AMDGPU/GlobalISel: Legalize G_[SU]DIVREM instructions

[cdevadas]

Ping

[arsenm]

LGTM except you don't need NoRegister all over the place

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

2827

Don't need the != NoRegister

2833–2837

Ditto

2836

Ditto

2983

Ditto

3068

This is default initialized

[cdevadas]

Rebase + comments addressed.

[foad]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

586–589

Matt usually objects to this indentation change on the grounds that clang-format is wrong :)

2985

Pull this buildICmp out of the "if"s, since it is identical in both cases, and you don't want to build it twice in the divrem case?

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.h

113

It looks odd that you did not change the name of this function, but I see it is not implemented, so maybe just remove the declaration here?

[cdevadas]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

586–589

Oh. It was auto-formatted by clang-format.

2985

Thanks Jay. It is addressed with D103083.

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.h

113

Yes, it was a dead declaration. Cleaned it up with e3b8e6d48251

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
   case TargetOpcode::G_SMAX:
     Observer.changingInstr(MI);
     widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
     widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
     widenScalarDst(MI, WideTy);
     Observer.changedInstr(MI);
     return Legalized;
 
+  case TargetOpcode::G_SDIVREM:
+    Observer.changingInstr(MI);
+    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
+    widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT);
+    widenScalarDst(MI, WideTy);
+    widenScalarDst(MI, WideTy, 1);
+    Observer.changedInstr(MI);
+    return Legalized;
+
   case TargetOpcode::G_ASHR:
   case TargetOpcode::G_LSHR:
     Observer.changingInstr(MI);
 
     if (TypeIdx == 0) {
       unsigned CvtOp = MI.getOpcode() == TargetOpcode::G_ASHR ?
         TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT;
 
LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
   case TargetOpcode::G_UMAX:
     Observer.changingInstr(MI);
     widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
     widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
     widenScalarDst(MI, WideTy);
     Observer.changedInstr(MI);
     return Legalized;
 
+  case TargetOpcode::G_UDIVREM:
+    Observer.changingInstr(MI);
+    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
+    widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT);
+    widenScalarDst(MI, WideTy);
+    widenScalarDst(MI, WideTy, 1);
+    Observer.changedInstr(MI);
+    return Legalized;
+
   case TargetOpcode::G_SELECT:
     Observer.changingInstr(MI);
     if (TypeIdx == 0) {
       // Perform operation at larger width (any extension is fine here, high
       // bits don't affect the result) and then truncate the result back to the
       // original type.
       widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
       widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT);
LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
   case G_FSIN:
   case G_FSQRT:
   case G_BSWAP:
   case G_BITREVERSE:
   case G_SDIV:
   case G_UDIV:
   case G_SREM:
   case G_UREM:
+  case G_SDIVREM:
+  case G_UDIVREM:
   case G_SMIN:
   case G_SMAX:
   case G_UMIN:
   case G_UMAX:
   case G_FMINNUM:
   case G_FMAXNUM:
   case G_FMINNUM_IEEE:
   case G_FMAXNUM_IEEE:

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.h

bool loadInputValue(Register DstReg, MachineIRBuilder &B,
                       const ArgDescriptor *Arg,
                       const TargetRegisterClass *ArgRC, LLT ArgTy) const;
   bool loadInputValue(Register DstReg, MachineIRBuilder &B,
                       AMDGPUFunctionArgInfo::PreloadedValue ArgType) const;
   bool legalizePreloadedArgIntrin(
     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B,
     AMDGPUFunctionArgInfo::PreloadedValue ArgType) const;
 
-  bool legalizeUDIV_UREM(MachineInstr &MI, MachineRegisterInfo &MRI,
+  bool legalizeUnsignedDIV_REM(MachineInstr &MI, MachineRegisterInfo &MRI,
                                MachineIRBuilder &B) const;
 
-  void legalizeUDIV_UREM32Impl(MachineIRBuilder &B,
-                               Register DstReg, Register Num, Register Den,
-                               bool IsRem) const;
+  void legalizeUnsignedDIV_REM32Impl(MachineIRBuilder &B, Register DstDivReg,
+                                     Register DstRemReg, Register Num,
+                                     Register Den) const;
 
