This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Optimize integral lt + blend into just blend (PR target/54700)
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Uros Bizjak <ubizjak at gmail dot com>
- Cc: "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>
- Date: Thu, 29 Nov 2018 22:40:52 +0100
- Subject: Re: [PATCH] Optimize integral lt + blend into just blend (PR target/54700)
- References: <20181129075956.GE12380@tucnak> <CAFULd4b_umLDY7FdNDf9T_FVk1zspV-8r1=fDL6=Qs__cj0wEw@mail.gmail.com> <CAFULd4ZmK+ZxaX8PLHH1tRQoR=Th9x+SJeSq+12tQUOV_CxLcQ@mail.gmail.com> <20181129162829.GP12380@tucnak> <CAFULd4Z4Ds_e8XN1TTcy10U6eGwk-BYOVUxWzFUQJgtQjTcOzg@mail.gmail.com>
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
On Thu, Nov 29, 2018 at 05:41:59PM +0100, Uros Bizjak wrote:
> On Thu, Nov 29, 2018 at 5:28 PM Jakub Jelinek <jakub@redhat.com> wrote:
> >
> > On Thu, Nov 29, 2018 at 05:21:53PM +0100, Uros Bizjak wrote:
> > > > > * g++.target/i386/avx2-check.h: New file.
> > > > > * g++.target/i386/m128-check.h: New file.
> > > > > * g++.target/i386/m256-check.h: New file.
> > > > > * g++.target/i386/avx-os-support.h: New file.
> > > >
> > > > OK.
> > >
> > > On a second thought, should we rather use (pre-reload?)
> > > define_insn_and_split to split the combination to the blend insn?
> >
> > I've already committed it. But can work on a patch that does that tomorrow.
>
> Thanks. You will probably need to split it after reload, since a
> change from intvec->FPvec is needed.
Like this? Bootstrapped/regtested on x86_64-linux and i686-linux.
2018-11-29 Jakub Jelinek <jakub@redhat.com>
PR target/54700
* config/i386/sse.md
(*<sse4_1>_blendv<ssemodesuffix><avxsizesuffix>_lt,
*<sse4_1>_blendv<ssefltmodesuffix><avxsizesuffix>_ltint,
*<sse4_1_avx2>_pblendvb_lt): Change define_insn into
define_insn_and_split.
--- gcc/config/i386/sse.md.jj 2018-11-29 15:32:27.597301378 +0100
+++ gcc/config/i386/sse.md 2018-11-29 18:52:42.747904630 +0100
@@ -15682,7 +15682,7 @@ (define_insn "sse4_1_blendv<ssemodesuffi
]
(const_string "<ssevecmode>")))])
-(define_insn "*<sse4_1>_blendv<ssemodesuffix><avxsizesuffix>_lt"
+(define_insn_and_split "*<sse4_1>_blendv<ssemodesuffix><avxsizesuffix>_lt"
[(set (match_operand:VF_128_256 0 "register_operand" "=Yr,*x,x")
(unspec:VF_128_256
[(match_operand:VF_128_256 1 "register_operand" "0,0,x")
@@ -15693,10 +15693,12 @@ (define_insn "*<sse4_1>_blendv<ssemodesu
(match_operand:<sseintvecmode> 4 "const0_operand" "C,C,C")) 0)]
UNSPEC_BLENDV))]
"TARGET_SSE4_1"
- "@
- blendv<ssemodesuffix>\t{%3, %2, %0|%0, %2, %3}
- blendv<ssemodesuffix>\t{%3, %2, %0|%0, %2, %3}
- vblendv<ssemodesuffix>\t{%3, %2, %1, %0|%0, %1, %2, %3}"
+ "#"
+ "&& reload_completed"
+ [(set (match_dup 0)
+ (unspec:VF_128_256
+ [(match_dup 1) (match_dup 2) (match_dup 3)] UNSPEC_BLENDV))]
+ "operands[3] = gen_lowpart (<MODE>mode, operands[3]);"
[(set_attr "isa" "noavx,noavx,avx")
(set_attr "type" "ssemov")
(set_attr "length_immediate" "1")
@@ -15712,7 +15714,7 @@ (define_mode_attr ssefltmodesuffix
(define_mode_attr ssefltvecmode
[(V2DI "V2DF") (V4DI "V4DF") (V4SI "V4SF") (V8SI "V8SF")])
-(define_insn "*<sse4_1>_blendv<ssefltmodesuffix><avxsizesuffix>_ltint"
+(define_insn_and_split "*<sse4_1>_blendv<ssefltmodesuffix><avxsizesuffix>_ltint"
[(set (match_operand:<ssebytemode> 0 "register_operand" "=Yr,*x,x")
(unspec:<ssebytemode>
[(match_operand:<ssebytemode> 1 "register_operand" "0,0,x")
@@ -15723,10 +15725,17 @@ (define_insn "*<sse4_1>_blendv<ssefltmod
(match_operand:VI48_AVX 4 "const0_operand" "C,C,C")) 0)]
UNSPEC_BLENDV))]
"TARGET_SSE4_1"
- "@
- blendv<ssefltmodesuffix>\t{%3, %2, %0|%0, %2, %3}
- blendv<ssefltmodesuffix>\t{%3, %2, %0|%0, %2, %3}
- vblendv<ssefltmodesuffix>\t{%3, %2, %1, %0|%0, %1, %2, %3}"
+ "#"
+ "&& reload_completed"
+ [(set (match_dup 0)
+ (unspec:<ssefltvecmode>
+ [(match_dup 1) (match_dup 2) (match_dup 3)] UNSPEC_BLENDV))]
+{
+ operands[0] = gen_lowpart (<ssefltvecmode>mode, operands[0]);
+ operands[1] = gen_lowpart (<ssefltvecmode>mode, operands[1]);
+ operands[2] = gen_lowpart (<ssefltvecmode>mode, operands[2]);
+ operands[3] = gen_lowpart (<ssefltvecmode>mode, operands[3]);
+}
[(set_attr "isa" "noavx,noavx,avx")
(set_attr "type" "ssemov")
(set_attr "length_immediate" "1")
@@ -15834,7 +15843,7 @@ (define_insn "<sse4_1_avx2>_pblendvb"
(set_attr "btver2_decode" "vector,vector,vector")
(set_attr "mode" "<sseinsnmode>")])
-(define_insn "*<sse4_1_avx2>_pblendvb_lt"
+(define_insn_and_split "*<sse4_1_avx2>_pblendvb_lt"
[(set (match_operand:VI1_AVX2 0 "register_operand" "=Yr,*x,x")
(unspec:VI1_AVX2
[(match_operand:VI1_AVX2 1 "register_operand" "0,0,x")
@@ -15843,10 +15852,12 @@ (define_insn "*<sse4_1_avx2>_pblendvb_lt
(match_operand:VI1_AVX2 4 "const0_operand" "C,C,C"))]
UNSPEC_BLENDV))]
"TARGET_SSE4_1"
- "@
- pblendvb\t{%3, %2, %0|%0, %2, %3}
- pblendvb\t{%3, %2, %0|%0, %2, %3}
- vpblendvb\t{%3, %2, %1, %0|%0, %1, %2, %3}"
+ "#"
+ ""
+ [(set (match_dup 0)
+ (unspec:VI1_AVX2
+ [(match_dup 1) (match_dup 2) (match_dup 3)] UNSPEC_BLENDV))]
+ ""
[(set_attr "isa" "noavx,noavx,avx")
(set_attr "type" "ssemov")
(set_attr "prefix_extra" "1")
Jakub