Hi!
We split OMP_TASKLOOP into 3 constructs, two GIMPLE_OMP_FOR
with GIMPLE_OMP_PARALLEL sandwiched in between them, so that
it is possible to compute number of iterations etc. before calling
GOMP_taskloop*. Using the original iterator in the outer gfor
doesn't play very well if the taskloop region is nested in other
OpenMP regions like parallel. This patch just creates a temporary
for that.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed
to trunk so far, queued for 6.x.
2017-03-08 Jakub Jelinek <ja...@redhat.com>
PR c/79940
* gimplify.c (gimplify_omp_for): Replace index var in outer
taskloop statement with an artificial variable and add
OMP_CLAUSE_PRIVATE clause for it.
* testsuite/libgomp.c/pr79940.c: New test.
--- gcc/gimplify.c.jj 2017-02-21 09:03:57.000000000 +0100
+++ gcc/gimplify.c 2017-03-08 10:06:11.926501447 +0100
@@ -10232,8 +10232,9 @@ gimplify_omp_for (tree *expr_p, gimple_s
gimple_omp_for_set_combined_into_p (gfor, true);
for (i = 0; i < (int) gimple_omp_for_collapse (gfor); i++)
{
- t = unshare_expr (gimple_omp_for_index (gfor, i));
- gimple_omp_for_set_index (gforo, i, t);
+ tree type = TREE_TYPE (gimple_omp_for_index (gfor, i));
+ tree v = create_tmp_var (type);
+ gimple_omp_for_set_index (gforo, i, v);
t = unshare_expr (gimple_omp_for_initial (gfor, i));
gimple_omp_for_set_initial (gforo, i, t);
gimple_omp_for_set_cond (gforo, i,
@@ -10241,7 +10242,13 @@ gimplify_omp_for (tree *expr_p, gimple_s
t = unshare_expr (gimple_omp_for_final (gfor, i));
gimple_omp_for_set_final (gforo, i, t);
t = unshare_expr (gimple_omp_for_incr (gfor, i));
+ gcc_assert (TREE_OPERAND (t, 0) == gimple_omp_for_index (gfor, i));
+ TREE_OPERAND (t, 0) = v;
gimple_omp_for_set_incr (gforo, i, t);
+ t = build_omp_clause (input_location, OMP_CLAUSE_PRIVATE);
+ OMP_CLAUSE_DECL (t) = v;
+ OMP_CLAUSE_CHAIN (t) = gimple_omp_for_clauses (gforo);
+ gimple_omp_for_set_clauses (gforo, t);
}
gimplify_seq_add_stmt (pre_p, gforo);
}
--- libgomp/testsuite/libgomp.c/pr79940.c.jj 2017-03-08 10:47:26.179154442
+0100
+++ libgomp/testsuite/libgomp.c/pr79940.c 2017-03-08 10:46:46.000000000
+0100
@@ -0,0 +1,47 @@
+/* PR c/79940 */
+
+int
+main ()
+{
+ int i, j, l, m;
+ int a[10000], b[10000], c[10000];
+ for (i = 0; i < 10000; i++)
+ {
+ a[i] = i;
+ b[i] = i & 31;
+ }
+#pragma omp parallel shared(a, b, c)
+#pragma omp single
+#pragma omp taskloop shared(a, b, c)
+ for (i = 0; i < 10000; i++)
+ c[i] = a[i] + b[i];
+#pragma omp parallel
+#pragma omp single
+ {
+ #pragma omp taskloop shared(a, b, c) lastprivate (i)
+ for (i = 0; i < 10000; i++)
+ c[i] += a[i] + b[i];
+ l = i;
+ }
+#pragma omp parallel
+#pragma omp single
+#pragma omp taskloop shared(a, b, c) collapse(2)
+ for (i = 0; i < 100; i++)
+ for (j = 0; j < 100; j++)
+ c[i * 100 + j] += a[i * 100 + j] + b[i * 100 + j];
+#pragma omp parallel
+#pragma omp single
+ {
+ #pragma omp taskloop shared(a, b, c) lastprivate (i, j)
+ for (i = 0; i < 100; i++)
+ for (j = 0; j < 100; j++)
+ c[i * 100 + j] += a[i * 100 + j] + b[i * 100 + j];
+ m = i * 100 + j;
+ }
+ for (i = 0; i < 10000; i++)
+ if (a[i] != i || b[i] != (i & 31) || c[i] != 4 * i + 4 * (i & 31))
+ __builtin_abort ();
+ if (l != 10000 || m != 10100)
+ __builtin_abort ();
+ return 0;
+}
Jakub
