https://gcc.gnu.org/bugzilla/show_bug.cgi?id=19792
--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> --- I think for foo the "solution" is to only consider indices with well-defined behavior, that is, for size_lookup[] indices in [0,257] which means we should be safely able to truncate the index to unsigned char (for the purpose of analysis only!). Note we're not value-numbering (int) t the same as t but even size_lookup[(short) t] should result in the same value to be loaded unless undefined behavior is invoked. Of course extern unsigned char size_lookup[257]; might be considered extern unsigned char size_lookup[]; for QOI reasons and existing broken code, so... but just change the testcase to non-extern size_lookup. The solution might ly in value-numbering which could, when value-numbering the two loads lookup (and insert!) VNs of (shortest-allowed-type)index. The danger is of course that the shortest-allowed-type might differ depentent on the shape of the lookup where otherwise we'd compute the same VN. I'm quite sure we can't simply trust TYPE_DOMAIN of TREE_TYPE of TREE_OPERAND (array-ref, 0) when constraining operand 1.
