[llvm] [RISCV] Don't use EVL/Mask for vid when lowering vp.reverse (PR #123048)

[Craig Topper via llvm-commits]

topperc wrote:

> This sorta feels like solving the problem in the wrong place. If the vectorizer is unprofitable emitting the vp.reverse, why not just change the vectorizer to emit a vp.reverse at vlmax or better an unpredicated vector.reverse?

The vectorizer can't emit a vlmax reverse. The reverse needs to put the element at EVL-1 into element 0 and element 0 into EVL-1. So the vrsub needs to do EVL-vid. It can't do VLMAX-vid.

https://github.com/llvm/llvm-project/pull/123048