[PATCH] D28934: Write a new SSAUpdater

[Daniel Berlin via Phabricator via llvm-commits]

dberlin added a comment.

It's algorithm was definitely not usable, time wise, for MemorySSA (i tried it), it was about 2-5x slower, just because of the amount of work it does per thing.  MemorySSA is, as i said, a bit weird.

I don't remember the llvmdev history, but my understanding was it was really bad for SROA (I'm sure chandler can saying more).

In any case, let me add the SCC post-pass to this (the do scc walking as we rename algorithm is ... more complex, and i'm not sure it's worth it speed wise) so someone can play with it more. We have an SCC iterator, so i think i can make this happen very easily and it should be fast.

My understanding, BTW (not verified), is that with one additional recursive simplification function above, the only difference between this and minimal form should be in irreducible control flow.

But i'll check :)


https://reviews.llvm.org/D28934