Index: lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -444,9 +444,22 @@
       return false;
 
     if (StoreInst *SI = dyn_cast_or_null<StoreInst>(UseInst)) {
+      if (SI->isVolatile())
+        return false;
+
       // Reject if the stored value is not the pointer operand.
       if (SI->getPointerOperand() != Val)
         return false;
+    } else if (LoadInst *LI = dyn_cast_or_null<LoadInst>(UseInst)) {
+      if (LI->isVolatile())
+        return false;
+    } else if (AtomicRMWInst *RMW = dyn_cast_or_null<AtomicRMWInst>(UseInst)) {
+      if (RMW->isVolatile())
+        return false;
+    } else if (AtomicCmpXchgInst *CAS
+               = dyn_cast_or_null<AtomicCmpXchgInst>(UseInst)) {
+      if (CAS->isVolatile())
+        return false;
     }
 
     if (!User->getType()->isPointerTy())
Index: test/CodeGen/AMDGPU/promote-alloca-volatile.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/promote-alloca-volatile.ll
@@ -0,0 +1,26 @@
+; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -amdgpu-promote-alloca < %s | FileCheck %s
+
+; CHECK-LABEL: @volatile_load(
+; CHECK: alloca [5 x i32]
+; CHECK load volatile i32, i32*
+define void @volatile_load(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
+entry:
+  %stack = alloca [5 x i32], align 4
+  %tmp = load i32, i32 addrspace(1)* %in, align 4
+  %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %tmp
+  %load = load volatile i32, i32* %arrayidx1
+  store i32 %load, i32 addrspace(1)* %out
+ ret void
+}
+
+; CHECK-LABEL: @volatile_store(
+; CHECK: alloca [5 x i32]
+; CHECK store volatile i32 %tmp, i32*
+define void @volatile_store(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
+entry:
+  %stack = alloca [5 x i32], align 4
+  %tmp = load i32, i32 addrspace(1)* %in, align 4
+  %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %tmp
+  store volatile i32 %tmp, i32* %arrayidx1
+ ret void
+}