Index: llvm/lib/Analysis/InstructionSimplify.cpp
===================================================================
--- llvm/lib/Analysis/InstructionSimplify.cpp
+++ llvm/lib/Analysis/InstructionSimplify.cpp
@@ -4095,6 +4095,11 @@
 
 /// See if we can fold the given phi. If not, returns null.
 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
+  // LCSSA PHIs only have one input value
+  // return nullptr here to avoid simplifying them away
+  // since they are necessary to hold out of the loop dependecies.
+  if (PN->getNumIncomingValues() < 2)
+    return nullptr;
   // If all of the PHI's incoming values are the same then replace the PHI node
   // with the common value.
   Value *CommonValue = nullptr;
Index: llvm/test/CodeGen/AMDGPU/lcssa-phi.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/lcssa-phi.mir
@@ -0,0 +1,31 @@
+# RUN: llc -march=amdgcn -run-pass early-cse -o - %s | FileCheck %s
+# Check that LCSSA PHI was not eliminated
+# CHECK:     %counter.lcssa = phi i32 [ %counter, %loop ]
+
+--- |
+  define amdgpu_kernel void @lcssa-phi(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+  entry:
+    %tid = call i32 @llvm.amdgcn.workitem.id.x()
+    br label %loop
+  
+  loop:                                             ; preds = %flow, %entry
+    %counter = phi i32 [ 0, %entry ], [ %inc, %flow ]
+    %cond = icmp ugt i32 %counter, %tid
+    br i1 %cond, label %endloop, label %flow
+  
+  flow:                                             ; preds = %loop
+    %inc = add i32 %counter, 1
+    br label %loop
+  
+  endloop:                                          ; preds = %loop
+    %counter.lcssa = phi i32 [ %counter, %loop ]
+    store i32 %counter.lcssa, i32 addrspace(1)* %out
+    ret void
+  }
+  
+  ; Function Attrs: nounwind readnone speculatable
+  declare i32 @llvm.amdgcn.workitem.id.x() #0
+  
+  attributes #0 = { nounwind readnone speculatable }
+  
+...