
+	    else
+	      __assert_unreachable<_Tp>();
+	  } // }}}
+	else if (__builtin_is_constant_evaluated())
+	  return _Base::_S_equal_to(__x, __y);
+	else if constexpr (sizeof(__x) =3D=3D 8) // {{{
+	  {
+	    const auto __r128 =3D __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
+				=3D=3D __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
+	    _MaskMember<_Tp> __r64;
+	    __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
+	    return __r64;
+	  } // }}}
+	else
+	  return _Base::_S_equal_to(__x, __y);
+      }
+
+    // }}}
+    // _S_not_equal_to {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_not_equal_to(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> _=
_y)
+      {
+	if constexpr (__is_avx512_abi<_Abi>()) // {{{
+	  {
+	    if (__builtin_is_constant_evaluated()
+		|| (__x._M_is_constprop() && __y._M_is_constprop()))
+	      return _MaskImpl::_S_to_bits(
+		__as_wrapper<_Np>(__x._M_data !=3D __y._M_data));
+
+	    constexpr auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    [[maybe_unused]] const auto __xi =3D __to_intrin(__x);
+	    [[maybe_unused]] const auto __yi =3D __to_intrin(__y);
+	    if constexpr (is_floating_point_v<_Tp>)
+	      {
+		if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+		  return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
+		else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+		  return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
+		else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+		  return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
+		else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+		  return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
+		else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+		  return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
+		else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+		  return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_UQ);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return ~_mm512_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return ~_mm512_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 2)
+	      return ~_mm512_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 1)
+	      return ~_mm512_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return ~_mm256_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return ~_mm256_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 2)
+	      return ~_mm256_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 1)
+	      return ~_mm256_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return ~_mm_mask_cmpeq_epi64_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return ~_mm_mask_cmpeq_epi32_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 2)
+	      return ~_mm_mask_cmpeq_epi16_mask(__k1, __xi, __yi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 1)
+	      return ~_mm_mask_cmpeq_epi8_mask(__k1, __xi, __yi);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }                                                   // }}}
+	else if constexpr (!__builtin_is_constant_evaluated() // {{{
+			   && sizeof(__x) =3D=3D 8)
+	  {
+	    const auto __r128 =3D __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
+				!=3D __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
+	    _MaskMember<_Tp> __r64;
+	    __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
+	    return __r64;
+	  } // }}}
+	else
+	  return _Base::_S_not_equal_to(__x, __y);
+      }
+
+    // }}}
+    // _S_less {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_less(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      {
+	if constexpr (__is_avx512_abi<_Abi>()) // {{{
+	  {
+	    if (__builtin_is_constant_evaluated()
+		|| (__x._M_is_constprop() && __y._M_is_constprop()))
+	      return _MaskImpl::_S_to_bits(
+		__as_wrapper<_Np>(__x._M_data < __y._M_data));
+
+	    constexpr auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    [[maybe_unused]] const auto __xi =3D __to_intrin(__x);
+	    [[maybe_unused]] const auto __yi =3D __to_intrin(__y);
+	    if constexpr (sizeof(__xi) =3D=3D 64)
+	      {
+		if constexpr (is_same_v<_Tp, float>)
+		  return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
+		else if constexpr (is_same_v<_Tp, double>)
+		  return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm512_mask_cmplt_epi8_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm512_mask_cmplt_epi16_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm512_mask_cmplt_epi32_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm512_mask_cmplt_epi64_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm512_mask_cmplt_epu8_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm512_mask_cmplt_epu16_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm512_mask_cmplt_epu32_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm512_mask_cmplt_epu64_mask(__k1, __xi, __yi);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (sizeof(__xi) =3D=3D 32)
+	      {
+		if constexpr (is_same_v<_Tp, float>)
+		  return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
+		else if constexpr (is_same_v<_Tp, double>)
+		  return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm256_mask_cmplt_epi8_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm256_mask_cmplt_epi16_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm256_mask_cmplt_epi32_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm256_mask_cmplt_epi64_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm256_mask_cmplt_epu8_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm256_mask_cmplt_epu16_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm256_mask_cmplt_epu32_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm256_mask_cmplt_epu64_mask(__k1, __xi, __yi);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (sizeof(__xi) =3D=3D 16)
+	      {
+		if constexpr (is_same_v<_Tp, float>)
+		  return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OS);
+		else if constexpr (is_same_v<_Tp, double>)
+		  return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OS);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm_mask_cmplt_epi8_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm_mask_cmplt_epi16_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm_mask_cmplt_epi32_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm_mask_cmplt_epi64_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm_mask_cmplt_epu8_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm_mask_cmplt_epu16_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm_mask_cmplt_epu32_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm_mask_cmplt_epu64_mask(__k1, __xi, __yi);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else
+	      __assert_unreachable<_Tp>();
+	  }                                                   // }}}
+	else if constexpr (!__builtin_is_constant_evaluated() // {{{
+			   && sizeof(__x) =3D=3D 8)
+	  {
+	    const auto __r128 =3D __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
+				< __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
+	    _MaskMember<_Tp> __r64;
+	    __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
+	    return __r64;
+	  } // }}}
+	else
+	  return _Base::_S_less(__x, __y);
+      }
+
+    // }}}
+    // _S_less_equal {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_less_equal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __y)
+      {
+	if constexpr (__is_avx512_abi<_Abi>()) // {{{
+	  {
+	    if (__builtin_is_constant_evaluated()
+		|| (__x._M_is_constprop() && __y._M_is_constprop()))
+	      return _MaskImpl::_S_to_bits(
+		__as_wrapper<_Np>(__x._M_data <=3D __y._M_data));
+
+	    constexpr auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    [[maybe_unused]] const auto __xi =3D __to_intrin(__x);
+	    [[maybe_unused]] const auto __yi =3D __to_intrin(__y);
+	    if constexpr (sizeof(__xi) =3D=3D 64)
+	      {
+		if constexpr (is_same_v<_Tp, float>)
+		  return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
+		else if constexpr (is_same_v<_Tp, double>)
+		  return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm512_mask_cmple_epi8_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm512_mask_cmple_epi16_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm512_mask_cmple_epi32_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm512_mask_cmple_epi64_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm512_mask_cmple_epu8_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm512_mask_cmple_epu16_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm512_mask_cmple_epu32_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm512_mask_cmple_epu64_mask(__k1, __xi, __yi);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (sizeof(__xi) =3D=3D 32)
+	      {
+		if constexpr (is_same_v<_Tp, float>)
+		  return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
+		else if constexpr (is_same_v<_Tp, double>)
+		  return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm256_mask_cmple_epi8_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm256_mask_cmple_epi16_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm256_mask_cmple_epi32_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm256_mask_cmple_epi64_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm256_mask_cmple_epu8_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm256_mask_cmple_epu16_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm256_mask_cmple_epu32_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm256_mask_cmple_epu64_mask(__k1, __xi, __yi);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else if constexpr (sizeof(__xi) =3D=3D 16)
+	      {
+		if constexpr (is_same_v<_Tp, float>)
+		  return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OS);
+		else if constexpr (is_same_v<_Tp, double>)
+		  return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OS);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm_mask_cmple_epi8_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm_mask_cmple_epi16_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm_mask_cmple_epi32_mask(__k1, __xi, __yi);
+		else if constexpr (is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm_mask_cmple_epi64_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1)
+		  return _mm_mask_cmple_epu8_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2)
+		  return _mm_mask_cmple_epu16_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4)
+		  return _mm_mask_cmple_epu32_mask(__k1, __xi, __yi);
+		else if constexpr (is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8)
+		  return _mm_mask_cmple_epu64_mask(__k1, __xi, __yi);
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else
+	      __assert_unreachable<_Tp>();
+	  }                                                   // }}}
+	else if constexpr (!__builtin_is_constant_evaluated() // {{{
+			   && sizeof(__x) =3D=3D 8)
+	  {
+	    const auto __r128 =3D __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__x)
+				<=3D __vector_bitcast<_Tp, 16 / sizeof(_Tp)>(__y);
+	    _MaskMember<_Tp> __r64;
+	    __builtin_memcpy(&__r64._M_data, &__r128, sizeof(__r64));
+	    return __r64;
+	  } // }}}
+	else
+	  return _Base::_S_less_equal(__x, __y);
+      }
+
+    // }}} }}}
+    // negation {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_negate(_SimdWrapper<_Tp, _Np> __x) noexcept
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  return _S_equal_to(__x, _SimdWrapper<_Tp, _Np>());
+	else
+	  return _Base::_S_negate(__x);
+      }
+
+    // }}}
+    // math {{{
+    using _Base::_S_abs;
+
+    // _S_sqrt {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_sqrt(_SimdWrapper<_Tp, _Np> __x)
+      {
+	if constexpr (__is_sse_ps<_Tp, _Np>())
+	  return __auto_bitcast(_mm_sqrt_ps(__to_intrin(__x)));
+	else if constexpr (__is_sse_pd<_Tp, _Np>())
+	  return _mm_sqrt_pd(__x);
+	else if constexpr (__is_avx_ps<_Tp, _Np>())
+	  return _mm256_sqrt_ps(__x);
+	else if constexpr (__is_avx_pd<_Tp, _Np>())
+	  return _mm256_sqrt_pd(__x);
+	else if constexpr (__is_avx512_ps<_Tp, _Np>())
+	  return _mm512_sqrt_ps(__x);
+	else if constexpr (__is_avx512_pd<_Tp, _Np>())
+	  return _mm512_sqrt_pd(__x);
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    // }}}
+    // _S_ldexp {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_ldexp(_SimdWrapper<_Tp, _Np> __x,
+	       __fixed_size_storage_t<int, _Np> __exp)
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    const auto __xi =3D __to_intrin(__x);
+	    constexpr _SimdConverter<int, simd_abi::fixed_size<_Np>, _Tp, _Abi>
+	      __cvt;
+	    const auto __expi =3D __to_intrin(__cvt(__exp));
+	    constexpr auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 16)
+	      {
+		if constexpr (sizeof(_Tp) =3D=3D 8)
+		  return _mm_maskz_scalef_pd(__k1, __xi, __expi);
+		else
+		  return _mm_maskz_scalef_ps(__k1, __xi, __expi);
+	      }
+	    else if constexpr (sizeof(__xi) =3D=3D 32)
+	      {
+		if constexpr (sizeof(_Tp) =3D=3D 8)
+		  return _mm256_maskz_scalef_pd(__k1, __xi, __expi);
+		else
+		  return _mm256_maskz_scalef_ps(__k1, __xi, __expi);
+	      }
+	    else
+	      {
+		static_assert(sizeof(__xi) =3D=3D 64);
+		if constexpr (sizeof(_Tp) =3D=3D 8)
+		  return _mm512_maskz_scalef_pd(__k1, __xi, __expi);
+		else
+		  return _mm512_maskz_scalef_ps(__k1, __xi, __expi);
+	      }
+	  }
+	else
+	  return _Base::_S_ldexp(__x, __exp);
+      }
+
+    // }}}
+    // _S_trunc {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_trunc(_SimdWrapper<_Tp, _Np> __x)
+      {
+	if constexpr (__is_avx512_ps<_Tp, _Np>())
+	  return _mm512_roundscale_ps(__x, 0x0b);
+	else if constexpr (__is_avx512_pd<_Tp, _Np>())
+	  return _mm512_roundscale_pd(__x, 0x0b);
+	else if constexpr (__is_avx_ps<_Tp, _Np>())
+	  return _mm256_round_ps(__x, 0x3);
+	else if constexpr (__is_avx_pd<_Tp, _Np>())
+	  return _mm256_round_pd(__x, 0x3);
+	else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
+	  return __auto_bitcast(_mm_round_ps(__to_intrin(__x), 0x3));
+	else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
+	  return _mm_round_pd(__x, 0x3);
+	else if constexpr (__is_sse_ps<_Tp, _Np>())
+	  {
+	    auto __truncated
+	      =3D _mm_cvtepi32_ps(_mm_cvttps_epi32(__to_intrin(__x)));
+	    const auto __no_fractional_values
+	      =3D __vector_bitcast<int>(__vector_bitcast<_UInt>(__to_intrin(__x))
+				      & 0x7f800000u)
+		< 0x4b000000; // the exponent is so large that no mantissa bits
+			      // signify fractional values (0x3f8 + 23*8 =3D
+			      // 0x4b0)
+	    return __no_fractional_values ? __truncated : __to_intrin(__x);
+	  }
+	else
+	  return _Base::_S_trunc(__x);
+      }
+
+    // }}}
+    // _S_round {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_round(_SimdWrapper<_Tp, _Np> __x)
+      {
+	// Note that _MM_FROUND_TO_NEAREST_INT rounds ties to even, not away
+	// from zero as required by std::round. Therefore this function is more
+	// complicated.
+	using _V =3D __vector_type_t<_Tp, _Np>;
+	_V __truncated;
+	if constexpr (__is_avx512_ps<_Tp, _Np>())
+	  __truncated =3D _mm512_roundscale_ps(__x._M_data, 0x0b);
+	else if constexpr (__is_avx512_pd<_Tp, _Np>())
+	  __truncated =3D _mm512_roundscale_pd(__x._M_data, 0x0b);
+	else if constexpr (__is_avx_ps<_Tp, _Np>())
+	  __truncated =3D _mm256_round_ps(__x._M_data,
+					_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
+	else if constexpr (__is_avx_pd<_Tp, _Np>())
+	  __truncated =3D _mm256_round_pd(__x._M_data,
+					_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
+	else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
+	  __truncated =3D __auto_bitcast(
+	    _mm_round_ps(__to_intrin(__x),
+			 _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC));
+	else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
+	  __truncated
+	    =3D _mm_round_pd(__x._M_data, _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
+	else if constexpr (__is_sse_ps<_Tp, _Np>())
+	  __truncated =3D __auto_bitcast(
+	    _mm_cvtepi32_ps(_mm_cvttps_epi32(__to_intrin(__x))));
+	else
+	  return _Base::_S_round(__x);
+
+	// x < 0 =3D> truncated <=3D 0 && truncated >=3D x =3D> x - truncated <=
=3D 0
+	// x > 0 =3D> truncated >=3D 0 && truncated <=3D x =3D> x - truncated >=
=3D 0
+
+	const _V __rounded
+	  =3D __truncated
+	    + (__and(_S_absmask<_V>, __x._M_data - __truncated) >=3D _Tp(.5)
+		 ? __or(__and(_S_signmask<_V>, __x._M_data), _V() + 1)
+		 : _V());
+	if constexpr (__have_sse4_1)
+	  return __rounded;
+	else // adjust for missing range in cvttps_epi32
+	  return __and(_S_absmask<_V>, __x._M_data) < 0x1p23f ? __rounded
+							      : __x._M_data;
+      }
+
+    // }}}
+    // _S_nearbyint {{{
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      _GLIBCXX_SIMD_INTRINSIC static _Tp _S_nearbyint(_Tp __x) noexcept
+      {
+	if constexpr (_TVT::template _S_is<float, 16>)
+	  return _mm512_roundscale_ps(__x, 0x0c);
+	else if constexpr (_TVT::template _S_is<double, 8>)
+	  return _mm512_roundscale_pd(__x, 0x0c);
+	else if constexpr (_TVT::template _S_is<float, 8>)
+	  return _mm256_round_ps(__x,
+				 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
+	else if constexpr (_TVT::template _S_is<double, 4>)
+	  return _mm256_round_pd(__x,
+				 _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
+	else if constexpr (__have_sse4_1 && _TVT::template _S_is<float, 4>)
+	  return _mm_round_ps(__x,
+			      _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
+	else if constexpr (__have_sse4_1 && _TVT::template _S_is<double, 2>)
+	  return _mm_round_pd(__x,
+			      _MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC);
+	else
+	  return _Base::_S_nearbyint(__x);
+      }
+
+    // }}}
+    // _S_rint {{{
+    template <typename _Tp, typename _TVT =3D _VectorTraits<_Tp>>
+      _GLIBCXX_SIMD_INTRINSIC static _Tp _S_rint(_Tp __x) noexcept
+      {
+	if constexpr (_TVT::template _S_is<float, 16>)
+	  return _mm512_roundscale_ps(__x, 0x04);
+	else if constexpr (_TVT::template _S_is<double, 8>)
+	  return _mm512_roundscale_pd(__x, 0x04);
+	else if constexpr (_TVT::template _S_is<float, 8>)
+	  return _mm256_round_ps(__x, _MM_FROUND_CUR_DIRECTION);
+	else if constexpr (_TVT::template _S_is<double, 4>)
+	  return _mm256_round_pd(__x, _MM_FROUND_CUR_DIRECTION);
+	else if constexpr (__have_sse4_1 && _TVT::template _S_is<float, 4>)
+	  return _mm_round_ps(__x, _MM_FROUND_CUR_DIRECTION);
+	else if constexpr (__have_sse4_1 && _TVT::template _S_is<double, 2>)
+	  return _mm_round_pd(__x, _MM_FROUND_CUR_DIRECTION);
+	else
+	  return _Base::_S_rint(__x);
+      }
+
+    // }}}
+    // _S_floor {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_floor(_SimdWrapper<_Tp, _Np> __x)
+      {
+	if constexpr (__is_avx512_ps<_Tp, _Np>())
+	  return _mm512_roundscale_ps(__x, 0x09);
+	else if constexpr (__is_avx512_pd<_Tp, _Np>())
+	  return _mm512_roundscale_pd(__x, 0x09);
+	else if constexpr (__is_avx_ps<_Tp, _Np>())
+	  return _mm256_round_ps(__x, 0x1);
+	else if constexpr (__is_avx_pd<_Tp, _Np>())
+	  return _mm256_round_pd(__x, 0x1);
+	else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
+	  return __auto_bitcast(_mm_floor_ps(__to_intrin(__x)));
+	else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
+	  return _mm_floor_pd(__x);
+	else
+	  return _Base::_S_floor(__x);
+      }
+
+    // }}}
+    // _S_ceil {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_ceil(_SimdWrapper<_Tp, _Np> __x)
+      {
+	if constexpr (__is_avx512_ps<_Tp, _Np>())
+	  return _mm512_roundscale_ps(__x, 0x0a);
+	else if constexpr (__is_avx512_pd<_Tp, _Np>())
+	  return _mm512_roundscale_pd(__x, 0x0a);
+	else if constexpr (__is_avx_ps<_Tp, _Np>())
+	  return _mm256_round_ps(__x, 0x2);
+	else if constexpr (__is_avx_pd<_Tp, _Np>())
+	  return _mm256_round_pd(__x, 0x2);
+	else if constexpr (__have_sse4_1 && __is_sse_ps<_Tp, _Np>())
+	  return __auto_bitcast(_mm_ceil_ps(__to_intrin(__x)));
+	else if constexpr (__have_sse4_1 && __is_sse_pd<_Tp, _Np>())
+	  return _mm_ceil_pd(__x);
+	else
+	  return _Base::_S_ceil(__x);
+      }
+
+    // }}}
+    // _S_signbit {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_signbit(_SimdWrapper<_Tp, _Np> __x)
+      {
+	if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
+	  {
+	    if constexpr (sizeof(__x) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_movepi32_mask(
+		__intrin_bitcast<__m512i>(__x._M_data));
+	    else if constexpr (sizeof(__x) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_movepi64_mask(
+		__intrin_bitcast<__m512i>(__x._M_data));
+	    else if constexpr (sizeof(__x) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_movepi32_mask(
+		__intrin_bitcast<__m256i>(__x._M_data));
+	    else if constexpr (sizeof(__x) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_movepi64_mask(
+		__intrin_bitcast<__m256i>(__x._M_data));
+	    else if constexpr (sizeof(__x) <=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_movepi32_mask(__intrin_bitcast<__m128i>(__x._M_data));
+	    else if constexpr (sizeof(__x) <=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_movepi64_mask(__intrin_bitcast<__m128i>(__x._M_data));
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    const auto __xi =3D __to_intrin(__x);
+	    [[maybe_unused]] constexpr auto __k1
+	      =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_movemask_ps(__xi);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_movemask_pd(__xi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_movemask_ps(__xi);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_movemask_pd(__xi);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmplt_epi32_mask(
+		__k1, __intrin_bitcast<__m512i>(__xi), __m512i());
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmplt_epi64_mask(
+		__k1, __intrin_bitcast<__m512i>(__xi), __m512i());
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_signbit(__x);
+	/*{
+	  using _I =3D __int_for_sizeof_t<_Tp>;
+	  if constexpr (sizeof(__x) =3D=3D 64)
+	    return _S_less(__vector_bitcast<_I>(__x), _I());
+	  else
+	    {
+	      const auto __xx =3D __vector_bitcast<_I>(__x._M_data);
+	      [[maybe_unused]] constexpr _I __signmask =3D __finite_min_v<_I>;
+	      if constexpr ((sizeof(_Tp) =3D=3D 4 &&
+			     (__have_avx2 || sizeof(__x) =3D=3D 16)) ||
+			    __have_avx512vl)
+		{
+		  return __vector_bitcast<_Tp>(__xx >> __digits_v<_I>);
+		}
+	      else if constexpr ((__have_avx2 ||
+				  (__have_ssse3 && sizeof(__x) =3D=3D 16)))
+		{
+		  return __vector_bitcast<_Tp>((__xx & __signmask) =3D=3D
+					       __signmask);
+		}
+	      else
+		{ // SSE2/3 or AVX (w/o AVX2)
+		  constexpr auto __one =3D __vector_broadcast<_Np, _Tp>(1);
+		  return __vector_bitcast<_Tp>(
+		    __vector_bitcast<_Tp>(
+		      (__xx & __signmask) |
+		      __vector_bitcast<_I>(__one)) // -1 or 1
+		    !=3D __one);
+		}
+	    }
+	}*/
+      }
+
+    // }}}
+    // _S_isnonzerovalue_mask {{{
+    // (isnormal | is subnormal =3D=3D !isinf & !isnan & !is zero)
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static auto _S_isnonzerovalue_mask(_Tp __x)
+      {
+	using _Traits =3D _VectorTraits<_Tp>;
+	if constexpr (__have_avx512dq_vl)
+	  {
+	    if constexpr (_Traits::template _S_is<
+			    float, 2> || _Traits::template _S_is<float, 4>)
