* [PATCH] [i386] Optimize bit_and op1 float_vector_all_ones_operands to op1.
@ 2021-12-16 5:59 liuhongt
2021-12-20 1:52 ` Hongtao Liu
0 siblings, 1 reply; 2+ messages in thread
From: liuhongt @ 2021-12-16 5:59 UTC (permalink / raw)
To: gcc-patches
Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
Ok for trunk?
gcc/ChangeLog:
PR target/98468
* config/i386/sse.md (*bit_and_float_vector_all_ones): New
pre-reload splitter.
gcc/testsuite/ChangeLog:
* gcc.target/i386/pr98468.c: New test.
---
gcc/config/i386/sse.md | 12 +++++++++
gcc/testsuite/gcc.target/i386/pr98468.c | 35 +++++++++++++++++++++++++
2 files changed, 47 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/i386/pr98468.c
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 5421fb51684..a715263740b 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -4432,6 +4432,18 @@ (define_insn "*<code><mode>3<mask_name>"
(const_string "<sseinsnmode>")
(const_string "XI")))])
+;; Generic part doesn't support the simpliciation of logic operation with
+;; float_vector_all_ones_operand since it's not valid rtl. Add combine spiltter
+;; for them, it should be safe since there's no SIMD Floating-Point Exceptions.
+(define_insn_and_split "*bit_and_float_vector_all_ones"
+ [(set (match_operand:VFB 0 "nonimmediate_operand")
+ (and:VFB (match_operand:VFB 1 "nonimmediate_operand")
+ (match_operand:VFB 2 "float_vector_all_ones_operand")))]
+ "TARGET_SSE && ix86_pre_reload_split ()"
+ "#"
+ "&& 1"
+ [(set (match_dup 0) (match_dup 1))])
+
(define_expand "copysign<mode>3"
[(set (match_dup 4)
(and:VFB
diff --git a/gcc/testsuite/gcc.target/i386/pr98468.c b/gcc/testsuite/gcc.target/i386/pr98468.c
new file mode 100644
index 00000000000..ca3aa256c52
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr98468.c
@@ -0,0 +1,35 @@
+/* { dg-do compile } */
+/* { dg-options "-mavx512dq -mavx512vl -O2" } */
+/* { dg-final { scan-assembler-not "vxorp" } } */
+/* { dg-final { scan-assembler-not "vandnp" } } */
+
+#include<immintrin.h>
+__m128 f(__m128 val)
+{
+ return _mm_andnot_ps(_mm_set_ps1(0.0f), val);
+}
+
+__m256 f2(__m256 val)
+{
+ return _mm256_andnot_ps(_mm256_set1_ps(0.0f), val);
+}
+
+__m512 f3(__m512 val)
+{
+ return _mm512_andnot_ps(_mm512_set1_ps(0.0f), val);
+}
+
+__m128d f4(__m128d val)
+{
+ return _mm_andnot_pd(_mm_set_pd1(0.0), val);
+}
+
+__m256d f5(__m256d val)
+{
+ return _mm256_andnot_pd(_mm256_set1_pd(0.0), val);
+}
+
+__m512d f6(__m512d val)
+{
+ return _mm512_andnot_pd(_mm512_set1_pd(0.0), val);
+}
--
2.18.1
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] [i386] Optimize bit_and op1 float_vector_all_ones_operands to op1.
2021-12-16 5:59 [PATCH] [i386] Optimize bit_and op1 float_vector_all_ones_operands to op1 liuhongt
@ 2021-12-20 1:52 ` Hongtao Liu
0 siblings, 0 replies; 2+ messages in thread
From: Hongtao Liu @ 2021-12-20 1:52 UTC (permalink / raw)
To: liuhongt; +Cc: GCC Patches, H. J. Lu
On Thu, Dec 16, 2021 at 1:59 PM liuhongt <hongtao.liu@intel.com> wrote:
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?
Pushed to trunk.
>
> gcc/ChangeLog:
>
> PR target/98468
> * config/i386/sse.md (*bit_and_float_vector_all_ones): New
> pre-reload splitter.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr98468.c: New test.
> ---
> gcc/config/i386/sse.md | 12 +++++++++
> gcc/testsuite/gcc.target/i386/pr98468.c | 35 +++++++++++++++++++++++++
> 2 files changed, 47 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr98468.c
>
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index 5421fb51684..a715263740b 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -4432,6 +4432,18 @@ (define_insn "*<code><mode>3<mask_name>"
> (const_string "<sseinsnmode>")
> (const_string "XI")))])
>
> +;; Generic part doesn't support the simpliciation of logic operation with
> +;; float_vector_all_ones_operand since it's not valid rtl. Add combine spiltter
> +;; for them, it should be safe since there's no SIMD Floating-Point Exceptions.
> +(define_insn_and_split "*bit_and_float_vector_all_ones"
> + [(set (match_operand:VFB 0 "nonimmediate_operand")
> + (and:VFB (match_operand:VFB 1 "nonimmediate_operand")
> + (match_operand:VFB 2 "float_vector_all_ones_operand")))]
> + "TARGET_SSE && ix86_pre_reload_split ()"
> + "#"
> + "&& 1"
> + [(set (match_dup 0) (match_dup 1))])
> +
> (define_expand "copysign<mode>3"
> [(set (match_dup 4)
> (and:VFB
> diff --git a/gcc/testsuite/gcc.target/i386/pr98468.c b/gcc/testsuite/gcc.target/i386/pr98468.c
> new file mode 100644
> index 00000000000..ca3aa256c52
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr98468.c
> @@ -0,0 +1,35 @@
> +/* { dg-do compile } */
> +/* { dg-options "-mavx512dq -mavx512vl -O2" } */
> +/* { dg-final { scan-assembler-not "vxorp" } } */
> +/* { dg-final { scan-assembler-not "vandnp" } } */
> +
> +#include<immintrin.h>
> +__m128 f(__m128 val)
> +{
> + return _mm_andnot_ps(_mm_set_ps1(0.0f), val);
> +}
> +
> +__m256 f2(__m256 val)
> +{
> + return _mm256_andnot_ps(_mm256_set1_ps(0.0f), val);
> +}
> +
> +__m512 f3(__m512 val)
> +{
> + return _mm512_andnot_ps(_mm512_set1_ps(0.0f), val);
> +}
> +
> +__m128d f4(__m128d val)
> +{
> + return _mm_andnot_pd(_mm_set_pd1(0.0), val);
> +}
> +
> +__m256d f5(__m256d val)
> +{
> + return _mm256_andnot_pd(_mm256_set1_pd(0.0), val);
> +}
> +
> +__m512d f6(__m512d val)
> +{
> + return _mm512_andnot_pd(_mm512_set1_pd(0.0), val);
> +}
> --
> 2.18.1
>
--
BR,
Hongtao
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2021-12-20 1:52 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-12-16 5:59 [PATCH] [i386] Optimize bit_and op1 float_vector_all_ones_operands to op1 liuhongt
2021-12-20 1:52 ` Hongtao Liu
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).