R600: Fix handling of vector kernel arguments

The SelectionDAGBuilder was promoting vector kernel arguments to legal
types, but this won't work for R600 and SI since kernel arguments are
stored in memory and can't be promoted.  In order to handle vector
arguments correctly we need to look at the original types from the LLVM IR
function.

10 files changed, 135 insertions, 109 deletions

diff --git a/lib/Target/R600/AMDGPUCallingConv.td b/lib/Target/R600/AMDGPUCallingConv.td
index a194e6d3c8b..3535e35f4fb 100644
--- a/lib/Target/R600/AMDGPUCallingConv.td
+++ b/lib/Target/R600/AMDGPUCallingConv.td
@@ -44,11 +44,7 @@ def CC_SI : CallingConv<[
 
 // Calling convention for compute kernels
 def CC_AMDGPU_Kernel : CallingConv<[
-  CCIfType<[v4i32, v4f32],               CCAssignToStack <16, 16>>,
-  CCIfType<[i64, f64, v2f32, v2i32],     CCAssignToStack < 8, 8>>,
-  CCIfType<[i32, f32],                   CCAssignToStack < 4, 4>>,
-  CCIfType<[i16],                        CCAssignToStack < 2, 4>>,
-  CCIfType<[i8],                         CCAssignToStack < 1, 4>>
+  CCCustom<"allocateStack">
 ]>;
 
 def CC_AMDGPU : CallingConv<[
diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index f8f0596c5a0..dfaabd7fd58 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -29,6 +29,14 @@
 #include "llvm/IR/DataLayout.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.getSizeInBits() / 8, ArgFlags.getOrigAlign());
+    State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
+
+  return true;
+}
 
 #include "AMDGPUGenCallingConv.inc"
 
@@ -208,6 +216,10 @@ SDValue AMDGPUTargetLowering::LowerReturn(
   return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
 }
 
+bool AMDGPUTargetLowering::hasArgumentsInMemory(const Function &F) const {
+  return true;
+}
+
 //===---------------------------------------------------------------------===//
 // Target specific lowering
 //===---------------------------------------------------------------------===//
@@ -673,6 +685,38 @@ SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
 // Helper functions
 //===----------------------------------------------------------------------===//
 
+void AMDGPUTargetLowering::getOriginalFunctionArgs(
+                               SelectionDAG &DAG,
+                               const Function *F,
+                               const SmallVectorImpl<ISD::InputArg> &Ins,
+                               SmallVectorImpl<ISD::InputArg> &OrigIns) const {
+
+  for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
+    if (Ins[i].ArgVT == Ins[i].VT) {
+      OrigIns.push_back(Ins[i]);
+      continue;
+    }
+
+    EVT VT;
+    if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
+      // Vector has been split into scalars.
+      VT = Ins[i].ArgVT.getVectorElementType();
+    } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
+               Ins[i].ArgVT.getVectorElementType() !=
+               Ins[i].VT.getVectorElementType()) {
+      // Vector elements have been promoted
+      VT = Ins[i].ArgVT;
+    } else {
+      // Vector has been spilt into smaller vectors.
+      VT = Ins[i].VT;
+    }
+
+    ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
+                      Ins[i].OrigArgIndex, Ins[i].PartOffset);
+    OrigIns.push_back(Arg);
+  }
+}
+
 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
   if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
     return CFP->isExactlyValue(1.0);
diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
index 43f6389fac7..845f3fe018f 100644
--- a/lib/Target/R600/AMDGPUISelLowering.h
+++ b/lib/Target/R600/AMDGPUISelLowering.h
@@ -37,7 +37,6 @@ private:
   SDValue MergeVectorStore(const SDValue &Op, SelectionDAG &DAG) const;
   /// \brief Split a vector store into multiple scalar stores.
   /// \returns The resulting chain. 
-  SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
 
 protected:
@@ -53,10 +52,21 @@ protected:
                              SelectionDAG &DAG) const;
   /// \brief Split a vector load into multiple scalar loads.
   SDValue SplitVectorLoad(const SDValue &Op, SelectionDAG &DAG) const;
+  SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
   bool isHWTrueValue(SDValue Op) const;
   bool isHWFalseValue(SDValue Op) const;
 