+	      return _knot_mask8(_mm_fpclass_ps_mask(__to_intrin(__x), 0x9f));
+	    else if constexpr (_Traits::template _S_is<float, 8>)
+	      return _knot_mask8(_mm256_fpclass_ps_mask(__x, 0x9f));
+	    else if constexpr (_Traits::template _S_is<float, 16>)
+	      return _knot_mask16(_mm512_fpclass_ps_mask(__x, 0x9f));
+	    else if constexpr (_Traits::template _S_is<double, 2>)
+	      return _knot_mask8(_mm_fpclass_pd_mask(__x, 0x9f));
+	    else if constexpr (_Traits::template _S_is<double, 4>)
+	      return _knot_mask8(_mm256_fpclass_pd_mask(__x, 0x9f));
+	    else if constexpr (_Traits::template _S_is<double, 8>)
+	      return _knot_mask8(_mm512_fpclass_pd_mask(__x, 0x9f));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  {
+	    using _Up =3D typename _Traits::value_type;
+	    constexpr size_t _Np =3D _Traits::_S_full_size;
+	    const auto __a =3D __x * __infinity_v<_Up>; // NaN if __x =3D=3D 0
+	    const auto __b =3D __x * _Up();             // NaN if __x =3D=3D inf
+	    if constexpr (__have_avx512vl && __is_sse_ps<_Up, _Np>())
+	      return _mm_cmp_ps_mask(__to_intrin(__a), __to_intrin(__b),
+				     _CMP_ORD_Q);
+	    else if constexpr (__have_avx512f && __is_sse_ps<_Up, _Np>())
+	      return __mmask8(0xf
+			      & _mm512_cmp_ps_mask(__auto_bitcast(__a),
+						   __auto_bitcast(__b),
+						   _CMP_ORD_Q));
+	    else if constexpr (__have_avx512vl && __is_sse_pd<_Up, _Np>())
+	      return _mm_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
+	    else if constexpr (__have_avx512f && __is_sse_pd<_Up, _Np>())
+	      return __mmask8(0x3
+			      & _mm512_cmp_pd_mask(__auto_bitcast(__a),
+						   __auto_bitcast(__b),
+						   _CMP_ORD_Q));
+	    else if constexpr (__have_avx512vl && __is_avx_ps<_Up, _Np>())
+	      return _mm256_cmp_ps_mask(__a, __b, _CMP_ORD_Q);
+	    else if constexpr (__have_avx512f && __is_avx_ps<_Up, _Np>())
+	      return __mmask8(_mm512_cmp_ps_mask(__auto_bitcast(__a),
+						 __auto_bitcast(__b),
+						 _CMP_ORD_Q));
+	    else if constexpr (__have_avx512vl && __is_avx_pd<_Up, _Np>())
+	      return _mm256_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
+	    else if constexpr (__have_avx512f && __is_avx_pd<_Up, _Np>())
+	      return __mmask8(0xf
+			      & _mm512_cmp_pd_mask(__auto_bitcast(__a),
+						   __auto_bitcast(__b),
+						   _CMP_ORD_Q));
+	    else if constexpr (__is_avx512_ps<_Up, _Np>())
+	      return _mm512_cmp_ps_mask(__a, __b, _CMP_ORD_Q);
+	    else if constexpr (__is_avx512_pd<_Up, _Np>())
+	      return _mm512_cmp_pd_mask(__a, __b, _CMP_ORD_Q);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+      }
+
+    // }}}
+    // _S_isfinite {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_isfinite(_SimdWrapper<_Tp, _Np> __x)
+      {
+	static_assert(is_floating_point_v<_Tp>);
+#if !__FINITE_MATH_ONLY__
+	if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
+	  {
+	    const auto __xi =3D __to_intrin(__x);
+	    constexpr auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return __k1 ^ _mm512_mask_fpclass_ps_mask(__k1, __xi, 0x99);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return __k1 ^ _mm512_mask_fpclass_pd_mask(__k1, __xi, 0x99);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return __k1 ^ _mm256_mask_fpclass_ps_mask(__k1, __xi, 0x99);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return __k1 ^ _mm256_mask_fpclass_pd_mask(__k1, __xi, 0x99);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return __k1 ^ _mm_mask_fpclass_ps_mask(__k1, __xi, 0x99);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return __k1 ^ _mm_mask_fpclass_pd_mask(__k1, __xi, 0x99);
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    // if all exponent bits are set, __x is either inf or NaN
+	    using _I =3D __int_for_sizeof_t<_Tp>;
+	    const auto __inf =3D __vector_bitcast<_I>(
+	      __vector_broadcast<_Np>(__infinity_v<_Tp>));
+	    return _S_less<_I, _Np>(__vector_bitcast<_I>(__x) & __inf, __inf);
+	  }
+	else
+#endif
+	  return _Base::_S_isfinite(__x);
+      }
+
+    // }}}
+    // _S_isinf {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_isinf(_SimdWrapper<_Tp, _Np> __x)
+      {
+#if !__FINITE_MATH_ONLY__
+	if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
+	  {
+	    const auto __xi =3D __to_intrin(__x);
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_fpclass_ps_mask(__xi, 0x18);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_fpclass_pd_mask(__xi, 0x18);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_fpclass_ps_mask(__xi, 0x18);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_fpclass_pd_mask(__xi, 0x18);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_fpclass_ps_mask(__xi, 0x18);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_fpclass_pd_mask(__xi, 0x18);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_avx512dq_vl)
+	  {
+	    if constexpr (__is_sse_pd<_Tp, _Np>())
+	      return _mm_movm_epi64(_mm_fpclass_pd_mask(__x, 0x18));
+	    else if constexpr (__is_avx_pd<_Tp, _Np>())
+	      return _mm256_movm_epi64(_mm256_fpclass_pd_mask(__x, 0x18));
+	    else if constexpr (__is_sse_ps<_Tp, _Np>())
+	      return _mm_movm_epi32(
+		_mm_fpclass_ps_mask(__to_intrin(__x), 0x18));
+	    else if constexpr (__is_avx_ps<_Tp, _Np>())
+	      return _mm256_movm_epi32(_mm256_fpclass_ps_mask(__x, 0x18));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+#endif
+	  return _Base::_S_isinf(__x);
+      }
+
+    // }}}
+    // _S_isnormal {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_isnormal(_SimdWrapper<_Tp, _Np> __x)
+      {
+#if __FINITE_MATH_ONLY__
+	[[maybe_unused]] constexpr int __mode =3D 0x26;
+#else
+      [[maybe_unused]] constexpr int __mode =3D 0xbf;
+#endif
+	if constexpr (__is_avx512_abi<_Abi>() && __have_avx512dq)
+	  {
+	    const auto __xi =3D __to_intrin(__x);
+	    const auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return __k1 ^ _mm512_mask_fpclass_ps_mask(__k1, __xi, __mode);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return __k1 ^ _mm512_mask_fpclass_pd_mask(__k1, __xi, __mode);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return __k1 ^ _mm256_mask_fpclass_ps_mask(__k1, __xi, __mode);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return __k1 ^ _mm256_mask_fpclass_pd_mask(__k1, __xi, __mode);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return __k1 ^ _mm_mask_fpclass_ps_mask(__k1, __xi, __mode);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return __k1 ^ _mm_mask_fpclass_pd_mask(__k1, __xi, __mode);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_avx512dq)
+	  {
+	    if constexpr (__have_avx512vl && __is_sse_ps<_Tp, _Np>())
+	      return _mm_movm_epi32(
+		_knot_mask8(_mm_fpclass_ps_mask(__to_intrin(__x), __mode)));
+	    else if constexpr (__have_avx512vl && __is_avx_ps<_Tp, _Np>())
+	      return _mm256_movm_epi32(
+		_knot_mask8(_mm256_fpclass_ps_mask(__x, __mode)));
+	    else if constexpr (__is_avx512_ps<_Tp, _Np>())
+	      return _knot_mask16(_mm512_fpclass_ps_mask(__x, __mode));
+	    else if constexpr (__have_avx512vl && __is_sse_pd<_Tp, _Np>())
+	      return _mm_movm_epi64(
+		_knot_mask8(_mm_fpclass_pd_mask(__x, __mode)));
+	    else if constexpr (__have_avx512vl && __is_avx_pd<_Tp, _Np>())
+	      return _mm256_movm_epi64(
+		_knot_mask8(_mm256_fpclass_pd_mask(__x, __mode)));
+	    else if constexpr (__is_avx512_pd<_Tp, _Np>())
+	      return _knot_mask8(_mm512_fpclass_pd_mask(__x, __mode));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    using _I =3D __int_for_sizeof_t<_Tp>;
+	    const auto absn =3D __vector_bitcast<_I>(_S_abs(__x));
+	    const auto minn =3D __vector_bitcast<_I>(
+	      __vector_broadcast<_Np>(__norm_min_v<_Tp>));
+#if __FINITE_MATH_ONLY__
+	    return _S_less_equal<_I, _Np>(minn, absn);
+#else
+	  const auto infn
+	    =3D __vector_bitcast<_I>(__vector_broadcast<_Np>(__infinity_v<_Tp>));
+	  return __and(_S_less_equal<_I, _Np>(minn, absn),
+		       _S_less<_I, _Np>(absn, infn));
+#endif
+	  }
+	else
+	  return _Base::_S_isnormal(__x);
+      }
+
+    // }}}
+    // _S_isnan {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_isnan(_SimdWrapper<_Tp, _Np> __x)
+      { return _S_isunordered(__x, __x); }
+
+    // }}}
+    // _S_isunordered {{{
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_isunordered([[maybe_unused]] _SimdWrapper<_Tp, _Np> __x,
+		     [[maybe_unused]] _SimdWrapper<_Tp, _Np> __y)
+      {
+#if __FINITE_MATH_ONLY__
+	return {}; // false
+#else
+	const auto __xi =3D __to_intrin(__x);
+	const auto __yi =3D __to_intrin(__y);
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    constexpr auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_UNORD_Q);
+	  }
+      else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_UNORD_Q));
+      else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_UNORD_Q));
+      else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	return __auto_bitcast(_mm_cmpunord_ps(__xi, __yi));
+      else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	return __to_masktype(_mm_cmpunord_pd(__xi, __yi));
+      else
+	__assert_unreachable<_Tp>();
+#endif
+      }
+
+    // }}}
+    // _S_isgreater {{{
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp> _S_isgreater(_SimdWrapper<_Tp, _Np=
> __x,
+						     _SimdWrapper<_Tp, _Np> __y)
+      {
+	const auto __xi =3D __to_intrin(__x);
+	const auto __yi =3D __to_intrin(__y);
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    const auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GT_OQ);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_avx)
+	  {
+	    if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_GT_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_GT_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_GT_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_GT_OQ));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 4)
+	  {
+	    const auto __xn =3D __vector_bitcast<int>(__xi);
+	    const auto __yn =3D __vector_bitcast<int>(__yi);
+	    const auto __xp =3D __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
+	    const auto __yp =3D __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
+	    return __auto_bitcast(
+	      __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp > __yp));
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 8)
+	  return __vector_type_t<__int_with_sizeof_t<8>, 2>{
+	    -_mm_ucomigt_sd(__xi, __yi),
+	    -_mm_ucomigt_sd(_mm_unpackhi_pd(__xi, __xi),
+			    _mm_unpackhi_pd(__yi, __yi))};
+	else
+	  return _Base::_S_isgreater(__x, __y);
+      }
+
+    // }}}
+    // _S_isgreaterequal {{{
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_isgreaterequal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np>=
 __y)
+      {
+	const auto __xi =3D __to_intrin(__x);
+	const auto __yi =3D __to_intrin(__y);
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    const auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_GE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_GE_OQ);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_avx)
+	  {
+	    if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_GE_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_GE_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_GE_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_GE_OQ));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 4)
+	  {
+	    const auto __xn =3D __vector_bitcast<int>(__xi);
+	    const auto __yn =3D __vector_bitcast<int>(__yi);
+	    const auto __xp =3D __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
+	    const auto __yp =3D __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
+	    return __auto_bitcast(
+	      __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp >=3D __yp));
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 8)
+	  return __vector_type_t<__int_with_sizeof_t<8>, 2>{
+	    -_mm_ucomige_sd(__xi, __yi),
+	    -_mm_ucomige_sd(_mm_unpackhi_pd(__xi, __xi),
+			    _mm_unpackhi_pd(__yi, __yi))};
+	else
+	  return _Base::_S_isgreaterequal(__x, __y);
+      }
+
+    // }}}
+    // _S_isless {{{
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp> _S_isless(_SimdWrapper<_Tp, _Np> _=
_x,
+						  _SimdWrapper<_Tp, _Np> __y)
+      {
+	const auto __xi =3D __to_intrin(__x);
+	const auto __yi =3D __to_intrin(__y);
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    const auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LT_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LT_OQ);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_avx)
+	  {
+	    if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_LT_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_LT_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_LT_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_LT_OQ));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 4)
+	  {
+	    const auto __xn =3D __vector_bitcast<int>(__xi);
+	    const auto __yn =3D __vector_bitcast<int>(__yi);
+	    const auto __xp =3D __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
+	    const auto __yp =3D __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
+	    return __auto_bitcast(
+	      __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp < __yp));
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 8)
+	  return __vector_type_t<__int_with_sizeof_t<8>, 2>{
+	    -_mm_ucomigt_sd(__yi, __xi),
+	    -_mm_ucomigt_sd(_mm_unpackhi_pd(__yi, __yi),
+			    _mm_unpackhi_pd(__xi, __xi))};
+	else
+	  return _Base::_S_isless(__x, __y);
+      }
+
+    // }}}
+    // _S_islessequal {{{
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_islessequal(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> __=
y)
+      {
+	const auto __xi =3D __to_intrin(__x);
+	const auto __yi =3D __to_intrin(__y);
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    const auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_LE_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_LE_OQ);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_avx)
+	  {
+	    if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_LE_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_LE_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_LE_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_LE_OQ));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 4)
+	  {
+	    const auto __xn =3D __vector_bitcast<int>(__xi);
+	    const auto __yn =3D __vector_bitcast<int>(__yi);
+	    const auto __xp =3D __xn < 0 ? -(__xn & 0x7fff'ffff) : __xn;
+	    const auto __yp =3D __yn < 0 ? -(__yn & 0x7fff'ffff) : __yn;
+	    return __auto_bitcast(
+	      __and(__to_masktype(_mm_cmpord_ps(__xi, __yi)), __xp <=3D __yp));
+	  }
+	else if constexpr (__have_sse2 && sizeof(__xi) =3D=3D 16
+			   && sizeof(_Tp) =3D=3D 8)
+	  return __vector_type_t<__int_with_sizeof_t<8>, 2>{
+	    -_mm_ucomige_sd(__yi, __xi),
+	    -_mm_ucomige_sd(_mm_unpackhi_pd(__yi, __yi),
+			    _mm_unpackhi_pd(__xi, __xi))};
+	else
+	  return _Base::_S_islessequal(__x, __y);
+      }
+
+    // }}}
+    // _S_islessgreater {{{
+    template <typename _Tp, size_t _Np>
+      static constexpr _MaskMember<_Tp>
+      _S_islessgreater(_SimdWrapper<_Tp, _Np> __x, _SimdWrapper<_Tp, _Np> =
__y)
+      {
+	const auto __xi =3D __to_intrin(__x);
+	const auto __yi =3D __to_intrin(__y);
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    const auto __k1 =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+	    if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 4)
+	      return _mm512_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 64 && sizeof(_Tp) =3D=3D 8)
+	      return _mm512_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return _mm256_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return _mm256_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return _mm_mask_cmp_ps_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return _mm_mask_cmp_pd_mask(__k1, __xi, __yi, _CMP_NEQ_OQ);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__have_avx)
+	  {
+	    if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 4)
+	      return __to_masktype(_mm256_cmp_ps(__xi, __yi, _CMP_NEQ_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 32 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm256_cmp_pd(__xi, __yi, _CMP_NEQ_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	      return __auto_bitcast(_mm_cmp_ps(__xi, __yi, _CMP_NEQ_OQ));
+	    else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	      return __to_masktype(_mm_cmp_pd(__xi, __yi, _CMP_NEQ_OQ));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 4)
+	  return __auto_bitcast(
+	    __and(_mm_cmpord_ps(__xi, __yi), _mm_cmpneq_ps(__xi, __yi)));
+	else if constexpr (sizeof(__xi) =3D=3D 16 && sizeof(_Tp) =3D=3D 8)
+	  return __to_masktype(
+	    __and(_mm_cmpord_pd(__xi, __yi), _mm_cmpneq_pd(__xi, __yi)));
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    //}}} }}}
+  };
+
+// }}}
+// _MaskImplX86Mixin {{{
+struct _MaskImplX86Mixin
+{
+  template <typename _Tp>
+    using _TypeTag =3D _Tp*;
+
+  using _Base =3D _MaskImplBuiltinMixin;
+
+  // _S_to_maskvector(bool) {{{
+  template <typename _Up, size_t _ToN =3D 1, typename _Tp>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr enable_if_t<
+      is_same_v<_Tp, bool>, _SimdWrapper<_Up, _ToN>>
+    _S_to_maskvector(_Tp __x)
+    {
+      static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
+      return __x ? __vector_type_t<_Up, _ToN>{~_Up()}
+		 : __vector_type_t<_Up, _ToN>();
+    }
+
+  // }}}
+  // _S_to_maskvector(_SanitizedBitMask) {{{
+  template <typename _Up, size_t _UpN =3D 0, size_t _Np,
+	    size_t _ToN =3D _UpN =3D=3D 0 ? _Np : _UpN>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
+    _S_to_maskvector(_SanitizedBitMask<_Np> __x)
+    {
+      static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
+      using _UV =3D __vector_type_t<_Up, _ToN>;
+      using _UI =3D __intrinsic_type_t<_Up, _ToN>;
+      [[maybe_unused]] const auto __k =3D __x._M_to_bits();
+      if constexpr (_Np =3D=3D 1)
+	return _S_to_maskvector<_Up, _ToN>(__k);
+      else if (__x._M_is_constprop() || __builtin_is_constant_evaluated())
+	return __generate_from_n_evaluations<std::min(_ToN, _Np), _UV>(
+	  [&](auto __i) -> _Up { return -__x[__i.value]; });
+      else if constexpr (sizeof(_Up) =3D=3D 1)
+	{
+	  if constexpr (sizeof(_UI) =3D=3D 16)
+	    {
+	      if constexpr (__have_avx512bw_vl)
+		return __intrin_bitcast<_UV>(_mm_movm_epi8(__k));
+	      else if constexpr (__have_avx512bw)
+		return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi8(__k)));
+	      else if constexpr (__have_avx512f)
+		{
+		  auto __as32bits =3D _mm512_maskz_mov_epi32(__k, ~__m512i());
+		  auto __as16bits
+		    =3D __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
+						__hi256(__as32bits)));
+		  return __intrin_bitcast<_UV>(
+		    _mm_packs_epi16(__lo128(__as16bits), __hi128(__as16bits)));
+		}
+	      else if constexpr (__have_ssse3)
+		{
+		  const auto __bitmask =3D __to_intrin(
+		    __make_vector<_UChar>(1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4,
+					  8, 16, 32, 64, 128));
+		  return __intrin_bitcast<_UV>(
+		    __vector_bitcast<_Up>(
+		      _mm_shuffle_epi8(__to_intrin(
+					 __vector_type_t<_ULLong, 2>{__k}),
+				       _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1,
+						     1, 1, 1, 1, 1, 1, 1))
+		      & __bitmask)
+		    !=3D 0);
+		}
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 32)
+	    {
+	      if constexpr (__have_avx512bw_vl)
+		return __vector_bitcast<_Up>(_mm256_movm_epi8(__k));
+	      else if constexpr (__have_avx512bw)
+		return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi8(__k)));
+	      else if constexpr (__have_avx512f)
+		{
+		  auto __as16bits =3D // 0 16 1 17 ... 15 31
+		    _mm512_srli_epi32(_mm512_maskz_mov_epi32(__k, ~__m512i()),
+				      16)
+		    | _mm512_slli_epi32(_mm512_maskz_mov_epi32(__k >> 16,
+							       ~__m512i()),
+					16);
+		  auto __0_16_1_17 =3D __xzyw(_mm256_packs_epi16(
+		    __lo256(__as16bits),
+		    __hi256(__as16bits)) // 0 16 1 17 2 18 3 19 8 24 9 25 ...
+		  );
+		  // deinterleave:
+		  return __vector_bitcast<_Up>(__xzyw(_mm256_shuffle_epi8(
+		    __0_16_1_17, // 0 16 1 17 2 ...
+		    _mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9,
+				     11, 13, 15, 0, 2, 4, 6, 8, 10, 12, 14, 1,
+				     3, 5, 7, 9, 11, 13,
+				     15)))); // 0-7 16-23 8-15 24-31 -> xzyw
+					     // 0-3  8-11 16-19 24-27
+					     // 4-7 12-15 20-23 28-31
+		}
+	      else if constexpr (__have_avx2)
+		{
+		  const auto __bitmask
+		    =3D _mm256_broadcastsi128_si256(__to_intrin(
+		      __make_vector<_UChar>(1, 2, 4, 8, 16, 32, 64, 128, 1, 2,
+					    4, 8, 16, 32, 64, 128)));
+		  return __vector_bitcast<_Up>(
+		    __vector_bitcast<_Up>(
+		      _mm256_shuffle_epi8(
+			_mm256_broadcastsi128_si256(
+			  __to_intrin(__vector_type_t<_ULLong, 2>{__k})),
+			_mm256_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,
+					 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3,
+					 3, 3, 3, 3, 3, 3))
+		      & __bitmask)
+		    !=3D 0);
+		}
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 64)
+	    return reinterpret_cast<_UV>(_mm512_movm_epi8(__k));
+	  if constexpr (std::min(_ToN, _Np) <=3D 4)
+	    {
+	      if constexpr (_Np > 7) // avoid overflow
+		__x &=3D _SanitizedBitMask<_Np>(0x0f);
+	      const _UInt __char_mask
+		=3D ((_UInt(__x.to_ulong()) * 0x00204081U) & 0x01010101ULL)
+		  * 0xff;
+	      _UV __r =3D {};
+	      __builtin_memcpy(&__r, &__char_mask,
+			       std::min(sizeof(__r), sizeof(__char_mask)));
+	      return __r;
+	    }
+	  else if constexpr (std::min(_ToN, _Np) <=3D 7)
+	    {
+	      if constexpr (_Np > 7) // avoid overflow
+		__x &=3D _SanitizedBitMask<_Np>(0x7f);
+	      const _ULLong __char_mask
+		=3D ((__x.to_ulong() * 0x40810204081ULL) & 0x0101010101010101ULL)
+		  * 0xff;
+	      _UV __r =3D {};
+	      __builtin_memcpy(&__r, &__char_mask,
+			       std::min(sizeof(__r), sizeof(__char_mask)));
+	      return __r;
+	    }
+	}
+      else if constexpr (sizeof(_Up) =3D=3D 2)
+	{
+	  if constexpr (sizeof(_UI) =3D=3D 16)
+	    {
+	      if constexpr (__have_avx512bw_vl)
+		return __intrin_bitcast<_UV>(_mm_movm_epi16(__k));
+	      else if constexpr (__have_avx512bw)
+		return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi16(__k)));
+	      else if constexpr (__have_avx512f)
+		{
+		  __m256i __as32bits =3D {};
+		  if constexpr (__have_avx512vl)
+		    __as32bits =3D _mm256_maskz_mov_epi32(__k, ~__m256i());
+		  else
+		    __as32bits
+		      =3D __lo256(_mm512_maskz_mov_epi32(__k, ~__m512i()));
+		  return __intrin_bitcast<_UV>(
+		    _mm_packs_epi32(__lo128(__as32bits), __hi128(__as32bits)));
+		}
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 32)
+	    {
+	      if constexpr (__have_avx512bw_vl)
+		return __vector_bitcast<_Up>(_mm256_movm_epi16(__k));
+	      else if constexpr (__have_avx512bw)
+		return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi16(__k)));
+	      else if constexpr (__have_avx512f)
+		{
+		  auto __as32bits =3D _mm512_maskz_mov_epi32(__k, ~__m512i());
+		  return __vector_bitcast<_Up>(
+		    __xzyw(_mm256_packs_epi32(__lo256(__as32bits),
