[llvm] [VPlan] Add vputils::reconstructSSA (PR #212209)

[Luke Lau via llvm-commits]

lukel97 wrote:

> Since this patch currently lacks landed real-world use cases, could we re-implement this block in `VPlanTransforms::foldTailByMasking`:
> ``` c++
>   // Insert phis for values coming past the end of the tail.
>   Builder.setInsertPoint(Latch, Latch->begin());
>   for (const auto &[V, Users] : NeedsPhi) {
>     if (isa<VPIRValue>(V))
>       continue;
>     VPValue *TailVal = Plan.getPoison(V->getScalarType());
>     VPIRFlags Flags;
>     assert(llvm::count_if(Users, IsaPred<VPReductionPHIRecipe>) <= 1 &&
>            "Value used by more than two reduction phis?");
>     auto *RedIt = find_if(Users, IsaPred<VPReductionPHIRecipe>);
>     auto *RdxPhi =
>         RedIt != Users.end() ? cast<VPReductionPHIRecipe>(*RedIt) : nullptr;
>     if (RdxPhi && !RdxPhi->isInLoop()) {
>       TailVal = RdxPhi;
>       Flags = *RdxPhi;
>     }
> 
>     VPInstruction *Phi = Builder.createScalarPhi({V, TailVal}, {}, "", Flags);
>     for (VPUser *U : Users)
>       U->replaceUsesOfWith(V, Phi);
>   }
> ```
> by `reconstructSSA({{OrigLatch, V}, {Header, TailVal}}, Latch)`?

This was discussed earlier in https://github.com/llvm/llvm-project/pull/201784#discussion_r3378409300

We could, but I decided against it because we only never need to insert one phi there. reconstructSSA doesn't replace any of the uses so it doesn't shorten the code

https://github.com/llvm/llvm-project/pull/212209