From: Hongtao Liu <crazylht@gmail.com>
To: "Hu, Lin1" <lin1.hu@intel.com>
Cc: gcc-patches@gcc.gnu.org, hongtao.liu@intel.com
Subject: Re: [PATCH] i386: Optimize code generation of __mm256_zextsi128_si256(__mm_set1_epi8(-1))
Date: Fri, 23 Sep 2022 09:47:33 +0800 [thread overview]
Message-ID: <CAMZc-bz8oZne7KV1BYU_AVDOchf-QbO-JxZpDZzreKfR5QZf-A@mail.gmail.com> (raw)
In-Reply-To: <20220922071944.1472123-1-lin1.hu@intel.com>
On Thu, Sep 22, 2022 at 3:20 PM Hu, Lin1 via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Hi all,
>
> This patch aims to optimize code generation of __mm256_zextsi128_si256(__mm_set1_epi8(-1)). Reduce the number of instructions required to achieve the final result.
>
> Regtested on x86_64-pc-linux-gnu. Ok for trunk?
>
> BRs,
> Lin
>
> gcc/ChangeLog:
>
> PR target/94962
> * config/i386/constraints.md (BH): New define_constraint.
> * config/i386/i386.cc (standard_sse_constant_p): Add return 3/4 when operand matches new predicate.
> (standard_sse_constant_opcode): Add new alternative branch to return "vpcmpeqd".
> * config/i386/predicates.md (vector_all_ones_zero_extend_half_operand): New define_predicate.
> (vector_all_ones_zero_extend_quarter_operand): Ditto.
> * config/i386/sse.md: Add constraint to insn "mov<mode>_internal".
(mov<mode>_internal): Add new constraint BH.
Put the insn name at first.
>
> gcc/testsuite/ChangeLog:
>
> PR target/94962
> * gcc.target/i386/avx256-unaligned-load-1.c: Modify test.
> * gcc.target/i386/avx256-unaligned-store-1.c: Ditto.
> * gcc.target/i386/avx256-unaligned-store-2.c: Ditto.
> * gcc.target/i386/avx256-unaligned-store-3.c: Ditto.
> * gcc.target/i386/pr94962-1.c: New test.
> * gcc.target/i386/pr94962-2.c: Ditto.
> * gcc.target/i386/pr94962-3.c: Ditto.
> * gcc.target/i386/pr94962-4.c: Ditto.
> ---
> gcc/config/i386/constraints.md | 8 +++
> gcc/config/i386/i386.cc | 26 +++++++-
> gcc/config/i386/predicates.md | 49 ++++++++++++++
> gcc/config/i386/sse.md | 8 +--
> .../gcc.target/i386/avx256-unaligned-load-1.c | 4 +-
> .../i386/avx256-unaligned-store-1.c | 4 +-
> .../i386/avx256-unaligned-store-2.c | 4 +-
> .../i386/avx256-unaligned-store-3.c | 4 +-
> gcc/testsuite/gcc.target/i386/pr94962-1.c | 11 ++++
> gcc/testsuite/gcc.target/i386/pr94962-2.c | 17 +++++
> gcc/testsuite/gcc.target/i386/pr94962-3.c | 64 +++++++++++++++++++
> gcc/testsuite/gcc.target/i386/pr94962-4.c | 49 ++++++++++++++
> 12 files changed, 235 insertions(+), 13 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr94962-1.c
> create mode 100644 gcc/testsuite/gcc.target/i386/pr94962-2.c
> create mode 100644 gcc/testsuite/gcc.target/i386/pr94962-3.c
> create mode 100644 gcc/testsuite/gcc.target/i386/pr94962-4.c
>
> diff --git a/gcc/config/i386/constraints.md b/gcc/config/i386/constraints.md
> index 7361687632f..95b2b142d41 100644
> --- a/gcc/config/i386/constraints.md
> +++ b/gcc/config/i386/constraints.md
> @@ -168,6 +168,9 @@
> ;; z Constant call address operand.
