[AMDGPU] Support mov dpp with 64 bit operands

[rampitec]

Summary

We define mov/update dpp intrinsics as overloaded but do not
support i64, which is a practically useful type. Fix the
selection and lowering.

[arsenm]

llvm/lib/Target/AMDGPU/SIInstructions.td

1869–1875

Why not do the split here? Why treat it as a post-RA pseudo? At latest I would have expected this to be expanded in FinalizeISel

[rampitec]

GCNDPPCombiner can split the new pseudo and then handle the split.
Post-RA split is needed anyway since combining is an optimization.
Tests are updated to handle case w/o optimization.

[rampitec]

Rebased.
Removed special handling of gfx10, it uses the same pseudo now.

[vpykhtin]

llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp

575

Changed should depend on the Split done.

llvm/lib/Target/AMDGPU/SIInstrInfo.h

232

please add short desc

[rampitec]

Addressed comments.

[arsenm]

llvm/lib/Target/AMDGPU/SIInstructions.td

1869–1875

I still don't see the point of the pseudo. You can just emit the reg_sequence directly here?

[rampitec]

Rebased.

[vpykhtin]

LGTM

llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp

bool GCNDPPCombine::runOnMachineFunction(MachineFunction &MF) {
 
   bool Changed = false;
   for (auto &MBB : MF) {
     for (auto I = MBB.rbegin(), E = MBB.rend(); I != E;) {
       auto &MI = *I++;
       if (MI.getOpcode() == AMDGPU::V_MOV_B32_dpp && combineDPPMov(MI)) {
         Changed = true;
         ++NumDPPMovsCombined;
+      } else if (MI.getOpcode() == AMDGPU::V_MOV_B64_DPP_PSEUDO) {
+        auto Split = TII->expandMovDPP64(MI);
+        for (auto M : { Split.first, Split.second }) {
+          if (combineDPPMov(*M))
+            ++NumDPPMovsCombined;
+        }
+        Changed = true;
       }
     }
   }
   return Changed;
 }

llvm/lib/Target/AMDGPU/SIInstrInfo.h

public:
 
   void loadRegFromStackSlot(MachineBasicBlock &MBB,
                             MachineBasicBlock::iterator MI, unsigned DestReg,
                             int FrameIndex, const TargetRegisterClass *RC,
                             const TargetRegisterInfo *TRI) const override;
 
   bool expandPostRAPseudo(MachineInstr &MI) const override;
 
+  // Splits a V_MOV_B64_DPP_PSEUDO opcode into a pair of v_mov_b32_dpp
+  // instructions. Returns a pair of generated instructions.
+  // Can split either post-RA with physical registers or pre-RA with
+  // virtual registers. In latter case IR needs to be in SSA form and
+  // and a REG_SEQUENCE is produced to define original register.
+  std::pair<MachineInstr*, MachineInstr*>
+  expandMovDPP64(MachineInstr &MI) const;
+
   // Returns an opcode that can be used to move a value to a \p DstRC
   // register.  If there is no hardware instruction that can store to \p
   // DstRC, then AMDGPU::COPY is returned.
   unsigned getMovOpcode(const TargetRegisterClass *DstRC) const;
 
   LLVM_READONLY
   int commuteOpcode(unsigned Opc) const;
 

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

if (SrcOp.isImm()) {
         .addReg(Dst, RegState::Implicit | RegState::Define);
       BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
         .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
         .addReg(Dst, RegState::Implicit | RegState::Define);
     }
     MI.eraseFromParent();
     break;
   }
+  case AMDGPU::V_MOV_B64_DPP_PSEUDO: {
+    expandMovDPP64(MI);
+    break;
+  }
   case AMDGPU::V_SET_INACTIVE_B32: {
     unsigned NotOpc = ST.isWave32() ? AMDGPU::S_NOT_B32 : AMDGPU::S_NOT_B64;
     unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
     BuildMI(MBB, MI, DL, get(NotOpc), Exec)
       .addReg(Exec);
     BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), MI.getOperand(0).getReg())
       .add(MI.getOperand(2));
     BuildMI(MBB, MI, DL, get(NotOpc), Exec)
case TargetOpcode::BUNDLE: {
 
