[llvm] [VPlan] Strip scalar/first-lane guard in removeRedundantCanIVs (PR #194338)

[Florian Hahn via llvm-commits]

https://github.com/fhahn commented:

> The guard is unnecessary, and stripping it leads to optimization benefit.

But it can also lead to regressions, as the transformation increase register pressure (one more vector register live across the loop, so in same cases it will cause extra spills which will likely be worse than an additional broadcast , e.g. in `llvm/test/Transforms/LoopVectorize/X86/widen-canonical-iv-register-pressure.ll`?

On some platforms, the broadcast may also be fold-able in to vector adds and effectively free (e.g. on RISCV https://llvm.godbolt.org/z/TfPqhTeq9)

Hence the motivation to make this driven by cost/register pressure checks.

https://github.com/llvm/llvm-project/pull/194338