> ;; C Integer SSE constant with all bits set operand.
> ;; F Floating-point SSE constant with all bits set operand.
> +;; H Integer SSE constant that is 128/256bit all ones
> +;; and zero-extand to 256/512bit, or 128bit all ones
> +;; and zero-extend to 512bit.
> ;; M x86-64 memory operand.
>
> (define_constraint "Bf"
> @@ -233,6 +236,11 @@
> (and (match_test "TARGET_SSE")
> (match_operand 0 "float_vector_all_ones_operand")))
>
> +(define_constraint "BH"
> + "@internal integer constant with last half/quarter bits set operand."
> + (ior (match_operand 0 "vector_all_ones_zero_extend_half_operand")
> + (match_operand 0 "vector_all_ones_zero_extend_quarter_operand")))
> +
> ;; NB: Similar to 'm', but don't use define_memory_constraint on x86-64
> ;; to prevent LRA from converting the operand to the form '(mem (reg X))'
> ;; where X is a base register.
> diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> index dadf453d6c0..ca799da5d7e 100644
> --- a/gcc/config/i386/i386.cc
> +++ b/gcc/config/i386/i386.cc
> @@ -5186,7 +5186,8 @@ standard_80387_constant_rtx (int idx)
> XFmode);
> }
>
> -/* Return 1 if X is all bits 0 and 2 if X is all bits 1
> +/* Return 1 if X is all bits 0, 2 if X is all bits 1
> + and 3 if X is all bits 1 with zero extend
> in supported SSE/AVX vector mode. */
>
> int
> @@ -5234,6 +5235,10 @@ standard_sse_constant_p (rtx x, machine_mode pred_mode)
> }
> }
>
> + if (vector_all_ones_zero_extend_half_operand (x, mode)
> + || vector_all_ones_zero_extend_quarter_operand (x, mode))
> + return 3;
> +
> return 0;
> }
>
> @@ -5341,6 +5346,25 @@ standard_sse_constant_opcode (rtx_insn *insn, rtx *operands)
> gcc_unreachable ();
> }
> }
> + else if (vector_all_ones_zero_extend_half_operand (x, mode))
> + {
> + if (GET_MODE_SIZE (mode) == 64)
> + {
> + gcc_assert (TARGET_AVX512F);
> + return "vpcmpeqd \t %t0, %t0, %t0";
> + }
> + else if (GET_MODE_SIZE (mode) == 32)
> + {
> + gcc_assert (TARGET_AVX);
> + return "vpcmpeqd \t %x0, %x0, %x0";
> + }
> + gcc_unreachable ();
> + }
> + else if (vector_all_ones_zero_extend_quarter_operand (x, mode))
> + {
> + gcc_assert (TARGET_AVX512F);
> + return "vpcmpeqd \t %x0, %x0, %x0";
> + }
>
Can we merge 2 vpcmpeqd \t %x0, %x0, %x0"; into 1?
