[llvm] [LoopVectorize] Enhance Vectorization decisions for predicate tail-folded loops with low trip counts (PR #69588)

[David Green via llvm-commits]

https://github.com/davemgreen commented:

I can verify that it doesn't look like there are any problems from this under MVE, which makes heavy use of predicted tails. (Or under the AArch64 tests I ran, both ran without PGO. There was a decent looking improvement in the llvm-test-suite, and there was one problem for MVE but I can fix that separately).

https://github.com/llvm/llvm-project/pull/69588