[AMDGPU] Support for v3i32/v3f32

[tpr]

Summary

Added support for dwordx3 for most load/store types, but not DS, and not
intrinsics yet.

SI (gfx6) does not have dwordx3 instructions, so they are not enabled
there.

Some of this patch is from Matt Arsenault, also of AMD.

Change-Id: I913ef54f1433a7149da8d72f4af54dbb13436bd9

[arsenm]

Needs a test for the 96-bit spills. An -O0 test case with a value live across a block should work

lib/Target/AMDGPU/SIISelLowering.cpp

6645 ↗

(On Diff #189144)

I think the widening case is missing an alignment check, but that's a pre-existing problem that should be fixed separately

lib/Target/AMDGPU/SIRegisterInfo.td

471 ↗

(On Diff #189144)

Yes, it needs to be allocatable.

The allocation priority of the SGPRs is supposed to be lower than VGPRs, but here it's the same as the VGPR. You might need to shift all of the values around to fix this

[rampitec]

What about flat and segmented memory operations?

lib/Target/AMDGPU/SIInstrInfo.cpp

577 ↗

(On Diff #189144)

Does that mean for a v3 we will have 3 s_mov_b32 instead of s_mov_b64 + s_mov_b32? That is suboptimal.

[rampitec]

What happens in asm parser/disasm? I thing these still create bogus v4 operands?

[arsenm]

In D58902#1417139, @rampitec wrote:

What happens in asm parser/disasm? I thing these still create bogus v4 operands?

Those should already be using the correct register classes. There were never issues after the DAG with these

[tpr]

What about flat and segmented memory operations?

Flat dwordx3 is supported, as shown by some test changes. What is segmented memory?

[arsenm]

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

1430 ↗

(On Diff #189144)

You can just hardcode i32

lib/Target/AMDGPU/SIISelLowering.cpp

334–335 ↗

(On Diff #189144)

Why do these need to be changed?

1332 ↗

(On Diff #189144)

In the backend you can just hardened the index type to i32

[rampitec]

In D58902#1418275, @tpr wrote:

What about flat and segmented memory operations?

Flat dwordx3 is supported, as shown by some test changes. What is segmented memory?

Segmented flat, like global_load.

[tpr]

lib/Target/AMDGPU/SIInstrInfo.cpp

577 ↗

(On Diff #189144)

Yes.

I can't find any lit test or graphics shader that attempts to copy a vec3 of sgprs here. Any idea how to provoke it so I can try it?

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

577 ↗

(On Diff #189144)

A v3 phi, and probably at -O0

[tpr]

In D58902#1418608, @rampitec wrote:

In D58902#1418275, @tpr wrote:

What about flat and segmented memory operations?

Flat dwordx3 is supported, as shown by some test changes. What is segmented memory?

Segmented flat, like global_load.

Yes, global dwordx3 is supported, as shown in test local-global-i32.ll (although the test does not check that it gets global rather than flat in the case that it should get global).

[tpr]

lib/Target/AMDGPU/SIISelLowering.cpp

334–335 ↗

(On Diff #189144)

Need the first pair to stop "Cannot select: t60: v3i32 = insert_subvector undef:v3i32, t55, Constant:i32<0>" on ds_read2_superreg.ll.

Need the second pair to stop "Cannot select: t129: v4i32 = insert_subvector undef:v4i32, t229, Constant:i32<0>" on setcc.ll.

lib/Target/AMDGPU/SIInstrInfo.cpp

577 ↗

(On Diff #189144)

I managed to provoke it, but fixing it looks less than trivial as we need to somehow allow for a vec3 having a sub0_sub1 and a sub2. And it looks pretty rare anyway. So I won't fix it here. Is that ok?

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

577 ↗

(On Diff #189144)

Yes

[tpr]

Needs a test for the 96-bit spills. An -O0 test case with a value live across a block should work

OK, I've got a change for this I will add in the next update. (Also the same for vec5 that I will add in the next update of that review.)

[tpr]

V2: Fixed broken v3f32 select.

Fixed reg class priorities.
Updated a fragile test.
Addressed review comments on DL::getIdxType.
Fixed vec3 sgpr inline asm constraint.
Added vec3 sgpr spill test.

[tpr]

Actually I only have a vec3 spilling test for sgprs. I didn't see any multi-dword vgpr spill tests.

[arsenm]

In D58902#1420345, @tpr wrote:

Actually I only have a vec3 spilling test for sgprs. I didn't see any multi-dword vgpr spill tests.

That can't be right. We should have some somewhere. We should have a new one at least for the 3 register case

[tpr]

V3: Added multi-dword vgpr spill test, inc vec3 and vec5.

[tpr]

V4: Fixed vec3 in sign_extend test.

[arsenm]

I'm not sure how you're getting away without updating AMDGPUCallingConv.td. We have to maintain a list of what to do for each legal type in each convention

[tpr]

V5: Added vec3 to AMDGPUCallingConvention.td. Fixed call-return-types test for dwordx3 buffer/flat.
(I will add vec5 to AMDGPUCallingConvention.td in the vec5 change.)

[arsenm]

LGTM

[hakzsam]

Hi,
This change breaks SI, at least with RADV.
Here's how to reproduce: ./deqp-vk --deqp-case=dEQP-VK.glsl.builtin.function.integer.bitfieldreverse.ivec3_highp_tess_control

Can you investigate please?
Thanks!

[tpr]

Hi Samuel

Oh dear, sorry about that.

I'm not in a position to immediately run RADV, so I'll try manual code inspection first. Could you please email over ([e-mail]) the .ll going in to the amdgcn backend, and the buggy .s you get out?

Thanks.

[tpr]

Hopefully fixed by D60457.

llvm/trunk/lib/Target/AMDGPU/AMDGPUCallingConv.td

def CC_AMDGPU_Func : CallingConv<[
   CCIfByVal<CCPassByVal<4, 4>>,
   CCIfType<[i1], CCPromoteToType<i32>>,
   CCIfType<[i1, i8, i16], CCIfExtend<CCPromoteToType<i32>>>,
   CCIfType<[i32, f32, i16, f16, v2i16, v2f16, i1], CCAssignToReg<[
     VGPR0, VGPR1, VGPR2, VGPR3, VGPR4, VGPR5, VGPR6, VGPR7,
     VGPR8, VGPR9, VGPR10, VGPR11, VGPR12, VGPR13, VGPR14, VGPR15,
     VGPR16, VGPR17, VGPR18, VGPR19, VGPR20, VGPR21, VGPR22, VGPR23,
     VGPR24, VGPR25, VGPR26, VGPR27, VGPR28, VGPR29, VGPR30, VGPR31]>>,
-  CCIfType<[i64, f64, v2i32, v2f32, v4i32, v4f32, v8i32, v8f32, v16i32, v16f32, v2i64, v2f64, v4i16, v4f16], CCCustom<"allocateVGPRTuple">>,
+  CCIfType<[i64, f64, v2i32, v2f32, v3i32, v3f32, v4i32, v4f32, v8i32, v8f32, v16i32, v16f32, v2i64, v2f64, v4i16, v4f16], CCCustom<"allocateVGPRTuple">>,
   CCIfType<[i32, f32, v2i16, v2f16, i16, f16, i1], CCAssignToStack<4, 4>>,
   CCIfType<[i64, f64, v2i32, v2f32], CCAssignToStack<8, 4>>,
+  CCIfType<[v3i32, v3f32], CCAssignToStack<12, 4>>,
   CCIfType<[v4i32, v4f32, v2i64, v2f64], CCAssignToStack<16, 4>>,
   CCIfType<[v8i32, v8f32], CCAssignToStack<32, 4>>,
   CCIfType<[v16i32, v16f32], CCAssignToStack<64, 4>>
 ]>;
 
 // Calling convention for leaf functions
 def RetCC_AMDGPU_Func : CallingConv<[
   CCIfType<[i1], CCPromoteToType<i32>>,

llvm/trunk/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

 }
 
 static unsigned selectSGPRVectorRegClassID(unsigned NumVectorElts) {
   switch (NumVectorElts) {
   case 1:
     return AMDGPU::SReg_32_XM0RegClassID;
   case 2:
     return AMDGPU::SReg_64RegClassID;
+  case 3:
+    return AMDGPU::SGPR_96RegClassID;
   case 4:
     return AMDGPU::SReg_128RegClassID;
   case 8:
     return AMDGPU::SReg_256RegClassID;
   case 16:
     return AMDGPU::SReg_512RegClassID;
   }
 

llvm/trunk/lib/Target/AMDGPU/AMDGPUISelLowering.h

virtual SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
                                      SelectionDAG &DAG) const;
 
   /// Return 64-bit value Op as two 32-bit integers.
   std::pair<SDValue, SDValue> split64BitValue(SDValue Op,
                                               SelectionDAG &DAG) const;
   SDValue getLoHalf64(SDValue Op, SelectionDAG &DAG) const;
   SDValue getHiHalf64(SDValue Op, SelectionDAG &DAG) const;
 
+  /// Split a vector type into two parts. The first part is a power of two
+  /// vector. The second part is whatever is left over, and is a scalar if it
+  /// would otherwise be a 1-vector.
+  std::pair<EVT, EVT> getSplitDestVTs(const EVT &VT, SelectionDAG &DAG) const;
+
+  /// Split a vector value into two parts of types LoVT and HiVT. HiVT could be
+  /// scalar.
+  std::pair<SDValue, SDValue> splitVector(const SDValue &N, const SDLoc &DL,
+                                          const EVT &LoVT, const EVT &HighVT,
+                                          SelectionDAG &DAG) const;
+
   /// Split a vector load into 2 loads of half the vector.
   SDValue SplitVectorLoad(SDValue Op, SelectionDAG &DAG) const;
 
+  /// Widen a vector load from vec3 to vec4.
+  SDValue WidenVectorLoad(SDValue Op, SelectionDAG &DAG) const;
+
   /// Split a vector store into 2 stores of half the vector.
   SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const;
   void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG,

llvm/trunk/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
   // Lower floating point store/load to integer store/load to reduce the number
   // of patterns in tablegen.
   setOperationAction(ISD::LOAD, MVT::f32, Promote);
   AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
 
   setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
   AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
 
+  setOperationAction(ISD::LOAD, MVT::v3f32, Promote);
+  AddPromotedToType(ISD::LOAD, MVT::v3f32, MVT::v3i32);
+
   setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
   AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
 
   setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
   AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
 
   setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
   AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
   setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
 
   setOperationAction(ISD::STORE, MVT::f32, Promote);
   AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
 
   setOperationAction(ISD::STORE, MVT::v2f32, Promote);
   AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
 
+  setOperationAction(ISD::STORE, MVT::v3f32, Promote);
+  AddPromotedToType(ISD::STORE, MVT::v3f32, MVT::v3i32);
+
   setOperationAction(ISD::STORE, MVT::v4f32, Promote);
   AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
 
   setOperationAction(ISD::STORE, MVT::v8f32, Promote);
   AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
 
   setOperationAction(ISD::STORE, MVT::v16f32, Promote);
   AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
 
   setOperationAction(ISD::FREM, MVT::f32, Custom);
   setOperationAction(ISD::FREM, MVT::f64, Custom);
 
   // Expand to fneg + fadd.
   setOperationAction(ISD::FSUB, MVT::f64, Expand);
 
+  setOperationAction(ISD::CONCAT_VECTORS, MVT::v3i32, Custom);
+  setOperationAction(ISD::CONCAT_VECTORS, MVT::v3f32, Custom);
   setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
   setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
   setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
   setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
+  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3f32, Custom);
+  setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v3i32, Custom);
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
   setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
 
   setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
   setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
   setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::UMAX, MVT::i32, Legal);
 
   setOperationAction(ISD::CTTZ, MVT::i64, Custom);
   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Custom);
   setOperationAction(ISD::CTLZ, MVT::i64, Custom);
   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
 
   static const MVT::SimpleValueType VectorIntTypes[] = {
-    MVT::v2i32, MVT::v4i32
+    MVT::v2i32, MVT::v3i32, MVT::v4i32
   };
 
   for (MVT VT : VectorIntTypes) {
     // Expand the following operations for the current type by default.
     setOperationAction(ISD::ADD,  VT, Expand);
     setOperationAction(ISD::AND,  VT, Expand);
     setOperationAction(ISD::FP_TO_SINT, VT, Expand);
     setOperationAction(ISD::FP_TO_UINT, VT, Expand);
for (MVT VT : VectorIntTypes) {
     setOperationAction(ISD::CTPOP, VT, Expand);
     setOperationAction(ISD::CTTZ, VT, Expand);
     setOperationAction(ISD::CTLZ, VT, Expand);
     setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
     setOperationAction(ISD::SETCC, VT, Expand);
   }
 
   static const MVT::SimpleValueType FloatVectorTypes[] = {
-    MVT::v2f32, MVT::v4f32
+     MVT::v2f32, MVT::v3f32, MVT::v4f32
   };
 
   for (MVT VT : FloatVectorTypes) {
     setOperationAction(ISD::FABS, VT, Expand);
     setOperationAction(ISD::FMINNUM, VT, Expand);
     setOperationAction(ISD::FMAXNUM, VT, Expand);
     setOperationAction(ISD::FADD, VT, Expand);
     setOperationAction(ISD::FCEIL, VT, Expand);
AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
   }
 
   // This causes using an unrolled select operation rather than expansion with
   // bit operations. This is in general better, but the alternative using BFI
   // instructions may be better if the select sources are SGPRs.
   setOperationAction(ISD::SELECT, MVT::v2f32, Promote);
   AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32);
 
+  setOperationAction(ISD::SELECT, MVT::v3f32, Promote);
+  AddPromotedToType(ISD::SELECT, MVT::v3f32, MVT::v3i32);
+
   setOperationAction(ISD::SELECT, MVT::v4f32, Promote);
   AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32);
 
   // There are no libcalls of any kind.
   for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I)
     setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr);
 
   setBooleanContents(ZeroOrNegativeOneBooleanContent);

 SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const {
   SDLoc SL(Op);
 
   SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
   const SDValue One = DAG.getConstant(1, SL, MVT::i32);
   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
 }
 
