[AMDGPU] Add intrinsics llvm.amdgcn.{raw|struct}.buffer.load.lds

[rampitec]

It uses a single mem operand with both load and store in addrspace(4). The addrspace(4) is common for all buffer intrinsics memops. In fact neither MemSDNode nor MemIntrinsicSDNode can have 2 mem ops. Only MachineSDNode and final MI can. I certainly do now want to create a MachineSDNode here and duplicate a lot of buffer operations logic. If we believe we really want 2 mem ops these can be split in the FinalizeLowering.

[rampitec]

It will also need to be rebased on top of D124550 to handle hazard between M0 initialization and LDS DMA.

[arsenm]

Why does this need an intrinsic? I thought the whole point of the LDS DMA thing was an optimization the backend would perform and doesn't need to be exposed directly

[rampitec]

In D124884#3489729, @arsenm wrote:

Why does this need an intrinsic? I thought the whole point of the LDS DMA thing was an optimization the backend would perform and doesn't need to be exposed directly

We cannot match this pattern. If you look at the addressing mode this is byzantine. Yet another addtid instruction on steroids.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

4289

Why is this needed if it's in the MMO?

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

Should use the return / data type?

[rampitec]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

4289

I believe to select based on the MMO I would need to write a complex pattern.

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

Changed to return. What do you mean by 'use data type'?

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

You're looking at the pointer element type instead of the return / data type. i.e. we would have i8/i16/i32 return types and you don't need to look at the pointer

1199–1200

I just noticed there is no return type so this is just introducing a dependency on typed pointers which is a no-go.

I don't actually see why can't we match these from the buffer intrinsic plus LDS access?

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

It does not return anything. This instruction does not have vdata. The only way to know the size is by looking at the overloaded LDS base pointer pointee.

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

LDS address = LDS_base + LDS_offset + inst_offset + (TIDinWave * 4)

We do not have TIDinWave, certainly not after selection. Even before selection it is extremely problematic.

Why typed pointer is a no-go if that works?

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

On top of that MEM_ADDR also depends on the TID. It not the same address as a normal buffer_load would use with the same operands.

[arsenm]

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

1268

This should be addrspace 3 pointers only also

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

Pointee types have been removed from the IR. If this really needs the type it would need to use an attribute on the parameter to carry it which may be new territory

[rampitec]

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

1268

It cannot be overloaded on the pointee type, infrastructure limitation. We are using llvm_anyptr_ty everywhere in such context.

If in turn we cannot use pointee types at all then this could be non overloaded pointer to void in addrspace 3.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1199–1200

It does not really need type but it needs size. I can add immediate to the intrinsic and switch to void* for LDS base.

[rampitec]

To confirm: is that OK to add yet another imm to the end of operands of the intrinsic to select a byte size? And then remove the overload. If yes I will do it tomorrow.

[rampitec]

In D124884#3489967, @rampitec wrote:

To confirm: is that OK to add yet another imm to the end of operands of the intrinsic to select a byte size? And then remove the overload. If yes I will do it tomorrow.

Or maybe after the pointer, although it is less convenient for lowering.

[arsenm]

In D124884#3489967, @rampitec wrote:

To confirm: is that OK to add yet another imm to the end of operands of the intrinsic to select a byte size? And then remove the overload. If yes I will do it tomorrow.

There's the elementtype attribute for this case which some arm intrinsics seem to be using. Not sure how you're supposed to define an intrinsic to use it though

[arsenm]

In D124884#3490643, @arsenm wrote:

In D124884#3489967, @rampitec wrote:

To confirm: is that OK to add yet another imm to the end of operands of the intrinsic to select a byte size? And then remove the overload. If yes I will do it tomorrow.

There's the elementtype attribute for this case which some arm intrinsics seem to be using. Not sure how you're supposed to define an intrinsic to use it though

Apparently this isn't well developed but works. The verifier is hardcoding these intrinsics (it's also looking at the call site instead of the intrinsic declaration attributes)

[nhaehnle]

In D124884#3490649, @arsenm wrote:

In D124884#3490643, @arsenm wrote:

In D124884#3489967, @rampitec wrote:

To confirm: is that OK to add yet another imm to the end of operands of the intrinsic to select a byte size? And then remove the overload. If yes I will do it tomorrow.

There's the elementtype attribute for this case which some arm intrinsics seem to be using. Not sure how you're supposed to define an intrinsic to use it though

Apparently this isn't well developed but works. The verifier is hardcoding these intrinsics (it's also looking at the call site instead of the intrinsic declaration attributes)

What's wrong with the idea of an i32 imm %size argument? That seems to me more in line with the philosophy of caring less about types.

[arsenm]

In D124884#3490731, @nhaehnle wrote:

What's wrong with the idea of an i32 imm %size argument? That seems to me more in line with the philosophy of caring less about types.

I'd prefer to keep any intrinsics that look like a load or store to look more like the regular load or store instructions. All of these arbitrary immediate parameters are uglier (e.g. the memory ordering arguments that don't actually work on some of the atomics)

[rampitec]

In D124884#3490643, @arsenm wrote:

In D124884#3489967, @rampitec wrote:

To confirm: is that OK to add yet another imm to the end of operands of the intrinsic to select a byte size? And then remove the overload. If yes I will do it tomorrow.

