Hi!
If collapse_bb != NULL, then cont_bb isn't a latch of schedule(static,C)
collapse(N) omp for loops, and the latch for simd loops was only correct
if the loop contained just two basic blocks and not more.
Fixed thusly, bootstrapped/regtested on x86_64-linux, committed to trunk.
2013-11-26 Jakub Jelinek <ja...@redhat.com>
PR middle-end/59152
* omp-low.c (expand_omp_for_static_chunk): Don't set loop->latch
for the inner loop if collapse_bb is non-NULL.
(expand_omp_simd): Use cont_bb rather than e->dest as latch.
* c-c++-common/gomp/pr59152.c: New test.
--- gcc/omp-low.c.jj 2013-11-25 18:30:18.000000000 +0100
+++ gcc/omp-low.c 2013-11-26 14:20:46.137171434 +0100
@@ -6491,7 +6491,8 @@ expand_omp_for_static_chunk (struct omp_
{
struct loop *loop = alloc_loop ();
loop->header = body_bb;
- loop->latch = cont_bb;
+ if (collapse_bb == NULL)
+ loop->latch = cont_bb;
add_loop (loop, trip_loop);
}
}
@@ -6771,7 +6772,7 @@ expand_omp_simd (struct omp_region *regi
{
struct loop *loop = alloc_loop ();
loop->header = l1_bb;
- loop->latch = e->dest;
+ loop->latch = cont_bb;
add_loop (loop, l1_bb->loop_father);
if (safelen == NULL_TREE)
loop->safelen = INT_MAX;
--- gcc/testsuite/c-c++-common/gomp/pr59152.c.jj 2013-11-26
14:15:32.099779366 +0100
+++ gcc/testsuite/c-c++-common/gomp/pr59152.c 2013-11-26 13:59:36.000000000
+0100
@@ -0,0 +1,40 @@
+/* PR middle-end/59152 */
+/* { dg-do compile } */
+/* { dg-options "-fopenmp -fipa-pure-const" } */
+
+extern int b[];
+void
+foo (void)
+{
+ unsigned long v1, v2, v3;
+ #pragma omp parallel for schedule(static, 32) collapse(3)
+ for (v1 = 0; v1 < 20; v1 += 2)
+ for (v2 = __LONG_MAX__; v2 > __LONG_MAX__ - 30; v2 -= 3)
+ for (v3 = 10; v3 > 0; v3--)
+ #pragma omp atomic
+ b[v3]++;
+}
+
+void
+bar (void)
+{
+ unsigned long v1, v2, v3;
+ #pragma omp parallel for schedule(static) collapse(3)
+ for (v1 = 0; v1 < 20; v1 += 2)
+ for (v2 = __LONG_MAX__; v2 > __LONG_MAX__ - 30; v2 -= 3)
+ for (v3 = 10; v3 > 0; v3--)
+ #pragma omp atomic
+ b[v3]++;
+}
+
+void
+baz (void)
+{
+ unsigned long v1, v2, v3;
+ #pragma omp parallel for schedule(runtime) collapse(3)
+ for (v1 = 0; v1 < 20; v1 += 2)
+ for (v2 = __LONG_MAX__; v2 > __LONG_MAX__ - 30; v2 -= 3)
+ for (v3 = 10; v3 > 0; v3--)
+ #pragma omp atomic
+ b[v3]++;
+}
Jakub
