AMDGPU/GlobalISel: Handle llvm.amdgcn.ds.{fadd|fmin|fmax}

[arsenm]

Summary

These intrinsics are missing mangling for both the pointer and data
type.

[arsenm]

ping

[madhur13490]

reviewing.

[madhur13490]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

3225

"Helper" can be marked as "const"?

3232

You could probably push this inside loop, run loop from 1 to 5 and skip 2 with a comment. That way it would be readable. If its possible then an uptick loop, please.

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.ds.fadd.ll

3

Why just mesa3d? No implication on hsa?

[foad]

Looks obviously fine to me once @madhur13490 is happy.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

3225

It wouldn't be if it were used as normal

3232

removing an operand in the middle of an instruction will force the instruction to shift all of its operands down each time, so it's much less efficient

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.ds.fadd.ll

3

The triple is irrelevant, but we error on the graphics shader conventions for HSA. We use the shaders for the simplest way to get sample SGPRs

[arsenm]

Reorder a RemoveOperand

[arsenm]

63cdc9a49f1fdd7df80dc964bf647f89982a2569

llvm/lib/Target/AMDGPU/AMDGPUGISel.td

 def : GINodeEquiv<G_AMDGPU_TBUFFER_STORE_FORMAT, SItbuffer_store>;
 def : GINodeEquiv<G_AMDGPU_TBUFFER_STORE_FORMAT_D16, SItbuffer_store_d16>;
 
 // FIXME: Check MMO is atomic
 def : GINodeEquiv<G_AMDGPU_ATOMIC_INC, SIatomic_inc>;
 def : GINodeEquiv<G_AMDGPU_ATOMIC_DEC, SIatomic_dec>;
 def : GINodeEquiv<G_AMDGPU_ATOMIC_INC, atomic_inc_glue>;
 def : GINodeEquiv<G_AMDGPU_ATOMIC_DEC, atomic_dec_glue>;
+def : GINodeEquiv<G_AMDGPU_ATOMIC_FMIN, SIatomic_fmin>;
+def : GINodeEquiv<G_AMDGPU_ATOMIC_FMAX, SIatomic_fmax>;
+def : GINodeEquiv<G_AMDGPU_ATOMIC_FMIN, atomic_load_fmin_glue>;
+def : GINodeEquiv<G_AMDGPU_ATOMIC_FMAX, atomic_load_fmax_glue>;
+
 
 def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SWAP, SIbuffer_atomic_swap>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_ADD, SIbuffer_atomic_add>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SUB, SIbuffer_atomic_sub>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SMIN, SIbuffer_atomic_smin>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_UMIN, SIbuffer_atomic_umin>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SMAX, SIbuffer_atomic_smax>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_UMAX, SIbuffer_atomic_umax>;

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

bool AMDGPUInstructionSelector::select(MachineInstr &I) {
   case TargetOpcode::G_ATOMICRMW_XOR:
   case TargetOpcode::G_ATOMICRMW_MIN:
   case TargetOpcode::G_ATOMICRMW_MAX:
   case TargetOpcode::G_ATOMICRMW_UMIN:
   case TargetOpcode::G_ATOMICRMW_UMAX:
   case TargetOpcode::G_ATOMICRMW_FADD:
   case AMDGPU::G_AMDGPU_ATOMIC_INC:
   case AMDGPU::G_AMDGPU_ATOMIC_DEC:
+  case AMDGPU::G_AMDGPU_ATOMIC_FMIN:
+  case AMDGPU::G_AMDGPU_ATOMIC_FMAX:
     return selectG_LOAD_ATOMICRMW(I);
   case AMDGPU::G_AMDGPU_ATOMIC_CMPXCHG:
     return selectG_AMDGPU_ATOMIC_CMPXCHG(I);
   case TargetOpcode::G_SELECT:
     return selectG_SELECT(I);
   case TargetOpcode::G_STORE:
     return selectG_STORE(I);
   case TargetOpcode::G_TRUNC:

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.h

public:
   bool legalizeFastUnsafeFDIV(MachineInstr &MI, MachineRegisterInfo &MRI,
                               MachineIRBuilder &B) const;
   bool legalizeFDIVFastIntrin(MachineInstr &MI, MachineRegisterInfo &MRI,
                               MachineIRBuilder &B) const;
 
   bool legalizeRsqClampIntrinsic(MachineInstr &MI, MachineRegisterInfo &MRI,
                                  MachineIRBuilder &B) const;
 
+  bool legalizeDSAtomicFPIntrinsic(LegalizerHelper &Helper,
+                                   MachineInstr &MI, Intrinsic::ID IID) const;
+
   bool getImplicitArgPtr(Register DstReg, MachineRegisterInfo &MRI,
                          MachineIRBuilder &B) const;
 
   bool legalizeImplicitArgPtr(MachineInstr &MI, MachineRegisterInfo &MRI,
                               MachineIRBuilder &B) const;
   bool legalizeIsAddrSpace(MachineInstr &MI, MachineRegisterInfo &MRI,
                            MachineIRBuilder &B, unsigned AddrSpace) const;
 

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

bool AMDGPULegalizerInfo::legalizeRsqClampIntrinsic(MachineInstr &MI,
   if (UseIEEE)
     B.buildFMaxNumIEEE(Dst, ClampMax, MinFlt, Flags);
   else
     B.buildFMaxNum(Dst, ClampMax, MinFlt, Flags);
   MI.eraseFromParent();
   return true;
 }
 
+static unsigned getDSFPAtomicOpcode(Intrinsic::ID IID) {
+  switch (IID) {
+  case Intrinsic::amdgcn_ds_fadd:
+    return AMDGPU::G_ATOMICRMW_FADD;
+  case Intrinsic::amdgcn_ds_fmin:
+    return AMDGPU::G_AMDGPU_ATOMIC_FMIN;
+  case Intrinsic::amdgcn_ds_fmax:
+    return AMDGPU::G_AMDGPU_ATOMIC_FMAX;
+  default:
+    llvm_unreachable("not a DS FP intrinsic");
+  }
+}
+
+bool AMDGPULegalizerInfo::legalizeDSAtomicFPIntrinsic(LegalizerHelper &Helper,
+                                                      MachineInstr &MI,
+                                                      Intrinsic::ID IID) const {
+  GISelChangeObserver &Observer = Helper.Observer;
+  Observer.changingInstr(MI);
+
+  MI.setDesc(ST.getInstrInfo()->get(getDSFPAtomicOpcode(IID)));
+
+  // The remaining operands were used to set fields in the MemOperand on
+  // construction.
+  for (int I = 6; I > 3; --I)
+    MI.RemoveOperand(I);
+
+  MI.RemoveOperand(1); // Remove the intrinsic ID.
+  Observer.changedInstr(MI);
+  return true;
+}
+
 bool AMDGPULegalizerInfo::getImplicitArgPtr(Register DstReg,
                                             MachineRegisterInfo &MRI,
                                             MachineIRBuilder &B) const {
   uint64_t Offset =
     ST.getTargetLowering()->getImplicitParameterOffset(
       B.getMF(), AMDGPUTargetLowering::FIRST_IMPLICIT);
   LLT DstTy = MRI.getType(DstReg);
   LLT IdxTy = LLT::scalar(DstTy.getSizeInBits());
bool AMDGPULegalizerInfo::legalizeIntrinsic(LegalizerHelper &Helper,
   case Intrinsic::amdgcn_atomic_dec:
     return legalizeAtomicIncDec(MI, B, false);
   case Intrinsic::trap:
     return legalizeTrapIntrinsic(MI, MRI, B);
   case Intrinsic::debugtrap:
     return legalizeDebugTrapIntrinsic(MI, MRI, B);
   case Intrinsic::amdgcn_rsq_clamp:
     return legalizeRsqClampIntrinsic(MI, MRI, B);
+  case Intrinsic::amdgcn_ds_fadd:
+  case Intrinsic::amdgcn_ds_fmin:
+  case Intrinsic::amdgcn_ds_fmax:
+    return legalizeDSAtomicFPIntrinsic(Helper, MI, IntrID);
   default: {
     if (const AMDGPU::ImageDimIntrinsicInfo *ImageDimIntr =
             AMDGPU::getImageDimIntrinsicInfo(IntrID))
       return legalizeImageIntrinsic(MI, B, Helper.Observer, ImageDimIntr);
     return true;
   }
   }
 
