[PATCH] D121788: [AArch64] Increase MaxInterleaveFactor to 4

Tue Mar 22 08:51:08 PDT 2022

jaykang10 added a comment.

Thanks for comment! @dmgreen

> Do you have any performance results for changing this? I've not had much luck with trying it in the past, and it obviously can change quite a lot. It can certainly help in places, but I've found that if you turn it up too high you just end up over-unrolling loops, not getting into the fast loop body as much. It can obviously depend on the input code and loop counts tough. Perhaps it needs some better costmodelling?

I was able to see the overall performance number slightly up for an internal benchmark on neoverse-n1 but we would need to tune something like cost model according to micro architectures.

> I was still hoping to get D118979 <https://reviews.llvm.org/D118979> in because it should help quite a bit - and it on it's own increases the number of items processed per vector element, and this will increase it further. We have cleaned up quite a few of the places it doesn't do as well, there are just a few that have been stuck in review a while. Perhaps it makes sense to try and push that through, then re-evaluate this on top?

I agree with you. Let's visit this patch later.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D121788/new/

https://reviews.llvm.org/D121788