There's the elementtype attribute for this case which some arm intrinsics seem to be using. Not sure how you're supposed to define an intrinsic to use it though

This will need a clang builtin to produce the attribute. I am not sure we really want to expose it as a clang builtin.

[rampitec]

In D124884#3490745, @arsenm wrote:

In D124884#3490731, @nhaehnle wrote:

What's wrong with the idea of an i32 imm %size argument? That seems to me more in line with the philosophy of caring less about types.

I'd prefer to keep any intrinsics that look like a load or store to look more like the regular load or store instructions. All of these arbitrary immediate parameters are uglier (e.g. the memory ordering arguments that don't actually work on some of the atomics)

It is more or less similar to memcpy, and memcpy uses size argument.

[asroy]

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.buffer.load.lds.ll

16

m0 holds the size of LDS, should we save the value of m0 before overwriting it, and write the value back before issuing ds_read?

[arsenm]

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.buffer.load.lds.ll

16

Every user of m0 is supposed to set it itself, and we hopefully clean up the redundant rewrites. It's not something that's generally saved and restored per operation

[rampitec]

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.buffer.load.lds.ll

16

DS_* do not read M0 since gfx9. These intrinsics are only available since gfx9.
Moreover, on gfx8 and earlier selection of DS opcodes takes care about M0 initialization right before the opcode.

[rampitec]

• Removed the overload and added i32 %size operand instead.
• LDS pointer is i8 addrspace(3) now qualified with the address space.
• Rebased on the change to handle hazards between m0 initialization and these operations.

[ramjana]

In D124884#3489729, @arsenm wrote:

Why does this need an intrinsic? I thought the whole point of the LDS DMA thing was an optimization the backend would perform and doesn't need to be exposed directly

In D124884#3492464, @rampitec wrote:

• Removed the overload and added i32 %size operand instead.
• LDS pointer is i8 addrspace(3) now qualified with the address space.
• Rebased on the change to handle hazards between m0 initialization and these operations.

In D124884#3492464, @rampitec wrote:

• Removed the overload and added i32 %size operand instead.
• LDS pointer is i8 addrspace(3) now qualified with the address space.
• Rebased on the change to handle hazards between m0 initialization and these operations.

Just to be clear , Is your expectation that intrinsic user to save and restore m0 before calling buffer_load lds intrinsic?

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.buffer.load.lds.ll

16

Just to be clear , Is your expectation that intrinsic user to save and restore m0 before calling buffer_load lds intrinsic?

[rampitec]

In D124884#3496841, @ramjana wrote:

Just to be clear , Is your expectation that intrinsic user to save and restore m0 before calling buffer_load lds intrinsic?

No, you do not have to.

[rampitec]

Do not split voffset because inst_offset is applied to both VMEM and LDS address and voffset is not. Add a separate operand instead.

[rampitec]

Removed support for wide than DWORD ops. See D125409.

[piotr]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

4277–4279

What is preventing this from clobbering M0? There are intrinsics like int_amdgcn_interp* that have a dependency on M0. Shouldn't the code save the existing M0 and restore it after the load?

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

4277–4279

m0 isn't treated as a preserved value. Each user is supposed to initialize m0 itself

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

8250

I don't see how / where this preserves the LDS bit

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

389

If you're going to rely on the memory operand, the verifier needs to start enforcing these have one memory operand (well, 2 actually with the same sizes)

[rampitec]

Return false from getMemOperandsWithOffsetWidth() instead of checking mem operand.

[rampitec]

Switched to direct select which allows to use 2 separate memory operands.
The patch now handles both raw and struct intrinsics.

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

8250

It has different number of operands comparing to the SIbuffer_load, so selects into _LDS versions of opcodes.

In fact after I have removed offset split because we cannot do it on one pointer only, and dropped multi-dword support I start thinking it might be better to drop SIbuffer_load_lds, patterns, and produce MachineSDNode right here (like in the D125279 for global load), it will not be so much code anymore and I will be able to produce 2 separate memory operands.

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

389

On a second thought it is better to just return false here. We cannot have a reasonable pointer here on either side anyway, and in fact even 2 memory operands which it should ideally have should be of a different size for a sub-dword operations. A load can be sub-dword, but the store is always extended to a dword.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

3080 ↗

(On Diff #429347)

Should return false, the verifier isn't enforcing this

3130 ↗

(On Diff #429347)

The verifier should probably be enforcing MMO ordering if you're going to rely on that

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

8243

Ditto, verifier isn't enforcing this so shouldn't assert

[rampitec]

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

3130 ↗

(On Diff #429347)

I do not rely on their order. Not anymore.

[rampitec]

Changed asserts to cannot select.

