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
@@ -36464,9 +36464,9 @@
         (V.getOpcode() == X86ISD::PSHUFLW ||
          V.getOpcode() == X86ISD::PSHUFHW) &&
         V.getOpcode() != N.getOpcode() &&
-        V.hasOneUse()) {
+        V.hasOneUse() && V.getOperand(0).hasOneUse()) {
       SDValue D = peekThroughOneUseBitcasts(V.getOperand(0));
-      if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
+      if (D.getOpcode() == X86ISD::PSHUFD) {
         SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
         SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
         int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
@@ -36903,10 +36903,11 @@
   // 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
   // 128-bit scalar_to_vector. This reduces the number of isel patterns.
-  if (N->getOpcode() == X86ISD::VZEXT_MOVL && !DCI.isBeforeLegalizeOps()) {
+  if (N->getOpcode() == X86ISD::VZEXT_MOVL && !DCI.isBeforeLegalizeOps() &&
+      N->getOperand(0).hasOneUse()) {
     SDValue V = peekThroughOneUseBitcasts(N->getOperand(0));
 
-    if (V.getOpcode() == ISD::INSERT_SUBVECTOR && V.hasOneUse() &&
+    if (V.getOpcode() == ISD::INSERT_SUBVECTOR &&
         V.getOperand(0).isUndef() && isNullConstant(V.getOperand(2))) {
       SDValue In = V.getOperand(1);
       MVT SubVT =