+// Split a vector type into two parts. The first part is a power of two vector.
+// The second part is whatever is left over, and is a scalar if it would
+// otherwise be a 1-vector.
+std::pair<EVT, EVT>
+AMDGPUTargetLowering::getSplitDestVTs(const EVT &VT, SelectionDAG &DAG) const {
+  EVT LoVT, HiVT;
+  EVT EltVT = VT.getVectorElementType();
+  unsigned NumElts = VT.getVectorNumElements();
+  unsigned LoNumElts = PowerOf2Ceil((NumElts + 1) / 2);
+  LoVT = EVT::getVectorVT(*DAG.getContext(), EltVT, LoNumElts);
+  HiVT = NumElts - LoNumElts == 1
+             ? EltVT
+             : EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts - LoNumElts);
+  return std::make_pair(LoVT, HiVT);
+}
+
+// Split a vector value into two parts of types LoVT and HiVT. HiVT could be
+// scalar.
+std::pair<SDValue, SDValue>
+AMDGPUTargetLowering::splitVector(const SDValue &N, const SDLoc &DL,
+                                  const EVT &LoVT, const EVT &HiVT,
+                                  SelectionDAG &DAG) const {
+  assert(LoVT.getVectorNumElements() +
+                 (HiVT.isVector() ? HiVT.getVectorNumElements() : 1) <=
+             N.getValueType().getVectorNumElements() &&
+         "More vector elements requested than available!");
+  auto IdxTy = getVectorIdxTy(DAG.getDataLayout());
+  SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, N,
+                           DAG.getConstant(0, DL, IdxTy));
+  SDValue Hi = DAG.getNode(
+      HiVT.isVector() ? ISD::EXTRACT_SUBVECTOR : ISD::EXTRACT_VECTOR_ELT, DL,
+      HiVT, N, DAG.getConstant(LoVT.getVectorNumElements(), DL, IdxTy));
+  return std::make_pair(Lo, Hi);
+}
+
 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
                                               SelectionDAG &DAG) const {
   LoadSDNode *Load = cast<LoadSDNode>(Op);
   EVT VT = Op.getValueType();
 
 
   // If this is a 2 element vector, we really want to scalarize and not create
   // weird 1 element vectors.
   if (VT.getVectorNumElements() == 2)
     return scalarizeVectorLoad(Load, DAG);
 
   SDValue BasePtr = Load->getBasePtr();
   EVT MemVT = Load->getMemoryVT();
   SDLoc SL(Op);
 
   const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
 
   EVT LoVT, HiVT;
   EVT LoMemVT, HiMemVT;
   SDValue Lo, Hi;
 
-  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
-  std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
-  std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
+  std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
+  std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
+  std::tie(Lo, Hi) = splitVector(Op, SL, LoVT, HiVT, DAG);
 
   unsigned Size = LoMemVT.getStoreSize();
   unsigned BaseAlign = Load->getAlignment();
   unsigned HiAlign = MinAlign(BaseAlign, Size);
 
   SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
                                   Load->getChain(), BasePtr, SrcValue, LoMemVT,
                                   BaseAlign, Load->getMemOperand()->getFlags());
   SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, Size);
   SDValue HiLoad =
       DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(),
                      HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()),
                      HiMemVT, HiAlign, Load->getMemOperand()->getFlags());
 
-  SDValue Ops[] = {
-    DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
-    DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
-                LoLoad.getValue(1), HiLoad.getValue(1))
-  };
+  auto IdxTy = getVectorIdxTy(DAG.getDataLayout());
+  SDValue Join;
+  if (LoVT == HiVT) {
+    // This is the case that the vector is power of two so was evenly split.
+    Join = DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad);
+  } else {
+    Join = DAG.getNode(ISD::INSERT_SUBVECTOR, SL, VT, DAG.getUNDEF(VT), LoLoad,
+                       DAG.getConstant(0, SL, IdxTy));
+    Join = DAG.getNode(HiVT.isVector() ? ISD::INSERT_SUBVECTOR
+                                       : ISD::INSERT_VECTOR_ELT,
+                       SL, VT, Join, HiLoad,
+                       DAG.getConstant(LoVT.getVectorNumElements(), SL, IdxTy));
+  }
+
+  SDValue Ops[] = {Join, DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
+                                     LoLoad.getValue(1), HiLoad.getValue(1))};
 
   return DAG.getMergeValues(Ops, SL);
 }
 
+// Widen a vector load from vec3 to vec4.
+SDValue AMDGPUTargetLowering::WidenVectorLoad(SDValue Op,
+                                              SelectionDAG &DAG) const {
+  LoadSDNode *Load = cast<LoadSDNode>(Op);
+  EVT VT = Op.getValueType();
+  assert(VT.getVectorNumElements() == 3);
+  SDValue BasePtr = Load->getBasePtr();
+  EVT MemVT = Load->getMemoryVT();
+  SDLoc SL(Op);
+  const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
+  unsigned BaseAlign = Load->getAlignment();
+
+  EVT WideVT =
+      EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), 4);
+  EVT WideMemVT =
+      EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(), 4);
+  SDValue WideLoad = DAG.getExtLoad(
+      Load->getExtensionType(), SL, WideVT, Load->getChain(), BasePtr, SrcValue,
+      WideMemVT, BaseAlign, Load->getMemOperand()->getFlags());
+  return DAG.getMergeValues(
+      {DAG.getNode(ISD::EXTRACT_SUBVECTOR, SL, VT, WideLoad,
+                   DAG.getConstant(0, SL, getVectorIdxTy(DAG.getDataLayout()))),
+       WideLoad.getValue(1)},
+      SL);
+}
+
 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
                                                SelectionDAG &DAG) const {
   StoreSDNode *Store = cast<StoreSDNode>(Op);
   SDValue Val = Store->getValue();
   EVT VT = Val.getValueType();
 
   // If this is a 2 element vector, we really want to scalarize and not create
   // weird 1 element vectors.
   if (VT.getVectorNumElements() == 2)
     return scalarizeVectorStore(Store, DAG);
 
   EVT MemVT = Store->getMemoryVT();
   SDValue Chain = Store->getChain();
   SDValue BasePtr = Store->getBasePtr();
   SDLoc SL(Op);
 
   EVT LoVT, HiVT;
   EVT LoMemVT, HiMemVT;
   SDValue Lo, Hi;
 
-  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
-  std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
-  std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
+  std::tie(LoVT, HiVT) = getSplitDestVTs(VT, DAG);
+  std::tie(LoMemVT, HiMemVT) = getSplitDestVTs(MemVT, DAG);
+  std::tie(Lo, Hi) = splitVector(Val, SL, LoVT, HiVT, DAG);
 
   SDValue HiPtr = DAG.getObjectPtrOffset(SL, BasePtr, LoMemVT.getStoreSize());
 
   const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
   unsigned BaseAlign = Store->getAlignment();
   unsigned Size = LoMemVT.getStoreSize();
   unsigned HiAlign = MinAlign(BaseAlign, Size);
 

llvm/trunk/lib/Target/AMDGPU/BUFInstructions.td

 >;
 defm BUFFER_LOAD_DWORD : MUBUF_Pseudo_Loads_Lds <
   "buffer_load_dword", VGPR_32, i32, mubuf_load
 >;
 defm BUFFER_LOAD_DWORDX2 : MUBUF_Pseudo_Loads <
   "buffer_load_dwordx2", VReg_64, v2i32, mubuf_load
 >;
 defm BUFFER_LOAD_DWORDX3 : MUBUF_Pseudo_Loads <
-  "buffer_load_dwordx3", VReg_96, untyped, mubuf_load
+  "buffer_load_dwordx3", VReg_96, v3i32, mubuf_load
 >;
 defm BUFFER_LOAD_DWORDX4 : MUBUF_Pseudo_Loads <
   "buffer_load_dwordx4", VReg_128, v4i32, mubuf_load
 >;
 
 // This is not described in AMD documentation,
 // but 'lds' versions of these opcodes are available
 // in at least GFX8+ chips. See Bug 37653.

 >;
 defm BUFFER_STORE_DWORD : MUBUF_Pseudo_Stores <
   "buffer_store_dword", VGPR_32, i32, store_global
 >;
 defm BUFFER_STORE_DWORDX2 : MUBUF_Pseudo_Stores <
   "buffer_store_dwordx2", VReg_64, v2i32, store_global
 >;
 defm BUFFER_STORE_DWORDX3 : MUBUF_Pseudo_Stores <
-  "buffer_store_dwordx3", VReg_96, untyped, store_global
+  "buffer_store_dwordx3", VReg_96, v3i32, store_global
 >;
 defm BUFFER_STORE_DWORDX4 : MUBUF_Pseudo_Stores <
   "buffer_store_dwordx4", VReg_128, v4i32, store_global
 >;
 defm BUFFER_ATOMIC_SWAP : MUBUF_Pseudo_Atomics <
   "buffer_atomic_swap", VGPR_32, i32, atomic_swap_global
 >;
 defm BUFFER_ATOMIC_CMPSWAP : MUBUF_Pseudo_Atomics <
let SubtargetPredicate = HasPackedD16VMem, D16Buf = 1 in {
   defm TBUFFER_STORE_FORMAT_D16_XYZW : MTBUF_Pseudo_Stores <"tbuffer_store_format_d16_xyzw", VReg_64>;
 } // End HasPackedD16VMem.
 
 let SubtargetPredicate = isCIVI in {
 
 //===----------------------------------------------------------------------===//
 // Instruction definitions for CI and newer.
 //===----------------------------------------------------------------------===//
-// Remaining instructions:
-// BUFFER_LOAD_DWORDX3
-// BUFFER_STORE_DWORDX3
 
 def BUFFER_WBINVL1_VOL : MUBUF_Invalidate <"buffer_wbinvl1_vol",
                                            int_amdgcn_buffer_wbinvl1_vol>;
 
 } // End let SubtargetPredicate = isCIVI
 
 //===----------------------------------------------------------------------===//
 // MUBUF Patterns

 defm : MUBUFScratchLoadPat <BUFFER_LOAD_UBYTE_OFFEN, BUFFER_LOAD_UBYTE_OFFSET, i32, az_extloadi8_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_SBYTE_OFFEN, BUFFER_LOAD_SBYTE_OFFSET, i16, sextloadi8_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_UBYTE_OFFEN, BUFFER_LOAD_UBYTE_OFFSET, i16, az_extloadi8_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_SSHORT_OFFEN, BUFFER_LOAD_SSHORT_OFFSET, i32, sextloadi16_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_USHORT_OFFEN, BUFFER_LOAD_USHORT_OFFSET, i32, az_extloadi16_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_USHORT_OFFEN, BUFFER_LOAD_USHORT_OFFSET, i16, load_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_DWORD_OFFEN, BUFFER_LOAD_DWORD_OFFSET, i32, load_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_DWORDX2_OFFEN, BUFFER_LOAD_DWORDX2_OFFSET, v2i32, load_private>;
+defm : MUBUFScratchLoadPat <BUFFER_LOAD_DWORDX3_OFFEN, BUFFER_LOAD_DWORDX3_OFFSET, v3i32, load_private>;
 defm : MUBUFScratchLoadPat <BUFFER_LOAD_DWORDX4_OFFEN, BUFFER_LOAD_DWORDX4_OFFSET, v4i32, load_private>;
 
 let OtherPredicates = [D16PreservesUnusedBits] in {
 defm : MUBUFScratchLoadPat_D16<BUFFER_LOAD_SHORT_D16_HI_OFFEN, BUFFER_LOAD_SHORT_D16_HI_OFFSET, v2i16, load_d16_hi_private>;
 defm : MUBUFScratchLoadPat_D16<BUFFER_LOAD_UBYTE_D16_HI_OFFEN, BUFFER_LOAD_UBYTE_D16_HI_OFFSET, v2i16, az_extloadi8_d16_hi_private>;
 defm : MUBUFScratchLoadPat_D16<BUFFER_LOAD_SBYTE_D16_HI_OFFEN, BUFFER_LOAD_SBYTE_D16_HI_OFFSET, v2i16, sextloadi8_d16_hi_private>;
 defm : MUBUFScratchLoadPat_D16<BUFFER_LOAD_SHORT_D16_HI_OFFEN, BUFFER_LOAD_SHORT_D16_HI_OFFSET, v2f16, load_d16_hi_private>;
 defm : MUBUFScratchLoadPat_D16<BUFFER_LOAD_UBYTE_D16_HI_OFFEN, BUFFER_LOAD_UBYTE_D16_HI_OFFSET, v2f16, az_extloadi8_d16_hi_private>;

