AMDGPU/SI: Add llvm.amdgcn.s.buffer.load intrinsic

[• tstellarAMD]

Summary

This patch also adds a new address space to represent loads from the constant
address space that use a buffer resource.

Note that this new address space does not support GEP instructions due to
limitations in the instruction selector.

[arsenm]

include/llvm/IR/IntrinsicsAMDGPU.td

421

This should probably be a more general type so that you can also mangle with FP types

[mareko]

include/llvm/IR/IntrinsicsAMDGPU.td

425

Do these flags prevent CSE? Mesa re-loads TGSI constants on each use. There is no reuse. It relies on CSE to do its the job.

[• tstellarAMD]

More test cases and rebase on top of latest master.

[• tstellarAMD]

include/llvm/IR/IntrinsicsAMDGPU.td

421

The only other option would be llvm_any_ty, I think tablegen needs some more work to make this happen. But we can always change this in the future without breaking backwards compatibility, because the intrinsic is already overloaded.

[mareko]

How is this different from using amdgcn.buffer.load if D28993 lands (which is not certain)?

[• tstellarAMD]

In D27586#662394, @mareko wrote:

How is this different from using amdgcn.buffer.load if D28993 lands (which is not certain)?

I don't think it's legal to select amdgcn.buffer.load to SMRD unless you can prove that it is uniform. llvm.amdgcn.s.buffer.load is known to always be uniform.

[arsenm]

lib/Target/AMDGPU/AMDGPU.h

172

Why 42?

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

167

Does this need to specify non-integral? Also there are a handful of places that assume 64-bit max we should take care of

lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

87

This should be last

[• tstellarAMD]

In D27586#662432, @tstellarAMD wrote:

In D27586#662394, @mareko wrote:

How is this different from using amdgcn.buffer.load if D28993 lands (which is not certain)?

I don't think it's legal to select amdgcn.buffer.load to SMRD unless you can prove that it is uniform. llvm.amdgcn.s.buffer.load is known to always be uniform.

lib/Target/AMDGPU/AMDGPU.h

172

4 dword resource size for address space 2

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

167

We want to be able to use inttoptr, so I don't think we can say non-integral. Do you know where any of these places are?

[mareko]

In D27586#662432, @tstellarAMD wrote:

In D27586#662394, @mareko wrote:

How is this different from using amdgcn.buffer.load if D28993 lands (which is not certain)?

I don't think it's legal to select amdgcn.buffer.load to SMRD unless you can prove that it is uniform. llvm.amdgcn.s.buffer.load is known to always be uniform.

Well, I can't prove that it is uniform, but neither does Mesa for non-constant offsets. The idea of my patch is that SMRD is selected first and moveToVALU will lower it if necessary.

[nhaehnle]

I haven't looked in too much detail yet. I assume getelementptr doesn't work with these pointers, so it would be good to have a negative test which ensures that GEP use fails.

[• tstellarAMD]

In D27586#662991, @nhaehnle wrote:

I haven't looked in too much detail yet. I assume getelementptr doesn't work with these pointers, so it would be good to have a negative test which ensures that GEP use fails.

That's correct. In order to support GEP, we would need to make i128 legal in the backend, which would be a pretty significant change. GlobalISel will make this much much easier, so I'm not sure it's even worth trying to support with SelectionDAG.

[mareko]

Would you please describe the purpose of this patch? It's not obvious why it's useful.

What is the behavior of the constant address space in LLVM? Note that vertex buffers and 'restrict' read-only buffers use immutable memory too, so those are also 'constant' and selecting SMEM for those is desirable if it's possible.

OpenGL Constant buffers aren't limited to SMEM. They can be read with a VGPR offset too. LLVM doesn't have the capability to recognize a non-constant SGPR offset at ISel. If you wanna select SMEM with a non-constant offset, you also need to update moveToVALU.

[• tstellarAMD]

In D27586#663090, @mareko wrote:

Would you please describe the purpose of this patch? It's not obvious why it's useful.

The main reason it is useful is because it tells the compiler that this is a load from a constant value without neededing any more analysis. It's also useful because s_buffer_load_* instructions have a much more simplified resource descriptor, so if then do end up getting selected to MUBUF you don't have to worry about swizzled addressing. It is true however, that you could just use a single llvm.amdgcn.buffer.load.i32 intrinsic for everything, but you may end up with worse code if you are unable to do the analysis required to select it to SMRD instructions.

What is the behavior of the constant address space in LLVM? Note that vertex buffers and 'restrict' read-only buffers use immutable memory too, so those are also 'constant' and selecting SMEM for those is desirable if it's possible.

