[llvm] [VPlan] Build initial VPlan 0 using HCFGBuilder for inner loops. (NFC) (PR #124432)

[via llvm-commits]

ayalz wrote:

> Use HCFGBuilder to build an initial VPlan 0, which wraps all input instructions in VPInstructions and update tryToBuildVPlanWithVPRecipes to replace the VPInstructions with widened recipes.
> 
> At the moment, widened recipes are created based on the underlying instruction of the VPInstruction. Masks are also still created based on the input IR basic blocks and the loop CFG is flattened in the main loop processing the VPInstructions.
> 
> This patch also incldues support for Switch instructions in HCFGBuilder using just a VPInstruction with Instruction::Switch opcode.
> 
> There are multiple follow-ups planned:
> 
> * Use VPIRInstructions instead of VPInstructions in HCFGBuilder,
> * Perform predication on the VPlan directly,
> * Unify code constructing VPlan 0 to be shared by both inner and outer loop code paths.
> * Construct VPlan 0 once, clone subsequent ones for VFs

This is a very nice step forward!
Trying to align with the roadmap sketched in https://llvm.org/devmtg/2023-10/slides/techtalks/Hahn-VPlan-StatusUpdateAndRoadmap.pdf#page=37 and its "Direction 1: Refactor VPlan Creation", the VPlan built by HCFGBuilder seems to correspond to the 2nd VPlan of "buildLoop". Several aspects deserves fixing for this VPlan to represent a scalar version of the to-be-vectorized loop (i.e., for UF=VF=1) more consistently:
- Phi's, as all other recipes of this VPlan, are scalar. However, this VPlan uses `VPWidenPhiRecipes` which represent vectorized and/or unrolled phi's.
- `VPInstructions` should rely on their opcode, operands and flags **only**. However, this VPlan uses `VPInstructions` which rely on their underlying `Instructions`, disregarding their flags.
Note that, in contrast, `VPIRInstructions`, which are associated with the 1st VPlan of "wrapInput", represent the immutable `Instructions` in-place.

https://github.com/llvm/llvm-project/pull/124432