[llvm] [AMDGPU] Fix set_inactive known bits to intersect both operands (PR #201817)

[Arseniy Obolenskiy via llvm-commits]

================
@@ -5848,13 +5848,28 @@ bool AMDGPUTargetLowering::SimplifyDemandedBitsForTargetNode(
     switch (Op.getConstantOperandVal(0)) {
     case Intrinsic::amdgcn_readfirstlane:
     case Intrinsic::amdgcn_readlane:
-    case Intrinsic::amdgcn_set_inactive:
     case Intrinsic::amdgcn_wwm: {
       if (SimplifyDemandedBits(Op.getOperand(1), OriginalDemandedBits,
                                OriginalDemandedElts, Known, TLO, Depth + 1))
         return true;
       break;
     }
+    case Intrinsic::amdgcn_set_inactive:
+    case Intrinsic::amdgcn_set_inactive_chain_arg: {
----------------
aobolensk wrote:

Done

https://github.com/llvm/llvm-project/pull/201817