+					      __hi256(__as32bits))));
+		}
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 64)
+	    return __vector_bitcast<_Up>(_mm512_movm_epi16(__k));
+	}
+      else if constexpr (sizeof(_Up) =3D=3D 4)
+	{
+	  if constexpr (sizeof(_UI) =3D=3D 16)
+	    {
+	      if constexpr (__have_avx512dq_vl)
+		return __intrin_bitcast<_UV>(_mm_movm_epi32(__k));
+	      else if constexpr (__have_avx512dq)
+		return __intrin_bitcast<_UV>(__lo128(_mm512_movm_epi32(__k)));
+	      else if constexpr (__have_avx512vl)
+		return __intrin_bitcast<_UV>(
+		  _mm_maskz_mov_epi32(__k, ~__m128i()));
+	      else if constexpr (__have_avx512f)
+		return __intrin_bitcast<_UV>(
+		  __lo128(_mm512_maskz_mov_epi32(__k, ~__m512i())));
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 32)
+	    {
+	      if constexpr (__have_avx512dq_vl)
+		return __vector_bitcast<_Up>(_mm256_movm_epi32(__k));
+	      else if constexpr (__have_avx512dq)
+		return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi32(__k)));
+	      else if constexpr (__have_avx512vl)
+		return __vector_bitcast<_Up>(
+		  _mm256_maskz_mov_epi32(__k, ~__m256i()));
+	      else if constexpr (__have_avx512f)
+		return __vector_bitcast<_Up>(
+		  __lo256(_mm512_maskz_mov_epi32(__k, ~__m512i())));
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 64)
+	    return __vector_bitcast<_Up>(
+	      __have_avx512dq ? _mm512_movm_epi32(__k)
+			      : _mm512_maskz_mov_epi32(__k, ~__m512i()));
+	}
+      else if constexpr (sizeof(_Up) =3D=3D 8)
+	{
+	  if constexpr (sizeof(_UI) =3D=3D 16)
+	    {
+	      if constexpr (__have_avx512dq_vl)
+		return __vector_bitcast<_Up>(_mm_movm_epi64(__k));
+	      else if constexpr (__have_avx512dq)
+		return __vector_bitcast<_Up>(__lo128(_mm512_movm_epi64(__k)));
+	      else if constexpr (__have_avx512vl)
+		return __vector_bitcast<_Up>(
+		  _mm_maskz_mov_epi64(__k, ~__m128i()));
+	      else if constexpr (__have_avx512f)
+		return __vector_bitcast<_Up>(
+		  __lo128(_mm512_maskz_mov_epi64(__k, ~__m512i())));
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 32)
+	    {
+	      if constexpr (__have_avx512dq_vl)
+		return __vector_bitcast<_Up>(_mm256_movm_epi64(__k));
+	      else if constexpr (__have_avx512dq)
+		return __vector_bitcast<_Up>(__lo256(_mm512_movm_epi64(__k)));
+	      else if constexpr (__have_avx512vl)
+		return __vector_bitcast<_Up>(
+		  _mm256_maskz_mov_epi64(__k, ~__m256i()));
+	      else if constexpr (__have_avx512f)
+		return __vector_bitcast<_Up>(
+		  __lo256(_mm512_maskz_mov_epi64(__k, ~__m512i())));
+	      // else fall through
+	    }
+	  else if constexpr (sizeof(_UI) =3D=3D 64)
+	    return __vector_bitcast<_Up>(
+	      __have_avx512dq ? _mm512_movm_epi64(__k)
+			      : _mm512_maskz_mov_epi64(__k, ~__m512i()));
+	}
+
+      using _UpUInt =3D make_unsigned_t<_Up>;
+      using _V =3D __vector_type_t<_UpUInt, _ToN>;
+      constexpr size_t __bits_per_element =3D sizeof(_Up) * __CHAR_BIT__;
+      if constexpr (_ToN =3D=3D 2)
+	{
+	  return __vector_bitcast<_Up>(_V{_UpUInt(-__x[0]), _UpUInt(-__x[1])});
+	}
+      else if constexpr (!__have_avx2 && __have_avx && sizeof(_V) =3D=3D 3=
2)
+	{
+	  if constexpr (sizeof(_Up) =3D=3D 4)
+	    return __vector_bitcast<_Up>(_mm256_cmp_ps(
+	      _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(__k)),
+			    _mm256_castsi256_ps(_mm256_setr_epi32(
+			      0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80))),
+	      _mm256_setzero_ps(), _CMP_NEQ_UQ));
+	  else if constexpr (sizeof(_Up) =3D=3D 8)
+	    return __vector_bitcast<_Up>(_mm256_cmp_pd(
+	      _mm256_and_pd(_mm256_castsi256_pd(_mm256_set1_epi64x(__k)),
+			    _mm256_castsi256_pd(
+			      _mm256_setr_epi64x(0x01, 0x02, 0x04, 0x08))),
+	      _mm256_setzero_pd(), _CMP_NEQ_UQ));
+	  else
+	    __assert_unreachable<_Up>();
+	}
+      else if constexpr (__bits_per_element >=3D _ToN)
+	{
+	  constexpr auto __bitmask
+	    =3D __generate_vector<_V>([](auto __i) constexpr->_UpUInt {
+		return __i < _ToN ? 1ull << __i : 0;
+	      });
+	  const auto __bits
+	    =3D __vector_broadcast<_ToN, _UpUInt>(__k) & __bitmask;
+	  if constexpr (__bits_per_element > _ToN)
+	    return __vector_bitcast<_Up>(__bits) > 0;
+	  else
+	    return __vector_bitcast<_Up>(__bits !=3D 0);
+	}
+      else
+	{
+	  const _V __tmp
+	    =3D __generate_vector<_V>([&](auto __i) constexpr {
+		return static_cast<_UpUInt>(
+		  __k >> (__bits_per_element * (__i / __bits_per_element)));
+	      })
+	      & __generate_vector<_V>([](auto __i) constexpr {
+		  return static_cast<_UpUInt>(1ull
+					      << (__i % __bits_per_element));
+		}); // mask bit index
+	  return __intrin_bitcast<_UV>(__tmp !=3D _V());
+	}
+    }
+
+  // }}}
+  // _S_to_maskvector(_SimdWrapper) {{{
+  template <typename _Up, size_t _UpN =3D 0, typename _Tp, size_t _Np,
+	    size_t _ToN =3D _UpN =3D=3D 0 ? _Np : _UpN>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Up, _ToN>
+    _S_to_maskvector(_SimdWrapper<_Tp, _Np> __x)
+    {
+      static_assert(is_same_v<_Up, __int_for_sizeof_t<_Up>>);
+      using _TW =3D _SimdWrapper<_Tp, _Np>;
+      using _UW =3D _SimdWrapper<_Up, _ToN>;
+      using _UI =3D __intrinsic_type_t<_Up, _ToN>;
+      if constexpr (is_same_v<_Tp, bool>) // bits -> vector
+	return _S_to_maskvector<_Up, _ToN>(
+	  _BitMask<_Np>(__x._M_data)._M_sanitized());
+      // vector -> vector bitcast
+      else if constexpr (sizeof(_Up) =3D=3D sizeof(_Tp)
+			 && sizeof(_TW) =3D=3D sizeof(_UW))
+	return __wrapper_bitcast<_Up, _ToN>(
+	  _ToN <=3D _Np
+	    ? __x
+	    : simd_abi::_VecBuiltin<sizeof(_Tp) * _Np>::_S_masked(__x));
+      else // vector -> vector {{{
+	{
+	  if (__x._M_is_constprop() || __builtin_is_constant_evaluated())
+	    {
+	      const auto __y =3D __vector_bitcast<__int_for_sizeof_t<_Tp>>(__x);
+	      return __generate_from_n_evaluations<std::min(_ToN, _Np),
+						   __vector_type_t<_Up, _ToN>>(
+		[&](auto __i) -> _Up { return __y[__i.value]; });
+	    }
+	  using _To =3D __vector_type_t<_Up, _ToN>;
+	  [[maybe_unused]] constexpr size_t _FromN =3D _Np;
+	  constexpr int _FromBytes =3D sizeof(_Tp);
+	  constexpr int _ToBytes =3D sizeof(_Up);
+	  const auto __k =3D __x._M_data;
+
+	  if constexpr (_FromBytes =3D=3D _ToBytes)
+	    return __intrin_bitcast<_To>(__k);
+	  else if constexpr (sizeof(_UI) =3D=3D 16 && sizeof(__k) =3D=3D 16)
+	    { // SSE -> SSE {{{
+	      if constexpr (_FromBytes =3D=3D 4 && _ToBytes =3D=3D 8)
+		return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
+	      else if constexpr (_FromBytes =3D=3D 2 && _ToBytes =3D=3D 8)
+		{
+		  const auto __y
+		    =3D __vector_bitcast<int>(__interleave128_lo(__k, __k));
+		  return __intrin_bitcast<_To>(__interleave128_lo(__y, __y));
+		}
+	      else if constexpr (_FromBytes =3D=3D 1 && _ToBytes =3D=3D 8)
+		{
+		  auto __y
+		    =3D __vector_bitcast<short>(__interleave128_lo(__k, __k));
+		  auto __z
+		    =3D __vector_bitcast<int>(__interleave128_lo(__y, __y));
+		  return __intrin_bitcast<_To>(__interleave128_lo(__z, __z));
+		}
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 4
+				 && __have_sse2)
+		return __intrin_bitcast<_To>(
+		  _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()));
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 4)
+		return __vector_shuffle<1, 3, 6, 7>(__vector_bitcast<_Up>(__k),
+						    _UI());
+	      else if constexpr (_FromBytes =3D=3D 2 && _ToBytes =3D=3D 4)
+		return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
+	      else if constexpr (_FromBytes =3D=3D 1 && _ToBytes =3D=3D 4)
+		{
+		  const auto __y
+		    =3D __vector_bitcast<short>(__interleave128_lo(__k, __k));
+		  return __intrin_bitcast<_To>(__interleave128_lo(__y, __y));
+		}
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 2)
+		{
+		  if constexpr (__have_sse2 && !__have_ssse3)
+		    return __intrin_bitcast<_To>(_mm_packs_epi32(
+		      _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()),
+		      __m128i()));
+		  else
+		    return __intrin_bitcast<_To>(
+		      __vector_permute<3, 7, -1, -1, -1, -1, -1, -1>(
+			__vector_bitcast<_Up>(__k)));
+		}
+	      else if constexpr (_FromBytes =3D=3D 4 && _ToBytes =3D=3D 2)
+		return __intrin_bitcast<_To>(
+		  _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i()));
+	      else if constexpr (_FromBytes =3D=3D 1 && _ToBytes =3D=3D 2)
+		return __intrin_bitcast<_To>(__interleave128_lo(__k, __k));
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 1
+				 && __have_ssse3)
+		return __intrin_bitcast<_To>(
+		  _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+				   _mm_setr_epi8(7, 15, -1, -1, -1, -1, -1, -1,
+						 -1, -1, -1, -1, -1, -1, -1,
+						 -1)));
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 1)
+		{
+		  auto __y
+		    =3D _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i());
+		  __y =3D _mm_packs_epi32(__y, __m128i());
+		  return __intrin_bitcast<_To>(_mm_packs_epi16(__y, __m128i()));
+		}
+	      else if constexpr (_FromBytes =3D=3D 4 && _ToBytes =3D=3D 1
+				 && __have_ssse3)
+		return __intrin_bitcast<_To>(
+		  _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+				   _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1,
+						 -1, -1, -1, -1, -1, -1, -1,
+						 -1)));
+	      else if constexpr (_FromBytes =3D=3D 4 && _ToBytes =3D=3D 1)
+		{
+		  const auto __y
+		    =3D _mm_packs_epi32(__vector_bitcast<_LLong>(__k), __m128i());
+		  return __intrin_bitcast<_To>(_mm_packs_epi16(__y, __m128i()));
+		}
+	      else if constexpr (_FromBytes =3D=3D 2 && _ToBytes =3D=3D 1)
+		return __intrin_bitcast<_To>(
+		  _mm_packs_epi16(__vector_bitcast<_LLong>(__k), __m128i()));
+	      else
+		__assert_unreachable<_Tp>();
+	    } // }}}
+	  else if constexpr (sizeof(_UI) =3D=3D 32 && sizeof(__k) =3D=3D 32)
+	    { // AVX -> AVX {{{
+	      if constexpr (_FromBytes =3D=3D _ToBytes)
+		__assert_unreachable<_Tp>();
+	      else if constexpr (_FromBytes =3D=3D _ToBytes * 2)
+		{
+		  const auto __y =3D __vector_bitcast<_LLong>(__k);
+		  return __intrin_bitcast<_To>(_mm256_castsi128_si256(
+		    _mm_packs_epi16(__lo128(__y), __hi128(__y))));
+		}
+	      else if constexpr (_FromBytes =3D=3D _ToBytes * 4)
+		{
+		  const auto __y =3D __vector_bitcast<_LLong>(__k);
+		  return __intrin_bitcast<_To>(_mm256_castsi128_si256(
+		    _mm_packs_epi16(_mm_packs_epi16(__lo128(__y), __hi128(__y)),
+				    __m128i())));
+		}
+	      else if constexpr (_FromBytes =3D=3D _ToBytes * 8)
+		{
+		  const auto __y =3D __vector_bitcast<_LLong>(__k);
+		  return __intrin_bitcast<_To>(
+		    _mm256_castsi128_si256(_mm_shuffle_epi8(
+		      _mm_packs_epi16(__lo128(__y), __hi128(__y)),
+		      _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1,
+				    -1, -1, -1, -1, -1))));
+		}
+	      else if constexpr (_FromBytes * 2 =3D=3D _ToBytes)
+		{
+		  auto __y =3D __xzyw(__to_intrin(__k));
+		  if constexpr (is_floating_point_v<
+				  _Tp> || (!__have_avx2 && _FromBytes =3D=3D 4))
+		    {
+		      const auto __yy =3D __vector_bitcast<float>(__y);
+		      return __intrin_bitcast<_To>(
+			_mm256_unpacklo_ps(__yy, __yy));
+		    }
+		  else
+		    return __intrin_bitcast<_To>(
+		      _mm256_unpacklo_epi8(__y, __y));
+		}
+	      else if constexpr (_FromBytes * 4 =3D=3D _ToBytes)
+		{
+		  auto __y
+		    =3D _mm_unpacklo_epi8(__lo128(__vector_bitcast<_LLong>(__k)),
+					__lo128(__vector_bitcast<_LLong>(
+					  __k))); // drops 3/4 of input
+		  return __intrin_bitcast<_To>(
+		    __concat(_mm_unpacklo_epi16(__y, __y),
+			     _mm_unpackhi_epi16(__y, __y)));
+		}
+	      else if constexpr (_FromBytes =3D=3D 1 && _ToBytes =3D=3D 8)
+		{
+		  auto __y
+		    =3D _mm_unpacklo_epi8(__lo128(__vector_bitcast<_LLong>(__k)),
+					__lo128(__vector_bitcast<_LLong>(
+					  __k))); // drops 3/4 of input
+		  __y
+		    =3D _mm_unpacklo_epi16(__y,
+					 __y); // drops another 1/2 =3D> 7/8 total
+		  return __intrin_bitcast<_To>(
+		    __concat(_mm_unpacklo_epi32(__y, __y),
+			     _mm_unpackhi_epi32(__y, __y)));
+		}
+	      else
+		__assert_unreachable<_Tp>();
+	    } // }}}
+	  else if constexpr (sizeof(_UI) =3D=3D 32 && sizeof(__k) =3D=3D 16)
+	    { // SSE -> AVX {{{
+	      if constexpr (_FromBytes =3D=3D _ToBytes)
+		return __intrin_bitcast<_To>(
+		  __intrinsic_type_t<_Tp, 32 / sizeof(_Tp)>(
+		    __zero_extend(__to_intrin(__k))));
+	      else if constexpr (_FromBytes * 2 =3D=3D _ToBytes)
+		{ // keep all
+		  return __intrin_bitcast<_To>(
+		    __concat(_mm_unpacklo_epi8(__vector_bitcast<_LLong>(__k),
+					       __vector_bitcast<_LLong>(__k)),
+			     _mm_unpackhi_epi8(__vector_bitcast<_LLong>(__k),
+					       __vector_bitcast<_LLong>(__k))));
+		}
+	      else if constexpr (_FromBytes * 4 =3D=3D _ToBytes)
+		{
+		  if constexpr (__have_avx2)
+		    {
+		      return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
+			__concat(__vector_bitcast<_LLong>(__k),
+				 __vector_bitcast<_LLong>(__k)),
+			_mm256_setr_epi8(0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3,
+					 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 6, 6,
+					 6, 6, 7, 7, 7, 7)));
+		    }
+		  else
+		    {
+		      return __intrin_bitcast<_To>(__concat(
+			_mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+					 _mm_setr_epi8(0, 0, 0, 0, 1, 1, 1, 1,
+						       2, 2, 2, 2, 3, 3, 3, 3)),
+			_mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+					 _mm_setr_epi8(4, 4, 4, 4, 5, 5, 5, 5,
+						       6, 6, 6, 6, 7, 7, 7,
+						       7))));
+		    }
+		}
+	      else if constexpr (_FromBytes * 8 =3D=3D _ToBytes)
+		{
+		  if constexpr (__have_avx2)
+		    {
+		      return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
+			__concat(__vector_bitcast<_LLong>(__k),
+				 __vector_bitcast<_LLong>(__k)),
+			_mm256_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,
+					 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3,
+					 3, 3, 3, 3, 3, 3)));
+		    }
+		  else
+		    {
+		      return __intrin_bitcast<_To>(__concat(
+			_mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+					 _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+						       1, 1, 1, 1, 1, 1, 1, 1)),
+			_mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+					 _mm_setr_epi8(2, 2, 2, 2, 2, 2, 2, 2,
+						       3, 3, 3, 3, 3, 3, 3,
+						       3))));
+		    }
+		}
+	      else if constexpr (_FromBytes =3D=3D _ToBytes * 2)
+		return __intrin_bitcast<_To>(__m256i(__zero_extend(
+		  _mm_packs_epi16(__vector_bitcast<_LLong>(__k), __m128i()))));
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 2)
+		{
+		  return __intrin_bitcast<_To>(__m256i(__zero_extend(
+		    _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+				     _mm_setr_epi8(6, 7, 14, 15, -1, -1, -1, -1,
+						   -1, -1, -1, -1, -1, -1, -1,
+						   -1)))));
+		}
+	      else if constexpr (_FromBytes =3D=3D 4 && _ToBytes =3D=3D 1)
+		{
+		  return __intrin_bitcast<_To>(__m256i(__zero_extend(
+		    _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+				     _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1,
+						   -1, -1, -1, -1, -1, -1, -1,
+						   -1)))));
+		}
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 1)
+		{
+		  return __intrin_bitcast<_To>(__m256i(__zero_extend(
+		    _mm_shuffle_epi8(__vector_bitcast<_LLong>(__k),
+				     _mm_setr_epi8(7, 15, -1, -1, -1, -1, -1,
+						   -1, -1, -1, -1, -1, -1, -1,
+						   -1, -1)))));
+		}
+	      else
+		static_assert(!is_same_v<_Tp, _Tp>, "should be unreachable");
+	    } // }}}
+	  else if constexpr (sizeof(_UI) =3D=3D 16 && sizeof(__k) =3D=3D 32)
+	    { // AVX -> SSE {{{
+	      if constexpr (_FromBytes =3D=3D _ToBytes)
+		{ // keep low 1/2
+		  return __intrin_bitcast<_To>(__lo128(__k));
+		}
+	      else if constexpr (_FromBytes =3D=3D _ToBytes * 2)
+		{ // keep all
+		  auto __y =3D __vector_bitcast<_LLong>(__k);
+		  return __intrin_bitcast<_To>(
+		    _mm_packs_epi16(__lo128(__y), __hi128(__y)));
+		}
+	      else if constexpr (_FromBytes =3D=3D _ToBytes * 4)
+		{ // add 1/2 undef
+		  auto __y =3D __vector_bitcast<_LLong>(__k);
+		  return __intrin_bitcast<_To>(
+		    _mm_packs_epi16(_mm_packs_epi16(__lo128(__y), __hi128(__y)),
+				    __m128i()));
+		}
+	      else if constexpr (_FromBytes =3D=3D 8 && _ToBytes =3D=3D 1)
+		{ // add 3/4 undef
+		  auto __y =3D __vector_bitcast<_LLong>(__k);
+		  return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+		    _mm_packs_epi16(__lo128(__y), __hi128(__y)),
+		    _mm_setr_epi8(3, 7, 11, 15, -1, -1, -1, -1, -1, -1, -1, -1,
+				  -1, -1, -1, -1)));
+		}
+	      else if constexpr (_FromBytes * 2 =3D=3D _ToBytes)
+		{ // keep low 1/4
+		  auto __y =3D __lo128(__vector_bitcast<_LLong>(__k));
+		  return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
+		}
+	      else if constexpr (_FromBytes * 4 =3D=3D _ToBytes)
+		{ // keep low 1/8
+		  auto __y =3D __lo128(__vector_bitcast<_LLong>(__k));
+		  __y =3D _mm_unpacklo_epi8(__y, __y);
+		  return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
+		}
+	      else if constexpr (_FromBytes * 8 =3D=3D _ToBytes)
+		{ // keep low 1/16
+		  auto __y =3D __lo128(__vector_bitcast<_LLong>(__k));
+		  __y =3D _mm_unpacklo_epi8(__y, __y);
+		  __y =3D _mm_unpacklo_epi8(__y, __y);
+		  return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__y, __y));
+		}
+	      else
+		static_assert(!is_same_v<_Tp, _Tp>, "should be unreachable");
+	    } // }}}
+	  else
+	    return _Base::template _S_to_maskvector<_Up, _ToN>(__x);
+	  /*
+	  if constexpr (_FromBytes > _ToBytes) {
+	      const _To     __y      =3D __vector_bitcast<_Up>(__k);
+	      return [&] <size_t... _Is> (index_sequence<_Is...>) {
+		constexpr int _Stride =3D _FromBytes / _ToBytes;
+		return _To{__y[(_Is + 1) * _Stride - 1]...};
+	      }(make_index_sequence<std::min(_ToN, _FromN)>());
+	  } else {
+	      // {0, 0, 1, 1} (_Dups =3D 2, _Is<4>)
+	      // {0, 0, 0, 0, 1, 1, 1, 1} (_Dups =3D 4, _Is<8>)
+	      // {0, 0, 1, 1, 2, 2, 3, 3} (_Dups =3D 2, _Is<8>)
+	      // ...
+	      return [&] <size_t... _Is> (index_sequence<_Is...>) {
+		constexpr int __dup =3D _ToBytes / _FromBytes;
+		return __intrin_bitcast<_To>(_From{__k[_Is / __dup]...});
+	      }(make_index_sequence<_FromN>());
+	  }
+	  */
+	} // }}}
+    }
+
+  // }}}
+  // _S_to_bits {{{
+  template <typename _Tp, size_t _Np>
+    _GLIBCXX_SIMD_INTRINSIC static constexpr _SanitizedBitMask<_Np>
+    _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
+    {
+      if constexpr (is_same_v<_Tp, bool>)
+	return _BitMask<_Np>(__x._M_data)._M_sanitized();
+      else
+	{
+	  static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
+	  if (__builtin_is_constant_evaluated()
+	      || __builtin_constant_p(__x._M_data))
+	    {
+	      const auto __bools =3D -__x._M_data;
+	      const _ULLong __k =3D __call_with_n_evaluations<_Np>(
+		[](auto... __bits) { return (__bits | ...); },
+		[&](auto __i) { return _ULLong(__bools[+__i]) << __i; });
+	      if (__builtin_is_constant_evaluated()
+		  || __builtin_constant_p(__k))
+		return __k;
+	    }
+	  const auto __xi =3D __to_intrin(__x);
+	  if constexpr (sizeof(_Tp) =3D=3D 1)
+	    if constexpr (sizeof(__xi) =3D=3D 16)
+	      if constexpr (__have_avx512bw_vl)
+		return _BitMask<_Np>(_mm_movepi8_mask(__xi));
+	      else // implies SSE2
+		return _BitMask<_Np>(_mm_movemask_epi8(__xi));
+	    else if constexpr (sizeof(__xi) =3D=3D 32)
+	      if constexpr (__have_avx512bw_vl)
+		return _BitMask<_Np>(_mm256_movepi8_mask(__xi));
+	      else // implies AVX2
+		return _BitMask<_Np>(_mm256_movemask_epi8(__xi));
+	    else // implies AVX512BW
+	      return _BitMask<_Np>(_mm512_movepi8_mask(__xi));
+
+	  else if constexpr (sizeof(_Tp) =3D=3D 2)
+	    if constexpr (sizeof(__xi) =3D=3D 16)
+	      if constexpr (__have_avx512bw_vl)
+		return _BitMask<_Np>(_mm_movepi16_mask(__xi));
+	      else if constexpr (__have_avx512bw)
+		return _BitMask<_Np>(_mm512_movepi16_mask(__zero_extend(__xi)));
+	      else // implies SSE2
+		return _BitMask<_Np>(
+		  _mm_movemask_epi8(_mm_packs_epi16(__xi, __m128i())));
+	    else if constexpr (sizeof(__xi) =3D=3D 32)
+	      if constexpr (__have_avx512bw_vl)
+		return _BitMask<_Np>(_mm256_movepi16_mask(__xi));
+	      else if constexpr (__have_avx512bw)
+		return _BitMask<_Np>(_mm512_movepi16_mask(__zero_extend(__xi)));
+	      else // implies SSE2
+		return _BitMask<_Np>(_mm_movemask_epi8(
+		  _mm_packs_epi16(__lo128(__xi), __hi128(__xi))));
+	    else // implies AVX512BW
+	      return _BitMask<_Np>(_mm512_movepi16_mask(__xi));
+
+	  else if constexpr (sizeof(_Tp) =3D=3D 4)
+	    if constexpr (sizeof(__xi) =3D=3D 16)
+	      if constexpr (__have_avx512dq_vl)
+		return _BitMask<_Np>(_mm_movepi32_mask(__xi));
+	      else if constexpr (__have_avx512vl)
+		return _BitMask<_Np>(_mm_cmplt_epi32_mask(__xi, __m128i()));
+	      else if constexpr (__have_avx512dq)
+		return _BitMask<_Np>(_mm512_movepi32_mask(__zero_extend(__xi)));
+	      else if constexpr (__have_avx512f)
+		return _BitMask<_Np>(
+		  _mm512_cmplt_epi32_mask(__zero_extend(__xi), __m512i()));
+	      else // implies SSE
+		return _BitMask<_Np>(
+		  _mm_movemask_ps(reinterpret_cast<__m128>(__xi)));
+	    else if constexpr (sizeof(__xi) =3D=3D 32)
+	      if constexpr (__have_avx512dq_vl)
+		return _BitMask<_Np>(_mm256_movepi32_mask(__xi));
+	      else if constexpr (__have_avx512dq)
+		return _BitMask<_Np>(_mm512_movepi32_mask(__zero_extend(__xi)));
+	      else if constexpr (__have_avx512vl)
+		return _BitMask<_Np>(_mm256_cmplt_epi32_mask(__xi, __m256i()));
+	      else if constexpr (__have_avx512f)
+		return _BitMask<_Np>(
+		  _mm512_cmplt_epi32_mask(__zero_extend(__xi), __m512i()));
+	      else // implies AVX
+		return _BitMask<_Np>(
+		  _mm256_movemask_ps(reinterpret_cast<__m256>(__xi)));
+	    else // implies AVX512??
+	      if constexpr (__have_avx512dq)
+	      return _BitMask<_Np>(_mm512_movepi32_mask(__xi));
+	    else // implies AVX512F
+	      return _BitMask<_Np>(_mm512_cmplt_epi32_mask(__xi, __m512i()));
+
+	  else if constexpr (sizeof(_Tp) =3D=3D 8)
+	    if constexpr (sizeof(__xi) =3D=3D 16)
+	      if constexpr (__have_avx512dq_vl)
+		return _BitMask<_Np>(_mm_movepi64_mask(__xi));
+	      else if constexpr (__have_avx512dq)
+		return _BitMask<_Np>(_mm512_movepi64_mask(__zero_extend(__xi)));
+	      else if constexpr (__have_avx512vl)
+		return _BitMask<_Np>(_mm_cmplt_epi64_mask(__xi, __m128i()));
+	      else if constexpr (__have_avx512f)
+		return _BitMask<_Np>(
+		  _mm512_cmplt_epi64_mask(__zero_extend(__xi), __m512i()));
+	      else // implies SSE2
+		return _BitMask<_Np>(
+		  _mm_movemask_pd(reinterpret_cast<__m128d>(__xi)));
+	    else if constexpr (sizeof(__xi) =3D=3D 32)
+	      if constexpr (__have_avx512dq_vl)
+		return _BitMask<_Np>(_mm256_movepi64_mask(__xi));
+	      else if constexpr (__have_avx512dq)
+		return _BitMask<_Np>(_mm512_movepi64_mask(__zero_extend(__xi)));
+	      else if constexpr (__have_avx512vl)
+		return _BitMask<_Np>(_mm256_cmplt_epi64_mask(__xi, __m256i()));
+	      else if constexpr (__have_avx512f)
+		return _BitMask<_Np>(
+		  _mm512_cmplt_epi64_mask(__zero_extend(__xi), __m512i()));
+	      else // implies AVX
+		return _BitMask<_Np>(
+		  _mm256_movemask_pd(reinterpret_cast<__m256d>(__xi)));
+	    else // implies AVX512??
