[Bug target/100312] New: __builtin_ia32_maskloadpd256 and friends should be pure

[Bug target/100312] New: __builtin_ia32_maskloadpd256 and friends should be pure

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100312

            Bug ID: 100312
           Summary: __builtin_ia32_maskloadpd256 and friends should be
                    pure
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

The testcase in PR99912 shows that we fail to eliminate dead AVX2 masked loads
since the x86 specific builtins are not marked pure and thus are considered
to clobber their arguments and have other unwanted side-effects.

It looks like the "way" to change this is to remove the builtins from
i386-builtin.def and manually add them to ix86_init_mmx_sse_builtins like
the AVX2 gather builtins (which are properly constructed as pure).

[Bug target/100312] __builtin_ia32_maskloadpd256 and friends should be pure

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100312

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
   Last reconfirmed|                            |2021-04-28
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |ASSIGNED
             Target|                            |x86_64-*-* i?86-*-*

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Working on a patch.

[Bug target/100312] __builtin_ia32_maskloadpd256 and friends should be pure

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100312

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 50699
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50699&action=edit
not working patch

This one misses the RTL expansion part.  I'm too lazy to writing duplicate code
to handle this.  The following works but is a bit ugly.  Meh, the builtins
stuff _is_ ugly (adding extra flags to bdesc would be nicer but a lot of
boiler-plate work and the question is whether all the rest is "nice enough"
or in need of serious overhaul anyway...)

diff --git a/gcc/config/i386/i386-builtins.c b/gcc/config/i386/i386-builtins.c
index 128bd39816c..fa05cd6cf65 100644
--- a/gcc/config/i386/i386-builtins.c
+++ b/gcc/config/i386/i386-builtins.c
@@ -521,7 +521,17 @@ ix86_init_mmx_sse_builtins (void)
        continue;

       ftype = (enum ix86_builtin_func_type) d->flag;
-      def_builtin (d->mask, d->mask2, d->name, ftype, d->code);
+      if (d->code == IX86_BUILTIN_MASKLOADPD
+         || d->code == IX86_BUILTIN_MASKLOADPS
+         || d->code == IX86_BUILTIN_MASKLOADPD256
+         || d->code == IX86_BUILTIN_MASKLOADPS256
+         || d->code == IX86_BUILTIN_MASKLOADD
+         || d->code == IX86_BUILTIN_MASKLOADQ
+         || d->code == IX86_BUILTIN_MASKLOADD256
+         || d->code == IX86_BUILTIN_MASKLOADQ256)
+       def_builtin_pure (d->mask, d->mask2, d->name, ftype, d->code);
+      else
+       def_builtin (d->mask, d->mask2, d->name, ftype, d->code);
     }
   BDESC_VERIFYS (IX86_BUILTIN__BDESC_SPECIAL_ARGS_LAST,
                 IX86_BUILTIN__BDESC_SPECIAL_ARGS_FIRST,

[Bug target/100312] __builtin_ia32_maskloadpd256 and friends should be pure

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100312

--- Comment #3 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Uros Bizjak <uros@gcc.gnu.org>:

https://gcc.gnu.org/g:fd5d57946036c967dae292330fa0aa856a58fb4b

commit r12-290-gfd5d57946036c967dae292330fa0aa856a58fb4b
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Thu Apr 29 16:43:33 2021 +0200

    i386: Mark x86 masked load builtins pure [PR100312]

    Mark x86 AVX and AVX2 masked load builtins pure to enable dead code
    elimination and more appropriate alias analysis.

    2021-04-29  Uroš Bizjak  <ubizjak@gmail.com>
                Richard Biener  <rguenther@suse.de>
    gcc/
            PR target/100312
            * config/i386/i386-builtin.def (IX86_BUILTIN_MASKLOADPD)
            (IX86_BUILTIN_MASKLOADPS, IX86_BUILTIN_MASKLOADPD256)
            (IX86_BUILTIN_MASKLOADPS256, IX86_BUILTIN_MASKLOADD)
            (IX86_BUILTIN_MASKLOADQ, IX86_BUILTIN_MASKLOADD256)
            (IX86_BUILTIN_MASKLOADQ256): Move from SPECIAL_ARGS
            to PURE_ARGS category.
            * config/i386/i386-builtins.c (ix86_init_mmx_sse_builtins):
            Handle PURE_ARGS category.
            * config/i386/i386-expand.c (ix86_expand_builtin): Ditto.

[Bug target/100312] __builtin_ia32_maskloadpd256 and friends should be pure

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100312

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|rguenth at gcc dot gnu.org         |ubizjak at gmail dot com
             Target|x86_64-*-* i?86-*-*         |x86
   Target Milestone|---                         |12.0
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
Fixed.