diff options
| author | Richard Biener <rguenther@suse.de> | 2025-02-12 14:18:06 +0100 |
|---|---|---|
| committer | Richard Biener <rguenth@gcc.gnu.org> | 2025-02-14 08:28:50 +0100 |
| commit | 27653070db35216d5115cc25672fcc6a51203d26 (patch) | |
| tree | 902b096ac125b500960271a4536e1705f91955f5 /gcc/tree-vect-loop.cc | |
| parent | 8caf67eea7e1b29a4437f07d13c300d9fdb04827 (diff) | |
| download | gcc-27653070db35216d5115cc25672fcc6a51203d26.zip gcc-27653070db35216d5115cc25672fcc6a51203d26.tar.gz gcc-27653070db35216d5115cc25672fcc6a51203d26.tar.bz2 | |
tree-optimization/90579 - avoid STLF fail by better optimizing
For the testcase in question which uses a fold-left vectorized
reduction of a reverse iterating loop we'd need two forwprop
invocations to first bypass the permute emitted for the reverse
iterating loop and then to decompose the vector load that only
feeds element extracts. The following moves the first transform
to a match.pd pattern and makes sure we fold the element extracts
when the vectorizer emits them so the single forwprop pass can
then pick up the vector load decomposition, avoiding the forwarding
fail that causes.
Moving simplify_bitfield_ref also makes forwprop remove the dead
VEC_PERM_EXPR via the simple-dce it uses - this was also
previously missing.
PR tree-optimization/90579
* tree-ssa-forwprop.cc (simplify_bitfield_ref): Move to
match.pd.
(pass_forwprop::execute): Adjust.
* match.pd (bit_field_ref (vec_perm ...)): New pattern
modeled after simplify_bitfield_ref.
* tree-vect-loop.cc (vect_expand_fold_left): Fold the
element extract stmt, combining it with the vector def.
* gcc.target/i386/pr90579.c: New testcase.
Diffstat (limited to 'gcc/tree-vect-loop.cc')
| -rw-r--r-- | gcc/tree-vect-loop.cc | 5 |
1 files changed, 5 insertions, 0 deletions
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc index eea0b89..07b19a2 100644 --- a/gcc/tree-vect-loop.cc +++ b/gcc/tree-vect-loop.cc @@ -7086,6 +7086,11 @@ vect_expand_fold_left (gimple_stmt_iterator *gsi, tree scalar_dest, rhs = make_ssa_name (scalar_dest, stmt); gimple_assign_set_lhs (stmt, rhs); gsi_insert_before (gsi, stmt, GSI_SAME_STMT); + /* Fold the vector extract, combining it with a previous reversal + like seen in PR90579. */ + auto gsi2 = gsi_for_stmt (stmt); + if (fold_stmt (&gsi2, follow_all_ssa_edges)) + update_stmt (gsi_stmt (gsi2)); stmt = gimple_build_assign (scalar_dest, code, lhs, rhs); tree new_name = make_ssa_name (scalar_dest, stmt); |