[gcc(refs/vendors/vrull/heads/slp-improvements)] match.pd: Look through view_convert when folding vec_perms
Philipp Tomsich
ptomsich@gcc.gnu.org
Tue Jan 23 20:57:16 GMT 2024
https://gcc.gnu.org/g:b896c82f141211437b5b5b83a3537d6bd2d243dc
commit b896c82f141211437b5b5b83a3537d6bd2d243dc
Author: Manolis Tsamis <manolis.tsamis@vrull.eu>
Date: Wed Nov 1 12:27:28 2023 +0100
match.pd: Look through view_convert when folding vec_perms
The match.pd patterns to merge two vector permutes into one fail to
match when a (potentially no-op) view convert expressions is
separating the two permutes. This is observable in the SLP tree for
the first loop of x264's satd and inhibits optimization.
This change adds a check whether a view_convert will be a no-op
(comparing the element precision) and rewrites the permute sequence
for those cases.
Ref #343
Diff:
---
gcc/match.pd | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/gcc/match.pd b/gcc/match.pd
index e42ecaf9ec7..be623d9773c 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -9965,7 +9965,7 @@ and,
d = VEC_PERM_EXPR <a, b, NEW_VCST>; */
(simplify
- (vec_perm (vec_perm@0 @1 @2 VECTOR_CST@3) @0 VECTOR_CST@4)
+ (vec_perm (view_convert?@0 (vec_perm @1 @2 VECTOR_CST@3)) @0 VECTOR_CST@4)
(if (TYPE_VECTOR_SUBPARTS (type).is_constant ())
(with
{
@@ -9977,7 +9977,8 @@ and,
vec_perm_builder builder2 (nelts, nelts, 1);
}
(if (tree_to_vec_perm_builder (&builder0, @3)
- && tree_to_vec_perm_builder (&builder1, @4))
+ && tree_to_vec_perm_builder (&builder1, @4)
+ && element_precision (type) == element_precision (TREE_TYPE (@0)))
(with
{
vec_perm_indices sel0 (builder0, 2, nelts);
@@ -10002,7 +10003,7 @@ and,
op0 = vec_perm_indices_to_tree (TREE_TYPE (@4), sel2);
}
(if (op0)
- (vec_perm @1 @2 { op0; })))))))
+ (view_convert (vec_perm @1 @2 { op0; }))))))))
/* Merge
c = VEC_PERM_EXPR <a, b, VCST0>;
More information about the Gcc-cvs
mailing list