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

class AMDGPUBufferAtomicFP : Intrinsic <
    llvm_i32_ty,      // vindex(VGPR)
    llvm_i32_ty,      // offset(SGPR/VGPR/imm)
    llvm_i1_ty],      // slc(imm)
   [ImmArg<ArgIndex<4>>, IntrWillReturn], "", [SDNPMemOperand]>,
   AMDGPURsrcIntrinsic<1, 0>;
 
 // Legacy form of the intrinsic. raw and struct forms should be preferred.
 def int_amdgcn_buffer_atomic_fadd : AMDGPUBufferAtomicFP;
+
+class AMDGPURawBufferLoadLDS : Intrinsic <
+  [],
+  [llvm_v4i32_ty,     // rsrc(SGPR)
+   llvm_anyptr_ty,    // LDS base offset
+   llvm_i32_ty,       // offset(VGPR/imm, included in bounds checking and swizzling)
+   llvm_i32_ty,       // soffset(SGPR/imm, excluded from bounds checking and swizzling)
+   llvm_i32_ty],      // auxiliary data (imm, cachepolicy     (bit 0 = glc,
+                      //                                       bit 1 = slc,
+                      //                                       bit 2 = dlc on gfx10+))
+                      //                      swizzled buffer (bit 3 = swz))
+  [IntrWillReturn, NoCapture<ArgIndex<1>>, ImmArg<ArgIndex<4>>], "", [SDNPMemOperand]>,
+  AMDGPURsrcIntrinsic<0>;
+def int_amdgcn_raw_buffer_load_lds : AMDGPURawBufferLoadLDS;
+
 } // defset AMDGPUBufferIntrinsics
 
 // Uses that do not set the done bit should set IntrWriteMem on the
 // call site.
 def int_amdgcn_exp : Intrinsic <[], [
   llvm_i32_ty,       // tgt,
   llvm_i32_ty,       // en
   llvm_any_ty,       // src0 (f32 or i32)

llvm/lib/Target/AMDGPU/AMDGPUGISel.td

 def : GINodeEquiv<G_AMDGPU_ATOMIC_CMPXCHG, AMDGPUatomic_cmp_swap>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD, SIbuffer_load>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_USHORT, SIbuffer_load_ushort>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_UBYTE, SIbuffer_load_ubyte>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_SSHORT, SIbuffer_load_short>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_SBYTE, SIbuffer_load_byte>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_FORMAT, SIbuffer_load_format>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_FORMAT_D16, SIbuffer_load_format_d16>;
+def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_LDS, SIbuffer_load_lds>;
 def : GINodeEquiv<G_AMDGPU_TBUFFER_LOAD_FORMAT, SItbuffer_load>;
 def : GINodeEquiv<G_AMDGPU_TBUFFER_LOAD_FORMAT_D16, SItbuffer_load_d16>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_STORE, SIbuffer_store>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_SHORT, SIbuffer_store_short>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_BYTE, SIbuffer_store_byte>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_FORMAT, SIbuffer_store_format>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_FORMAT_D16, SIbuffer_store_format_d16>;
 def : GINodeEquiv<G_AMDGPU_TBUFFER_STORE_FORMAT, SItbuffer_store>;

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.h

Register fixStoreSourceType(MachineIRBuilder &B, Register VData,
                               bool IsFormat) const;
 
   bool legalizeBufferStore(MachineInstr &MI, MachineRegisterInfo &MRI,
                            MachineIRBuilder &B, bool IsTyped,
                            bool IsFormat) const;
   bool legalizeBufferLoad(MachineInstr &MI, MachineRegisterInfo &MRI,
                           MachineIRBuilder &B, bool IsFormat,
                           bool IsTyped) const;
+  bool legalizeBufferLoadLds(MachineInstr &MI, MachineRegisterInfo &MRI,
+                             MachineIRBuilder &B) const;
   bool legalizeBufferAtomic(MachineInstr &MI, MachineIRBuilder &B,
                             Intrinsic::ID IID) const;
 
   bool legalizeBVHIntrinsic(MachineInstr &MI, MachineIRBuilder &B) const;
 
   bool legalizeFPTruncRound(MachineInstr &MI, MachineIRBuilder &B) const;
 
