[Bug c/99100] New: Inconsistent vector length used in autovectorizer for AVX-512

[Bug c/99100] New: Inconsistent vector length used in autovectorizer for AVX-512

[shibatch.sf.net at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99100

            Bug ID: 99100
           Summary: Inconsistent vector length used in autovectorizer for
                    AVX-512
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: shibatch.sf.net at gmail dot com
  Target Milestone: ---

When AVX512 instruction is available, the auto-vectorizer in gcc 
sometimes generates calls to AVX2 functions instead of AVX512 functions.
-mprefer-vector-width=512 does not affect the result.


$ cat vabitest.c
#include <stdio.h>
#include <math.h>

_Pragma ("omp declare simd simdlen(8) notinbranch") 
__attribute__((const)) double myfunc(double x);

#define N 1024
__attribute__ ((__aligned__(256))) double a[N], b[N], c[N];

int main(void) {
   for (int i = 0; i < N; i++) a[i] = myfunc(b[i]);
   for (int i = 0; i < N; i++) c[i] = sin(b[i]);
}

$ gcc-10 -ffast-math -O3 -mavx512f -fopenmp vabitest.c -S -o- | grep _ZGV
         call    _ZGVdN8v_myfunc@PLT
         call    _ZGVeN8v_sin@PLT

[Bug tree-optimization/99100] Inconsistent vector length used in autovectorizer for AVX-512

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99100

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-02-15
           Keywords|                            |missed-optimization
          Component|target                      |tree-optimization
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
             Blocks|                            |53947

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  I suspect the logic selecting the overload to use in the vectorizer
is flawed.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations

[Bug target/99100] Inconsistent vector length used in autovectorizer for AVX-512

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99100

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |target

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
What's odd is that with -mavx512f -mtune=skylake-avx512 we don't vectorize it
at all.  I'm looking at a smaller test:

#pragma omp declare simd simdlen(8) notinbranch
extern double __attribute__((const)) myfunc(double x);

#define N 1024
__attribute__ ((__aligned__(256))) double a[N], b[N], c[N];

void foo()
{
   for (int i = 0; i < N; i++) a[i] = myfunc(b[i]);
}


we have 3 simd clones here and the only discriminator is the target badness
returned by targetm.simd_clone.usable (n).  I suspect

  switch (node->simdclone->vecsize_mangle)
    {
    case 'b':
      if (!TARGET_SSE2)
        return -1;
      if (!TARGET_AVX)
        return 0;
      return TARGET_AVX2 ? 2 : 1;
    case 'c':
      if (!TARGET_AVX)
        return -1;
      return TARGET_AVX2 ? 1 : 0;
    case 'd':
      if (!TARGET_AVX2)
        return -1;
      return 0;
    case 'e':
      if (!TARGET_AVX512F)
        return -1;
      return 0;

since 'd' behaves the same as 'e' here we prefer 'd' or 'e' dependent on
the order in the simd clones.

Note the odd thing is we produce a 'c' mangling clone _and_ a 'd' mangling
one.  But indeed once we reach 'd' badness is zero and can't improve anymore.

There's also no sign of handling of prefered SIMD modes in the above
function...

Target issue.  Not sure about the -mtune=skylake-avx512 issue though.

[Bug target/99100] Inconsistent vector length used in autovectorizer for AVX-512

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99100

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 50184
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50184&action=edit
gcc11-pr99100.patch

Untested fix.

[Bug target/99100] Inconsistent vector length used in autovectorizer for AVX-512

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99100

--- Comment #4 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:1531f39268c8973cf9478585fba5c5bbdb6e9c4c

commit r11-7253-g1531f39268c8973cf9478585fba5c5bbdb6e9c4c
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Feb 16 08:59:03 2021 +0100

    openmp: Fix up vectorization simd call badness computation [PR99100]

    As mentioned in the PR, ix86_simd_clone_usable didn't make it more
desirable
    to use 'e' mangled AVX512F entrypoints over 'd' mangled ones (AVX2) with
the
    same simdlen.  This patch fixes that.  I have tweaked the generic code too
    to make more room for these target specific badness factors.

    2021-02-16  Jakub Jelinek  <jakub@redhat.com>

            PR target/99100
            * tree-vect-stmts.c (vectorizable_simd_clone_call): For num_calls
!= 1
            multiply by 4096 and for inbranch by 8192.
            * config/i386/i386.c (ix86_simd_clone_usable): For TARGET_AVX512F,
            return 3, 2 or 1 for mangle letters 'b', 'c' or 'd'.

            * gcc.target/i386/pr99100.c: New test.

[Bug target/99100] Inconsistent vector length used in autovectorizer for AVX-512

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99100

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed on the trunk, not sure about backporting yet.