[llvm] [LV, VP]VP intrinsics support for the Loop Vectorizer + adding new tail-folding mode using EVL. (PR #76172)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> > > It looks like there's a UBSan failure when running some of the tests (comment inline, looks unrelated to the patch, but the tests need to force a vector width > 1), would be good to check with an UBSan build.
> > 
> > 
> > Possible, will double check.
> > > Also, the target-independent tests don't exercise the EVL code path due to missing support for VP intrinsics at the moment AFACIT; I think we need to have them with `RISCV` targets as well, as otherwise some code paths aren't exercised in the EVL code path (e.g. interleave groups, gather/scatter, reverse memory ops, masked ops)
> > 
> > 
> > I had all these tests under RISCV. You asked to make them target independent. Should I again move them to RISCV?
> 
> I'd just move the individual files to RISCV. Earlier comment was if this would retain test coverage (i.e. by adding a something similar like `-force-target-supports-scalable-vectors` for vector-predication), sorry that wasn't clearer from the previous comment

And I answered that all these tests are target-specific, but you still insisted on making them independent. What's changed?

https://github.com/llvm/llvm-project/pull/76172