   bool legalizeImageIntrinsic(

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

else {
       B.buildMerge(Dst, Repack);
     }
   }
 
   MI.eraseFromParent();
   return true;
 }
 
+bool AMDGPULegalizerInfo::legalizeBufferLoadLds(MachineInstr &MI,
+                                                MachineRegisterInfo &MRI,
+                                                MachineIRBuilder &B) const {
+  MachineMemOperand *MMO = *MI.memoperands_begin();
+  const LLT MemTy = MMO->getMemoryType();
+  const LLT S32 = LLT::scalar(32);
+
+  Register RSrc = MI.getOperand(1).getReg();
+
+  // The struct intrinsic variants add one additional operand over raw.
+  const bool HasVIndex = MI.getNumOperands() == 7;
+  Register VIndex;
+  int OpOffset = 0;
+  if (HasVIndex) {
+    VIndex = MI.getOperand(3).getReg();
+    OpOffset = 1;
+  } else {
+    VIndex = B.buildConstant(S32, 0).getReg(0);
+  }
+
+  Register VOffset = MI.getOperand(3 + OpOffset).getReg();
+  Register SOffset = MI.getOperand(4 + OpOffset).getReg();
+  Register M0Val = MI.getOperand(2).getReg();
+
+  unsigned AuxiliaryData = MI.getOperand(5 + OpOffset).getImm();
+  unsigned ImmOffset;
+
+  std::tie(VOffset, ImmOffset) = splitBufferOffsets(B, VOffset);
+  updateBufferMMO(MMO, VOffset, SOffset, ImmOffset, VIndex, MRI);
+
+  B.buildInstr(AMDGPU::COPY)
+    .addDef(AMDGPU::M0)
+    .addUse(M0Val);
+  B.buildInstr(AMDGPU::G_AMDGPU_BUFFER_LOAD_LDS)
+    .addUse(RSrc)                   // rsrc
+    .addUse(VIndex)                 // vindex
+    .addUse(VOffset)                // voffset
+    .addUse(SOffset)                // soffset
+    .addImm(ImmOffset)              // offset(imm)
+    .addImm(AuxiliaryData)          // cachepolicy, swizzled buffer(imm)
+    .addImm(HasVIndex ? -1 : 0)     // idxen(imm)
+    .addImm(MemTy.getSizeInBytes()) // data byte size
+    .addMemOperand(MMO);
+
+  MI.eraseFromParent();
+  return true;
+}
+
 bool AMDGPULegalizerInfo::legalizeAtomicIncDec(MachineInstr &MI,
                                                MachineIRBuilder &B,
                                                bool IsInc) const {
   unsigned Opc = IsInc ? AMDGPU::G_AMDGPU_ATOMIC_INC :
                          AMDGPU::G_AMDGPU_ATOMIC_DEC;
   B.buildInstr(Opc)
     .addDef(MI.getOperand(0).getReg())
     .addUse(MI.getOperand(2).getReg())
bool AMDGPULegalizerInfo::legalizeIntrinsic(LegalizerHelper &Helper,
   case Intrinsic::amdgcn_struct_buffer_store_format:
     return legalizeBufferStore(MI, MRI, B, false, true);
   case Intrinsic::amdgcn_raw_tbuffer_store:
   case Intrinsic::amdgcn_struct_tbuffer_store:
     return legalizeBufferStore(MI, MRI, B, true, true);
   case Intrinsic::amdgcn_raw_buffer_load:
   case Intrinsic::amdgcn_struct_buffer_load:
     return legalizeBufferLoad(MI, MRI, B, false, false);
+  case Intrinsic::amdgcn_raw_buffer_load_lds:
+    return legalizeBufferLoadLds(MI, MRI, B);
   case Intrinsic::amdgcn_raw_buffer_load_format:
   case Intrinsic::amdgcn_struct_buffer_load_format:
     return legalizeBufferLoad(MI, MRI, B, true, false);
   case Intrinsic::amdgcn_raw_tbuffer_load:
   case Intrinsic::amdgcn_struct_tbuffer_load:
     return legalizeBufferLoad(MI, MRI, B, true, true);
   case Intrinsic::amdgcn_raw_buffer_atomic_swap:
   case Intrinsic::amdgcn_struct_buffer_atomic_swap:

llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

void AMDGPURegisterBankInfo::applyMappingImpl(
   case AMDGPU::G_AMDGPU_BUFFER_STORE_FORMAT:
   case AMDGPU::G_AMDGPU_BUFFER_STORE_FORMAT_D16:
   case AMDGPU::G_AMDGPU_TBUFFER_STORE_FORMAT:
   case AMDGPU::G_AMDGPU_TBUFFER_STORE_FORMAT_D16: {
     applyDefaultMapping(OpdMapper);
     executeInWaterfallLoop(MI, MRI, {1, 4});
     return;
   }
+  case AMDGPU::G_AMDGPU_BUFFER_LOAD_LDS: {
+    applyDefaultMapping(OpdMapper);
+    executeInWaterfallLoop(MI, MRI, {0, 3});
+    return;
+  }
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SWAP:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_ADD:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SUB:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SMIN:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_UMIN:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SMAX:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_UMAX:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_AND:
case AMDGPU::G_AMDGPU_BUFFER_STORE_FORMAT_D16: {
 
     // soffset
     OpdsMapping[4] = getSGPROpMapping(MI.getOperand(4).getReg(), MRI, *TRI);
 