 }
 
 defm : MUBUFScratchStorePat <BUFFER_STORE_BYTE_OFFEN, BUFFER_STORE_BYTE_OFFSET, i32, truncstorei8_private>;
 defm : MUBUFScratchStorePat <BUFFER_STORE_SHORT_OFFEN, BUFFER_STORE_SHORT_OFFSET, i32, truncstorei16_private>;
 defm : MUBUFScratchStorePat <BUFFER_STORE_BYTE_OFFEN, BUFFER_STORE_BYTE_OFFSET, i16, truncstorei8_private>;
 defm : MUBUFScratchStorePat <BUFFER_STORE_SHORT_OFFEN, BUFFER_STORE_SHORT_OFFSET, i16, store_private>;
 defm : MUBUFScratchStorePat <BUFFER_STORE_DWORD_OFFEN, BUFFER_STORE_DWORD_OFFSET, i32, store_private>;
 defm : MUBUFScratchStorePat <BUFFER_STORE_DWORDX2_OFFEN, BUFFER_STORE_DWORDX2_OFFSET, v2i32, store_private>;
+defm : MUBUFScratchStorePat <BUFFER_STORE_DWORDX3_OFFEN, BUFFER_STORE_DWORDX3_OFFSET, v3i32, store_private>;
 defm : MUBUFScratchStorePat <BUFFER_STORE_DWORDX4_OFFEN, BUFFER_STORE_DWORDX4_OFFSET, v4i32, store_private>;
 
 
 let OtherPredicates = [D16PreservesUnusedBits] in {
  // Hiding the extract high pattern in the PatFrag seems to not
  // automatically increase the complexity.
 let AddedComplexity = 1 in {
 defm : MUBUFScratchStorePat <BUFFER_STORE_SHORT_D16_HI_OFFEN, BUFFER_STORE_SHORT_D16_HI_OFFSET, i32, store_hi16_private>;

llvm/trunk/lib/Target/AMDGPU/FLATInstructions.td

 def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i32>;
 def : FlatLoadPat <FLAT_LOAD_UBYTE, az_extloadi8_flat, i16>;
 def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i16>;
 def : FlatLoadPat <FLAT_LOAD_USHORT, az_extloadi16_flat, i32>;
 def : FlatLoadPat <FLAT_LOAD_USHORT, load_flat, i16>;
 def : FlatLoadPat <FLAT_LOAD_SSHORT, sextloadi16_flat, i32>;
 def : FlatLoadPat <FLAT_LOAD_DWORD, load_flat, i32>;
 def : FlatLoadPat <FLAT_LOAD_DWORDX2, load_flat, v2i32>;
+def : FlatLoadPat <FLAT_LOAD_DWORDX3, load_flat, v3i32>;
 def : FlatLoadPat <FLAT_LOAD_DWORDX4, load_flat, v4i32>;
 
 def : FlatLoadAtomicPat <FLAT_LOAD_DWORD, atomic_load_flat, i32>;
 def : FlatLoadAtomicPat <FLAT_LOAD_DWORDX2, atomic_load_flat, i64>;
 
 def : FlatStorePat <FLAT_STORE_BYTE, truncstorei8_flat, i32>;
 def : FlatStorePat <FLAT_STORE_SHORT, truncstorei16_flat, i32>;
 def : FlatStorePat <FLAT_STORE_DWORD, store_flat, i32>;
 def : FlatStorePat <FLAT_STORE_DWORDX2, store_flat, v2i32>;
+def : FlatStorePat <FLAT_STORE_DWORDX3, store_flat, v3i32>;
 def : FlatStorePat <FLAT_STORE_DWORDX4, store_flat, v4i32>;
 
 def : FlatStoreAtomicPat <FLAT_STORE_DWORD, atomic_store_flat, i32>;
 def : FlatStoreAtomicPat <FLAT_STORE_DWORDX2, atomic_store_flat, i64>;
 
 def : FlatAtomicPat <FLAT_ATOMIC_ADD_RTN, atomic_add_global, i32>;
 def : FlatAtomicPat <FLAT_ATOMIC_SUB_RTN, atomic_sub_global, i32>;
 def : FlatAtomicPat <FLAT_ATOMIC_INC_RTN, atomic_inc_global, i32>;

 def : FlatLoadSignedPat <GLOBAL_LOAD_UBYTE, az_extloadi8_global, i16>;
 def : FlatLoadSignedPat <GLOBAL_LOAD_SBYTE, sextloadi8_global, i16>;
 def : FlatLoadSignedPat <GLOBAL_LOAD_USHORT, az_extloadi16_global, i32>;
 def : FlatLoadSignedPat <GLOBAL_LOAD_SSHORT, sextloadi16_global, i32>;
 def : FlatLoadSignedPat <GLOBAL_LOAD_USHORT, load_global, i16>;
 
 def : FlatLoadSignedPat <GLOBAL_LOAD_DWORD, load_global, i32>;
 def : FlatLoadSignedPat <GLOBAL_LOAD_DWORDX2, load_global, v2i32>;
+def : FlatLoadSignedPat <GLOBAL_LOAD_DWORDX3, load_global, v3i32>;
 def : FlatLoadSignedPat <GLOBAL_LOAD_DWORDX4, load_global, v4i32>;
 
 def : FlatLoadAtomicPat <GLOBAL_LOAD_DWORD, atomic_load_global, i32>;
 def : FlatLoadAtomicPat <GLOBAL_LOAD_DWORDX2, atomic_load_global, i64>;
 
 def : FlatStoreSignedPat <GLOBAL_STORE_BYTE, truncstorei8_global, i32>;
 def : FlatStoreSignedPat <GLOBAL_STORE_BYTE, truncstorei8_global, i16>;
 def : FlatStoreSignedPat <GLOBAL_STORE_SHORT, truncstorei16_global, i32>;
 def : FlatStoreSignedPat <GLOBAL_STORE_SHORT, store_global, i16>;
 def : FlatStoreSignedPat <GLOBAL_STORE_DWORD, store_global, i32>;
 def : FlatStoreSignedPat <GLOBAL_STORE_DWORDX2, store_global, v2i32>;
+def : FlatStoreSignedPat <GLOBAL_STORE_DWORDX3, store_global, v3i32>;
 def : FlatStoreSignedPat <GLOBAL_STORE_DWORDX4, store_global, v4i32>;
 
 let OtherPredicates = [D16PreservesUnusedBits] in {
 def : FlatStoreSignedPat <GLOBAL_STORE_SHORT_D16_HI, truncstorei16_hi16_global, i32>;
 def : FlatStoreSignedPat <GLOBAL_STORE_BYTE_D16_HI, truncstorei8_hi16_global, i32>;
 
 def : FlatSignedLoadPat_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2i16>;
 def : FlatSignedLoadPat_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2f16>;

llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp

SITargetLowering::SITargetLowering(const TargetMachine &TM,
 
   addRegisterClass(MVT::i32, &AMDGPU::SReg_32_XM0RegClass);
   addRegisterClass(MVT::f32, &AMDGPU::VGPR_32RegClass);
 
   addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
   addRegisterClass(MVT::v2i32, &AMDGPU::SReg_64RegClass);
   addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
 
+  addRegisterClass(MVT::v3i32, &AMDGPU::SGPR_96RegClass);
+  addRegisterClass(MVT::v3f32, &AMDGPU::VReg_96RegClass);
+
   addRegisterClass(MVT::v2i64, &AMDGPU::SReg_128RegClass);
   addRegisterClass(MVT::v2f64, &AMDGPU::SReg_128RegClass);
 
   addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
   addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
 
   addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
   addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
if (Subtarget->has16BitInsts()) {
     addRegisterClass(MVT::v4i16, &AMDGPU::SReg_64RegClass);
     addRegisterClass(MVT::v4f16, &AMDGPU::SReg_64RegClass);
   }
 
   computeRegisterProperties(Subtarget->getRegisterInfo());
 
   // We need to custom lower vector stores from local memory
   setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
+  setOperationAction(ISD::LOAD, MVT::v3i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
   setOperationAction(ISD::LOAD, MVT::i1, Custom);
   setOperationAction(ISD::LOAD, MVT::v32i32, Custom);
 
   setOperationAction(ISD::STORE, MVT::v2i32, Custom);
+  setOperationAction(ISD::STORE, MVT::v3i32, Custom);
   setOperationAction(ISD::STORE, MVT::v4i32, Custom);
   setOperationAction(ISD::STORE, MVT::v8i32, Custom);
   setOperationAction(ISD::STORE, MVT::v16i32, Custom);
   setOperationAction(ISD::STORE, MVT::i1, Custom);
   setOperationAction(ISD::STORE, MVT::v32i32, Custom);
 
   setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
   setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
#endif
   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i8, Custom);
   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i8, Custom);
 
   setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i16, Custom);
   setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f16, Custom);
   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f16, Custom);
 
+  // Deal with vec3 vector operations when widened to vec4.
+  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v3i32, Expand);
+  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v3f32, Expand);
+  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i32, Expand);
+  setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4f32, Expand);
+
   // BUFFER/FLAT_ATOMIC_CMP_SWAP on GCN GPUs needs input marshalling,
   // and output demarshalling
   setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
   setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
 
   // We can't return success/failure, only the old value,
   // let LLVM add the comparison
   setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i32, Expand);
uint64_t Offset = getImplicitParameterOffset(DAG.getMachineFunction(),
                                                FIRST_IMPLICIT);
   return lowerKernArgParameterPtr(DAG, SL, DAG.getEntryNode(), Offset);
 }
 
 SDValue SITargetLowering::convertArgType(SelectionDAG &DAG, EVT VT, EVT MemVT,
                                          const SDLoc &SL, SDValue Val,
                                          bool Signed,
                                          const ISD::InputArg *Arg) const {
+  // First, if it is a widened vector, narrow it.
+  if (VT.isVector() &&
+      VT.getVectorNumElements() != MemVT.getVectorNumElements()) {
+    EVT NarrowedVT =
+        EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(),
+                         VT.getVectorNumElements());
+    Val = DAG.getNode(ISD::EXTRACT_SUBVECTOR, SL, NarrowedVT, Val,
+                      DAG.getConstant(0, SL, MVT::i32));
+  }
+
+  // Then convert the vector elements or scalar value.
   if (Arg && (Arg->Flags.isSExt() || Arg->Flags.isZExt()) &&
       VT.bitsLT(MemVT)) {
     unsigned Opc = Arg->Flags.isZExt() ? ISD::AssertZext : ISD::AssertSext;
     Val = DAG.getNode(Opc, SL, MemVT, Val, DAG.getValueType(VT));
   }
 
   if (MemVT.isFloatingPoint())
     Val = getFPExtOrFPTrunc(DAG, Val, SL, VT);
if (ExtType == ISD::NON_EXTLOAD && MemVT.getSizeInBits() < 32) {
     SDValue BasePtr = Load->getBasePtr();
     MachineMemOperand *MMO = Load->getMemOperand();
 
     EVT RealMemVT = (MemVT == MVT::i1) ? MVT::i8 : MVT::i16;
 
     SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
                                    BasePtr, RealMemVT, MMO);
 
+    if (!MemVT.isVector()) {
       SDValue Ops[] = {
         DAG.getNode(ISD::TRUNCATE, DL, MemVT, NewLD),
         NewLD.getValue(1)
       };
 
       return DAG.getMergeValues(Ops, DL);
     }
 
+    SmallVector<SDValue, 3> Elts;
+    for (unsigned I = 0, N = MemVT.getVectorNumElements(); I != N; ++I) {
+      SDValue Elt = DAG.getNode(ISD::SRL, DL, MVT::i32, NewLD,
+                                DAG.getConstant(I, DL, MVT::i32));
+
+      Elts.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Elt));
+    }
+
+    SDValue Ops[] = {
+      DAG.getBuildVector(MemVT, DL, Elts),
+      NewLD.getValue(1)
+    };
+
+    return DAG.getMergeValues(Ops, DL);
+  }
+
   if (!MemVT.isVector())
     return SDValue();
 
   assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
          "Custom lowering for non-i32 vectors hasn't been implemented.");
 
   unsigned Alignment = Load->getAlignment();
   unsigned AS = Load->getAddressSpace();
SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
   if (AS == AMDGPUAS::FLAT_ADDRESS)
     AS = MFI->hasFlatScratchInit() ?
          AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
 
   unsigned NumElements = MemVT.getVectorNumElements();
 
   if (AS == AMDGPUAS::CONSTANT_ADDRESS ||
       AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT) {
-    if (!Op->isDivergent() && Alignment >= 4 && NumElements < 32)
+    if (!Op->isDivergent() && Alignment >= 4 && NumElements < 32) {
+      if (MemVT.isPow2VectorType())
         return SDValue();
+      if (NumElements == 3)
+        return WidenVectorLoad(Op, DAG);
+      return SplitVectorLoad(Op, DAG);
+    }
     // Non-uniform loads will be selected to MUBUF instructions, so they
     // have the same legalization requirements as global and private
     // loads.
     //
   }
 
   if (AS == AMDGPUAS::CONSTANT_ADDRESS ||
       AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
       AS == AMDGPUAS::GLOBAL_ADDRESS) {
     if (Subtarget->getScalarizeGlobalBehavior() && !Op->isDivergent() &&
         !Load->isVolatile() && isMemOpHasNoClobberedMemOperand(Load) &&
-        Alignment >= 4 && NumElements < 32)
+        Alignment >= 4 && NumElements < 32) {
+      if (MemVT.isPow2VectorType())
         return SDValue();
+      if (NumElements == 3)
+        return WidenVectorLoad(Op, DAG);
+      return SplitVectorLoad(Op, DAG);
+    }
     // Non-uniform loads will be selected to MUBUF instructions, so they
     // have the same legalization requirements as global and private
     // loads.
     //
   }
   if (AS == AMDGPUAS::CONSTANT_ADDRESS ||
       AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
       AS == AMDGPUAS::GLOBAL_ADDRESS ||
       AS == AMDGPUAS::FLAT_ADDRESS) {
     if (NumElements > 4)
       return SplitVectorLoad(Op, DAG);
-    // v4 loads are supported for private and global memory.
+    // v3 loads not supported on SI.
+    if (NumElements == 3 && !Subtarget->hasDwordx3LoadStores())
+      return WidenVectorLoad(Op, DAG);
+    // v3 and v4 loads are supported for private and global memory.
     return SDValue();
   }
   if (AS == AMDGPUAS::PRIVATE_ADDRESS) {
     // Depending on the setting of the private_element_size field in the
     // resource descriptor, we can only make private accesses up to a certain
     // size.
     switch (Subtarget->getMaxPrivateElementSize()) {
     case 4:
       return scalarizeVectorLoad(Load, DAG);
     case 8:
       if (NumElements > 2)
         return SplitVectorLoad(Op, DAG);
       return SDValue();
     case 16:
       // Same as global/flat
       if (NumElements > 4)
         return SplitVectorLoad(Op, DAG);
+      // v3 loads not supported on SI.
+      if (NumElements == 3 && !Subtarget->hasDwordx3LoadStores())
+        return WidenVectorLoad(Op, DAG);
       return SDValue();
     default:
       llvm_unreachable("unsupported private_element_size");
     }
   } else if (AS == AMDGPUAS::LOCAL_ADDRESS) {
     // Use ds_read_b128 if possible.
     if (Subtarget->useDS128() && Load->getAlignment() >= 16 &&
         MemVT.getStoreSize() == 16)
if (AS == AMDGPUAS::FLAT_ADDRESS)
     AS = MFI->hasFlatScratchInit() ?
          AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
 
   unsigned NumElements = VT.getVectorNumElements();
   if (AS == AMDGPUAS::GLOBAL_ADDRESS ||
       AS == AMDGPUAS::FLAT_ADDRESS) {
     if (NumElements > 4)
       return SplitVectorStore(Op, DAG);
+    // v3 stores not supported on SI.
+    if (NumElements == 3 && !Subtarget->hasDwordx3LoadStores())
+      return SplitVectorStore(Op, DAG);
     return SDValue();
   } else if (AS == AMDGPUAS::PRIVATE_ADDRESS) {
     switch (Subtarget->getMaxPrivateElementSize()) {
     case 4:
       return scalarizeVectorStore(Store, DAG);
     case 8:
       if (NumElements > 2)
         return SplitVectorStore(Op, DAG);
       return SDValue();
     case 16:
-      if (NumElements > 4)
+      if (NumElements > 4 || NumElements == 3)
         return SplitVectorStore(Op, DAG);
       return SDValue();
     default:
       llvm_unreachable("unsupported private_element_size");
     }
   } else if (AS == AMDGPUAS::LOCAL_ADDRESS) {
     // Use ds_write_b128 if possible.
     if (Subtarget->useDS128() && Store->getAlignment() >= 16 &&
-        VT.getStoreSize() == 16)
+        VT.getStoreSize() == 16 && NumElements != 3)
       return SDValue();
 
     if (NumElements > 2)
       return SplitVectorStore(Op, DAG);
 
