https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65220
Jakub Jelinek <jakub at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |uros at gcc dot gnu.org --- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Ah, I meant UINTVAL (operands[3]) - 2 < <MODE_SIZE> * BITS_PER_UNIT instead of UINTVAL (operands[3]) + 2 < <MODE_SIZE> * BITS_PER_UNIT so that it rules out both 0 and 1 (for the latter, not sure how well would & 0 work). For splitting post-reload, the "0" constraint should make the other change unnecessary, but if we want to split also before reload, something like what you've done is needed. Though I wonder if there couldn't be e.g. the case that before reload operand 1 is the same as operand 2 (one pseudo) and operand 0 is different. Perhaps what exact operand should be used should depend on something like reg_overlap_mentioned_p and depending on that set operands[5] to either operands[2] or operands[1] for the second insn source? Or just add reload_completed condition to the splitter.