Hi all, So i was playing around with adding an intel-like cmov insn to a gcc 7.5 based backend. syntax is cmov %a,%b,%c,%d semantics is a = (%b ? %c : %d) Now I am aware of two predefined rtl insns for this purpose (from optabs.def): mov$acc cmov$a6 my question is, what are the differences, and in what context is one selected over the other? Now cmov sounds more intuitively what I want (conditional move), but my initial RTL .md template implementation didn't get caught. So i build x86_64 and looked how _it_ emitted cmov. This led me to look at ifcvt.c, and discovered it makes its decision to convert a conditional to a move by looking if the target has a movMMcc insn. ifcvt.c:3577 tries cmov'ing if we HAVE_conditional_move (we do), and then ifcvt.c:1775 here we call noce_emit_cmove if the args are right, then ifcvt.c:1700 we actually call emit_conditional_move, then optabs.c:4275 this line reads "icode = direct_optab_handler (movcc_optab, mode)" which was surprising. I couldn't find much uses of cmov_optab either. Maxim