+	      if constexpr (__have_avx512dq)
+	      return _BitMask<_Np>(_mm512_movepi64_mask(__xi));
+	    else // implies AVX512F
+	      return _BitMask<_Np>(_mm512_cmplt_epi64_mask(__xi, __m512i()));
+
+	  else
+	    __assert_unreachable<_Tp>();
+	}
+    }
+  // }}}
+};
+
+// }}}
+// _MaskImplX86 {{{
+template <typename _Abi>
+  struct _MaskImplX86 : _MaskImplX86Mixin, _MaskImplBuiltin<_Abi>
+  {
+    using _MaskImplX86Mixin::_S_to_bits;
+    using _MaskImplX86Mixin::_S_to_maskvector;
+    using _MaskImplBuiltin<_Abi>::_S_convert;
+
+    // member types {{{
+    template <typename _Tp>
+      using _SimdMember =3D typename _Abi::template __traits<_Tp>::_SimdMe=
mber;
+
+    template <typename _Tp>
+      using _MaskMember =3D typename _Abi::template _MaskMember<_Tp>;
+
+    template <typename _Tp>
+      static constexpr size_t _S_size =3D simd_size_v<_Tp, _Abi>;
+
+    using _Base =3D _MaskImplBuiltin<_Abi>;
+
+    // }}}
+    // _S_broadcast {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_broadcast(bool __x)
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  return __x ? _Abi::_S_masked(_MaskMember<_Tp>(-1))
+		     : _MaskMember<_Tp>();
+	else
+	  return _Base::template _S_broadcast<_Tp>(__x);
+      }
+
+    // }}}
+    // _S_load {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _MaskMember<_Tp>
+      _S_load(const bool* __mem)
+      {
+	static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
+	if constexpr (__have_avx512bw)
+	  {
+	    const auto __to_vec_or_bits =3D [](auto __bits) -> decltype(auto) {
+	      if constexpr (__is_avx512_abi<_Abi>())
+		return __bits;
+	      else
+		return _S_to_maskvector<_Tp>(
+		  _BitMask<_S_size<_Tp>>(__bits)._M_sanitized());
+	    };
+
+	    if constexpr (_S_size<_Tp> <=3D 16 && __have_avx512vl)
+	      {
+		__m128i __a =3D {};
+		__builtin_memcpy(&__a, __mem, _S_size<_Tp>);
+		return __to_vec_or_bits(_mm_test_epi8_mask(__a, __a));
+	      }
+	    else if constexpr (_S_size<_Tp> <=3D 32 && __have_avx512vl)
+	      {
+		__m256i __a =3D {};
+		__builtin_memcpy(&__a, __mem, _S_size<_Tp>);
+		return __to_vec_or_bits(_mm256_test_epi8_mask(__a, __a));
+	      }
+	    else if constexpr (_S_size<_Tp> <=3D 64)
+	      {
+		__m512i __a =3D {};
+		__builtin_memcpy(&__a, __mem, _S_size<_Tp>);
+		return __to_vec_or_bits(_mm512_test_epi8_mask(__a, __a));
+	      }
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    if constexpr (_S_size<_Tp> <=3D 8)
+	      {
+		__m128i __a =3D {};
+		__builtin_memcpy(&__a, __mem, _S_size<_Tp>);
+		const auto __b =3D _mm512_cvtepi8_epi64(__a);
+		return _mm512_test_epi64_mask(__b, __b);
+	      }
+	    else if constexpr (_S_size<_Tp> <=3D 16)
+	      {
+		__m128i __a =3D {};
+		__builtin_memcpy(&__a, __mem, _S_size<_Tp>);
+		const auto __b =3D _mm512_cvtepi8_epi32(__a);
+		return _mm512_test_epi32_mask(__b, __b);
+	      }
+	    else if constexpr (_S_size<_Tp> <=3D 32)
+	      {
+		__m128i __a =3D {};
+		__builtin_memcpy(&__a, __mem, 16);
+		const auto __b =3D _mm512_cvtepi8_epi32(__a);
+		__builtin_memcpy(&__a, __mem + 16, _S_size<_Tp> - 16);
+		const auto __c =3D _mm512_cvtepi8_epi32(__a);
+		return _mm512_test_epi32_mask(__b, __b)
+		       | (_mm512_test_epi32_mask(__c, __c) << 16);
+	      }
+	    else if constexpr (_S_size<_Tp> <=3D 64)
+	      {
+		__m128i __a =3D {};
+		__builtin_memcpy(&__a, __mem, 16);
+		const auto __b =3D _mm512_cvtepi8_epi32(__a);
+		__builtin_memcpy(&__a, __mem + 16, 16);
+		const auto __c =3D _mm512_cvtepi8_epi32(__a);
+		if constexpr (_S_size<_Tp> <=3D 48)
+		  {
+		    __builtin_memcpy(&__a, __mem + 32, _S_size<_Tp> - 32);
+		    const auto __d =3D _mm512_cvtepi8_epi32(__a);
+		    return _mm512_test_epi32_mask(__b, __b)
+			   | (_mm512_test_epi32_mask(__c, __c) << 16)
+			   | (_ULLong(_mm512_test_epi32_mask(__d, __d)) << 32);
+		  }
+		else
+		  {
+		    __builtin_memcpy(&__a, __mem + 16, 16);
+		    const auto __d =3D _mm512_cvtepi8_epi32(__a);
+		    __builtin_memcpy(&__a, __mem + 32, _S_size<_Tp> - 48);
+		    const auto __e =3D _mm512_cvtepi8_epi32(__a);
+		    return _mm512_test_epi32_mask(__b, __b)
+			   | (_mm512_test_epi32_mask(__c, __c) << 16)
+			   | (_ULLong(_mm512_test_epi32_mask(__d, __d)) << 32)
+			   | (_ULLong(_mm512_test_epi32_mask(__e, __e)) << 48);
+		  }
+	      }
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (sizeof(_Tp) =3D=3D 8 && _S_size<_Tp> =3D=3D 2)
+	  return __vector_bitcast<_Tp>(
+	    __vector_type16_t<int>{-int(__mem[0]), -int(__mem[0]),
+				   -int(__mem[1]), -int(__mem[1])});
+	else if constexpr (sizeof(_Tp) =3D=3D 8 && _S_size<_Tp> <=3D 4 && __have_=
avx)
+	  {
+	    int __bool4 =3D 0;
+	    __builtin_memcpy(&__bool4, __mem, _S_size<_Tp>);
+	    const auto __k =3D __to_intrin(
+	      (__vector_broadcast<4>(__bool4)
+	       & __make_vector<int>(0x1, 0x100, 0x10000,
+				    _S_size<_Tp> =3D=3D 4 ? 0x1000000 : 0))
+	      !=3D 0);
+	    return __vector_bitcast<_Tp>(
+	      __concat(_mm_unpacklo_epi32(__k, __k),
+		       _mm_unpackhi_epi32(__k, __k)));
+	  }
+	else if constexpr (sizeof(_Tp) =3D=3D 4 && _S_size<_Tp> <=3D 4)
+	  {
+	    int __bools =3D 0;
+	    __builtin_memcpy(&__bools, __mem, _S_size<_Tp>);
+	    if constexpr (__have_sse2)
+	      {
+		__m128i __k =3D _mm_cvtsi32_si128(__bools);
+		__k =3D _mm_cmpgt_epi16(_mm_unpacklo_epi8(__k, __k), __m128i());
+		return __vector_bitcast<_Tp, _S_size<_Tp>>(
+		  _mm_unpacklo_epi16(__k, __k));
+	      }
+	    else
+	      {
+		__m128 __k =3D _mm_cvtpi8_ps(_mm_cvtsi32_si64(__bools));
+		_mm_empty();
+		return __vector_bitcast<_Tp, _S_size<_Tp>>(
+		  _mm_cmpgt_ps(__k, __m128()));
+	      }
+	  }
+	else if constexpr (sizeof(_Tp) =3D=3D 4 && _S_size<_Tp> <=3D 8)
+	  {
+	    __m128i __k =3D {};
+	    __builtin_memcpy(&__k, __mem, _S_size<_Tp>);
+	    __k =3D _mm_cmpgt_epi16(_mm_unpacklo_epi8(__k, __k), __m128i());
+	    return __vector_bitcast<_Tp>(
+	      __concat(_mm_unpacklo_epi16(__k, __k),
+		       _mm_unpackhi_epi16(__k, __k)));
+	  }
+	else if constexpr (sizeof(_Tp) =3D=3D 2 && _S_size<_Tp> <=3D 16)
+	  {
+	    __m128i __k =3D {};
+	    __builtin_memcpy(&__k, __mem, _S_size<_Tp>);
+	    __k =3D _mm_cmpgt_epi8(__k, __m128i());
+	    if constexpr (_S_size<_Tp> <=3D 8)
+	      return __vector_bitcast<_Tp, _S_size<_Tp>>(
+		_mm_unpacklo_epi8(__k, __k));
+	    else
+	      return __concat(_mm_unpacklo_epi8(__k, __k),
+			      _mm_unpackhi_epi8(__k, __k));
+	  }
+	else
+	  return _Base::template _S_load<_Tp>(__mem);
+      }
+
+    // }}}
+    // _S_from_bitmask{{{
+    template <size_t _Np, typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static _MaskMember<_Tp>
+      _S_from_bitmask(_SanitizedBitMask<_Np> __bits, _TypeTag<_Tp>)
+      {
+	static_assert(is_same_v<_Tp, __int_for_sizeof_t<_Tp>>);
+	if constexpr (__is_avx512_abi<_Abi>())
+	  return __bits._M_to_bits();
+	else
+	  return _S_to_maskvector<_Tp, _S_size<_Tp>>(__bits);
+      }
+
+    // }}}
+    // _S_masked_load {{{2
+    template <typename _Tp, size_t _Np>
+      static inline _SimdWrapper<_Tp, _Np>
+      _S_masked_load(_SimdWrapper<_Tp, _Np> __merge,
+		     _SimdWrapper<_Tp, _Np> __mask, const bool* __mem) noexcept
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    if constexpr (__have_avx512bw_vl)
+	      {
+		if constexpr (_Np <=3D 16)
+		  {
+		    const auto __a
+		      =3D _mm_mask_loadu_epi8(__m128i(), __mask, __mem);
+		    return (__merge & ~__mask) | _mm_test_epi8_mask(__a, __a);
+		  }
+		else if constexpr (_Np <=3D 32)
+		  {
+		    const auto __a
+		      =3D _mm256_mask_loadu_epi8(__m256i(), __mask, __mem);
+		    return (__merge & ~__mask)
+			   | _mm256_test_epi8_mask(__a, __a);
+		  }
+		else if constexpr (_Np <=3D 64)
+		  {
+		    const auto __a
+		      =3D _mm512_mask_loadu_epi8(__m512i(), __mask, __mem);
+		    return (__merge & ~__mask)
+			   | _mm512_test_epi8_mask(__a, __a);
+		  }
+		else
+		  __assert_unreachable<_Tp>();
+	      }
+	    else
+	      {
+		_BitOps::_S_bit_iteration(__mask, [&](auto __i) {
+		  __merge._M_set(__i, __mem[__i]);
+		});
+		return __merge;
+	      }
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 32 && sizeof(_Tp) =3D=
=3D 1)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge =3D _mm256_mask_sub_epi8(__to_intrin(__merge), __k, __m256i(),
+					   _mm256_mask_loadu_epi8(__m256i(),
+								  __k, __mem));
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 16 && sizeof(_Tp) =3D=
=3D 1)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge
+	      =3D _mm_mask_sub_epi8(__vector_bitcast<_LLong>(__merge), __k,
+				  __m128i(),
+				  _mm_mask_loadu_epi8(__m128i(), __k, __mem));
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 16 && sizeof(_Tp) =3D=
=3D 2)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge =3D _mm256_mask_sub_epi16(
+	      __vector_bitcast<_LLong>(__merge), __k, __m256i(),
+	      _mm256_cvtepi8_epi16(_mm_mask_loadu_epi8(__m128i(), __k, __mem)));
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 8 && sizeof(_Tp) =3D=
=3D 2)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge =3D _mm_mask_sub_epi16(
+	      __vector_bitcast<_LLong>(__merge), __k, __m128i(),
+	      _mm_cvtepi8_epi16(_mm_mask_loadu_epi8(__m128i(), __k, __mem)));
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 8 && sizeof(_Tp) =3D=
=3D 4)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge =3D __vector_bitcast<_Tp>(_mm256_mask_sub_epi32(
+	      __vector_bitcast<_LLong>(__merge), __k, __m256i(),
+	      _mm256_cvtepi8_epi32(
+		_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 4 && sizeof(_Tp) =3D=
=3D 4)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge =3D __vector_bitcast<_Tp>(_mm_mask_sub_epi32(
+	      __vector_bitcast<_LLong>(__merge), __k, __m128i(),
+	      _mm_cvtepi8_epi32(_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 4 && sizeof(_Tp) =3D=
=3D 8)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge =3D __vector_bitcast<_Tp>(_mm256_mask_sub_epi64(
+	      __vector_bitcast<_LLong>(__merge), __k, __m256i(),
+	      _mm256_cvtepi8_epi64(
+		_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
+	  }
+	else if constexpr (__have_avx512bw_vl && _Np =3D=3D 2 && sizeof(_Tp) =3D=
=3D 8)
+	  {
+	    const auto __k =3D _S_to_bits(__mask)._M_to_bits();
+	    __merge =3D __vector_bitcast<_Tp>(_mm_mask_sub_epi64(
+	      __vector_bitcast<_LLong>(__merge), __k, __m128i(),
+	      _mm_cvtepi8_epi64(_mm_mask_loadu_epi8(__m128i(), __k, __mem))));
+	  }
+	else
+	  return _Base::_S_masked_load(__merge, __mask, __mem);
+	return __merge;
+      }
+
+    // _S_store {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void _S_store(_SimdWrapper<_Tp, _Np> =
__v,
+						   bool* __mem) noexcept
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    if constexpr (__have_avx512bw_vl)
+	      _CommonImplX86::_S_store<_Np>(
+		__vector_bitcast<char>([](auto __data) {
+		  if constexpr (_Np <=3D 16)
+		    return _mm_maskz_set1_epi8(__data, 1);
+		  else if constexpr (_Np <=3D 32)
+		    return _mm256_maskz_set1_epi8(__data, 1);
+		  else
+		    return _mm512_maskz_set1_epi8(__data, 1);
+		}(__v._M_data)),
+		__mem);
+	    else if constexpr (_Np <=3D 8)
+	      _CommonImplX86::_S_store<_Np>(
+		__vector_bitcast<char>(
+#if defined __x86_64__
+		  __make_wrapper<_ULLong>(
+		    _pdep_u64(__v._M_data, 0x0101010101010101ULL), 0ull)
+#else
+		  __make_wrapper<_UInt>(_pdep_u32(__v._M_data, 0x01010101U),
+					_pdep_u32(__v._M_data >> 4,
+						  0x01010101U))
+#endif
+		    ),
+		__mem);
+	    else if constexpr (_Np <=3D 16)
+	      _mm512_mask_cvtepi32_storeu_epi8(
+		__mem, 0xffffu >> (16 - _Np),
+		_mm512_maskz_set1_epi32(__v._M_data, 1));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else if constexpr (__is_sse_abi<_Abi>()) //{{{
+	  {
+	    if constexpr (_Np =3D=3D 2 && sizeof(_Tp) =3D=3D 8)
+	      {
+		const auto __k =3D __vector_bitcast<int>(__v);
+		__mem[0] =3D -__k[1];
+		__mem[1] =3D -__k[3];
+	      }
+	    else if constexpr (_Np <=3D 4 && sizeof(_Tp) =3D=3D 4)
+	      {
+		if constexpr (__have_sse2)
+		  {
+		    const unsigned __bool4
+		      =3D __vector_bitcast<_UInt>(_mm_packs_epi16(
+			  _mm_packs_epi32(__intrin_bitcast<__m128i>(
+					    __to_intrin(__v)),
+					  __m128i()),
+			  __m128i()))[0]
+			& 0x01010101u;
+		    __builtin_memcpy(__mem, &__bool4, _Np);
+		  }
+		else if constexpr (__have_mmx)
+		  {
+		    const __m64 __k =3D _mm_cvtps_pi8(
+		      __and(__to_intrin(__v), _mm_set1_ps(1.f)));
+		    __builtin_memcpy(__mem, &__k, _Np);
+		    _mm_empty();
+		  }
+		else
+		  return _Base::_S_store(__v, __mem);
+	      }
+	    else if constexpr (_Np <=3D 8 && sizeof(_Tp) =3D=3D 2)
+	      {
+		_CommonImplX86::_S_store<_Np>(
+		  __vector_bitcast<char>(_mm_packs_epi16(
+		    __to_intrin(__vector_bitcast<_UShort>(__v) >> 15),
+		    __m128i())),
+		  __mem);
+	      }
+	    else if constexpr (_Np <=3D 16 && sizeof(_Tp) =3D=3D 1)
+	      _CommonImplX86::_S_store<_Np>(__v._M_data & 1, __mem);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }                                      // }}}
+	else if constexpr (__is_avx_abi<_Abi>()) // {{{
+	  {
+	    if constexpr (_Np <=3D 4 && sizeof(_Tp) =3D=3D 8)
+	      {
+		auto __k =3D __intrin_bitcast<__m256i>(__to_intrin(__v));
+		int __bool4;
+		if constexpr (__have_avx2)
+		  __bool4 =3D _mm256_movemask_epi8(__k);
+		else
+		  __bool4 =3D (_mm_movemask_epi8(__lo128(__k))
+			     | (_mm_movemask_epi8(__hi128(__k)) << 16));
+		__bool4 &=3D 0x01010101;
+		__builtin_memcpy(__mem, &__bool4, _Np);
+	      }
+	    else if constexpr (_Np <=3D 8 && sizeof(_Tp) =3D=3D 4)
+	      {
+		const auto __k =3D __intrin_bitcast<__m256i>(__to_intrin(__v));
+		const auto __k2
+		  =3D _mm_srli_epi16(_mm_packs_epi16(__lo128(__k), __hi128(__k)),
+				   15);
+		const auto __k3
+		  =3D __vector_bitcast<char>(_mm_packs_epi16(__k2, __m128i()));
+		_CommonImplX86::_S_store<_Np>(__k3, __mem);
+	      }
+	    else if constexpr (_Np <=3D 16 && sizeof(_Tp) =3D=3D 2)
+	      {
+		if constexpr (__have_avx2)
+		  {
+		    const auto __x =3D _mm256_srli_epi16(__to_intrin(__v), 15);
+		    const auto __bools =3D __vector_bitcast<char>(
+		      _mm_packs_epi16(__lo128(__x), __hi128(__x)));
+		    _CommonImplX86::_S_store<_Np>(__bools, __mem);
+		  }
+		else
+		  {
+		    const auto __bools
+		      =3D 1
+			& __vector_bitcast<_UChar>(
+			  _mm_packs_epi16(__lo128(__to_intrin(__v)),
+					  __hi128(__to_intrin(__v))));
+		    _CommonImplX86::_S_store<_Np>(__bools, __mem);
+		  }
+	      }
+	    else if constexpr (_Np <=3D 32 && sizeof(_Tp) =3D=3D 1)
+	      _CommonImplX86::_S_store<_Np>(1 & __v._M_data, __mem);
+	    else
+	      __assert_unreachable<_Tp>();
+	  } // }}}
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    // _S_masked_store {{{2
+    template <typename _Tp, size_t _Np>
+      static inline void
+      _S_masked_store(const _SimdWrapper<_Tp, _Np> __v, bool* __mem,
+		      const _SimdWrapper<_Tp, _Np> __k) noexcept
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    static_assert(is_same_v<_Tp, bool>);
+	    if constexpr (_Np <=3D 16 && __have_avx512bw_vl)
+	      _mm_mask_storeu_epi8(__mem, __k, _mm_maskz_set1_epi8(__v, 1));
+	    else if constexpr (_Np <=3D 16)
+	      _mm512_mask_cvtepi32_storeu_epi8(__mem, __k,
+					       _mm512_maskz_set1_epi32(__v, 1));
+	    else if constexpr (_Np <=3D 32 && __have_avx512bw_vl)
+	      _mm256_mask_storeu_epi8(__mem, __k,
+				      _mm256_maskz_set1_epi8(__v, 1));
+	    else if constexpr (_Np <=3D 32 && __have_avx512bw)
+	      _mm256_mask_storeu_epi8(__mem, __k,
+				      __lo256(_mm512_maskz_set1_epi8(__v, 1)));
+	    else if constexpr (_Np <=3D 64 && __have_avx512bw)
+	      _mm512_mask_storeu_epi8(__mem, __k,
+				      _mm512_maskz_set1_epi8(__v, 1));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  _Base::_S_masked_store(__v, __mem, __k);
+      }
+
+    // logical and bitwise operators {{{2
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_logical_and(const _SimdWrapper<_Tp, _Np>& __x,
+		     const _SimdWrapper<_Tp, _Np>& __y)
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  {
+	    if constexpr (__have_avx512dq && _Np <=3D 8)
+	      return _kand_mask8(__x._M_data, __y._M_data);
+	    else if constexpr (_Np <=3D 16)
+	      return _kand_mask16(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 32)
+	      return _kand_mask32(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 64)
+	      return _kand_mask64(__x._M_data, __y._M_data);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_logical_and(__x, __y);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_logical_or(const _SimdWrapper<_Tp, _Np>& __x,
+		    const _SimdWrapper<_Tp, _Np>& __y)
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  {
+	    if constexpr (__have_avx512dq && _Np <=3D 8)
+	      return _kor_mask8(__x._M_data, __y._M_data);
+	    else if constexpr (_Np <=3D 16)
+	      return _kor_mask16(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 32)
+	      return _kor_mask32(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 64)
+	      return _kor_mask64(__x._M_data, __y._M_data);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_logical_or(__x, __y);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_not(const _SimdWrapper<_Tp, _Np>& __x)
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  {
+	    if constexpr (__have_avx512dq && _Np <=3D 8)
+	      return _kandn_mask8(__x._M_data,
+				  _Abi::template __implicit_mask_n<_Np>());
+	    else if constexpr (_Np <=3D 16)
+	      return _kandn_mask16(__x._M_data,
+				   _Abi::template __implicit_mask_n<_Np>());
+	    else if constexpr (__have_avx512bw && _Np <=3D 32)
+	      return _kandn_mask32(__x._M_data,
+				   _Abi::template __implicit_mask_n<_Np>());
+	    else if constexpr (__have_avx512bw && _Np <=3D 64)
+	      return _kandn_mask64(__x._M_data,
+				   _Abi::template __implicit_mask_n<_Np>());
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_bit_not(__x);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_and(const _SimdWrapper<_Tp, _Np>& __x,
+		 const _SimdWrapper<_Tp, _Np>& __y)
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  {
+	    if constexpr (__have_avx512dq && _Np <=3D 8)
+	      return _kand_mask8(__x._M_data, __y._M_data);
+	    else if constexpr (_Np <=3D 16)
+	      return _kand_mask16(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 32)
+	      return _kand_mask32(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 64)
+	      return _kand_mask64(__x._M_data, __y._M_data);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_bit_and(__x, __y);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_or(const _SimdWrapper<_Tp, _Np>& __x,
+		const _SimdWrapper<_Tp, _Np>& __y)
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  {
+	    if constexpr (__have_avx512dq && _Np <=3D 8)
+	      return _kor_mask8(__x._M_data, __y._M_data);
+	    else if constexpr (_Np <=3D 16)
+	      return _kor_mask16(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 32)
+	      return _kor_mask32(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 64)
+	      return _kor_mask64(__x._M_data, __y._M_data);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_bit_or(__x, __y);
+      }
+
+    template <typename _Tp, size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static constexpr _SimdWrapper<_Tp, _Np>
+      _S_bit_xor(const _SimdWrapper<_Tp, _Np>& __x,
+		 const _SimdWrapper<_Tp, _Np>& __y)
+      {
+	if constexpr (is_same_v<_Tp, bool>)
+	  {
+	    if constexpr (__have_avx512dq && _Np <=3D 8)
+	      return _kxor_mask8(__x._M_data, __y._M_data);
+	    else if constexpr (_Np <=3D 16)
+	      return _kxor_mask16(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 32)
+	      return _kxor_mask32(__x._M_data, __y._M_data);
+	    else if constexpr (__have_avx512bw && _Np <=3D 64)
+	      return _kxor_mask64(__x._M_data, __y._M_data);
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return _Base::_S_bit_xor(__x, __y);
+      }
+
+    //}}}2
+    // _S_masked_assign{{{
+    template <size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(_SimdWrapper<bool, _Np> __k,
+		       _SimdWrapper<bool, _Np>& __lhs,
+		       _SimdWrapper<bool, _Np> __rhs)
+      {
+	__lhs._M_data
+	  =3D (~__k._M_data & __lhs._M_data) | (__k._M_data & __rhs._M_data);
+      }
+
+    template <size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static void
+      _S_masked_assign(_SimdWrapper<bool, _Np> __k,
+		       _SimdWrapper<bool, _Np>& __lhs, bool __rhs)
+      {
+	if (__rhs)
+	  __lhs._M_data =3D __k._M_data | __lhs._M_data;
+	else
+	  __lhs._M_data =3D ~__k._M_data & __lhs._M_data;
+      }
+
+    using _MaskImplBuiltin<_Abi>::_S_masked_assign;
+
+    //}}}
+    // _S_all_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_all_of(simd_mask<_Tp, _Abi> _=
_k)
+      {
+	if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
+	  {
+	    constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	    using _TI =3D __intrinsic_type_t<_Tp, _Np>;
+	    const _TI __a =3D reinterpret_cast<_TI>(__to_intrin(__data(__k)));
+	    if constexpr (__have_sse4_1)
+	      {
+		_GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
+		  =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+		return 0 !=3D __testc(__a, __b);
+	      }
+	    else if constexpr (is_same_v<_Tp, float>)
+	      return (_mm_movemask_ps(__a) & ((1 << _Np) - 1))
+		     =3D=3D (1 << _Np) - 1;
+	    else if constexpr (is_same_v<_Tp, double>)
+	      return (_mm_movemask_pd(__a) & ((1 << _Np) - 1))
+		     =3D=3D (1 << _Np) - 1;
+	    else
+	      return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) - 1))
+		     =3D=3D (1 << (_Np * sizeof(_Tp))) - 1;
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    constexpr auto _Mask =3D _Abi::template _S_implicit_mask<_Tp>();
+	    const auto __kk =3D __k._M_data._M_data;
+	    if constexpr (sizeof(__kk) =3D=3D 1)
+	      {
+		if constexpr (__have_avx512dq)
+		  return _kortestc_mask8_u8(__kk, _Mask =3D=3D 0xff
+						    ? __kk
+						    : __mmask8(~_Mask));
+		else
+		  return _kortestc_mask16_u8(__kk, __mmask16(~_Mask));
+	      }
+	    else if constexpr (sizeof(__kk) =3D=3D 2)
+	      return _kortestc_mask16_u8(__kk, _Mask =3D=3D 0xffff
+						 ? __kk
+						 : __mmask16(~_Mask));
+	    else if constexpr (sizeof(__kk) =3D=3D 4 && __have_avx512bw)
+	      return _kortestc_mask32_u8(__kk, _Mask =3D=3D 0xffffffffU
+						 ? __kk
+						 : __mmask32(~_Mask));
+	    else if constexpr (sizeof(__kk) =3D=3D 8 && __have_avx512bw)
+	      return _kortestc_mask64_u8(__kk, _Mask =3D=3D 0xffffffffffffffffULL
+						 ? __kk
+						 : __mmask64(~_Mask));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+      }
+
+    // }}}
+    // _S_any_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_any_of(simd_mask<_Tp, _Abi> _=
_k)
+      {
+	if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
+	  {