     // Any remaining operands are immediates and were correctly null
     // initialized.
     break;
   }
+  case AMDGPU::G_AMDGPU_BUFFER_LOAD_LDS: {
+    // rsrc
+    OpdsMapping[0] = getSGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
+
+    // vindex
+    OpdsMapping[1] = getVGPROpMapping(MI.getOperand(1).getReg(), MRI, *TRI);
+
+    // voffset
+    OpdsMapping[2] = getVGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);
+
+    // soffset
+    OpdsMapping[3] = getSGPROpMapping(MI.getOperand(3).getReg(), MRI, *TRI);
+
+    // Any remaining operands are immediates and were correctly null
+    // initialized.
+    break;
+  }
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SWAP:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_ADD:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SUB:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SMIN:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_UMIN:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SMAX:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_UMAX:
   case AMDGPU::G_AMDGPU_BUFFER_ATOMIC_AND:
case Intrinsic::amdgcn_raw_tbuffer_load: {
       // FIXME: Should make intrinsic ID the last operand of the instruction,
       // then this would be the same as store
       OpdsMapping[0] = getVGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
       OpdsMapping[2] = getSGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);
       OpdsMapping[3] = getVGPROpMapping(MI.getOperand(3).getReg(), MRI, *TRI);
       OpdsMapping[4] = getSGPROpMapping(MI.getOperand(4).getReg(), MRI, *TRI);
       break;
     }
+    case Intrinsic::amdgcn_raw_buffer_load_lds: {
+      OpdsMapping[1] = getSGPROpMapping(MI.getOperand(1).getReg(), MRI, *TRI);
+      OpdsMapping[2] = getVGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);
+      OpdsMapping[3] = getVGPROpMapping(MI.getOperand(3).getReg(), MRI, *TRI);
+      OpdsMapping[4] = getSGPROpMapping(MI.getOperand(4).getReg(), MRI, *TRI);
+      break;
+    }
     case Intrinsic::amdgcn_raw_buffer_store:
     case Intrinsic::amdgcn_raw_buffer_store_format:
     case Intrinsic::amdgcn_raw_tbuffer_store: {
       OpdsMapping[1] = getVGPROpMapping(MI.getOperand(1).getReg(), MRI, *TRI);
       OpdsMapping[2] = getSGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);
       OpdsMapping[3] = getVGPROpMapping(MI.getOperand(3).getReg(), MRI, *TRI);
       OpdsMapping[4] = getSGPROpMapping(MI.getOperand(4).getReg(), MRI, *TRI);
       break;

llvm/lib/Target/AMDGPU/BUFInstructions.td

 defm : MUBUF_LoadIntrinsicPat<SIbuffer_load, v3i32, "BUFFER_LOAD_DWORDX3">;
 defm : MUBUF_LoadIntrinsicPat<SIbuffer_load, v4f32, "BUFFER_LOAD_DWORDX4">;
 defm : MUBUF_LoadIntrinsicPat<SIbuffer_load, v4i32, "BUFFER_LOAD_DWORDX4">;
 defm : MUBUF_LoadIntrinsicPat<SIbuffer_load_byte, i32, "BUFFER_LOAD_SBYTE">;
 defm : MUBUF_LoadIntrinsicPat<SIbuffer_load_short, i32, "BUFFER_LOAD_SSHORT">;
 defm : MUBUF_LoadIntrinsicPat<SIbuffer_load_ubyte, i32, "BUFFER_LOAD_UBYTE">;
 defm : MUBUF_LoadIntrinsicPat<SIbuffer_load_ushort,  i32, "BUFFER_LOAD_USHORT">;
 
+multiclass MUBUF_LoadLDSIntrinsicPat<SDPatternOperator ld, int size, string opcode> {
+  def : GCNPat<
+    (ld v4i32:$rsrc, 0, 0, i32:$soffset, timm:$offset,
+              timm:$auxiliary, 0, (i32 size)),
+    (!cast<MUBUF_Pseudo>(opcode # _OFFSET) SReg_128:$rsrc, SCSrc_b32:$soffset, (as_i16timm $offset),
+      (extract_cpol $auxiliary), (extract_swz $auxiliary))
+  >;
+
+  def : GCNPat<
+    (ld v4i32:$rsrc, 0, i32:$voffset, i32:$soffset, timm:$offset,
+              timm:$auxiliary, 0, (i32 size)),
+    (!cast<MUBUF_Pseudo>(opcode # _OFFEN) VGPR_32:$voffset, SReg_128:$rsrc, SCSrc_b32:$soffset, (as_i16timm $offset),
+      (extract_cpol $auxiliary), (extract_swz $auxiliary))
+  >;
+
+  def : GCNPat<
+    (ld v4i32:$rsrc, i32:$vindex, 0, i32:$soffset, timm:$offset,
+              timm:$auxiliary, timm, (i32 size)),
+    (!cast<MUBUF_Pseudo>(opcode # _IDXEN) VGPR_32:$vindex, SReg_128:$rsrc, SCSrc_b32:$soffset, (as_i16timm $offset),
+      (extract_cpol $auxiliary), (extract_swz $auxiliary))
+  >;
+
+  def : GCNPat<
+    (ld v4i32:$rsrc, i32:$vindex, i32:$voffset, i32:$soffset, timm:$offset,
+              timm:$auxiliary, timm, (i32 size)),
+    (!cast<MUBUF_Pseudo>(opcode # _BOTHEN)
+      (REG_SEQUENCE VReg_64, VGPR_32:$vindex, sub0, VGPR_32:$voffset, sub1),
+      SReg_128:$rsrc, SCSrc_b32:$soffset, (as_i16timm $offset),
+      (extract_cpol $auxiliary), (extract_swz $auxiliary))
+  >;
+}
+
+defm : MUBUF_LoadLDSIntrinsicPat<SIbuffer_load_lds,  4, "BUFFER_LOAD_DWORD_LDS">;
+defm : MUBUF_LoadLDSIntrinsicPat<SIbuffer_load_lds,  8, "BUFFER_LOAD_DWORDX2_LDS">;
+defm : MUBUF_LoadLDSIntrinsicPat<SIbuffer_load_lds, 12, "BUFFER_LOAD_DWORDX3_LDS">;
+defm : MUBUF_LoadLDSIntrinsicPat<SIbuffer_load_lds, 16, "BUFFER_LOAD_DWORDX4_LDS">;
+defm : MUBUF_LoadLDSIntrinsicPat<SIbuffer_load_lds,  1, "BUFFER_LOAD_UBYTE_LDS">;
+defm : MUBUF_LoadLDSIntrinsicPat<SIbuffer_load_lds,  2, "BUFFER_LOAD_USHORT_LDS">;
+
 multiclass MUBUF_StoreIntrinsicPat<SDPatternOperator name, ValueType vt,
                                    string opcode, ValueType memoryVt = vt> {
   defvar st = !if(!eq(memoryVt, vt), name, mubuf_intrinsic_store<name, memoryVt>);
 
