This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[committed] Fix OpenMP array reduction handling in orphaned workshare (PR middle-end/80853)
- From: Jakub Jelinek <jakub at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: Mon, 22 May 2017 20:58:29 +0200
- Subject: [committed] Fix OpenMP array reduction handling in orphaned workshare (PR middle-end/80853)
- Authentication-results: sourceware.org; auth=none
- Authentication-results: ext-mx02.extmail.prod.ext.phx2.redhat.com; dmarc=none (p=none dis=none) header.from=redhat.com
- Authentication-results: ext-mx02.extmail.prod.ext.phx2.redhat.com; spf=pass smtp.mailfrom=jakub at redhat dot com
- Dkim-filter: OpenDKIM Filter v2.11.0 mx1.redhat.com BCA4580C20
- Dmarc-filter: OpenDMARC Filter v1.3.2 mx1.redhat.com BCA4580C20
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
For pointer based array sections, we are not requiring the pointer to be
shared in outer context and it might be not shared (e.g. function argument
or just a private pointer initialized to point to something shared etc.),
which is right, but we shouldn't require it during lowering either.
This patch fixes ICE on it. Bootstrapped/regtested on x86_64-linux and
i686-linux, committed to trunk, queued for release branches.
2017-05-22 Jakub Jelinek <jakub@redhat.com>
PR middle-end/80853
* omp-low.c (lower_reduction_clauses): Pass OMP_CLAUSE_PRIVATE
as last argument to build_outer_var_ref for pointer bases of array
section reductions.
* testsuite/libgomp.c/pr80853.c: New test.
--- gcc/omp-low.c.jj 2017-04-21 08:14:29.436790613 +0200
+++ gcc/omp-low.c 2017-05-22 11:34:51.085746641 +0200
@@ -5140,15 +5140,25 @@ lower_reduction_clauses (tree clauses, g
if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
continue;
+ enum omp_clause_code ccode = OMP_CLAUSE_REDUCTION;
orig_var = var = OMP_CLAUSE_DECL (c);
if (TREE_CODE (var) == MEM_REF)
{
var = TREE_OPERAND (var, 0);
if (TREE_CODE (var) == POINTER_PLUS_EXPR)
var = TREE_OPERAND (var, 0);
- if (TREE_CODE (var) == INDIRECT_REF
- || TREE_CODE (var) == ADDR_EXPR)
+ if (TREE_CODE (var) == ADDR_EXPR)
var = TREE_OPERAND (var, 0);
+ else
+ {
+ /* If this is a pointer or referenced based array
+ section, the var could be private in the outer
+ context e.g. on orphaned loop construct. Pretend this
+ is private variable's outer reference. */
+ ccode = OMP_CLAUSE_PRIVATE;
+ if (TREE_CODE (var) == INDIRECT_REF)
+ var = TREE_OPERAND (var, 0);
+ }
orig_var = var;
if (is_variable_sized (var))
{
@@ -5162,7 +5172,7 @@ lower_reduction_clauses (tree clauses, g
new_var = lookup_decl (var, ctx);
if (var == OMP_CLAUSE_DECL (c) && omp_is_reference (var))
new_var = build_simple_mem_ref_loc (clause_loc, new_var);
- ref = build_outer_var_ref (var, ctx);
+ ref = build_outer_var_ref (var, ctx, ccode);
code = OMP_CLAUSE_REDUCTION_CODE (c);
/* reduction(-:var) sums up the partial results, so it acts
--- libgomp/testsuite/libgomp.c/pr80853.c.jj 2017-05-22 11:28:50.060287195 +0200
+++ libgomp/testsuite/libgomp.c/pr80853.c 2017-05-22 11:28:29.000000000 +0200
@@ -0,0 +1,29 @@
+/* PR middle-end/80853 */
+/* { dg-do run } */
+
+__attribute__((noinline, noclone)) void
+foo (int *p)
+{
+ #pragma omp for reduction(+:p[:4])
+ for (int i = 0; i < 64; i++)
+ {
+ p[0] += i;
+ p[1] += i / 2;
+ p[2] += 2 * i;
+ p[3] += 3 * i;
+ }
+}
+
+int
+main ()
+{
+ int p[4] = { 0, 0, 0, 0 };
+ #pragma omp parallel
+ foo (p);
+ if (p[0] != 63 * 64 / 2
+ || p[1] != 31 * 32
+ || p[2] != 63 * 64
+ || p[3] != 3 * 63 * 64 / 2)
+ __builtin_abort ();
+ return 0;
+}
Jakub