[PATCH] D44919: [LoopUnroll][NFC] Remove redundant canPeel check

[Max Kazantsev via Phabricator via llvm-commits]

mkazantsev added a comment.

Note that we might not reach `computePeelCount` not because of full unrolling, but also because of user-forced unroll count. For example, if I write a loop

  for (int i = 0; i < 400; i++)

and then force its unrolling by 4, with your patch we will first peel off 2 iterations, then unroll 396 iterations by 4 and 2 in tail computation. I don't know what motivation stands behind https://reviews.llvm.org/D44880 in sense of performance situations, but at least in some cases your patch will cause harm. If you agree that it's a problem, then it should be reworked.


Repository:
  rL LLVM

https://reviews.llvm.org/D44919