like
else if (vector_all_ones_zero_extend_half_operand (x, mode)
&& GET_MODE_SIZE(mode) == 64))
return "vpcmpeqd \t %t0, %t0, %t0";
else if ((vector_all_ones_zero_extend_half_operand (x, mode)
&& GET_MODE_SIZE (mode) == 32)
|| vector_all_ones_zero_extend_quarter_operand (x, mode))
return "vpcmpeqd \t %x0, %x0, %x0";
> gcc_unreachable ();
> }
> diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
> index 4f16bb748b5..655eabf793b 100644
> --- a/gcc/config/i386/predicates.md
> +++ b/gcc/config/i386/predicates.md
> @@ -1159,6 +1159,55 @@
> (match_test "INTEGRAL_MODE_P (GET_MODE (op))")
> (match_test "op == CONSTM1_RTX (GET_MODE (op))")))
>
> +/* Return true if operand is an 128/256bit all ones vector
> + that zero-extends to 256/512bit. */
> +(define_predicate "vector_all_ones_zero_extend_half_operand"
> + (match_code "const_vector")
> +{
> + mode = GET_MODE (op);
> + if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT
> + || (GET_MODE_SIZE (mode) != 32
> + && GET_MODE_SIZE (mode) != 64))
> + return false;
> +
> + int nelts = CONST_VECTOR_NUNITS (op);
> + for (int i = 0; i != nelts; i++)
> + {
> + rtx elt = CONST_VECTOR_ELT (op, i);
> + if (i < nelts / 2
> + && elt != CONSTM1_RTX (GET_MODE_INNER (mode)))
> + return false;
> + if (i >= nelts / 2
> + && elt != CONST0_RTX (GET_MODE_INNER (mode)))
> + return false;
> + }
> + return true;
> +})
> +
> +/* Return true if operand is an 128bit all ones vector
> + that zero extends to 512bit. */
> +(define_predicate "vector_all_ones_zero_extend_quarter_operand"
> + (match_code "const_vector")
> +{
> + mode = GET_MODE (op);
> + if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT
> + || GET_MODE_SIZE (mode) != 64)
> + return false;
> +
> + int nelts = CONST_VECTOR_NUNITS (op);
> + for (int i = 0; i != nelts; i++)
> + {
> + rtx elt = CONST_VECTOR_ELT (op, i);
> + if (i < nelts / 4
> + && elt != CONSTM1_RTX (GET_MODE_INNER (mode)))
> + return false;
> + if (i >= nelts / 4
> + && elt != CONST0_RTX (GET_MODE_INNER (mode)))
> + return false;
> + }
> + return true;
> +})
> +
> ; Return true when OP is operand acceptable for vector memory operand.
> ; Only AVX can have misaligned memory operand.
> (define_predicate "vector_memory_operand"
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index d535c0af043..f804dbe9b7a 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -1318,9 +1318,9 @@
>
> (define_insn "mov<mode>_internal"
> [(set (match_operand:VMOVE 0 "nonimmediate_operand"
> - "=v,v ,v ,m")
> + "=v,v ,v,v ,m")
> (match_operand:VMOVE 1 "nonimmediate_or_sse_const_operand"
> - " C,<sseconstm1>,vm,v"))]
> + " C,<sseconstm1>,BH,vm,v"))]
> "TARGET_SSE
> && (register_operand (operands[0], <MODE>mode)
> || register_operand (operands[1], <MODE>mode))
> @@ -1338,7 +1338,7 @@
> gcc_unreachable ();
> }
> }
> - [(set_attr "type" "sselog1,sselog1,ssemov,ssemov")
> + [(set_attr "type" "sselog1,sselog1,sselog1,ssemov,ssemov")
> (set_attr "prefix" "maybe_vex")
> (set (attr "mode")
> (cond [(match_test "TARGET_AVX")
> @@ -1349,7 +1349,7 @@
> (and (match_test "<MODE>mode == V2DFmode")
> (match_test "TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL"))
> (const_string "V4SF")
> - (and (eq_attr "alternative" "3")
> + (and (eq_attr "alternative" "4")
> (match_test "TARGET_SSE_TYPELESS_STORES"))
> (const_string "V4SF")
> (and (eq_attr "alternative" "0")
> diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-1.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-1.c
> index 68378a556fb..7115b0a9dde 100644
> --- a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-1.c
> +++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-1.c
> @@ -14,6 +14,6 @@ avx_test (void)
