diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index ccb0072..733cff7 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -2518,12 +2518,17 @@
       int32_t Offset = LocMemOffset;
 
       SDValue PtrOff = DAG.getConstant(Offset, DL, PtrVT);
+      unsigned Align = 0;
 
       if (IsTailCall) {
         ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
         unsigned OpSize = Flags.isByVal() ?
           Flags.getByValSize() : VA.getValVT().getStoreSize();
 
+        // FIXME: We can have better than the minimum byval required alignment.
+        Align = Flags.isByVal() ? Flags.getByValAlign() :
+          MinAlign(Subtarget->getStackAlignment(), Offset);
+
         Offset = Offset + FPDiff;
         int FI = MFI.CreateFixedObject(OpSize, Offset, true);
 
@@ -2541,6 +2546,7 @@
       } else {
         DstAddr = PtrOff;
         DstInfo = MachinePointerInfo::getStack(MF, LocMemOffset);
+        Align = MinAlign(Subtarget->getStackAlignment(), LocMemOffset);
       }
 
       if (Outs[i].Flags.isByVal()) {
@@ -2555,7 +2561,7 @@
 
         MemOpChains.push_back(Cpy);
       } else {
-        SDValue Store = DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo);
+        SDValue Store = DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, Align);
         MemOpChains.push_back(Store);
       }
     }