i386: Fix up the vzeroupper REG_DEAD/REG_UNUSED note workaround [PR114190]
Checks
Commit Message
Hi!
When writing the rest_of_handle_insert_vzeroupper workaround to manually
remove all the REG_DEAD/REG_UNUSED notes from the IL, I've missed that
there is a df_analyze () call right after it and that the problems added
earlier in the pass, like df_note_add_problem () done during mode switching,
doesn't affect just the next df_analyze () call right after it, but all
other df_analyze () calls until the end of the current pass where
df_finish_pass removes the optional problems.
So, as can be seen on the following patch, the workaround doesn't actually
work there, because while rest_of_handle_insert_vzeroupper carefully removes
all REG_DEAD/REG_UNUSED notes, the df_analyze () call at the end of the
function immediately adds them in again (so, I must say I have no idea
why the workaround worked on the earlier testcases).
Now, I could move the df_analyze () call just before the REG_DEAD/REG_UNUSED
note removal loop, but I think the following patch is better, because
the df_analyze () call doesn't have to recompute the problem when we don't
care about it and will actively strip all traces of it away.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2024-03-06 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/114190
* config/i386/i386-features.cc (rest_of_handle_insert_vzeroupper):
Call df_remove_problem for df_note before calling df_analyze.
* gcc.target/i386/avx-pr114190.c: New test.
Jakub
Comments
On Wed, Mar 6, 2024 at 9:10 AM Jakub Jelinek <jakub@redhat.com> wrote:
>
> Hi!
>
> When writing the rest_of_handle_insert_vzeroupper workaround to manually
> remove all the REG_DEAD/REG_UNUSED notes from the IL, I've missed that
> there is a df_analyze () call right after it and that the problems added
> earlier in the pass, like df_note_add_problem () done during mode switching,
> doesn't affect just the next df_analyze () call right after it, but all
> other df_analyze () calls until the end of the current pass where
> df_finish_pass removes the optional problems.
>
> So, as can be seen on the following patch, the workaround doesn't actually
> work there, because while rest_of_handle_insert_vzeroupper carefully removes
> all REG_DEAD/REG_UNUSED notes, the df_analyze () call at the end of the
> function immediately adds them in again (so, I must say I have no idea
> why the workaround worked on the earlier testcases).
>
> Now, I could move the df_analyze () call just before the REG_DEAD/REG_UNUSED
> note removal loop, but I think the following patch is better, because
> the df_analyze () call doesn't have to recompute the problem when we don't
> care about it and will actively strip all traces of it away.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2024-03-06 Jakub Jelinek <jakub@redhat.com>
>
> PR rtl-optimization/114190
> * config/i386/i386-features.cc (rest_of_handle_insert_vzeroupper):
> Call df_remove_problem for df_note before calling df_analyze.
>
> * gcc.target/i386/avx-pr114190.c: New test.
OK.
Thanks,
Uros.
>
> --- gcc/config/i386/i386-features.cc.jj 2024-02-22 10:10:18.658032517 +0100
> +++ gcc/config/i386/i386-features.cc 2024-03-05 09:23:54.496112264 +0100
> @@ -2690,6 +2690,7 @@ rest_of_handle_insert_vzeroupper (void)
> }
> }
>
> + df_remove_problem (df_note);
> df_analyze ();
> return 0;
> }
> --- gcc/testsuite/gcc.target/i386/avx-pr114190.c.jj 2024-03-05 10:07:24.869454305 +0100
> +++ gcc/testsuite/gcc.target/i386/avx-pr114190.c 2024-03-05 10:06:52.870889687 +0100
> @@ -0,0 +1,27 @@
> +/* PR rtl-optimization/114190 */
> +/* { dg-do run { target avx } } */
> +/* { dg-options "-O2 -fno-dce -fharden-compares -mavx --param=max-rtl-if-conversion-unpredictable-cost=136 -mno-avx512f -Wno-psabi" } */
> +
> +#include "avx-check.h"
> +
> +typedef unsigned char U __attribute__((vector_size (64)));
> +typedef unsigned int V __attribute__((vector_size (64)));
> +U u;
> +
> +V
> +foo (V a, V b)
> +{
> + u[0] = __builtin_sub_overflow (0, (int) a[0], &a[b[7] & 5]) ? -u[1] : -b[3];
> + b ^= 0 != b;
> + return (V) u + (V) a + (V) b;
> +}
> +
> +static void
> +avx_test (void)
> +{
> + V x = foo ((V) { 1 }, (V) { 0, 0, 0, 1 });
> + if (x[0] != -1U)
> + __builtin_abort ();
> + if (x[3] != -2U)
> + __builtin_abort ();
> +}
>
> Jakub
>
@@ -2690,6 +2690,7 @@ rest_of_handle_insert_vzeroupper (void)
}
}
+ df_remove_problem (df_note);
df_analyze ();
return 0;
}
@@ -0,0 +1,27 @@
+/* PR rtl-optimization/114190 */
+/* { dg-do run { target avx } } */
+/* { dg-options "-O2 -fno-dce -fharden-compares -mavx --param=max-rtl-if-conversion-unpredictable-cost=136 -mno-avx512f -Wno-psabi" } */
+
+#include "avx-check.h"
+
+typedef unsigned char U __attribute__((vector_size (64)));
+typedef unsigned int V __attribute__((vector_size (64)));
+U u;
+
+V
+foo (V a, V b)
+{
+ u[0] = __builtin_sub_overflow (0, (int) a[0], &a[b[7] & 5]) ? -u[1] : -b[3];
+ b ^= 0 != b;
+ return (V) u + (V) a + (V) b;
+}
+
+static void
+avx_test (void)
+{
+ V x = foo ((V) { 1 }, (V) { 0, 0, 0, 1 });
+ if (x[0] != -1U)
+ __builtin_abort ();
+ if (x[3] != -2U)
+ __builtin_abort ();
+}