Constant address space just means that the memory is unchanged for the life of the program.

OpenGL Constant buffers aren't limited to SMEM. They can be read with a VGPR offset too. LLVM doesn't have the capability to recognize a non-constant SGPR offset at ISel. If you wanna select SMEM with a non-constant offset, you also need to update moveToVALU.

[mareko]

In D27586#663420, @tstellarAMD wrote:

In D27586#663090, @mareko wrote:

Would you please describe the purpose of this patch? It's not obvious why it's useful.

The main reason it is useful is because it tells the compiler that this is a load from a constant value without neededing any more analysis. It's also useful because s_buffer_load_* instructions have a much more simplified resource descriptor, so if then do end up getting selected to MUBUF you don't have to worry about swizzled addressing. It is true however, that you could just use a single llvm.amdgcn.buffer.load.i32 intrinsic for everything, but you may end up with worse code if you are unable to do the analysis required to select it to SMRD instructions.

For OpenGL, we'll also need to support non-constant SGPR offsets. And for those, moveToVALU needs to have the corresponding lowering. I don't think having an intrinsic that only accepts constant nodes for the offset is useful.

Also for OpenGL, llvm.amdgcn.buffer.load for SMRD might not be the best choice, because Mesa also needs to determine whether SMRD can be selected with regard to SQC L1 coherency. Therefore, I'd define llvm.amdgcn.s.buffer.load as: "do llvm.amdgcn.buffer.load, but allow SMRD selection for constant and non-constant offsets, and also the referenced memory is immutable (which is why we can use SMRD)". I don't think any other definition is useful for Mesa. Do we agree?

What is the behavior of the constant address space in LLVM? Note that vertex buffers and 'restrict' read-only buffers use immutable memory too, so those are also 'constant' and selecting SMEM for those is desirable if it's possible.

Constant address space just means that the memory is unchanged for the life of the program.

Yes, but what effect on the compiler does it have? Will it allow code sinking and arbitrary scheduling?

[dstuttard]

I've got a scenario that will benefit from this change. It seems to me that this might be a more workable solution than D28993 (although a more generic approach like that is attractive). This change doesn't necessarily preclude something like D28993 at a later stage does it?

Is it possible to make the non-const offset change that mareko talks about as well?

[mareko]

In D27586#744594, @dstuttard wrote:

I've got a scenario that will benefit from this change. It seems to me that this might be a more workable solution than D28993 (although a more generic approach like that is attractive). This change doesn't necessarily preclude something like D28993 at a later stage does it?

Is it possible to make the non-const offset change that mareko talks about as well?

Yes, we'd like to support non-const SGPR offsets as well. Constant offsets are not that interesting to us. Constant offsets can be supported first just to get things going. Non-const offsets can be added afterwards, and the SALU->VALU lowering for s_load_dword should also do inst_offset folding at least. The idea is that the lowering shouldn't generate worse code than amdgcn.buffer.load.dword used directly.

include/llvm/IR/IntrinsicsAMDGPU.td

class AMDGPUBufferLoad : Intrinsic <
    llvm_i32_ty,       // vindex(VGPR)
    llvm_i32_ty,       // offset(SGPR/VGPR/imm)
    llvm_i1_ty,        // glc(imm)
    llvm_i1_ty],       // slc(imm)
   [IntrReadMem]>;
 def int_amdgcn_buffer_load_format : AMDGPUBufferLoad;
 def int_amdgcn_buffer_load : AMDGPUBufferLoad;
 
+
+def int_amdgcn_s_buffer_load : Intrinsic <
+  [llvm_anyint_ty],
+  [LLVMQualPointerType<LLVMMatchType<0>, 42>,
+   llvm_i32_ty, // byte offset
+   llvm_i1_ty], // glc
+  [IntrReadMem, IntrArgMemOnly, NoCapture<0>]>;
+
 class AMDGPUBufferStore : Intrinsic <
   [],
   [llvm_anyfloat_ty,  // vdata(VGPR) -- can currently only select f32, v2f32, v4f32
    llvm_v4i32_ty,     // rsrc(SGPR)
    llvm_i32_ty,       // vindex(VGPR)
    llvm_i32_ty,       // offset(SGPR/VGPR/imm)
    llvm_i1_ty,        // glc(imm)
    llvm_i1_ty],       // slc(imm)

lib/Target/AMDGPU/AMDGPU.h

enum AddressSpaces : unsigned {
   CONSTANT_BUFFER_9 = 17,
   CONSTANT_BUFFER_10 = 18,
   CONSTANT_BUFFER_11 = 19,
   CONSTANT_BUFFER_12 = 20,
   CONSTANT_BUFFER_13 = 21,
   CONSTANT_BUFFER_14 = 22,
   CONSTANT_BUFFER_15 = 23,
 