+	    constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	    using _TI =3D __intrinsic_type_t<_Tp, _Np>;
+	    const _TI __a =3D reinterpret_cast<_TI>(__to_intrin(__data(__k)));
+	    if constexpr (__have_sse4_1)
+	      {
+		if constexpr (_Abi::template _S_is_partial<
+				_Tp> || sizeof(__k) < 16)
+		  {
+		    _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
+		      =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+		    return 0 =3D=3D __testz(__a, __b);
+		  }
+		else
+		  return 0 =3D=3D __testz(__a, __a);
+	      }
+	    else if constexpr (is_same_v<_Tp, float>)
+	      return (_mm_movemask_ps(__a) & ((1 << _Np) - 1)) !=3D 0;
+	    else if constexpr (is_same_v<_Tp, double>)
+	      return (_mm_movemask_pd(__a) & ((1 << _Np) - 1)) !=3D 0;
+	    else
+	      return (_mm_movemask_epi8(__a) & ((1 << (_Np * sizeof(_Tp))) - 1))
+		     !=3D 0;
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  return (__k._M_data._M_data & _Abi::template _S_implicit_mask<_Tp>())
+		 !=3D 0;
+      }
+
+    // }}}
+    // _S_none_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_none_of(simd_mask<_Tp, _Abi> =
__k)
+      {
+	if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
+	  {
+	    constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	    using _TI =3D __intrinsic_type_t<_Tp, _Np>;
+	    const _TI __a =3D reinterpret_cast<_TI>(__to_intrin(__data(__k)));
+	    if constexpr (__have_sse4_1)
+	      {
+		if constexpr (_Abi::template _S_is_partial<
+				_Tp> || sizeof(__k) < 16)
+		  {
+		    _GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
+		      =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+		    return 0 !=3D __testz(__a, __b);
+		  }
+		else
+		  return 0 !=3D __testz(__a, __a);
+	      }
+	    else if constexpr (is_same_v<_Tp, float>)
+	      return (__movemask(__a) & ((1 << _Np) - 1)) =3D=3D 0;
+	    else if constexpr (is_same_v<_Tp, double>)
+	      return (__movemask(__a) & ((1 << _Np) - 1)) =3D=3D 0;
+	    else
+	      return (__movemask(__a) & int((1ull << (_Np * sizeof(_Tp))) - 1))
+		     =3D=3D 0;
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  return (__k._M_data._M_data & _Abi::template _S_implicit_mask<_Tp>())
+		 =3D=3D 0;
+      }
+
+    // }}}
+    // _S_some_of {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static bool _S_some_of(simd_mask<_Tp, _Abi> =
__k)
+      {
+	if constexpr (__is_sse_abi<_Abi>() || __is_avx_abi<_Abi>())
+	  {
+	    constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	    using _TI =3D __intrinsic_type_t<_Tp, _Np>;
+	    const _TI __a =3D reinterpret_cast<_TI>(__to_intrin(__data(__k)));
+	    if constexpr (__have_sse4_1)
+	      {
+		_GLIBCXX_SIMD_USE_CONSTEXPR _TI __b
+		  =3D _Abi::template _S_implicit_mask_intrin<_Tp>();
+		return 0 !=3D __testnzc(__a, __b);
+	      }
+	    else if constexpr (is_same_v<_Tp, float>)
+	      {
+		constexpr int __allbits =3D (1 << _Np) - 1;
+		const auto __tmp =3D _mm_movemask_ps(__a) & __allbits;
+		return __tmp > 0 && __tmp < __allbits;
+	      }
+	    else if constexpr (is_same_v<_Tp, double>)
+	      {
+		constexpr int __allbits =3D (1 << _Np) - 1;
+		const auto __tmp =3D _mm_movemask_pd(__a) & __allbits;
+		return __tmp > 0 && __tmp < __allbits;
+	      }
+	    else
+	      {
+		constexpr int __allbits =3D (1 << (_Np * sizeof(_Tp))) - 1;
+		const auto __tmp =3D _mm_movemask_epi8(__a) & __allbits;
+		return __tmp > 0 && __tmp < __allbits;
+	      }
+	  }
+	else if constexpr (__is_avx512_abi<_Abi>())
+	  return _S_any_of(__k) && !_S_all_of(__k);
+	else
+	  __assert_unreachable<_Tp>();
+      }
+
+    // }}}
+    // _S_popcount {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int _S_popcount(simd_mask<_Tp, _Abi> =
__k)
+      {
+	constexpr size_t _Np =3D simd_size_v<_Tp, _Abi>;
+	const auto __kk =3D _Abi::_S_masked(__k._M_data)._M_data;
+	if constexpr (__is_avx512_abi<_Abi>())
+	  {
+	    if constexpr (_Np > 32)
+	      return __builtin_popcountll(__kk);
+	    else
+	      return __builtin_popcount(__kk);
+	  }
+	else
+	  {
+	    if constexpr (__have_popcnt)
+	      {
+		int __bits
+		  =3D __movemask(__to_intrin(__vector_bitcast<_Tp>(__kk)));
+		const int __count =3D __builtin_popcount(__bits);
+		return is_integral_v<_Tp> ? __count / sizeof(_Tp) : __count;
+	      }
+	    else if constexpr (_Np =3D=3D 2 && sizeof(_Tp) =3D=3D 8)
+	      {
+		const int mask =3D _mm_movemask_pd(__auto_bitcast(__kk));
+		return mask - (mask >> 1);
+	      }
+	    else if constexpr (_Np <=3D 4 && sizeof(_Tp) =3D=3D 8)
+	      {
+		auto __x =3D -(__lo128(__kk) + __hi128(__kk));
+		return __x[0] + __x[1];
+	      }
+	    else if constexpr (_Np <=3D 4 && sizeof(_Tp) =3D=3D 4)
+	      {
+		if constexpr (__have_sse2)
+		  {
+		    __m128i __x =3D __intrin_bitcast<__m128i>(__to_intrin(__kk));
+		    __x =3D _mm_add_epi32(
+		      __x, _mm_shuffle_epi32(__x, _MM_SHUFFLE(0, 1, 2, 3)));
+		    __x =3D _mm_add_epi32(
+		      __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(1, 0, 3, 2)));
+		    return -_mm_cvtsi128_si32(__x);
+		  }
+		else
+		  return __builtin_popcount(
+		    _mm_movemask_ps(__auto_bitcast(__kk)));
+	      }
+	    else if constexpr (_Np <=3D 8 && sizeof(_Tp) =3D=3D 2)
+	      {
+		auto __x =3D __to_intrin(__kk);
+		__x =3D _mm_add_epi16(__x,
+				    _mm_shuffle_epi32(__x,
+						      _MM_SHUFFLE(0, 1, 2, 3)));
+		__x =3D _mm_add_epi16(
+		  __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(0, 1, 2, 3)));
+		__x =3D _mm_add_epi16(
+		  __x, _mm_shufflelo_epi16(__x, _MM_SHUFFLE(2, 3, 0, 1)));
+		return -short(_mm_extract_epi16(__x, 0));
+	      }
+	    else if constexpr (_Np <=3D 16 && sizeof(_Tp) =3D=3D 1)
+	      {
+		auto __x =3D __to_intrin(__kk);
+		__x =3D _mm_add_epi8(__x,
+				   _mm_shuffle_epi32(__x,
+						     _MM_SHUFFLE(0, 1, 2, 3)));
+		__x =3D _mm_add_epi8(__x,
+				   _mm_shufflelo_epi16(__x, _MM_SHUFFLE(0, 1, 2,
+									3)));
+		__x =3D _mm_add_epi8(__x,
+				   _mm_shufflelo_epi16(__x, _MM_SHUFFLE(2, 3, 0,
+									1)));
+		auto __y =3D -__vector_bitcast<_UChar>(__x);
+		if constexpr (__have_sse4_1)
+		  return __y[0] + __y[1];
+		else
+		  {
+		    unsigned __z =3D _mm_extract_epi16(__to_intrin(__y), 0);
+		    return (__z & 0xff) + (__z >> 8);
+		  }
+	      }
+	    else if constexpr (sizeof(__kk) =3D=3D 32)
+	      {
+		// The following works only as long as the implementations above
+		// use a summation
+		using _I =3D __int_for_sizeof_t<_Tp>;
+		const auto __as_int =3D __vector_bitcast<_I>(__kk);
+		_MaskImplX86<simd_abi::__sse>::_S_popcount(
+		  simd_mask<_I, simd_abi::__sse>(__private_init,
+						 __lo128(__as_int)
+						   + __hi128(__as_int)));
+	      }
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+      }
+
+    // }}}
+    // _S_find_first_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_first_set(simd_mask<_Tp, _Abi> __k)
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  return std::__countr_zero(__k._M_data._M_data);
+	else
+	  return _Base::_S_find_first_set(__k);
+      }
+
+    // }}}
+    // _S_find_last_set {{{
+    template <typename _Tp>
+      _GLIBCXX_SIMD_INTRINSIC static int
+      _S_find_last_set(simd_mask<_Tp, _Abi> __k)
+      {
+	if constexpr (__is_avx512_abi<_Abi>())
+	  return std::__bit_width(__k._M_data._M_data) - 1;
+	else
+	  return _Base::_S_find_last_set(__k);
+      }
+
+    // }}}
+  };
+
+// }}}
+
+_GLIBCXX_SIMD_END_NAMESPACE
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_X86_H_
+
+// vim: foldmethod=3Dmarker sw=3D2 noet ts=3D8 sts=3D2 tw=3D80
diff --git a/libstdc++-v3/include/experimental/bits/simd_x86_conversions.h =
b/libstdc++-v3/include/experimental/bits/simd_x86_conversions.h
new file mode 100644
index 00000000000..09f898b7ec9
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/bits/simd_x86_conversions.h
@@ -0,0 +1,2029 @@
+// x86 specific conversion optimizations -*- C++ -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_X86_CONVERSIONS_H
+#define _GLIBCXX_EXPERIMENTAL_SIMD_X86_CONVERSIONS_H
+
+#if __cplusplus >=3D 201703L
+
+// work around PR85827
+// 1-arg __convert_x86 {{{1
+template <typename _To, typename _V, typename _Traits>
+  _GLIBCXX_SIMD_INTRINSIC _To
+  __convert_x86(_V __v)
+  {
+    static_assert(__is_vector_type_v<_V>);
+    using _Tp =3D typename _Traits::value_type;
+    constexpr size_t _Np =3D _Traits::_S_full_size;
+    [[maybe_unused]] const auto __intrin =3D __to_intrin(__v);
+    using _Up =3D typename _VectorTraits<_To>::value_type;
+    constexpr size_t _M =3D _VectorTraits<_To>::_S_full_size;
+
+    // [xyz]_to_[xyz] {{{2
+    [[maybe_unused]] constexpr bool __x_to_x
+      =3D sizeof(__v) <=3D 16 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __x_to_y
+      =3D sizeof(__v) <=3D 16 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __x_to_z
+      =3D sizeof(__v) <=3D 16 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __y_to_x
+      =3D sizeof(__v) =3D=3D 32 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __y_to_y
+      =3D sizeof(__v) =3D=3D 32 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __y_to_z
+      =3D sizeof(__v) =3D=3D 32 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __z_to_x
+      =3D sizeof(__v) =3D=3D 64 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __z_to_y
+      =3D sizeof(__v) =3D=3D 64 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __z_to_z
+      =3D sizeof(__v) =3D=3D 64 && sizeof(_To) =3D=3D 64;
+
+    // iX_to_iX {{{2
+    [[maybe_unused]] constexpr bool __i_to_i
+      =3D is_integral_v<_Up> && is_integral_v<_Tp>;
+    [[maybe_unused]] constexpr bool __i8_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i8_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __i8_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i16_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i16_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __i16_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i32_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i32_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i32_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i64_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i64_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i64_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 4;
+
+    // [fsu]X_to_[fsu]X {{{2
+    // ibw =3D integral && byte or word, i.e. char and short with any sign=
edness
+    [[maybe_unused]] constexpr bool __s64_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s32_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s16_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s8_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u64_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u32_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u16_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u8_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s64_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __s32_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u64_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u32_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_s64
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_s32
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_u64
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_u32
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f64_to_s64
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_s32
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_u64
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_u32
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __ibw_to_f32
+      =3D is_integral_v<_Tp> && sizeof(_Tp) <=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __ibw_to_f64
+      =3D is_integral_v<_Tp> && sizeof(_Tp) <=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_ibw
+      =3D is_integral_v<_Up> && sizeof(_Up) <=3D 2
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f64_to_ibw
+      =3D is_integral_v<_Up> && sizeof(_Up) <=3D 2
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_f64
+      =3D is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_f32
+      =3D is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+
+    if constexpr (__i_to_i && __y_to_x && !__have_avx2) //{{{2
+      return __convert_x86<_To>(__lo128(__v), __hi128(__v));
+    else if constexpr (__i_to_i && __x_to_y && !__have_avx2) //{{{2
+      return __concat(__convert_x86<__vector_type_t<_Up, _M / 2>>(__v),
+		      __convert_x86<__vector_type_t<_Up, _M / 2>>(
+			__extract_part<1, _Np / _M * 2>(__v)));
+    else if constexpr (__i_to_i) //{{{2
+      {
+	static_assert(__x_to_x || __have_avx2,
+		      "integral conversions with ymm registers require AVX2");
+	static_assert(__have_avx512bw
+			|| ((sizeof(_Tp) >=3D 4 || sizeof(__v) < 64)
+			    && (sizeof(_Up) >=3D 4 || sizeof(_To) < 64)),
+		      "8/16-bit integers in zmm registers require AVX512BW");
+	static_assert((sizeof(__v) < 64 && sizeof(_To) < 64) || __have_avx512f,
+		      "integral conversions with ymm registers require AVX2");
+      }
+    if constexpr (is_floating_point_v<_Tp> =3D=3D is_floating_point_v<_Up>=
 && //{{{2
+		  sizeof(_Tp) =3D=3D sizeof(_Up))
+      {
+	// conversion uses simple bit reinterpretation (or no conversion at all)
+	if constexpr (_Np >=3D _M)
+	  return __intrin_bitcast<_To>(__v);
+	else
+	  return __zero_extend(__vector_bitcast<_Up>(__v));
+      }
+    else if constexpr (_Np < _M && sizeof(_To) > 16) //{{{2
+      // zero extend (eg. xmm -> ymm)
+      return __zero_extend(
+	__convert_x86<__vector_type_t<
+	  _Up, (16 / sizeof(_Up) > _Np) ? 16 / sizeof(_Up) : _Np>>(__v));
+    else if constexpr (_Np > _M && sizeof(__v) > 16) //{{{2
+      // partial input (eg. ymm -> xmm)
+      return __convert_x86<_To>(__extract_part<0, _Np / _M>(__v));
+    else if constexpr (__i64_to_i32) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi64_epi32(__intrin));
+	else if constexpr (__x_to_x)
+	  return __auto_bitcast(
+	    _mm_shuffle_ps(__vector_bitcast<float>(__v), __m128(), 8));
+	else if constexpr (__y_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi64_epi32(__intrin));
+	else if constexpr (__y_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi64_epi32(__auto_bitcast(__v))));
+	else if constexpr (__y_to_x)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm256_permute4x64_epi64(_mm256_shuffle_epi32(__intrin, 8),
+					     0 + 4 * 2)));
+	else if constexpr (__z_to_y)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi64_epi32(__intrin));
+      }
+    else if constexpr (__i64_to_i16) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi64_epi16(__intrin));
+	else if constexpr (__x_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi64_epi16(__auto_bitcast(__v))));
+	else if constexpr (__x_to_x && __have_ssse3)
+	  {
+	    return __intrin_bitcast<_To>(
+	      _mm_shuffle_epi8(__intrin,
+			       _mm_setr_epi8(0, 1, 8, 9, -0x80, -0x80, -0x80,
+					     -0x80, -0x80, -0x80, -0x80, -0x80,
+					     -0x80, -0x80, -0x80, -0x80)));
+	    // fallback without SSSE3
+	  }
+	else if constexpr (__y_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi64_epi16(__intrin));
+	else if constexpr (__y_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi64_epi16(__auto_bitcast(__v))));
+	else if constexpr (__y_to_x)
+	  {
+	    const auto __a =3D _mm256_shuffle_epi8(
+	      __intrin,
+	      _mm256_setr_epi8(0, 1, 8, 9, -0x80, -0x80, -0x80, -0x80, -0x80,
+			       -0x80, -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
+			       -0x80, -0x80, -0x80, -0x80, 0, 1, 8, 9, -0x80,
+			       -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
+			       -0x80));
+	    return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
+	  }
+	else if constexpr (__z_to_x)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi64_epi16(__intrin));
+      }
+    else if constexpr (__i64_to_i8) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi64_epi8(__intrin));
+	else if constexpr (__x_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi64_epi8(__zero_extend(__intrin))));
+	else if constexpr (__y_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi64_epi8(__intrin));
+	else if constexpr (__y_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    _mm512_cvtepi64_epi8(__zero_extend(__intrin)));
+	else if constexpr (__z_to_x)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi64_epi8(__intrin));
+      }
+    else if constexpr (__i32_to_i64) //{{{2
+      {
+	if constexpr (__have_sse4_1 && __x_to_x)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm_cvtepi32_epi64(__intrin)
+					 : _mm_cvtepu32_epi64(__intrin));
+	else if constexpr (__x_to_x)
+	  {
+	    return __intrin_bitcast<_To>(
+	      _mm_unpacklo_epi32(__intrin, is_signed_v<_Tp>
+					     ? _mm_srai_epi32(__intrin, 31)
+					     : __m128i()));
+	  }
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm256_cvtepi32_epi64(__intrin)
+					 : _mm256_cvtepu32_epi64(__intrin));
+	else if constexpr (__y_to_z)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm512_cvtepi32_epi64(__intrin)
+					 : _mm512_cvtepu32_epi64(__intrin));
+      }
+    else if constexpr (__i32_to_i16) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi32_epi16(__intrin));
+	else if constexpr (__x_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi32_epi16(__auto_bitcast(__v))));
+	else if constexpr (__x_to_x && __have_ssse3)
+	  return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+	    __intrin, _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80,
+				    -0x80, -0x80, -0x80, -0x80, -0x80, -0x80)));
+	else if constexpr (__x_to_x)
+	  {
+	    auto __a =3D _mm_unpacklo_epi16(__intrin, __m128i()); // 0o.o 1o.o
+	    auto __b =3D _mm_unpackhi_epi16(__intrin, __m128i()); // 2o.o 3o.o
+	    auto __c =3D _mm_unpacklo_epi16(__a, __b);            // 02oo ..oo
+	    auto __d =3D _mm_unpackhi_epi16(__a, __b);            // 13oo ..oo
+	    return __intrin_bitcast<_To>(
+	      _mm_unpacklo_epi16(__c, __d)); // 0123 oooo
+	  }
+	else if constexpr (__y_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi32_epi16(__intrin));
+	else if constexpr (__y_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi32_epi16(__auto_bitcast(__v))));
+	else if constexpr (__y_to_x)
+	  {
+	    auto __a =3D _mm256_shuffle_epi8(
+	      __intrin,
+	      _mm256_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80, -0x80,
+			       -0x80, -0x80, -0x80, -0x80, -0x80, 0, 1, 4, 5, 8,
+			       9, 12, 13, -0x80, -0x80, -0x80, -0x80, -0x80,
+			       -0x80, -0x80, -0x80));
+	    return __intrin_bitcast<_To>(__lo128(
+	      _mm256_permute4x64_epi64(__a,
+				       0xf8))); // __a[0] __a[2] | __a[3] __a[3]
+	  }
+	else if constexpr (__z_to_y)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi32_epi16(__intrin));
+      }
+    else if constexpr (__i32_to_i8) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi32_epi8(__intrin));
+	else if constexpr (__x_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi32_epi8(__zero_extend(__intrin))));
+	else if constexpr (__x_to_x && __have_ssse3)
+	  {
+	    return __intrin_bitcast<_To>(
+	      _mm_shuffle_epi8(__intrin,
+			       _mm_setr_epi8(0, 4, 8, 12, -0x80, -0x80, -0x80,
+					     -0x80, -0x80, -0x80, -0x80, -0x80,
+					     -0x80, -0x80, -0x80, -0x80)));
+	  }
+	else if constexpr (__x_to_x)
+	  {
+	    const auto __a
+	      =3D _mm_unpacklo_epi8(__intrin, __intrin); // 0... .... 1... ....
+	    const auto __b
+	      =3D _mm_unpackhi_epi8(__intrin, __intrin);    // 2... .... 3... ....
+	    const auto __c =3D _mm_unpacklo_epi8(__a, __b); // 02.. .... .... ....
+	    const auto __d =3D _mm_unpackhi_epi8(__a, __b); // 13.. .... .... ....
+	    const auto __e =3D _mm_unpacklo_epi8(__c, __d); // 0123 .... .... ....
+	    return __intrin_bitcast<_To>(__e & _mm_cvtsi32_si128(-1));
+	  }
+	else if constexpr (__y_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi32_epi8(__intrin));
+	else if constexpr (__y_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    _mm512_cvtepi32_epi8(__zero_extend(__intrin)));
+	else if constexpr (__z_to_x)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi32_epi8(__intrin));
+      }
+    else if constexpr (__i16_to_i64) //{{{2
+      {
+	if constexpr (__x_to_x && __have_sse4_1)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm_cvtepi16_epi64(__intrin)
+					 : _mm_cvtepu16_epi64(__intrin));
+	else if constexpr (__x_to_x && is_signed_v<_Tp>)
+	  {
+	    auto __x =3D _mm_srai_epi16(__intrin, 15);
+	    auto __y =3D _mm_unpacklo_epi16(__intrin, __x);
+	    __x =3D _mm_unpacklo_epi16(__x, __x);
+	    return __intrin_bitcast<_To>(_mm_unpacklo_epi32(__y, __x));
+	  }
+	else if constexpr (__x_to_x)
+	  return __intrin_bitcast<_To>(
+	    _mm_unpacklo_epi32(_mm_unpacklo_epi16(__intrin, __m128i()),
+			       __m128i()));
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm256_cvtepi16_epi64(__intrin)
+					 : _mm256_cvtepu16_epi64(__intrin));
+	else if constexpr (__x_to_z)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm512_cvtepi16_epi64(__intrin)
+					 : _mm512_cvtepu16_epi64(__intrin));
+      }
+    else if constexpr (__i16_to_i32) //{{{2
+      {
+	if constexpr (__x_to_x && __have_sse4_1)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm_cvtepi16_epi32(__intrin)
+					 : _mm_cvtepu16_epi32(__intrin));
+	else if constexpr (__x_to_x && is_signed_v<_Tp>)
+	  return __intrin_bitcast<_To>(
+	    _mm_srai_epi32(_mm_unpacklo_epi16(__intrin, __intrin), 16));
+	else if constexpr (__x_to_x && is_unsigned_v<_Tp>)
+	  return __intrin_bitcast<_To>(_mm_unpacklo_epi16(__intrin, __m128i()));
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm256_cvtepi16_epi32(__intrin)
+					 : _mm256_cvtepu16_epi32(__intrin));
+	else if constexpr (__y_to_z)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm512_cvtepi16_epi32(__intrin)
+					 : _mm512_cvtepu16_epi32(__intrin));
+      }
+    else if constexpr (__i16_to_i8) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512bw_vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi16_epi8(__intrin));
+	else if constexpr (__x_to_x && __have_avx512bw)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepi16_epi8(__zero_extend(__intrin))));
+	else if constexpr (__x_to_x && __have_ssse3)
+	  return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+	    __intrin, _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, -0x80, -0x80,
+				    -0x80, -0x80, -0x80, -0x80, -0x80, -0x80)));
+	else if constexpr (__x_to_x)
+	  {
+	    auto __a
+	      =3D _mm_unpacklo_epi8(__intrin, __intrin); // 00.. 11.. 22.. 33..
+	    auto __b
+	      =3D _mm_unpackhi_epi8(__intrin, __intrin); // 44.. 55.. 66.. 77..
+	    auto __c =3D _mm_unpacklo_epi8(__a, __b);    // 0404 .... 1515 ....
+	    auto __d =3D _mm_unpackhi_epi8(__a, __b);    // 2626 .... 3737 ....
+	    auto __e =3D _mm_unpacklo_epi8(__c, __d);    // 0246 0246 .... ....
+	    auto __f =3D _mm_unpackhi_epi8(__c, __d);    // 1357 1357 .... ....
