On Mon, 21 Jul 2025, Robin Dapp wrote: > > There is currently no way to mimic this, the original idea would have been > > that you record the per stmt info during add_stmt cost hook time and then > > process that data at finish_cost time. > > > > With SLP you could in theory walk the SLP graph via the instances vector of > > the vinfo. But I’m not sure what exactly you compute at the moment and how > > that even makes sense with SLP in mind. You can of course compute register > > pressure of the original scalar code, but that doesn’t seem like what you > > are doing. > > Although the heuristic would need more rework to be "SLP first" I think we can > get away with a hash_map that stores the necessary information, at least for > the time being. Currently testing a local workaround.
Thanks a lot. I would expect the heuristic to get SLP loads/stores wrong given when you do stmt-based heuristics you count insns multiple times. Richard.
