[PATCH] Do less (redudant) constant propagation during unrolling
Richard Biener
rguenther@suse.de
Tue Jul 10 13:09:00 GMT 2018
The following avoids constant propagating both on a loop and its
children.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
I'm going to remove the checking code I added before committing
(and I will of course re-test).
Richard.
2018-07-10 Richard Biener <rguenther@suse.de>
* tree-ssa-loop-ivcanon.c (tree_unroll_loops_completely_1):
Rework father_bb setting in a way to avoid propagating constants
multiple times on a loop body.
diff --git a/gcc/tree-ssa-loop-ivcanon.c b/gcc/tree-ssa-loop-ivcanon.c
index c951f17f9ba..4f6080b339f 100644
--- a/gcc/tree-ssa-loop-ivcanon.c
+++ b/gcc/tree-ssa-loop-ivcanon.c
@@ -1378,20 +1378,53 @@ tree_unroll_loops_completely_1 (bool may_increase_size, bool unroll_outer,
enum unroll_level ul;
unsigned num = number_of_loops (cfun);
+ gcc_assert (bitmap_empty_p (father_bbs));
+
/* Process inner loops first. Don't walk loops added by the recursive
calls because SSA form is not up-to-date. They can be handled in the
next iteration. */
+ bitmap child_father_bbs = NULL;
for (inner = loop->inner; inner != NULL; inner = inner->next)
if ((unsigned) inner->num < num)
- changed |= tree_unroll_loops_completely_1 (may_increase_size,
- unroll_outer, father_bbs,
- inner);
+ {
+ if (!child_father_bbs)
+ child_father_bbs = BITMAP_ALLOC (NULL);
+ if (tree_unroll_loops_completely_1 (may_increase_size, unroll_outer,
+ child_father_bbs, inner))
+ {
+ bitmap_ior_into (father_bbs, child_father_bbs);
+ bitmap_clear (child_father_bbs);
+ changed = true;
+ }
+ }
+ if (child_father_bbs)
+ BITMAP_FREE (child_father_bbs);
/* If we changed an inner loop we cannot process outer loops in this
iteration because SSA form is not up-to-date. Continue with
siblings of outer loops instead. */
if (changed)
- return true;
+ {
+ /* If we are recorded as father clear all other fathers that
+ are necessarily covered already to avoid redundant work. */
+ if (bitmap_bit_p (father_bbs, loop->header->index))
+ {
+ if (flag_checking)
+ {
+ bitmap_iterator bi;
+ unsigned i;
+ EXECUTE_IF_SET_IN_BITMAP (father_bbs, 0, i, bi)
+ {
+ loop_p floop = BASIC_BLOCK_FOR_FN (cfun, i)->loop_father;
+ gcc_assert (floop == loop
+ || flow_loop_nested_p (loop, floop));
+ }
+ }
+ bitmap_clear (father_bbs);
+ bitmap_set_bit (father_bbs, loop->header->index);
+ }
+ return true;
+ }
/* Don't unroll #pragma omp simd loops until the vectorizer
attempts to vectorize those. */
@@ -1421,7 +1454,23 @@ tree_unroll_loops_completely_1 (bool may_increase_size, bool unroll_outer,
computations; otherwise, the size might blow up before the
iteration is complete and the IR eventually cleaned up. */
if (loop_outer (loop_father))
- bitmap_set_bit (father_bbs, loop_father->header->index);
+ {
+ /* Once we process our father we will have processed
+ the fathers of our children as well, so avoid doing
+ redundant work and clear fathers we've gathered sofar. */
+ if (flag_checking)
+ {
+ bitmap_iterator bi;
+ unsigned i;
+ EXECUTE_IF_SET_IN_BITMAP (father_bbs, 0, i, bi)
+ {
+ loop_p floop = BASIC_BLOCK_FOR_FN (cfun, i)->loop_father;
+ gcc_assert (flow_loop_nested_p (loop_father, floop));
+ }
+ }
+ bitmap_clear (father_bbs);
+ bitmap_set_bit (father_bbs, loop_father->header->index);
+ }
return true;
}
More information about the Gcc-patches
mailing list