+  CONSTANT_ADDRESS_W_RSRC = 42,
   // Some places use this if the address space can't be determined.
   UNKNOWN_ADDRESS_SPACE = ~0u
 };
 
 } // namespace AMDGPUAS
 
 #endif

lib/Target/AMDGPU/AMDGPUISelLowering.h

enum NodeType : unsigned {
   STORE_MSKOR,
   LOAD_CONSTANT,
   TBUFFER_STORE_FORMAT,
   ATOMIC_CMP_SWAP,
   ATOMIC_INC,
   ATOMIC_DEC,
   BUFFER_LOAD,
   BUFFER_LOAD_FORMAT,
+  SBUFFER_LOAD,
   LAST_AMDGPU_ISD_NUMBER
 };
 
 
 } // End namespace AMDGPUISD
 
 } // End namespace llvm
 
 #endif

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(STORE_MSKOR)
   NODE_NAME_CASE(LOAD_CONSTANT)
   NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
   NODE_NAME_CASE(ATOMIC_CMP_SWAP)
   NODE_NAME_CASE(ATOMIC_INC)
   NODE_NAME_CASE(ATOMIC_DEC)
   NODE_NAME_CASE(BUFFER_LOAD)
   NODE_NAME_CASE(BUFFER_LOAD_FORMAT)
+  NODE_NAME_CASE(SBUFFER_LOAD)
   case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
   }
   return nullptr;
 }
 
 SDValue AMDGPUTargetLowering::getSqrtEstimate(SDValue Operand,
                                               SelectionDAG &DAG, int Enabled,
                                               int &RefinementSteps,

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

static StringRef computeDataLayout(const Triple &TT) {
   if (TT.getArch() == Triple::r600) {
     // 32-bit pointers.
     return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
             "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
   }
 
   // 32-bit private, local, and region pointers. 64-bit global, constant and
   // flat.
-  return "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32"
+  return "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p42:128:128"
          "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
          "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
 }
 
 LLVM_READNONE
 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
   if (!GPU.empty())
     return GPU;

lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h

unsigned getFlatAddressSpace() const {
     // Don't bother running InferAddressSpaces pass on graphics shaders which
     // don't use flat addressing.
     if (IsGraphicsShader)
       return -1;
     return ST->hasFlatAddressSpace() ? AMDGPUAS::FLAT_ADDRESS : -1;
   }
 
   unsigned getVectorSplitCost() { return 0; }
+
+  bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info);
+
+  Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
+                                           Type *ExpectedType);
+
 };
 
 } // end namespace llvm
 
 #endif

lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

 unsigned AMDGPUTTIImpl::getRegisterBitWidth(bool Vector) {
   return Vector ? 0 : 32;
 }
 
 unsigned AMDGPUTTIImpl::getLoadStoreVecRegBitWidth(unsigned AddrSpace) const {
   switch (AddrSpace) {
   case AMDGPUAS::GLOBAL_ADDRESS:
   case AMDGPUAS::CONSTANT_ADDRESS:
+  case AMDGPUAS::CONSTANT_ADDRESS_W_RSRC:
   case AMDGPUAS::FLAT_ADDRESS:
     return 128;
   case AMDGPUAS::LOCAL_ADDRESS:
   case AMDGPUAS::REGION_ADDRESS:
     return 64;
   case AMDGPUAS::PRIVATE_ADDRESS:
     return 8 * ST->getMaxPrivateElementSize();
   default:
bool AMDGPUTTIImpl::isSourceOfDivergence(const Value *V) const {
   }
 