   return true;
 }

llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

case AMDGPU::G_INTRINSIC_W_SIDE_EFFECTS: {
     case Intrinsic::amdgcn_s_getreg:
     case Intrinsic::amdgcn_s_memtime:
     case Intrinsic::amdgcn_s_memrealtime:
     case Intrinsic::amdgcn_s_get_waveid_in_workgroup: {
       unsigned Size = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
       OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::SGPRRegBankID, Size);
       break;
     }
-    case Intrinsic::amdgcn_ds_fadd:
-    case Intrinsic::amdgcn_ds_fmin:
-    case Intrinsic::amdgcn_ds_fmax:
     case Intrinsic::amdgcn_global_atomic_csub:
       return getDefaultMappingAllVGPR(MI);
     case Intrinsic::amdgcn_ds_ordered_add:
     case Intrinsic::amdgcn_ds_ordered_swap: {
       unsigned DstSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
       OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, DstSize);
       unsigned M0Bank = getRegBankID(MI.getOperand(2).getReg(), MRI,
                                  AMDGPU::SGPRRegBankID);

   case AMDGPU::G_ATOMICRMW_XOR:
   case AMDGPU::G_ATOMICRMW_MAX:
   case AMDGPU::G_ATOMICRMW_MIN:
   case AMDGPU::G_ATOMICRMW_UMAX:
   case AMDGPU::G_ATOMICRMW_UMIN:
   case AMDGPU::G_ATOMICRMW_FADD:
   case AMDGPU::G_AMDGPU_ATOMIC_CMPXCHG:
   case AMDGPU::G_AMDGPU_ATOMIC_INC:
-  case AMDGPU::G_AMDGPU_ATOMIC_DEC: {
+  case AMDGPU::G_AMDGPU_ATOMIC_DEC:
+  case AMDGPU::G_AMDGPU_ATOMIC_FMIN:
+  case AMDGPU::G_AMDGPU_ATOMIC_FMAX: {
     OpdsMapping[0] = getVGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
     OpdsMapping[1] = getValueMappingForPtr(MRI, MI.getOperand(1).getReg());
     OpdsMapping[2] = getVGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);
     break;
   }
   case AMDGPU::G_ATOMIC_CMPXCHG: {
     OpdsMapping[0] = getVGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
     OpdsMapping[1] = getValueMappingForPtr(MRI, MI.getOperand(1).getReg());

llvm/lib/Target/AMDGPU/SIInstructions.td

def G_AMDGPU_ATOMIC_CMPXCHG : AMDGPUGenericInstruction {
   let hasSideEffects = 0;
   let mayLoad = 1;
   let mayStore = 1;
 }
 
 let Namespace = "AMDGPU" in {
 def G_AMDGPU_ATOMIC_INC : G_ATOMICRMW_OP;
 def G_AMDGPU_ATOMIC_DEC : G_ATOMICRMW_OP;
+def G_AMDGPU_ATOMIC_FMIN : G_ATOMICRMW_OP;
+def G_AMDGPU_ATOMIC_FMAX : G_ATOMICRMW_OP;
 }
 
 class BufferAtomicGenericInstruction : AMDGPUGenericInstruction {
   let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins type0:$vdata, type1:$rsrc, type2:$vindex, type2:$voffset,
                            type2:$soffset, untyped_imm_0:$offset,
                            untyped_imm_0:$cachepolicy, untyped_imm_0:$idxen);
   let hasSideEffects = 0;

