On Wed, Aug 02, 2017 at 12:17:44PM -0600, Jeff Law wrote: > So this has probably been hashed to death, but just a couple thoughts. > > I think realistically one has to look at the entirety of the PARALLEL to > get any reasonable costing. > > Summing the individual components is wrong. Taking the cost of any > individual component is wrong. You might be able to make a case for the > maximum cost of the individual components, but even that is almost > certain to be wrong in some cases.
I have tried all of these over the years. Nothing is good for all targets and all insns for a target, alas. > Presumably this is part of what Segher is trying to address with his > costing changes. Sure, seq_cost is trivial to express in terms of insn_cost. I'll add a patch for it. (set_rtx_cost, which is what seq_cost currently uses, is most often used for RTL that isn't (yet) in an insn. We'll have to look per case how to best handle that. Ideally rtx_cost will go away or only continue in a much slimmed down form, for targets that implement insn_cost). Segher