https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94873
--- Comment #17 from Eric Botcazou <ebotcazou at gcc dot gnu.org> --- > Ok, so what we do about this bug then if it ought to be combine.c that needs > changing? For REG_EQUAL notes in combine_instructions check for the > auto-incdec side-effects in the pattern (I'd hope we don't have them in > REG_EQUAL notes content) and if there are any, punt? At least for > backporting that seems like the right solution, and given that it seems > try_combine also punts on these if we remove or add any in the patterns, it > seems in line with what the rest of combiner does. At least the fix for the branches seems to be clear: do not do the temporary replacement in combine_instructions if the SET_SRC has side effects; there are a bunch of similar side_effects_p tests in the combiner. > Or, shall it e.g. queue the side-effects in some new argument to try_combine > and if the combination would succeed, add the side-effect as yet another > instruction (if it can match and is cheaper than what we have previously)? That seems overkill to me, even on the mainline, but others might disagree.
