On Mon, 12 Jan 2015, Richard Biener wrote:
>
> I am testing the following patch to fix a latent bug in the vectorizer
> dealing with redundant DRs.
>
> Bootstrap and regtest pending on x86_64-unknown-linux-gnu.
Which shows the patch is bogus. Instead we are not prepared to
handle this situation. Thus the following patch rejects it, making
the testcase a runtime one as well.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied.
Richard.
2015-01-13 Richard Biener <rguent...@suse.de>
PR tree-optimization/64404
* tree-vect-stmts.c (vectorizable_load): Reject conflicting
SLP types for CSEd loads.
* gcc.dg/vect/pr64404.c: New testcase.
Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c (revision 219520)
+++ gcc/tree-vect-stmts.c (working copy)
@@ -5791,6 +5791,20 @@ vectorizable_load (gimple stmt, gimple_s
"group loads with negative dependence distance\n");
return false;
}
+
+ /* Similarly when the stmt is a load that is both part of a SLP
+ instance and a loop vectorized stmt via the same-dr mechanism
+ we have to give up. */
+ if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info)
+ && (STMT_SLP_TYPE (stmt_info)
+ != STMT_SLP_TYPE (vinfo_for_stmt
+ (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info)))))
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+ "conflicting SLP types for CSEd load\n");
+ return false;
+ }
}
Index: gcc/testsuite/gcc.dg/vect/pr64404.c
===================================================================
--- gcc/testsuite/gcc.dg/vect/pr64404.c (revision 0)
+++ gcc/testsuite/gcc.dg/vect/pr64404.c (working copy)
@@ -0,0 +1,59 @@
+/* { dg-do run } */
+/* { dg-additional-options "--param=sccvn-max-alias-queries-per-access=1" } */
+
+#include "tree-vect.h"
+
+extern void abort (void);
+
+typedef struct
+{
+ int l, h;
+} tFPinterval;
+
+tFPinterval X[1024];
+tFPinterval Y[1024];
+tFPinterval Z[1024];
+
+void __attribute__((noinline))
+Compute (void)
+{
+ int d;
+ for (d = 0; d < 1024; d++)
+ {
+ Y[d].l = X[d].l + X[d].h;
+ Y[d].h = Y[d].l;
+ Z[d].l = X[d].l;
+ Z[d].h = X[d].h;
+ }
+}
+
+int
+main (void)
+{
+ int d;
+
+ check_vect ();
+
+ for (d = 0; d < 1024; d++)
+ {
+ X[d].l = d;
+ X[d].h = d + 1;
+ __asm__ volatile ("");
+ }
+
+ Compute ();
+
+ for (d = 0; d < 1024; d++)
+ {
+ if (Y[d].l != X[d].l + X[d].h
+ || Y[d].h != Y[d].l
+ || Z[d].l != X[d].l
+ || Z[d].h != X[d].h)
+ abort ();
+ __asm__ volatile ("");
+ }
+
+ return 0;
+}
+
+/* { dg-final { cleanup-tree-dump "vect" } } */