   def : GCNPat<
     (st vt:$vdata, v4i32:$rsrc, 0, 0, i32:$soffset, timm:$offset,
               timm:$auxiliary, 0),
     (!cast<MUBUF_Pseudo>(opcode # _OFFSET_exact) getVregSrcForVT<vt>.ret:$vdata, SReg_128:$rsrc, SCSrc_b32:$soffset, (as_i16timm $offset),

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (Attr.hasFnAttr(Attribute::ReadOnly)) {
                                             ISD::INTRINSIC_W_CHAIN;
       Info.memVT = MVT::getVT(CI.getArgOperand(0)->getType());
       Info.flags |= MachineMemOperand::MOLoad |
                     MachineMemOperand::MOStore |
                     MachineMemOperand::MODereferenceable;
 
       // XXX - Should this be volatile without known ordering?
       Info.flags |= MachineMemOperand::MOVolatile;
+
+      switch (IntrID) {
+      default:
+        break;
+      case Intrinsic::amdgcn_raw_buffer_load_lds:
+        Info.memVT = MVT::getVT(CI.getArgOperand(1)->getType()->
+                                    getNonOpaquePointerElementType());
+        return true;
+      }
     }
     return true;
   }
 
   switch (IntrID) {
   case Intrinsic::amdgcn_atomic_inc:
   case Intrinsic::amdgcn_atomic_dec:
   case Intrinsic::amdgcn_ds_ordered_add:
case Intrinsic::amdgcn_struct_buffer_store_format: {
     // Handle BUFFER_STORE_BYTE/SHORT overloaded intrinsics
     EVT VDataType = VData.getValueType().getScalarType();
     if (!IsD16 && !VDataVT.isVector() && EltType.getSizeInBits() < 32)
       return handleByteShortBufferStores(DAG, VDataType, DL, Ops, M);
 
     return DAG.getMemIntrinsicNode(Opc, DL, Op->getVTList(), Ops,
                                    M->getMemoryVT(), M->getMemOperand());
   }
+  case Intrinsic::amdgcn_raw_buffer_load_lds: {
+    auto Offsets = splitBufferOffsets(Op.getOperand(4), DAG);
+    auto *M = cast<MemSDNode>(Op);
+    MachineMemOperand *MMO = M->getMemOperand();
+
+    SDValue Ops[] = {
+      Op.getOperand(0), // Chain
+      Op.getOperand(2), // rsrc
+      DAG.getConstant(0, DL, MVT::i32), // vindex
+      Offsets.first,    // voffset
+      Op.getOperand(5), // soffset
+      Offsets.second,   // offset
+      Op.getOperand(6), // cachepolicy, swizzled buffer
+      DAG.getTargetConstant(0, DL, MVT::i1), // idxen
+      DAG.getTargetConstant(MMO->getSize(), DL, MVT::i32), // data byte size
+      copyToM0(DAG, Chain, DL, Op.getOperand(3)).getValue(1) // Glue
+    };
+
+    updateBufferMMO(MMO, Ops[3], Ops[4], Ops[5]);
+
+    return DAG.getMemIntrinsicNode(AMDGPUISD::BUFFER_LOAD, DL,
+                                   M->getVTList(), Ops, M->getMemoryVT(), MMO);
+  }
   case Intrinsic::amdgcn_end_cf:
     return SDValue(DAG.getMachineNode(AMDGPU::SI_END_CF, DL, MVT::Other,
                                       Op->getOperand(2), Chain), 0);
 
   default: {
     if (const AMDGPU::ImageDimIntrinsicInfo *ImageDimIntr =
             AMDGPU::getImageDimIntrinsicInfo(IntrinsicID))
       return lowerImage(Op, ImageDimIntr, DAG, true);

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

if (SOffset) {
         BaseOps.push_back(SOffset);
       else
         Offset += SOffset->getImm();
     }
     // Get appropriate operand, and compute width accordingly.
     DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
     if (DataOpIdx == -1)
       DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
+    if (DataOpIdx == -1) { // LDS DMA
+      Width = (*LdSt.memoperands_begin())->getSize();
+      return true;
+    }
     Width = getOpSize(LdSt, DataOpIdx);
     return true;
   }
 
   if (isMIMG(LdSt)) {
     int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::srsrc);
     BaseOps.push_back(&LdSt.getOperand(SRsrcIdx));
     int VAddr0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr0);

llvm/lib/Target/AMDGPU/SIInstrInfo.td

def SDTBufferLoad : SDTypeProfile<1, 7,
      SDTCisVT<1, v4i32>, // rsrc
      SDTCisVT<2, i32>,   // vindex(VGPR)
      SDTCisVT<3, i32>,   // voffset(VGPR)
      SDTCisVT<4, i32>,   // soffset(SGPR)
      SDTCisVT<5, i32>,   // offset(imm)
      SDTCisVT<6, i32>,   // cachepolicy, swizzled buffer(imm)
      SDTCisVT<7, i1>]>;  // idxen(imm)
 
