Hans-Peter Nilsson via Gcc-patches <gcc-patches@gcc.gnu.org> writes: > Most comments, including the second sentence in the head comment > of combine_validate_cost, the main decision-maker of the combine > pass, refer to the function as returning true if the new > insns(s) *cheaper* than the old insns, when in fact the function > returned true also if the cost was the same. Returning true for > cheaper also seems more sane than as-cheap-as considering the > need to avoid oscillation between same-cost combinations. Also, > it makes the job of later passes harder, having combine make > more complex combinations of the same cost. > > Right, you can affect this with your target TARGET_RTX_COSTS and > TARGET_INSN_COST hooks, but only for trivial cases, and you have > increasingly more complex combinations (many-to-many > combinations) where you have to twist simple logic to appease > combine (stop it from combining) or give up. > > Main-interest ports are unsurprisingly pretty tied to this > effect. I'd love to install the following patch, adjusting the > function and the two opposing comments. But...it causes > hundreds of regressions for each of x86_64-linux and > aarch64-linux (tens for ppc64le-linux), so I'm just not up to > the task, at least not without previous buy-in from reviewers.
Out of interest, how do the results change if we still allow the combination for equal costs if the new sequence is shorter than the original? I think that still counts as “cheaper than”, but maybe I'm too RISC-centric. ;-) (I'm not saying that's what we should do, I'm just curious.) Originally combine always produced shorter sequences, so by the above reasoning, combining for == was correct. These days we allow N-to-N replacements too, which are obviously a good thing when the new cost is lower, but are more of a wash when the costs are the same. But even then, the combination should have a “canonicalisation” effect. (Unfortunately that effect includes the result of expand_compound_operation/make_compound_operation.) Do you have specific examples of where things go wrong? Thanks, Richard