     MI.eraseFromParent();
     break;
   }
   }
   return true;
 }
 
+std::pair<MachineInstr*, MachineInstr*>
+SIInstrInfo::expandMovDPP64(MachineInstr &MI) const {
+  assert (MI.getOpcode() == AMDGPU::V_MOV_B64_DPP_PSEUDO);
+
+  MachineBasicBlock &MBB = *MI.getParent();
+  DebugLoc DL = MBB.findDebugLoc(MI);
+  MachineFunction *MF = MBB.getParent();
+  MachineRegisterInfo &MRI = MF->getRegInfo();
+  Register Dst = MI.getOperand(0).getReg();
+  unsigned Part = 0;
+  MachineInstr *Split[2];
+
+
+  for (auto Sub : { AMDGPU::sub0, AMDGPU::sub1 }) {
+    auto MovDPP = BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_dpp));
+    if (Dst.isPhysical()) {
+      MovDPP.addDef(RI.getSubReg(Dst, Sub));
+    } else {
+      assert(MRI.isSSA());
+      auto Tmp = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+      MovDPP.addDef(Tmp);
+    }
+
+    for (unsigned I = 1; I <= 2; ++I) { // old and src operands.
+      const MachineOperand &SrcOp = MI.getOperand(I);
+      assert(!SrcOp.isFPImm());
+      if (SrcOp.isImm()) {
+        APInt Imm(64, SrcOp.getImm());
+        Imm.ashrInPlace(Part * 32);
+        MovDPP.addImm(Imm.getLoBits(32).getZExtValue());
+      } else {
+        assert(SrcOp.isReg());
+        Register Src = SrcOp.getReg();
+        if (Src.isPhysical())
+          MovDPP.addReg(RI.getSubReg(Src, Sub));
+        else
+          MovDPP.addReg(Src, SrcOp.isUndef() ? RegState::Undef : 0, Sub);
+      }
+    }
+
+    for (unsigned I = 3; I < MI.getNumExplicitOperands(); ++I)
+      MovDPP.addImm(MI.getOperand(I).getImm());
+
+    Split[Part] = MovDPP;
+    ++Part;
+  }
+
+  if (Dst.isVirtual())
+    BuildMI(MBB, MI, DL, get(AMDGPU::REG_SEQUENCE), Dst)
+      .addReg(Split[0]->getOperand(0).getReg())
+      .addImm(AMDGPU::sub0)
+      .addReg(Split[1]->getOperand(0).getReg())
+      .addImm(AMDGPU::sub1);
+
+  MI.eraseFromParent();
+  return std::make_pair(Split[0], Split[1]);
+}
+
 bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
                                       MachineOperand &Src0,
                                       unsigned Src0OpName,
                                       MachineOperand &Src1,
                                       unsigned Src1OpName) const {
   MachineOperand *Src0Mods = getNamedOperand(MI, Src0OpName);
   if (!Src0Mods)
     return false;

llvm/lib/Target/AMDGPU/SIInstructions.td

 def ATOMIC_FENCE : SPseudoInstSI<
   (outs), (ins i32imm:$ordering, i32imm:$scope),
   [(atomic_fence (i32 imm:$ordering), (i32 imm:$scope))],
   "ATOMIC_FENCE $ordering, $scope"> {
   let hasSideEffects = 1;
   let maybeAtomic = 1;
 }
 
+def VOP_I64_I64_DPP : VOPProfile <[i64, i64, untyped, untyped]> {
+  let HasExt = 1;
+  let HasExtDPP = 1;
+}
+
 let hasSideEffects = 0, mayLoad = 0, mayStore = 0, Uses = [EXEC] in {
 
 // For use in patterns
 def V_CNDMASK_B64_PSEUDO : VOP3Common <(outs VReg_64:$vdst),
   (ins VSrc_b64:$src0, VSrc_b64:$src1, SSrc_b64:$src2), "", []> {
   let isPseudo = 1;
   let isCodeGenOnly = 1;
   let usesCustomInserter = 1;
 }
 
 // 64-bit vector move instruction. This is mainly used by the
 // SIFoldOperands pass to enable folding of inline immediates.
 def V_MOV_B64_PSEUDO : VPseudoInstSI <(outs VReg_64:$vdst),
                                       (ins VSrc_b64:$src0)>;
 
+// 64-bit vector move with dpp. Expanded post-RA.
+def V_MOV_B64_DPP_PSEUDO : VOP_DPP_Pseudo <"v_mov_b64_dpp", VOP_I64_I64_DPP> {
+  let Size = 16; // Requires two 8-byte v_mov_b32_dpp to complete.
+}
+
 // Pseudoinstruction for @llvm.amdgcn.wqm. It is turned into a copy after the
 // WQM pass processes it.
 def WQM : PseudoInstSI <(outs unknown:$vdst), (ins unknown:$src0)>;
 
