As the following example shows, this optimization cannot be performed legally
if the parallel loop spawns tasks which reference vars local to the parallel
region. Without the optimization, all tasks would be executed at the implicit
barrier at the end of the loop (where "array" is still intact) rather than the
barrier at the end of the parallel region (thread 0's stack has been popped and
"array" deallocated).
% gcc -v
Using built-in specs.
Target: i686-pc-linux-gnu
Configured with: /net/gnu-13/export/gnu/src/gcc/gcc/configure --enable
clocale=gnu --with-system-zlib --enable-checking=assert --with-demangler-in-ld
--enable-shared --enable-threads=posix --enable-haifa --prefix=/usr/gcc4.4
--with-local-prefix=/usr/local
Thread model: posix
gcc version 4.4.0 20090311 (experimental) [trunk revision 144788] (GCC)
% cat test.c
#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
int flag = 0;
int foo(int *array)
{
int i;
printf("array = %p\n", array);
for (i = 0; i < 10; i++) {
printf("i: %d 0x%x\n", i, array[i]);
}
}
int main(int argc, char *argv[])
{
#pragma omp parallel shared(argc)
{
int array[10];
int i;
for (i = 0; i < 10; i++) {
array[i] = 0x55555555;
}
#pragma omp for schedule(dynamic)
for (i = 0; i < argc; i++)
#pragma omp task default(shared) firstprivate(i)
{
if (omp_get_thread_num() == 0) {
if (! flag) {
#pragma omp critical(cs)
foo(array);
flag = 1;
}
}
else {
sleep(1);
}
#pragma omp critical(cs)
printf("i = %d\n", i);
}
}
}
%
% gcc -fopenmp test.c
% a.out x x x
array = 0xf75bf380
i: 0 0x55555555
i: 1 0x55555555
i: 2 0x55555555
i: 3 0x1312d00
i: 4 0x55555555
i: 5 0x55555555
i: 6 0x55555555
i: 7 0xf7ffc908
i: 8 0xffffffcc
i: 9 0x804872c
i = 3
i = 2
i = 1
i = 0
% nm a.out | grep GOMP_loop
U GOMP_loop_dynamic_next@@GOMP_1.0
U GOMP_loop_dynamic_start@@GOMP_1.0
U GOMP_loop_end_nowait@@GOMP_1.0
%
--
Summary: GOMP_loop_end illegally optmized into
GOMP_loop_end_nowait
Product: gcc
Version: 4.4.0
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: other
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: brian dot e dot bliss at intel dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39591
