[Bug c/60117] New: simd reduction clause suppresses simd auto-vectorization when -fopenmp is set

[Bug c/60117] New: simd reduction clause suppresses simd auto-vectorization when -fopenmp is set

[tprince at computer dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60117

            Bug ID: 60117
           Summary: simd reduction clause suppresses simd
                    auto-vectorization when -fopenmp is set
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tprince at computer dot org

Created attachment 32082
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=32082&action=edit
source code reproducer

gcc version 4.9.0 20140203
gcc -O2 -ftree-vectorize -std=c99 -march=core-avx2 -fopt-info -S -fopenmp
s314.c
This uses vmaxss instruction in the main loop body, in spite of the fairly
positive vectorization report.
-O3 makes no significant difference, so -O2 is used in practice for stability
elsewhere in gcc source code.
If the omp simd is disabled by removing -fopenmp, excellent code is produced
using vmaxps.  Performance test shows 10x speedup with max-unroll-times=2.

[Bug c/60117] simd reduction clause suppresses simd auto-vectorization when -fopenmp is set

[tprince at computer dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60117

--- Comment #1 from tprince at computer dot org ---
Created attachment 32083
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=32083&action=edit
f2c style include file

[Bug c/60117] simd reduction clause suppresses simd auto-vectorization when -fopenmp is set

[tprince at computer dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60117

--- Comment #2 from tprince at computer dot org ---
Jakub Jelinek requested this PR filing

[Bug c/60117] simd reduction clause suppresses simd auto-vectorization when -fopenmp is set

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60117

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
If the inner loop body is written as
x = a[i__] > x ? a[i__] : x;
then we actually do vectorize with -Ofast -fopenmp-simd, but for some reason
the vectorized loop is optimized away as dead later on.
For vectorization of the original loop, not sure if we can safely transform omp
simd array conditional stores into unconditional read + conditional operation +
unconditional store, or if we should (and why don't we) transform those into
MASK_STORE.