[llvm] [LoopPeel] Peel last iteration to enable load widening (PR #173420)

[Nikita Popov via llvm-commits]

https://github.com/nikic commented:

As a very high level question, why do we want to peel here instead of unroll?

For example, if we're accessing 3 elements, we could unroll twice and use a 4 + 2 byte load -- instead of two overlapping 4 byte loads.

Esp for the case of unrolling twice (where code size impact should be about the same as peeling), it seems like unrolling would be more beneficial, esp as it can also enable other optimizations?

https://github.com/llvm/llvm-project/pull/173420