------- Comment #1 from irar at il dot ibm dot com 2009-04-08 11:17 -------
A testcase for 4.5:
#define N 128
int out[N*4], out2[N], in[N*4];
void
foo ()
{
int i, a0, a1, a2, a3;
for (i = 0; i < N; i++)
{
a0 = in[i*4];
a1 = in[i*4 + 1];
a2 = in[i*4 + 2];
a3 = in[i*4 + 3];
out[i*4] = a0;
out[i*4 + 1] = a1;
out[i*4 + 2] = a2;
out[i*4 + 3] = a3;
out2[i] = a0;
}
}
Fails with
internal compiler error: in vect_get_vec_def_for_operand, at
tree-vect-stmts.c:940
(on powerpc64-suse-linux with -O3 -maltivec)
The bug is in decision whether a statement must be "loop-vectorized" in a
addition to SLP. Here is a patch (for 4.4) that I am testing:
Index: tree-vect-transform.c
===================================================================
--- tree-vect-transform.c (revision 145711)
+++ tree-vect-transform.c (working copy)
@@ -8459,20 +8459,11 @@
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "=== scheduling SLP instances ===");
- is_store = vect_schedule_slp (loop_vinfo);
-
- /* IS_STORE is true if STMT is a store. Stores cannot be of
- hybrid SLP type. They are removed in
- vect_schedule_slp_instance and their vinfo is destroyed.
*/
- if (is_store)
- {
- gsi_next (&si);
- continue;
- }
- }
+ vect_schedule_slp (loop_vinfo);
+ }
/* Hybrid SLP stmts must be vectorized in addition to SLP. */
- if (PURE_SLP_STMT (stmt_info))
+ if (!vinfo_for_stmt (stmt) || PURE_SLP_STMT (stmt_info))
{
gsi_next (&si);
continue;
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39675
