[PATCH] D103362: [LoopUnroll] Separate peeling from unrolling

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

In D103362#2796508 <https://reviews.llvm.org/D103362#2796508>, @reames wrote:

> I'm not sure we can drop the behavior of both peeling and unrolling.  The tests use the command lines, but I believe there's a pragma/metadata mechanism which can be used to achieve the same effect.  (Vague memory of prior conversations w/ @Meinersbur)

While there is indeed metadata to set an unroll count, I don't believe there is metadata to set a peel count. If an unroll count is set via metadata, we will not attempt to compute a peel count.

> Supporting the existing behavior with the new code structure doesn't seem too hard, would you mind updating to continue supporting both?  (Well, while fixing the miscompile noted in D103620 <https://reviews.llvm.org/D103620>.)

I don't think the existing behavior makes sense, so I would prefer not to preserve it. It seems pointless to go out of our way to allow a scenario in testing that cannot appear as part of the optimization pipeline.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D103362/new/

https://reviews.llvm.org/D103362