https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126328
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Segher Boessenkool from comment #5) > (In reply to Richard Biener from comment #3) > > so we are back to a cost of 4, because combine_validate_cost doesn't > > re-recog before costing :/ > > What does this mean? The result of recog () should depend only on the RTL of > the insn, always. So why would you ever want to run recog () more often than > required? Because you want to write the insn matching code once. Like you could add a 'cost' insn attribute to patterns in the .md file and in the insn_cost hook simply return that. But - combine calls insn_cost when the insn isn't recognized. Now, I suppose the hook isn't supposed to modify the insn (even though it doesnt' get a const_rtx_insn *), or do you say the insn_cost hook should simply call recog () to match up a pattern with costs? See https://gcc.gnu.org/pipermail/gcc-patches/2026-July/724831.html how I propose to "fix" the x86 insn_cost hook to recognize the hadd patterns. Relying on recog () would be way easier and more maintainable?
