https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56049
--- Comment #23 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we are _nearly_ there on trunk. LIM has improved up to the point the only
blocker is mismatching alias-sets in mem_ref_hasher::equal and the case in
question for this PR is even handled correctly because refs arrive in the
correct order and the one we'd chose has the ref-all case we'd have to
canonicalize to anyways. So the following works (and can be slightly
enhanced).
Index: gcc/tree-ssa-loop-im.c
===================================================================
--- gcc/tree-ssa-loop-im.c (revision 270366)
+++ gcc/tree-ssa-loop-im.c (working copy)
@@ -178,7 +178,15 @@ mem_ref_hasher::equal (const im_mem_ref
&& known_eq (mem1->mem.size, obj2->size)
&& known_eq (mem1->mem.max_size, obj2->max_size)
&& mem1->mem.volatile_p == obj2->volatile_p
- && mem1->mem.ref_alias_set == obj2->ref_alias_set
+ && (mem1->mem.ref_alias_set == obj2->ref_alias_set
+ /* We are not canonicalizing alias-sets but for the
+ special-case we didn't canonicalize yet and the
+ incoming ref is a alias-set zero MEM we pick
+ the correct one already. */
+ || (!mem1->ref_canonical
+ && (TREE_CODE (obj2->ref) == MEM_REF
+ || TREE_CODE (obj2->ref) == TARGET_MEM_REF)
+ && obj2->ref_alias_set == 0))
&& types_compatible_p (TREE_TYPE (mem1->mem.ref),
TREE_TYPE (obj2->ref)));
else
