This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Improve 128-bit to 256-bit broadcasts
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Uros Bizjak <ubizjak at gmail dot com>, Kirill Yukhin <kirill dot yukhin at gmail dot com>
- Cc: gcc-patches at gcc dot gnu dot org
- Date: Wed, 18 May 2016 23:00:23 +0200
- Subject: [PATCH] Improve 128-bit to 256-bit broadcasts
- Authentication-results: sourceware.org; auth=none
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
vbroadcast[fi]32x4 and vinsert[fi]32x4 are in AVX512VL,
vbroadcast[fi]64x2 and vinsert[fi]64x2 are in AVX512VL & AVX512DQ.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2016-05-18 Jakub Jelinek <jakub@redhat.com>
* config/i386/sse.md (i128vldq): New mode iterator.
(avx2_vbroadcasti128_<mode>, avx_vbroadcastf128_<mode>): Add
avx512dq and avx512vl altenratives.
* gcc.target/i386/avx512dq-vbroadcast-2.c: New test.
* gcc.target/i386/avx512vl-vbroadcast-2.c: New test.
--- gcc/config/i386/sse.md.jj 2016-05-18 12:30:50.000000000 +0200
+++ gcc/config/i386/sse.md 2016-05-18 13:21:35.339616623 +0200
@@ -778,6 +778,12 @@ (define_mode_attr i128
(V64QI "i64x4") (V32QI "%~128") (V32HI "i64x4") (V16HI "%~128")
(V16SI "i64x4") (V8SI "%~128") (V8DI "i64x4") (V4DI "%~128")])
+;; For 256-bit modes for TARGET_AVX512VL && TARGET_AVX512DQ
+;; i32x4, f32x4, i64x2 or f64x2 suffixes.
+(define_mode_attr i128vldq
+ [(V8SF "f32x4") (V4DF "f64x2")
+ (V32QI "i32x4") (V16HI "i32x4") (V8SI "i32x4") (V4DI "i64x2")])
+
;; Mix-n-match
(define_mode_iterator AVX256MODE2P [V8SI V8SF V4DF])
(define_mode_iterator AVX512MODE2P [V16SI V16SF V8DF])
@@ -17038,15 +17044,19 @@ (define_insn "*vec_dupv2di"
(set_attr "mode" "TI,TI,DF,V4SF")])
(define_insn "avx2_vbroadcasti128_<mode>"
- [(set (match_operand:VI_256 0 "register_operand" "=x")
+ [(set (match_operand:VI_256 0 "register_operand" "=x,v,v")
(vec_concat:VI_256
- (match_operand:<ssehalfvecmode> 1 "memory_operand" "m")
+ (match_operand:<ssehalfvecmode> 1 "memory_operand" "m,m,m")
(match_dup 1)))]
"TARGET_AVX2"
- "vbroadcasti128\t{%1, %0|%0, %1}"
- [(set_attr "type" "ssemov")
+ "@
+ vbroadcasti128\t{%1, %0|%0, %1}
+ vbroadcast<i128vldq>\t{%1, %0|%0, %1}
+ vbroadcast<shuffletype>32x4\t{%1, %0|%0, %1}"
+ [(set_attr "isa" "*,avx512dq,avx512vl")
+ (set_attr "type" "ssemov")
(set_attr "prefix_extra" "1")
- (set_attr "prefix" "vex")
+ (set_attr "prefix" "vex,evex,evex")
(set_attr "mode" "OI")])
;; Modes handled by AVX vec_dup patterns.
@@ -17123,19 +17133,24 @@ (define_split
"operands[2] = gen_lowpart (<ssehalfvecmode>mode, operands[0]);")
(define_insn "avx_vbroadcastf128_<mode>"
- [(set (match_operand:V_256 0 "register_operand" "=x,x,x")
+ [(set (match_operand:V_256 0 "register_operand" "=x,x,x,v,v,v,v")
(vec_concat:V_256
- (match_operand:<ssehalfvecmode> 1 "nonimmediate_operand" "m,0,?x")
+ (match_operand:<ssehalfvecmode> 1 "nonimmediate_operand" "m,0,?x,m,0,m,0")
(match_dup 1)))]
"TARGET_AVX"
"@
vbroadcast<i128>\t{%1, %0|%0, %1}
vinsert<i128>\t{$1, %1, %0, %0|%0, %0, %1, 1}
- vperm2<i128>\t{$0, %t1, %t1, %0|%0, %t1, %t1, 0}"
- [(set_attr "type" "ssemov,sselog1,sselog1")
+ vperm2<i128>\t{$0, %t1, %t1, %0|%0, %t1, %t1, 0}
+ vbroadcast<i128vldq>\t{%1, %0|%0, %1}
+ vinsert<i128vldq>\t{$1, %1, %0, %0|%0, %0, %1, 1}
+ vbroadcast<shuffletype>32x4\t{%1, %0|%0, %1}
+ vinsert<shuffletype>32x4\t{$1, %1, %0, %0|%0, %0, %1, 1}"
+ [(set_attr "isa" "*,*,*,avx512dq,avx512dq,avx512vl,avx512vl")
+ (set_attr "type" "ssemov,sselog1,sselog1,ssemov,sselog1,ssemov,sselog1")
(set_attr "prefix_extra" "1")
- (set_attr "length_immediate" "0,1,1")
- (set_attr "prefix" "vex")
+ (set_attr "length_immediate" "0,1,1,0,1,0,1")
+ (set_attr "prefix" "vex,vex,vex,evex,evex,evex,evex")
(set_attr "mode" "<sseinsnmode>")])
;; For broadcast[i|f]32x2. Yes there is no v4sf version, only v4si.
