Index: lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -397,13 +397,18 @@
 // instructions.
 static bool canVectorizeInst(Instruction *Inst, User *User) {
   switch (Inst->getOpcode()) {
-  case Instruction::Load:
+  case Instruction::Load: {
+    LoadInst *LI = cast<LoadInst>(Inst);
+    return !LI->isVolatile();
+  }
   case Instruction::BitCast:
   case Instruction::AddrSpaceCast:
     return true;
   case Instruction::Store: {
     // Must be the stored pointer operand, not a stored value.
     StoreInst *SI = cast<StoreInst>(Inst);
+    if (SI->isVolatile())
+      return false;
     return SI->getPointerOperand() == User;
   }
   default:
Index: test/CodeGen/AMDGPU/promote-alloca-volatile.ll
===================================================================
--- test/CodeGen/AMDGPU/promote-alloca-volatile.ll
+++ test/CodeGen/AMDGPU/promote-alloca-volatile.ll
@@ -1,26 +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: alloca [4 x i32]
 ; CHECK: load volatile i32, i32*
 define amdgpu_kernel void @volatile_load(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
 entry:
-  %stack = alloca [5 x i32], align 4
+  %stack = alloca [4 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
+  %arrayidx1 = getelementptr inbounds [4 x i32], [4 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: alloca [4 x i32]
 ; CHECK: store volatile i32 %tmp, i32*
 define amdgpu_kernel void @volatile_store(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
 entry:
-  %stack = alloca [5 x i32], align 4
+  %stack = alloca [4 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
+  %arrayidx1 = getelementptr inbounds [4 x i32], [4 x i32]* %stack, i32 0, i32 %tmp
   store volatile i32 %tmp, i32* %arrayidx1
  ret void
 }