+  /// The SelectionDAGBuilder will automatically promote function arguments
+  /// with illegal types.  However, this does not work for the AMDGPU targets
+  /// since the function arguments are stored in memory as these illegal types.
+  /// In order to handle this properly we need to get the origianl types sizes
+  /// from the LLVM IR Function and fixup the ISD:InputArg values before
+  /// passing them to AnalyzeFormalArguments()
+  void getOriginalFunctionArgs(SelectionDAG &DAG,
+                               const Function *F,
+                               const SmallVectorImpl<ISD::InputArg> &Ins,
+                               SmallVectorImpl<ISD::InputArg> &OrigIns) const;
   void AnalyzeFormalArguments(CCState &State,
                               const SmallVectorImpl<ISD::InputArg> &Ins) const;
 
@@ -71,6 +81,7 @@ public:
                               const SmallVectorImpl<ISD::OutputArg> &Outs,
                               const SmallVectorImpl<SDValue> &OutVals,
                               SDLoc DL, SelectionDAG &DAG) const;
+  bool hasArgumentsInMemory(const Function &F) const;
   virtual SDValue LowerCall(CallLoweringInfo &CLI,
                             SmallVectorImpl<SDValue> &InVals) const {
     CLI.Callee.dump();
diff --git a/lib/Target/R600/AMDGPUTargetMachine.cpp b/lib/Target/R600/AMDGPUTargetMachine.cpp
index 9722e7dddf0..b19277d97be 100644
--- a/lib/Target/R600/AMDGPUTargetMachine.cpp
+++ b/lib/Target/R600/AMDGPUTargetMachine.cpp
@@ -59,8 +59,9 @@ AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, StringRef TT,
   LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
   Subtarget(TT, CPU, FS),
   Layout(Subtarget.getDataLayout()),
-  FrameLowering(TargetFrameLowering::StackGrowsUp, 16 // Stack Alignment
-                                                 , 0),
+  FrameLowering(TargetFrameLowering::StackGrowsUp,
+                64 * 16 // Maximum stack alignment (long16)
+               , 0),
   IntrinsicInfo(this),
   InstrItins(&Subtarget.getInstrItineraryData()) {
   // TLInfo uses InstrInfo so it must be initialized after.
diff --git a/lib/Target/R600/R600ISelLowering.cpp b/lib/Target/R600/R600ISelLowering.cpp
index 1765261acde..b46a2834f29 100644
--- a/lib/Target/R600/R600ISelLowering.cpp
+++ b/lib/Target/R600/R600ISelLowering.cpp
@@ -1194,7 +1194,7 @@ SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
   }
 
   int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
-  if (ConstantBlock > -1) {
+  if (ConstantBlock > -1 && LoadNode->getExtensionType() != ISD::SEXTLOAD) {
     SDValue Result;
     if (dyn_cast<ConstantExpr>(LoadNode->getSrcValue()) ||
         dyn_cast<Constant>(LoadNode->getSrcValue()) ||
@@ -1325,22 +1325,29 @@ SDValue R600TargetLowering::LowerFormalArguments(
   CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
                  getTargetMachine(), ArgLocs, *DAG.getContext());
 
-  AnalyzeFormalArguments(CCInfo, Ins);
+  SmallVector<ISD::InputArg, 8> LocalIns;
+
+  getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
+                          LocalIns);
+
+  AnalyzeFormalArguments(CCInfo, LocalIns);
 
   for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
     CCValAssign &VA = ArgLocs[i];
-    EVT VT = VA.getLocVT();
+    EVT VT = Ins[i].VT;
+    EVT MemVT = LocalIns[i].VT;
 
     PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
                                                    AMDGPUAS::CONSTANT_BUFFER_0);
 
     // The first 36 bytes of the input buffer contains information about
     // thread group and global sizes.
-    SDValue Arg = DAG.getLoad(VT, DL, Chain,
-                           DAG.getConstant(36 + VA.getLocMemOffset(), MVT::i32),
-                           MachinePointerInfo(UndefValue::get(PtrTy)), false,
-                           false, false, 4); // 4 is the prefered alignment for
-                                             // the CONSTANT memory space.
+    SDValue Arg = DAG.getExtLoad(ISD::SEXTLOAD, DL, VT, Chain,
+                                 DAG.getConstant(36 + VA.getLocMemOffset(), MVT::i32),
+                                 MachinePointerInfo(UndefValue::get(PtrTy)),
+                                 MemVT, false, false, 4);
+                                 // 4 is the prefered alignment for
+                                 // the CONSTANT memory space.
     InVals.push_back(Arg);
   }
   return Chain;
diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 2e7e3a49c89..ad88cae8a5a 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -75,13 +75,16 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::LOAD, MVT::i64, Custom);
   setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
+  setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
+  setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
 
