Alias analysis was treating .MASK_STORE as storing a full vector
which means we disambiguate against decls of smaller than vector size.
That's of course wrong and a similar issue was fixed for DSE already.
The following makes sure we set the size of the access to unknown
and only constrain max_size.
This fixes runtime execution FAILs of gfortran.dg/matmul_2.f90,
gfortran.dg/matmul_6.f90 and gfortran.dg/pr91577.f90 when using
AVX512 with full masked loop vectorization on Zen4.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed to
trunk sofar.
* tree-ssa-alias.cc (call_may_clobber_ref_p_1): For
.MASK_STORE and friend set the size of the access to
unknown.
---
gcc/tree-ssa-alias.cc | 3 +++
1 file changed, 3 insertions(+)
diff --git a/gcc/tree-ssa-alias.cc b/gcc/tree-ssa-alias.cc
index 79ed956e300..b5476e8b41e 100644
--- a/gcc/tree-ssa-alias.cc
+++ b/gcc/tree-ssa-alias.cc
@@ -3072,6 +3072,9 @@ call_may_clobber_ref_p_1 (gcall *call, ao_ref *ref, bool
tbaa_p)
ao_ref lhs_ref;
ao_ref_init_from_ptr_and_size (&lhs_ref, gimple_call_arg (call, 0),
TYPE_SIZE_UNIT (TREE_TYPE (rhs)));
+ /* We cannot make this a known-size access since otherwise
+ we disambiguate against stores to decls that are smaller. */
+ lhs_ref.size = -1;
lhs_ref.ref_alias_set = lhs_ref.base_alias_set
= tbaa_p ? get_deref_alias_set
(TREE_TYPE (gimple_call_arg (call, 1))) : 0;
--
2.35.3