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.ds.fadd.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX8 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX9 %s
+
+define amdgpu_ps float @ds_fadd_f32_ss(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fadd_f32_ss:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    ; return to shader part epilog
+;
+; GFX9-LABEL: ds_fadd_f32_ss:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v0, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    ; return to shader part epilog
+  %ret = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define amdgpu_ps float @ds_fadd_f32_ss_offset(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fadd_f32_ss_offset:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    ; return to shader part epilog
+;
+; GFX9-LABEL: ds_fadd_f32_ss_offset:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    ; return to shader part epilog
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define amdgpu_ps void @ds_fadd_f32_ss_nortn(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fadd_f32_ss_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: ds_fadd_f32_ss_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v0, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_endpgm
+  %unused = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define amdgpu_ps void @ds_fadd_f32_ss_offset_nortn(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fadd_f32_ss_offset_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: ds_fadd_f32_ss_offset_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_endpgm
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %unused = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define float @ds_fadd_f32_vv(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fadd_f32_vv:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fadd_f32_vv:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define float @ds_fadd_f32_vv_offset(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fadd_f32_vv_offset:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1 offset:512
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fadd_f32_vv_offset:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1 offset:512
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define void @ds_fadd_f32_vv_nortn(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fadd_f32_vv_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fadd_f32_vv_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define void @ds_fadd_f32_vv_offset_nortn(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fadd_f32_vv_offset_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1 offset:512
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fadd_f32_vv_offset_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1 offset:512
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define float @ds_fadd_f32_vv_volatile(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fadd_f32_vv_volatile:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fadd_f32_vv_volatile:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_add_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call float @llvm.amdgcn.ds.fadd(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 true)
+  ret float %ret
+}
+
+declare float @llvm.amdgcn.ds.fadd(float addrspace(3)* nocapture, float, i32 immarg, i32 immarg, i1 immarg) #0
+
+attributes #0 = { argmemonly nounwind willreturn }