   setOperationAction(ISD::STORE, MVT::i32, Custom);
   setOperationAction(ISD::STORE, MVT::i64, Custom);
   setOperationAction(ISD::STORE, MVT::i128, Custom);
   setOperationAction(ISD::STORE, MVT::v2i32, Custom);
   setOperationAction(ISD::STORE, MVT::v4i32, Custom);
-
+  setOperationAction(ISD::STORE, MVT::v8i32, Custom);
+  setOperationAction(ISD::STORE, MVT::v16i32, Custom);
 
   setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
   setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
@@ -91,6 +94,7 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
   setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
 
+  setOperationAction(ISD::ANY_EXTEND, MVT::i64, Custom);
   setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
   setOperationAction(ISD::ZERO_EXTEND, MVT::i64, Custom);
 
@@ -103,10 +107,15 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
 
   setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
   setLoadExtAction(ISD::EXTLOAD, MVT::i32, Expand);
+  setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, Expand);
+  setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, Expand);
 
   setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
   setTruncStoreAction(MVT::i64, MVT::i32, Expand);
+  setTruncStoreAction(MVT::i128, MVT::i64, Expand);
+  setTruncStoreAction(MVT::v8i32, MVT::v8i16, Expand);
+  setTruncStoreAction(MVT::v16i32, MVT::v16i16, Expand);
 
   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
   setOperationAction(ISD::FrameIndex, MVT::i64, Custom);
@@ -132,23 +141,22 @@ bool SITargetLowering::allowsUnalignedMemoryAccesses(EVT  VT,
 }
 
 bool SITargetLowering::shouldSplitVectorElementType(EVT VT) const {
-  return VT.bitsLE(MVT::i8);
+  return VT.bitsLE(MVT::i16);
 }
 
-SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT,
+SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
                                          SDLoc DL, SDValue Chain,
                                          unsigned Offset) const {
   MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
   PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
                                             AMDGPUAS::CONSTANT_ADDRESS);
-  EVT ArgVT = MVT::getIntegerVT(VT.getSizeInBits());
   SDValue BasePtr =  DAG.getCopyFromReg(Chain, DL,
                            MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
   SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
                                              DAG.getConstant(Offset, MVT::i64));
-  return DAG.getLoad(VT, DL, Chain, Ptr,
-                            MachinePointerInfo(UndefValue::get(PtrTy)),
-                            false, false, false, ArgVT.getSizeInBits() >> 3);
+  return DAG.getExtLoad(ISD::SEXTLOAD, DL, VT, Chain, Ptr,
+                            MachinePointerInfo(UndefValue::get(PtrTy)), MemVT,
+                            false, false, MemVT.getSizeInBits() >> 3);
 
 }
 
@@ -207,7 +215,7 @@ SDValue SITargetLowering::LowerFormalArguments(
         NewArg.PartOffset += NewArg.VT.getStoreSize();
       }
 
-    } else {
+    } else if (Info->ShaderType != ShaderType::COMPUTE) {
       Splits.push_back(Arg);
     }
   }
@@ -230,6 +238,11 @@ SDValue SITargetLowering::LowerFormalArguments(
     MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
   }
 
+  if (Info->ShaderType == ShaderType::COMPUTE) {
+    getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
+                            Splits);
+  }
+
   AnalyzeFormalArguments(CCInfo, Splits);
 
   for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
@@ -241,12 +254,13 @@ SDValue SITargetLowering::LowerFormalArguments(
     }
 
