[clang] [llvm] [HLSL] Implement `WaveReadLaneAt` intrinsic (PR #111010)

[Steven Perron via cfe-commits]

================
@@ -428,6 +431,7 @@ bool SPIRVInstructionSelector::spvSelect(Register ResVReg,
 
   case TargetOpcode::G_INTRINSIC:
   case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
+  case TargetOpcode::G_INTRINSIC_CONVERGENT:
----------------
s-perron wrote:

The extra convergence information is being used when we generate the merge instructions. They should not change how the instrinsic itself is generated. The convergence token help us distinguish between:

```
int i;
for(i = 0; i < N; ++i) {
  if (cond(i)) {
    // The "active" lanes are those where cond was true in the same iteration.
    WaveReadLaneAt(i);
    break;
  }
}


for(i = 0; i < N; ++i) {
  if (cond(i))
    break;
}
// All lanes are active because they all reconverged at the end of the loop.
WaveReadLaneAt(i);
```

https://github.com/llvm/llvm-project/pull/111010