[Bug c/66862] New: OpenMP SIMD does not work (use SIMD instructions) on conditional code

[Bug c/66862] New: OpenMP SIMD does not work (use SIMD instructions) on conditional code

[bill.jordan at intel dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66862

            Bug ID: 66862
           Summary: OpenMP SIMD does not work (use SIMD instructions) on
                    conditional code
           Product: gcc
           Version: 5.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: bill.jordan at intel dot com
  Target Milestone: ---

Created attachment 35971
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35971&action=edit
SIMD sample code

In the following loop,the omp simd pragma has no affect:

 #pragma omp simd private(value) linear(ij, kj) safelen(64)
 for (j = 0; j < switches; ++j) {
     value = cost[ik] + cost[kj];
     if (value < cost[ij]) {
         cost[ij] = value;
     }
     ++ij; ++kj;
 }

Without the "if" clause, the compiler will generate SIMD code.

Compilation line is:
 gcc -Wall -Wextra -O3 -save-temps -fopenmp-simd -c floyds_simd.c

Attached source includes 4 routines:
 floyds_wo_omp_simd: Base function.
 floyds_w_omp_simd: Base function with #pragma omp simd directive
 not_floyds_wo_omp_simd: Base function without if clause
 not_floyds_w_omp_simd: Base function without if clause with #pragma omp simd

[Bug c/66862] OpenMP SIMD does not work (use SIMD instructions) on conditional code

[bill.jordan at intel dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66862

--- Comment #1 from William Jordan <bill.jordan at intel dot com> ---
Created attachment 35972
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35972&action=edit
Disassembled code

Assembly shows that code generated with #pragma omp simd is the same as without
when the "for" loop contains an "if" statement. Without the "if" statement, the
#pragma omp simd uses SIMD operations seen in not_floyds_w_omp_simd function.

[Bug tree-optimization/66862] OpenMP SIMD does not work (use SIMD instructions) on conditional code

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66862

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

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

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
There is an optimization which should have converted the loop to use
uncondtional stores.  I wonder why it is not happening in this case.

[Bug tree-optimization/66862] OpenMP SIMD does not work (use SIMD instructions) on conditional code

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66862

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
store data-races

[Bug tree-optimization/66862] OpenMP SIMD does not work (use SIMD instructions) on conditional code

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66862

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

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

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
If you sed -i 's/short/int/' on the testcase, then e.g. with -mavx2 it is
vectorized with vmaskmovd.  But AVX2 does not have a masked store for packed
16-bit integers, and as Richard mentioned, using vpminuw/vmovdqu that icc emits
is IMHO invalid, as it introduces a store data race and I see no wording in the
OpenMP standard that would allow introducing store data races, even in omp simd
regions.

Now, it seems AVX512BW (and AVX512VL in some cases) has the needed
instructions,
in particular VMOVDQU{8,16}, but it is not reflected in maskload<mode> and
maskstore<mode> expanders.  CCing Kyrill and Uros on this.

[Bug tree-optimization/66862] OpenMP SIMD does not work (use SIMD instructions) on conditional code

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66862

--- Comment #5 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
> Now, it seems AVX512BW (and AVX512VL in some cases) has the needed
> instructions,
> in particular VMOVDQU{8,16}, but it is not reflected in maskload<mode> and
> maskstore<mode> expanders.  CCing Kyrill and Uros on this.

w/ -mavx512bw and -mavx512vl, the loop is vectorized since GCC 8.1.

[Bug tree-optimization/66862] OpenMP SIMD does not work (use SIMD instructions) on conditional code

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66862

Hongtao Liu <liuhongt at gcc dot gnu.org> changed:

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

--- Comment #6 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
Fixed since GCC8 w/ avx512bw avx512vl.
w/o avx512bw, x86 doesn't support packed int16 mask{load,store}, and can't
vectorize the loop.