[llvm] [AArch64][NeoverseV1] Fix integer divide throughput (PR #224592)

[Julien Villette via llvm-commits]

jvillette38 wrote:

> The distribution was actually a little more spread than I remember it being, but the average tends to be much closer to the minimum than the maximum. And trying to schedule 20 * dispatch width instructions around a divide can end up being detrimental to performance.
> 
> https://gist.github.com/davemgreen/54fb478687ace14281c8cb14236fd7e3 was written more than a year ago now, but shows how I was measuring it and getting distributions.

Hello,
Very interesting! What do you expect for this PR? Do we add these results to update throughput/latency?
I guess we have to update integer and floating point divisions...

https://github.com/llvm/llvm-project/pull/224592