[llvm] [AArch64] Improve scheduling latency into Bundles (PR #86310)

[David Green via llvm-commits]

================
@@ -235,7 +235,9 @@ class TargetSubtargetInfo : public MCSubtargetInfo {
   // and UseOpIdx are the indices of the operands in Def and Use, respectively.
   // Otherwise, either may be -1.
   virtual void adjustSchedDependency(SUnit *Def, int DefOpIdx, SUnit *Use,
-                                     int UseOpIdx, SDep &Dep) const {}
+                                     int UseOpIdx, SDep &Dep,
+                                     const TargetSchedModel *SchedModel) const {
----------------
davemgreen wrote:

It looks like the AMDGPU InstrInfo will create a second TargetScheduleModel that it can use from InstrInfo. I think it would be better if there was only one, the one from the MachineScheduler/MachinePipeliner that it is being called from.

I'm open to suggestions if the AMD folks want it to work differently (preferably so long as it doesn't involve creating a new TargetScheduleModel!)

https://github.com/llvm/llvm-project/pull/86310