public inbox for libstdc++@gcc.gnu.org
 help / color / mirror / Atom feed
From: Matthias Kretz <m.kretz@gsi.de>
To: Matthias Kretz via Libstdc++ <libstdc++@gcc.gnu.org>,
	<gcc-patches@gcc.gnu.org>, Alexandre Oliva <oliva@adacore.com>
Subject: Re: [PATCH] libstdc++: Correct NTTP and simd_mask ctor call
Date: Fri, 2 Jun 2023 10:50:33 +0200	[thread overview]
Message-ID: <4768566.iIbC2pHGDl@minbar> (raw)
In-Reply-To: <orwn0m1cyf.fsf@lxoliva.fsfla.org>

Hello Alexandre,

On Friday, 2 June 2023 10:32:40 CEST Alexandre Oliva wrote:
> On May 26, 2023, Matthias Kretz via Libstdc++ <libstdc++@gcc.gnu.org> wrote:
> > OK for master and all backports (after 11.4 is done)?
> > tested on powerpc64le-linux-gnu and x86_64-pc-linux-gnu
> > 
> > 	* testsuite/experimental/simd/pr109822_cast_functions.cc: New
> > 	test.
> 
> This testcase fails to compile on PowerPC targets without VSX: 64-bit
> integer and floating-point types cannot be vectorized.

Yes, and the simd implementation already encodes that both in 
__vectorized_sizeof<T>() and __intrinsic_type<T>.

> I wonder if the test is malformed (and should be amended to test for
> available simd types), or whether a patch like this would be desirable
> to make simd constructs more portable.  I'm not sure about the
> requirements.

The test is correct. The stdx::simd implementation has a latent bug (my 
dejagnu boards included only POWER7-POWER9; I'm at POWER5-POWER10 by now). The 
_S_store function is trying to work around bad code-gen but fails to notice 
that long long vectors can't be used.

I'm looking at that function again, also in light of recent improvements wrt. 
code-gen, and will remove that assumption, that long long is vectorizable.

__intrinsic_type_t<T> should never be T, but always the type that can be 
passed to corresponding platform intrinsics. There are traits for the 
implementation to detect whether the intrinsics types are available.

- Matthias

> 
> 
> [libstdc++] [simd] [ppc] use nonvector intrinsic fallback types
> 
> From: Alexandre Oliva <oliva@adacore.com>
> 
> Compiling such tests as pr109822_cast_functions.cc on powerpc targets
> that don't support VSX fails because some intrinsic types that are
> expected to be vectorizable are not defined without VSX.
> 
> Introduce fallback non-vector types to enable the code to compile.
> 
> 
> for  libstdc++-v3/ChangeLog
> 
> 	* include/experimental/bits/simd.h: Introduce fallback
> 	non-vector intrinsic_type_impl specializations for PowerPC
> 	without VSX.
> ---
>  libstdc++-v3/include/experimental/bits/simd.h |   12 ++++++++++++
>  1 file changed, 12 insertions(+)
> 
> diff --git a/libstdc++-v3/include/experimental/bits/simd.h
> b/libstdc++-v3/include/experimental/bits/simd.h index
> 834fe923065bd..2691823e869e8 100644
> --- a/libstdc++-v3/include/experimental/bits/simd.h
> +++ b/libstdc++-v3/include/experimental/bits/simd.h
> @@ -2431,9 +2431,14 @@ template <typename _Tp>
>  #define _GLIBCXX_SIMD_PPC_INTRIN(_Tp)                                      
>    \ template <>                                                           
>       \ struct __intrinsic_type_impl<_Tp> { using type = __vector _Tp; }
> +#define _GLIBCXX_SIMD_PPC_INTRIN_NOVEC(_Tp)				
       \
> +  template <>                                                              
>    \ +    struct __intrinsic_type_impl<_Tp> { using type = _Tp; }
>  _GLIBCXX_SIMD_PPC_INTRIN(float);
>  #ifdef __VSX__
>  _GLIBCXX_SIMD_PPC_INTRIN(double);
> +#else
> +_GLIBCXX_SIMD_PPC_INTRIN_NOVEC(double);
>  #endif
>  _GLIBCXX_SIMD_PPC_INTRIN(signed char);
>  _GLIBCXX_SIMD_PPC_INTRIN(unsigned char);
> @@ -2444,12 +2449,19 @@ _GLIBCXX_SIMD_PPC_INTRIN(unsigned int);
>  #if defined __VSX__ || __SIZEOF_LONG__ == 4
>  _GLIBCXX_SIMD_PPC_INTRIN(signed long);
>  _GLIBCXX_SIMD_PPC_INTRIN(unsigned long);
> +#else
> +_GLIBCXX_SIMD_PPC_INTRIN_NOVEC(signed long);
> +_GLIBCXX_SIMD_PPC_INTRIN_NOVEC(unsigned long);
>  #endif
>  #ifdef __VSX__
>  _GLIBCXX_SIMD_PPC_INTRIN(signed long long);
>  _GLIBCXX_SIMD_PPC_INTRIN(unsigned long long);
> +#else
> +_GLIBCXX_SIMD_PPC_INTRIN_NOVEC(signed long long);
> +_GLIBCXX_SIMD_PPC_INTRIN_NOVEC(unsigned long long);
>  #endif
>  #undef _GLIBCXX_SIMD_PPC_INTRIN
> +#undef _GLIBCXX_SIMD_PPC_INTRIN_NOVEC
> 
>  template <typename _Tp, size_t _Bytes>
>    struct __intrinsic_type<_Tp, _Bytes, enable_if_t<__is_vectorizable_v<_Tp>
> && _Bytes <= 16>>


-- 
──────────────────────────────────────────────────────────────────────────
 Dr. Matthias Kretz                           https://mattkretz.github.io
 GSI Helmholtz Centre for Heavy Ion Research               https://gsi.de
 stdₓ::simd
──────────────────────────────────────────────────────────────────────────




  reply	other threads:[~2023-06-02  8:50 UTC|newest]

Thread overview: 7+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-05-26 11:10 Matthias Kretz
2023-05-26 15:46 ` Jonathan Wakely
2023-06-02  8:32 ` Alexandre Oliva
2023-06-02  8:50   ` Matthias Kretz [this message]
2023-06-02  9:30     ` Alexandre Oliva
2023-06-02  9:41       ` Matthias Kretz
2023-06-02 12:06       ` Jonathan Wakely

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=4768566.iIbC2pHGDl@minbar \
    --to=m.kretz@gsi.de \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=libstdc++@gcc.gnu.org \
    --cc=oliva@adacore.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).