tree-optimization/95761 - fix vector insertion place compute

I missed that indeed SLP permutation code generation can end up
refering to a non-last vectorized stmt in the last SLP_TREE_VEC_STMTS
element as optimization.  So walk them all.

2020-06-19  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/95761
	* tree-vect-slp.c (vect_schedule_slp_instance): Walk all
	vectorized stmts for finding the last one.

	* gcc.dg/torture/pr95761.c: New testcase.
This commit is contained in:
Richard Biener 2020-06-19 10:03:46 +02:00
parent bafd12cb22
commit f8f5715606
2 changed files with 34 additions and 6 deletions

View File

@ -0,0 +1,25 @@
/* { dg-do compile } */
typedef int a[10];
typedef struct {
a b;
a c;
a d;
a e;
} f;
f g;
int *j;
void k() {
for (;;) {
a l;
j[0] = g.b[0];
int *h = g.d;
int i = 0;
for (; i < 10; i++)
h[i] = l[0] - g.e[0];
h = g.e;
i = 0;
for (; i < 10; i++)
h[i] = l[1] + g.e[i];
}
}

View File

@ -4199,7 +4199,6 @@ vect_schedule_slp_instance (vec_info *vinfo,
{
/* Or if we do not have 1:1 matching scalar stmts emit after the
children vectorized defs. */
gimple *last_in_child;
gimple *last_stmt = NULL;
FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
/* ??? With only external defs the following breaks. Note
@ -4208,11 +4207,15 @@ vect_schedule_slp_instance (vec_info *vinfo,
if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
{
/* We are emitting all vectorized stmts in the same place and
the last one is the last. */
last_in_child = SLP_TREE_VEC_STMTS (child).last ();
if (!last_stmt
|| vect_stmt_dominates_stmt_p (last_stmt, last_in_child))
last_stmt = last_in_child;
the last one is the last.
??? Unless we have a load permutation applied and that
figures to re-use an earlier generated load. */
unsigned j;
gimple *vstmt;
FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (child), j, vstmt)
if (!last_stmt
|| vect_stmt_dominates_stmt_p (last_stmt, vstmt))
last_stmt = vstmt;
}
if (is_a <gphi *> (last_stmt))
si = gsi_after_labels (gimple_bb (last_stmt));