public inbox for gcc-bugs@sourceware.org
help / color / mirror / Atom feed
* [Bug c/94366] New: OpenMP Parallel Reduction with "&&" operator does not compute correct result
@ 2020-03-27 18:27 satzeni at nvidia dot com
  2021-06-30  7:55 ` [Bug c/94366] " pal0009 at uah dot edu
                   ` (7 more replies)
  0 siblings, 8 replies; 9+ messages in thread
From: satzeni at nvidia dot com @ 2020-03-27 18:27 UTC (permalink / raw)
  To: gcc-bugs

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94366

            Bug ID: 94366
           Summary: OpenMP Parallel Reduction with "&&" operator does not
                    compute correct result
           Product: gcc
           Version: 9.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: satzeni at nvidia dot com
  Target Milestone: ---

Created attachment 48137
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48137&action=edit
reproducer

Consider this program:

#include <assert.h>
#include <omp.h>
#include <stdio.h>

static int expect = 1;
static int result;

int main() {
    int a = 2;  /* or any another even number */

#pragma omp parallel reduction(&& : a)
    {
        a = a && 1;
    }
    result = a ? 1 : 0;
    assert(result == 1);
}

Compiler and run with:

$ gcc -fopenmp test.c
$ OMP_NUM_THREADS=2 ./a.out

GCC is not correctly computing the logical-and reduction on variable "a".

Inside the parallel region "a" value is 1, but right after the parallel region
is 0 failing the assertion. The correct result should be 1 as the initial value
is 2 and the logical-and between 2 && 1 should be non zero.

Looking at the assembly code, GCC is using a bitwise “and” instruction to
perform the “&&” operation:

<main._omp_fn.0+26>    mov    (%rbx),%edx
<main._omp_fn.0+28>    jmp    0x400782 <main._omp_fn.0+34>
<main._omp_fn.0+30>    xchg   %ax,%ax
<main._omp_fn.0+32>    mov    %eax,%edx
<main._omp_fn.0+34>    mov    %edx,%ecx
<main._omp_fn.0+36>    mov    %edx,%eax
<main._omp_fn.0+38>    and    $0x1,%ecx
<main._omp_fn.0+41>    lock cmpxchg %ecx,(%rbx)

However the the value of “a” is not reduced to “1” before it is passed to the
parallel region.

<main+15>              movl   $0x2,(%rsp)
<main+22>              callq  0x4005c0 <GOMP_parallel_start@plt>

^ permalink raw reply	[flat|nested] 9+ messages in thread
end of thread, other threads:[~2022-05-11  6:33 UTC | newest]

Thread overview: 9+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-03-27 18:27 [Bug c/94366] New: OpenMP Parallel Reduction with "&&" operator does not compute correct result satzeni at nvidia dot com
2021-06-30  7:55 ` [Bug c/94366] " pal0009 at uah dot edu
2021-06-30 10:33 ` [Bug middle-end/94366] " jakub at gcc dot gnu.org
2021-07-01  7:00 ` cvs-commit at gcc dot gnu.org
2021-07-01  7:54 ` jakub at gcc dot gnu.org
2021-07-18 23:29 ` cvs-commit at gcc dot gnu.org
2022-05-10  8:19 ` cvs-commit at gcc dot gnu.org
2022-05-11  6:21 ` cvs-commit at gcc dot gnu.org
2022-05-11  6:33 ` jakub at gcc dot gnu.org

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).