https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120747
--- Comment #20 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Andrew Pinski from comment #16) > Looks like the same issue as those others listed as `see also`. Basically > reassociation is tied to the ssa #s and any small improvements to other code > earlier might change the ssa versions #s and that alone causes the > difference when it comes to forming the FMAs as described in comment #14. I > am not sure what we should do here though. Note the SSA number is only used as fallback when we run into an impossible condition (which isn't impossible in case both are PHI defs. Trying to use equal in sort_by_operand_rank and using a stable sort might be worth trying?
