diff mbox

Fix SLP vectorization ICE (PR tree-optimization/71823)

Message ID 20160711192843.GO7387@tucnak.redhat.com
State New
Headers show

Commit Message

Jakub Jelinek July 11, 2016, 7:28 p.m. UTC
Hi!

vect_get_vec_defs handles only one or two arguments, for ternary
ops like fma vectorizable_operation calls vect_get_vec_defs on the
first two arguments, and then did
             vec_oprnds2.create (1);
             vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2,   
                                                                   stmt));
which is what vect_get_vec_defs does, but only for loop vectorization.
The following patch uses vect_get_vec_defs again so that it does what it
did before for loop vectorization, but additionally handles properly
slp vectorization.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk/6.2?

2016-07-11  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/71823
	* tree-vect-stmts.c (vectorizable_operation): Use vect_get_vec_defs
	to get vec_oprnds2 from op2.

	* gcc.dg/vect/pr71823.c: New test.


	Jakub

Comments

Richard Biener July 11, 2016, 8:07 p.m. UTC | #1
On July 11, 2016 9:28:43 PM GMT+02:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>vect_get_vec_defs handles only one or two arguments, for ternary
>ops like fma vectorizable_operation calls vect_get_vec_defs on the
>first two arguments, and then did
>             vec_oprnds2.create (1);
>          vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2,   
>                                                                stmt));
>which is what vect_get_vec_defs does, but only for loop vectorization.
>The following patch uses vect_get_vec_defs again so that it does what
>it
>did before for loop vectorization, but additionally handles properly
>slp vectorization.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for
>trunk/6.2?

OK.

Thanks,
Richard.

>2016-07-11  Jakub Jelinek  <jakub@redhat.com>
>
>	PR tree-optimization/71823
>	* tree-vect-stmts.c (vectorizable_operation): Use vect_get_vec_defs
>	to get vec_oprnds2 from op2.
>
>	* gcc.dg/vect/pr71823.c: New test.
>
>--- gcc/tree-vect-stmts.c.jj	2016-07-07 20:40:43.000000000 +0200
>+++ gcc/tree-vect-stmts.c	2016-07-11 12:05:52.501823209 +0200
>@@ -5362,11 +5362,8 @@ vectorizable_operation (gimple *stmt, gi
> 	    vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
> 			       slp_node, -1);
> 	  if (op_type == ternary_op)
>-	    {
>-	      vec_oprnds2.create (1);
>-	      vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2,
>-		                                                    stmt));
>-	    }
>+	    vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL,
>+			       slp_node, -1);
> 	}
>       else
> 	{
>--- gcc/testsuite/gcc.dg/vect/pr71823.c.jj	2016-07-11
>12:07:41.687485519 +0200
>+++ gcc/testsuite/gcc.dg/vect/pr71823.c	2016-07-11 12:08:42.449741939
>+0200
>@@ -0,0 +1,14 @@
>+/* PR tree-optimization/71823 */
>+/* { dg-do compile } */
>+/* { dg-additional-options "-mfma" { target i?86-*-* x86_64-*-* } } */
>+
>+float a[4], b[4];
>+
>+int
>+main ()
>+{
>+  int i;
>+  for (i = 0; i < 4; ++i)
>+    b[i] = __builtin_fma (1024.0f, 1024.0f, a[i]);
>+  return 0;
>+}
>
>	Jakub
diff mbox

Patch

--- gcc/tree-vect-stmts.c.jj	2016-07-07 20:40:43.000000000 +0200
+++ gcc/tree-vect-stmts.c	2016-07-11 12:05:52.501823209 +0200
@@ -5362,11 +5362,8 @@  vectorizable_operation (gimple *stmt, gi
 	    vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
 			       slp_node, -1);
 	  if (op_type == ternary_op)
-	    {
-	      vec_oprnds2.create (1);
-	      vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2,
-		                                                    stmt));
-	    }
+	    vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL,
+			       slp_node, -1);
 	}
       else
 	{
--- gcc/testsuite/gcc.dg/vect/pr71823.c.jj	2016-07-11 12:07:41.687485519 +0200
+++ gcc/testsuite/gcc.dg/vect/pr71823.c	2016-07-11 12:08:42.449741939 +0200
@@ -0,0 +1,14 @@ 
+/* PR tree-optimization/71823 */
+/* { dg-do compile } */
+/* { dg-additional-options "-mfma" { target i?86-*-* x86_64-*-* } } */
+
+float a[4], b[4];
+
+int
+main ()
+{
+  int i;
+  for (i = 0; i < 4; ++i)
+    b[i] = __builtin_fma (1024.0f, 1024.0f, a[i]);
+  return 0;
+}