[llvm] [AMDGPU][SIInsertWaitCnts] Use RegUnits-based tracking (PR #162077)

[Fabian Ritter via llvm-commits]

================
@@ -1035,9 +991,9 @@ void WaitcntBrackets::updateByEvent(WaitEventType E, MachineInstr &Inst) {
     // Special cases where implicit register defs exists, such as M0 or VCC,
     // but none with memory instructions.
     for (const MachineOperand &Op : Inst.defs()) {
-      RegInterval Interval = getRegInterval(&Inst, Op);
       if (T == LOAD_CNT || T == SAMPLE_CNT || T == BVH_CNT) {
-        if (Interval.first >= NUM_ALL_VGPRS)
+        if (!Context->TRI->isVectorRegister(*Context->MRI,
+                                            Op.getReg())) // TODO: add wrapper
----------------
ritter-x2a wrote:

```suggestion
        if (!TRI->isVectorRegister(*MRI, Op.getReg())) // TODO: add wrapper
```
`TRI` and `MRI` from the `Context` are available in local  variables.
A `bool SIInsertWaitcnts::isVectorRegister(Register)` wrapper might clean this and other places even further, though.

https://github.com/llvm/llvm-project/pull/162077