[llvm] [VPlan] Thread plan to VPBuilder (NFC) (PR #125364)

[Ramkumar Ramachandra via llvm-commits]

artagnon wrote:

> It seems like there is more to this PR than just passing a VPlan object reference to the VPBuilder constructor, since it's also changing VPRecipeBuilder to instantiate a `VPBuilder` object every time a `VPRecipeBuilder` object is constructed. Unless I'm missing something I can't see any reason why they are tied together in the same patch?

This is due to the way they're inherently tied up. If `LoopVectorizationPlanner` were to own a `VPBuilder`, then it needs to be passed a `Plan` so it can construct a `VPBuilder`, and passing a plan to the lv-planner is just wrong: lv-planner keeps a list of candidate plans. The changes then cascade to `VPRecipeBuilder` having to own a `VPBuilder` instead.

> Can you explain more what the patch is trying to achieve with the refactoring so we can see the bigger picture?

VPlan-level constant folding: https://github.com/llvm/llvm-project/pull/125365. The constant-folder shows real improvements in some cases, with no regressions.

https://github.com/llvm/llvm-project/pull/125364