+def SDTBufferLoadLDS : SDTypeProfile<0, 8,
+    [SDTCisVT<0, v4i32>, // rsrc
+     SDTCisVT<1, i32>,   // vindex(VGPR)
+     SDTCisVT<2, i32>,   // voffset(VGPR)
+     SDTCisVT<3, i32>,   // soffset(SGPR)
+     SDTCisVT<4, i32>,   // offset(imm)
+     SDTCisVT<5, i32>,   // cachepolicy, swizzled buffer(imm)
+     SDTCisVT<6, i1>,    // idxen(imm)
+     SDTCisVT<7, i32>]>; // data byte size
+
 def SIbuffer_load : SDNode <"AMDGPUISD::BUFFER_LOAD", SDTBufferLoad,
                             [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
 def SIbuffer_load_ubyte : SDNode <"AMDGPUISD::BUFFER_LOAD_UBYTE", SDTBufferLoad,
                             [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
 def SIbuffer_load_ushort : SDNode <"AMDGPUISD::BUFFER_LOAD_USHORT", SDTBufferLoad,
                             [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
 def SIbuffer_load_byte : SDNode <"AMDGPUISD::BUFFER_LOAD_BYTE", SDTBufferLoad,
                             [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
 def SIbuffer_load_short: SDNode <"AMDGPUISD::BUFFER_LOAD_SHORT", SDTBufferLoad,
                             [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
 def SIbuffer_load_format : SDNode <"AMDGPUISD::BUFFER_LOAD_FORMAT", SDTBufferLoad,
                             [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
 def SIbuffer_load_format_d16 : SDNode <"AMDGPUISD::BUFFER_LOAD_FORMAT_D16",
                                 SDTBufferLoad,
                                 [SDNPMemOperand, SDNPHasChain, SDNPMayLoad]>;
+def SIbuffer_load_lds : SDNode <"AMDGPUISD::BUFFER_LOAD", SDTBufferLoadLDS,
+                            [SDNPMemOperand, SDNPHasChain, SDNPMayLoad, SDNPMayStore, SDNPInGlue]>;
 
 def SDTBufferStore : SDTypeProfile<0, 8,
     [                    // vdata
      SDTCisVT<1, v4i32>, // rsrc
      SDTCisVT<2, i32>,   // vindex(VGPR)
      SDTCisVT<3, i32>,   // voffset(VGPR)
      SDTCisVT<4, i32>,   // soffset(SGPR)
      SDTCisVT<5, i32>,   // offset(imm)

llvm/lib/Target/AMDGPU/SIInstructions.td

class TBufferLoadGenericInstruction : AMDGPUGenericInstruction {
   let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins type1:$rsrc, type2:$vindex, type2:$voffset,
                            type2:$soffset, untyped_imm_0:$offset, untyped_imm_0:$format,
                            untyped_imm_0:$cachepolicy, untyped_imm_0:$idxen);
   let hasSideEffects = 0;
   let mayLoad = 1;
 }
 
+class BufferLoadLdsGenericInstruction : AMDGPUGenericInstruction {
+  let OutOperandList = (outs);
+  let InOperandList = (ins type1:$rsrc, type2:$vindex, type2:$voffset,
+                           type2:$soffset, untyped_imm_0:$offset,
+                           untyped_imm_0:$cachepolicy, untyped_imm_0:$idxen,
+                           untyped_imm_0:$size);
+  let hasSideEffects = 0;
+  let mayLoad = 1;
+  let mayStore = 1;
+}
+
 def G_AMDGPU_BUFFER_LOAD_UBYTE : BufferLoadGenericInstruction;
 def G_AMDGPU_BUFFER_LOAD_SBYTE : BufferLoadGenericInstruction;
 def G_AMDGPU_BUFFER_LOAD_USHORT : BufferLoadGenericInstruction;
 def G_AMDGPU_BUFFER_LOAD_SSHORT : BufferLoadGenericInstruction;
 def G_AMDGPU_BUFFER_LOAD : BufferLoadGenericInstruction;
 def G_AMDGPU_BUFFER_LOAD_FORMAT : BufferLoadGenericInstruction;
 def G_AMDGPU_BUFFER_LOAD_FORMAT_D16 : BufferLoadGenericInstruction;
 def G_AMDGPU_TBUFFER_LOAD_FORMAT : TBufferLoadGenericInstruction;
 def G_AMDGPU_TBUFFER_LOAD_FORMAT_D16 : TBufferLoadGenericInstruction;
+def G_AMDGPU_BUFFER_LOAD_LDS : BufferLoadLdsGenericInstruction;
 
 class BufferStoreGenericInstruction : AMDGPUGenericInstruction {
   let OutOperandList = (outs);
   let InOperandList = (ins type0:$vdata, type1:$rsrc, type2:$vindex, type2:$voffset,
                            type2:$soffset, untyped_imm_0:$offset,
                            untyped_imm_0:$cachepolicy, untyped_imm_0:$idxen);
   let hasSideEffects = 0;
   let mayStore = 1;

