[llvm] [SCEV] Apply loop guards to End in computeMaxBECountForLT (PR #116187)

[Luke Lau via llvm-commits]

lukel97 wrote:

I'm going to close this for now since I'm can't really think of another way of avoiding the compile time hit, and this was kind of just a follow-up that fell out of #115705.

I benchmarked this on SPEC CPU 2017 on RISC-V and whilst the generated code was different (I think I saw one loop vectorized where it wasn't previously?), it didn't have any measurable performance impact so I'm happy to go without it for now. 

If we end up moving to EVL tail folding eventually this might be worth picking up again to remove the overflow check, but on RISC-V vscale will always be a power-of-2 so there might be other ways of avoiding it.

https://github.com/llvm/llvm-project/pull/116187