[PATCH] D15559: [LoopUnrollRuntime] Do unroll when the unroll factor is 2

Junmo Park via llvm-commits llvm-commits at lists.llvm.org
Thu Dec 17 13:46:25 PST 2015


flyingforyou added a comment.

Thanks Charlie for comment.

> I had a look at A57 in A64 mode and noticed a 4.81% regression in spec2006's hmmer benchmark. Do you have access to this suite?


I will figure it out what's wrong. I set up spec2006, recently.

> @hfinkel This needs a test case (either in test/Transforms/LoopUnroll/AArch64 or in test/Transforms/LoopUnroll). Also, why does this logic belong here and not in Expander.isHighCostExpansion?




> @zzheng I'm OK with this change for AArch64. Like Hal said, please add a test case.


I also think cheking the unroll factor only is not enough for figuring out that TripCountSC is really expensive or not. 
So I will upload new logic & test case next week. I hope that could be acceptable.

Junmo.


http://reviews.llvm.org/D15559





More information about the llvm-commits mailing list