--- gcc/testsuite/gcc.target/i386/avx512dq-vbroadcast-2.c.jj 2016-05-18 13:46:05.757523635 +0200
+++ gcc/testsuite/gcc.target/i386/avx512dq-vbroadcast-2.c 2016-05-18 13:50:31.330891648 +0200
@@ -0,0 +1,49 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2 -mavx512vl -mavx512dq" } */
+
+#include <x86intrin.h>
+
+void
+f1 (__m128i x)
+{
+ register __m128i a __asm ("xmm16");
+ register __m256i c;
+ a = x;
+ asm volatile ("" : "+v" (a));
+ c = _mm256_broadcastsi128_si256 (a);
+ register __m256i b __asm ("xmm16");
+ b = c;
+ asm volatile ("" : "+v" (b));
+}
+
+/* { dg-final { scan-assembler "vinserti64x2\[^\n\r]*(xmm16\[^\n\r]*ymm16\[^\n\r]*ymm16|ymm16\[^\n\r]*ymm16\[^\n\r]*xmm16)" } } */
+
+void
+f2 (__m128i *x)
+{
+ register __m256i a __asm ("xmm16");
+ a = _mm256_broadcastsi128_si256 (*x);
+ asm volatile ("" : "+v" (a));
+}
+
+/* { dg-final { scan-assembler "vbroadcasti64x2\[^\n\r]*ymm16" } } */
+
+void
+f3 (__m128 *x)
+{
+ register __m256 a __asm ("xmm16");
+ a = _mm256_broadcast_ps (x);
+ asm volatile ("" : "+v" (a));
+}
+
+/* { dg-final { scan-assembler "vbroadcastf32x4\[^\n\r]*ymm16" } } */
+
+void
+f4 (__m128d *x)
+{
+ register __m256d a __asm ("xmm16");
+ a = _mm256_broadcast_pd (x);
+ asm volatile ("" : "+v" (a));
+}
+
+/* { dg-final { scan-assembler "vbroadcastf64x2\[^\n\r]*ymm16" } } */
--- gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-2.c.jj 2016-05-18 13:45:40.449869743 +0200
+++ gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-2.c 2016-05-18 13:50:46.922678414 +0200
@@ -0,0 +1,47 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2 -mavx512vl -mno-avx512dq" } */
+
+#include <x86intrin.h>
+
+void
+f1 (__m128i x)
+{
+ register __m128i a __asm ("xmm16");
+ register __m256i c;
+ a = x;
+ asm volatile ("" : "+v" (a));
+ c = _mm256_broadcastsi128_si256 (a);
+ register __m256i b __asm ("xmm16");
+ b = c;
+ asm volatile ("" : "+v" (b));
+}
+
+/* { dg-final { scan-assembler "vinserti32x4\[^\n\r]*(xmm16\[^\n\r]*ymm16\[^\n\r]*ymm16|ymm16\[^\n\r]*ymm16\[^\n\r]*xmm16)" } } */
+
+void
+f2 (__m128i *x)
+{
+ register __m256i a __asm ("xmm16");
+ a = _mm256_broadcastsi128_si256 (*x);
+ asm volatile ("" : "+v" (a));
+}
+
+/* { dg-final { scan-assembler "vbroadcasti32x4\[^\n\r]*ymm16" } } */
+
+void
+f3 (__m128 *x)
+{
+ register __m256 a __asm ("xmm16");
+ a = _mm256_broadcast_ps (x);
+ asm volatile ("" : "+v" (a));
+}
+
+void
+f4 (__m128d *x)
+{
+ register __m256d a __asm ("xmm16");
+ a = _mm256_broadcast_pd (x);
+ asm volatile ("" : "+v" (a));
+}
+
+/* { dg-final { scan-assembler-times "vbroadcastf32x4\[^\n\r]*ymm16" 2 } } */
Jakub