 // Pseudoinstruction for @llvm.amdgcn.softwqm. Like @llvm.amdgcn.wqm it is
 // turned into a copy by WQM pass, but does not seed WQM requirements.
 def SOFT_WQM : PseudoInstSI <(outs unknown:$vdst), (ins unknown:$src0)>;
 
def : GCNPat <
   (INSERT_SUBREG (IMPLICIT_DEF), $src0, sub0)
 >;
 
 def : GCNPat <
   (v4f16 (scalar_to_vector f16:$src0)),
   (INSERT_SUBREG (IMPLICIT_DEF), $src0, sub0)
 >;
 
+def : GCNPat <
+  (i64 (int_amdgcn_mov_dpp i64:$src, timm:$dpp_ctrl, timm:$row_mask, timm:$bank_mask,
+                           timm:$bound_ctrl)),
+  (V_MOV_B64_DPP_PSEUDO $src, $src, (as_i32imm $dpp_ctrl),
+                        (as_i32imm $row_mask), (as_i32imm $bank_mask),
+                        (as_i1imm $bound_ctrl))
+>;
+
+def : GCNPat <
+  (i64 (int_amdgcn_update_dpp i64:$old, i64:$src, timm:$dpp_ctrl, timm:$row_mask,
+                              timm:$bank_mask, timm:$bound_ctrl)),
+  (V_MOV_B64_DPP_PSEUDO $old, $src, (as_i32imm $dpp_ctrl),
+                        (as_i32imm $row_mask), (as_i32imm $bank_mask),
+                        (as_i1imm $bound_ctrl))
+>;
+
 //===----------------------------------------------------------------------===//
 // Fract Patterns
 //===----------------------------------------------------------------------===//
 
 let SubtargetPredicate = isGFX6 in {
 
 // V_FRACT is buggy on SI, so the F32 version is never used and (x-floor(x)) is
 // used instead. However, SI doesn't have V_FLOOR_F64, so the most efficient

llvm/test/CodeGen/AMDGPU/dpp_combine.mir

bb.0:
     %8:vgpr_32 = V_MOV_B32_e32 5, implicit $exec
     %2:vgpr_32 = V_MOV_B32_dpp %0.sub0, %1.sub0, 1, 15, 15, 1, implicit $exec
     %3:vgpr_32 = V_MOV_B32_dpp %0.sub1, %1.sub1, 1, 15, 15, 1, implicit $exec
     %4:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, %3, %subreg.sub1
     %5:vreg_64 = REG_SEQUENCE %4.sub0, %subreg.sub0, %4.sub1, %subreg.sub1
     %6:vgpr_32 = V_ADD_I32_e32 %5.sub0, %8, implicit-def $vcc, implicit $exec
     %7:vgpr_32 = V_ADDC_U32_e32 %5.sub1, %8, implicit-def $vcc, implicit $vcc, implicit $exec
 ...
+
+# GCN-LABEL: name: dpp64_add64_impdef
+# GCN: %3:vgpr_32 = V_ADD_I32_dpp %1.sub0, %0.sub0, undef %4:vgpr_32, 1, 15, 15, 1, implicit-def $vcc, implicit $exec
+# GCN: %5:vgpr_32 = V_ADDC_U32_dpp %1.sub1, %0.sub1, undef %4:vgpr_32, 1, 15, 15, 1, implicit-def $vcc, implicit $vcc, implicit $exec
+name: dpp64_add64_impdef
+tracksRegLiveness: true
+body: |
+  bb.0:
+    %0:vreg_64 = IMPLICIT_DEF
+    %1:vreg_64 = IMPLICIT_DEF
+    %2:vreg_64 = V_MOV_B64_DPP_PSEUDO %1:vreg_64, %0:vreg_64, 1, 15, 15, 1, implicit $exec
+    %5:vgpr_32 = V_ADD_I32_e32 %2.sub0, undef %4:vgpr_32, implicit-def $vcc, implicit $exec
+    %6:vgpr_32 = V_ADDC_U32_e32 %2.sub1, undef %4, implicit-def $vcc, implicit $vcc, implicit $exec
+...
+
+# GCN-LABEL: name:  dpp64_add64_undef
+# GCN: %3:vgpr_32 = V_ADD_I32_dpp undef %1.sub0:vreg_64, undef %2.sub0:vreg_64, undef %4:vgpr_32, 1, 15, 15, 1, implicit-def $vcc, implicit $exec
+# GCN: %5:vgpr_32 = V_ADDC_U32_dpp undef %1.sub1:vreg_64, undef %2.sub1:vreg_64, undef %4:vgpr_32, 1, 15, 15, 1, implicit-def $vcc, implicit $vcc, implicit $exec
+name: dpp64_add64_undef
+tracksRegLiveness: true
+body: |
+  bb.0:
+    %2:vreg_64 = V_MOV_B64_DPP_PSEUDO undef %1:vreg_64, undef %0:vreg_64, 1, 15, 15, 1, implicit $exec
+    %5:vgpr_32 = V_ADD_I32_e32 %2.sub0, undef %4:vgpr_32, implicit-def $vcc, implicit $exec
+    %6:vgpr_32 = V_ADDC_U32_e32 %2.sub1, undef %4, implicit-def $vcc, implicit $vcc, implicit $exec
+...
+
+# GCN-LABEL: name: dpp64_add64_first_combined
+# GCN: %8:vgpr_32 = V_MOV_B32_dpp undef %1.sub1:vreg_64, undef %2.sub1:vreg_64, 1, 15, 15, 1, implicit $exec
+# GCN: %0:vreg_64 = REG_SEQUENCE undef %7:vgpr_32, %subreg.sub0, %8, %subreg.sub1
+# GCN: %3:vgpr_32 = V_ADD_I32_dpp undef %1.sub0:vreg_64, undef %2.sub0:vreg_64, undef %4:vgpr_32, 1, 15, 15, 1, implicit-def $vcc, implicit $exec
+# GCN: %5:vgpr_32, dead %6:sreg_64_xexec = V_ADDC_U32_e64 1, %0.sub1, undef $vcc, 0, implicit $exec
+name: dpp64_add64_first_combined
+tracksRegLiveness: true
+body: |
+  bb.0:
+    %2:vreg_64 = V_MOV_B64_DPP_PSEUDO undef %1:vreg_64, undef %0:vreg_64, 1, 15, 15, 1, implicit $exec
+    %4:vgpr_32 = V_ADD_I32_e32 %2.sub0, undef %3:vgpr_32, implicit-def $vcc, implicit $exec
+    %5:vgpr_32, dead %6:sreg_64_xexec = V_ADDC_U32_e64 1, %2.sub1, undef $vcc, 0, implicit $exec
+...

