https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107855
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> --- Can you attach the ifcvt dump as well? The issue seems to be that only the first loop is if-converted, so floats are not handled at all (only a single loop is vectorized). Not sure about the runtime FAIL with 32bit - does that also happen with -O0?