> c[i] = a[i] * b[i+3];
> }
>
> -/* { dg-final { scan-assembler-not "vmovups\[^\n\r]*movv8sf_internal/2" } } */
> -/* { dg-final { scan-assembler "movv4sf_internal/2" } } */
> +/* { dg-final { scan-assembler-not "vmovups\[^\n\r]*movv8sf_internal/3" } } */
> +/* { dg-final { scan-assembler "movv4sf_internal/3" } } */
> /* { dg-final { scan-assembler "vinsertf128" } } */
> diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-1.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-1.c
> index d82aecffda9..4c713959df2 100644
> --- a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-1.c
> +++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-1.c
> @@ -17,6 +17,6 @@ avx_test (void)
> d[i] = c[i] * 20.0;
> }
>
> -/* { dg-final { scan-assembler-not "vmovups.*movv8sf_internal/3" } } */
> -/* { dg-final { scan-assembler "vmovups.*movv4sf_internal/3" } } */
> +/* { dg-final { scan-assembler-not "vmovups.*movv8sf_internal/4" } } */
> +/* { dg-final { scan-assembler "vmovups.*movv4sf_internal/4" } } */
> /* { dg-final { scan-assembler "vextractf128" } } */
> diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-2.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-2.c
> index be12529e8d5..4978c37f526 100644
> --- a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-2.c
> +++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-2.c
> @@ -23,6 +23,6 @@ avx_test (void)
> }
> }
>
> -/* { dg-final { scan-assembler-not "vmovdqu.*movv32qi_internal/3" } } */
> -/* { dg-final { scan-assembler "vmovdqu.*movv16qi_internal/3" } } */
> +/* { dg-final { scan-assembler-not "vmovdqu.*movv32qi_internal/4" } } */
> +/* { dg-final { scan-assembler "vmovdqu.*movv16qi_internal/4" } } */
> /* { dg-final { scan-assembler "vextract.128" } } */
> diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-3.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-3.c
> index 918028df9ed..f909099bcb1 100644
> --- a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-3.c
> +++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-3.c
> @@ -17,6 +17,6 @@ avx_test (void)
> d[i] = c[i] * 20.0;
> }
>
> -/* { dg-final { scan-assembler-not "vmovupd.*movv4df_internal/3" } } */
> -/* { dg-final { scan-assembler "vmovupd.*movv2df_internal/3" } } */
> +/* { dg-final { scan-assembler-not "vmovupd.*movv4df_internal/4" } } */
> +/* { dg-final { scan-assembler "vmovupd.*movv2df_internal/4" } } */
> /* { dg-final { scan-assembler "vextractf128" } } */
> diff --git a/gcc/testsuite/gcc.target/i386/pr94962-1.c b/gcc/testsuite/gcc.target/i386/pr94962-1.c
> new file mode 100644
> index 00000000000..e3b01249421
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr94962-1.c
> @@ -0,0 +1,11 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx" } */
> +/* { dg-final { scan-assembler-not "vmovdqa" } } */
> +/* { dg-final { scan-assembler-times "vpcmpeqd\[ \\t\]+%xmm\[0-9\]" 1 } } */
> +
> +#include <immintrin.h>
> +
> +__m256i mask()
> +{
> + return _mm256_zextsi128_si256(_mm_set1_epi8(-1));
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr94962-2.c b/gcc/testsuite/gcc.target/i386/pr94962-2.c
> new file mode 100644
> index 00000000000..4e10e927ba1
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr94962-2.c
> @@ -0,0 +1,17 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx512f" } */
> +/* { dg-final { scan-assembler-not "vmovdqa" } } */
> +/* { dg-final { scan-assembler-times "vpcmpeqd\[ \\t\]+%xmm\[0-9\]" 1 } } */
> +/* { dg-final { scan-assembler-times "vpcmpeqd\[ \\t\]+%ymm\[0-9\]" 1 } } */
> +
> +#include <immintrin.h>
> +
> +__m512i mask1()
> +{
> + return _mm512_zextsi128_si512(_mm_set1_epi8(-1));
> +}
> +
> +__m512i mask2()