+	    return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__e, __f));
+	  }
+	else if constexpr (__y_to_x && __have_avx512bw_vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi16_epi8(__intrin));
+	else if constexpr (__y_to_x && __have_avx512bw)
+	  return __intrin_bitcast<_To>(
+	    __lo256(_mm512_cvtepi16_epi8(__zero_extend(__intrin))));
+	else if constexpr (__y_to_x)
+	  {
+	    auto __a =3D _mm256_shuffle_epi8(
+	      __intrin,
+	      _mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, -0x80, -0x80, -0x80,
+			       -0x80, -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
+			       -0x80, -0x80, -0x80, -0x80, -0x80, -0x80, 0, 2,
+			       4, 6, 8, 10, 12, 14));
+	    return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
+	  }
+	else if constexpr (__z_to_y && __have_avx512bw)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi16_epi8(__intrin));
+	else if constexpr (__z_to_y)
+	  __assert_unreachable<_Tp>();
+      }
+    else if constexpr (__i8_to_i64) //{{{2
+      {
+	if constexpr (__x_to_x && __have_sse4_1)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm_cvtepi8_epi64(__intrin)
+					 : _mm_cvtepu8_epi64(__intrin));
+	else if constexpr (__x_to_x && is_signed_v<_Tp>)
+	  {
+	    if constexpr (__have_ssse3)
+	      {
+		auto __dup =3D _mm_unpacklo_epi8(__intrin, __intrin);
+		auto __epi16 =3D _mm_srai_epi16(__dup, 8);
+		_mm_shuffle_epi8(__epi16,
+				 _mm_setr_epi8(0, 1, 1, 1, 1, 1, 1, 1, 2, 3, 3,
+					       3, 3, 3, 3, 3));
+	      }
+	    else
+	      {
+		auto __x =3D _mm_unpacklo_epi8(__intrin, __intrin);
+		__x =3D _mm_unpacklo_epi16(__x, __x);
+		return __intrin_bitcast<_To>(
+		  _mm_unpacklo_epi32(_mm_srai_epi32(__x, 24),
+				     _mm_srai_epi32(__x, 31)));
+	      }
+	  }
+	else if constexpr (__x_to_x)
+	  {
+	    return __intrin_bitcast<_To>(_mm_unpacklo_epi32(
+	      _mm_unpacklo_epi16(_mm_unpacklo_epi8(__intrin, __m128i()),
+				 __m128i()),
+	      __m128i()));
+	  }
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm256_cvtepi8_epi64(__intrin)
+					 : _mm256_cvtepu8_epi64(__intrin));
+	else if constexpr (__x_to_z)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm512_cvtepi8_epi64(__intrin)
+					 : _mm512_cvtepu8_epi64(__intrin));
+      }
+    else if constexpr (__i8_to_i32) //{{{2
+      {
+	if constexpr (__x_to_x && __have_sse4_1)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm_cvtepi8_epi32(__intrin)
+					 : _mm_cvtepu8_epi32(__intrin));
+	else if constexpr (__x_to_x && is_signed_v<_Tp>)
+	  {
+	    const auto __x =3D _mm_unpacklo_epi8(__intrin, __intrin);
+	    return __intrin_bitcast<_To>(
+	      _mm_srai_epi32(_mm_unpacklo_epi16(__x, __x), 24));
+	  }
+	else if constexpr (__x_to_x && is_unsigned_v<_Tp>)
+	  return __intrin_bitcast<_To>(
+	    _mm_unpacklo_epi16(_mm_unpacklo_epi8(__intrin, __m128i()),
+			       __m128i()));
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm256_cvtepi8_epi32(__intrin)
+					 : _mm256_cvtepu8_epi32(__intrin));
+	else if constexpr (__x_to_z)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm512_cvtepi8_epi32(__intrin)
+					 : _mm512_cvtepu8_epi32(__intrin));
+      }
+    else if constexpr (__i8_to_i16) //{{{2
+      {
+	if constexpr (__x_to_x && __have_sse4_1)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm_cvtepi8_epi16(__intrin)
+					 : _mm_cvtepu8_epi16(__intrin));
+	else if constexpr (__x_to_x && is_signed_v<_Tp>)
+	  return __intrin_bitcast<_To>(
+	    _mm_srai_epi16(_mm_unpacklo_epi8(__intrin, __intrin), 8));
+	else if constexpr (__x_to_x && is_unsigned_v<_Tp>)
+	  return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__intrin, __m128i()));
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm256_cvtepi8_epi16(__intrin)
+					 : _mm256_cvtepu8_epi16(__intrin));
+	else if constexpr (__y_to_z && __have_avx512bw)
+	  return __intrin_bitcast<_To>(is_signed_v<_Tp>
+					 ? _mm512_cvtepi8_epi16(__intrin)
+					 : _mm512_cvtepu8_epi16(__intrin));
+	else if constexpr (__y_to_z)
+	  __assert_unreachable<_Tp>();
+      }
+    else if constexpr (__f32_to_s64) //{{{2
+      {
+	if constexpr (__have_avx512dq_vl && __x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvttps_epi64(__intrin));
+	else if constexpr (__have_avx512dq_vl && __x_to_y)
+	  return __intrin_bitcast<_To>(_mm256_cvttps_epi64(__intrin));
+	else if constexpr (__have_avx512dq && __y_to_z)
+	  return __intrin_bitcast<_To>(_mm512_cvttps_epi64(__intrin));
+	// else use scalar fallback
+      }
+    else if constexpr (__f32_to_u64) //{{{2
+      {
+	if constexpr (__have_avx512dq_vl && __x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvttps_epu64(__intrin));
+	else if constexpr (__have_avx512dq_vl && __x_to_y)
+	  return __intrin_bitcast<_To>(_mm256_cvttps_epu64(__intrin));
+	else if constexpr (__have_avx512dq && __y_to_z)
+	  return __intrin_bitcast<_To>(_mm512_cvttps_epu64(__intrin));
+	// else use scalar fallback
+      }
+    else if constexpr (__f32_to_s32) //{{{2
+      {
+	if constexpr (__x_to_x || __y_to_y || __z_to_z)
+	  {
+	    // go to fallback, it does the right thing
+	  }
+	else
+	  __assert_unreachable<_Tp>();
+      }
+    else if constexpr (__f32_to_u32) //{{{2
+      {
+	if constexpr (__have_avx512vl && __x_to_x)
+	  return __auto_bitcast(_mm_cvttps_epu32(__intrin));
+	else if constexpr (__have_avx512f && __x_to_x)
+	  return __auto_bitcast(
+	    __lo128(_mm512_cvttps_epu32(__auto_bitcast(__v))));
+	else if constexpr (__have_avx512vl && __y_to_y)
+	  return __vector_bitcast<_Up>(_mm256_cvttps_epu32(__intrin));
+	else if constexpr (__have_avx512f && __y_to_y)
+	  return __vector_bitcast<_Up>(
+	    __lo256(_mm512_cvttps_epu32(__auto_bitcast(__v))));
+	else if constexpr (__x_to_x || __y_to_y || __z_to_z)
+	  {
+	    // go to fallback, it does the right thing. We can't use the
+	    // _mm_floor_ps - 0x8000'0000 trick for f32->u32 because it would
+	    // discard small input values (only 24 mantissa bits)
+	  }
+	else
+	  __assert_unreachable<_Tp>();
+      }
+    else if constexpr (__f32_to_ibw) //{{{2
+      return __convert_x86<_To>(__convert_x86<__vector_type_t<int, _Np>>(_=
_v));
+    else if constexpr (__f64_to_s64) //{{{2
+      {
+	if constexpr (__have_avx512dq_vl && __x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvttpd_epi64(__intrin));
+	else if constexpr (__have_avx512dq_vl && __y_to_y)
+	  return __intrin_bitcast<_To>(_mm256_cvttpd_epi64(__intrin));
+	else if constexpr (__have_avx512dq && __z_to_z)
+	  return __intrin_bitcast<_To>(_mm512_cvttpd_epi64(__intrin));
+	// else use scalar fallback
+      }
+    else if constexpr (__f64_to_u64) //{{{2
+      {
+	if constexpr (__have_avx512dq_vl && __x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvttpd_epu64(__intrin));
+	else if constexpr (__have_avx512dq_vl && __y_to_y)
+	  return __intrin_bitcast<_To>(_mm256_cvttpd_epu64(__intrin));
+	else if constexpr (__have_avx512dq && __z_to_z)
+	  return __intrin_bitcast<_To>(_mm512_cvttpd_epu64(__intrin));
+	// else use scalar fallback
+      }
+    else if constexpr (__f64_to_s32) //{{{2
+      {
+	if constexpr (__x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvttpd_epi32(__intrin));
+	else if constexpr (__y_to_x)
+	  return __intrin_bitcast<_To>(_mm256_cvttpd_epi32(__intrin));
+	else if constexpr (__z_to_y)
+	  return __intrin_bitcast<_To>(_mm512_cvttpd_epi32(__intrin));
+      }
+    else if constexpr (__f64_to_u32) //{{{2
+      {
+	if constexpr (__have_avx512vl && __x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvttpd_epu32(__intrin));
+	else if constexpr (__have_sse4_1 && __x_to_x)
+	  return __vector_bitcast<_Up, _M>(
+		   _mm_cvttpd_epi32(_mm_floor_pd(__intrin) - 0x8000'0000u))
+		 ^ 0x8000'0000u;
+	else if constexpr (__x_to_x)
+	  {
+	    // use scalar fallback: it's only 2 values to convert, can't get
+	    // much better than scalar decomposition
+	  }
+	else if constexpr (__have_avx512vl && __y_to_x)
+	  return __intrin_bitcast<_To>(_mm256_cvttpd_epu32(__intrin));
+	else if constexpr (__y_to_x)
+	  {
+	    return __intrin_bitcast<_To>(
+	      __vector_bitcast<_Up>(
+		_mm256_cvttpd_epi32(_mm256_floor_pd(__intrin) - 0x8000'0000u))
+	      ^ 0x8000'0000u);
+	  }
+	else if constexpr (__z_to_y)
+	  return __intrin_bitcast<_To>(_mm512_cvttpd_epu32(__intrin));
+      }
+    else if constexpr (__f64_to_ibw) //{{{2
+      {
+	return __convert_x86<_To>(
+	  __convert_x86<__vector_type_t<int, (_Np < 4 ? 4 : _Np)>>(__v));
+      }
+    else if constexpr (__s64_to_f32) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi64_ps(__intrin));
+	else if constexpr (__y_to_x && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi64_ps(__intrin));
+	else if constexpr (__z_to_y && __have_avx512dq)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi64_ps(__intrin));
+	else if constexpr (__z_to_y)
+	  return __intrin_bitcast<_To>(
+	    _mm512_cvtpd_ps(__convert_x86<__vector_type_t<double, 8>>(__v)));
+      }
+    else if constexpr (__u64_to_f32) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepu64_ps(__intrin));
+	else if constexpr (__y_to_x && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepu64_ps(__intrin));
+	else if constexpr (__z_to_y && __have_avx512dq)
+	  return __intrin_bitcast<_To>(_mm512_cvtepu64_ps(__intrin));
+	else if constexpr (__z_to_y)
+	  {
+	    return __intrin_bitcast<_To>(
+	      __lo256(_mm512_cvtepu32_ps(__auto_bitcast(
+		_mm512_cvtepi64_epi32(_mm512_srai_epi64(__intrin, 32)))))
+		* 0x100000000LL
+	      + __lo256(_mm512_cvtepu32_ps(
+		__auto_bitcast(_mm512_cvtepi64_epi32(__intrin)))));
+	  }
+      }
+    else if constexpr (__s32_to_f32) //{{{2
+      {
+	// use fallback (builtin conversion)
+      }
+    else if constexpr (__u32_to_f32) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512vl)
+	  {
+	    // use fallback
+	  }
+	else if constexpr (__x_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepu32_ps(__auto_bitcast(__v))));
+	else if constexpr (__x_to_x && (__have_fma || __have_fma4))
+	  // work around PR85819
+	  return __auto_bitcast(0x10000
+				  * _mm_cvtepi32_ps(__to_intrin(__v >> 16))
+				+ _mm_cvtepi32_ps(__to_intrin(__v & 0xffff)));
+	else if constexpr (__y_to_y && __have_avx512vl)
+	  {
+	    // use fallback
+	  }
+	else if constexpr (__y_to_y && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo256(_mm512_cvtepu32_ps(__auto_bitcast(__v))));
+	else if constexpr (__y_to_y)
+	  // work around PR85819
+	  return 0x10000 * _mm256_cvtepi32_ps(__to_intrin(__v >> 16))
+		 + _mm256_cvtepi32_ps(__to_intrin(__v & 0xffff));
+	// else use fallback (builtin conversion)
+      }
+    else if constexpr (__ibw_to_f32) //{{{2
+      {
+	if constexpr (_M <=3D 4 || __have_avx2)
+	  return __convert_x86<_To>(
+	    __convert_x86<__vector_type_t<int, _M>>(__v));
+	else
+	  {
+	    static_assert(__x_to_y);
+	    __m128i __a, __b;
+	    if constexpr (__have_sse4_1)
+	      {
+		__a =3D sizeof(_Tp) =3D=3D 2
+			? (is_signed_v<_Tp> ? _mm_cvtepi16_epi32(__intrin)
+					    : _mm_cvtepu16_epi32(__intrin))
+			: (is_signed_v<_Tp> ? _mm_cvtepi8_epi32(__intrin)
+					    : _mm_cvtepu8_epi32(__intrin));
+		const auto __w
+		  =3D _mm_shuffle_epi32(__intrin, sizeof(_Tp) =3D=3D 2 ? 0xee : 0xe9);
+		__b =3D sizeof(_Tp) =3D=3D 2
+			? (is_signed_v<_Tp> ? _mm_cvtepi16_epi32(__w)
+					    : _mm_cvtepu16_epi32(__w))
+			: (is_signed_v<_Tp> ? _mm_cvtepi8_epi32(__w)
+					    : _mm_cvtepu8_epi32(__w));
+	      }
+	    else
+	      {
+		__m128i __tmp;
+		if constexpr (sizeof(_Tp) =3D=3D 1)
+		  {
+		    __tmp =3D is_signed_v<_Tp>
+			      ? _mm_srai_epi16(_mm_unpacklo_epi8(__intrin,
+								 __intrin),
+					       8)
+			      : _mm_unpacklo_epi8(__intrin, __m128i());
+		  }
+		else
+		  {
+		    static_assert(sizeof(_Tp) =3D=3D 2);
+		    __tmp =3D __intrin;
+		  }
+		__a =3D is_signed_v<_Tp>
+			? _mm_srai_epi32(_mm_unpacklo_epi16(__tmp, __tmp), 16)
+			: _mm_unpacklo_epi16(__tmp, __m128i());
+		__b =3D is_signed_v<_Tp>
+			? _mm_srai_epi32(_mm_unpackhi_epi16(__tmp, __tmp), 16)
+			: _mm_unpackhi_epi16(__tmp, __m128i());
+	      }
+	    return __convert_x86<_To>(__vector_bitcast<int>(__a),
+				      __vector_bitcast<int>(__b));
+	  }
+      }
+    else if constexpr (__s64_to_f64) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepi64_pd(__intrin));
+	else if constexpr (__y_to_y && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi64_pd(__intrin));
+	else if constexpr (__z_to_z && __have_avx512dq)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi64_pd(__intrin));
+	else if constexpr (__z_to_z)
+	  {
+	    return __intrin_bitcast<_To>(
+	      _mm512_cvtepi32_pd(_mm512_cvtepi64_epi32(__to_intrin(__v >> 32)))
+		* 0x100000000LL
+	      + _mm512_cvtepu32_pd(_mm512_cvtepi64_epi32(__intrin)));
+	  }
+      }
+    else if constexpr (__u64_to_f64) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepu64_pd(__intrin));
+	else if constexpr (__y_to_y && __have_avx512dq_vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepu64_pd(__intrin));
+	else if constexpr (__z_to_z && __have_avx512dq)
+	  return __intrin_bitcast<_To>(_mm512_cvtepu64_pd(__intrin));
+	else if constexpr (__z_to_z)
+	  {
+	    return __intrin_bitcast<_To>(
+	      _mm512_cvtepu32_pd(_mm512_cvtepi64_epi32(__to_intrin(__v >> 32)))
+		* 0x100000000LL
+	      + _mm512_cvtepu32_pd(_mm512_cvtepi64_epi32(__intrin)));
+	  }
+      }
+    else if constexpr (__s32_to_f64) //{{{2
+      {
+	if constexpr (__x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvtepi32_pd(__intrin));
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(_mm256_cvtepi32_pd(__intrin));
+	else if constexpr (__y_to_z)
+	  return __intrin_bitcast<_To>(_mm512_cvtepi32_pd(__intrin));
+      }
+    else if constexpr (__u32_to_f64) //{{{2
+      {
+	if constexpr (__x_to_x && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm_cvtepu32_pd(__intrin));
+	else if constexpr (__x_to_x && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo128(_mm512_cvtepu32_pd(__auto_bitcast(__v))));
+	else if constexpr (__x_to_x)
+	  return __intrin_bitcast<_To>(
+	    _mm_cvtepi32_pd(__to_intrin(__v ^ 0x8000'0000u)) + 0x8000'0000u);
+	else if constexpr (__x_to_y && __have_avx512vl)
+	  return __intrin_bitcast<_To>(_mm256_cvtepu32_pd(__intrin));
+	else if constexpr (__x_to_y && __have_avx512f)
+	  return __intrin_bitcast<_To>(
+	    __lo256(_mm512_cvtepu32_pd(__auto_bitcast(__v))));
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(
+	    _mm256_cvtepi32_pd(__to_intrin(__v ^ 0x8000'0000u)) + 0x8000'0000u);
+	else if constexpr (__y_to_z)
+	  return __intrin_bitcast<_To>(_mm512_cvtepu32_pd(__intrin));
+      }
+    else if constexpr (__ibw_to_f64) //{{{2
+      {
+	return __convert_x86<_To>(
+	  __convert_x86<__vector_type_t<int, std::max(size_t(4), _M)>>(__v));
+      }
+    else if constexpr (__f32_to_f64) //{{{2
+      {
+	if constexpr (__x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvtps_pd(__intrin));
+	else if constexpr (__x_to_y)
+	  return __intrin_bitcast<_To>(_mm256_cvtps_pd(__intrin));
+	else if constexpr (__y_to_z)
+	  return __intrin_bitcast<_To>(_mm512_cvtps_pd(__intrin));
+      }
+    else if constexpr (__f64_to_f32) //{{{2
+      {
+	if constexpr (__x_to_x)
+	  return __intrin_bitcast<_To>(_mm_cvtpd_ps(__intrin));
+	else if constexpr (__y_to_x)
+	  return __intrin_bitcast<_To>(_mm256_cvtpd_ps(__intrin));
+	else if constexpr (__z_to_y)
+	  return __intrin_bitcast<_To>(_mm512_cvtpd_ps(__intrin));
+      }
+    else //{{{2
+      __assert_unreachable<_Tp>();
+
+    // fallback:{{{2
+    return __vector_convert<_To>(__v, make_index_sequence<std::min(_M, _Np=
)>());
+    //}}}
+  }
+
+// }}}
+// 2-arg __convert_x86 {{{1
+template <typename _To, typename _V, typename _Traits>
+  _GLIBCXX_SIMD_INTRINSIC _To
+  __convert_x86(_V __v0, _V __v1)
+  {
+    static_assert(__is_vector_type_v<_V>);
+    using _Tp =3D typename _Traits::value_type;
+    constexpr size_t _Np =3D _Traits::_S_full_size;
+    [[maybe_unused]] const auto __i0 =3D __to_intrin(__v0);
+    [[maybe_unused]] const auto __i1 =3D __to_intrin(__v1);
+    using _Up =3D typename _VectorTraits<_To>::value_type;
+    constexpr size_t _M =3D _VectorTraits<_To>::_S_full_size;
+
+    static_assert(2 * _Np <=3D _M,
+		  "__v1 would be discarded; use the one-argument "
+		  "__convert_x86 overload instead");
+
+    // [xyz]_to_[xyz] {{{2
+    [[maybe_unused]] constexpr bool __x_to_x
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __x_to_y
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __x_to_z
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __y_to_x
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __y_to_y
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __y_to_z
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __z_to_x
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __z_to_y
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __z_to_z
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) =3D=3D 64;
+
+    // iX_to_iX {{{2
+    [[maybe_unused]] constexpr bool __i_to_i
+      =3D is_integral_v<_Up> && is_integral_v<_Tp>;
+    [[maybe_unused]] constexpr bool __i8_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i8_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __i8_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i16_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i16_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __i16_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i32_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i32_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i32_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i64_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i64_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i64_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 4;
+
+    // [fsu]X_to_[fsu]X {{{2
+    // ibw =3D integral && byte or word, i.e. char and short with any sign=
edness
+    [[maybe_unused]] constexpr bool __i64_to_f32
+      =3D is_integral_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s32_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s16_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s8_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u32_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u16_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u8_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s64_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __s32_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __s16_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __s8_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u64_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u32_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u16_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u8_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_s64
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_s32
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_u64
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_u32
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f64_to_s64
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_s32
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_u64
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_u32
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_ibw
+      =3D is_integral_v<_Up> && sizeof(_Up) <=3D 2
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f64_to_ibw
+      =3D is_integral_v<_Up> && sizeof(_Up) <=3D 2
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_f64
+      =3D is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_f32
+      =3D is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+
+    if constexpr (__i_to_i && __y_to_x && !__have_avx2) //{{{2
+      // <double, 4>, <double, 4> =3D> <short, 8>
+      return __convert_x86<_To>(__lo128(__v0), __hi128(__v0), __lo128(__v1=
),
+				__hi128(__v1));
+    else if constexpr (__i_to_i) // assert ISA {{{2
+      {
+	static_assert(__x_to_x || __have_avx2,
+		      "integral conversions with ymm registers require AVX2");
+	static_assert(__have_avx512bw
+			|| ((sizeof(_Tp) >=3D 4 || sizeof(__v0) < 64)
+			    && (sizeof(_Up) >=3D 4 || sizeof(_To) < 64)),
+		      "8/16-bit integers in zmm registers require AVX512BW");
+	static_assert((sizeof(__v0) < 64 && sizeof(_To) < 64) || __have_avx512f,
+		      "integral conversions with ymm registers require AVX2");
+      }
+    // concat =3D> use 1-arg __convert_x86 {{{2
+    if constexpr (sizeof(__v0) < 16 || (sizeof(__v0) =3D=3D 16 && __have_a=
vx2)
+		  || (sizeof(__v0) =3D=3D 16 && __have_avx
+		      && is_floating_point_v<_Tp>)
+		  || (sizeof(__v0) =3D=3D 32 && __have_avx512f
+		      && (sizeof(_Tp) >=3D 4 || __have_avx512bw)))
+      {
+	// The ISA can handle wider input registers, so concat and use one-arg
+	// implementation. This reduces code duplication considerably.
+	return __convert_x86<_To>(__concat(__v0, __v1));
+      }
+    else //{{{2
+      {
+	// conversion using bit reinterpretation (or no conversion at all)
+	// should all go through the concat branch above:
+	static_assert(
+	  !(is_floating_point_v<
+	      _Tp> =3D=3D is_floating_point_v<_Up> && sizeof(_Tp) =3D=3D sizeof(_=
Up)));
+	// handle all zero extension{{{2
+	if constexpr (2 * _Np < _M && sizeof(_To) > 16)
+	  {
+	    constexpr size_t Min =3D 16 / sizeof(_Up);
+	    return __zero_extend(
+	      __convert_x86<
+		__vector_type_t<_Up, (Min > 2 * _Np) ? Min : 2 * _Np>>(__v0,
+								       __v1));
+	  }
+	else if constexpr (__i64_to_i32) //{{{2
+	  {
+	    if constexpr (__x_to_x)
+	      return __auto_bitcast(_mm_shuffle_ps(__auto_bitcast(__v0),
+						   __auto_bitcast(__v1), 0x88));
+	    else if constexpr (__y_to_y)
+	      {
+		// AVX512F is not available (would concat otherwise)
+		return __auto_bitcast(
+		  __xzyw(_mm256_shuffle_ps(__auto_bitcast(__v0),
+					   __auto_bitcast(__v1), 0x88)));
+		// alternative:
+		// const auto v0_abxxcdxx =3D _mm256_shuffle_epi32(__v0, 8);
+		// const auto v1_efxxghxx =3D _mm256_shuffle_epi32(__v1, 8);
+		// const auto v_abefcdgh =3D _mm256_unpacklo_epi64(v0_abxxcdxx,
+		// v1_efxxghxx); return _mm256_permute4x64_epi64(v_abefcdgh,
+		// 0x01 * 0 + 0x04 * 2 + 0x10 * 1 + 0x40 * 3);  // abcdefgh
+	      }
+	    else if constexpr (__z_to_z)
+	      return __intrin_bitcast<_To>(
+		__concat(_mm512_cvtepi64_epi32(__i0),
+			 _mm512_cvtepi64_epi32(__i1)));
+	  }
+	else if constexpr (__i64_to_i16) //{{{2
+	  {
+	    if constexpr (__x_to_x)
+	      {
+		// AVX2 is not available (would concat otherwise)
+		if constexpr (__have_sse4_1)
+		  {
+		    return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+		      _mm_blend_epi16(__i0, _mm_slli_si128(__i1, 4), 0x44),
+		      _mm_setr_epi8(0, 1, 8, 9, 4, 5, 12, 13, -0x80, -0x80,
+				    -0x80, -0x80, -0x80, -0x80, -0x80, -0x80)));
+		  }
+		else
+		  {
+		    return __vector_type_t<_Up, _M>{_Up(__v0[0]), _Up(__v0[1]),
+						    _Up(__v1[0]), _Up(__v1[1])};
+		  }
+	      }
+	    else if constexpr (__y_to_x)
+	      {
+		auto __a
+		  =3D _mm256_unpacklo_epi16(__i0, __i1); // 04.. .... 26.. ....
