[llvm] [VPlan] Model address separately. (PR #72164)

[via llvm-commits]

https://github.com/ayalz approved this pull request.

Thanks, looks good to me, with a final cleanup nit.

Note that VPVectorPointerRecipes could be quite easily introduced as an "adjusting" VPlan-to-VPlan pass, placed between every consecutive VPWidenMemoryInstructionRecipe and its address operand. With some indication marking a VPWidenMemoryInstructionRecipe as adjusted or not, validated at execute time. But further breaking up of VPWidenMemoryInstructionRecipe, unrolling as a VPlan-to-VPlan pass, and starting with initial scalar VPlan, should eventually clarify this.

Hope subsequent scheduling is indifferent to the clustering of loads/stores separate from their GEPs.

https://github.com/llvm/llvm-project/pull/72164