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.ds.fmax.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX8 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX9 %s
+
+; Make sure the memory operand information is preserved.
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -stop-after=instruction-select -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX8-MIR %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -stop-after=instruction-select -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX9-MIR %s
+
+
+define amdgpu_ps float @ds_fmax_f32_ss(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmax_f32_ss:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    ; return to shader part epilog
+;
+; GFX9-LABEL: ds_fmax_f32_ss:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v0, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    ; return to shader part epilog
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_ss
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[COPY]]
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY2]], [[COPY3]], 0, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX8-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_]]
+  ; GFX8-MIR:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_ss
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[COPY]]
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY2]], [[COPY3]], 0, 0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX9-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_gfx9_]]
+  ; GFX9-MIR:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define amdgpu_ps float @ds_fmax_f32_ss_offset(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmax_f32_ss_offset:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    ; return to shader part epilog
+;
+; GFX9-LABEL: ds_fmax_f32_ss_offset:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    ; return to shader part epilog
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_ss_offset
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX8-MIR:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 512
+  ; GFX8-MIR:   [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY]], [[S_MOV_B32_]], implicit-def $scc
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_ADD_U32_]]
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY2]], [[COPY3]], 0, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX8-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_]]
+  ; GFX8-MIR:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_ss_offset
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX9-MIR:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 512
+  ; GFX9-MIR:   [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY]], [[S_MOV_B32_]], implicit-def $scc
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_ADD_U32_]]
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY2]], [[COPY3]], 0, 0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX9-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_gfx9_]]
+  ; GFX9-MIR:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define amdgpu_ps void @ds_fmax_f32_ss_nortn(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmax_f32_ss_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: ds_fmax_f32_ss_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v0, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_endpgm
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_ss_nortn
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[COPY]]
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY2]], [[COPY3]], 0, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX8-MIR:   S_ENDPGM 0
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_ss_nortn
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[COPY]]
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY2]], [[COPY3]], 0, 0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX9-MIR:   S_ENDPGM 0
+  %unused = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define amdgpu_ps void @ds_fmax_f32_ss_offset_nortn(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmax_f32_ss_offset_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: ds_fmax_f32_ss_offset_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_endpgm
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_ss_offset_nortn
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX8-MIR:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 512
+  ; GFX8-MIR:   [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY]], [[S_MOV_B32_]], implicit-def $scc
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_ADD_U32_]]
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY2]], [[COPY3]], 0, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX8-MIR:   S_ENDPGM 0
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_ss_offset_nortn
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $sgpr2, $sgpr3
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX9-MIR:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 512
+  ; GFX9-MIR:   [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY]], [[S_MOV_B32_]], implicit-def $scc
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_ADD_U32_]]
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[COPY1]]
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY2]], [[COPY3]], 0, 0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX9-MIR:   S_ENDPGM 0
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %unused = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define float @ds_fmax_f32_vv(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmax_f32_vv:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmax_f32_vv:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_vv
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY]], [[COPY1]], 0, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX8-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_]]
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX8-MIR:   S_SETPC_B64_return [[COPY3]], implicit $vgpr0
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_vv
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY]], [[COPY1]], 0, 0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX9-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_gfx9_]]
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX9-MIR:   S_SETPC_B64_return [[COPY3]], implicit $vgpr0
+  %ret = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define float @ds_fmax_f32_vv_offset(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmax_f32_vv_offset:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1 offset:512
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmax_f32_vv_offset:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1 offset:512
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_vv_offset
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY]], [[COPY1]], 512, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX8-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_]]
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX8-MIR:   S_SETPC_B64_return [[COPY3]], implicit $vgpr0
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_vv_offset
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY]], [[COPY1]], 512, 0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX9-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_gfx9_]]
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX9-MIR:   S_SETPC_B64_return [[COPY3]], implicit $vgpr0
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define void @ds_fmax_f32_vv_nortn(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmax_f32_vv_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmax_f32_vv_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_vv_nortn
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY]], [[COPY1]], 0, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX8-MIR:   S_SETPC_B64_return [[COPY3]]
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_vv_nortn
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY]], [[COPY1]], 0, 0, implicit $exec :: (load store 4 on %ir.ptr, addrspace 3)
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX9-MIR:   S_SETPC_B64_return [[COPY3]]
+  %ret = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define void @ds_fmax_f32_vv_offset_nortn(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmax_f32_vv_offset_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1 offset:512
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmax_f32_vv_offset_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1 offset:512
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_vv_offset_nortn
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY]], [[COPY1]], 512, 0, implicit $m0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX8-MIR:   S_SETPC_B64_return [[COPY3]]
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_vv_offset_nortn
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY]], [[COPY1]], 512, 0, implicit $exec :: (load store 4 on %ir.gep, addrspace 3)
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX9-MIR:   S_SETPC_B64_return [[COPY3]]
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define float @ds_fmax_f32_vv_volatile(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmax_f32_vv_volatile:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmax_f32_vv_volatile:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_max_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  ; GFX8-MIR-LABEL: name: ds_fmax_f32_vv_volatile
+  ; GFX8-MIR: bb.1 (%ir-block.0):
+  ; GFX8-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX8-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX8-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX8-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX8-MIR:   $m0 = S_MOV_B32 -1
+  ; GFX8-MIR:   [[DS_MAX_RTN_F32_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32 [[COPY]], [[COPY1]], 0, 0, implicit $m0, implicit $exec :: (volatile load store 4 on %ir.ptr, addrspace 3)
+  ; GFX8-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_]]
+  ; GFX8-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX8-MIR:   S_SETPC_B64_return [[COPY3]], implicit $vgpr0
+  ; GFX9-MIR-LABEL: name: ds_fmax_f32_vv_volatile
+  ; GFX9-MIR: bb.1 (%ir-block.0):
+  ; GFX9-MIR:   liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31
+  ; GFX9-MIR:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX9-MIR:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX9-MIR:   [[COPY2:%[0-9]+]]:sgpr_64 = COPY $sgpr30_sgpr31
+  ; GFX9-MIR:   [[DS_MAX_RTN_F32_gfx9_:%[0-9]+]]:vgpr_32 = DS_MAX_RTN_F32_gfx9 [[COPY]], [[COPY1]], 0, 0, implicit $exec :: (volatile load store 4 on %ir.ptr, addrspace 3)
+  ; GFX9-MIR:   $vgpr0 = COPY [[DS_MAX_RTN_F32_gfx9_]]
+  ; GFX9-MIR:   [[COPY3:%[0-9]+]]:ccr_sgpr_64 = COPY [[COPY2]]
+  ; GFX9-MIR:   S_SETPC_B64_return [[COPY3]], implicit $vgpr0
+  %ret = call float @llvm.amdgcn.ds.fmax(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 true)
+  ret float %ret
+}
+
+declare float @llvm.amdgcn.ds.fmax(float addrspace(3)* nocapture, float, i32 immarg, i32 immarg, i1 immarg) #0
+
+attributes #0 = { argmemonly nounwind willreturn }