-  void legalizeUDIV_UREM64Impl(MachineIRBuilder &B,
-                               Register DstReg, Register Numer, Register Denom,
-                               bool IsDiv) const;
+  void legalizeUnsignedDIV_REM64Impl(MachineIRBuilder &B, Register DstDivReg,
+                                     Register DstRemReg, Register Numer,
+                                     Register Denom) const;
 
   bool legalizeUDIV_UREM64(MachineInstr &MI, MachineRegisterInfo &MRI,
                            MachineIRBuilder &B) const;
-  bool legalizeSDIV_SREM(MachineInstr &MI, MachineRegisterInfo &MRI,
+  bool legalizeSignedDIV_REM(MachineInstr &MI, MachineRegisterInfo &MRI,
                              MachineIRBuilder &B) const;
 
   bool legalizeFDIV(MachineInstr &MI, MachineRegisterInfo &MRI,
                     MachineIRBuilder &B) const;
   bool legalizeFDIV16(MachineInstr &MI, MachineRegisterInfo &MRI,
                       MachineIRBuilder &B) const;
   bool legalizeFDIV32(MachineInstr &MI, MachineRegisterInfo &MRI,
                       MachineIRBuilder &B) const;
   bool legalizeFDIV64(MachineInstr &MI, MachineRegisterInfo &MRI,

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

if (ST.hasVOP3PInsts() && ST.hasAddNoCarry() && ST.hasIntClamp()) {
     // FIXME: DAG expansion gets better results. The widening uses the smaller
     // range values and goes for the min/max lowering directly.
     getActionDefinitionsBuilder({G_SADDSAT, G_SSUBSAT})
       .minScalar(0, S32)
       .scalarize(0)
       .lower();
   }
 
-  getActionDefinitionsBuilder({G_SDIV, G_UDIV, G_SREM, G_UREM})
+  getActionDefinitionsBuilder(
+      {G_SDIV, G_UDIV, G_SREM, G_UREM, G_SDIVREM, G_UDIVREM})
       .customFor({S32, S64})
       .clampScalar(0, S32, S64)
       .widenScalarToNextPow2(0, 32)
       .scalarize(0);
 
   auto &Mulh = getActionDefinitionsBuilder({G_UMULH, G_SMULH})
                    .legalFor({S32})
                    .maxScalarOrElt(0, S32);
 
   if (ST.hasVOP3PInsts()) {
     Mulh
       .clampMaxNumElements(0, S8, 2)
bool AMDGPULegalizerInfo::legalizeCustom(LegalizerHelper &Helper,
   case TargetOpcode::G_LOAD:
     return legalizeLoad(Helper, MI);
   case TargetOpcode::G_FMAD:
     return legalizeFMad(MI, MRI, B);
   case TargetOpcode::G_FDIV:
     return legalizeFDIV(MI, MRI, B);
   case TargetOpcode::G_UDIV:
   case TargetOpcode::G_UREM:
-    return legalizeUDIV_UREM(MI, MRI, B);
+  case TargetOpcode::G_UDIVREM:
+    return legalizeUnsignedDIV_REM(MI, MRI, B);
   case TargetOpcode::G_SDIV:
   case TargetOpcode::G_SREM:
-    return legalizeSDIV_SREM(MI, MRI, B);
+  case TargetOpcode::G_SDIVREM:
+    return legalizeSignedDIV_REM(MI, MRI, B);
   case TargetOpcode::G_ATOMIC_CMPXCHG:
     return legalizeAtomicCmpXChg(MI, MRI, B);
   case TargetOpcode::G_FLOG:
     return legalizeFlog(MI, B, numbers::ln2f);
   case TargetOpcode::G_FLOG10:
     return legalizeFlog(MI, B, numbers::ln2f / numbers::ln10f);
   case TargetOpcode::G_FEXP:
     return legalizeFExp(MI, B);
bool AMDGPULegalizerInfo::legalizeFDIV(MachineInstr &MI,
   if (DstTy == S32)
     return legalizeFDIV32(MI, MRI, B);
   if (DstTy == S64)
     return legalizeFDIV64(MI, MRI, B);
 
   return false;
 }
 
-void AMDGPULegalizerInfo::legalizeUDIV_UREM32Impl(MachineIRBuilder &B,
-                                                  Register DstReg,
+void AMDGPULegalizerInfo::legalizeUnsignedDIV_REM32Impl(MachineIRBuilder &B,
+                                                        Register DstDivReg,
+                                                        Register DstRemReg,
                                                         Register X,
-                                                  Register Y,
-                                                  bool IsDiv) const {
+                                                        Register Y) const {
   const LLT S1 = LLT::scalar(1);
   const LLT S32 = LLT::scalar(32);
 
   // See AMDGPUCodeGenPrepare::expandDivRem32 for a description of the
   // algorithm used here.
 
