https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108764
Jeffrey A. Law <law at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |law at gcc dot gnu.org --- Comment #6 from Jeffrey A. Law <law at gcc dot gnu.org> --- For Veyron V1, those shNadds are single cycle and can issue to any of the 4 ALUs which makes the first sequence better for that uarch. As folks have noted, costing is highly dependent on the uarch. So in cases like this the right way to go is describe the costing in the uarch specific structure and query that much like we do for other operations.