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.buffer.load.lds.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=GCN
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=GCN
+
+declare void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32>, float addrspace(3)* nocapture, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.load.lds.p3v2f32(<4 x i32>, <2 x float> addrspace(3)* nocapture, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.load.lds.p3v3f32(<4 x i32>, <3 x float> addrspace(3)* nocapture, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.load.lds.p3v4f32(<4 x i32>, <4 x float> addrspace(3)* nocapture, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.load.lds.p3i16(<4 x i32>, i16 addrspace(3)* nocapture, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.load.lds.p3f16(<4 x i32>, half addrspace(3)* nocapture, i32, i32, i32)
+declare void @llvm.amdgcn.raw.buffer.load.lds.p3i8(<4 x i32>, i8 addrspace(3)* nocapture, i32, i32, i32)
+
+define amdgpu_ps float @buffer_load_lds(<4 x i32> inreg %rsrc, float addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dword off, s[0:3], 0 lds
+; GCN-NEXT:    buffer_load_dword off, s[0:3], 0 glc lds
+; GCN-NEXT:    buffer_load_dword off, s[0:3], 0 slc lds
+; GCN-NEXT:    v_mov_b32_e32 v0, s4
+; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    ds_read_b32 v0, v0
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    ; return to shader part epilog
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32> %rsrc, float addrspace(3)* %lds, i32 0, i32 0, i32 0)
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32> %rsrc, float addrspace(3)* %lds, i32 0, i32 0, i32 1)
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32> %rsrc, float addrspace(3)* %lds, i32 0, i32 0, i32 2)
+  %res = load float, float addrspace(3)* %lds
+  ret float %res
+}
+
+define amdgpu_ps void @buffer_load_lds_imm_offset(<4 x i32> inreg %rsrc, float addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds_imm_offset:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dword off, s[0:3], 0 offset:2048 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32> %rsrc, float addrspace(3)* %lds, i32 2048, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_v_offset(<4 x i32> inreg %rsrc, float addrspace(3)* inreg %lds, i32 %voffset) {
+; GCN-LABEL: buffer_load_lds_v_offset:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dword v0, s[0:3], 0 offen lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32> %rsrc, float addrspace(3)* %lds, i32 %voffset, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_s_offset(<4 x i32> inreg %rsrc, float addrspace(3)* inreg %lds, i32 inreg %soffset) {
+; GCN-LABEL: buffer_load_lds_s_offset:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dword off, s[0:3], s5 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32> %rsrc, float addrspace(3)* %lds, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_vs_offset(<4 x i32> inreg %rsrc, float addrspace(3)* inreg %lds, i32 %voffset, i32 inreg %soffset) {
+; GCN-LABEL: buffer_load_lds_vs_offset:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dword v0, s[0:3], s5 offen lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f32(<4 x i32> %rsrc, float addrspace(3)* %lds, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_v2f32(<4 x i32> inreg %rsrc, <2 x float> addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds_v2f32:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dwordx2 off, s[0:3], 0 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3v2f32(<4 x i32> %rsrc, <2 x float> addrspace(3)* %lds, i32 0, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_v3f32(<4 x i32> inreg %rsrc, <3 x float> addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds_v3f32:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dwordx3 off, s[0:3], 0 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3v3f32(<4 x i32> %rsrc, <3 x float> addrspace(3)* %lds, i32 0, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_v4f32(<4 x i32> inreg %rsrc, <4 x float> addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds_v4f32:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_dwordx4 off, s[0:3], 0 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3v4f32(<4 x i32> %rsrc, <4 x float> addrspace(3)* %lds, i32 0, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_i16(<4 x i32> inreg %rsrc, i16 addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds_i16:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_ushort off, s[0:3], 0 offset:2048 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3i16(<4 x i32> %rsrc, i16 addrspace(3)* %lds, i32 2048, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_f16(<4 x i32> inreg %rsrc, half addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds_f16:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_ushort off, s[0:3], 0 offset:2048 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3f16(<4 x i32> %rsrc, half addrspace(3)* %lds, i32 2048, i32 0, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_load_lds_i8(<4 x i32> inreg %rsrc, i8 addrspace(3)* inreg %lds) {
+; GCN-LABEL: buffer_load_lds_i8:
+; GCN:       ; %bb.0: ; %main_body
+; GCN-NEXT:    s_mov_b32 m0, s4
+; GCN-NEXT:    buffer_load_ubyte off, s[0:3], 0 offset:2048 lds
+; GCN-NEXT:    s_endpgm
+main_body:
+  call void @llvm.amdgcn.raw.buffer.load.lds.p3i8(<4 x i32> %rsrc, i8 addrspace(3)* %lds, i32 2048, i32 0, i32 0)
+  ret void
+}