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mov.dpp.ll

endif:
   %tmp0 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %val_i32, i32 1, i32 1, i32 1, i1 1) #0
   %tmp1 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp0, i32 1, i32 1, i32 1, i1 1) #0
   %tmp2 = call i32 @llvm.amdgcn.mov.dpp.i32(i32 %tmp1, i32 1, i32 1, i32 1, i1 1) #0
   %tmp_float = bitcast i32 %tmp2 to float
   store float %tmp_float, float addrspace(1)* %out
   ret void
 }
 
+; VI-LABEL: {{^}}mov_dpp64_test:
+; VI: v_mov_b32_dpp v{{[0-9]+}}, v{{[0-9]+}} quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1
+; VI: v_mov_b32_dpp v{{[0-9]+}}, v{{[0-9]+}} quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1
+define amdgpu_kernel void @mov_dpp64_test(i64 addrspace(1)* %out, i64 %in1) {
+  %tmp0 = call i64 @llvm.amdgcn.mov.dpp.i64(i64 %in1, i32 1, i32 1, i32 1, i1 0) #0
+  store i64 %tmp0, i64 addrspace(1)* %out
+  ret void
+}
+
+; VI-LABEL: {{^}}mov_dpp64_imm_test:
+; VI-OPT-DAG: s_mov_b32 s[[SOLD_LO:[0-9]+]], 0x3afaedd9
+; VI-OPT-DAG: s_movk_i32 s[[SOLD_HI:[0-9]+]], 0x7047
+; VI-OPT-DAG: v_mov_b32_e32 v[[OLD_LO:[0-9]+]], s[[SOLD_LO]]
+; VI-OPT-DAG: v_mov_b32_e32 v[[OLD_HI:[0-9]+]], s[[SOLD_HI]]
+; VI-OPT-DAG: v_mov_b32_dpp v[[OLD_LO]], v[[OLD_LO]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1
+; VI-OPT-DAG: v_mov_b32_dpp v[[OLD_HI]], v[[OLD_HI]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1
+; VI-NOOPT-COUNT2: v_mov_b32_dpp v{{[0-9]+}}, v{{[0-9]+}} quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1
+define amdgpu_kernel void @mov_dpp64_imm_test(i64 addrspace(1)* %out) {
+  %tmp0 = call i64 @llvm.amdgcn.mov.dpp.i64(i64 123451234512345, i32 1, i32 1, i32 1, i1 0) #0
+  store i64 %tmp0, i64 addrspace(1)* %out
+  ret void
+}
+
 declare i32 @llvm.amdgcn.mov.dpp.i32(i32, i32, i32, i32, i1) #0
+declare i64 @llvm.amdgcn.mov.dpp.i64(i64, i32, i32, i32, i1) #0
 