+		auto __b
+		  =3D _mm256_unpackhi_epi16(__i0, __i1); // 15.. .... 37.. ....
+		auto __c
+		  =3D _mm256_unpacklo_epi16(__a, __b); // 0145 .... 2367 ....
+		return __intrin_bitcast<_To>(
+		  _mm_unpacklo_epi32(__lo128(__c), __hi128(__c))); // 0123 4567
+	      }
+	    else if constexpr (__z_to_y)
+	      return __intrin_bitcast<_To>(
+		__concat(_mm512_cvtepi64_epi16(__i0),
+			 _mm512_cvtepi64_epi16(__i1)));
+	  }
+	else if constexpr (__i64_to_i8) //{{{2
+	  {
+	    if constexpr (__x_to_x && __have_sse4_1)
+	      {
+		return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+		  _mm_blend_epi16(__i0, _mm_slli_si128(__i1, 4), 0x44),
+		  _mm_setr_epi8(0, 8, 4, 12, -0x80, -0x80, -0x80, -0x80, -0x80,
+				-0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
+				-0x80)));
+	      }
+	    else if constexpr (__x_to_x && __have_ssse3)
+	      {
+		return __intrin_bitcast<_To>(_mm_unpacklo_epi16(
+		  _mm_shuffle_epi8(
+		    __i0, _mm_setr_epi8(0, 8, -0x80, -0x80, -0x80, -0x80, -0x80,
+					-0x80, -0x80, -0x80, -0x80, -0x80,
+					-0x80, -0x80, -0x80, -0x80)),
+		  _mm_shuffle_epi8(
+		    __i1, _mm_setr_epi8(0, 8, -0x80, -0x80, -0x80, -0x80, -0x80,
+					-0x80, -0x80, -0x80, -0x80, -0x80,
+					-0x80, -0x80, -0x80, -0x80))));
+	      }
+	    else if constexpr (__x_to_x)
+	      {
+		return __vector_type_t<_Up, _M>{_Up(__v0[0]), _Up(__v0[1]),
+						_Up(__v1[0]), _Up(__v1[1])};
+	      }
+	    else if constexpr (__y_to_x)
+	      {
+		const auto __a =3D _mm256_shuffle_epi8(
+		  _mm256_blend_epi32(__i0, _mm256_slli_epi64(__i1, 32), 0xAA),
+		  _mm256_setr_epi8(0, 8, -0x80, -0x80, 4, 12, -0x80, -0x80,
+				   -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
+				   -0x80, -0x80, -0x80, -0x80, 0, 8, -0x80,
+				   -0x80, 4, 12, -0x80, -0x80, -0x80, -0x80,
+				   -0x80, -0x80, -0x80, -0x80));
+		return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
+	      } // __z_to_x uses concat fallback
+	  }
+	else if constexpr (__i32_to_i16) //{{{2
+	  {
+	    if constexpr (__x_to_x)
+	      {
+		// AVX2 is not available (would concat otherwise)
+		if constexpr (__have_sse4_1)
+		  {
+		    return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+		      _mm_blend_epi16(__i0, _mm_slli_si128(__i1, 2), 0xaa),
+		      _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10,
+				    11, 14, 15)));
+		  }
+		else if constexpr (__have_ssse3)
+		  {
+		    return __intrin_bitcast<_To>(
+		      _mm_hadd_epi16(__to_intrin(__v0 << 16),
+				     __to_intrin(__v1 << 16)));
+		    /*
+		    return _mm_unpacklo_epi64(
+			_mm_shuffle_epi8(__i0, _mm_setr_epi8(0, 1, 4, 5, 8, 9,
+		    12, 13, 8, 9, 12, 13, 12, 13, 14, 15)),
+		    _mm_shuffle_epi8(__i1, _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12,
+		    13, 8, 9, 12, 13, 12, 13, 14, 15)));
+							   */
+		  }
+		else
+		  {
+		    auto __a =3D _mm_unpacklo_epi16(__i0, __i1); // 04.. 15..
+		    auto __b =3D _mm_unpackhi_epi16(__i0, __i1); // 26.. 37..
+		    auto __c =3D _mm_unpacklo_epi16(__a, __b);   // 0246 ....
+		    auto __d =3D _mm_unpackhi_epi16(__a, __b);   // 1357 ....
+		    return __intrin_bitcast<_To>(
+		      _mm_unpacklo_epi16(__c, __d)); // 0123 4567
+		  }
+	      }
+	    else if constexpr (__y_to_y)
+	      {
+		const auto __shuf
+		  =3D _mm256_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80,
+				     -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
+				     0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80,
+				     -0x80, -0x80, -0x80, -0x80, -0x80, -0x80);
+		auto __a =3D _mm256_shuffle_epi8(__i0, __shuf);
+		auto __b =3D _mm256_shuffle_epi8(__i1, __shuf);
+		return __intrin_bitcast<_To>(
+		  __xzyw(_mm256_unpacklo_epi64(__a, __b)));
+	      } // __z_to_z uses concat fallback
+	  }
+	else if constexpr (__i32_to_i8) //{{{2
+	  {
+	    if constexpr (__x_to_x && __have_ssse3)
+	      {
+		const auto shufmask
+		  =3D _mm_setr_epi8(0, 4, 8, 12, -0x80, -0x80, -0x80, -0x80,
+				  -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
+				  -0x80, -0x80);
+		return __intrin_bitcast<_To>(
+		  _mm_unpacklo_epi32(_mm_shuffle_epi8(__i0, shufmask),
+				     _mm_shuffle_epi8(__i1, shufmask)));
+	      }
+	    else if constexpr (__x_to_x)
+	      {
+		auto __a =3D _mm_unpacklo_epi8(__i0, __i1); // 04.. .... 15.. ....
+		auto __b =3D _mm_unpackhi_epi8(__i0, __i1); // 26.. .... 37.. ....
+		auto __c =3D _mm_unpacklo_epi8(__a, __b);   // 0246 .... .... ....
+		auto __d =3D _mm_unpackhi_epi8(__a, __b);   // 1357 .... .... ....
+		auto __e =3D _mm_unpacklo_epi8(__c, __d);   // 0123 4567 .... ....
+		return __intrin_bitcast<_To>(__e & __m128i{-1, 0});
+	      }
+	    else if constexpr (__y_to_x)
+	      {
+		const auto __a =3D _mm256_shuffle_epi8(
+		  _mm256_blend_epi16(__i0, _mm256_slli_epi32(__i1, 16), 0xAA),
+		  _mm256_setr_epi8(0, 4, 8, 12, -0x80, -0x80, -0x80, -0x80, 2,
+				   6, 10, 14, -0x80, -0x80, -0x80, -0x80, -0x80,
+				   -0x80, -0x80, -0x80, 0, 4, 8, 12, -0x80,
+				   -0x80, -0x80, -0x80, 2, 6, 10, 14));
+		return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
+	      } // __z_to_y uses concat fallback
+	  }
+	else if constexpr (__i16_to_i8) //{{{2
+	  {
+	    if constexpr (__x_to_x && __have_ssse3)
+	      {
+		const auto __shuf =3D reinterpret_cast<__m128i>(
+		  __vector_type_t<_UChar, 16>{0, 2, 4, 6, 8, 10, 12, 14, 0x80,
+					      0x80, 0x80, 0x80, 0x80, 0x80,
+					      0x80, 0x80});
+		return __intrin_bitcast<_To>(
+		  _mm_unpacklo_epi64(_mm_shuffle_epi8(__i0, __shuf),
+				     _mm_shuffle_epi8(__i1, __shuf)));
+	      }
+	    else if constexpr (__x_to_x)
+	      {
+		auto __a =3D _mm_unpacklo_epi8(__i0, __i1); // 08.. 19.. 2A.. 3B..
+		auto __b =3D _mm_unpackhi_epi8(__i0, __i1); // 4C.. 5D.. 6E.. 7F..
+		auto __c =3D _mm_unpacklo_epi8(__a, __b);   // 048C .... 159D ....
+		auto __d =3D _mm_unpackhi_epi8(__a, __b);   // 26AE .... 37BF ....
+		auto __e =3D _mm_unpacklo_epi8(__c, __d);   // 0246 8ACE .... ....
+		auto __f =3D _mm_unpackhi_epi8(__c, __d);   // 1357 9BDF .... ....
+		return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__e, __f));
+	      }
+	    else if constexpr (__y_to_y)
+	      {
+		return __intrin_bitcast<_To>(__xzyw(_mm256_shuffle_epi8(
+		  (__to_intrin(__v0) & _mm256_set1_epi32(0x00ff00ff))
+		    | _mm256_slli_epi16(__i1, 8),
+		  _mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11,
+				   13, 15, 0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5,
+				   7, 9, 11, 13, 15))));
+	      } // __z_to_z uses concat fallback
+	  }
+	else if constexpr (__i64_to_f32) //{{{2
+	  {
+	    if constexpr (__x_to_x)
+	      return __make_wrapper<float>(__v0[0], __v0[1], __v1[0], __v1[1]);
+	    else if constexpr (__y_to_y)
+	      {
+		static_assert(__y_to_y && __have_avx2);
+		const auto __a =3D _mm256_unpacklo_epi32(__i0, __i1); // aeAE cgCG
+		const auto __b =3D _mm256_unpackhi_epi32(__i0, __i1); // bfBF dhDH
+		const auto __lo32
+		  =3D _mm256_unpacklo_epi32(__a, __b); // abef cdgh
+		const auto __hi32 =3D __vector_bitcast<
+		  conditional_t<is_signed_v<_Tp>, int, _UInt>>(
+		  _mm256_unpackhi_epi32(__a, __b)); // ABEF CDGH
+		const auto __hi
+		  =3D 0x100000000LL
+		    * __convert_x86<__vector_type_t<float, 8>>(__hi32);
+		const auto __mid
+		  =3D 0x10000 * _mm256_cvtepi32_ps(_mm256_srli_epi32(__lo32, 16));
+		const auto __lo
+		  =3D _mm256_cvtepi32_ps(_mm256_set1_epi32(0x0000ffffu) & __lo32);
+		return __xzyw((__hi + __mid) + __lo);
+	      }
+	    else if constexpr (__z_to_z && __have_avx512dq)
+	      {
+		return is_signed_v<_Tp> ? __concat(_mm512_cvtepi64_ps(__i0),
+						   _mm512_cvtepi64_ps(__i1))
+					: __concat(_mm512_cvtepu64_ps(__i0),
+						   _mm512_cvtepu64_ps(__i1));
+	      }
+	    else if constexpr (__z_to_z && is_signed_v<_Tp>)
+	      {
+		const __m512 __hi32 =3D _mm512_cvtepi32_ps(
+		  __concat(_mm512_cvtepi64_epi32(__to_intrin(__v0 >> 32)),
+			   _mm512_cvtepi64_epi32(__to_intrin(__v1 >> 32))));
+		const __m512i __lo32 =3D __concat(_mm512_cvtepi64_epi32(__i0),
+						_mm512_cvtepi64_epi32(__i1));
+		// split low 32-bits, because if __hi32 is a small negative
+		// number, the 24-bit mantissa may lose important information if
+		// any of the high 8 bits of __lo32 is set, leading to
+		// catastrophic cancelation in the FMA
+		const __m512 __hi16
+		  =3D _mm512_cvtepu32_ps(_mm512_set1_epi32(0xffff0000u) & __lo32);
+		const __m512 __lo16
+		  =3D _mm512_cvtepi32_ps(_mm512_set1_epi32(0x0000ffffu) & __lo32);
+		return (__hi32 * 0x100000000LL + __hi16) + __lo16;
+	      }
+	    else if constexpr (__z_to_z && is_unsigned_v<_Tp>)
+	      {
+		return __intrin_bitcast<_To>(
+		  _mm512_cvtepu32_ps(__concat(
+		    _mm512_cvtepi64_epi32(_mm512_srai_epi64(__i0, 32)),
+		    _mm512_cvtepi64_epi32(_mm512_srai_epi64(__i1, 32))))
+		    * 0x100000000LL
+		  + _mm512_cvtepu32_ps(__concat(_mm512_cvtepi64_epi32(__i0),
+						_mm512_cvtepi64_epi32(__i1))));
+	      }
+	  }
+	else if constexpr (__f64_to_s32) //{{{2
+	  {
+	    // use concat fallback
+	  }
+	else if constexpr (__f64_to_u32) //{{{2
+	  {
+	    if constexpr (__x_to_x && __have_sse4_1)
+	      {
+		return __vector_bitcast<_Up, _M>(_mm_unpacklo_epi64(
+			 _mm_cvttpd_epi32(_mm_floor_pd(__i0) - 0x8000'0000u),
+			 _mm_cvttpd_epi32(_mm_floor_pd(__i1) - 0x8000'0000u)))
+		       ^ 0x8000'0000u;
+		// without SSE4.1 just use the scalar fallback, it's only four
+		// values
+	      }
+	    else if constexpr (__y_to_y)
+	      {
+		return __vector_bitcast<_Up>(
+			 __concat(_mm256_cvttpd_epi32(_mm256_floor_pd(__i0)
+						      - 0x8000'0000u),
+				  _mm256_cvttpd_epi32(_mm256_floor_pd(__i1)
+						      - 0x8000'0000u)))
+		       ^ 0x8000'0000u;
+	      } // __z_to_z uses fallback
+	  }
+	else if constexpr (__f64_to_ibw) //{{{2
+	  {
+	    // one-arg __f64_to_ibw goes via _SimdWrapper<int, ?>. The fallback
+	    // would go via two independet conversions to _SimdWrapper<_To> and
+	    // subsequent interleaving. This is better, because f64->__i32
+	    // allows to combine __v0 and __v1 into one register: if constexpr
+	    // (__z_to_x || __y_to_x) {
+	    return __convert_x86<_To>(
+	      __convert_x86<__vector_type_t<int, _Np * 2>>(__v0, __v1));
+	    //}
+	  }
+	else if constexpr (__f32_to_ibw) //{{{2
+	  {
+	    return __convert_x86<_To>(
+	      __convert_x86<__vector_type_t<int, _Np>>(__v0),
+	      __convert_x86<__vector_type_t<int, _Np>>(__v1));
+	  } //}}}
+
+	// fallback: {{{2
+	if constexpr (sizeof(_To) >=3D 32)
+	  // if _To is ymm or zmm, then _SimdWrapper<_Up, _M / 2> is xmm or ymm
+	  return __concat(__convert_x86<__vector_type_t<_Up, _M / 2>>(__v0),
+			  __convert_x86<__vector_type_t<_Up, _M / 2>>(__v1));
+	else if constexpr (sizeof(_To) =3D=3D 16)
+	  {
+	    const auto __lo =3D __to_intrin(__convert_x86<_To>(__v0));
+	    const auto __hi =3D __to_intrin(__convert_x86<_To>(__v1));
+	    if constexpr (sizeof(_Up) * _Np =3D=3D 8)
+	      {
+		if constexpr (is_floating_point_v<_Up>)
+		  return __auto_bitcast(
+		    _mm_unpacklo_pd(__vector_bitcast<double>(__lo),
+				    __vector_bitcast<double>(__hi)));
+		else
+		  return __intrin_bitcast<_To>(_mm_unpacklo_epi64(__lo, __hi));
+	      }
+	    else if constexpr (sizeof(_Up) * _Np =3D=3D 4)
+	      {
+		if constexpr (is_floating_point_v<_Up>)
+		  return __auto_bitcast(
+		    _mm_unpacklo_ps(__vector_bitcast<float>(__lo),
+				    __vector_bitcast<float>(__hi)));
+		else
+		  return __intrin_bitcast<_To>(_mm_unpacklo_epi32(__lo, __hi));
+	      }
+	    else if constexpr (sizeof(_Up) * _Np =3D=3D 2)
+	      return __intrin_bitcast<_To>(_mm_unpacklo_epi16(__lo, __hi));
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return __vector_convert<_To>(__v0, __v1, make_index_sequence<_Np>());
+	//}}}
+      }
+  }
+
+//}}}1
+// 4-arg __convert_x86 {{{1
+template <typename _To, typename _V, typename _Traits>
+  _GLIBCXX_SIMD_INTRINSIC _To
+  __convert_x86(_V __v0, _V __v1, _V __v2, _V __v3)
+  {
+    static_assert(__is_vector_type_v<_V>);
+    using _Tp =3D typename _Traits::value_type;
+    constexpr size_t _Np =3D _Traits::_S_full_size;
+    [[maybe_unused]] const auto __i0 =3D __to_intrin(__v0);
+    [[maybe_unused]] const auto __i1 =3D __to_intrin(__v1);
+    [[maybe_unused]] const auto __i2 =3D __to_intrin(__v2);
+    [[maybe_unused]] const auto __i3 =3D __to_intrin(__v3);
+    using _Up =3D typename _VectorTraits<_To>::value_type;
+    constexpr size_t _M =3D _VectorTraits<_To>::_S_full_size;
+
+    static_assert(4 * _Np <=3D _M,
+		  "__v2/__v3 would be discarded; use the two/one-argument "
+		  "__convert_x86 overload instead");
+
+    // [xyz]_to_[xyz] {{{2
+    [[maybe_unused]] constexpr bool __x_to_x
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __x_to_y
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __x_to_z
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __y_to_x
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __y_to_y
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __y_to_z
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __z_to_x
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __z_to_y
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __z_to_z
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) =3D=3D 64;
+
+    // iX_to_iX {{{2
+    [[maybe_unused]] constexpr bool __i_to_i
+      =3D is_integral_v<_Up> && is_integral_v<_Tp>;
+    [[maybe_unused]] constexpr bool __i8_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i8_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __i8_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 1 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i16_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i16_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __i16_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 2 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i32_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i32_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i32_to_i64
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 4 && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __i64_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __i64_to_i16
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 2;
+    [[maybe_unused]] constexpr bool __i64_to_i32
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 4;
+
+    // [fsu]X_to_[fsu]X {{{2
+    // ibw =3D integral && byte or word, i.e. char and short with any sign=
edness
+    [[maybe_unused]] constexpr bool __i64_to_f32
+      =3D is_integral_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s32_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s16_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s8_to_f32
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u32_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u16_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __u8_to_f32
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __s64_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __s32_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __s16_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __s8_to_f64
+      =3D is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u64_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u32_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u16_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 2
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __u8_to_f64
+      =3D is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) =3D=3D 1
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_s64
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_s32
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_u64
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f32_to_u32
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f64_to_s64
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_s32
+      =3D is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_u64
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 8
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_u32
+      =3D is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) =3D=3D 4
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_ibw
+      =3D is_integral_v<_Up> && sizeof(_Up) <=3D 2
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4;
+    [[maybe_unused]] constexpr bool __f64_to_ibw
+      =3D is_integral_v<_Up> && sizeof(_Up) <=3D 2
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f32_to_f64
+      =3D is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 4
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 8;
+    [[maybe_unused]] constexpr bool __f64_to_f32
+      =3D is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8
+	&& is_floating_point_v<_Up> && sizeof(_Up) =3D=3D 4;
+
+    if constexpr (__i_to_i && __y_to_x && !__have_avx2) //{{{2
+      {
+	// <double, 4>, <double, 4>, <double, 4>, <double, 4> =3D> <char, 16>
+	return __convert_x86<_To>(__lo128(__v0), __hi128(__v0), __lo128(__v1),
+				  __hi128(__v1), __lo128(__v2), __hi128(__v2),
+				  __lo128(__v3), __hi128(__v3));
+      }
+    else if constexpr (__i_to_i) // assert ISA {{{2
+      {
+	static_assert(__x_to_x || __have_avx2,
+		      "integral conversions with ymm registers require AVX2");
+	static_assert(__have_avx512bw
+			|| ((sizeof(_Tp) >=3D 4 || sizeof(__v0) < 64)
+			    && (sizeof(_Up) >=3D 4 || sizeof(_To) < 64)),
+		      "8/16-bit integers in zmm registers require AVX512BW");
+	static_assert((sizeof(__v0) < 64 && sizeof(_To) < 64) || __have_avx512f,
+		      "integral conversions with ymm registers require AVX2");
+      }
+    // concat =3D> use 2-arg __convert_x86 {{{2
+    if constexpr (sizeof(__v0) < 16 || (sizeof(__v0) =3D=3D 16 && __have_a=
vx2)
+		  || (sizeof(__v0) =3D=3D 16 && __have_avx
+		      && is_floating_point_v<_Tp>)
+		  || (sizeof(__v0) =3D=3D 32 && __have_avx512f))
+      {
+	// The ISA can handle wider input registers, so concat and use two-arg
+	// implementation. This reduces code duplication considerably.
+	return __convert_x86<_To>(__concat(__v0, __v1), __concat(__v2, __v3));
+      }
+    else //{{{2
+      {
+	// conversion using bit reinterpretation (or no conversion at all)
+	// should all go through the concat branch above:
+	static_assert(
+	  !(is_floating_point_v<
+	      _Tp> =3D=3D is_floating_point_v<_Up> && sizeof(_Tp) =3D=3D sizeof(_=
Up)));
+	// handle all zero extension{{{2
+	if constexpr (4 * _Np < _M && sizeof(_To) > 16)
+	  {
+	    constexpr size_t Min =3D 16 / sizeof(_Up);
+	    return __zero_extend(
+	      __convert_x86<
+		__vector_type_t<_Up, (Min > 4 * _Np) ? Min : 4 * _Np>>(
+		__v0, __v1, __v2, __v3));
+	  }
+	else if constexpr (__i64_to_i16) //{{{2
+	  {
+	    if constexpr (__x_to_x && __have_sse4_1)
+	      {
+		return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+		  _mm_blend_epi16(
+		    _mm_blend_epi16(__i0, _mm_slli_si128(__i1, 2), 0x22),
+		    _mm_blend_epi16(_mm_slli_si128(__i2, 4),
+				    _mm_slli_si128(__i3, 6), 0x88),
+		    0xcc),
+		  _mm_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7,
+				14, 15)));
+	      }
+	    else if constexpr (__y_to_y && __have_avx2)
+	      {
+		return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
+		  __xzyw(_mm256_blend_epi16(
+		    __auto_bitcast(
+		      _mm256_shuffle_ps(__vector_bitcast<float>(__v0),
+					__vector_bitcast<float>(__v2),
+					0x88)), // 0.1. 8.9. 2.3. A.B.
+		    __to_intrin(__vector_bitcast<int>(_mm256_shuffle_ps(
+				  __vector_bitcast<float>(__v1),
+				  __vector_bitcast<float>(__v3), 0x88))
+				<< 16), // .4.5 .C.D .6.7 .E.F
+		    0xaa)               // 0415 8C9D 2637 AEBF
+			 ),             // 0415 2637 8C9D AEBF
+		  _mm256_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11,
+				   14, 15, 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7,
+				   10, 11, 14, 15)));
+		/*
+		auto __a =3D _mm256_unpacklo_epi16(__v0, __v1);  // 04.. .... 26..
+		.... auto __b =3D _mm256_unpackhi_epi16(__v0, __v1);  // 15..
+		.... 37.. .... auto __c =3D _mm256_unpacklo_epi16(__v2, __v3);  //
+		8C.. .... AE.. .... auto __d =3D _mm256_unpackhi_epi16(__v2,
+		__v3);
+		// 9D.. .... BF.. .... auto __e =3D _mm256_unpacklo_epi16(__a,
+		__b);
+		// 0145 .... 2367 .... auto __f =3D _mm256_unpacklo_epi16(__c,
+		__d);
+		// 89CD .... ABEF .... auto __g =3D _mm256_unpacklo_epi64(__e,
+		__f);
+		// 0145 89CD 2367 ABEF return __concat(
+		    _mm_unpacklo_epi32(__lo128(__g), __hi128(__g)),
+		    _mm_unpackhi_epi32(__lo128(__g), __hi128(__g)));  // 0123
+		4567 89AB CDEF
+		    */
+	      } // else use fallback
+	  }
+	else if constexpr (__i64_to_i8) //{{{2
+	  {
+	    if constexpr (__x_to_x)
+	      {
+		// TODO: use fallback for now
+	      }
+	    else if constexpr (__y_to_x)
+	      {
+		auto __a
+		  =3D _mm256_srli_epi32(_mm256_slli_epi32(__i0, 24), 24)
+		    | _mm256_srli_epi32(_mm256_slli_epi32(__i1, 24), 16)
+		    | _mm256_srli_epi32(_mm256_slli_epi32(__i2, 24), 8)
+		    | _mm256_slli_epi32(
+		      __i3, 24); // 048C .... 159D .... 26AE .... 37BF ....
+		/*return _mm_shuffle_epi8(
+		    _mm_blend_epi32(__lo128(__a) << 32, __hi128(__a), 0x5),
+		    _mm_setr_epi8(4, 12, 0, 8, 5, 13, 1, 9, 6, 14, 2, 10, 7, 15,
+		   3, 11));*/
+		auto __b =3D _mm256_unpackhi_epi64(
+		  __a, __a); // 159D .... 159D .... 37BF .... 37BF ....
+		auto __c =3D _mm256_unpacklo_epi8(
+		  __a, __b); // 0145 89CD .... .... 2367 ABEF .... ....
