[Bug target/95791] New: Unnecessary vzeroupper when only using zmm16 through zmm31

[Bug target/95791] New: Unnecessary vzeroupper when only using zmm16 through zmm31

[josephcsible at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95791

            Bug ID: 95791
           Summary: Unnecessary vzeroupper when only using zmm16 through
                    zmm31
           Product: gcc
           Version: 10.1.0
            Status: UNCONFIRMED
          Keywords: missed-optimization, ssemmx
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: josephcsible at gmail dot com
  Target Milestone: ---
            Target: x86_64-linux-gnu

Consider this C code:

void f(void) {
    __asm__ __volatile__("" ::: "zmm16");
}

When compiled with "-O2 -mavx512f", it generates a vzeroupper instruction, but
this is unnecessary, since zmm16 through zmm31 don't cause the performance
penalty, and in fact they aren't even affected by vzeroupper.

[Bug target/95791] Unnecessary vzeroupper when only using zmm16 through zmm31

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95791

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2020-06-20
                URL|                            |https://gcc.gnu.org/piperma
                   |                            |il/gcc-patches/2020-June/54
                   |                            |8648.html
           Keywords|                            |patch
     Ever confirmed|0                           |1
   Target Milestone|---                         |11.0

[Bug target/95791] Unnecessary vzeroupper when only using zmm16 through zmm31

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95791

--- Comment #1 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by H.J. Lu <hjl@gcc.gnu.org>:

https://gcc.gnu.org/g:9302421e71e85b4d3766a534ed9e1c4ae1e7a6ca

commit r11-1585-g9302421e71e85b4d3766a534ed9e1c4ae1e7a6ca
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Sat Jun 20 16:02:42 2020 -0700

    x86: Skip EXT_REX_SSE_REG_P for vzeroupper optimization

    Skip EXT_REX_SSE_REG_P for vzeroupper optimization since upper 16 vector
    registers don't trigger SSE <-> AVX transition penalty.

    gcc/

            PR target/95791
            * config/i386/i386.c (ix86_dirflag_mode_needed): Skip
            EXT_REX_SSE_REG_P.

    gcc/testsuite/

            PR target/95791
            * gcc.target/i386/pr95791.c: New test.

[Bug target/95791] Unnecessary vzeroupper when only using zmm16 through zmm31

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95791

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|NEW                         |RESOLVED

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> ---
Fixed for GCC 11.