   // Assume all function calls are a source of divergence.
   if (isa<CallInst>(V) || isa<InvokeInst>(V))
     return true;
 
   return false;
 }
+
+bool AMDGPUTTIImpl::getTgtMemIntrinsic(IntrinsicInst *Inst,
+                                      MemIntrinsicInfo &Info) {
+  IRBuilder<> Builder(Inst);
+  switch (Inst->getIntrinsicID()) {
+  default:
+    return false;
+  case Intrinsic::amdgcn_s_buffer_load:
+    Info.ReadMem = true;
+    Info.WriteMem = false;
+    Info.IsSimple = true;
+    Info.NumMemRefs = 1;
+
+    // We can only set this if the intrinsic is functionally equivalent to a
+    // load/store.
+    if (auto Offset = dyn_cast<ConstantInt>(Inst->getArgOperand(1))) {
+      if (Offset->isZero() &&
+          static_cast<ConstantInt*>(Inst->getArgOperand(2))->isZero()) {
+        Info.PtrVal = Inst->getArgOperand(0);
+      }
+    }
+    break;
+  }
+  return true;
+}
+
+Value *AMDGPUTTIImpl::getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
+                                                        Type *ExpectedType) {
+  if (Inst->getType() == ExpectedType)
+    return Inst;
+
+  return nullptr;
+}

lib/Target/AMDGPU/BUFInstructions.td

 // int_SI_vs_load_input
 def : Pat<
   (SIload_input v4i32:$tlst, imm:$attr_offset, i32:$buf_idx_vgpr),
   (BUFFER_LOAD_FORMAT_XYZW_IDXEN $buf_idx_vgpr, $tlst, (i32 0), imm:$attr_offset, 0, 0, 0)
 >;
 
 // Offset in an 32-bit VGPR
 def : Pat <
-  (SIload_constant v4i32:$sbase, i32:$voff),
-  (BUFFER_LOAD_DWORD_OFFEN $voff, $sbase, (i32 0), 0, 0, 0, 0)
+  (i32 (SIsbuffer_load v4i32:$sbase, i32:$offset, i1:$glc)),
+  (BUFFER_LOAD_DWORD_OFFEN $offset, $sbase, (i32 0), 0, (as_i1imm $glc), 0, 0)
 >;
 
 
 //===----------------------------------------------------------------------===//
 // buffer_load/store_format patterns
 //===----------------------------------------------------------------------===//
 
 multiclass MUBUF_LoadIntrinsicPat<SDPatternOperator name, ValueType vt,

lib/Target/AMDGPU/SIISelLowering.h

public:
   MachineBasicBlock *
   EmitInstrWithCustomInserter(MachineInstr &MI,
                               MachineBasicBlock *BB) const override;
   bool enableAggressiveFMAFusion(EVT VT) const override;
   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
                          EVT VT) const override;
   MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override;
   bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
+  void LowerOperationWrapper(SDNode *N, SmallVectorImpl<SDValue> &Results,
+                             SelectionDAG &DAG) const override;
   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
                           SelectionDAG &DAG) const override;
 
   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
   SDNode *PostISelFolding(MachineSDNode *N, SelectionDAG &DAG) const override;
   void AdjustInstrPostInstrSelection(MachineInstr &MI,
                                      SDNode *Node) const override;

lib/Target/AMDGPU/SIISelLowering.cpp

 #include "llvm/ADT/APFloat.h"
 #include "llvm/ADT/APInt.h"
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/StringSwitch.h"
 #include "llvm/ADT/Twine.h"
+#include "llvm/Analysis/Loads.h"
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/CodeGen/DAGCombine.h"
 #include "llvm/CodeGen/ISDOpcodes.h"
 #include "llvm/CodeGen/MachineBasicBlock.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstr.h"
case Intrinsic::amdgcn_atomic_dec:
     Info.opc = ISD::INTRINSIC_W_CHAIN;
     Info.memVT = MVT::getVT(CI.getType());
     Info.ptrVal = CI.getOperand(0);
     Info.align = 0;
     Info.vol = false;
     Info.readMem = true;
     Info.writeMem = true;
     return true;
+  case Intrinsic::amdgcn_s_buffer_load:
+    Info.opc = AMDGPUISD::SBUFFER_LOAD;
+    Info.memVT = MVT::getVT(CI.getType());
+    Info.ptrVal = CI.getOperand(0);
+    Info.align = 0;
+    Info.vol = false;
+    Info.readMem = true;
+    Info.writeMem = false;
+    return true;
   default:
     return false;
   }
 }
 
 bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &,
                                           EVT) const {
   // SI has some legal vector types, but no legal vector operations. Say no
if (Subtarget->getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
       // by setting the stride value in the resource descriptor which would
       // increase the size limit to (stride * 4GB).  However, this is risky,
       // because it has never been validated.
       return isLegalFlatAddressingMode(AM);
     }
 
     return isLegalMUBUFAddressingMode(AM);
 
