https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114999
--- Comment #15 from Andrew Pinski <pinskia at gcc dot gnu.org> --- (In reply to Jennifer Schmitz from comment #12) > Created attachment 60149 [details] > Proposed patch for detecting abs diff for signed integers Note the change for `-(A - B) -> B - A` with the addition of a nop_convert can introduce a signed integer overflow if the cast is from signed to unsigned. That is: `-(unsigned)(int1 = int2)` is not the same as `(unsigned)(int2 - int1)` if int1 is INT_MIN and int2 is 0. The subtraction has to be done in unsigned type to be valid transformation. The other way around is a problem too: `-(signed)(unsigned1 - unsigned2)` that is take unsigned1 == 1 and unsigned2 = (unsigned)INT_MIN. INT_MIN-1 would have an overflow; but the original (1-(unsigned)INT_MIN) is well defined (for 32bit it is -2147483647) and taking negative of that also would be well defined in signed integer.
