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

[Davide Italiano via Phabricator via llvm-commits]

davide added a comment.

If I read these numbers correctly, this makes Instruction selection ~ 4-5% slower on large testcases (in your case, an LTO build of clang).
This is, quite a bit, and I need requires further justifications (i.e. needs to be backed by the performance improvement we get on these testcases for the additional compile time we pay.


https://reviews.llvm.org/D33587