On Mon, 20 Feb 2012, Jakub Jelinek wrote:
> On Mon, Feb 20, 2012 at 04:11:13PM +0100, Richard Guenther wrote:
> > This fixes PR52298, we need to use the proper DR step for outer
> > loop vectorization.
> >
> > Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
>
> Thanks.
>
> > *************** vectorizable_load (gimple stmt, gimple_s
> > *** 4266,4272 ****
> > if (!STMT_VINFO_DATA_REF (stmt_info))
> > return false;
> >
> > ! negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0;
> > if (negative && ncopies > 1)
> > {
> > if (vect_print_dump_info (REPORT_DETAILS))
> > --- 4268,4277 ----
> > if (!STMT_VINFO_DATA_REF (stmt_info))
> > return false;
> >
> > ! negative = tree_int_cst_compare (loop && nested_in_vect_loop_p (loop,
> > stmt)
> > ! ? STMT_VINFO_DR_STEP (stmt_info)
> > ! : DR_STEP (dr),
> > ! size_zero_node) < 0;
>
> vectorizable_loop, unlike vectorizable_store, has nested_in_vect_loop
> bool flag, so you could just test that instead of
> loop && nested_in_vect_loop_p (loop, stmt) above and below.
Fixed as follows.
Richard.
2012-02-21 Richard Guenther <rguent...@suse.de>
* tree-vect-stmts.c (vectorizable_load): Use pre-computed
nested_in_vect_loop.
Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c (revision 184396)
+++ gcc/tree-vect-stmts.c (working copy)
@@ -4268,7 +4268,7 @@ vectorizable_load (gimple stmt, gimple_s
if (!STMT_VINFO_DATA_REF (stmt_info))
return false;
- negative = tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt)
+ negative = tree_int_cst_compare (nested_in_vect_loop
? STMT_VINFO_DR_STEP (stmt_info)
: DR_STEP (dr),
size_zero_node) < 0;
@@ -4658,7 +4658,7 @@ vectorizable_load (gimple stmt, gimple_s
This can only occur when vectorizing memory accesses in the inner-loop
nested within an outer-loop that is being vectorized. */
- if (loop && nested_in_vect_loop_p (loop, stmt)
+ if (nested_in_vect_loop
&& (TREE_INT_CST_LOW (STMT_VINFO_DR_STEP (stmt_info))
% GET_MODE_SIZE (TYPE_MODE (vectype)) != 0))
{
