[Bug tree-optimization/97707] New: avx12 math function invoked even if -mprefer-vector-width=256 specified

[Bug tree-optimization/97707] New: avx12 math function invoked even if -mprefer-vector-width=256 specified

[vincenzo.innocente at cern dot ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97707

            Bug ID: 97707
           Summary: avx12 math function invoked even if
                    -mprefer-vector-width=256 specified
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vincenzo.innocente at cern dot ch
  Target Milestone: ---

this code will invoke _ZGVeN8v_sin instead of _ZGVdN4v_sin making use of zmm
registers
#include<cmath>

int main() {

  double res=0;

  for (int x=0; x<1024;x++) {
    double y = x; 
    res += std::sin(y);
  }


 return res > 0.5;

}

NOTE if I specify
for (long long x=0; x<1024;x++) {

it will correcty invoke _ZGVdN4v_sin (no zmm)


compiler options
-Ofast -march=skylake-avx512 -mprefer-vector-width=256

[Bug tree-optimization/97707] avx512 math function invoked even if -mprefer-vector-width=256 specified

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97707

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
                 CC|                            |jakub at gcc dot gnu.org
             Blocks|                            |53947
             Target|                            |x86_64-*-* i?86-*-*
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2020-11-04

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
-mprefer-vector-width=256 means it _prefers_ AVX256 but then in case
vectorization fails (for whatever bogus reason) with AVX256 it will still try
AVX512:

static unsigned int
ix86_autovectorize_vector_modes (vector_modes *modes, bool all)
{
  if (TARGET_AVX512F && !TARGET_PREFER_AVX256)
    {
      modes->safe_push (V64QImode);
      modes->safe_push (V32QImode);
      modes->safe_push (V16QImode);
    }
  else if (TARGET_AVX512F && all)
    {
      modes->safe_push (V32QImode);
      modes->safe_push (V16QImode);
      modes->safe_push (V64QImode);
    }

the odd thing is that the dumps suggest we use V4DFmode.  We use a VF of 8
because of V8SI and it seems that call vectorization then chooses the
larger vector mode because it is supported - it seems to have its own
idea of costing there (probably reasonable).

Do you think this is a bad idea?


Referenced Bugs:

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

[Bug tree-optimization/97707] avx512 math function invoked even if -mprefer-vector-width=256 specified

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97707

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The prefer option is exactly that, try that first and if it fails, use the
others.
So I guess the main question is why it failed for the VF=4 case.
E.g. for -mavx512f it wouldn't surprise me as much, because then AVX512VL is
not in and it could be that AVX2 isn't able to deal with that.

[Bug tree-optimization/97707] avx512 math function invoked even if -mprefer-vector-width=256 specified

[vincenzo.innocente at cern dot ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97707

--- Comment #3 from vincenzo Innocente <vincenzo.innocente at cern dot ch> ---
the main point in using -mprefer-vector-width=256 is to avoid clock throttling
in "mixed" workloads.
In small benchmarks like this one avx512 is faster (even on an old Silver) even
if trigger a slower clock. (and the test should be performed with the machine
fully loaded). Still if I ask  -mprefer-vector-width=256 I would like to see no
512-wide instructions to be used.

A disturbing feature is also the difference between using int or long long as
loop index.

[Bug tree-optimization/97707] avx512 math function invoked even if -mprefer-vector-width=256 specified

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97707

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
If you want no AVX512* instructions, the option to say that is -mno-avx512f