[llvm] [AArch64] Disable Pre-RA Scheduler for Neoverse V2 (PR #127784)

[David Green via llvm-commits]

davemgreen wrote:

Hi - Another of the things that is possibly quite important is to schedule for register pressure. That might be related to the copy elimination you mention, I wasn't sure, it might be different. It's one those things that could happen to be right from the original order of the instructions (people tend to write code which uses variables close together), but if it is needed then something should be improving the order if it can.

@c-rhodes has been looking at a couple of cases recently where we have found scheduling has not been performing as well as it could. Either the inorder scheduling model or no scheduling was better for specific cases, but when we tried that on a larger selection of benchmarks the results ended up as a wash I believe, with some things getting better and some worse in about equal proportions. The GCC team recently tried this (they have very old scheduling models too), but ended up going back on it I believe. It still worries me to remove it entirely but at the end of the day it is the data that matters if we can show it reliably.

https://github.com/llvm/llvm-project/pull/127784