This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Improve XMM16-31 handling in various *vec_dup* patterns
- 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 22:58:02 +0200
- Subject: [PATCH] Improve XMM16-31 handling in various *vec_dup* patterns
- Authentication-results: sourceware.org; auth=none
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
These instructions are available in AVX512VL, so we can use
XMM16+ in there.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2016-05-18 Jakub Jelinek <jakub@redhat.com>
* config/i386/sse.md (avx2_vec_dupv4df): Use v instead of x
constraint, use maybe_evex prefix instead of vex.
(vec_dupv4sf): Use v constraint instead of x for output
operand except for noavx alternative, use Yv constraint
instead of x for input. Use maybe_evex prefix instead of vex.
(*vec_dupv4si): Likewise.
(*vec_dupv2di): Likewise.
* gcc.target/i386/avx512vl-vbroadcast-1.c: New test.
--- gcc/config/i386/sse.md.jj 2016-05-18 11:24:21.000000000 +0200
+++ gcc/config/i386/sse.md 2016-05-18 12:30:50.929220572 +0200
@@ -16880,15 +16880,15 @@ (define_insn "avx2_permv2ti"
(set_attr "mode" "OI")])
(define_insn "avx2_vec_dupv4df"
- [(set (match_operand:V4DF 0 "register_operand" "=x")
+ [(set (match_operand:V4DF 0 "register_operand" "=v")
(vec_duplicate:V4DF
(vec_select:DF
- (match_operand:V2DF 1 "register_operand" "x")
+ (match_operand:V2DF 1 "register_operand" "v")
(parallel [(const_int 0)]))))]
"TARGET_AVX2"
"vbroadcastsd\t{%1, %0|%0, %1}"
[(set_attr "type" "sselog1")
- (set_attr "prefix" "vex")
+ (set_attr "prefix" "maybe_evex")
(set_attr "mode" "V4DF")])
(define_insn "<avx512>_vec_dup<mode>_1"
@@ -16991,9 +16991,9 @@ (define_insn "<mask_codefor><avx512>_vec
(const_int 1)))])
(define_insn "vec_dupv4sf"
- [(set (match_operand:V4SF 0 "register_operand" "=x,x,x")
+ [(set (match_operand:V4SF 0 "register_operand" "=v,v,x")
(vec_duplicate:V4SF
- (match_operand:SF 1 "nonimmediate_operand" "x,m,0")))]
+ (match_operand:SF 1 "nonimmediate_operand" "Yv,m,0")))]
"TARGET_SSE"
"@
vshufps\t{$0, %1, %1, %0|%0, %1, %1, 0}
@@ -17003,13 +17003,13 @@ (define_insn "vec_dupv4sf"
(set_attr "type" "sseshuf1,ssemov,sseshuf1")
(set_attr "length_immediate" "1,0,1")
(set_attr "prefix_extra" "0,1,*")
- (set_attr "prefix" "vex,vex,orig")
+ (set_attr "prefix" "maybe_evex,maybe_evex,orig")
(set_attr "mode" "V4SF")])
(define_insn "*vec_dupv4si"
- [(set (match_operand:V4SI 0 "register_operand" "=x,x,x")
+ [(set (match_operand:V4SI 0 "register_operand" "=v,v,x")
(vec_duplicate:V4SI
- (match_operand:SI 1 "nonimmediate_operand" " x,m,0")))]
+ (match_operand:SI 1 "nonimmediate_operand" "Yv,m,0")))]
"TARGET_SSE"
"@
%vpshufd\t{$0, %1, %0|%0, %1, 0}
@@ -17019,13 +17019,13 @@ (define_insn "*vec_dupv4si"
(set_attr "type" "sselog1,ssemov,sselog1")
(set_attr "length_immediate" "1,0,1")
(set_attr "prefix_extra" "0,1,*")
- (set_attr "prefix" "maybe_vex,vex,orig")
+ (set_attr "prefix" "maybe_vex,maybe_evex,orig")
(set_attr "mode" "TI,V4SF,V4SF")])
(define_insn "*vec_dupv2di"
- [(set (match_operand:V2DI 0 "register_operand" "=x,x,x,x")
+ [(set (match_operand:V2DI 0 "register_operand" "=x,v,v,x")
(vec_duplicate:V2DI
- (match_operand:DI 1 "nonimmediate_operand" " 0,x,m,0")))]
+ (match_operand:DI 1 "nonimmediate_operand" " 0,Yv,m,0")))]
"TARGET_SSE"
"@
punpcklqdq\t%0, %0
@@ -17034,7 +17034,7 @@ (define_insn "*vec_dupv2di"
movlhps\t%0, %0"
[(set_attr "isa" "sse2_noavx,avx,sse3,noavx")
(set_attr "type" "sselog1,sselog1,sselog1,ssemov")
- (set_attr "prefix" "orig,vex,maybe_vex,orig")
+ (set_attr "prefix" "orig,maybe_evex,maybe_vex,orig")
(set_attr "mode" "TI,TI,DF,V4SF")])
(define_insn "avx2_vbroadcasti128_<mode>"
--- gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-1.c.jj 2016-05-18 12:31:29.486693255 +0200
+++ gcc/testsuite/gcc.target/i386/avx512vl-vbroadcast-1.c 2016-05-18 12:33:41.202891888 +0200
@@ -0,0 +1,41 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2 -mavx512vl" } */
+
+#include <x86intrin.h>
+
+void
+f1 (__m128d x)
+{
+ register __m128d a __asm ("xmm16");
+ register __m256d b __asm ("xmm17");
+ a = x;
+ asm volatile ("" : "+v" (a));
+ b = _mm256_broadcastsd_pd (a);
+ asm volatile ("" : "+v" (b));
+}
+
+/* { dg-final { scan-assembler "vbroadcastsd\[^\n\r]*(xmm16\[^\n\r]*ymm17|ymm17\[^\n\r]*xmm16)" } } */
+
+void
+f2 (float const *x)
+{
+ register __m128 a __asm ("xmm16");
+ a = _mm_broadcast_ss (x);
+ asm volatile ("" : "+v" (a));
+}
+
+/* { dg-final { scan-assembler "vbroadcastss\[^\n\r]*(\\)\[^\n\r]*xmm16|xmm16\[^\n\r]*PTR)" } } */
+
+void
+f3 (float x)
+{
+ register float a __asm ("xmm16");
+ register __m128 b __asm ("xmm17");
+ a = x;
+ asm volatile ("" : "+v" (a));
+ float c = a;
+ b = _mm_broadcast_ss (&c);
+ asm volatile ("" : "+v" (b));
+}
+
+/* { dg-final { scan-assembler "vbroadcastss\[^\n\r]*xmm1\[67]\[^\n\r]*xmm1\[67]" } } */
Jakub