     // SI has a hardware bug in the LDS / GDS boounds checking: if the base
     // address is negative, then the instruction is incorrectly treated as
     // out-of-bounds even if base + offsets is in bounds. Split vectorized
case 'r':
         return std::make_pair(0U, nullptr);
       case 32:
       case 16:
         RC = &AMDGPU::SReg_32_XM0RegClass;
         break;
       case 64:
         RC = &AMDGPU::SGPR_64RegClass;
         break;
+      case 96:
+        RC = &AMDGPU::SReg_96RegClass;
+        break;
       case 128:
         RC = &AMDGPU::SReg_128RegClass;
         break;
       case 256:
         RC = &AMDGPU::SReg_256RegClass;
         break;
       case 512:
         RC = &AMDGPU::SReg_512RegClass;

llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp

BuildMI(MBB, MI, DL, get(AMDGPU::S_CMP_LG_U32))
       .addReg(SrcReg, getKillRegState(KillSrc))
       .addImm(0);
     return;
   }
 
   unsigned EltSize = 4;
   unsigned Opcode = AMDGPU::V_MOV_B32_e32;
   if (RI.isSGPRClass(RC)) {
-    if (RI.getRegSizeInBits(*RC) > 32) {
+    // TODO: Copy vec3/vec5 with s_mov_b64s then final s_mov_b32.
+    if (!(RI.getRegSizeInBits(*RC) % 64)) {
       Opcode =  AMDGPU::S_MOV_B64;
       EltSize = 8;
     } else {
       Opcode = AMDGPU::S_MOV_B32;
       EltSize = 4;
     }
 
     if (!RI.isSGPRClass(RI.getPhysRegClass(SrcReg))) {

 }
 
 static unsigned getSGPRSpillSaveOpcode(unsigned Size) {
   switch (Size) {
   case 4:
     return AMDGPU::SI_SPILL_S32_SAVE;
   case 8:
     return AMDGPU::SI_SPILL_S64_SAVE;
+  case 12:
+    return AMDGPU::SI_SPILL_S96_SAVE;
   case 16:
     return AMDGPU::SI_SPILL_S128_SAVE;
   case 32:
     return AMDGPU::SI_SPILL_S256_SAVE;
   case 64:
     return AMDGPU::SI_SPILL_S512_SAVE;
   default:
     llvm_unreachable("unknown register size");

 }
 
 static unsigned getSGPRSpillRestoreOpcode(unsigned Size) {
   switch (Size) {
   case 4:
     return AMDGPU::SI_SPILL_S32_RESTORE;
   case 8:
     return AMDGPU::SI_SPILL_S64_RESTORE;
+  case 12:
+    return AMDGPU::SI_SPILL_S96_RESTORE;
   case 16:
     return AMDGPU::SI_SPILL_S128_RESTORE;
   case 32:
     return AMDGPU::SI_SPILL_S256_RESTORE;
   case 64:
     return AMDGPU::SI_SPILL_S512_RESTORE;
   default:
     llvm_unreachable("unknown register size");
static const int16_t Sub0_15_64[] = {
     AMDGPU::sub12_sub13, AMDGPU::sub14_sub15,
   };
 
   unsigned SelOp = AMDGPU::V_CNDMASK_B32_e32;
   const TargetRegisterClass *EltRC = &AMDGPU::VGPR_32RegClass;
   const int16_t *SubIndices = Sub0_15;
   int NElts = DstSize / 32;
 
-  // 64-bit select is only avaialble for SALU.
+  // 64-bit select is only available for SALU.
+  // TODO: Split 96-bit into 64-bit and 32-bit, not 3x 32-bit.
   if (Pred == SCC_TRUE) {
+    if (NElts % 2) {
+      SelOp = AMDGPU::S_CSELECT_B32;
+      EltRC = &AMDGPU::SGPR_32RegClass;
+    } else {
       SelOp = AMDGPU::S_CSELECT_B64;
       EltRC = &AMDGPU::SGPR_64RegClass;
       SubIndices = Sub0_15_64;
-
-    assert(NElts % 2 == 0);
       NElts /= 2;
     }
+  }
 
   MachineInstrBuilder MIB = BuildMI(
     MBB, I, DL, get(AMDGPU::REG_SEQUENCE), DstReg);
 
   I = MIB->getIterator();
 
   SmallVector<unsigned, 8> Regs;
   for (int Idx = 0; Idx != NElts; ++Idx) {

llvm/trunk/lib/Target/AMDGPU/SIInstructions.td

 }
 
 // You cannot use M0 as the output of v_readlane_b32 instructions or
 // use it in the sdata operand of SMEM instructions. We still need to
 // be able to spill the physical register m0, so allow it for
 // SI_SPILL_32_* instructions.
 defm SI_SPILL_S32  : SI_SPILL_SGPR <SReg_32>;
 defm SI_SPILL_S64  : SI_SPILL_SGPR <SReg_64>;
+defm SI_SPILL_S96  : SI_SPILL_SGPR <SReg_96>;
 defm SI_SPILL_S128 : SI_SPILL_SGPR <SReg_128>;
 defm SI_SPILL_S256 : SI_SPILL_SGPR <SReg_256>;
 defm SI_SPILL_S512 : SI_SPILL_SGPR <SReg_512>;
 
 multiclass SI_SPILL_VGPR <RegisterClass vgpr_class> {
   let UseNamedOperandTable = 1, VGPRSpill = 1,
        SchedRW = [WriteVMEM] in {
     def _SAVE : VPseudoInstSI <
foreach Index = 0-2 in {
   def Extract_Element_v2f32_#Index : Extract_Element <
     f32, v2f32, Index, !cast<SubRegIndex>(sub#Index)
   >;
   def Insert_Element_v2f32_#Index : Insert_Element <
     f32, v2f32, Index, !cast<SubRegIndex>(sub#Index)
   >;
 }
 
+foreach Index = 0-2 in {
+  def Extract_Element_v3i32_#Index : Extract_Element <
+    i32, v3i32, Index, !cast<SubRegIndex>(sub#Index)
+  >;
+  def Insert_Element_v3i32_#Index : Insert_Element <
+    i32, v3i32, Index, !cast<SubRegIndex>(sub#Index)
+  >;
+
+  def Extract_Element_v3f32_#Index : Extract_Element <
+    f32, v3f32, Index, !cast<SubRegIndex>(sub#Index)
+  >;
+  def Insert_Element_v3f32_#Index : Insert_Element <
+    f32, v3f32, Index, !cast<SubRegIndex>(sub#Index)
+  >;
+}
+
 foreach Index = 0-3 in {
   def Extract_Element_v4i32_#Index : Extract_Element <
     i32, v4i32, Index, !cast<SubRegIndex>(sub#Index)
   >;
   def Insert_Element_v4i32_#Index : Insert_Element <
     i32, v4i32, Index, !cast<SubRegIndex>(sub#Index)
   >;
 

 def : BitConvert <v4i16, i64, VReg_64>;
 def : BitConvert <v4f16, i64, VReg_64>;
 def : BitConvert <i64, v4i16, VReg_64>;
 def : BitConvert <i64, v4f16, VReg_64>;
 
 def : BitConvert <v4i32, v4f32, VReg_128>;
 def : BitConvert <v4f32, v4i32, VReg_128>;
 
+// 96-bit bitcast
+def : BitConvert <v3i32, v3f32, SGPR_96>;
+def : BitConvert <v3f32, v3i32, SGPR_96>;
+
 // 128-bit bitcast
 def : BitConvert <v2i64, v4i32, SReg_128>;
 def : BitConvert <v4i32, v2i64, SReg_128>;
 def : BitConvert <v2f64, v4f32, VReg_128>;
 def : BitConvert <v2f64, v4i32, VReg_128>;
 def : BitConvert <v4f32, v2f64, VReg_128>;
 def : BitConvert <v4i32, v2f64, VReg_128>;
 def : BitConvert <v2i64, v2f64, VReg_128>;

llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp

static unsigned getNumSubRegsForSpillOp(unsigned Op) {
   case AMDGPU::SI_SPILL_V256_SAVE:
   case AMDGPU::SI_SPILL_V256_RESTORE:
     return 8;
   case AMDGPU::SI_SPILL_S128_SAVE:
   case AMDGPU::SI_SPILL_S128_RESTORE:
   case AMDGPU::SI_SPILL_V128_SAVE:
   case AMDGPU::SI_SPILL_V128_RESTORE:
     return 4;
+  case AMDGPU::SI_SPILL_S96_SAVE:
+  case AMDGPU::SI_SPILL_S96_RESTORE:
   case AMDGPU::SI_SPILL_V96_SAVE:
   case AMDGPU::SI_SPILL_V96_RESTORE:
     return 3;
   case AMDGPU::SI_SPILL_S64_SAVE:
   case AMDGPU::SI_SPILL_S64_RESTORE:
   case AMDGPU::SI_SPILL_V64_SAVE:
   case AMDGPU::SI_SPILL_V64_RESTORE:
     return 2;

 bool SIRegisterInfo::eliminateSGPRToVGPRSpillFrameIndex(
   MachineBasicBlock::iterator MI,
   int FI,
   RegScavenger *RS) const {
   switch (MI->getOpcode()) {
   case AMDGPU::SI_SPILL_S512_SAVE:
   case AMDGPU::SI_SPILL_S256_SAVE:
   case AMDGPU::SI_SPILL_S128_SAVE:
+  case AMDGPU::SI_SPILL_S96_SAVE:
   case AMDGPU::SI_SPILL_S64_SAVE:
   case AMDGPU::SI_SPILL_S32_SAVE:
     return spillSGPR(MI, FI, RS, true);
   case AMDGPU::SI_SPILL_S512_RESTORE:
   case AMDGPU::SI_SPILL_S256_RESTORE:
   case AMDGPU::SI_SPILL_S128_RESTORE:
+  case AMDGPU::SI_SPILL_S96_RESTORE:
   case AMDGPU::SI_SPILL_S64_RESTORE:
   case AMDGPU::SI_SPILL_S32_RESTORE:
     return restoreSGPR(MI, FI, RS, true);
   default:
     llvm_unreachable("not an SGPR spill instruction");
   }
 }
 
void SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI,
   MachineOperand &FIOp = MI->getOperand(FIOperandNum);
   int Index = MI->getOperand(FIOperandNum).getIndex();
 
   switch (MI->getOpcode()) {
     // SGPR register spill
     case AMDGPU::SI_SPILL_S512_SAVE:
     case AMDGPU::SI_SPILL_S256_SAVE:
     case AMDGPU::SI_SPILL_S128_SAVE:
+    case AMDGPU::SI_SPILL_S96_SAVE:
     case AMDGPU::SI_SPILL_S64_SAVE:
     case AMDGPU::SI_SPILL_S32_SAVE: {
       spillSGPR(MI, Index, RS);
       break;
     }
 
     // SGPR register restore
     case AMDGPU::SI_SPILL_S512_RESTORE:
     case AMDGPU::SI_SPILL_S256_RESTORE:
     case AMDGPU::SI_SPILL_S128_RESTORE:
+    case AMDGPU::SI_SPILL_S96_RESTORE:
     case AMDGPU::SI_SPILL_S64_RESTORE:
     case AMDGPU::SI_SPILL_S32_RESTORE: {
       restoreSGPR(MI, Index, RS);
       break;
     }
 
     // VGPR register spill
     case AMDGPU::SI_SPILL_V512_SAVE:
const TargetRegisterClass *SIRegisterInfo::getPhysRegClass(unsigned Reg) const {
   assert(!TargetRegisterInfo::isVirtualRegister(Reg));
 
   static const TargetRegisterClass *const BaseClasses[] = {
     &AMDGPU::VGPR_32RegClass,
     &AMDGPU::SReg_32RegClass,
     &AMDGPU::VReg_64RegClass,
     &AMDGPU::SReg_64RegClass,
     &AMDGPU::VReg_96RegClass,
+    &AMDGPU::SReg_96RegClass,
     &AMDGPU::VReg_128RegClass,
     &AMDGPU::SReg_128RegClass,
     &AMDGPU::VReg_256RegClass,
     &AMDGPU::SReg_256RegClass,
     &AMDGPU::VReg_512RegClass,
     &AMDGPU::SReg_512RegClass,
     &AMDGPU::SCC_CLASSRegClass,
     &AMDGPU::Pseudo_SReg_32RegClass,

 
 const TargetRegisterClass *SIRegisterInfo::getEquivalentSGPRClass(
                                          const TargetRegisterClass *VRC) const {
   switch (getRegSizeInBits(*VRC)) {
   case 32:
     return &AMDGPU::SGPR_32RegClass;
   case 64:
     return &AMDGPU::SReg_64RegClass;
+  case 96:
+    return &AMDGPU::SReg_96RegClass;
   case 128:
     return &AMDGPU::SReg_128RegClass;
   case 256:
     return &AMDGPU::SReg_256RegClass;
   case 512:
     return &AMDGPU::SReg_512RegClass;
   default:
     llvm_unreachable("Invalid register class size");
   }
 }
 
 const TargetRegisterClass *SIRegisterInfo::getSubRegClass(
                          const TargetRegisterClass *RC, unsigned SubIdx) const {
   if (SubIdx == AMDGPU::NoSubRegister)
     return RC;
 
