This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Fix wrong-code x86 issue with avx512{f,vl} fma (PR target/93009)
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Uros Bizjak <ubizjak at gmail dot com>, Jeff Law <law at redhat dot com>
- Cc: gcc-patches at gcc dot gnu dot org, hjl dot tools at gmail dot com
- Date: Sat, 21 Dec 2019 00:34:54 +0100
- Subject: [PATCH] Fix wrong-code x86 issue with avx512{f,vl} fma (PR target/93009)
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
As mentioned in the PR, the following testcase is miscompiled with avx512vl.
The reason is that the fma *_bcst_1 define_insns have two alternatives:
"=v,v" "0,v" "v,0" "m,m" and use the same
vfmadd213* %3<avx512bcst>, %2, %0<sd_mask_op4>
pattern. If the first alternative is chosen, everything is ok, but if the
second alternative is chosen, %2 and %0 are the same register, so instead
of doing dest=dest*another+membcst we do dest=dest*dest+membcst.
Now, to fix this, either we'd need separate:
"vfmadd213<ssemodesuffix>\t{%3<avx512bcst>, %2, %0<sd_mask_op4>|%0<sd_mask_op4>, %2, %3<avx512bcst>}
vfmadd213<ssemodesuffix>\t{%3<avx512bcst>, %1, %0<sd_mask_op4>|%0<sd_mask_op4>, %1, %3<avx512bcst>}"
where for the second alternative, we'd just use %1 instead of %2, but
what I think is actually cleaner is just use a single alternative and
make the two multiplication operands commutative, which they really are.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2019-12-20 Jakub Jelinek <jakub@redhat.com>
PR target/93009
* config/i386/sse.md
(*<sd_mask_codefor>fma_fmadd_<mode><sd_maskz_name>_bcst_1,
*<sd_mask_codefor>fma_fmsub_<mode><sd_maskz_name>_bcst_1,
*<sd_mask_codefor>fma_fnmadd_<mode><sd_maskz_name>_bcst_1,
*<sd_mask_codefor>fma_fnmsub_<mode><sd_maskz_name>_bcst_1): Use
just a single alternative instead of two, make operands 1 and 2
commutative.
* gcc.target/i386/avx512vl-pr93009.c: New test.
--- gcc/config/i386/sse.md.jj 2019-12-09 11:12:30.000000000 +0100
+++ gcc/config/i386/sse.md 2019-12-20 21:44:58.165432773 +0100
@@ -4172,12 +4172,12 @@ (define_insn "<sd_mask_codefor>fma_fmadd
(set_attr "mode" "<MODE>")])
(define_insn "*<sd_mask_codefor>fma_fmadd_<mode><sd_maskz_name>_bcst_1"
- [(set (match_operand:VF_AVX512 0 "register_operand" "=v,v")
+ [(set (match_operand:VF_AVX512 0 "register_operand" "=v")
(fma:VF_AVX512
- (match_operand:VF_AVX512 1 "register_operand" "0,v")
- (match_operand:VF_AVX512 2 "register_operand" "v,0")
+ (match_operand:VF_AVX512 1 "register_operand" "%0")
+ (match_operand:VF_AVX512 2 "register_operand" "v")
(vec_duplicate:VF_AVX512
- (match_operand:<ssescalarmode> 3 "memory_operand" "m,m"))))]
+ (match_operand:<ssescalarmode> 3 "memory_operand" "m"))))]
"TARGET_AVX512F && <sd_mask_mode512bit_condition>"
"vfmadd213<ssemodesuffix>\t{%3<avx512bcst>, %2, %0<sd_mask_op4>|%0<sd_mask_op4>, %2, %3<avx512bcst>}"
[(set_attr "type" "ssemuladd")
@@ -4289,13 +4289,13 @@ (define_insn "<sd_mask_codefor>fma_fmsub
(set_attr "mode" "<MODE>")])
(define_insn "*<sd_mask_codefor>fma_fmsub_<mode><sd_maskz_name>_bcst_1"
- [(set (match_operand:VF_AVX512 0 "register_operand" "=v,v")
+ [(set (match_operand:VF_AVX512 0 "register_operand" "=v")
(fma:VF_AVX512
- (match_operand:VF_AVX512 1 "register_operand" "0,v")