 attributes #0 = { nounwind readnone convergent }
 

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.update.dpp.ll

-; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX8 %s
-; RUN: llc -march=amdgcn -mcpu=gfx1010 -mattr=-flat-for-global -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10 %s
+; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX8,GFX8-OPT,GCN-OPT %s
+; RUN: llc -march=amdgcn -mcpu=tonga -O0 -mattr=-flat-for-global -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX8,GFX8-NOOPT %s
+; RUN: llc -march=amdgcn -mcpu=gfx1010 -mattr=-flat-for-global -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10,GCN-OPT %s
 
 ; GCN-LABEL: {{^}}dpp_test:
 ; GCN:  v_mov_b32_e32 [[DST:v[0-9]+]], s{{[0-9]+}}
 ; GCN:  v_mov_b32_e32 [[SRC:v[0-9]+]], s{{[0-9]+}}
-; GFX8: s_nop 1
+; GFX8-OPT: s_nop 1
+; GFX8-NOOPT: s_nop 0
+; GFX8-NOOPT-NEXT: s_nop 0
 ; GCN:  v_mov_b32_dpp [[DST]], [[SRC]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
 define amdgpu_kernel void @dpp_test(i32 addrspace(1)* %out, i32 %in1, i32 %in2) {
   %tmp0 = call i32 @llvm.amdgcn.update.dpp.i32(i32 %in1, i32 %in2, i32 1, i32 1, i32 1, i1 0) #0
   store i32 %tmp0, i32 addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}dpp_test_bc:
 ; GCN:  v_mov_b32_e32 [[DST:v[0-9]+]], s{{[0-9]+}}
 ; GCN:  v_mov_b32_e32 [[SRC:v[0-9]+]], s{{[0-9]+}}
-; GFX8: s_nop 1
+; GFX8-OPT: s_nop 1
+; GFX8-NOOPT: s_nop 0
+; GFX8-NOOPT-NEXT: s_nop 0
 ; GCN:  v_mov_b32_dpp [[DST]], [[SRC]] quad_perm:[2,0,0,0] row_mask:0x1 bank_mask:0x1 bound_ctrl:0{{$}}
 define amdgpu_kernel void @dpp_test_bc(i32 addrspace(1)* %out, i32 %in1, i32 %in2) {
   %tmp0 = call i32 @llvm.amdgcn.update.dpp.i32(i32 %in1, i32 %in2, i32 2, i32 1, i32 1, i1 1) #0
   store i32 %tmp0, i32 addrspace(1)* %out
   ret void
 }
 
 
-; VI-LABEL: {{^}}dpp_test1:
+; GCN-LABEL: {{^}}dpp_test1:
 ; GFX10: v_add_nc_u32_e32 [[REG:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}
 ; GFX8-OPT: v_add_u32_e32 [[REG:v[0-9]+]], vcc, v{{[0-9]+}}, v{{[0-9]+}}
 ; GFX8-NOOPT: v_add_u32_e64 [[REG:v[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}, v{{[0-9]+}}
 ; GFX8-NOOPT: v_mov_b32_e32 v{{[0-9]+}}, 0
 ; GFX8: s_nop 0
 ; GFX8-NEXT: s_nop 0
-; GFX8-OPT-NEXT: v_mov_b32_dpp {{v[0-9]+}}, [[REG]] quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
+; GFX8-NEXT: v_mov_b32_dpp {{v[0-9]+}}, [[REG]] quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
 @0 = internal unnamed_addr addrspace(3) global [448 x i32] undef, align 4
 define weak_odr amdgpu_kernel void @dpp_test1(i32* %arg) local_unnamed_addr {
 bb:
   %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
   %tmp1 = zext i32 %tmp to i64
   %tmp2 = getelementptr inbounds [448 x i32], [448 x i32] addrspace(3)* @0, i32 0, i32 %tmp
   %tmp3 = load i32, i32 addrspace(3)* %tmp2, align 4
   fence syncscope("workgroup-one-as") release
   tail call void @llvm.amdgcn.s.barrier()
   fence syncscope("workgroup-one-as") acquire
   %tmp4 = add nsw i32 %tmp3, %tmp3
   %tmp5 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp4, i32 177, i32 15, i32 15, i1 zeroext false)
   %tmp6 = add nsw i32 %tmp5, %tmp4
   %tmp7 = getelementptr inbounds i32, i32* %arg, i64 %tmp1
   store i32 %tmp6, i32* %tmp7, align 4
   ret void
 }
 
