[Bug c/66740] New: omp simd reduction miscompiles at -O3 with AVX (recent regression)

[Bug c/66740] New: omp simd reduction miscompiles at -O3 with AVX (recent regression)

[tprince at computer dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66740

            Bug ID: 66740
           Summary: omp simd reduction miscompiles at -O3 with AVX (recent
                    regression)
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tprince at computer dot org
  Target Milestone: ---

Created attachment 35897
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35897&action=edit
f2c include file

This regression is new in the last 2 weeks.
gcc -O3 -march=sandybridge -std=c99 -D__assume_aligned=  -fopenmp -c s115.c
is miscompiled (bad numerical results when run under test harness).
It is OK at -O2 or with -march=westmere, or with the omp directive disabled.
It doesn't vectorize without the omp directive.  It fails the same with
-march=haswell.
Similar versions using inner_product or Fortran dot_product are vectorizing
correctly without omp simd.

[Bug c/66740] omp simd reduction miscompiles at -O3 with AVX (recent regression)

[tprince at computer dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66740

--- Comment #1 from tprince at computer dot org ---
Created attachment 35898
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35898&action=edit
C source

[Bug c/66740] omp simd reduction miscompiles at -O3 with AVX (recent regression)

[tprince at computer dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66740

--- Comment #2 from tprince at computer dot org ---
$ gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/local/gcc6.0/libexec/gcc/x86_64-unknown-cygwin/6.0.0/lt
o-wrapper.exe
Target: x86_64-unknown-cygwin
Configured with: ../configure --prefix=/usr/local/gcc6.0 --enable-languages='c
c
++ fortran' --enable-libgomp --enable-threads=posix --with-dwarf2
--without-libi
conv-prefix --without-libintl-prefix --with-system-zlib --disable-werror
--witho
ut-cloog --without-isl
Thread model: posix
gcc version 6.0.0 20150630 (experimental) (GCC)

[Bug c/66740] omp simd reduction miscompiles at -O3 with AVX (recent regression)

[tprince at computer dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66740

--- Comment #3 from tprince at computer dot org ---
$ gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/local/gcc6.0/libexec/gcc/x86_64-unknown-cygwin/6.0.0/lt
o-wrapper.exe
Target: x86_64-unknown-cygwin
Configured with: ../configure --prefix=/usr/local/gcc6.0 --enable-languages='c
c
++ fortran' --enable-libgomp --enable-threads=posix --with-dwarf2
--without-libi
conv-prefix --without-libintl-prefix --with-system-zlib --disable-werror
--witho
ut-cloog --without-isl
Thread model: posix
gcc version 6.0.0 20150630 (experimental) (GCC)

[Bug tree-optimization/66740] [6 Regression] omp simd reduction miscompiles at -O3 with AVX (recent regression)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66740

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Target|                            |x86_64-*-*
          Component|c                           |tree-optimization
   Target Milestone|---                         |6.0
            Summary|omp simd reduction          |[6 Regression] omp simd
                   |miscompiles at -O3 with AVX |reduction miscompiles at
                   |(recent regression)         |-O3 with AVX (recent
                   |                            |regression)

[Bug tree-optimization/66740] [6 Regression] omp simd reduction miscompiles at -O3 with AVX (recent regression)

[tprince at computer dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66740

--- Comment #4 from tprince at computer dot org ---
It optimizes with implicit simd reduction and gives correct result with option
-ffast-math in place of -fopenmp.