[PATCH] D59129: [SROA] Don't lower expensive allocas with minsize

[Sjoerd Meijer via Phabricator via llvm-commits]

SjoerdMeijer added a comment.

> My concern isn't really the size of the change, it's with the variance. With only four benchmarks which change, and only by a small amount each, then if you happened to run one more or one fewer benchmark you could come to a completely different conclusion.

Okay, understood. But I have tested 3 I would say quite large code bases.  2 of them don't show any changes, which I interpret as good news (because there's definitely memcpys in these codebases). The other, CSiBe, shows this little up and down behaviour, but improvements overall. Before I continue fine tuning, how can I make this more convincing for you? Do I need more code bases, or just getting rid of the regressions?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D59129/new/

https://reviews.llvm.org/D59129