https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117601
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Andrew Pinski from comment #2) > looking into this slightly more, there needs to have a check on > INTEGER_CST@1 if less than the size of the array if o is an array. > (otherwise only allow 1). Yeah, we don't necessarily want to bring back "optimistic" back-conversion to pointer arithmetic from integer one. I'd instead have approached this one from re-association as well where for both signed integer and pointer add/sub chains we can perform elimination if we then rewrite the result to a no-overflow variant. Your pattern would only solve the issue for address of decl bases. You also have to be careful to make sure the decl binds locally given a DSO override might cause the arithmetic to spill over to another decl and integer arithmetic for that would have been valid!
