diff --git a/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp b/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
--- a/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
@@ -53,7 +53,8 @@
     const bool IsOffset = ValuePair.second;
     for (auto &U : ValuePair.first->uses()) {
       auto *I = cast<Instruction>(U.getUser());
-
+      if (isa<PHINode>(I))
+        continue;
       if (auto *LI = dyn_cast<LoadInst>(I)) {
         // Ignore non-volatile loads, they are always ok.
         if (!LI->isSimple()) return false;
@@ -256,6 +257,10 @@
       if (Load->isVolatile())
         return false;
       Worklist.insert(Load);
+    } else if (auto *PHI = dyn_cast<PHINode>(Inst)) {
+      for (unsigned int I = 0; I < PHI->getNumOperands(); ++I)
+        Worklist.insert(cast<Instruction>(PHI->getOperand(I)));
+      Worklist.insert(PHI);
     } else if (isa<GetElementPtrInst>(Inst) || isa<BitCastInst>(Inst)) {
       Worklist.insert(Inst);
       if (!collectUsers(*Inst))
@@ -293,6 +298,18 @@
     IC.InsertNewInstWith(NewI, *LT);
     IC.replaceInstUsesWith(*LT, NewI);
     WorkMap[LT] = NewI;
+  } else if (auto *PHI = dyn_cast<PHINode>(I)) {
+    SmallVector<Value *, 8> ReplacedOperands;
+    for (unsigned int I = 0; I < PHI->getNumOperands(); ++I) {
+      ReplacedOperands.push_back(getReplacement(PHI->getOperand(I)));
+      assert(ReplacedOperands.back() && "Operand not replaced");
+    }
+    auto *NewPHI = PHINode::Create(PHI->getType(), PHI->getNumIncomingValues(),
+                                   PHI->getName());
+    NewPHI->takeName(PHI);
+    IC.InsertNewInstWith(NewPHI, *PHI);
+    IC.replaceInstUsesWith(*PHI, NewPHI);
+    WorkMap[PHI] = NewPHI;
   } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
     auto *V = getReplacement(GEP->getPointerOperand());
     assert(V && "Operand not replaced");
diff --git a/llvm/test/Transforms/InstCombine/replace-alloca-phi.ll b/llvm/test/Transforms/InstCombine/replace-alloca-phi.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/replace-alloca-phi.ll
@@ -0,0 +1,21 @@
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -passes=instcombine -S -o - %s | FileCheck %s
+
+; CHECK-LABEL: @remove_alloca
+define i8 @remove_alloca(i1 %flag, i32 %idx, [32 x i8] addrspace(4)* align 8 %arg) {
+entry:
+  %xp = alloca [32 x i8], align 8, addrspace(5)
+  call void @llvm.memcpy.p5.p4.i64([32 x i8] addrspace(5)* align 8 %xp, [32 x i8] addrspace(4)* align 8 %arg, i64 1360, i1 false)
+  br i1 %flag, label %true, label %false
+true:
+  %gep.true = getelementptr inbounds [32 x i8], [32 x i8] addrspace(5)* %xp, i32 0, i32 0
+  br label %sink
+false:
+  %gep.false = getelementptr inbounds [32 x i8], [32 x i8] addrspace(5)* %xp, i32 0, i32 0
+  br label %sink
+sink:
+  %x.2 = phi i8 addrspace(5)* [%gep.true, %true], [%gep.false, %false]
+  %x.3 = load i8, i8 addrspace(5)* %x.2, align 4
+  ret i8 %x.3
+}
+
+declare void @llvm.memcpy.p5.p4.i64([32 x i8] addrspace(5)* align 8, [32 x i8] addrspace(4)* align 8, i64, i1)