On 11/20/20 12:41 PM, Martin Jambor wrote: > Hi, > > this is an updated patch based on our conversation on IRC today. So far > I have had a look at the effects on only tramp3d and although it makes > the heuristics more pessimistic more times than optimistic (number of > clones at -Ofast drops from 559 to 557), there are also lattices which > are massively boosted. I don't think that's inherently a bad thing. > > When looking at the testcase of PR 97816 I realized that the reason > why we were hitting overflows in size growth estimates in IPA-CP is > not because the chains of how lattices feed values to each other are > so long but mainly because we add estimates in callee lattices to > caller lattices for each value source, which roughly corresponds to a > call graph edge, and therefore if there are multiple calls between two > functions passing the same value in a parameter we end up doing it > more than once, sometimes actually quite many times. > > This patch avoids it by using a has_set to remember the source values > we have already updated and not increasing their size again. > Furhtermore, to improve estimation of times we scale the propagated > time benefits with edge frequencies as we accumulate them. > > This should make any overflows very unlikely but not impossible, so I > still included checks for overflows but decided to restructure the > code to only need it in the propagate_effects function and modified it > so that it does not need to perform the check before each sum. > > This is because I decided to add local estimates to propagated > estimates already in propagate_effects and not at the evaluation time. > The function can then do the sums in a wide type and discard them in > the unlikely case of an overflow. I also decided to use the > opportunity to make propagated effect stats now include stats from > other values in the same SCCs. In the dumps I have seen this tended > to increase size cost a tiny bit more than the estimated time benefit > but both increases were small. > > Bootstrapped and LTO bootstrapped on x86_64-linux. OK for trunk? > > Thanks, > > Martin > > > gcc/ChangeLog: > > 2020-11-20 Martin Jambor <mjam...@suse.cz> > > PR ipa/97816 > * ipa-cp.c (safe_add): Removed. > (good_cloning_opportunity_p): Remove special handling of INT_MAX. > (value_topo_info<valtype>::propagate_effects): Take care not to > propagate from size one value to another through more sources. Scale > propagated times with edge frequencies. Include local time and size > in propagates ones here. Take care not to overflow size. > (decide_about_value): Do not add local and propagated effects when > passing them to good_cloning_opportunity_p. OK Jeff