   // Initial estimate of inv(y).
   auto FloatY = B.buildUITOFP(S32, Y);
void AMDGPULegalizerInfo::legalizeUnsignedDIV_REM32Impl(MachineIRBuilder &B,
 
   // Quotient/remainder estimate.
   auto Q = B.buildUMulH(S32, X, Z);
   auto R = B.buildSub(S32, X, B.buildMul(S32, Q, Y));
 
   // First quotient/remainder refinement.
   auto One = B.buildConstant(S32, 1);
   auto Cond = B.buildICmp(CmpInst::ICMP_UGE, S1, R, Y);
-  if (IsDiv)
+  if (DstDivReg)
     Q = B.buildSelect(S32, Cond, B.buildAdd(S32, Q, One), Q);
   R = B.buildSelect(S32, Cond, B.buildSub(S32, R, Y), R);
 
   // Second quotient/remainder refinement.
   Cond = B.buildICmp(CmpInst::ICMP_UGE, S1, R, Y);
-  if (IsDiv)
-    B.buildSelect(DstReg, Cond, B.buildAdd(S32, Q, One), Q);
-  else
-    B.buildSelect(DstReg, Cond, B.buildSub(S32, R, Y), R);
+  if (DstDivReg)
+    B.buildSelect(DstDivReg, Cond, B.buildAdd(S32, Q, One), Q);
+
+  if (DstRemReg)
+    B.buildSelect(DstRemReg, Cond, B.buildSub(S32, R, Y), R);
 }
 
 // Build integer reciprocal sequence arounud V_RCP_IFLAG_F32
 //
 // Return lo, hi of result
 //
 // %cvt.lo = G_UITOFP Val.lo
 // %cvt.hi = G_UITOFP Val.hi
auto Mad2 = B.buildFMAD(S32, Trunc,
                           B.buildFConstant(S32, BitsToFloat(0xcf800000)), Mul1);
 
   auto ResultLo = B.buildFPTOUI(S32, Mad2);
   auto ResultHi = B.buildFPTOUI(S32, Trunc);
 
   return {ResultLo.getReg(0), ResultHi.getReg(0)};
 }
 
-void AMDGPULegalizerInfo::legalizeUDIV_UREM64Impl(MachineIRBuilder &B,
-                                                  Register DstReg,
+void AMDGPULegalizerInfo::legalizeUnsignedDIV_REM64Impl(MachineIRBuilder &B,
+                                                        Register DstDivReg,
+                                                        Register DstRemReg,
                                                         Register Numer,
-                                                  Register Denom,
-                                                  bool IsDiv) const {
+                                                        Register Denom) const {
   const LLT S32 = LLT::scalar(32);
   const LLT S64 = LLT::scalar(64);
   const LLT S1 = LLT::scalar(1);
   Register RcpLo, RcpHi;
 
   std::tie(RcpLo, RcpHi) = emitReciprocalU64(B, Denom);
 
   auto Rcp = B.buildMerge(S64, {RcpLo, RcpHi});
void AMDGPULegalizerInfo::legalizeUnsignedDIV_REM64Impl(MachineIRBuilder &B,
 
   auto Sub3_Mi = B.buildUSube(S32, S1, Sub2_Mi, DenomHi, Sub2_Lo.getReg(1));
   auto Sub3_Hi = B.buildUSube(S32, S1, Sub3_Mi, Zero32, Sub3_Lo.getReg(1));
   auto Sub3 = B.buildMerge(S64, {Sub3_Lo, Sub3_Hi});
 
   // endif C6
   // endif C3
 
-  if (IsDiv) {
+  if (DstDivReg) {
     auto Sel1 = B.buildSelect(
         S64, B.buildICmp(CmpInst::ICMP_NE, S1, C6, Zero32), Add4, Add3);
-    B.buildSelect(DstReg,
-                  B.buildICmp(CmpInst::ICMP_NE, S1, C3, Zero32), Sel1, MulHi3);
-  } else {
+    B.buildSelect(DstDivReg, B.buildICmp(CmpInst::ICMP_NE, S1, C3, Zero32),
+                  Sel1, MulHi3);
+  }
+
+  if (DstRemReg) {
     auto Sel2 = B.buildSelect(
         S64, B.buildICmp(CmpInst::ICMP_NE, S1, C6, Zero32), Sub3, Sub2);
-    B.buildSelect(DstReg,
-                  B.buildICmp(CmpInst::ICMP_NE, S1, C3, Zero32), Sel2, Sub1);
+    B.buildSelect(DstRemReg, B.buildICmp(CmpInst::ICMP_NE, S1, C3, Zero32),
+                  Sel2, Sub1);
   }
 }
 
