[llvm] [SLP]Initial support for non-power-of-2 (but still whole register) number of elements in operands. (PR #107273)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> (Meta comment)
> 
> You have several cases where you are relaxing a power-of-two guard for code only given the new VL=N*VLMAX class where I struggle to believe that if that code is correct, it's not also correct for VL=2^N-1. I strongly request that if you believe a piece of code is correct for a non-power-of-two vector, that you remove the guard in a separate commit which adds a test for that particular path.
> 
> Once you've done so, this patch becomes fairly trivial as it's just a new "type" of non-power-of-twos which can flow through.

I'll drop for now most of non-tested stuff, will include just some basic support.

https://github.com/llvm/llvm-project/pull/107273