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

[Jonas Paulsson via Phabricator via llvm-commits]

jonpa added a comment.

In https://reviews.llvm.org/D38128#908337, @efriedma wrote:

> What specifically do you want me to comment on?  I don't really understand the register allocator well enough to comment on your code changes.


Given that this is something that Quentin has been meaning to fix before I did, and that this is *generally* good, the question is now if we can accept minor issues like this. On SystemZ, I think this is quite clear since we get 12500 less COPYs and also less spilling (on SPEC). Could you perhaps try it on your target and see if you likewise would agree this is good enough despite the swifterror test issue?

Also, it could be that other tests with a similar code actually now improve, since this was a random result. So just one regression alone should not have to stop the patch, unless it is a sign of a more general issue.


https://reviews.llvm.org/D38128