[PATCH] D33587: [DAGCombine] Do several rounds of combine.

[Amaury SECHET via Phabricator via llvm-commits]

deadalnix added a comment.

@inouehrs That wouldn't be the same as this will bail when no more combine is found.
@davide It's more like 3% as far as I can tell. The sad truth here, looking into it, is that there are a lot of combine that and undo themselves and most of the perf hit come from there. These transform are the very reason why i limited the number of iterations to begin with.

As far as benefit goes, it's very helpful for code that's legalized. I've been caring about this a lot lately, because I have workload that involve a lot of cryptography, and the gains are pretty substancial. In addition, I have various transform that I haven't published yet because they simply cannot kick in reliably without the mechanism in this diff. Any pattern that match over more than 2 level deep suffer from not having this patch.

One things that could be done is to only enable this when optimizations are on. We can then weed out the case that loop over time and enable it consistently when there are not so many of them anymore.


https://reviews.llvm.org/D33587