* [PATCH] Fix -O0 AVX512 comparison ICE (PR target/78227)
@ 2016-11-07 13:02 Jakub Jelinek
2016-11-07 15:08 ` Uros Bizjak
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2016-11-07 13:02 UTC (permalink / raw)
To: Uros Bizjak; +Cc: gcc-patches
Hi!
The following testcases ICE at -O0, because ix86_expand_sse_cmp avoid using
the passed in dest only if optimize or if there is some value overlap, but
we actually need to do that also if we have a maskcmp where we want to use
a different mode than dest has.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2016-11-07 Jakub Jelinek <jakub@redhat.com>
PR target/78227
* config/i386/i386.c (ix86_expand_sse_cmp): Force dest into
cmp_mode argument even for -O0 if cmp_mode != mode and maskcmp.
* gcc.target/i386/pr78227-1.c: New test.
* gcc.target/i386/pr78227-2.c: New test.
--- gcc/config/i386/i386.c.jj 2016-11-04 20:09:48.000000000 +0100
+++ gcc/config/i386/i386.c 2016-11-07 10:14:15.625018144 +0100
@@ -23561,6 +23561,7 @@ ix86_expand_sse_cmp (rtx dest, enum rtx_
cmp_op1 = force_reg (cmp_ops_mode, cmp_op1);
if (optimize
+ || (cmp_mode != mode && maskcmp)
|| (op_true && reg_overlap_mentioned_p (dest, op_true))
|| (op_false && reg_overlap_mentioned_p (dest, op_false)))
dest = gen_reg_rtx (maskcmp ? cmp_mode : mode);
--- gcc/testsuite/gcc.target/i386/pr78227-1.c.jj 2016-11-07 10:15:52.606762613 +0100
+++ gcc/testsuite/gcc.target/i386/pr78227-1.c 2016-11-07 10:24:58.821480125 +0100
@@ -0,0 +1,30 @@
+/* PR target/78227 */
+/* { dg-do compile } */
+/* { dg-options "-mavx512f -O0 -Wno-psabi" } */
+
+typedef int V __attribute__((vector_size (64)));
+typedef long long int W __attribute__((vector_size (64)));
+
+V
+foo1 (V v)
+{
+ return v > 0;
+}
+
+V
+bar1 (V v)
+{
+ return v != 0;
+}
+
+W
+foo2 (W w)
+{
+ return w > 0;
+}
+
+W
+bar2 (W w)
+{
+ return w != 0;
+}
--- gcc/testsuite/gcc.target/i386/pr78227-2.c.jj 2016-11-07 10:22:17.055670476 +0100
+++ gcc/testsuite/gcc.target/i386/pr78227-2.c 2016-11-07 10:25:03.722413765 +0100
@@ -0,0 +1,30 @@
+/* PR target/78227 */
+/* { dg-do compile } */
+/* { dg-options "-mavx512bw -O0 -Wno-psabi" } */
+
+typedef signed char V __attribute__((vector_size (64)));
+typedef short int W __attribute__((vector_size (64)));
+
+V
+foo1 (V v)
+{
+ return v > 0;
+}
+
+V
+bar1 (V v)
+{
+ return v != 0;
+}
+
+W
+foo2 (W w)
+{
+ return w > 0;
+}
+
+W
+bar2 (W w)
+{
+ return w != 0;
+}
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Fix -O0 AVX512 comparison ICE (PR target/78227)
2016-11-07 13:02 [PATCH] Fix -O0 AVX512 comparison ICE (PR target/78227) Jakub Jelinek
@ 2016-11-07 15:08 ` Uros Bizjak
0 siblings, 0 replies; 2+ messages in thread
From: Uros Bizjak @ 2016-11-07 15:08 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc-patches
On Mon, Nov 7, 2016 at 2:02 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> The following testcases ICE at -O0, because ix86_expand_sse_cmp avoid using
> the passed in dest only if optimize or if there is some value overlap, but
> we actually need to do that also if we have a maskcmp where we want to use
> a different mode than dest has.
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2016-11-07 Jakub Jelinek <jakub@redhat.com>
>
> PR target/78227
> * config/i386/i386.c (ix86_expand_sse_cmp): Force dest into
> cmp_mode argument even for -O0 if cmp_mode != mode and maskcmp.
>
> * gcc.target/i386/pr78227-1.c: New test.
> * gcc.target/i386/pr78227-2.c: New test.
OK with a small nit, please see inline ...
Thanks,
Uros.
> --- gcc/config/i386/i386.c.jj 2016-11-04 20:09:48.000000000 +0100
> +++ gcc/config/i386/i386.c 2016-11-07 10:14:15.625018144 +0100
> @@ -23561,6 +23561,7 @@ ix86_expand_sse_cmp (rtx dest, enum rtx_
> cmp_op1 = force_reg (cmp_ops_mode, cmp_op1);
>
> if (optimize
> + || (cmp_mode != mode && maskcmp)
Maybe beter to switch condition around, so:
"(maskcmp && cmp_mode != mode)"
> || (op_true && reg_overlap_mentioned_p (dest, op_true))
> || (op_false && reg_overlap_mentioned_p (dest, op_false)))
> dest = gen_reg_rtx (maskcmp ? cmp_mode : mode);
> --- gcc/testsuite/gcc.target/i386/pr78227-1.c.jj 2016-11-07 10:15:52.606762613 +0100
> +++ gcc/testsuite/gcc.target/i386/pr78227-1.c 2016-11-07 10:24:58.821480125 +0100
> @@ -0,0 +1,30 @@
> +/* PR target/78227 */
> +/* { dg-do compile } */
> +/* { dg-options "-mavx512f -O0 -Wno-psabi" } */
> +
> +typedef int V __attribute__((vector_size (64)));
> +typedef long long int W __attribute__((vector_size (64)));
> +
> +V
> +foo1 (V v)
> +{
> + return v > 0;
> +}
> +
> +V
> +bar1 (V v)
> +{
> + return v != 0;
> +}
> +
> +W
> +foo2 (W w)
> +{
> + return w > 0;
> +}
> +
> +W
> +bar2 (W w)
> +{
> + return w != 0;
> +}
> --- gcc/testsuite/gcc.target/i386/pr78227-2.c.jj 2016-11-07 10:22:17.055670476 +0100
> +++ gcc/testsuite/gcc.target/i386/pr78227-2.c 2016-11-07 10:25:03.722413765 +0100
> @@ -0,0 +1,30 @@
> +/* PR target/78227 */
> +/* { dg-do compile } */
> +/* { dg-options "-mavx512bw -O0 -Wno-psabi" } */
> +
> +typedef signed char V __attribute__((vector_size (64)));
> +typedef short int W __attribute__((vector_size (64)));
> +
> +V
> +foo1 (V v)
> +{
> + return v > 0;
> +}
> +
> +V
> +bar1 (V v)
> +{
> + return v != 0;
> +}
> +
> +W
> +foo2 (W w)
> +{
> + return w > 0;
> +}
> +
> +W
> +bar2 (W w)
> +{
> + return w != 0;
> +}
>
> Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2016-11-07 15:08 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-11-07 13:02 [PATCH] Fix -O0 AVX512 comparison ICE (PR target/78227) Jakub Jelinek
2016-11-07 15:08 ` 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).