+		return __intrin_bitcast<_To>(
+		  _mm_unpacklo_epi16(__lo128(__c),
+				     __hi128(__c))); // 0123 4567 89AB CDEF
+	      }
+	  }
+	else if constexpr (__i32_to_i8) //{{{2
+	  {
+	    if constexpr (__x_to_x)
+	      {
+		if constexpr (__have_ssse3)
+		  {
+		    const auto __x0 =3D __vector_bitcast<_UInt>(__v0) & 0xff;
+		    const auto __x1 =3D (__vector_bitcast<_UInt>(__v1) & 0xff)
+				      << 8;
+		    const auto __x2 =3D (__vector_bitcast<_UInt>(__v2) & 0xff)
+				      << 16;
+		    const auto __x3 =3D __vector_bitcast<_UInt>(__v3) << 24;
+		    return __intrin_bitcast<_To>(
+		      _mm_shuffle_epi8(__to_intrin(__x0 | __x1 | __x2 | __x3),
+				       _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13,
+						     2, 6, 10, 14, 3, 7, 11,
+						     15)));
+		  }
+		else
+		  {
+		    auto __a
+		      =3D _mm_unpacklo_epi8(__i0, __i2); // 08.. .... 19.. ....
+		    auto __b
+		      =3D _mm_unpackhi_epi8(__i0, __i2); // 2A.. .... 3B.. ....
+		    auto __c
+		      =3D _mm_unpacklo_epi8(__i1, __i3); // 4C.. .... 5D.. ....
+		    auto __d
+		      =3D _mm_unpackhi_epi8(__i1, __i3); // 6E.. .... 7F.. ....
+		    auto __e
+		      =3D _mm_unpacklo_epi8(__a, __c); // 048C .... .... ....
+		    auto __f
+		      =3D _mm_unpackhi_epi8(__a, __c); // 159D .... .... ....
+		    auto __g
+		      =3D _mm_unpacklo_epi8(__b, __d); // 26AE .... .... ....
+		    auto __h
+		      =3D _mm_unpackhi_epi8(__b, __d); // 37BF .... .... ....
+		    return __intrin_bitcast<_To>(_mm_unpacklo_epi8(
+		      _mm_unpacklo_epi8(__e, __g), // 0246 8ACE .... ....
+		      _mm_unpacklo_epi8(__f, __h)  // 1357 9BDF .... ....
+		      ));                          // 0123 4567 89AB CDEF
+		  }
+	      }
+	    else if constexpr (__y_to_y)
+	      {
+		const auto __a =3D _mm256_shuffle_epi8(
+		  __to_intrin((__vector_bitcast<_UShort>(_mm256_blend_epi16(
+				 __i0, _mm256_slli_epi32(__i1, 16), 0xAA))
+			       & 0xff)
+			      | (__vector_bitcast<_UShort>(_mm256_blend_epi16(
+				   __i2, _mm256_slli_epi32(__i3, 16), 0xAA))
+				 << 8)),
+		  _mm256_setr_epi8(0, 4, 8, 12, 2, 6, 10, 14, 1, 5, 9, 13, 3, 7,
+				   11, 15, 0, 4, 8, 12, 2, 6, 10, 14, 1, 5, 9,
+				   13, 3, 7, 11, 15));
+		return __intrin_bitcast<_To>(_mm256_permutevar8x32_epi32(
+		  __a, _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7)));
+	      }
+	  }
+	else if constexpr (__i64_to_f32) //{{{2
+	  {
+	    // this branch is only relevant with AVX and w/o AVX2 (i.e. no ymm
+	    // integers)
+	    if constexpr (__x_to_y)
+	      {
+		return __make_wrapper<float>(__v0[0], __v0[1], __v1[0], __v1[1],
+					     __v2[0], __v2[1], __v3[0],
+					     __v3[1]);
+
+		const auto __a =3D _mm_unpacklo_epi32(__i0, __i1);   // acAC
+		const auto __b =3D _mm_unpackhi_epi32(__i0, __i1);   // bdBD
+		const auto __c =3D _mm_unpacklo_epi32(__i2, __i3);   // egEG
+		const auto __d =3D _mm_unpackhi_epi32(__i2, __i3);   // fhFH
+		const auto __lo32a =3D _mm_unpacklo_epi32(__a, __b); // abcd
+		const auto __lo32b =3D _mm_unpacklo_epi32(__c, __d); // efgh
+		const auto __hi32 =3D __vector_bitcast<
+		  conditional_t<is_signed_v<_Tp>, int, _UInt>>(
+		  __concat(_mm_unpackhi_epi32(__a, __b),
+			   _mm_unpackhi_epi32(__c, __d))); // ABCD EFGH
+		const auto __hi
+		  =3D 0x100000000LL
+		    * __convert_x86<__vector_type_t<float, 8>>(__hi32);
+		const auto __mid
+		  =3D 0x10000
+		    * _mm256_cvtepi32_ps(__concat(_mm_srli_epi32(__lo32a, 16),
+						  _mm_srli_epi32(__lo32b, 16)));
+		const auto __lo =3D _mm256_cvtepi32_ps(
+		  __concat(_mm_set1_epi32(0x0000ffffu) & __lo32a,
+			   _mm_set1_epi32(0x0000ffffu) & __lo32b));
+		return (__hi + __mid) + __lo;
+	      }
+	  }
+	else if constexpr (__f64_to_ibw) //{{{2
+	  {
+	    return __convert_x86<_To>(
+	      __convert_x86<__vector_type_t<int, _Np * 2>>(__v0, __v1),
+	      __convert_x86<__vector_type_t<int, _Np * 2>>(__v2, __v3));
+	  }
+	else if constexpr (__f32_to_ibw) //{{{2
+	  {
+	    return __convert_x86<_To>(
+	      __convert_x86<__vector_type_t<int, _Np>>(__v0),
+	      __convert_x86<__vector_type_t<int, _Np>>(__v1),
+	      __convert_x86<__vector_type_t<int, _Np>>(__v2),
+	      __convert_x86<__vector_type_t<int, _Np>>(__v3));
+	  } //}}}
+
+	// fallback: {{{2
+	if constexpr (sizeof(_To) >=3D 32)
+	  // if _To is ymm or zmm, then _SimdWrapper<_Up, _M / 2> is xmm or ymm
+	  return __concat(__convert_x86<__vector_type_t<_Up, _M / 2>>(__v0,
+								      __v1),
+			  __convert_x86<__vector_type_t<_Up, _M / 2>>(__v2,
+								      __v3));
+	else if constexpr (sizeof(_To) =3D=3D 16)
+	  {
+	    const auto __lo =3D __to_intrin(__convert_x86<_To>(__v0, __v1));
+	    const auto __hi =3D __to_intrin(__convert_x86<_To>(__v2, __v3));
+	    if constexpr (sizeof(_Up) * _Np * 2 =3D=3D 8)
+	      {
+		if constexpr (is_floating_point_v<_Up>)
+		  return __auto_bitcast(_mm_unpacklo_pd(__lo, __hi));
+		else
+		  return __intrin_bitcast<_To>(_mm_unpacklo_epi64(__lo, __hi));
+	      }
+	    else if constexpr (sizeof(_Up) * _Np * 2 =3D=3D 4)
+	      {
+		if constexpr (is_floating_point_v<_Up>)
+		  return __auto_bitcast(_mm_unpacklo_ps(__lo, __hi));
+		else
+		  return __intrin_bitcast<_To>(_mm_unpacklo_epi32(__lo, __hi));
+	      }
+	    else
+	      __assert_unreachable<_Tp>();
+	  }
+	else
+	  return __vector_convert<_To>(__v0, __v1, __v2, __v3,
+				       make_index_sequence<_Np>());
+	//}}}2
+      }
+  }
+
+//}}}
+// 8-arg __convert_x86 {{{1
+template <typename _To, typename _V, typename _Traits>
+  _GLIBCXX_SIMD_INTRINSIC _To
+  __convert_x86(_V __v0, _V __v1, _V __v2, _V __v3, _V __v4, _V __v5, _V _=
_v6,
+		_V __v7)
+  {
+    static_assert(__is_vector_type_v<_V>);
+    using _Tp =3D typename _Traits::value_type;
+    constexpr size_t _Np =3D _Traits::_S_full_size;
+    [[maybe_unused]] const auto __i0 =3D __to_intrin(__v0);
+    [[maybe_unused]] const auto __i1 =3D __to_intrin(__v1);
+    [[maybe_unused]] const auto __i2 =3D __to_intrin(__v2);
+    [[maybe_unused]] const auto __i3 =3D __to_intrin(__v3);
+    [[maybe_unused]] const auto __i4 =3D __to_intrin(__v4);
+    [[maybe_unused]] const auto __i5 =3D __to_intrin(__v5);
+    [[maybe_unused]] const auto __i6 =3D __to_intrin(__v6);
+    [[maybe_unused]] const auto __i7 =3D __to_intrin(__v7);
+    using _Up =3D typename _VectorTraits<_To>::value_type;
+    constexpr size_t _M =3D _VectorTraits<_To>::_S_full_size;
+
+    static_assert(8 * _Np <=3D _M,
+		  "__v4-__v7 would be discarded; use the four/two/one-argument "
+		  "__convert_x86 overload instead");
+
+    // [xyz]_to_[xyz] {{{2
+    [[maybe_unused]] constexpr bool __x_to_x
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __x_to_y
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __x_to_z
+      =3D sizeof(__v0) <=3D 16 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __y_to_x
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __y_to_y
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __y_to_z
+      =3D sizeof(__v0) =3D=3D 32 && sizeof(_To) =3D=3D 64;
+    [[maybe_unused]] constexpr bool __z_to_x
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) <=3D 16;
+    [[maybe_unused]] constexpr bool __z_to_y
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) =3D=3D 32;
+    [[maybe_unused]] constexpr bool __z_to_z
+      =3D sizeof(__v0) =3D=3D 64 && sizeof(_To) =3D=3D 64;
+
+    // [if]X_to_i8 {{{2
+    [[maybe_unused]] constexpr bool __i_to_i
+      =3D is_integral_v<_Up> && is_integral_v<_Tp>;
+    [[maybe_unused]] constexpr bool __i64_to_i8
+      =3D __i_to_i && sizeof(_Tp) =3D=3D 8 && sizeof(_Up) =3D=3D 1;
+    [[maybe_unused]] constexpr bool __f64_to_i8
+      =3D is_integral_v<_Up> && sizeof(_Up) =3D=3D 1
+	&& is_floating_point_v<_Tp> && sizeof(_Tp) =3D=3D 8;
+
+    if constexpr (__i_to_i) // assert ISA {{{2
+      {
+	static_assert(__x_to_x || __have_avx2,
+		      "integral conversions with ymm registers require AVX2");
+	static_assert(__have_avx512bw
+			|| ((sizeof(_Tp) >=3D 4 || sizeof(__v0) < 64)
+			    && (sizeof(_Up) >=3D 4 || sizeof(_To) < 64)),
+		      "8/16-bit integers in zmm registers require AVX512BW");
+	static_assert((sizeof(__v0) < 64 && sizeof(_To) < 64) || __have_avx512f,
+		      "integral conversions with ymm registers require AVX2");
+      }
+    // concat =3D> use 4-arg __convert_x86 {{{2
+    if constexpr (sizeof(__v0) < 16 || (sizeof(__v0) =3D=3D 16 && __have_a=
vx2)
+		  || (sizeof(__v0) =3D=3D 16 && __have_avx
+		      && is_floating_point_v<_Tp>)
+		  || (sizeof(__v0) =3D=3D 32 && __have_avx512f))
+      {
+	// The ISA can handle wider input registers, so concat and use two-arg
+	// implementation. This reduces code duplication considerably.
+	return __convert_x86<_To>(__concat(__v0, __v1), __concat(__v2, __v3),
+				  __concat(__v4, __v5), __concat(__v6, __v7));
+      }
+    else //{{{2
+      {
+	// conversion using bit reinterpretation (or no conversion at all)
+	// should all go through the concat branch above:
+	static_assert(
+	  !(is_floating_point_v<
+	      _Tp> =3D=3D is_floating_point_v<_Up> && sizeof(_Tp) =3D=3D sizeof(_=
Up)));
+	static_assert(!(8 * _Np < _M && sizeof(_To) > 16),
+		      "zero extension should be impossible");
+	if constexpr (__i64_to_i8) //{{{2
+	  {
+	    if constexpr (__x_to_x && __have_ssse3)
+	      {
+		// unsure whether this is better than the variant below
+		return __intrin_bitcast<_To>(_mm_shuffle_epi8(
+		  __to_intrin(
+		    (((__v0 & 0xff) | ((__v1 & 0xff) << 8))
+		     | (((__v2 & 0xff) << 16) | ((__v3 & 0xff) << 24)))
+		    | ((((__v4 & 0xff) << 32) | ((__v5 & 0xff) << 40))
+		       | (((__v6 & 0xff) << 48) | (__v7 << 56)))),
+		  _mm_setr_epi8(0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14,
+				7, 15)));
+	      }
+	    else if constexpr (__x_to_x)
+	      {
+		const auto __a =3D _mm_unpacklo_epi8(__i0, __i1); // ac
+		const auto __b =3D _mm_unpackhi_epi8(__i0, __i1); // bd
+		const auto __c =3D _mm_unpacklo_epi8(__i2, __i3); // eg
+		const auto __d =3D _mm_unpackhi_epi8(__i2, __i3); // fh
+		const auto __e =3D _mm_unpacklo_epi8(__i4, __i5); // ik
+		const auto __f =3D _mm_unpackhi_epi8(__i4, __i5); // jl
+		const auto __g =3D _mm_unpacklo_epi8(__i6, __i7); // mo
+		const auto __h =3D _mm_unpackhi_epi8(__i6, __i7); // np
+		return __intrin_bitcast<_To>(_mm_unpacklo_epi64(
+		  _mm_unpacklo_epi32(_mm_unpacklo_epi8(__a, __b),  // abcd
+				     _mm_unpacklo_epi8(__c, __d)), // efgh
+		  _mm_unpacklo_epi32(_mm_unpacklo_epi8(__e, __f),  // ijkl
+				     _mm_unpacklo_epi8(__g, __h))  // mnop
+		  ));
+	      }
+	    else if constexpr (__y_to_y)
+	      {
+		auto __a =3D // 048C GKOS 159D HLPT 26AE IMQU 37BF JNRV
+		  __to_intrin(
+		    (((__v0 & 0xff) | ((__v1 & 0xff) << 8))
+		     | (((__v2 & 0xff) << 16) | ((__v3 & 0xff) << 24)))
+		    | ((((__v4 & 0xff) << 32) | ((__v5 & 0xff) << 40))
+		       | (((__v6 & 0xff) << 48) | ((__v7 << 56)))));
+		/*
+		auto __b =3D _mm256_unpackhi_epi64(__a, __a);  // 159D HLPT 159D
+		HLPT 37BF JNRV 37BF JNRV auto __c =3D _mm256_unpacklo_epi8(__a,
+		__b);  // 0145 89CD GHKL OPST 2367 ABEF IJMN QRUV auto __d =3D
+		__xzyw(__c); // 0145 89CD 2367 ABEF GHKL OPST IJMN QRUV return
+		_mm256_shuffle_epi8(
+		    __d, _mm256_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12,
+		13, 6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7,
+		14, 15));
+					*/
+		auto __b =3D _mm256_shuffle_epi8( // 0145 89CD GHKL OPST 2367 ABEF
+						// IJMN QRUV
+		  __a, _mm256_setr_epi8(0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13,
+					6, 14, 7, 15, 0, 8, 1, 9, 2, 10, 3, 11,
+					4, 12, 5, 13, 6, 14, 7, 15));
+		auto __c
+		  =3D __xzyw(__b); // 0145 89CD 2367 ABEF GHKL OPST IJMN QRUV
+		return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
+		  __c, _mm256_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13,
+					6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11,
+					4, 5, 12, 13, 6, 7, 14, 15)));
+	      }
+	    else if constexpr (__z_to_z)
+	      {
+		return __concat(
+		  __convert_x86<__vector_type_t<_Up, _M / 2>>(__v0, __v1, __v2,
+							      __v3),
+		  __convert_x86<__vector_type_t<_Up, _M / 2>>(__v4, __v5, __v6,
+							      __v7));
+	      }
+	  }
+	else if constexpr (__f64_to_i8) //{{{2
+	  {
+	    return __convert_x86<_To>(
+	      __convert_x86<__vector_type_t<int, _Np * 2>>(__v0, __v1),
+	      __convert_x86<__vector_type_t<int, _Np * 2>>(__v2, __v3),
+	      __convert_x86<__vector_type_t<int, _Np * 2>>(__v4, __v5),
+	      __convert_x86<__vector_type_t<int, _Np * 2>>(__v6, __v7));
+	  }
+	else // unreachable {{{2
+	  __assert_unreachable<_Tp>();
+	//}}}
+
+	// fallback: {{{2
+	if constexpr (sizeof(_To) >=3D 32)
+	  // if _To is ymm or zmm, then _SimdWrapper<_Up, _M / 2> is xmm or ymm
+	  return __concat(
+	    __convert_x86<__vector_type_t<_Up, _M / 2>>(__v0, __v1, __v2, __v3),
+	    __convert_x86<__vector_type_t<_Up, _M / 2>>(__v4, __v5, __v6,
+							__v7));
+	else if constexpr (sizeof(_To) =3D=3D 16)
+	  {
+	    const auto __lo
+	      =3D __to_intrin(__convert_x86<_To>(__v0, __v1, __v2, __v3));
+	    const auto __hi
+	      =3D __to_intrin(__convert_x86<_To>(__v4, __v5, __v6, __v7));
+	    static_assert(sizeof(_Up) =3D=3D 1 && _Np =3D=3D 2);
+	    return __intrin_bitcast<_To>(_mm_unpacklo_epi64(__lo, __hi));
+	  }
+	else
+	  {
+	    __assert_unreachable<_Tp>();
+	    // return __vector_convert<_To>(__v0, __v1, __v2, __v3, __v4, __v5,
+	    // __v6, __v7,
+	    //                             make_index_sequence<_Np>());
+	  } //}}}2
+      }
+  }
+
+//}}}
+// 16-arg __convert_x86 {{{1
+template <typename _To, typename _V, typename _Traits>
+  _GLIBCXX_SIMD_INTRINSIC _To
+  __convert_x86(_V __v0, _V __v1, _V __v2, _V __v3, _V __v4, _V __v5, _V _=
_v6,
+		_V __v7, _V __v8, _V __v9, _V __v10, _V __v11, _V __v12,
+		_V __v13, _V __v14, _V __v15)
+  {
+    // concat =3D> use 8-arg __convert_x86
+    return __convert_x86<_To>(__concat(__v0, __v1), __concat(__v2, __v3),
+			      __concat(__v4, __v5), __concat(__v6, __v7),
+			      __concat(__v8, __v9), __concat(__v10, __v11),
+			      __concat(__v12, __v13), __concat(__v14, __v15));
+  }
+
+//}}}
+
+#endif // __cplusplus >=3D 201703L
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD_X86_CONVERSIONS_H
+
+// vim: foldmethod=3Dmarker
diff --git a/libstdc++-v3/include/experimental/simd b/libstdc++-v3/include/=
experimental/simd
new file mode 100644
index 00000000000..43d2ba19903
=2D-- /dev/null
+++ b/libstdc++-v3/include/experimental/simd
@@ -0,0 +1,70 @@
+// Components for element-wise operations on data-parallel objects -*- C++=
 -*-
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// Under Section 7 of GPL version 3, you are granted additional
+// permissions described in the GCC Runtime Library Exception, version
+// 3.1, as published by the Free Software Foundation.
+
+// You should have received a copy of the GNU General Public License and
+// a copy of the GCC Runtime Library Exception along with this program;
+// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+// <http://www.gnu.org/licenses/>.
+
+/** @file experimental/simd
+ *  This is a TS C++ Library header.
+ */
+
+//
+// N4773 =C2=A79 data-parallel types library
+//
+
+#ifndef _GLIBCXX_EXPERIMENTAL_SIMD
+#define _GLIBCXX_EXPERIMENTAL_SIMD
+
+#define __cpp_lib_experimental_parallel_simd 201803
+
+#pragma GCC diagnostic push
+// Many [[gnu::vector_size(N)]] types might lead to a -Wpsabi warning whic=
h is
+// irrelevant as those functions never appear on ABI borders
+#ifndef __clang__
+#pragma GCC diagnostic ignored "-Wpsabi"
+#endif
+
+// If __OPTIMIZE__ is not defined some intrinsics are defined as macros, m=
aking
+// use of C casts internally. This requires us to disable the warning as it
+// would otherwise yield many false positives.
+#ifndef __OPTIMIZE__
+#pragma GCC diagnostic ignored "-Wold-style-cast"
+#endif
+
+#include "bits/simd_detail.h"
+#include "bits/simd.h"
+#include "bits/simd_fixed_size.h"
+#include "bits/simd_scalar.h"
+#include "bits/simd_builtin.h"
+#include "bits/simd_converter.h"
+#if _GLIBCXX_SIMD_X86INTRIN
+#include "bits/simd_x86.h"
+#elif _GLIBCXX_SIMD_HAVE_NEON
+#include "bits/simd_neon.h"
+#elif __ALTIVEC__
+#include "bits/simd_ppc.h"
+#endif
+#include "bits/simd_math.h"
+
+#pragma GCC diagnostic pop
+
+#endif // _GLIBCXX_EXPERIMENTAL_SIMD
+// vim: ft=3Dcpp
diff --git a/libstdc++-v3/testsuite/experimental/simd/standard_abi_usable.c=
c b/libstdc++-v3/testsuite/experimental/simd/standard_abi_usable.cc
new file mode 100644
index 00000000000..b055eee9c04
=2D-- /dev/null
+++ b/libstdc++-v3/testsuite/experimental/simd/standard_abi_usable.cc
@@ -0,0 +1,64 @@
+// { dg-options "-std=3Dc++17 -fno-fast-math" }
+// { dg-do compile { target c++17 } }
+
+// Copyright (C) 2020 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3.  If not see
+// <http://www.gnu.org/licenses/>.
+
+#include <experimental/simd>
+
+template <typename V>
+  void
+  is_usable()
+  {
+    static_assert(std::is_default_constructible_v<V>);
+    static_assert(std::is_destructible_v         <V>);
+    static_assert(std::is_default_constructible_v<typename V::mask_type>);
+    static_assert(std::is_destructible_v         <typename V::mask_type>);
+  }
+
+template <typename T>
+  void
+  test01()
+  {
+    namespace stdx =3D std::experimental;
+    is_usable<stdx::simd<T>>();
+    is_usable<stdx::native_simd<T>>();
+    is_usable<stdx::fixed_size_simd<T, 3>>();
+    is_usable<stdx::fixed_size_simd<T, stdx::simd_abi::max_fixed_size<T>>>=
();
+  }
+
+int main()
+{
+  test01<char>();
+  test01<wchar_t>();
+  test01<char16_t>();
+  test01<char32_t>();
+
+  test01<signed char>();
+  test01<unsigned char>();
+  test01<short>();
+  test01<unsigned short>();
+  test01<int>();
+  test01<unsigned int>();
+  test01<long>();
+  test01<unsigned long>();
+  test01<long long>();
+  test01<unsigned long long>();
+  test01<float>();
+  test01<double>();
+  test01<long double>();
+}
diff --git a/libstdc++-v3/testsuite/experimental/simd/standard_abi_usable_2=
=2Ecc b/libstdc++-v3/testsuite/experimental/simd/standard_abi_usable_2.cc
new file mode 100644
index 00000000000..a609adaf000
=2D-- /dev/null
+++ b/libstdc++-v3/testsuite/experimental/simd/standard_abi_usable_2.cc
@@ -0,0 +1,4 @@
+// { dg-options "-std=3Dc++17 -ffast-math" }
+// { dg-do compile }
+
+#include "standard_abi_usable.cc"
diff --git a/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp b/libstdc+=
+-v3/testsuite/libstdc++-dg/conformance.exp
index 11fdc8d340b..90264f5bfa9 100644
=2D-- a/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp
+++ b/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp
@@ -89,12 +89,14 @@ if {[info exists tests_file] && [file exists $tests_fil=
e]} {
 	    # 3. wchar_t tests, if not supported.
 	    # 4. thread tests, if not supported.=20
 	    # 5. *_filebuf, if file I/O is not supported.
+	    # 6. simd tests.
 	    if { [string first _xin $t] =3D=3D -1
 		 && [string first performance $t] =3D=3D -1
 		 && (${v3-wchar_t} || [string first wchar_t $t] =3D=3D -1)=20
 		 && (${v3-threads} || [string first thread $t] =3D=3D -1) =20
 		 && ([string first "_filebuf" $t] =3D=3D -1
=2D		     || [check_v3_target_fileio]) } {
+		     || [check_v3_target_fileio])
+		 && [string first "/experimental/simd/" $t] =3D=3D -1 } {
 		lappend tests $t
 	    }
 	}
@@ -107,5 +109,19 @@ global DEFAULT_CXXFLAGS
 global PCH_CXXFLAGS
 dg-runtest $tests "" "$DEFAULT_CXXFLAGS $PCH_CXXFLAGS"
=20
+# Finally run simd tests with extra SIMD-relevant flags
+global DEFAULT_VECTCFLAGS
+global EFFECTIVE_TARGETS
+set DEFAULT_VECTCFLAGS ""
+set EFFECTIVE_TARGETS ""
+
+if [check_vect_support_and_set_flags] {
+  lappend DEFAULT_VECTCFLAGS "-O2"
+  lappend DEFAULT_VECTCFLAGS "-Wno-psabi"
+  et-dg-runtest dg-runtest [lsort \
+    [glob -nocomplain $srcdir/experimental/simd/*.cc]] \
+    "$DEFAULT_VECTCFLAGS" "$DEFAULT_CXXFLAGS $PCH_CXXFLAGS"
+}
+
 # All done.
 dg-finish

--nextPart3621190.fJnDbHsrEp--