On Fri, 2026-06-05 at 14:40 +0200, Georg-Johann Lay via Gcc wrote: > Am 05.06.26 um 13:55 schrieb Richard Biener: > > > > In general I agree with Andrew that match.pd, for this kind of > > "optimizations", > > needs to apply costing. > > As far as I understand, that approach works at tree -> rtl lowering when > rtx costs are available. IIUC there's nothing like costs at match.pd time. >
I'd be great if we didn't have to implement another cost function in the backend for that ;) Best regards, Oleg Endo
