[llvm] [VPlan] Run removeDeadRecipes early. (PR #190191)

[Florian Hahn via llvm-commits]

================

----------------
fhahn wrote:

Thanks, I think most of the regressions should be gone now, and the only cases were we pick VPWidenCanonicalIV are ccases where we would spill (mostly in target-inpendent tests, where TTI returns a low vector register count)

https://github.com/llvm/llvm-project/pull/190191