   // We can assume that each lane corresponds to one 32-bit register.
   unsigned Count = getSubRegIndexLaneMask(SubIdx).getNumLanes();
   if (isSGPRClass(RC)) {
     switch (Count) {
     case 1:
       return &AMDGPU::SGPR_32RegClass;
     case 2:
       return &AMDGPU::SReg_64RegClass;
+    case 3:
+      return &AMDGPU::SReg_96RegClass;
     case 4:
       return &AMDGPU::SReg_128RegClass;
     case 8:
       return &AMDGPU::SReg_256RegClass;
     case 16: /* fall-through */
     default:
       llvm_unreachable("Invalid sub-register class size");
     }
SIRegisterInfo::getConstrainedRegClassForOperand(const MachineOperand &MO,
   case 32:
     return RB->getID() == AMDGPU::VGPRRegBankID ? &AMDGPU::VGPR_32RegClass :
                                                   &AMDGPU::SReg_32_XM0RegClass;
   case 64:
     return RB->getID() == AMDGPU::VGPRRegBankID ? &AMDGPU::VReg_64RegClass :
                                                    &AMDGPU::SReg_64_XEXECRegClass;
   case 96:
     return RB->getID() == AMDGPU::VGPRRegBankID ? &AMDGPU::VReg_96RegClass :
-                                                  nullptr;
+                                                  &AMDGPU::SReg_96RegClass;
   case 128:
     return RB->getID() == AMDGPU::VGPRRegBankID ? &AMDGPU::VReg_128RegClass :
                                                   &AMDGPU::SReg_128RegClass;
   case 256:
     return RB->getID() == AMDGPU::VGPRRegBankID ? &AMDGPU::VReg_256RegClass :
                                                   &AMDGPU::SReg_256RegClass;
   case 512:
     return RB->getID() == AMDGPU::VGPRRegBankID ? &AMDGPU::VReg_512RegClass :

llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.td

def SGPR_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
   let AllocationPriority = 7;
 }
 
 // SGPR 64-bit registers
 def SGPR_64Regs : RegisterTuples<getSubRegs<2>.ret,
                              [(add (decimate SGPR_32, 2)),
                               (add (decimate (shl SGPR_32, 1), 2))]>;
 
+// SGPR 96-bit registers. No operations use these, but for symmetry with 96-bit VGPRs.
+def SGPR_96Regs : RegisterTuples<getSubRegs<3>.ret,
+                            [(add (decimate SGPR_32, 3)),
+                             (add (decimate (shl SGPR_32, 1), 3)),
+                             (add (decimate (shl SGPR_32, 2), 3))]>;
+
 // SGPR 128-bit registers
 def SGPR_128Regs : RegisterTuples<getSubRegs<4>.ret,
                               [(add (decimate SGPR_32, 4)),
                                (add (decimate (shl SGPR_32, 1), 4)),
                                (add (decimate (shl SGPR_32, 2), 4)),
                                (add (decimate (shl SGPR_32, 3), 4))]>;
 
 // SGPR 256-bit registers

 }
 
 // Subset of SReg_32 without M0 for SMRD instructions and alike.
 // See comments in SIInstructions.td for more info.
 def SReg_32_XM0_XEXEC : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
   (add SGPR_32, VCC_LO, VCC_HI, FLAT_SCR_LO, FLAT_SCR_HI, XNACK_MASK_LO, XNACK_MASK_HI,
    TTMP_32, TMA_LO, TMA_HI, TBA_LO, TBA_HI, SRC_SHARED_BASE, SRC_SHARED_LIMIT,
    SRC_PRIVATE_BASE, SRC_PRIVATE_LIMIT, SRC_POPS_EXITING_WAVE_ID)> {
-  let AllocationPriority = 7;
+  let AllocationPriority = 8;
 }
 
 def SReg_32_XEXEC_HI : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
   (add SReg_32_XM0_XEXEC, EXEC_LO, M0_CLASS)> {
-  let AllocationPriority = 7;
+  let AllocationPriority = 8;
 }
 
 def SReg_32_XM0 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
   (add SReg_32_XM0_XEXEC, EXEC_LO, EXEC_HI)> {
-  let AllocationPriority = 7;
+  let AllocationPriority = 8;
 }
 
 // Register class for all scalar registers (SGPRs + Special Registers)
 def SReg_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
   (add SReg_32_XM0, M0_CLASS, EXEC_LO, EXEC_HI, SReg_32_XEXEC_HI)> {
-  let AllocationPriority = 7;
+  let AllocationPriority = 8;
 }
 
 def SRegOrLds_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
   (add SReg_32_XM0, M0_CLASS, EXEC_LO, EXEC_HI, SReg_32_XEXEC_HI, LDS_DIRECT_CLASS)> {
   let isAllocatable = 0;
 }
 
 def SGPR_64 : RegisterClass<"AMDGPU", [v2i32, i64, v2f32, f64, v4i16, v4f16], 32, (add SGPR_64Regs)> {
   let CopyCost = 1;
-  let AllocationPriority = 8;
+  let AllocationPriority = 9;
 }
 
 def TTMP_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64, v4i16, v4f16], 32, (add TTMP_64Regs)> {
   let isAllocatable = 0;
 }
 
 def SReg_64_XEXEC : RegisterClass<"AMDGPU", [v2i32, i64, v2f32, f64, i1, v4i16, v4f16], 32,
   (add SGPR_64, VCC, FLAT_SCR, XNACK_MASK, TTMP_64, TBA, TMA)> {
   let CopyCost = 1;
-  let AllocationPriority = 8;
+  let AllocationPriority = 9;
 }
 
 def SReg_64 : RegisterClass<"AMDGPU", [v2i32, i64, v2f32, f64, i1, v4i16, v4f16], 32,
   (add SReg_64_XEXEC, EXEC)> {
   let CopyCost = 1;
-  let AllocationPriority = 8;
+  let AllocationPriority = 9;
 }
 
 // Requires 2 s_mov_b64 to copy
 let CopyCost = 2 in {
 
-def SGPR_128 : RegisterClass<"AMDGPU", [v4i32, v4f32, v2i64], 32, (add SGPR_128Regs)> {
+// There are no 3-component scalar instructions, but this is needed
+// for symmetry with VGPRs.
+def SGPR_96 : RegisterClass<"AMDGPU", [v3i32, v3f32], 32,
+  (add SGPR_96Regs)> {
   let AllocationPriority = 10;
 }
 
+def SReg_96 : RegisterClass<"AMDGPU", [v3i32, v3f32], 32,
+  (add SGPR_96)> {
+  let AllocationPriority = 10;
+}
+
+def SGPR_128 : RegisterClass<"AMDGPU", [v4i32, v4f32, v2i64], 32, (add SGPR_128Regs)> {
+  let AllocationPriority = 11;
+}
+
 def TTMP_128 : RegisterClass<"AMDGPU", [v4i32, v4f32, v2i64], 32, (add TTMP_128Regs)> {
   let isAllocatable = 0;
 }
 
 def SReg_128 : RegisterClass<"AMDGPU", [v4i32, v4f32, v2i64, v2f64], 32,
   (add SGPR_128, TTMP_128)> {
-  let AllocationPriority = 10;
+  let AllocationPriority = 11;
 }
 
 } // End CopyCost = 2
 
 def SGPR_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 32, (add SGPR_256Regs)> {
-  let AllocationPriority = 11;
+  let AllocationPriority = 13;
 }
 
 def TTMP_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 32, (add TTMP_256Regs)> {
   let isAllocatable = 0;
 }
 
 def SReg_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 32,
   (add SGPR_256, TTMP_256)> {
   // Requires 4 s_mov_b64 to copy
   let CopyCost = 4;
-  let AllocationPriority = 11;
+  let AllocationPriority = 13;
 }
 
 def SGPR_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32, (add SGPR_512Regs)> {
-  let AllocationPriority = 12;
+  let AllocationPriority = 14;
 }
 
 def TTMP_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32, (add TTMP_512Regs)> {
   let isAllocatable = 0;
 }
 
 def SReg_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32,
   (add SGPR_512, TTMP_512)> {
   // Requires 8 s_mov_b64 to copy
   let CopyCost = 8;
-  let AllocationPriority = 12;
+  let AllocationPriority = 14;
 }
 
 def VRegOrLds_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
                                  (add VGPR_32, LDS_DIRECT_CLASS)> {
   let isAllocatable = 0;
 }
 
 // Register class for all vector registers (VGPRs + Interploation Registers)
 def VReg_64 : RegisterClass<"AMDGPU", [i64, f64, v2i32, v2f32, v4f16, v4i16], 32, (add VGPR_64)> {
   let Size = 64;
 
   // Requires 2 v_mov_b32 to copy
   let CopyCost = 2;
   let AllocationPriority = 2;
 }
 
-def VReg_96 : RegisterClass<"AMDGPU", [untyped], 32, (add VGPR_96)> {
+def VReg_96 : RegisterClass<"AMDGPU", [v3i32, v3f32], 32, (add VGPR_96)> {
   let Size = 96;
 
   // Requires 3 v_mov_b32 to copy
   let CopyCost = 3;
   let AllocationPriority = 3;
 }
 
 def VReg_128 : RegisterClass<"AMDGPU", [v4i32, v4f32, v2i64, v2f64], 32, (add VGPR_128)> {
   let Size = 128;
 
   // Requires 4 v_mov_b32 to copy
   let CopyCost = 4;
   let AllocationPriority = 4;
 }
 
 def VReg_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 32, (add VGPR_256)> {
   let Size = 256;
   let CopyCost = 8;
-  let AllocationPriority = 5;
+  let AllocationPriority = 6;
 }
 
 def VReg_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32, (add VGPR_512)> {
   let Size = 512;
   let CopyCost = 16;
-  let AllocationPriority = 6;
+  let AllocationPriority = 7;
 }
 
 def VReg_1 : RegisterClass<"AMDGPU", [i1], 32, (add VGPR_32)> {
   let Size = 32;
 }
 
 def VS_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
                           (add VGPR_32, SReg_32, LDS_DIRECT_CLASS)> {

llvm/trunk/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

unsigned getRegBitWidth(unsigned RCID) {
   case AMDGPU::SRegOrLds_32RegClassID:
     return 32;
   case AMDGPU::SGPR_64RegClassID:
   case AMDGPU::VS_64RegClassID:
   case AMDGPU::SReg_64RegClassID:
   case AMDGPU::VReg_64RegClassID:
   case AMDGPU::SReg_64_XEXECRegClassID:
     return 64;
+  case AMDGPU::SGPR_96RegClassID:
+  case AMDGPU::SReg_96RegClassID:
   case AMDGPU::VReg_96RegClassID:
     return 96;
   case AMDGPU::SGPR_128RegClassID:
   case AMDGPU::SReg_128RegClassID:
   case AMDGPU::VReg_128RegClassID:
     return 128;
   case AMDGPU::SReg_256RegClassID:
   case AMDGPU::VReg_256RegClassID:

llvm/trunk/test/CodeGen/AMDGPU/call-return-types.ll

 define amdgpu_kernel void @test_call_external_v2i32_func_void() #0 {
   %val = call <2 x i32> @external_v2i32_func_void()
   store volatile <2 x i32> %val, <2 x i32> addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}test_call_external_v3i32_func_void:
 ; GCN: s_swappc
-; GFX7-DAG: flat_store_dwordx2 {{.*}}, v[0:1]
-; GFX7-DAG: flat_store_dword {{.*}}, v2
-; GFX89-DAG: buffer_store_dwordx2 v[0:1]
-; GFX89-DAG: buffer_store_dword v2
+; GFX7-DAG: flat_store_dwordx3 {{.*}}, v[0:2]
+; GFX89-DAG: buffer_store_dwordx3 v[0:2]
 define amdgpu_kernel void @test_call_external_v3i32_func_void() #0 {
   %val = call <3 x i32> @external_v3i32_func_void()
   store volatile <3 x i32> %val, <3 x i32> addrspace(1)* undef, align 8
   ret void
 }
 
 ; GCN-LABEL: {{^}}test_call_external_v4i32_func_void:
 define amdgpu_kernel void @test_call_external_v4i32_func_void() #0 {

 define amdgpu_kernel void @test_call_external_v4f16_func_void() #0 {
   %val = call <4 x half> @external_v4f16_func_void()
   store volatile <4 x half> %val, <4 x half> addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}test_call_external_v3f32_func_void:
 ; GCN: s_swappc
-; GFX7-DAG: flat_store_dwordx2 {{.*}}, v[0:1]
-; GFX7-DAG: flat_store_dword {{.*}}, v2
-; GFX89-DAG: buffer_store_dwordx2 v[0:1]
-; GFX89-DAG: buffer_store_dword v2
+; GFX7-DAG: flat_store_dwordx3 {{.*}}, v[0:2]
+; GFX89-DAG: buffer_store_dwordx3 v[0:2]
 define amdgpu_kernel void @test_call_external_v3f32_func_void() #0 {
   %val = call <3 x float> @external_v3f32_func_void()
   store volatile <3 x float> %val, <3 x float> addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}test_call_external_v5f32_func_void:
 ; GCN: s_swappc

llvm/trunk/test/CodeGen/AMDGPU/early-if-convert-cost.ll

-; RUN:  llc -amdgpu-scalarize-global-loads=false  -stress-early-ifcvt -amdgpu-early-ifcvt=1 -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -stress-early-ifcvt -amdgpu-early-ifcvt=1 -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,SI %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -stress-early-ifcvt -amdgpu-early-ifcvt=1 -march=amdgcn -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GCNX3 %s
 
 ; FIXME: Most of these cases that don't trigger because of broken cost
 ; heuristics. Should not need -stress-early-ifcvt
 
 ; GCN-LABEL: {{^}}test_vccnz_ifcvt_triangle64:
 ; GCN: buffer_load_dwordx2 v{{\[}}[[VAL_LO:[0-9]+]]:[[VAL_HI:[0-9]+]]{{\]}}
 ; GCN: v_cmp_neq_f64_e32 vcc, 1.0, v{{\[}}[[VAL_LO]]:[[VAL_HI]]{{\]}}
 ; GCN: v_add_f64 v{{\[}}[[ADD_LO:[0-9]+]]:[[ADD_HI:[0-9]+]]{{\]}}, v{{\[}}[[VAL_LO]]:[[VAL_HI]]{{\]}}, v{{\[}}[[VAL_LO]]:[[VAL_HI]]{{\]}}

 ; GCN: v_add_i32_e32
 ; GCN: v_add_i32_e32
 ; GCN: s_mov_b64 vcc, [[CMP]]
 
 ; GCN: v_cndmask_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, vcc
 ; GCN: v_cndmask_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, vcc
 ; GCN: v_cndmask_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, vcc
 
