[Bug rtl-optimization/109585] [10/11/12/13/14 regression] Carla/sord miscompiled with -O2 on ARM64 with flexible array member

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109585

--- Comment #21 from Richard Biener <rguenth at gcc dot gnu.org> ---
aliasing_component_refs_p seems to try handle trailing arrays via
component_ref_to_zero_sized_trailing_array_p.  It properly detects

f_5->fam[0].n

but fails on

MEM[(struct P *)f_3(D) + 8B].n

but it would also in all cases happily continue with the base ref type
when comparing sizes:

  /* Now search for the type1 in the access path of ref2.  This
     would be a common base for doing offset based disambiguation on.
     This however only makes sense if type2 is big enough to hold type1.  */
  int cmp_outer = compare_type_sizes (type2, type1);

I also wonder if

  /* If we didn't find a common base, try the other way around.  */
  if (cmp_outer <= 0
      || (end_struct_ref1
          && compare_type_sizes (TREE_TYPE (end_struct_ref1), type1) <= 0))

here type1 shouldn't be type2?

In any case we end up failing the access_path_may_continue_p checks and
disambiguate.

diff --git a/gcc/tree-ssa-alias.cc b/gcc/tree-ssa-alias.cc
index 81bc51ed4ad..8a1ec9091fa 100644
--- a/gcc/tree-ssa-alias.cc
+++ b/gcc/tree-ssa-alias.cc
@@ -1330,7 +1330,7 @@ aliasing_component_refs_p (tree ref1,
   /* If we didn't find a common base, try the other way around.  */
   if (cmp_outer <= 0 
       || (end_struct_ref1
-         && compare_type_sizes (TREE_TYPE (end_struct_ref1), type1) <= 0))
+         && compare_type_sizes (TREE_TYPE (end_struct_ref1), type2) <= 0))
     {
       int res = aliasing_component_refs_walk (ref2, type2, base2,
                                              offset2, max_size2,

seems to fix the testcase - can anybody confirm?

Honza, is that indeed a typo or do I just hide the issue?