[Bug libgomp/80394] New: Empty OpenMP task is wrongly removed when optimizing

[sergi.mateo at bsc dot es]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80394

            Bug ID: 80394
           Summary: Empty OpenMP task is wrongly removed when optimizing
           Product: gcc
           Version: 7.0.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libgomp
          Assignee: unassigned at gcc dot gnu.org
          Reporter: sergi.mateo at bsc dot es
                CC: jakub at gcc dot gnu.org
  Target Milestone: ---

Hi,

In the following code:

#include<unistd.h>
#include<stdio.h>

void foo(int foo_id) {
    int x = 0;
    for (int i = 0; i < 5; ++i)
    {
        #pragma omp task depend(inout: x) shared(x)
        {
            printf("%d -> hola : %d\n", foo_id, i);
            usleep(1000000);
            x++;
        }
    }

    #pragma omp task if(0) depend(inout: x)
    {                                                                           
        //printf("if\n");                                                       
    }                                                                           
}                                                                               

int main() {                                                                    
    #pragma omp parallel                                                        
    #pragma omp single                                                          
    {                                                                           
        foo(0);                                                                 
        foo(1);                                                                 
    }                                                                           
}


GCC wrongly removes the empty second task when optimizing (-O2 or higher).
Then, tasks created by the first call to 'foo' may be executed concurrently
with the tasks created by the second call, which is wrong: when a thread
reaches the empty task, the current task region should be suspended until all
the predecessor tasks of that task are executed.

Uncommenting the 'printf(...);' statement or compiling with a lower
optimization level make it work.

Thanks!,

Sergi