-; GCN-DAG: buffer_store_dword v
-; GCN-DAG: buffer_store_dwordx2
+; SI-DAG: buffer_store_dwordx2
+; SI-DAG: buffer_store_dword v
+; GCNX3: buffer_store_dwordx3
 define amdgpu_kernel void @test_vccnz_ifcvt_triangle96(<3 x i32> addrspace(1)* %out, <3 x i32> addrspace(1)* %in, float %cnd) #0 {
 entry:
   %v = load <3 x i32>, <3 x i32> addrspace(1)* %in
   %cc = fcmp oeq float %cnd, 1.000000e+00
   br i1 %cc, label %if, label %endif
 
 if:
   %u = add <3 x i32> %v, %v

llvm/trunk/test/CodeGen/AMDGPU/early-if-convert.ll

endif:
   ret void
 }
 
 ; TODO: Can do s_cselect_b64; s_cselect_b32
 ; GCN-LABEL: {{^}}test_scc1_sgpr_ifcvt_triangle96:
 ; GCN: s_add_i32
 ; GCN: s_add_i32
 ; GCN: s_add_i32
-; GCN: s_add_i32
 ; GCN: s_cmp_lg_u32 s{{[0-9]+}}, 1
-; GCN-NEXT: s_cselect_b64 s{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}
-; GCN-NEXT: s_cselect_b64 s{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}
+; GCN-NEXT: s_cselect_b32 s
+; GCN-NEXT: s_cselect_b32 s
+; GCN-NEXT: s_cselect_b32 s
 define amdgpu_kernel void @test_scc1_sgpr_ifcvt_triangle96(<3 x i32> addrspace(4)* %in, i32 %cond) #0 {
 entry:
   %v = load <3 x i32>, <3 x i32> addrspace(4)* %in
   %cc = icmp eq i32 %cond, 1
   br i1 %cc, label %if, label %endif
 
 if:
   %u = add <3 x i32> %v, %v

llvm/trunk/test/CodeGen/AMDGPU/function-args.ll

 ; GCN-NOT: v1
 ; GCN: buffer_store_dwordx2 v[0:1], off
 define void @void_func_v2i32(<2 x i32> %arg0) #0 {
   store <2 x i32> %arg0, <2 x i32> addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}void_func_v3i32:
-; GCN-DAG: buffer_store_dword v2, off
-; GCN-DAG: buffer_store_dwordx2 v[0:1], off
+; GCN-DAG: buffer_store_dwordx3 v[0:2], off
 define void @void_func_v3i32(<3 x i32> %arg0) #0 {
   store <3 x i32> %arg0, <3 x i32> addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}void_func_v4i32:
 ; GCN: buffer_store_dwordx4 v[0:3], off
 define void @void_func_v4i32(<4 x i32> %arg0) #0 {

 ; GCN-NOT: v1
 ; GCN: buffer_store_dwordx2 v[0:1], off
 define void @void_func_v2f32(<2 x float> %arg0) #0 {
   store <2 x float> %arg0, <2 x float> addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}void_func_v3f32:
-; GCN-DAG: buffer_store_dword v2, off
-; GCN-DAG: buffer_store_dwordx2 v[0:1], off
+; GCN-DAG: buffer_store_dwordx3 v[0:2], off
 define void @void_func_v3f32(<3 x float> %arg0) #0 {
   store <3 x float> %arg0, <3 x float> addrspace(1)* undef
   ret void
 }
 
 ; GCN-LABEL: {{^}}void_func_v4f32:
 ; GCN: buffer_store_dwordx4 v[0:3], off
 define void @void_func_v4f32(<4 x float> %arg0) #0 {

llvm/trunk/test/CodeGen/AMDGPU/function-returns.ll

 ; GCN-NEXT: s_waitcnt vmcnt(0)
 ; GCN-NEXT: s_setpc_b64
 define <2 x i32> @v2i32_func_void() #0 {
   %val = load <2 x i32>, <2 x i32> addrspace(1)* undef
   ret <2 x i32> %val
 }
 
 ; GCN-LABEL: {{^}}v3i32_func_void:
-; GCN: buffer_load_dwordx4 v[0:3], off
+; GCN: buffer_load_dwordx3 v[0:2], off
 ; GCN-NEXT: s_waitcnt vmcnt(0)
 ; GCN-NEXT: s_setpc_b64
 define <3 x i32> @v3i32_func_void() #0 {
   %val = load <3 x i32>, <3 x i32> addrspace(1)* undef
   ret <3 x i32> %val
 }
 
 ; GCN-LABEL: {{^}}v4i32_func_void:

llvm/trunk/test/CodeGen/AMDGPU/half.ll

 define amdgpu_kernel void @extload_v2f16_to_v2f32_arg(<2 x float> addrspace(1)* %out, <2 x half> %arg) #0 {
   %ext = fpext <2 x half> %arg to <2 x float>
   store <2 x float> %ext, <2 x float> addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}extload_v3f16_to_v3f32_arg:
 ; GCN: s_load_dwordx2 s
-; GCN: s_load_dwordx2 s
-; GCN-NOT: _load
 ; GCN: v_cvt_f32_f16_e32
 ; GCN: v_cvt_f32_f16_e32
+; GCN: s_load_dwordx2 s
+; GCN-NOT: _load
 ; GCN: v_cvt_f32_f16_e32
 ; GCN-NOT: v_cvt_f32_f16
-; GCN-DAG: _store_dword
-; GCN-DAG: _store_dwordx2
+; GCN-DAG: _store_dwordx3
 ; GCN: s_endpgm
 define amdgpu_kernel void @extload_v3f16_to_v3f32_arg(<3 x float> addrspace(1)* %out, <3 x half> %arg) #0 {
   %ext = fpext <3 x half> %arg to <3 x float>
   store <3 x float> %ext, <3 x float> addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}extload_v4f16_to_v4f32_arg:

 define amdgpu_kernel void @global_truncstore_v2f32_to_v2f16(<2 x half> addrspace(1)* %out, <2 x float> addrspace(1)* %in) #0 {
   %val = load <2 x float>, <2 x float> addrspace(1)* %in
   %cvt = fptrunc <2 x float> %val to <2 x half>
   store <2 x half> %cvt, <2 x half> addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}global_truncstore_v3f32_to_v3f16:
-; GCN: flat_load_dwordx4
+; GCN: flat_load_dwordx3
 ; GCN-DAG: v_cvt_f16_f32_e32
 ; SI-DAG:  v_cvt_f16_f32_e32
 ; VI-DAG:  v_cvt_f16_f32_sdwa
 ; GCN-DAG: v_cvt_f16_f32_e32
 ; GCN: flat_store_short
 ; GCN: flat_store_dword
 ; GCN: s_endpgm
 define amdgpu_kernel void @global_truncstore_v3f32_to_v3f16(<3 x half> addrspace(1)* %out, <3 x float> addrspace(1)* %in) #0 {

llvm/trunk/test/CodeGen/AMDGPU/idot4u.ll

 ; GFX7-NEXT:    v_add_i32_e32 v0, vcc, v4, v0
 ; GFX7-NEXT:    buffer_store_byte v0, off, s[0:3], 0
 ; GFX7-NEXT:    s_endpgm
 ;
 ; GFX8-LABEL: udot4_acc8_vecMul:
 ; GFX8:       ; %bb.0: ; %entry
 ; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    s_movk_i32 s8, 0xff
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX8-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX8-NEXT:    v_mov_b32_e32 v1, s1
 ; GFX8-NEXT:    flat_load_ubyte v2, v[0:1]
-; GFX8-NEXT:    s_load_dword s1, s[4:5], 0x0
-; GFX8-NEXT:    s_load_dword s2, s[6:7], 0x0
-; GFX8-NEXT:    s_movk_i32 s0, 0xff
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX8-NEXT:    s_load_dword s1, s[6:7], 0x0
 ; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s2, s0, 24
 ; GFX8-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX8-NEXT:    s_lshr_b32 s4, s2, 24
-; GFX8-NEXT:    s_and_b32 s6, s1, s0
-; GFX8-NEXT:    s_bfe_u32 s7, s2, 0x80010
-; GFX8-NEXT:    s_and_b32 s0, s2, s0
-; GFX8-NEXT:    v_mov_b32_e32 v3, s1
-; GFX8-NEXT:    v_mov_b32_e32 v4, s2
+; GFX8-NEXT:    s_bfe_u32 s6, s1, 0x80010
+; GFX8-NEXT:    s_and_b32 s7, s1, s8
+; GFX8-NEXT:    v_mov_b32_e32 v3, s0
+; GFX8-NEXT:    v_mov_b32_e32 v4, s1
 ; GFX8-NEXT:    v_mul_u32_u24_sdwa v3, v3, v4 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
-; GFX8-NEXT:    v_mov_b32_e32 v4, s0
-; GFX8-NEXT:    s_bfe_u32 s5, s1, 0x80010
-; GFX8-NEXT:    v_mov_b32_e32 v5, s7
-; GFX8-NEXT:    v_mov_b32_e32 v6, s4
-; GFX8-NEXT:    v_mov_b32_e32 v7, s3
-; GFX8-NEXT:    v_mul_u32_u24_e32 v4, s6, v4
-; GFX8-NEXT:    v_mul_u32_u24_e32 v5, s5, v5
+; GFX8-NEXT:    s_bfe_u32 s4, s0, 0x80010
+; GFX8-NEXT:    v_mov_b32_e32 v5, s6
+; GFX8-NEXT:    s_and_b32 s5, s0, s8
+; GFX8-NEXT:    v_mov_b32_e32 v4, s7
+; GFX8-NEXT:    v_mov_b32_e32 v6, s3
+; GFX8-NEXT:    v_mov_b32_e32 v7, s2
+; GFX8-NEXT:    v_mul_u32_u24_e32 v4, s5, v4
+; GFX8-NEXT:    v_mul_u32_u24_e32 v5, s4, v5
 ; GFX8-NEXT:    v_mul_u32_u24_sdwa v6, v7, v6 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX8-NEXT:    v_or_b32_sdwa v5, v5, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX8-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
 ; GFX8-NEXT:    v_or_b32_sdwa v3, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 8, v3
 ; GFX8-NEXT:    s_waitcnt vmcnt(0)
 ; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
 ; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v4, v2

llvm/trunk/test/CodeGen/AMDGPU/load-global-f32.ll

-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=FUNC %s
-; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn-amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=FUNC %s
-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=FUNC %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=FUNC -check-prefix=SI-NOHSA %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn-amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=FUNC -check-prefix=GCNX3-HSA %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=FUNC -check-prefix=GCNX3-NOHSA %s
 
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=redwood < %s | FileCheck -check-prefix=R600 -check-prefix=FUNC %s
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=cayman < %s | FileCheck -check-prefix=R600 -check-prefix=FUNC %s
 
 ; FUNC-LABEL: {{^}}global_load_f32:
 ; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}
 ; GCN-HSA: flat_load_dword
 

 define amdgpu_kernel void @global_load_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %in) #0 {
 entry:
   %tmp0 = load <2 x float>, <2 x float> addrspace(1)* %in
   store <2 x float> %tmp0, <2 x float> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}global_load_v3f32:
-; GCN-NOHSA: buffer_load_dwordx4
-; GCN-HSA: flat_load_dwordx4
+; SI-NOHSA: buffer_load_dwordx4
+; GCNX3-NOHSA: buffer_load_dwordx3
+; GCNX3-HSA: flat_load_dwordx3
 
 ; R600: VTX_READ_128
 define amdgpu_kernel void @global_load_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %in) #0 {
 entry:
   %tmp0 = load <3 x float>, <3 x float> addrspace(1)* %in
   store <3 x float> %tmp0, <3 x float> addrspace(1)* %out
   ret void
 }

llvm/trunk/test/CodeGen/AMDGPU/load-global-i32.ll

-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=FUNC %s
-; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn--amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=FUNC %s
-; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=FUNC %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=SI-NOHSA -check-prefix=FUNC %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn--amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=GCNX3-HSA -check-prefix=FUNC %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-NOHSA -check-prefix=GCNX3-NOHSA -check-prefix=FUNC %s
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=redwood < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=EG -check-prefix=FUNC %s
-; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn--amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=FUNC %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -mtriple=amdgcn--amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=GCN -check-prefix=GCN-HSA -check-prefix=GCNX3-HSA -check-prefix=FUNC %s
 
 
 ; FUNC-LABEL: {{^}}global_load_i32:
 ; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}
 ; GCN-HSA: {{flat|global}}_load_dword
 
 ; EG: VTX_READ_32 T{{[0-9]+}}.X, T{{[0-9]+}}.X, 0
 define amdgpu_kernel void @global_load_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) #0 {

 define amdgpu_kernel void @global_load_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in) #0 {
 entry:
   %ld = load <2 x i32>, <2 x i32> addrspace(1)* %in
   store <2 x i32> %ld, <2 x i32> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}global_load_v3i32:
-; GCN-NOHSA: buffer_load_dwordx4
-; GCN-HSA: {{flat|global}}_load_dwordx4
+; SI-NOHSA: buffer_load_dwordx4
+; GCNX3-NOHSA: buffer_load_dwordx3
+; GCNX3-HSA: {{flat|global}}_load_dwordx3
 
 ; EG: VTX_READ_128
 define amdgpu_kernel void @global_load_v3i32(<3 x i32> addrspace(1)* %out, <3 x i32> addrspace(1)* %in) #0 {
 entry:
   %ld = load <3 x i32>, <3 x i32> addrspace(1)* %in
   store <3 x i32> %ld, <3 x i32> addrspace(1)* %out
   ret void
 }

llvm/trunk/test/CodeGen/AMDGPU/mad-mix-lo.ll

define <2 x half> @v_mad_mix_v2f32_clamp_postcvt(<2 x half> %src0, <2 x half> %src1, <2 x half> %src2) #0 {
   %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %cvt.result, <2 x half> zeroinitializer)
   %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
   ret <2 x half> %clamp
 }
 
