[PATCH] D144485: [LV][RISCV] Don't interleave scalable vector loops

[Philip Reames via Phabricator via llvm-commits]

reames added a comment.

In D144485#4142300 <https://reviews.llvm.org/D144485#4142300>, @craig.topper wrote:

> I'm not sure we should leave this on for fixed vectors either once we go to LMUL=2.

I think I agree with this point.  In a follow on patch, we should probably revisit this part of the heuristic as well.

A more general framing here is that interleave and LMUL both increase register pressure and tail effects, and that as we increase one we probably need to decrease the other by a roughly equal ratio.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D144485/new/

https://reviews.llvm.org/D144485