-bool AMDGPULegalizerInfo::legalizeUDIV_UREM(MachineInstr &MI,
+bool AMDGPULegalizerInfo::legalizeUnsignedDIV_REM(MachineInstr &MI,
                                                   MachineRegisterInfo &MRI,
                                                   MachineIRBuilder &B) const {
+  Register DstDivReg, DstRemReg;
+  switch (MI.getOpcode()) {
+  default:
+    llvm_unreachable("Unexpected opcode!");
+  case AMDGPU::G_UDIV: {
+    DstDivReg = MI.getOperand(0).getReg();
+    break;
+  }
+  case AMDGPU::G_UREM: {
+    DstRemReg = MI.getOperand(0).getReg();
+    break;
+  }
+  case AMDGPU::G_UDIVREM: {
+    DstDivReg = MI.getOperand(0).getReg();
+    DstRemReg = MI.getOperand(1).getReg();
+    break;
+  }
+  }
+
   const LLT S64 = LLT::scalar(64);
   const LLT S32 = LLT::scalar(32);
-  const bool IsDiv = MI.getOpcode() == AMDGPU::G_UDIV;
-  Register DstReg = MI.getOperand(0).getReg();
-  Register Num = MI.getOperand(1).getReg();
-  Register Den = MI.getOperand(2).getReg();
-  LLT Ty = MRI.getType(DstReg);
+  const unsigned FirstSrcOpIdx = MI.getNumExplicitDefs();
+  Register Num = MI.getOperand(FirstSrcOpIdx).getReg();
+  Register Den = MI.getOperand(FirstSrcOpIdx + 1).getReg();
+  LLT Ty = MRI.getType(MI.getOperand(0).getReg());
 
   if (Ty == S32)
-    legalizeUDIV_UREM32Impl(B, DstReg, Num, Den, IsDiv);
+    legalizeUnsignedDIV_REM32Impl(B, DstDivReg, DstRemReg, Num, Den);
   else if (Ty == S64)
-    legalizeUDIV_UREM64Impl(B, DstReg, Num, Den, IsDiv);
+    legalizeUnsignedDIV_REM64Impl(B, DstDivReg, DstRemReg, Num, Den);
   else
     return false;
 
   MI.eraseFromParent();
   return true;
-
 }
 
-bool AMDGPULegalizerInfo::legalizeSDIV_SREM(MachineInstr &MI,
+bool AMDGPULegalizerInfo::legalizeSignedDIV_REM(MachineInstr &MI,
                                                 MachineRegisterInfo &MRI,
                                                 MachineIRBuilder &B) const {
   const LLT S64 = LLT::scalar(64);
   const LLT S32 = LLT::scalar(32);
 
-  Register DstReg = MI.getOperand(0).getReg();
-  const LLT Ty = MRI.getType(DstReg);
+  LLT Ty = MRI.getType(MI.getOperand(0).getReg());
   if (Ty != S32 && Ty != S64)
     return false;
 
-  const bool IsDiv = MI.getOpcode() == AMDGPU::G_SDIV;
-
-  Register LHS = MI.getOperand(1).getReg();
-  Register RHS = MI.getOperand(2).getReg();
+  const unsigned FirstSrcOpIdx = MI.getNumExplicitDefs();
+  Register LHS = MI.getOperand(FirstSrcOpIdx).getReg();
+  Register RHS = MI.getOperand(FirstSrcOpIdx + 1).getReg();
 
   auto SignBitOffset = B.buildConstant(S32, Ty.getSizeInBits() - 1);
   auto LHSign = B.buildAShr(Ty, LHS, SignBitOffset);
   auto RHSign = B.buildAShr(Ty, RHS, SignBitOffset);
 
   LHS = B.buildAdd(Ty, LHS, LHSign).getReg(0);
   RHS = B.buildAdd(Ty, RHS, RHSign).getReg(0);
 
   LHS = B.buildXor(Ty, LHS, LHSign).getReg(0);
   RHS = B.buildXor(Ty, RHS, RHSign).getReg(0);
 
