https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96963
Jakub Jelinek <jakub at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |jakub at gcc dot gnu.org --- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> --- The PR93200 "fix" was incorrect, as you should know in GIMPLE most pointer conversions are useless, so the pointer type of the MEM_REF's first operand is completely arbitrary. If you wanted to exclude vector types from the overflow warnings, you should have punted for MEM_REFs with vector type being TREE_TYPE (memref). It is still just a hack, because the whole relying on &something->field vs. something in addresses stands on sand, in GIMPLE those are all just about address values and nothing else, they don't carry semantic information.