+  case AMDGPUAS::CONSTANT_ADDRESS_W_RSRC:
   case AMDGPUAS::CONSTANT_ADDRESS:
     // If the offset isn't a multiple of 4, it probably isn't going to be
     // correctly aligned.
     // FIXME: Can we get the real alignment here?
     if (AM.BaseOffs % 4 != 0)
       return isLegalMUBUFAddressingMode(AM);
 
     // There are no SMRD extloads, so if we have to do a small type access we
bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
 
   return false;
 }
 
 //===----------------------------------------------------------------------===//
 // Custom DAG Lowering Operations
 //===----------------------------------------------------------------------===//
 
+
+void SITargetLowering::LowerOperationWrapper(SDNode *N,
+                                             SmallVectorImpl<SDValue> &Results,
+                                             SelectionDAG &DAG) const {
+
+  if (N->getOpcode() != AMDGPUISD::SBUFFER_LOAD) {
+    TargetLowering::LowerOperationWrapper(N, Results, DAG);
+    return;
+  }
+
+  SDLoc DL(N);
+  MemSDNode *M = cast<MemSDNode>(N);
+  SDValue Ops[] = {
+    M->getOperand(0), // Chain
+    DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, M->getOperand(1)), // Ptr
+    M->getOperand(2), // Offset
+    DAG.getTargetConstant(cast<ConstantSDNode>(
+      M->getOperand(3))->getZExtValue(), DL, MVT::i1) // glc
+  };
+
+  auto MMO = M->getMemOperand();
+  if (isDereferenceablePointer(MMO->getValue(), DAG.getDataLayout()))
+    MMO->setFlags(MachineMemOperand::MODereferenceable);
+
+  SDValue LD = DAG.getMemIntrinsicNode(AMDGPUISD::SBUFFER_LOAD, DL,
+                                       M->getVTList(), Ops, M->getMemoryVT(),
+                                       M->getMemOperand());
+  Results.push_back(LD);
+  Results.push_back(LD.getValue(1));
+}
+
 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   switch (Op.getOpcode()) {
   default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
   case ISD::BRCOND: return LowerBRCOND(Op, DAG);
   case ISD::LOAD: {
     SDValue Result = LowerLOAD(Op, DAG);
     assert((!Result.getNode() ||
             Result.getNode()->getNumValues() == 2) &&
case Intrinsic::r600_read_tidig_y:
     return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
       TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Y), VT);
   case Intrinsic::amdgcn_workitem_id_z:
   case Intrinsic::r600_read_tidig_z:
     return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
       TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z), VT);
   case AMDGPUIntrinsic::SI_load_const: {
     SDValue Ops[] = {
-      Op.getOperand(1),
-      Op.getOperand(2)
+      DAG.getEntryNode(), // Chain
+      Op.getOperand(1), // Ptr
+      Op.getOperand(2), // Offset
+      DAG.getTargetConstant(0, DL, MVT::i1) // glc
     };
 
     MachineMemOperand *MMO = MF.getMachineMemOperand(
         MachinePointerInfo(),
         MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
             MachineMemOperand::MOInvariant,
         VT.getStoreSize(), 4);
-    return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
-                                   Op->getVTList(), Ops, VT, MMO);
+    SDVTList VTList = DAG.getVTList(MVT::i32, MVT::Other);
+    SDValue Load = DAG.getMemIntrinsicNode(AMDGPUISD::SBUFFER_LOAD, DL,
+                                           VTList, Ops, MVT::i32, MMO);
+
+    SDValue MergeOps[] = {
+      DAG.getNode(ISD::BITCAST, DL, MVT::f32, Load),
+      Load.getValue(1)
+    };
+
+    return DAG.getMergeValues(MergeOps, DL);
   }
   case AMDGPUIntrinsic::amdgcn_fdiv_fast:
     return lowerFDIV_FAST(Op, DAG);
   case AMDGPUIntrinsic::SI_vs_load_input:
     return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
                        Op.getOperand(1),
                        Op.getOperand(2),
                        Op.getOperand(3));

lib/Target/AMDGPU/SIInstrInfo.td

def SIEncodingFamily {
   int SI = 0;
   int VI = 1;
 }
 
 //===----------------------------------------------------------------------===//
 // SI DAG Nodes
 //===----------------------------------------------------------------------===//
 
-def SIload_constant : SDNode<"AMDGPUISD::LOAD_CONSTANT",
-  SDTypeProfile<1, 2, [SDTCisVT<0, f32>, SDTCisVT<1, v4i32>, SDTCisVT<2, i32>]>,
-                      [SDNPMayLoad, SDNPMemOperand]
->;
-
 def SIatomic_inc : SDNode<"AMDGPUISD::ATOMIC_INC", SDTAtomic2,
   [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain]
 >;
 
 def SIatomic_dec : SDNode<"AMDGPUISD::ATOMIC_DEC", SDTAtomic2,
   [SDNPMayLoad, SDNPMayStore, SDNPMemOperand, SDNPHasChain]
 >;
 