     CCValAssign &VA = ArgLocs[ArgIdx++];
-    EVT VT = VA.getLocVT();
+    EVT VT = Ins[i].VT;
+    EVT MemVT = Splits[i].VT;
 
     if (VA.isMemLoc()) {
       // The first 36 bytes of the input buffer contains information about
       // thread group and global sizes.
-      SDValue Arg = LowerParameter(DAG, VT, DL, DAG.getRoot(),
+      SDValue Arg = LowerParameter(DAG, VT, MemVT,  DL, DAG.getRoot(),
                                    36 + VA.getLocMemOffset());
       InVals.push_back(Arg);
       continue;
@@ -408,9 +422,10 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::BRCOND: return LowerBRCOND(Op, DAG);
   case ISD::LOAD: {
     LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
-    if ((Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
-         Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
-        Op.getValueType().isVector()) {
+    if (Op.getValueType().isVector() &&
+        (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
+         Load->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
+         Load->getMemoryVT().getVectorNumElements() >= 8)) {
       SDValue MergedValues[2] = {
         SplitVectorLoad(Op, DAG),
         Load->getChain()
@@ -423,6 +438,7 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
   case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
   case ISD::STORE: return LowerSTORE(Op, DAG);
+  case ISD::ANY_EXTEND: // Fall-through
   case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
   case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
   case ISD::INTRINSIC_WO_CHAIN: {
@@ -435,23 +451,23 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
     switch (IntrinsicID) {
     default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
     case Intrinsic::r600_read_ngroups_x:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 0);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0);
     case Intrinsic::r600_read_ngroups_y:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 4);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4);
     case Intrinsic::r600_read_ngroups_z:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 8);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8);
     case Intrinsic::r600_read_global_size_x:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 12);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12);
     case Intrinsic::r600_read_global_size_y:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 16);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16);
     case Intrinsic::r600_read_global_size_z:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 20);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20);
     case Intrinsic::r600_read_local_size_x:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 24);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24);
     case Intrinsic::r600_read_local_size_y:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 28);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28);
     case Intrinsic::r600_read_local_size_z:
-      return LowerParameter(DAG, VT, DL, DAG.getEntryNode(), 32);
+      return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32);
     case Intrinsic::r600_read_tgid_x:
       return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
                      AMDGPU::SReg_32RegClass.getRegister(NumUserSGPRs + 0), VT);
@@ -722,6 +738,12 @@ SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
   if (Ret.getNode())
     return Ret;
 
+  if (VT.isVector() && VT.getVectorNumElements() >= 8) {
+    Ret = SplitVectorStore(Op, DAG);
+    if (Ret.getNode())
+      return Ret;
+  }
+
   if (Store->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
     return SDValue();
 
diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
index ecfea15e612..384caf4bc93 100644
--- a/lib/Target/R600/SIISelLowering.h
+++ b/lib/Target/R600/SIISelLowering.h
@@ -21,7 +21,7 @@
 namespace llvm {
 
 class SITargetLowering : public AMDGPUTargetLowering {
-  SDValue LowerParameter(SelectionDAG &DAG, EVT VT, SDLoc DL,
+  SDValue LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT, SDLoc DL,
                          SDValue Chain, unsigned Offset) const;
   SDValue LowerSampleIntrinsic(unsigned Opcode, const SDValue &Op,
                                SelectionDAG &DAG) const;
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 460c1ce0b09..2d0a11e4af5 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1614,9 +1614,12 @@ def : BitConvert <f64, i64, VReg_64>;
 
 def : BitConvert <v2f32, v2i32, VReg_64>;
 def : BitConvert <v2i32, v2f32, VReg_64>;
+def : BitConvert <v2i32, i64, VReg_64>;
 
 def : BitConvert <v4f32, v4i32, VReg_128>;
 def : BitConvert <v4i32, v4f32, VReg_128>;
+def : BitConvert <v4i32, i128,  VReg_128>;
+def : BitConvert <i128, v4i32,  VReg_128>;
 
 def : BitConvert <v8i32, v32i8, SReg_256>;
 def : BitConvert <v32i8, v8i32, SReg_256>;
@@ -2034,6 +2037,17 @@ def : Pat <
     (V_OR_B32_e32 (EXTRACT_SUBREG $a, sub1), (EXTRACT_SUBREG $b, sub1)), sub1)
 >;
 
+//===----------------------------------------------------------------------===//
+// Miscellaneous Pattens
+//===----------------------------------------------------------------------===//
+
+def : Pat <
+  (i64 (trunc i128:$x)),
+  (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
+    (i32 (EXTRACT_SUBREG $x, sub0)), sub0),
+    (i32 (EXTRACT_SUBREG $x, sub1)), sub1)
+>;
+
 //============================================================================//
 // Miscellaneous Optimization Patterns
 //============================================================================//
diff --git a/lib/Target/R600/SIRegisterInfo.td b/lib/Target/R600/SIRegisterInfo.td
index 2d7bff076dc..e40f34b2c6e 100644
--- a/lib/Target/R600/SIRegisterInfo.td
+++ b/lib/Target/R600/SIRegisterInfo.td
@@ -174,7 +174,7 @@ def VReg_96 : RegisterClass<"AMDGPU", [untyped], 96, (add VGPR_96)> {
   let Size = 96;
 }
 
