On 08/12/15 10:22, Richard Biener wrote:
On Tue, 8 Dec 2015, Jakub Jelinek wrote:
On Tue, Dec 08, 2015 at 09:49:49AM +0100, Tom de Vries wrote:
diff --git a/gcc/omp-low.c b/gcc/omp-low.c
index d1d1e3c..ac4a94d 100644
--- a/gcc/omp-low.c
+++ b/gcc/omp-low.c
@@ -1389,6 +1389,12 @@ install_var_field (tree var, bool by_ref, int mask,
omp_context *ctx,
|| !is_gimple_omp_oacc (ctx->stmt));
type = TREE_TYPE (var);
+ /* Prevent redeclaring the var in the split-off function with a restrict
+ pointer type. Note that we only clear type itself, restrict qualifiers in
+ the pointed-to type will be ignored by points-to analysis. */
+ if (POINTER_TYPE_P (type))
+ type = build_qualified_type (type, TYPE_QUALS (type) &
~TYPE_QUAL_RESTRICT);
Is it necessary to call build_qualified_type in the common case (when it is
not restrict)? I'd think if (POINTER_TYPE_P (type) && TYPE_RESTRICT (type))
might be better.
Will do.
That said, I forgot when exactly are the cliques computed, it would be nice
if then computing those and seeing a
GOMP_parallel/GOACC_parallel_keyed/GOMP_target_ext/GOMP_task/GOMP_taskloop*
builtins we would just continue walking the outlined functions they refer to
as if it was a part of the current function for the purpose of the restrict
computation.
Cliques are not computed in IPA PTA and the above would require IPA.
FTR, this is a generic (not openmp-related) example, and AFAIU a more
generic form of the issue that is being raised.
...
static void __attribute__((noinline, noclone))
bar (int *a, int * b)
{
*a = 1;
*b = 2;
*a = *a + 1;
}
void
foo (int *__restrict__ a, int *__restrict__ b)
{
bar (a, b);
}
...
Thanks,
- Tom
