On 8/9/19 2:13 PM, Michael Matz wrote: > Hi, > > On Fri, 9 Aug 2019, Richard Biener wrote: > >> Of course I'm still afraid that the other code exists for a reason >> (tuning/hack/whatever...). >> >> Note that with the patch we're now applying LOOP_ALIGN to L2 here: >> if (a) >> foo = bar; >> L2: >> blah; >> >> because there's a jump-around and a fallthru. > > Yeah, and I think that would be wrong. That's why the existing code (not > sure about after the patch) does this only when L2 is reached by one edge > much more often than by the other edges. > >> So I'm not sure we don't need to apply some condition on fallthru_count >> (which is unused after your patch btw). > > > Ciao, > Michael. >
I'm sending numbers for the opposite condition. > Of course I'm still afraid that the other code exists for a reason > (tuning/hack/whatever...). I fully agree that the current code is quite hacking and was probably subject of some tuning. I'm leaving the decision about simplification to you? You're much more experienced in the area :) Martin
lnt-loop-alignment-v2.pdf.bz2
Description: application/bzip