 ; FIXME: Should be packed into 2 registers per argument?
 ; GCN-LABEL: {{^}}v_mad_mix_v3f32_clamp_postcvt:
 ; GCN: s_waitcnt
-; GFX9-NEXT: v_mad_mixlo_f16 v6, v0, v2, v4 op_sel_hi:[1,1,1] clamp
-; GFX9-NEXT: v_mad_mixhi_f16 v6, v0, v2, v4 op_sel:[1,1,1] op_sel_hi:[1,1,1] clamp
-; GFX9-NEXT: v_mad_mixlo_f16 v2, v1, v3, v5 op_sel_hi:[1,1,1] clamp
-; GFX9-NEXT: v_mad_mixhi_f16 v2, v1, v3, v5 op_sel:[1,1,1] op_sel_hi:[1,1,1] clamp
-; GFX9-NEXT: v_mov_b32_e32 v0, v6
-; GFX9-NEXT: v_mov_b32_e32 v1, v2
+; GFX9-DAG: v_mad_mixlo_f16 v{{[0-9]+}}, v0, v2, v4 op_sel_hi:[1,1,1] clamp
+; GFX9-DAG: v_mad_mixhi_f16 v{{[0-9]+}}, v0, v2, v4 op_sel:[1,1,1] op_sel_hi:[1,1,1] clamp
+; GFX9-DAG: v_mad_mixlo_f16 v{{[0-9]+}}, v1, v3, v5 op_sel_hi:[1,1,1]
+; GFX9-DAG: v_pk_max_f16 v1, v1, v1 clamp
+; GFX9: v_mov_b32_e32 v0, v{{[0-9]+}}
 ; GFX9-NEXT: s_setpc_b64
 define <3 x half> @v_mad_mix_v3f32_clamp_postcvt(<3 x half> %src0, <3 x half> %src1, <3 x half> %src2) #0 {
   %src0.ext = fpext <3 x half> %src0 to <3 x float>
   %src1.ext = fpext <3 x half> %src1 to <3 x float>
   %src2.ext = fpext <3 x half> %src2 to <3 x float>
   %result = tail call <3 x float> @llvm.fmuladd.v3f32(<3 x float> %src0.ext, <3 x float> %src1.ext, <3 x float> %src2.ext)
   %cvt.result = fptrunc <3 x float> %result to <3 x half>
   %max = call <3 x half> @llvm.maxnum.v3f16(<3 x half> %cvt.result, <3 x half> zeroinitializer)

llvm/trunk/test/CodeGen/AMDGPU/merge-stores.ll

define amdgpu_kernel void @merge_global_store_4_adjacent_loads_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) #0 {
   store i32 %x, i32 addrspace(1)* %out
   store i32 %y, i32 addrspace(1)* %out.gep.1
   store i32 %z, i32 addrspace(1)* %out.gep.2
   store i32 %w, i32 addrspace(1)* %out.gep.3
   ret void
 }
 
 ; GCN-LABEL: {{^}}merge_global_store_3_adjacent_loads_i32:
-; SI-DAG: buffer_load_dwordx2
-; SI-DAG: buffer_load_dword v
+; SI-DAG: buffer_load_dwordx4
 ; CI-DAG: buffer_load_dwordx3
 ; GCN: s_waitcnt
 ; SI-DAG: buffer_store_dwordx2
 ; SI-DAG: buffer_store_dword v
 ; CI-DAG: buffer_store_dwordx3
 ; GCN: s_endpgm
 define amdgpu_kernel void @merge_global_store_3_adjacent_loads_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) #0 {
   %out.gep.1 = getelementptr i32, i32 addrspace(1)* %out, i32 1
define amdgpu_kernel void @merge_global_store_6_constants_i32(i32 addrspace(1)* %out) {
   store i32 11, i32 addrspace(1)* %idx4, align 4
   %idx5 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 5
   store i32 123, i32 addrspace(1)* %idx5, align 4
   ret void
 }
 
 ; GCN-LABEL: {{^}}merge_global_store_7_constants_i32:
 ; GCN: buffer_store_dwordx4
-; SI: buffer_store_dwordx2
-; SI: buffer_store_dword v
+; SI-DAG: buffer_store_dwordx2
+; SI-DAG: buffer_store_dword v
 ; CI: buffer_store_dwordx3
 define amdgpu_kernel void @merge_global_store_7_constants_i32(i32 addrspace(1)* %out) {
   store i32 34, i32 addrspace(1)* %out, align 4
   %idx1 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 1
   store i32 999, i32 addrspace(1)* %idx1, align 4
   %idx2 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 2
   store i32 65, i32 addrspace(1)* %idx2, align 4
   %idx3 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 3

 }
 
 ; This requires handling of scalar_to_vector for v2i64 to avoid
 ; scratch usage.
 ; FIXME: Should do single load and store
 
 ; GCN-LABEL: {{^}}copy_v3i32_align4:
 ; GCN-NOT: SCRATCH_RSRC_DWORD
-; SI-DAG: buffer_load_dword v{{[0-9]+}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:8
-; SI-DAG: buffer_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
+; SI-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; CI-DAG: buffer_load_dwordx3 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; GCN-NOT: offen
 ; GCN: s_waitcnt vmcnt
 ; GCN-NOT: offen
 ; SI-DAG: buffer_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; SI-DAG: buffer_store_dword v{{[0-9]+}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:8
 ; CI-DAG: buffer_store_dwordx3 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 

 define amdgpu_kernel void @copy_v3i64_align4(<3 x i64> addrspace(1)* noalias %out, <3 x i64> addrspace(1)* noalias %in) #0 {
   %vec = load <3 x i64>, <3 x i64> addrspace(1)* %in, align 4
   store <3 x i64> %vec, <3 x i64> addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}copy_v3f32_align4:
 ; GCN-NOT: SCRATCH_RSRC_DWORD
-; SI-DAG: buffer_load_dword v{{[0-9]+}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:8
-; SI-DAG: buffer_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
+; SI-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; CI-DAG: buffer_load_dwordx3 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; GCN-NOT: offen
 ; GCN: s_waitcnt vmcnt
 ; GCN-NOT: offen
 ; SI-DAG: buffer_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; SI-DAG: buffer_store_dword v{{[0-9]+}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:8
 ; CI-DAG: buffer_store_dwordx3 v{{\[[0-9]+:[0-9]+\]}}, off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
 ; GCN: ScratchSize: 0{{$}}

llvm/trunk/test/CodeGen/AMDGPU/multi-dword-vgpr-spill.ll

+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -enable-misched=0 -post-RA-scheduler=0 -stress-regalloc=8 < %s | FileCheck %s
+
+; CHECK-LABEL: spill_v2i32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:24 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:28 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:24 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:28 ; 4-byte Folded Reload
+
+define void @spill_v2i32() {
+entry:
+  %alloca = alloca <2 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <2 x i32> %a, <2 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+; CHECK-LABEL: spill_v2f32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:24 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:28 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:24 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:28 ; 4-byte Folded Reload
+
+define void @spill_v2f32() {
+entry:
+  %alloca = alloca <2 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <2 x i32> %a, <2 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+; CHECK-LABEL: spill_v3i32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:48 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:52 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:56 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:48 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:52 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:56 ; 4-byte Folded Reload
+
+define void @spill_v3i32() {
+entry:
+  %alloca = alloca <3 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <3 x i32>, <3 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <3 x i32>, <3 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <3 x i32>, <3 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <3 x i32> %a, <3 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+; CHECK-LABEL: spill_v3f32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:48 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:52 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:56 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:48 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:52 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:56 ; 4-byte Folded Reload
+
+define void @spill_v3f32() {
+entry:
+  %alloca = alloca <3 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <3 x i32>, <3 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <3 x i32>, <3 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <3 x i32>, <3 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <3 x i32> %a, <3 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+; CHECK-LABEL: spill_v4i32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:48 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:52 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:56 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:60 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:48 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:52 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:56 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:60 ; 4-byte Folded Reload
+
+define void @spill_v4i32() {
+entry:
+  %alloca = alloca <4 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <4 x i32>, <4 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <4 x i32>, <4 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <4 x i32>, <4 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <4 x i32> %a, <4 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+; CHECK-LABEL: spill_v4f32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:48 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:52 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:56 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:60 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:48 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:52 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:56 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:60 ; 4-byte Folded Reload
+
+define void @spill_v4f32() {
+entry:
+  %alloca = alloca <4 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <4 x i32>, <4 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <4 x i32>, <4 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <4 x i32>, <4 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <4 x i32> %a, <4 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+; CHECK-LABEL: spill_v5i32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:96 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:100 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:104 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:108 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:96 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:100 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:104 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:108 ; 4-byte Folded Reload
+
+define void @spill_v5i32() {
+entry:
+  %alloca = alloca <5 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <5 x i32>, <5 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <5 x i32>, <5 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <5 x i32>, <5 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <5 x i32> %a, <5 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+; CHECK-LABEL: spill_v5f32:
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:96 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:100 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:104 ; 4-byte Folded Spill
+; CHECK-DAG: buffer_store_dword v{{.*}} offset:108 ; 4-byte Folded Spill
+; CHECK: ;;#ASMSTART
+; CHECK-NEXT: ;;#ASMEND
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:96 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:100 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:104 ; 4-byte Folded Reload
+; CHECK-DAG: buffer_load_dword v{{.*}} offset:108 ; 4-byte Folded Reload
+
+define void @spill_v5f32() {
+entry:
+  %alloca = alloca <5 x i32>, i32 2, align 4, addrspace(5)
+
+  %aptr = getelementptr <5 x i32>, <5 x i32> addrspace(5)* %alloca, i32 1
+  %a = load volatile <5 x i32>, <5 x i32> addrspace(5)* %aptr
+
+  ; Force %a to spill.
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
+
+  %outptr = getelementptr <5 x i32>, <5 x i32> addrspace(5)* %alloca, i32 1
+  store volatile <5 x i32> %a, <5 x i32> addrspace(5)* %outptr
+
+  ret void
+}
+
+
+

llvm/trunk/test/CodeGen/AMDGPU/select-vectors.ll

 
 ; GCN-LABEL: {{^}}v_select_v3i16:
 ; SI: v_cndmask_b32_e32
 ; SI: cndmask
 ; SI-NOT: cndmask
 
 ; GFX89: v_cndmask_b32_e32
 ; GFX89: cndmask
+; VI: cndmask
 ; GFX89-NOT: cndmask
 define amdgpu_kernel void @v_select_v3i16(<3 x i16> addrspace(1)* %out, <3 x i16> addrspace(1)* %a.ptr, <3 x i16> addrspace(1)* %b.ptr, i32 %c) #0 {
   %a = load <3 x i16>, <3 x i16> addrspace(1)* %a.ptr
   %b = load <3 x i16>, <3 x i16> addrspace(1)* %b.ptr
   %cmp = icmp eq i32 %c, 0
   %select = select i1 %cmp, <3 x i16> %a, <3 x i16> %b
   store <3 x i16> %select, <3 x i16> addrspace(1)* %out, align 4
   ret void

 ; GCN: buffer_store_dwordx2
 define amdgpu_kernel void @s_select_v2f32(<2 x float> addrspace(1)* %out, <2 x float> %a, <2 x float> %b, i32 %c) #0 {
   %cmp = icmp eq i32 %c, 0
   %select = select i1 %cmp, <2 x float> %a, <2 x float> %b
   store <2 x float> %select, <2 x float> addrspace(1)* %out, align 16
   ret void
 }
 
+; GCN-LABEL: {{^}}s_select_v3f32:
+; GCN: v_cmp_eq_u32_e64 vcc, s{{[0-9]+}}, 0{{$}}
+
+; GCN: v_cndmask_b32_e32
+; GCN: v_cndmask_b32_e32
+; GCN: v_cndmask_b32_e32
+
+; GCN: buffer_store_dwordx
+define amdgpu_kernel void @s_select_v3f32(<3 x float> addrspace(1)* %out, <3 x float> %a, <3 x float> %b, i32 %c) #0 {
+  %cmp = icmp eq i32 %c, 0
+  %select = select i1 %cmp, <3 x float> %a, <3 x float> %b
+  store <3 x float> %select, <3 x float> addrspace(1)* %out, align 16
+  ret void
+}
+
 ; GCN-LABEL: {{^}}s_select_v4f32:
 ; GCN: s_load_dwordx4
 ; GCN: s_load_dwordx4
 ; GCN: v_cmp_eq_u32_e64 vcc, s{{[0-9]+}}, 0{{$}}
 
 ; GCN: v_cndmask_b32_e32
 ; GCN: v_cndmask_b32_e32
 ; GCN: v_cndmask_b32_e32

llvm/trunk/test/CodeGen/AMDGPU/shader-addr64-nonuniform.ll

-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx700 -verify-machineinstrs <%s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SICI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx600 -verify-machineinstrs <%s | FileCheck -enable-var-scope -check-prefixes=GCN,SICI,SI %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx700 -verify-machineinstrs <%s | FileCheck -enable-var-scope -check-prefixes=GCN,SICI,CI %s
 
 ; Check that an addrspace(1) (const) load with various combinations of
 ; uniform, nonuniform and constant address components all load with an
 ; addr64 mubuf with no readfirstlane.
 