- (match_operand:VF_AVX512 2 "register_operand" "v,0")
+ (match_operand:VF_AVX512 1 "register_operand" "%0")
+ (match_operand:VF_AVX512 2 "register_operand" "v")
(neg:VF_AVX512
(vec_duplicate:VF_AVX512
- (match_operand:<ssescalarmode> 3 "memory_operand" "m,m")))))]
+ (match_operand:<ssescalarmode> 3 "memory_operand" "m")))))]
"TARGET_AVX512F && <sd_mask_mode512bit_condition>"
"vfmsub213<ssemodesuffix>\t{%3<avx512bcst>, %2, %0<sd_mask_op4>|%0<sd_mask_op4>, %2, %3<avx512bcst>}"
[(set_attr "type" "ssemuladd")
@@ -4411,13 +4411,13 @@ (define_insn "<sd_mask_codefor>fma_fnmad
(set_attr "mode" "<MODE>")])
(define_insn "*<sd_mask_codefor>fma_fnmadd_<mode><sd_maskz_name>_bcst_1"
- [(set (match_operand:VF_AVX512 0 "register_operand" "=v,v")
+ [(set (match_operand:VF_AVX512 0 "register_operand" "=v")
(fma:VF_AVX512
(neg:VF_AVX512
- (match_operand:VF_AVX512 1 "register_operand" "0,v"))
- (match_operand:VF_AVX512 2 "register_operand" "v,0")
+ (match_operand:VF_AVX512 1 "register_operand" "%0"))
+ (match_operand:VF_AVX512 2 "register_operand" "v")
(vec_duplicate:VF_AVX512
- (match_operand:<ssescalarmode> 3 "memory_operand" "m,m"))))]
+ (match_operand:<ssescalarmode> 3 "memory_operand" "m"))))]
"TARGET_AVX512F && <sd_mask_mode512bit_condition>"
"vfnmadd213<ssemodesuffix>\t{%3<avx512bcst>, %2, %0<sd_mask_op4>|%0<sd_mask_op4>, %2, %3<avx512bcst>}"
[(set_attr "type" "ssemuladd")
@@ -4535,14 +4535,14 @@ (define_insn "<sd_mask_codefor>fma_fnmsu
(set_attr "mode" "<MODE>")])
(define_insn "*<sd_mask_codefor>fma_fnmsub_<mode><sd_maskz_name>_bcst_1"
- [(set (match_operand:VF_AVX512 0 "register_operand" "=v,v")
+ [(set (match_operand:VF_AVX512 0 "register_operand" "=v")
(fma:VF_AVX512
(neg:VF_AVX512
- (match_operand:VF_AVX512 1 "register_operand" "0,v"))
- (match_operand:VF_AVX512 2 "register_operand" "v,0")
+ (match_operand:VF_AVX512 1 "register_operand" "%0"))
+ (match_operand:VF_AVX512 2 "register_operand" "v")
(neg:VF_AVX512
(vec_duplicate:VF_AVX512
- (match_operand:<ssescalarmode> 3 "memory_operand" "m,m")))))]
+ (match_operand:<ssescalarmode> 3 "memory_operand" "m")))))]
"TARGET_AVX512F && <sd_mask_mode512bit_condition>"
"vfnmsub213<ssemodesuffix>\t{%3<avx512bcst>, %2, %0<sd_mask_op4>|%0<sd_mask_op4>, %2, %3<avx512bcst>}"
[(set_attr "type" "ssemuladd")
--- gcc/testsuite/gcc.target/i386/avx512vl-pr93009.c.jj 2019-12-20 21:54:11.158011954 +0100
+++ gcc/testsuite/gcc.target/i386/avx512vl-pr93009.c 2019-12-20 21:53:50.375328425 +0100
@@ -0,0 +1,38 @@
+/* PR target/93009 */
+/* { dg-do run { target { avx512vl } } } */
+/* { dg-options "-O2 -mavx512vl" } */
+
+#define AVX512VL
+#define AVX512F_LEN 512
+#define AVX512F_LEN_HALF 256
+
+#include "avx512f-check.h"
+
+typedef double v2df __attribute__((vector_size (16)));
+
+__attribute__((noipa)) v2df
+foo (v2df x, v2df y, double *z)
+{
+ return x * y + (v2df) { z[0], z[0] };
+}
+
+__attribute__((noipa)) v2df
+bar (v2df x, v2df y, double *z)
+{
+ return y * x + (v2df) { z[0], z[0] };
+}
+
+static void
+test_256 (void)
+{
+}
+
+static void
+test_128 (void)
+{
+ double z = 5.0;
+ v2df x = foo ((v2df) { 1.0, 2.0 }, (v2df) { 3.0, 4.0 }, &z);
+ v2df y = bar ((v2df) { 6.0, 7.0 }, (v2df) { 8.0, 9.0 }, &z);
+ if (x[0] != 8.0 || x[1] != 13.0 || y[0] != 53.0 || y[1] != 68.0)
+ abort ();
+}
Jakub