[PATCH] D38128: Handle COPYs of physregs better (regalloc hints)

[Martin Storsjö via Phabricator via llvm-commits]

mstorsjo added a subscriber: aemerson.
mstorsjo added a comment.

Yes, this looks like mostly no-op reorderings. I don't remember exactly what part I thought was a potential performance regression in swifterror.ll before - it moves one "mov" instruction (in two places) to before a branch, so potentially executing one instruction more than before. I would say it's most probably ok (and the gains in some of the other register shuffling tests would make it sound like a net gain in any case).

I'm in no way authoritative for this though, so perhaps @aemerson or @qcolombet would like to give such a stamp of approval?


https://reviews.llvm.org/D38128