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.ds.fmin.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX8 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX9 %s
+
+define amdgpu_ps float @ds_fmin_f32_ss(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmin_f32_ss:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    ; return to shader part epilog
+;
+; GFX9-LABEL: ds_fmin_f32_ss:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v0, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    ; return to shader part epilog
+  %ret = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define amdgpu_ps float @ds_fmin_f32_ss_offset(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmin_f32_ss_offset:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    ; return to shader part epilog
+;
+; GFX9-LABEL: ds_fmin_f32_ss_offset:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    ; return to shader part epilog
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define amdgpu_ps void @ds_fmin_f32_ss_nortn(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmin_f32_ss_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: ds_fmin_f32_ss_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v0, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_endpgm
+  %unused = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define amdgpu_ps void @ds_fmin_f32_ss_offset_nortn(float addrspace(3)* inreg %ptr, float inreg %val) {
+; GFX8-LABEL: ds_fmin_f32_ss_offset_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: ds_fmin_f32_ss_offset_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_add_u32 s0, s2, 0x200
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_endpgm
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %unused = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define float @ds_fmin_f32_vv(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmin_f32_vv:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmin_f32_vv:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define float @ds_fmin_f32_vv_offset(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmin_f32_vv_offset:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1 offset:512
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmin_f32_vv_offset:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1 offset:512
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret float %ret
+}
+
+define void @ds_fmin_f32_vv_nortn(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmin_f32_vv_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmin_f32_vv_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define void @ds_fmin_f32_vv_offset_nortn(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmin_f32_vv_offset_nortn:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1 offset:512
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmin_f32_vv_offset_nortn:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1 offset:512
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %gep = getelementptr float, float addrspace(3)* %ptr, i32 128
+  %ret = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %gep, float %val, i32 0, i32 0, i1 false)
+  ret void
+}
+
+define float @ds_fmin_f32_vv_volatile(float addrspace(3)* %ptr, float %val) {
+; GFX8-LABEL: ds_fmin_f32_vv_volatile:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b32 m0, -1
+; GFX8-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: ds_fmin_f32_vv_volatile:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    ds_min_rtn_f32 v0, v0, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call float @llvm.amdgcn.ds.fmin(float addrspace(3)* %ptr, float %val, i32 0, i32 0, i1 true)
+  ret float %ret
+}
+
+declare float @llvm.amdgcn.ds.fmin(float addrspace(3)* nocapture, float, i32 immarg, i32 immarg, i1 immarg) #0
+
+attributes #0 = { argmemonly nounwind willreturn }

llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.ds.fmax.mir

-# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect -regbankselect-fast -verify-machineinstrs %s -o - | FileCheck %s
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect -regbankselect-greedy -verify-machineinstrs %s -o - | FileCheck %s
-
----
-name: ds_fmax_ss
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $sgpr0, $sgpr1
-
-    ; CHECK-LABEL: name: ds_fmax_ss
-    ; CHECK: liveins: $sgpr0, $sgpr1
-    ; CHECK: [[COPY:%[0-9]+]]:sgpr(p3) = COPY $sgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
-    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(p3) = COPY [[COPY]](p3)
-    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), [[COPY2]](p3), [[COPY3]](s32), 0, 0, 0
-    %0:_(p3) = COPY $sgpr0
-    %1:_(s32) = COPY $sgpr1
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), %0, %1, 0, 0, 0
-
-...
-
----
-name: ds_fmax_sv
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $sgpr0, $vgpr0
-
-    ; CHECK-LABEL: name: ds_fmax_sv
-    ; CHECK: liveins: $sgpr0, $vgpr0
-    ; CHECK: [[COPY:%[0-9]+]]:sgpr(p3) = COPY $sgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
-    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(p3) = COPY [[COPY]](p3)
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), [[COPY2]](p3), [[COPY1]](s32), 0, 0, 0
-    %0:_(p3) = COPY $sgpr0
-    %1:_(s32) = COPY $vgpr0
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), %0, %1, 0, 0, 0
-
-...
-
----
-name: ds_fmax_vs
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $vgpr0, $sgpr0
-
-    ; CHECK-LABEL: name: ds_fmax_vs
-    ; CHECK: liveins: $vgpr0, $sgpr0
-    ; CHECK: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
-    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), [[COPY]](p3), [[COPY2]](s32), 0, 0, 0
-    %0:_(p3) = COPY $vgpr0
-    %1:_(s32) = COPY $sgpr0
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), %0, %1, 0, 0, 0
-
-...
-
----
-name: ds_fmax_vv
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $vgpr0, $vgpr1
-
-    ; CHECK-LABEL: name: ds_fmax_vv
-    ; CHECK: liveins: $vgpr0, $vgpr1
-    ; CHECK: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), [[COPY]](p3), [[COPY1]](s32), 0, 0, 0
-    %0:_(p3) = COPY $vgpr0
-    %1:_(s32) = COPY $vgpr1
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmax), %0, %1, 0, 0, 0
-
-...

llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.ds.fmin.mir

-# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect -regbankselect-fast -verify-machineinstrs %s -o - | FileCheck %s
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect -regbankselect-greedy -verify-machineinstrs %s -o - | FileCheck %s
-
----
-name: ds_fmin_ss
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $sgpr0, $sgpr1
-
-    ; CHECK-LABEL: name: ds_fmin_ss
-    ; CHECK: liveins: $sgpr0, $sgpr1
-    ; CHECK: [[COPY:%[0-9]+]]:sgpr(p3) = COPY $sgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
-    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(p3) = COPY [[COPY]](p3)
-    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), [[COPY2]](p3), [[COPY3]](s32), 0, 0, 0
-    %0:_(p3) = COPY $sgpr0
-    %1:_(s32) = COPY $sgpr1
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), %0, %1, 0, 0, 0
-
-...
-
----
-name: ds_fmin_sv
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $sgpr0, $vgpr0
-
-    ; CHECK-LABEL: name: ds_fmin_sv
-    ; CHECK: liveins: $sgpr0, $vgpr0
-    ; CHECK: [[COPY:%[0-9]+]]:sgpr(p3) = COPY $sgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
-    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(p3) = COPY [[COPY]](p3)
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), [[COPY2]](p3), [[COPY1]](s32), 0, 0, 0
-    %0:_(p3) = COPY $sgpr0
-    %1:_(s32) = COPY $vgpr0
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), %0, %1, 0, 0, 0
-
-...
-
----
-name: ds_fmin_vs
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $vgpr0, $sgpr0
-
-    ; CHECK-LABEL: name: ds_fmin_vs
-    ; CHECK: liveins: $vgpr0, $sgpr0
-    ; CHECK: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
-    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[COPY1]](s32)
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), [[COPY]](p3), [[COPY2]](s32), 0, 0, 0
-    %0:_(p3) = COPY $vgpr0
-    %1:_(s32) = COPY $sgpr0
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), %0, %1, 0, 0, 0
-
-...
-
----
-name: ds_fmin_vv
-legalized: true
-tracksRegLiveness: true
-body: |
-  bb.0:
-    liveins: $vgpr0, $vgpr1
-
-    ; CHECK-LABEL: name: ds_fmin_vv
-    ; CHECK: liveins: $vgpr0, $vgpr1
-    ; CHECK: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
-    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
-    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), [[COPY]](p3), [[COPY1]](s32), 0, 0, 0
-    %0:_(p3) = COPY $vgpr0
-    %1:_(s32) = COPY $vgpr1
-    %2:_(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.ds.fmin), %0, %1, 0, 0, 0
-
-...