[llvm] VectorWiden pass to widen aleady vectorized instrctions (PR #67029)

[Sander de Smalen via llvm-commits]

https://github.com/sdesmalen-arm commented:

Left a bunch of comments which will hopefully clarify the code a bit more when addressed, because I'm still trying to make sense of how things work. Other than that, the code in this pass should conceptually distinguish between:
* Target-agnostic vector widening
* Target-specific choices about whether to widen

You've added a TargetTransformInfo interface for the latter, which is good, but the code in this pass makes various assumptions based on the implementation of `AArch64TargetTransformInfo::considerToWiden`. For example, the code in the generic pass assumes that the vector length is a `[vscale x] 4` and the tests only seem to cover whatever AArch64 enables with its TTI hook.

https://github.com/llvm/llvm-project/pull/67029