 
 //===----------------------------------------------------------------------===//
 // PatFrags for global memory operations
 //===----------------------------------------------------------------------===//
 
 defm atomic_inc_global : global_binary_atomic_op<SIatomic_inc>;
 defm atomic_dec_global : global_binary_atomic_op<SIatomic_dec>;
 
+def SIsbuffer_load : SDNode<"AMDGPUISD::SBUFFER_LOAD",
+  SDTypeProfile<1, 3, [SDTCisVT<1, v4i32>, SDTCisVT<2, i32>, SDTCisVT<3, i1>]>,
+  [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]
+>;
+
 //===----------------------------------------------------------------------===//
 // SDNodes and PatFrag for local loads and stores to enable s_mov_b32 m0, -1
 // to be glued to the memory instructions.
 //===----------------------------------------------------------------------===//
 
 def SIld_local : SDNode <"ISD::LOAD", SDTLoad,
   [SDNPHasChain, SDNPMayLoad, SDNPMemOperand, SDNPInGlue]
 >;

lib/Target/AMDGPU/SMInstructions.td

 
 defm : SMRD_Pattern <"S_LOAD_DWORD",    i32>;
 defm : SMRD_Pattern <"S_LOAD_DWORDX2",  v2i32>;
 defm : SMRD_Pattern <"S_LOAD_DWORDX4",  v4i32>;
 defm : SMRD_Pattern <"S_LOAD_DWORDX8",  v8i32>;
 defm : SMRD_Pattern <"S_LOAD_DWORDX16", v16i32>;
 
 // 1. Offset as an immediate
-def SM_LOAD_PATTERN : Pat <  // name this pattern to reuse AddedComplexity on CI
-  (SIload_constant v4i32:$sbase, (SMRDBufferImm i32:$offset)),
-  (S_BUFFER_LOAD_DWORD_IMM $sbase, $offset, 0)
+// name this pattern to reuse AddedComplexity on CI
+def SM_LOAD_PATTERN : Pat <
+  (i32 (SIsbuffer_load v4i32:$sbase, (SMRDBufferImm i32:$offset), i1:$glc)),
+  (S_BUFFER_LOAD_DWORD_IMM $sbase, $offset, (as_i1imm $glc))
 >;
 
 // 2. Offset loaded in an 32bit SGPR
 def : Pat <
-  (SIload_constant v4i32:$sbase, (SMRDBufferSgpr i32:$offset)),
-  (S_BUFFER_LOAD_DWORD_SGPR $sbase, $offset, 0)
+  (i32 (SIsbuffer_load v4i32:$sbase, (SMRDBufferSgpr i32:$offset), i1:$glc)),
+  (S_BUFFER_LOAD_DWORD_SGPR $sbase, $offset, (as_i1imm $glc))
 >;
 
 } // End let AddedComplexity = 100
 
 } // let Predicates = [isGCN]
 
 let Predicates = [isVI] in {
 

 
 def : SMRD_Pattern_ci <"S_LOAD_DWORD",    i32>;
 def : SMRD_Pattern_ci <"S_LOAD_DWORDX2",  v2i32>;
 def : SMRD_Pattern_ci <"S_LOAD_DWORDX4",  v4i32>;
 def : SMRD_Pattern_ci <"S_LOAD_DWORDX8",  v8i32>;
 def : SMRD_Pattern_ci <"S_LOAD_DWORDX16", v16i32>;
 
 def : Pat <
-  (SIload_constant v4i32:$sbase, (SMRDBufferImm32 i32:$offset)),
-  (S_BUFFER_LOAD_DWORD_IMM_ci $sbase, $offset, 0)> {
+  (i32 (SIsbuffer_load v4i32:$sbase, (SMRDBufferImm32 i32:$offset), i1:$glc)),
+  (S_BUFFER_LOAD_DWORD_IMM_ci $sbase, $offset, $glc)> {
   let Predicates = [isCI]; // should this be isCIOnly?
 }
 
