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

via llvm-commits llvm-commits at lists.llvm.org
Wed Jan 22 11:03:06 PST 2025


================
@@ -55447,6 +55447,82 @@ 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);
+  EVT LoadVT = Load.getSimpleValueType();
+
+  APInt UndefElts;
+  SmallVector<APInt, 32> EltBits;
+  if (!getTargetConstantBitsFromNode(Load, LoadVT.getScalarSizeInBits(),
+                                     UndefElts, EltBits,
+                                     /*AllowWholeUndefs*/ true,
+                                     /*AllowPartialUndefs*/ false) ||
+      UndefElts[0] || !EltBits[0].isPowerOf2())
+    return SDValue();
+
+  unsigned N = EltBits[0].logBase2();
+  unsigned Len = UndefElts.getBitWidth();
+  for (unsigned I = 1; I != Len; ++I) {
+    if (UndefElts[I]) {
+      for (unsigned J = I + 1; J != Len; ++J)
+        if (!UndefElts[J])
+          return SDValue();
+      break;
+    }
+
+    if (EltBits[I] != 1 << (N + I))
+      return SDValue();
+  }
+
+  SDValue BroadcastOp = Broadcast.getOpcode() == X86ISD::VBROADCAST
+                            ? Broadcast.getOperand(0)
+                            : Broadcast.getOperand(1);
+  MVT BroadcastOpVT = BroadcastOp.getSimpleValueType();
+  SDValue Masked = BroadcastOp;
+  if (N != 0) {
+    unsigned Mask = (1ULL << Len) - 1;
+    SDValue ShiftedValue = DAG.getNode(ISD::SRL, DL, BroadcastOpVT, BroadcastOp,
+                                       DAG.getConstant(N, DL, BroadcastOpVT));
+    Masked = DAG.getNode(ISD::AND, DL, BroadcastOpVT, ShiftedValue,
----------------
abhishek-kaushik22 wrote:

Sorry I missed that, added the test for this case as well.

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


More information about the llvm-commits mailing list