[llvm] [X86] Generate `kmov` for masking integers (PR #120593)

[Simon Pilgrim via llvm-commits]

================
@@ -55293,6 +55293,95 @@ static SDValue truncateAVX512SetCCNoBWI(EVT VT, EVT OpVT, SDValue LHS,
   return SDValue();
 }
 
+static SDValue combineAVX512SetCCToKMOV(EVT VT, SDValue Op0, ISD::CondCode CC,
+                                        const SDLoc &DL, SelectionDAG &DAG,
+                                        const X86Subtarget &Subtarget) {
+  if (CC != ISD::SETNE && CC != ISD::SETEQ)
+    return SDValue();
+
+  if (!Subtarget.hasAVX512())
+    return SDValue();
+
+  if (Op0.getOpcode() != ISD::AND)
+    return SDValue();
+
+  SDValue Broadcast = Op0.getOperand(0);
+  if (Broadcast.getOpcode() != X86ISD::VBROADCAST &&
+      Broadcast.getOpcode() != X86ISD::VBROADCAST_LOAD)
+    return SDValue();
+
+  SDValue Load = Op0.getOperand(1);
+  if (Load.getOpcode() != ISD::LOAD)
----------------
RKSimon wrote:

You should be able to use getTargetConstantBitsFromNode to extract the element bits/undefs for the load:
```cpp
APInt UndefElts;
SmallVector<APInt> EltBits;
if (!getTargetConstantBitsFromNode(Op0.getOperand(1), VT.getScalarSizeInBits(), UndefElts, EltBits))
  return SDValue();
```

https://github.com/llvm/llvm-project/pull/120593