* [PATCH] Fix vec_concatv2di pattern for SSE4 (PR target/80819)
@ 2017-11-29 8:35 Jakub Jelinek
2017-11-29 9:36 ` Uros Bizjak
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2017-11-29 8:35 UTC (permalink / raw)
To: Uros Bizjak; +Cc: gcc-patches
Hi!
Before r218303 we had just (=x,0,rm) alternative for SSE4 (no AVX),
that change turned it into (=Yr,0,*rm) and (=*x,0,rm) alternatives,
so that we avoid too many prefixes if possible.
The latter alternative is fine, we want the *, because that is the point,
Yr class is the subset of x registers that don't need the REX prefix.
The * in the first alternative makes no sense, with it IRA is effectively
forced to allocate the second vec_concat pseudo into NO_REGS - memory,
and while postreload can fix it up afterwards, we end up with dead stores
that nothing ever removes afterwards.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?
2017-11-29 Jakub Jelinek <jakub@redhat.com>
PR target/80819
* config/i386/sse.md (vec_concatv2di): Remove * from (=Yr,0,*rm)
alternative.
* gcc.target/i386/pr80819-1.c: New test.
* gcc.target/i386/pr80819-2.c: New test.
--- gcc/config/i386/sse.md.jj 2017-11-24 08:58:05.000000000 +0100
+++ gcc/config/i386/sse.md 2017-11-28 18:04:20.739396199 +0100
@@ -13915,7 +13915,7 @@ (define_insn "vec_concatv2di"
(match_operand:DI 1 "nonimmediate_operand"
" 0, 0,x ,Yv,r ,vm,?!*Yn,0,Yv,0,0,v")
(match_operand:DI 2 "vector_move_operand"
- "*rm,rm,rm,rm,C ,C ,C ,x,Yv,x,m,m")))]
+ " rm,rm,rm,rm,C ,C ,C ,x,Yv,x,m,m")))]
"TARGET_SSE"
"@
pinsrq\t{$1, %2, %0|%0, %2, 1}
--- gcc/testsuite/gcc.target/i386/pr80819-1.c.jj 2017-11-28 18:11:09.452482042 +0100
+++ gcc/testsuite/gcc.target/i386/pr80819-1.c 2017-11-28 18:09:57.000000000 +0100
@@ -0,0 +1,13 @@
+/* PR target/80819 */
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2 -msse4 -mno-avx -mtune=haswell -masm=att" } */
+
+typedef unsigned long long V __attribute__((vector_size (16)));
+
+V
+foo (unsigned long long x, unsigned long long y)
+{
+ return (V) { x, y };
+}
+
+/* { dg-final { scan-assembler-not "movq\[ \t]*%rsi, \[-0-9]*\\(" } } */
--- gcc/testsuite/gcc.target/i386/pr80819-2.c.jj 2017-11-28 18:11:15.942404034 +0100
+++ gcc/testsuite/gcc.target/i386/pr80819-2.c 2017-11-28 18:11:21.915332239 +0100
@@ -0,0 +1,13 @@
+/* PR target/80819 */
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2 -msse4 -mno-avx -mtune=generic -masm=att" } */
+
+typedef unsigned long long V __attribute__((vector_size (16)));
+
+V
+foo (unsigned long long x, unsigned long long y)
+{
+ return (V) { x, y };
+}
+
+/* { dg-final { scan-assembler-not "movq\[ \t]*%rsi, \[-0-9]*\\(" } } */
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Fix vec_concatv2di pattern for SSE4 (PR target/80819)
2017-11-29 8:35 [PATCH] Fix vec_concatv2di pattern for SSE4 (PR target/80819) Jakub Jelinek
@ 2017-11-29 9:36 ` Uros Bizjak
0 siblings, 0 replies; 2+ messages in thread
From: Uros Bizjak @ 2017-11-29 9:36 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc-patches
On Wed, Nov 29, 2017 at 9:24 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> Before r218303 we had just (=x,0,rm) alternative for SSE4 (no AVX),
> that change turned it into (=Yr,0,*rm) and (=*x,0,rm) alternatives,
> so that we avoid too many prefixes if possible.
> The latter alternative is fine, we want the *, because that is the point,
> Yr class is the subset of x registers that don't need the REX prefix.
> The * in the first alternative makes no sense, with it IRA is effectively
> forced to allocate the second vec_concat pseudo into NO_REGS - memory,
> and while postreload can fix it up afterwards, we end up with dead stores
> that nothing ever removes afterwards.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?
>
> 2017-11-29 Jakub Jelinek <jakub@redhat.com>
>
> PR target/80819
> * config/i386/sse.md (vec_concatv2di): Remove * from (=Yr,0,*rm)
> alternative.
>
> * gcc.target/i386/pr80819-1.c: New test.
> * gcc.target/i386/pr80819-2.c: New test.
OK.
Thanks,
Uros.
> --- gcc/config/i386/sse.md.jj 2017-11-24 08:58:05.000000000 +0100
> +++ gcc/config/i386/sse.md 2017-11-28 18:04:20.739396199 +0100
> @@ -13915,7 +13915,7 @@ (define_insn "vec_concatv2di"
> (match_operand:DI 1 "nonimmediate_operand"
> " 0, 0,x ,Yv,r ,vm,?!*Yn,0,Yv,0,0,v")
> (match_operand:DI 2 "vector_move_operand"
> - "*rm,rm,rm,rm,C ,C ,C ,x,Yv,x,m,m")))]
> + " rm,rm,rm,rm,C ,C ,C ,x,Yv,x,m,m")))]
> "TARGET_SSE"
> "@
> pinsrq\t{$1, %2, %0|%0, %2, 1}
> --- gcc/testsuite/gcc.target/i386/pr80819-1.c.jj 2017-11-28 18:11:09.452482042 +0100
> +++ gcc/testsuite/gcc.target/i386/pr80819-1.c 2017-11-28 18:09:57.000000000 +0100
> @@ -0,0 +1,13 @@
> +/* PR target/80819 */
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-O2 -msse4 -mno-avx -mtune=haswell -masm=att" } */
> +
> +typedef unsigned long long V __attribute__((vector_size (16)));
> +
> +V
> +foo (unsigned long long x, unsigned long long y)
> +{
> + return (V) { x, y };
> +}
> +
> +/* { dg-final { scan-assembler-not "movq\[ \t]*%rsi, \[-0-9]*\\(" } } */
> --- gcc/testsuite/gcc.target/i386/pr80819-2.c.jj 2017-11-28 18:11:15.942404034 +0100
> +++ gcc/testsuite/gcc.target/i386/pr80819-2.c 2017-11-28 18:11:21.915332239 +0100
> @@ -0,0 +1,13 @@
> +/* PR target/80819 */
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-O2 -msse4 -mno-avx -mtune=generic -masm=att" } */
> +
> +typedef unsigned long long V __attribute__((vector_size (16)));
> +
> +V
> +foo (unsigned long long x, unsigned long long y)
> +{
> + return (V) { x, y };
> +}
> +
> +/* { dg-final { scan-assembler-not "movq\[ \t]*%rsi, \[-0-9]*\\(" } } */
>
> Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2017-11-29 8:35 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-11-29 8:35 [PATCH] Fix vec_concatv2di pattern for SSE4 (PR target/80819) Jakub Jelinek
2017-11-29 9:36 ` Uros Bizjak
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).