> +{
> + return _mm512_zextsi256_si512(_mm256_set1_epi8(-1));
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr94962-3.c b/gcc/testsuite/gcc.target/i386/pr94962-3.c
> new file mode 100644
> index 00000000000..8d0b9974435
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr94962-3.c
> @@ -0,0 +1,64 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx512f" } */
> +/* { dg-final { scan-assembler-not "vmovdqa" } } */
> +/* { dg-final { scan-assembler-times "vpcmpeqd\[ \\t\]+%ymm\[0-9\]" 4 } } */
> +
> +typedef long long __v8di __attribute__ ((__vector_size__ (64)));
> +typedef int __v16si __attribute__ ((__vector_size__ (64)));
> +typedef short __v32hi __attribute__ ((__vector_size__ (64)));
> +typedef char __v64qi __attribute__ ((__vector_size__ (64)));
> +typedef long long __m512i __attribute__ ((__vector_size__ (64), __may_alias__));
> +
> +__m512i
> +__attribute__ ((noinline, noclone))
> +foo1 ()
> +{
> + return __extension__ (__m512i)(__v8di) { -1, -1, -1, -1,
> + 0, 0, 0, 0 };
> +}
> +
> +__m512i
> +__attribute__ ((noinline, noclone))
> +foo2 ()
> +{
> + return __extension__ (__m512i)(__v16si) { -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0 };
> +}
> +
> +__m512i
> +__attribute__ ((noinline, noclone))
> +foo3 ()
> +{
> + return __extension__ (__m512i)(__v32hi) { -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0 };
> +}
> +
> +__m512i
> +__attribute__ ((noinline, noclone))
> +foo4 ()
> +{
> + return __extension__ (__m512i)(__v64qi) { -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0 };
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr94962-4.c b/gcc/testsuite/gcc.target/i386/pr94962-4.c
> new file mode 100644
> index 00000000000..5502c39910b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr94962-4.c
> @@ -0,0 +1,49 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx" } */
> +/* { dg-final { scan-assembler-not "vmovdqa" } } */
> +/* { dg-final { scan-assembler-times "vpcmpeqd\[ \\t\]+%xmm\[0-9\]" 4 } } */
> +
> +typedef long long __v4di __attribute__ ((__vector_size__ (32)));
> +typedef int __v8si __attribute__ ((__vector_size__ (32)));
> +typedef short __v16hi __attribute__ ((__vector_size__ (32)));
> +typedef char __v32qi __attribute__ ((__vector_size__ (32)));
> +typedef long long __m256i __attribute__ ((__vector_size__ (32), __may_alias__));
> +
> +__m256i
> +__attribute__ ((noinline, noclone))
> +foo1 ()
> +{
> + return __extension__ (__m256i)(__v4di) { -1, -1, 0, 0 };
> +}
> +
> +__m256i
> +__attribute__ ((noinline, noclone))
> +foo2 ()
> +{
> + return __extension__ (__m256i)(__v8si) { -1, -1, -1, -1,
> + 0, 0, 0, 0 };
> +}
> +
> +__m256i
> +__attribute__ ((noinline, noclone))
> +foo3 ()
> +{
> + return __extension__ (__m256i)(__v16hi) { -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0 };
> +}
> +
> +__m256i
> +__attribute__ ((noinline, noclone))
> +foo4 ()
> +{
> + return __extension__ (__m256i)(__v32qi) { -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + -1, -1, -1, -1,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0,
> + 0, 0, 0, 0 };
> +}
> --
> 2.18.2
>
Others LGTM.
--
BR,
Hongtao
next prev parent reply other threads:[~2022-09-23 1:44 UTC|newest]
Thread overview: 4+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-09-22 7:19 Hu, Lin1
2022-09-23 1:47 ` Hongtao Liu [this message]
2022-09-23 3:07 ` Hu, Lin1
2022-09-23 3:09 ` Hongtao Liu
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=CAMZc-bz8oZne7KV1BYU_AVDOchf-QbO-JxZpDZzreKfR5QZf-A@mail.gmail.com \
--to=crazylht@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hongtao.liu@intel.com \
--cc=lin1.hu@intel.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).