Richard Biener <rguent...@suse.de> writes: >> > So, may I go with the original patch? >> >> Still feels like we're counting spills on targets that shouldn't need them. >> But going back to: >> >> /* Assume that a reg-reg move is possible and cheap, >> do not account for vector to gp register move cost. */ >> >> I guess we could gloss over the "unnecessary" spill cost by saying that, >> even if the spill isn't needed, this is a conservative estimate of the >> vector to gp register move cost? > > Yes. What I really tried to ask is - am I going to need the > vectorized result piecewise in the end (but not being in control > of the chopping into pieces)? I wanted to pessimize that with > an estimate of the "chopping cost". I probably shouldn't have > talked about spilling but that's the usual straight-forward > solution of extracting sth from a larger register that works > everywhere. > > So I guess I'll update the comment and install as-is?
Sounds good to me. Richard > > I still hope for a better solution either on the target or the > RTL optimization side (undoing the vectorization). > > Richard.
