https://gcc.gnu.org/g:95f3686d4788d7420422d514996c5b6e7a8facfd
commit 95f3686d4788d7420422d514996c5b6e7a8facfd
Author: Richard Biener <rguent...@suse.de>
Date: Thu Oct 5 14:07:02 2023 +0200
Avoid SLP build failure for unsupported shifts
When asserting that SLP _build_ can succeed we run into the SLP
shift discovery code trying to be clever doing vectorizable_shifts
work and failing discovery already. That gives a false impression
for now, so disable that when we do single-lane builds.
* tree-vect-slp.cc (vect_build_slp_tree_1): Do not fail
fatally for shifts not supported by the target when discovering
a single lane.
Diff:
---
gcc/tree-vect-slp.cc | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 6bfc59dc2131..ecc185aae885 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -1215,7 +1215,13 @@ vect_build_slp_tree_1 (vec_info *vinfo, unsigned char
*swap,
if (!directly_supported_p (rhs_code, vectype, optab_vector))
{
/* No vector/vector shift, try for a vector/scalar shift. */
- if (!directly_supported_p (rhs_code, vectype, optab_scalar))
+ if (!directly_supported_p (rhs_code, vectype, optab_scalar)
+ /* ??? We are using this to guide operand swapping to
+ eventually make all shift operands the same but we
+ shouldn't fail in the end - that's be business of
+ vectorizable_shift.
+ Avoid spurious ICEs for single-lane discovery. */
+ && group_size != 1)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