+; GCN-LABEL: {{^}}update_dpp64_test:
+; GCN:     load_dwordx2 v{{\[}}[[SRC_LO:[0-9]+]]:[[SRC_HI:[0-9]+]]]
+; GCN-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[SRC_LO]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+; GCN-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[SRC_HI]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+define amdgpu_kernel void @update_dpp64_test(i64 addrspace(1)* %arg, i64 %in1, i64 %in2) {
+  %id = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %gep = getelementptr inbounds i64, i64 addrspace(1)* %arg, i32 %id
+  %load = load i64, i64 addrspace(1)* %gep
+  %tmp0 = call i64 @llvm.amdgcn.update.dpp.i64(i64 %in1, i64 %load, i32 1, i32 1, i32 1, i1 0) #0
+  store i64 %tmp0, i64 addrspace(1)* %gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}update_dpp64_imm_old_test:
+; GCN-OPT-DAG: v_mov_b32_e32 v[[OLD_LO:[0-9]+]], 0x3afaedd9
+; GCN-OPT-DAG: v_mov_b32_e32 v[[OLD_HI:[0-9]+]], 0x7047
+; GFX8-NOOPT-DAG: s_mov_b32 s[[SOLD_LO:[0-9]+]], 0x3afaedd9
+; GFX8-NOOPT-DAG: s_movk_i32 s[[SOLD_HI:[0-9]+]], 0x7047
+; GCN-DAG: load_dwordx2 v{{\[}}[[SRC_LO:[0-9]+]]:[[SRC_HI:[0-9]+]]]
+; GCN-OPT-DAG: v_mov_b32_dpp v[[OLD_LO]], v[[SRC_LO]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+; GCN-OPT-DAG: v_mov_b32_dpp v[[OLD_HI]], v[[SRC_HI]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+; GCN-NOOPT-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[SRC_LO]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+; GCN-NOOPT-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[SRC_HI]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+define amdgpu_kernel void @update_dpp64_imm_old_test(i64 addrspace(1)* %arg, i64 %in2) {
+  %id = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %gep = getelementptr inbounds i64, i64 addrspace(1)* %arg, i32 %id
+  %load = load i64, i64 addrspace(1)* %gep
+  %tmp0 = call i64 @llvm.amdgcn.update.dpp.i64(i64 123451234512345, i64 %load, i32 1, i32 1, i32 1, i1 0) #0
+  store i64 %tmp0, i64 addrspace(1)* %gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}update_dpp64_imm_src_test:
+; GCN-OPT-DAG: v_mov_b32_e32 v[[OLD_LO:[0-9]+]], 0x3afaedd9
+; GCN-OPT-DAG: v_mov_b32_e32 v[[OLD_HI:[0-9]+]], 0x7047
+; GFX8-NOOPT-DAG: s_mov_b32 s[[SOLD_LO:[0-9]+]], 0x3afaedd9
+; GFX8-NOOPT-DAG: s_movk_i32 s[[SOLD_HI:[0-9]+]], 0x7047
+; GCN-OPT-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[OLD_LO]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+; GCN-OPT-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[OLD_HI]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+; GCN-NOOPT-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[SRC_LO]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+; GCN-NOOPT-DAG: v_mov_b32_dpp v{{[0-9]+}}, v[[SRC_HI]] quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1{{$}}
+define amdgpu_kernel void @update_dpp64_imm_src_test(i64 addrspace(1)* %out, i64 %in1) {
+  %tmp0 = call i64 @llvm.amdgcn.update.dpp.i64(i64 %in1, i64 123451234512345, i32 1, i32 1, i32 1, i1 0) #0
+  store i64 %tmp0, i64 addrspace(1)* %out
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x()
 declare void @llvm.amdgcn.s.barrier()
 declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32, i32, i32, i1) #0
+declare i64 @llvm.amdgcn.update.dpp.i64(i64, i64, i32, i32, i32, i1) #0
 
 attributes #0 = { nounwind readnone convergent }