diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -120,9 +120,9 @@
   void addUsersToMoveToVALUWorklist(unsigned Reg, MachineRegisterInfo &MRI,
                                     SetVectorType &Worklist) const;
 
-  void
-  addSCCDefUsersToVALUWorklist(MachineInstr &SCCDefInst,
-                               SetVectorType &Worklist) const;
+  void addSCCDefUsersToVALUWorklist(MachineOperand &Op,
+                                    MachineInstr &SCCDefInst,
+                                    SetVectorType &Worklist) const;
 
   const TargetRegisterClass *
   getDestEquivalentVGPRClass(const MachineInstr &Inst) const;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -4320,8 +4320,10 @@
     for (unsigned i = Inst.getNumOperands() - 1; i > 0; --i) {
       MachineOperand &Op = Inst.getOperand(i);
       if (Op.isReg() && Op.getReg() == AMDGPU::SCC) {
+        // Only propagate through live-def of SCC.
+        if (Op.isDef() && !Op.isDead())
+          addSCCDefUsersToVALUWorklist(Op, Inst, Worklist);
         Inst.RemoveOperand(i);
-        addSCCDefUsersToVALUWorklist(Inst, Worklist);
       }
     }
 
@@ -5014,19 +5016,23 @@
   addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
 }
 
-void SIInstrInfo::addSCCDefUsersToVALUWorklist(
-    MachineInstr &SCCDefInst, SetVectorType &Worklist) const {
+void SIInstrInfo::addSCCDefUsersToVALUWorklist(MachineOperand &Op,
+                                               MachineInstr &SCCDefInst,
+                                               SetVectorType &Worklist) const {
+  // Ensure that def inst defines SCC, which is still live.
+  assert(Op.isReg() && Op.getReg() == AMDGPU::SCC && Op.isDef() &&
+         !Op.isDead() && Op.getParent() == &SCCDefInst);
   // This assumes that all the users of SCC are in the same block
   // as the SCC def.
-  for (MachineInstr &MI :
-       make_range(MachineBasicBlock::iterator(SCCDefInst),
-                      SCCDefInst.getParent()->end())) {
+  for (MachineInstr &MI : // Skip the def inst itself.
+       make_range(std::next(MachineBasicBlock::iterator(SCCDefInst)),
+                  SCCDefInst.getParent()->end())) {
+    // Check if SCC is used first.
+    if (MI.findRegisterUseOperandIdx(AMDGPU::SCC, false, &RI) != -1)
+      Worklist.insert(&MI);
     // Exit if we find another SCC def.
     if (MI.findRegisterDefOperandIdx(AMDGPU::SCC, false, false, &RI) != -1)
       return;
-
-    if (MI.findRegisterUseOperandIdx(AMDGPU::SCC, false, &RI) != -1)
-      Worklist.insert(&MI);
   }
 }
 
diff --git a/llvm/test/CodeGen/AMDGPU/udivrem64.ll b/llvm/test/CodeGen/AMDGPU/udivrem64.ll
--- a/llvm/test/CodeGen/AMDGPU/udivrem64.ll
+++ b/llvm/test/CodeGen/AMDGPU/udivrem64.ll
@@ -168,3 +168,17 @@
   store i64 %result, i64 addrspace(1)* %out
   ret void
 }
+
+;FUNC-LABEL: {{^}}test_udiv_k:
+;GCN: v_mul{{.+}} v{{[0-9]+}}, v{{[0-9]+}}, 24
+;GCN: v_mul{{.+}} v{{[0-9]+}}, v{{[0-9]+}}, 24
+;GCN: v_mul{{.+}} v{{[0-9]+}}, v{{[0-9]+}}, 24
+;GCN: v_add
+;GCN: v_addc
+;GCN: v_addc
+;GCN: s_endpgm
+define amdgpu_kernel void @test_udiv_k(i64 addrspace(1)* %out, i64 %x) {
+  %result = udiv i64 24, %x
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}