diff mbox series

[committed] Improve simd with a single lastprivate iterator

Message ID 20190720064318.GA2125@tucnak
State New
Headers show
Series [committed] Improve simd with a single lastprivate iterator | expand

Commit Message

Jakub Jelinek July 20, 2019, 6:43 a.m. UTC 
Hi!

While the iterator of a simd collapse(1) loop is predetermined linear, in
OpenMP 5 one can specify it also explicitly in a linear, lastprivate or
private clause.  The following testcase shows that we weren't vectorizing
those if the iterator wasn't addressable and has been explicitly
lastprivate, as the magic simd arrays prevented number of iterations
computation.  Fixed by not forcing it into a simd array in that case.

Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.

2019-07-20  Jakub Jelinek  <jakub@redhat.com>

	* omp-low.c (lower_rec_input_clauses): Don't force simd arrays for
	lastprivate non-addressable iterator of a collapse(1) simd.

	* gcc.dg/vect/vect-simd-16.c: New test.


	Jakub
diff mbox series

Patch

--- gcc/omp-low.c.jj	2019-07-19 13:25:52.001314547 +0200
+++ gcc/omp-low.c	2019-07-19 17:01:17.168704782 +0200
@@ -5123,7 +5123,10 @@  lower_rec_input_clauses (tree clauses, g
 		{
 		  tree y = lang_hooks.decls.omp_clause_dtor (c, new_var);
 		  if ((TREE_ADDRESSABLE (new_var) || nx || y
-		       || OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE
+		       || (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE
+			   && (gimple_omp_for_collapse (ctx->stmt) != 1
+			       || (gimple_omp_for_index (ctx->stmt, 0)
+				   != new_var)))
 		       || OMP_CLAUSE_CODE (c) == OMP_CLAUSE__CONDTEMP_
 		       || omp_is_reference (var))
 		      && lower_rec_simd_input_clauses (new_var, ctx, &sctx,
--- gcc/testsuite/gcc.dg/vect/vect-simd-16.c.jj	2019-07-19 17:16:24.035069658 +0200
+++ gcc/testsuite/gcc.dg/vect/vect-simd-16.c	2019-07-19 17:20:48.034099757 +0200
@@ -0,0 +1,61 @@ 
+/* { dg-additional-options "-fopenmp-simd" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[1-3] loops" 3 "vect" { target i?86-*-* x86_64-*-* } } } */
+
+#include "tree-vect.h"
+
+__attribute__((noipa)) int
+foo (int *a)
+{
+  int i;
+  #pragma omp simd lastprivate (i)
+  for (i = 0; i < 64; i++)
+    a[i] = i;
+  return i;
+}
+
+__attribute__((noipa)) void
+bar (int *a)
+{
+  int i;
+  #pragma omp simd private (i)
+  for (i = 0; i < 64; i++)
+    a[i] = i + 1;
+}
+
+__attribute__((noipa)) int
+baz (int *a)
+{
+  int i;
+  #pragma omp simd linear (i)
+  for (i = 0; i < 64; i++)
+    a[i] = i + 2;
+  return i;
+}
+
+int
+main ()
+{
+  int i;
+  int a[64];
+  check_vect ();
+  if (foo (a) != 64)
+    abort ();
+  for (i = 0; i < 64; ++i)
+    if (a[i] != i)
+      abort ();
+    else
+      a[i] = -8;
+  bar (a);
+  for (i = 0; i < 64; ++i)
+    if (a[i] != i + 1)
+      abort ();
+    else
+      a[i] = -8;
+  if (baz (a) != 64)
+    abort ();
+  for (i = 0; i < 64; ++i)
+    if (a[i] != i + 2)
+      abort ();
+  return 0;
+}