 @indexable = internal unnamed_addr addrspace(1) constant [6 x <3 x float>] [<3 x float> <float 1.000000e+00, float 0.000000e+00, float 0.000000e+00>, <3 x float> <float 0.000000e+00, float 1.000000e+00, float 0.000000e+00>, <3 x float> <float 0.000000e+00, float 0.000000e+00, float 1.000000e+00>, <3 x float> <float 0.000000e+00, float 1.000000e+00, float 1.000000e+00>, <3 x float> <float 1.000000e+00, float 0.000000e+00, float 1.000000e+00>, <3 x float> <float 1.000000e+00, float 1.000000e+00, float 0.000000e+00>]
 
 ; GCN-LABEL: {{^}}nonuniform_uniform:
 ; GCN-NOT: readfirstlane
-; SICI: buffer_load_dwordx4 {{.*}} addr64
+; SI: buffer_load_dwordx4 {{.*}} addr64
+; CI: buffer_load_dwordx3 {{.*}} addr64
 
 define amdgpu_ps float @nonuniform_uniform(i32 %arg18) {
 .entry:
   %tmp31 = sext i32 %arg18 to i64
   %tmp32 = getelementptr [6 x <3 x float>], [6 x <3 x float>] addrspace(1)* @indexable, i64 0, i64 %tmp31
   %tmp33 = load <3 x float>, <3 x float> addrspace(1)* %tmp32, align 16
   %tmp34 = extractelement <3 x float> %tmp33, i32 0
   ret float %tmp34
 }
 
 ; GCN-LABEL: {{^}}uniform_nonuniform:
 ; GCN-NOT: readfirstlane
-; SICI: buffer_load_dwordx4 {{.*}} addr64
+; SI: buffer_load_dwordx4 {{.*}} addr64
+; CI: buffer_load_dwordx3 {{.*}} addr64
 
 define amdgpu_ps float @uniform_nonuniform(i32 inreg %offset, i32 %arg18) {
 .entry:
   %tmp1 = zext i32 %arg18 to i64
   %tmp2 = inttoptr i64 %tmp1 to [6 x <3 x float>] addrspace(1)*
   %tmp32 = getelementptr [6 x <3 x float>], [6 x <3 x float>] addrspace(1)* %tmp2, i32 0, i32 %offset
   %tmp33 = load <3 x float>, <3 x float> addrspace(1)* %tmp32, align 16
   %tmp34 = extractelement <3 x float> %tmp33, i32 0
   ret float %tmp34
 }
 
 ; GCN-LABEL: {{^}}const_nonuniform:
 ; GCN-NOT: readfirstlane
-; SICI: buffer_load_dwordx4 {{.*}} addr64
+; SI: buffer_load_dwordx4 {{.*}} addr64
+; CI: buffer_load_dwordx3 {{.*}} addr64
 
 define amdgpu_ps float @const_nonuniform(i32 %arg18) {
 .entry:
   %tmp1 = zext i32 %arg18 to i64
   %tmp2 = inttoptr i64 %tmp1 to [6 x <3 x float>] addrspace(1)*
   %tmp32 = getelementptr [6 x <3 x float>], [6 x <3 x float>] addrspace(1)* %tmp2, i32 0, i32 1
   %tmp33 = load <3 x float>, <3 x float> addrspace(1)* %tmp32, align 16
   %tmp34 = extractelement <3 x float> %tmp33, i32 0
   ret float %tmp34
 }
 
 ; GCN-LABEL: {{^}}nonuniform_nonuniform:
 ; GCN-NOT: readfirstlane
-; SICI: buffer_load_dwordx4 {{.*}} addr64
+; SI: buffer_load_dwordx4 {{.*}} addr64
+; CI: buffer_load_dwordx3 {{.*}} addr64
 
 define amdgpu_ps float @nonuniform_nonuniform(i32 %offset, i32 %arg18) {
 .entry:
   %tmp1 = zext i32 %arg18 to i64
   %tmp2 = inttoptr i64 %tmp1 to [6 x <3 x float>] addrspace(1)*
   %tmp32 = getelementptr [6 x <3 x float>], [6 x <3 x float>] addrspace(1)* %tmp2, i32 0, i32 %offset
   %tmp33 = load <3 x float>, <3 x float> addrspace(1)* %tmp32, align 16
   %tmp34 = extractelement <3 x float> %tmp33, i32 0

llvm/trunk/test/CodeGen/AMDGPU/sign_extend.ll

; VI-NEXT:    s_endpgm
   %sext = sext i1 %cmp to i32
   store i32 %sext, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 define amdgpu_kernel void @test_s_sext_i32_to_i64(i64 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) nounwind {
 ; SI-LABEL: test_s_sext_i32_to_i64:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xb
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0xb
 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mul_i32 s4, s4, s5
-; SI-NEXT:    s_add_i32 s4, s4, s2
+; SI-NEXT:    s_mul_i32 s2, s4, s5
+; SI-NEXT:    s_add_i32 s4, s2, s6
 ; SI-NEXT:    s_ashr_i32 s5, s4, 31
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    v_mov_b32_e32 v1, s5
 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: test_s_sext_i32_to_i64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; VI-NEXT:    s_load_dword s0, s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x2c
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mul_i32 s1, s2, s3
-; VI-NEXT:    s_add_i32 s1, s1, s0
-; VI-NEXT:    s_ashr_i32 s0, s1, 31
-; VI-NEXT:    v_mov_b32_e32 v0, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s0
+; VI-NEXT:    s_mul_i32 s0, s0, s1
+; VI-NEXT:    s_add_i32 s0, s0, s2
+; VI-NEXT:    s_ashr_i32 s1, s0, 31
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 entry:
   %mul = mul i32 %a, %b
   %add = add i32 %mul, %c
   %sext = sext i32 %add to i64
   store i64 %sext, i64 addrspace(1)* %out, align 8
   ret void
; VI-NEXT:    s_endpgm
   store i16 %sext, i16 addrspace(1)* %out
   ret void
 }
 
 define amdgpu_kernel void @v_sext_i1_to_i16_with_and(i16 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) nounwind {
 ; SI-LABEL: v_sext_i1_to_i16_with_and:
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xb
-; SI-NEXT:    s_load_dword s0, s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0xb
 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 ; SI-NEXT:    s_mov_b32 s6, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_cmp_eq_u32_e32 vcc, s2, v0
-; SI-NEXT:    v_mov_b32_e32 v0, s0
-; SI-NEXT:    v_cmp_eq_u32_e64 s[0:1], s3, v0
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, s0, v0
+; SI-NEXT:    v_mov_b32_e32 v0, s2
+; SI-NEXT:    v_cmp_eq_u32_e64 s[0:1], s1, v0
 ; SI-NEXT:    s_and_b64 s[0:1], vcc, s[0:1]
 ; SI-NEXT:    v_cndmask_b32_e64 v0, 0, -1, s[0:1]
 ; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_sext_i1_to_i16_with_and:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; VI-NEXT:    s_load_dword s0, s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x2c
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_cmp_eq_u32_e32 vcc, s2, v0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_cmp_eq_u32_e64 s[0:1], s3, v0
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, s0, v0
+; VI-NEXT:    v_cmp_eq_u32_e64 s[0:1], s1, v1
 ; VI-NEXT:    s_and_b64 s[0:1], vcc, s[0:1]
 ; VI-NEXT:    v_cndmask_b32_e64 v0, 0, -1, s[0:1]
 ; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
   %tid = tail call i32 @llvm.amdgcn.workitem.id.x() #1
   %cmp0 = icmp eq i32 %a, %tid
   %cmp1 = icmp eq i32 %b, %c
   %cmp = and i1 %cmp0, %cmp1

llvm/trunk/test/CodeGen/AMDGPU/spill-wide-sgpr.ll

 bb0:
   call void asm sideeffect "; use $0", "s"(<2 x i32> %wide.sgpr) #0
   br label %ret
 
 ret:
   ret void
 }
 
+; ALL-LABEL: {{^}}spill_sgpr_x3:
+; SMEM: s_add_u32 m0, s3, 0x100{{$}}
+; SMEM: s_buffer_store_dword s
+; SMEM: s_buffer_store_dword s
+; SMEM: s_buffer_store_dword s
+; SMEM: s_cbranch_scc1
+
+; SMEM: s_add_u32 m0, s3, 0x100{{$}}
+; SMEM: s_buffer_load_dword s
+; SMEM: s_buffer_load_dword s
+; SMEM: s_buffer_load_dword s
+; SMEM: s_dcache_wb
+; SMEM: s_endpgm
+
+; FIXME: Should only need 4 bytes
+; SMEM: ScratchSize: 16
+
+; VGPR: v_writelane_b32 v{{[0-9]+}}, s{{[0-9]+}}, 0
+; VGPR: v_writelane_b32 v{{[0-9]+}}, s{{[0-9]+}}, 1
+; VGPR: v_writelane_b32 v{{[0-9]+}}, s{{[0-9]+}}, 2
+; VGPR: s_cbranch_scc1
+
+; VGPR: v_readlane_b32 s{{[0-9]+}}, v{{[0-9]+}}, 0
+; VGPR: v_readlane_b32 s{{[0-9]+}}, v{{[0-9]+}}, 1
+; VGPR: v_readlane_b32 s{{[0-9]+}}, v{{[0-9]+}}, 2
+
+
+; VMEM: buffer_store_dword
+; VMEM: buffer_store_dword
+; VMEM: buffer_store_dword
+; VMEM: s_cbranch_scc1
+
+; VMEM: buffer_load_dword
+; VMEM: buffer_load_dword
+; VMEM: buffer_load_dword
+define amdgpu_kernel void @spill_sgpr_x3(i32 addrspace(1)* %out, i32 %in) #0 {
+  %wide.sgpr = call <3 x i32>  asm sideeffect "; def $0", "=s" () #0
+  %cmp = icmp eq i32 %in, 0
+  br i1 %cmp, label %bb0, label %ret
+
+bb0:
+  call void asm sideeffect "; use $0", "s"(<3 x i32> %wide.sgpr) #0
+  br label %ret
+
+ret:
+  ret void
+}
+
 ; ALL-LABEL: {{^}}spill_sgpr_x4:
 ; SMEM: s_add_u32 m0, s3, 0x100{{$}}
 ; SMEM: s_buffer_store_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[VALS:[0-9]+:[0-9]+]]{{\]}}, m0 ; 16-byte Folded Spill
 ; SMEM: s_cbranch_scc1
 
 ; SMEM: s_add_u32 m0, s3, 0x100{{$}}
 ; SMEM: s_buffer_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s{{\[}}[[VALS]]{{\]}}, m0 ; 16-byte Folded Reload
 ; SMEM: s_dcache_wb

llvm/trunk/test/CodeGen/AMDGPU/store-global.ll

 entry:
   %0 = insertelement <2 x float> <float 0.0, float 0.0>, float %a, i32 0
   %1 = insertelement <2 x float> %0, float %b, i32 1
   store <2 x float> %1, <2 x float> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}store_v3i32:
-; SI-DAG: buffer_store_dwordx2
 ; SI-DAG: buffer_store_dword v
+; SI-DAG: buffer_store_dwordx2
 
-; VI-DAG: buffer_store_dwordx3
+; VI: buffer_store_dwordx3
 
-; GFX9-DAG: global_store_dwordx2
-; GFX9-DAG: global_store_dword v
+; GFX9: global_store_dwordx3
 
 ; EG-DAG: MEM_RAT_CACHELESS STORE_RAW {{T[0-9]+\.[XYZW]}}, {{T[0-9]+\.[XYZW]}},
 ; EG-DAG: MEM_RAT_CACHELESS STORE_RAW {{T[0-9]+\.XY}}, {{T[0-9]+\.[XYZW]}},
 define amdgpu_kernel void @store_v3i32(<3 x i32> addrspace(1)* %out, <3 x i32> %a) nounwind {
   store <3 x i32> %a, <3 x i32> addrspace(1)* %out, align 16
   ret void
 }
 

llvm/trunk/test/CodeGen/AMDGPU/v_mac.ll

entry:
   %tmp1 = fadd float %tmp0, %c
 
   store float %tmp1, float addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}safe_mad_sub0_src0:
 ; GCN: v_sub_f32_e32 [[SUB0:v[0-9]+]], 0,
-; GCN: v_mac_f32_e32 v{{[0-9]+}}, [[SUB0]], v{{[0-9]+}}
+; GCN: v_ma{{[cd]}}_f32{{[_e32]*}} v{{[0-9]+}}, [[SUB0]], v{{[0-9]+}}
 define amdgpu_kernel void @safe_mad_sub0_src0(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
 entry:
   %b_ptr = getelementptr float, float addrspace(1)* %in, i32 1
   %c_ptr = getelementptr float, float addrspace(1)* %in, i32 2
 
   %a = load float, float addrspace(1)* %in
   %b = load float, float addrspace(1)* %b_ptr
   %c = load float, float addrspace(1)* %c_ptr

llvm/trunk/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll

 ; HSA: enable_sgpr_private_segment_buffer = 1
 ; HSA: enable_sgpr_flat_scratch_init = 0
 ; HSA: workitem_private_segment_byte_size = 1536
 
 ; GCN-NOT: flat_scr
 ; MESA-NOT: s_mov_b32 s3
 ; HSA-NOT: s_mov_b32 s7
 
-; GCNMESA-DAG: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
-; GCNMESA-DAG: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
-; GCNMESA-DAG: s_mov_b32 s14, -1
-; SIMESA-DAG: s_mov_b32 s15, 0xe8f000
-; VIMESA-DAG: s_mov_b32 s15, 0xe80000
-; GFX9MESA-DAG: s_mov_b32 s15, 0xe00000
-
-
-; GCNMESAMESA: buffer_store_dword {{v[0-9]+}}, off, s[12:15], s3 offset:{{[0-9]+}} ; 4-byte Folded Spill
-
-; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[12:15], s3 offset:{{[0-9]+}}
-; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[12:15], s3 offset:{{[0-9]+}}
-; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[12:15], s3 offset:{{[0-9]+}}
-; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[12:15], s3 offset:{{[0-9]+}}
-
-; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[12:15], s3 offset:{{[0-9]+}}
-; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[12:15], s3 offset:{{[0-9]+}}
-; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[12:15], s3 offset:{{[0-9]+}}
-; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[12:15], s3 offset:{{[0-9]+}}
+; GCNMESA-DAG: s_mov_b32 s16, SCRATCH_RSRC_DWORD0
+; GCNMESA-DAG: s_mov_b32 s17, SCRATCH_RSRC_DWORD1
+; GCNMESA-DAG: s_mov_b32 s18, -1
+; SIMESA-DAG: s_mov_b32 s19, 0xe8f000
+; VIMESA-DAG: s_mov_b32 s19, 0xe80000
+; GFX9MESA-DAG: s_mov_b32 s19, 0xe00000
+
+
+; GCNMESAMESA: buffer_store_dword {{v[0-9]+}}, off, s[16:19], s3 offset:{{[0-9]+}} ; 4-byte Folded Spill
+
+; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[16:19], s3 offset:{{[0-9]+}}
+; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[16:19], s3 offset:{{[0-9]+}}
+; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[16:19], s3 offset:{{[0-9]+}}
+; GCNMESA: buffer_store_dword {{v[0-9]}}, off, s[16:19], s3 offset:{{[0-9]+}}
+
+; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[16:19], s3 offset:{{[0-9]+}}
+; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[16:19], s3 offset:{{[0-9]+}}
+; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[16:19], s3 offset:{{[0-9]+}}
+; GCNMESA: buffer_load_dword {{v[0-9]+}}, off, s[16:19], s3 offset:{{[0-9]+}}
 
 
 
 ; HSA: buffer_store_dword {{v[0-9]+}}, off, s[0:3], s7 offset:{{[0-9]+}} ; 4-byte Folded Spill
 
 ; HSA: buffer_store_dword {{v[0-9]}}, off, s[0:3], s7 offset:{{[0-9]+}}
 ; HSA: buffer_store_dword {{v[0-9]}}, off, s[0:3], s7 offset:{{[0-9]+}}
 ; HSA: buffer_store_dword {{v[0-9]}}, off, s[0:3], s7 offset:{{[0-9]+}}