-def VReg_128 : RegisterClass<"AMDGPU", [v4i32, v4f32], 128, (add VGPR_128)>;
+def VReg_128 : RegisterClass<"AMDGPU", [v4i32, v4f32, i128], 128, (add VGPR_128)>;
 
 def VReg_256 : RegisterClass<"AMDGPU", [v32i8, v8i32, v8f32], 256, (add VGPR_256)>;
 
diff --git a/test/CodeGen/R600/short-args.ll b/test/CodeGen/R600/short-args.ll
deleted file mode 100644
index 8882978d750..00000000000
--- a/test/CodeGen/R600/short-args.ll
+++ /dev/null
@@ -1,69 +0,0 @@
-; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s --check-prefix=EG-CHECK
-; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck %s --check-prefix=EG-CHECK
-; RUN: llc < %s -march=r600 -mcpu=SI -verify-machineinstrs | FileCheck %s --check-prefix=SI-CHECK
-
-; EG-CHECK: @i8_arg
-; EG-CHECK: MOV {{[ *]*}}T{{[0-9]+\.[XYZW]}}, KC0[2].Z
-; SI-CHECK: BUFFER_LOAD_UBYTE
-
-define void @i8_arg(i32 addrspace(1)* nocapture %out, i8 %in) nounwind {
-entry:
-  %0 = zext i8 %in to i32
-  store i32 %0, i32 addrspace(1)* %out, align 4
-  ret void
-}
-
-; EG-CHECK: @i8_zext_arg
-; EG-CHECK: MOV {{[ *]*}}T{{[0-9]+\.[XYZW]}}, KC0[2].Z
-; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]}}, SGPR0_SGPR1, 11
-
-define void @i8_zext_arg(i32 addrspace(1)* nocapture %out, i8 zeroext %in) nounwind {
-entry:
-  %0 = zext i8 %in to i32
-  store i32 %0, i32 addrspace(1)* %out, align 4
-  ret void
-}
-
-; EG-CHECK: @i8_sext_arg
-; EG-CHECK: MOV {{[ *]*}}T{{[0-9]+\.[XYZW]}}, KC0[2].Z
-; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]}}, SGPR0_SGPR1, 11
-
-define void @i8_sext_arg(i32 addrspace(1)* nocapture %out, i8 signext %in) nounwind {
-entry:
-  %0 = sext i8 %in to i32
-  store i32 %0, i32 addrspace(1)* %out, align 4
-  ret void
-}
-
-; EG-CHECK: @i16_arg
-; EG-CHECK: MOV {{[ *]*}}T{{[0-9]+\.[XYZW]}}, KC0[2].Z
-; SI-CHECK: BUFFER_LOAD_USHORT
-
-define void @i16_arg(i32 addrspace(1)* nocapture %out, i16 %in) nounwind {
-entry:
-  %0 = zext i16 %in to i32
-  store i32 %0, i32 addrspace(1)* %out, align 4
-  ret void
-}
-
-; EG-CHECK: @i16_zext_arg
-; EG-CHECK: MOV {{[ *]*}}T{{[0-9]+\.[XYZW]}}, KC0[2].Z
-; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]}}, SGPR0_SGPR1, 11
-
-define void @i16_zext_arg(i32 addrspace(1)* nocapture %out, i16 zeroext %in) nounwind {
-entry:
-  %0 = zext i16 %in to i32
-  store i32 %0, i32 addrspace(1)* %out, align 4
-  ret void
-}
-
-; EG-CHECK: @i16_sext_arg
-; EG-CHECK: MOV {{[ *]*}}T{{[0-9]+\.[XYZW]}}, KC0[2].Z
-; SI-CHECK: S_LOAD_DWORD SGPR{{[0-9]}}, SGPR0_SGPR1, 11
-
-define void @i16_sext_arg(i32 addrspace(1)* nocapture %out, i16 signext %in) nounwind {
-entry:
-  %0 = sext i16 %in to i32
-  store i32 %0, i32 addrspace(1)* %out, align 4
-  ret void
-}