In PR53346 we vectorize a simple memset loop very inefficiently.
But of course we should have detected this and transformed the
loop into a memset! Seems like we only do that if the original
loop does sth else than memset as well.
Fixed as follows.
Bootstrap and regtest on x86_64-unknown-linux-gnu ongoing
(I suppose that will really stress loop distribution now ;))
Richard.
2012-05-18 Richard Guenther <rguent...@suse.de>
PR tree-optimization/53346
* tree-loop-distribution.c (ldist_gen): Make sure to apply
builtin transform even when only a single partition with
all reads/writes exists.
* gcc.dg/tree-ssa/ldist-18.c: New testcase.
Index: gcc/tree-loop-distribution.c
===================================================================
*** gcc/tree-loop-distribution.c (revision 187650)
--- gcc/tree-loop-distribution.c (working copy)
*************** ldist_gen (struct loop *loop, struct gra
*** 1131,1138 ****
BITMAP_FREE (processed);
nbp = VEC_length (bitmap, partitions);
! if (nbp <= 1
! || partition_contains_all_rw (rdg, partitions))
goto ldist_done;
if (dump_file && (dump_flags & TDF_DETAILS))
--- 1131,1141 ----
BITMAP_FREE (processed);
nbp = VEC_length (bitmap, partitions);
! if (nbp == 0
! || (nbp == 1
! && !can_generate_builtin (rdg, VEC_index (bitmap, partitions, 0)))
! || (nbp > 1
! && partition_contains_all_rw (rdg, partitions)))
goto ldist_done;
if (dump_file && (dump_flags & TDF_DETAILS))
Index: gcc/testsuite/gcc.dg/tree-ssa/ldist-18.c
===================================================================
*** gcc/testsuite/gcc.dg/tree-ssa/ldist-18.c (revision 0)
--- gcc/testsuite/gcc.dg/tree-ssa/ldist-18.c (revision 0)
***************
*** 0 ****
--- 1,12 ----
+ /* { dg-do compile } */
+ /* { dg-options "-O2 -ftree-loop-distribute-patterns
-fdump-tree-ldist-details" } */
+
+ void foo (int *p, int n)
+ {
+ int i;
+ for (i = 0; i < n; ++i)
+ p[i] = 0;
+ }
+
+ /* { dg-final { scan-tree-dump "generated memset zero" "ldist" } } */
+ /* { dg-final { cleanup-tree-dump "ldist" } } */
