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

[Simon Pilgrim via llvm-commits]

================
@@ -55688,6 +55688,95 @@ static SDValue truncateAVX512SetCCNoBWI(EVT VT, EVT OpVT, SDValue LHS,
   return SDValue();
 }
 
+// The pattern (setcc (and (broadcast x), (2^n, 2^{n+1}, ...)), (0, 0, ...),
+// eq/ne) is generated when using an integer as a mask. Instead of generating a
+// broadcast + vptest, we can directly move the integer to a mask register.
+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);
+  EVT LoadVT = Load.getSimpleValueType();
+
+  APInt UndefElts;
+  SmallVector<APInt, 32> EltBits;
+  if (!getTargetConstantBitsFromNode(Load, LoadVT.getScalarSizeInBits(),
----------------
RKSimon wrote:

Use Op0.getOperand(1) instead of Load - and remove the Load variable

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