[Bug rtl-optimization/49390] [4.6/4.7 Regression] GCSE miscompilation

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49390

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-06-13 13:53:23 UTC ---
Perhaps we should have some exceptions where we allow different MEM_ATTRS, but
they need to be carefully chosen.  E.g. if both refs are indirect refs and are
similar, with the same points-to info, it would be ok.

  if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
    return 0;

  if (MEM_ATTRS (x) != MEM_ATTRS (y) && mem_attrs_equiv_p (x, y))
    return 0;

where mem_attrs_equiv_p would call ao_ref_from_mem on both x and y,
if at least one of them returns false, fail, compare all integer fields for
equality (alias sets using ao_ref_base_alias_set/ao_ref_alias_set) and compare
base kinds, for indirect we could check for same type and same points-to info
(couldn't find a points-to set comparison function, e.g. for pt->vars can just
pointer equality be tested or does it need bitmap_equal_p?).