[llvm] [AMDGPU] Add new llvm.amdgcn.wave.shuffle intrinsic (PR #167372)

[Jay Foad via llvm-commits]

================
@@ -7280,6 +7280,81 @@ static SDValue lowerLaneOp(const SITargetLowering &TLI, SDNode *N,
   return DAG.getBitcast(VT, UnrolledLaneOp);
 }
 
+static SDValue lowerWaveShuffle(const SITargetLowering &TLI, SDNode *N,
+                                    SelectionDAG &DAG) {
+  EVT VT = N->getValueType(0);
+  assert(VT.getSizeInBits() == 32);
----------------
jayfoad wrote:

Nit: this shouldn't be an assert because there is nothing to stop the user from calling the intrinsic with a different type like i64. Instead you can `return SDValue()` here and the compiler should fail with a nicer "cannot select" diagnostic.

https://github.com/llvm/llvm-project/pull/167372