Index: lib/Target/AMDGPU/AMDGPUCallingConv.td
===================================================================
--- lib/Target/AMDGPU/AMDGPUCallingConv.td
+++ lib/Target/AMDGPU/AMDGPUCallingConv.td
@@ -110,7 +110,7 @@
 
 // Calling convention for compute kernels
 def CC_AMDGPU_Kernel : CallingConv<[
-  CCCustom<"allocateStack">
+  CCCustom<"allocateKernArg">
 ]>;
 
 def CC_AMDGPU : CallingConv<[
Index: lib/Target/AMDGPU/AMDGPUISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -31,13 +31,15 @@
 #include "SIInstrInfo.h"
 using namespace llvm;
 
-static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
-                      CCValAssign::LocInfo LocInfo,
-                      ISD::ArgFlagsTy ArgFlags, CCState &State) {
-  unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
-                                        ArgFlags.getOrigAlign());
-  State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
-
+static bool allocateKernArg(unsigned ValNo, MVT ValVT, MVT LocVT,
+                            CCValAssign::LocInfo LocInfo,
+                            ISD::ArgFlagsTy ArgFlags, CCState &State) {
+  MachineFunction &MF = State.getMachineFunction();
+  AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
+
+  uint64_t Offset = MFI->allocateKernArg(ValVT.getStoreSize(),
+                                         ArgFlags.getOrigAlign());
+  State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT, LocInfo));
   return true;
 }
 
Index: lib/Target/AMDGPU/AMDGPUMachineFunction.h
===================================================================
--- lib/Target/AMDGPU/AMDGPUMachineFunction.h
+++ lib/Target/AMDGPU/AMDGPUMachineFunction.h
@@ -16,10 +16,25 @@
 namespace llvm {
 
 class AMDGPUMachineFunction : public MachineFunctionInfo {
+  uint64_t KernArgSize;
+  unsigned MaxKernArgAlign;
+
   virtual void anchor();
 
 public:
   AMDGPUMachineFunction(const MachineFunction &MF);
+
+  uint64_t allocateKernArg(uint64_t Size, unsigned Align) {
+    assert(isPowerOf2_32(Align));
+    KernArgSize = alignTo(KernArgSize, Align);
+
+    uint64_t Result = KernArgSize;
+    KernArgSize += Size;
+
+    MaxKernArgAlign = std::max(Align, MaxKernArgAlign);
+    return Result;
+  }
+
   /// A map to keep track of local memory objects and their offsets within
   /// the local memory space.
   std::map<const GlobalValue *, unsigned> LocalMemoryObjects;
Index: lib/Target/AMDGPU/AMDGPUMachineFunction.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUMachineFunction.cpp
+++ lib/Target/AMDGPU/AMDGPUMachineFunction.cpp
@@ -1,8 +1,5 @@
 #include "AMDGPUMachineFunction.h"
-#include "AMDGPU.h"
-#include "Utils/AMDGPUBaseInfo.h"
-#include "llvm/IR/Attributes.h"
-#include "llvm/IR/Function.h"
+
 using namespace llvm;
 
 // Pin the vtable to this file.
@@ -10,8 +7,9 @@
 
 AMDGPUMachineFunction::AMDGPUMachineFunction(const MachineFunction &MF) :
   MachineFunctionInfo(),
+  KernArgSize(0),
+  MaxKernArgAlign(0),
   LDSSize(0),
   ABIArgOffset(0),
   ScratchSize(0),
-  IsKernel(true) {
-}
+  IsKernel(true) {}
Index: test/CodeGen/AMDGPU/kernarg-stack-alignment.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/kernarg-stack-alignment.ll
@@ -0,0 +1,44 @@
+; RUN: llc -O0 -march=amdgcn -verify-machineinstrs < %s | FileCheck %s
+
+; Test that the alignment of kernel arguments does not impact the
+; alignment of the stack
+
+; CHECK-LABEL: {{^}}no_args:
+; CHECK: ScratchSize: 8{{$}}
+define void @no_args() {
+  %alloca = alloca i8
+  store volatile i8 0, i8* %alloca
+  ret void
+}
+
+; CHECK-LABEL: {{^}}force_align32:
+; CHECK: ScratchSize: 8{{$}}
+define void @force_align32(<8 x i32>) {
+  %alloca = alloca i8
+  store volatile i8 0, i8* %alloca
+  ret void
+}
+
+; CHECK-LABEL: {{^}}force_align64:
+; CHECK: ScratchSize: 8{{$}}
+define void @force_align64(<16 x i32>) {
+  %alloca = alloca i8
+  store volatile i8 0, i8* %alloca
+  ret void
+}
+
+; CHECK-LABEL: {{^}}force_align128:
+; CHECK: ScratchSize: 8{{$}}
+define void @force_align128(<32 x i32>) {
+  %alloca = alloca i8
+  store volatile i8 0, i8* %alloca
+  ret void
+}
+
+; CHECK-LABEL: {{^}}force_align256:
+; CHECK: ScratchSize: 8{{$}}
+define void @force_align256(<64 x i32>) {
+  %alloca = alloca i8
+  store volatile i8 0, i8* %alloca
+  ret void
+}