diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -36402,51 +36402,6 @@
       return SDValue(N, 0);
   }
 
-  // Look for a v2i64/v2f64 VZEXT_MOVL of a node that already produces zeros
-  // in the upper 64 bits.
-  // TODO: Can we generalize this using computeKnownBits.
-  if (N->getOpcode() == X86ISD::VZEXT_MOVL &&
-      (VT == MVT::v2f64 || VT == MVT::v2i64) &&
-      N->getOperand(0).getOpcode() == ISD::BITCAST) {
-    SDValue In = N->getOperand(0).getOperand(0);
-    EVT InVT = In.getValueType();
-    switch (In.getOpcode()) {
-    default:
-      break;
-    case X86ISD::CVTP2SI:   case X86ISD::CVTP2UI:
-    case X86ISD::MCVTP2SI:  case X86ISD::MCVTP2UI:
-    case X86ISD::CVTTP2SI:  case X86ISD::CVTTP2UI:
-    case X86ISD::MCVTTP2SI: case X86ISD::MCVTTP2UI:
-    case X86ISD::CVTSI2P:   case X86ISD::CVTUI2P:
-    case X86ISD::MCVTSI2P:  case X86ISD::MCVTUI2P:
-    case X86ISD::VFPROUND:  case X86ISD::VMFPROUND:
-      if ((InVT == MVT::v4f32 || InVT == MVT::v4i32) &&
-          (In.getOperand(0).getValueType() == MVT::v2f64 ||
-           In.getOperand(0).getValueType() == MVT::v2i64))
-        return N->getOperand(0); // return the bitcast
-      break;
-    case X86ISD::STRICT_CVTTP2SI:
-    case X86ISD::STRICT_CVTTP2UI:
-    case X86ISD::STRICT_CVTSI2P:
-    case X86ISD::STRICT_CVTUI2P:
-    case X86ISD::STRICT_VFPROUND:
-      if ((InVT == MVT::v4f32 || InVT == MVT::v4i32) &&
-          (In.getOperand(1).getValueType() == MVT::v2f64 ||
-           In.getOperand(1).getValueType() == MVT::v2i64))
-        return N->getOperand(0); // return the bitcast
-      break;
-    case X86ISD::CVTPS2PH:
-    case X86ISD::MCVTPS2PH:
-      if (InVT == MVT::v8i16 && In.getOperand(0).getValueType() == MVT::v4f32)
-        return N->getOperand(0); // return the bitcast
-      break;
-    case X86ISD::STRICT_CVTPS2PH:
-      if (InVT == MVT::v8i16 && In.getOperand(1).getValueType() == MVT::v4f32)
-        return N->getOperand(0); // return the bitcast
-      break;
-    }
-  }
-
   // Pull subvector inserts into undef through VZEXT_MOVL by making it an
   // insert into a zero vector. This helps get VZEXT_MOVL closer to
   // scalar_to_vectors where 256/512 are canonicalized to an insert and a
@@ -36614,6 +36569,38 @@
     if (SimplifyDemandedVectorElts(Src, SrcElts, SrcUndef, SrcZero, TLO,
                                    Depth + 1))
       return true;
+    LLVM_FALLTHROUGH;
+  }
+  case X86ISD::CVTP2SI:
+  case X86ISD::CVTP2UI:
+  case X86ISD::MCVTP2SI:
+  case X86ISD::MCVTP2UI:
+  case X86ISD::CVTTP2SI:
+  case X86ISD::CVTTP2UI:
+  case X86ISD::MCVTTP2SI:
+  case X86ISD::MCVTTP2UI:
+  case X86ISD::MCVTSI2P:
+  case X86ISD::MCVTUI2P:
+  case X86ISD::VFPROUND:
+  case X86ISD::VMFPROUND:
+  case X86ISD::CVTPS2PH:
+  case X86ISD::MCVTPS2PH: {
+    // Conversions - upper elements are known zero.
+    int NumSrcElts = Op.getOperand(0).getValueType().getVectorNumElements();
+    if (NumElts > NumSrcElts)
+      KnownZero = APInt::getHighBitsSet(NumElts, NumElts - NumSrcElts);
+    break;
+  }
+  case X86ISD::STRICT_CVTTP2SI:
+  case X86ISD::STRICT_CVTTP2UI:
+  case X86ISD::STRICT_CVTSI2P:
+  case X86ISD::STRICT_CVTUI2P:
+  case X86ISD::STRICT_VFPROUND:
+  case X86ISD::STRICT_CVTPS2PH: {
+    // Strict Conversions - upper elements are known zero.
+    int NumSrcElts = Op.getOperand(1).getValueType().getVectorNumElements();
+    if (NumElts > NumSrcElts)
+      KnownZero = APInt::getHighBitsSet(NumElts, NumElts - NumSrcElts);
     break;
   }
   case X86ISD::PACKSS:
@@ -36702,6 +36689,18 @@
     KnownUndef = LHSUndef & RHSUndef;
     break;
   }
+  case X86ISD::VZEXT_MOVL: {
+    SDValue Src = Op.getOperand(0);
+    if (SimplifyDemandedVectorElts(Src, DemandedElts, KnownUndef, KnownZero,
+                                   TLO, Depth + 1))
+      return true;
+
+    // If upper demanded elements are already zero then we have nothing to do.
+    // TODO - move this into shuffle code below.
+    if ((NumElts - 1) <= (int)(KnownZero | ~DemandedElts).countLeadingOnes())
+      return TLO.CombineTo(Op, Src);
+    break;
+  }
   case X86ISD::VBROADCAST: {
     SDValue Src = Op.getOperand(0);
     MVT SrcVT = Src.getSimpleValueType();