 } // End let AddedComplexity = SM_LOAD_PATTERN.AddedComplexity
 

test/CodeGen/AMDGPU/mubuf.ll

main_body:
   %tmp1 = load <16 x i8>, <16 x i8> addrspace(2)* %tmp0
   %tmp2 = shl i32 %6, 2
   %tmp3 = call i32 @llvm.SI.buffer.load.dword.i32.i32(<16 x i8> %tmp1, i32 %tmp2, i32 65, i32 0, i32 1, i32 0, i32 1, i32 0, i32 0)
   %tmp4 = add i32 %6, 16
   call void @llvm.SI.tbuffer.store.i32(<16 x i8> %tmp1, i32 %tmp3, i32 1, i32 %tmp4, i32 %4, i32 0, i32 4, i32 4, i32 1, i32 0, i32 1, i32 1, i32 0)
   ret void
 }
 
+; Using the load.const intrinsic with an vgpr offset
+; CHECK-LABEL: {{^}}s_buffer_load:
+; CHECK: buffer_load_dword v{{[0-9+]}}, v{{[0-9+]}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen
+; CHECK: buffer_load_dword v{{[0-9+]}}, v{{[0-9+]}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen
+define amdgpu_ps void @s_buffer_load(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float, i32 addrspace(42)* addrspace(2)* inreg %in) {
+main_body:
+  %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)
+  %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
+  %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
+  %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 %tid)
+  %23 = load i32 addrspace(42)*, i32 addrspace(42)* addrspace(2)* %in
+  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %23, i32 %tid, i1 false)
+  %s.buffer.float = bitcast i32 %s.buffer to float
+  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %s.buffer.float)
+  ret void
+}
+
 ;;;==========================================================================;;;
 ;;; MUBUF STORE TESTS
 ;;;==========================================================================;;;
 
 ; MUBUF store with an immediate byte offset that fits into 12-bits
 ; CHECK-LABEL: {{^}}mubuf_store0:
 ; CHECK: buffer_store_dword v{{[0-9]}}, off, s[{{[0-9]:[0-9]}}], 0 offset:4 ; encoding: [0x04,0x00,0x70,0xe0
 define void @mubuf_store0(i32 addrspace(1)* %out) {

 ; CHECK: buffer_store_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64
 define void @store_vgpr_ptr(i32 addrspace(1)* %out) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x() readnone
   %out.gep = getelementptr i32, i32 addrspace(1)* %out, i32 %tid
   store i32 99, i32 addrspace(1)* %out.gep, align 4
   ret void
 }
 
+declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #1
 declare i32 @llvm.SI.buffer.load.dword.i32.i32(<16 x i8>, i32, i32, i32, i32, i32, i32, i32, i32) #0
 declare void @llvm.SI.tbuffer.store.i32(<16 x i8>, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32)
+declare float @llvm.SI.load.const(<16 x i8>, i32) #1
+declare i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* nocapture, i32, i1)
+declare void @llvm.SI.export(i32, i32, i32, i32, i32, float, float, float, float)
 
 attributes #0 = { nounwind readonly }
+attributes #1 = { nounwind readnone }

test/CodeGen/AMDGPU/smrd.ll

entry:
   %1 = load i32, i32 addrspace(2)* %0
   store i32 %1, i32 addrspace(1)* %out
   ret void
 }
 
 ; SMRD load using the load.const intrinsic with an immediate offset
 ; GCN-LABEL: {{^}}smrd_load_const0:
 ; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4 ; encoding: [0x04
+; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4 ; encoding: [0x04
+; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x10
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x10
-define amdgpu_ps void @smrd_load_const0(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) {
+define amdgpu_ps void @smrd_load_const0(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float, i32 addrspace(42)* addrspace(2)* inreg %in) {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 16)
-  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
+  %23 = load i32 addrspace(42)*, i32 addrspace(42)* addrspace(2)* %in
+  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %23, i32 16, i1 false)
+  %s.buffer.float = bitcast i32 %s.buffer to float
+  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %s.buffer.float)
   ret void
 }
 
