------- Comment #22 from rguenth at gcc dot gnu dot org 2010-04-02 12:33 ------- Or a combination of both. In particular the code doesn't account for the negative offsets we offset the spill-slot decl with when trying to handle a NULL MEM_OFFSET conservatively.
I wonder what sizex / sizey is for the spill-slot-decl, or rather what mode or size it has ... Least pessimizing patch follows, it looks like nobody else would disambiguate spill slots otherwise (apart from the tree oracle calls). Index: gcc/alias.c =================================================================== --- gcc/alias.c (revision 157942) +++ gcc/alias.c (working copy) @@ -2147,6 +2147,13 @@ nonoverlapping_memrefs_p (const_rtx x, c if (exprx == 0 || expry == 0) return 0; + /* For spill-slot accesses make sure we have valid offsets. */ + if ((exprx == get_spill_slot_decl (false) + && ! MEM_OFFSET (x)) + || (expry == get_spill_slot_decl (false) + && ! MEM_OFFSET (y))) + return 0; + /* If both are field references, we may be able to determine something. */ if (TREE_CODE (exprx) == COMPONENT_REF && TREE_CODE (expry) == COMPONENT_REF -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42509