Index: lib/Target/R600/R600ISelLowering.cpp
===================================================================
--- lib/Target/R600/R600ISelLowering.cpp
+++ lib/Target/R600/R600ISelLowering.cpp
@@ -1400,8 +1400,9 @@
 
   for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
     CCValAssign &VA = ArgLocs[i];
-    EVT VT = Ins[i].VT;
-    EVT MemVT = LocalIns[i].VT;
+    const ISD::InputArg &In = Ins[i];
+    EVT VT = In.VT;
+    EVT MemVT = In.VT;
 
     if (ShaderType != ShaderType::COMPUTE) {
       unsigned Reg = MF.addLiveIn(VA.getLocReg(), &AMDGPU::R600_Reg128RegClass);
@@ -1423,11 +1424,20 @@
 
     // FIXME: This should really check the extload type, but the handling of
     // extload vecto parameters seems to be broken.
-    //ISD::LoadExtType Ext = Ins[i].Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
+    //ISD::LoadExtType Ext = In.Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
     ISD::LoadExtType Ext = ISD::SEXTLOAD;
+
+    // Compute the offset from the value.
+    // XXX - I think PartOffset should give you this, but it seems to give the
+    // size of the register which isn't useful.
+
+    unsigned ValBase = ArgLocs[In.OrigArgIndex].getLocMemOffset();
+    unsigned PartOffset = VA.getLocMemOffset();
+
+    MachinePointerInfo PtrInfo(UndefValue::get(PtrTy), PartOffset - ValBase);
     SDValue Arg = DAG.getExtLoad(Ext, DL, VT, Chain,
-                                 DAG.getConstant(36 + VA.getLocMemOffset(), MVT::i32),
-                                 MachinePointerInfo(UndefValue::get(PtrTy)),
+                                 DAG.getConstant(36 + PartOffset, MVT::i32),
+                                 PtrInfo,
                                  MemVT, false, false, 4);
 
     // 4 is the preferred alignment for the CONSTANT memory space.
Index: test/CodeGen/R600/kernel-args.ll
===================================================================
--- test/CodeGen/R600/kernel-args.ll
+++ test/CodeGen/R600/kernel-args.ll
@@ -453,3 +453,21 @@
   store <16 x float> %in, <16 x float> addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: @kernel_arg_i64
+; SI: S_LOAD_DWORDX2
+; SI: S_LOAD_DWORDX2
+; SI: BUFFER_STORE_DWORDX2
+define void @kernel_arg_i64(i64 addrspace(1)* %out, i64 %a) nounwind {
+  store i64 %a, i64 addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @kernel_arg_v1i64
+; SI: S_LOAD_DWORDX2
+; SI: S_LOAD_DWORDX2
+; SI: BUFFER_STORE_DWORDX2
+define void @kernel_arg_v1i64(<1 x i64> addrspace(1)* %out, <1 x i64> %a) nounwind {
+  store <1 x i64> %a, <1 x i64> addrspace(1)* %out, align 8
+  ret void
+}
Index: test/CodeGen/R600/v1i64-kernel-arg.ll
===================================================================
--- test/CodeGen/R600/v1i64-kernel-arg.ll
+++ /dev/null
@@ -1,17 +0,0 @@
-; REQUIRES: asserts
-; XFAIL: *
-; RUN: llc -march=r600 -mcpu=cypress < %s | FileCheck %s
-
-; CHECK-LABEL: @kernel_arg_i64
-define void @kernel_arg_i64(i64 addrspace(1)* %out, i64 %a) nounwind {
-  store i64 %a, i64 addrspace(1)* %out, align 8
-  ret void
-}
-
-; i64 arg works, v1i64 arg does not.
-; CHECK-LABEL: @kernel_arg_v1i64
-define void @kernel_arg_v1i64(<1 x i64> addrspace(1)* %out, <1 x i64> %a) nounwind {
-  store <1 x i64> %a, <1 x i64> addrspace(1)* %out, align 8
-  ret void
-}
-