 ; SMRD load using the load.const intrinsic with the largest possible immediate
 ; offset.
 ; GCN-LABEL: {{^}}smrd_load_const1:
 ; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff
+; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff
+; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
-define amdgpu_ps void @smrd_load_const1(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) {
+define amdgpu_ps void @smrd_load_const1(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float, i32 addrspace(42)* addrspace(2)* inreg %in) {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1020)
-  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
+  %23 = load i32 addrspace(42)*, i32 addrspace(42)* addrspace(2)* %in
+  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %23, i32 1020, i1 false)
+  %s.buffer.float = bitcast i32 %s.buffer to float
+  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %s.buffer.float)
   ret void
 }
 ; SMRD load using the load.const intrinsic with an offset greater than the
 ; largets possible immediate.
 ; immediate offset.
 ; GCN-LABEL: {{^}}smrd_load_const2:
 ; SI: s_movk_i32 s[[OFFSET:[0-9]]], 0x400
 ; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
+; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
+; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
 ; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
-define amdgpu_ps void @smrd_load_const2(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) {
+; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
+define amdgpu_ps void @smrd_load_const2(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float, i32 addrspace(42)* addrspace(2)* inreg %in) {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1024)
-  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
+  %23 = load i32 addrspace(42)*, i32 addrspace(42)* addrspace(2)* %in
+  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %23, i32 1024, i1 false)
+  %s.buffer.float = bitcast i32 %s.buffer to float
+  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %s.buffer.float)
   ret void
 }
 
 ; SMRD load with the largest possible immediate offset on VI
 ; GCN-LABEL: {{^}}smrd_load_const3:
 ; SI: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
 ; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
 ; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
 ; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
-define amdgpu_ps void @smrd_load_const3(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) {
+; VI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
+define amdgpu_ps void @smrd_load_const3(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float, i32 addrspace(42)* addrspace(2)* inreg %in) {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1048572)
-  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
+  %23 = load i32 addrspace(42)*, i32 addrspace(42)* addrspace(2)* %in
+  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %23, i32 1048572, i1 false)
+  %s.buffer.float = bitcast i32 %s.buffer to float
+  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %s.buffer.float)
   ret void
 }
 
 ; SMRD load with an offset greater than the largest possible immediate on VI
 ; GCN-LABEL: {{^}}smrd_load_const4:
 ; SIVI: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
 ; SIVI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; SIVI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
 ; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
 ; GCN: s_endpgm
-define amdgpu_ps void @smrd_load_const4(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) {
+define amdgpu_ps void @smrd_load_const4(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float, i32 addrspace(42)* addrspace(2)* inreg %in) {
 main_body:
   %20 = getelementptr <16 x i8>, <16 x i8> addrspace(2)* %0, i32 0
   %21 = load <16 x i8>, <16 x i8> addrspace(2)* %20
   %22 = call float @llvm.SI.load.const(<16 x i8> %21, i32 1048576)
-  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %22)
+  %23 = load i32 addrspace(42)*, i32 addrspace(42)* addrspace(2)* %in
+  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %23, i32 1048576, i1 false)
+  %s.buffer.float = bitcast i32 %s.buffer to float
+  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 0, float %22, float %22, float %22, float %s.buffer.float)
   ret void
 }
 
 ; Function Attrs: nounwind readnone
 declare float @llvm.SI.load.const(<16 x i8>, i32) #0
 
 declare void @llvm.SI.export(i32, i32, i32, i32, i32, float, float, float, float)
 
+declare i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* nocapture, i32, i1)
+
 attributes #0 = { nounwind readnone }

test/Transforms/EarlyCSE/AMDGPU/intrinsics.ll

+; RUN: opt < %s -S -mtriple=amdgcn-- -early-cse | FileCheck %s
+
+; CHECK-LABEL: @no_cse
+; CHECK: call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 0, i1 false)
+; CHECK: call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 4, i1 false)
+define void @no_cse(i32 addrspace(1)* %out, i32 addrspace(42)* %in) {
+  %a = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 0, i1 false)
+  %b = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 4, i1 false)
+  %c = add i32 %a, %b
+  store i32 %c, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: @cse_zero_offset
+; CHECK: [[CSE:%[a-z0-9A-Z]+]] = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 0, i1 false)
+; CHECK: add i32 [[CSE]], [[CSE]]
+define void @cse_zero_offset(i32 addrspace(1)* %out, i32 addrspace(42)* %in) {
+  %a = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 0, i1 false)
+  %b = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 0, i1 false)
+  %c = add i32 %a, %b
+  store i32 %c, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK-LABEL: @cse_nonzero_offset
+; CHECK: [[CSE:%[a-z0-9A-Z]+]] = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 4, i1 false)
+; CHECK: add i32 [[CSE]], [[CSE]]
+define void @cse_nonzero_offset(i32 addrspace(1)* %out, i32 addrspace(42)* %in) {
+  %a = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 4, i1 false)
+  %b = call i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* %in, i32 4, i1 false)
+  %c = add i32 %a, %b
+  store i32 %c, i32 addrspace(1)* %out
+  ret void
+}
+
+declare i32 @llvm.amdgcn.s.buffer.load.i32(i32 addrspace(42)* nocapture, i32, i1)