-  Register UDivRem = MRI.createGenericVirtualRegister(Ty);
+  Register DstDivReg, DstRemReg, TmpDivReg, TmpRemReg;
+  switch (MI.getOpcode()) {
+  default:
+    llvm_unreachable("Unexpected opcode!");
+  case AMDGPU::G_SDIV: {
+    DstDivReg = MI.getOperand(0).getReg();
+    TmpDivReg = MRI.createGenericVirtualRegister(Ty);
+    break;
+  }
+  case AMDGPU::G_SREM: {
+    DstRemReg = MI.getOperand(0).getReg();
+    TmpRemReg = MRI.createGenericVirtualRegister(Ty);
+    break;
+  }
+  case AMDGPU::G_SDIVREM: {
+    DstDivReg = MI.getOperand(0).getReg();
+    DstRemReg = MI.getOperand(1).getReg();
+    TmpDivReg = MRI.createGenericVirtualRegister(Ty);
+    TmpRemReg = MRI.createGenericVirtualRegister(Ty);
+    break;
+  }
+  }
+
   if (Ty == S32)
-    legalizeUDIV_UREM32Impl(B, UDivRem, LHS, RHS, IsDiv);
+    legalizeUnsignedDIV_REM32Impl(B, TmpDivReg, TmpRemReg, LHS, RHS);
   else
-    legalizeUDIV_UREM64Impl(B, UDivRem, LHS, RHS, IsDiv);
+    legalizeUnsignedDIV_REM64Impl(B, TmpDivReg, TmpRemReg, LHS, RHS);
 
-  Register Sign;
-  if (IsDiv)
-    Sign = B.buildXor(Ty, LHSign, RHSign).getReg(0);
-  else
-    Sign = LHSign.getReg(0); // Remainder sign is the same as LHS
+  if (DstDivReg) {
+    auto Sign = B.buildXor(Ty, LHSign, RHSign).getReg(0);
+    auto SignXor = B.buildXor(Ty, TmpDivReg, Sign).getReg(0);
+    B.buildSub(DstDivReg, SignXor, Sign);
+  }
 
-  UDivRem = B.buildXor(Ty, UDivRem, Sign).getReg(0);
-  B.buildSub(DstReg, UDivRem, Sign);
+  if (DstRemReg) {
+    auto Sign = LHSign.getReg(0); // Remainder sign is the same as LHS
+    auto SignXor = B.buildXor(Ty, TmpRemReg, Sign).getReg(0);
+    B.buildSub(DstRemReg, SignXor, Sign);
+  }
 
   MI.eraseFromParent();
   return true;
 }
 
 bool AMDGPULegalizerInfo::legalizeFastUnsafeFDIV(MachineInstr &MI,
                                                  MachineRegisterInfo &MRI,
                                                  MachineIRBuilder &B) const {

llvm/test/CodeGen/AMDGPU/GlobalISel/sdivrem.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck --check-prefix=GFX8 %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10 %s
+
+define amdgpu_kernel void @sdivrem_i32(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 %x, i32 %y) {
+; GFX8-LABEL: sdivrem_i32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX8-NEXT:    s_add_i32 s1, s1, s6
+; GFX8-NEXT:    s_xor_b32 s7, s1, s6
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX8-NEXT:    s_sub_i32 s1, 0, s7
+; GFX8-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX8-NEXT:    s_add_i32 s0, s0, s8
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_xor_b32 s9, s0, s8
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    s_xor_b32 s4, s8, s6
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s9, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s7
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s9, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_xor_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s4, v2
+; GFX8-NEXT:    v_xor_b32_e32 v3, s8, v3
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s8, v3
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_dword v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_i32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX9-NEXT:    s_add_i32 s1, s1, s6
+; GFX9-NEXT:    s_xor_b32 s7, s1, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_sub_i32 s1, 0, s7
+; GFX9-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX9-NEXT:    s_add_i32 s0, s0, s8
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_xor_b32 s9, s0, s8
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_xor_b32 s4, s8, s6
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s9, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s4, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s8, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_i32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX10-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX10-NEXT:    s_add_i32 s1, s1, s6
+; GFX10-NEXT:    s_add_i32 s0, s0, s8
+; GFX10-NEXT:    s_xor_b32 s7, s1, s6
+; GFX10-NEXT:    s_xor_b32 s0, s0, s8
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX10-NEXT:    s_sub_i32 s1, 0, s7
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    s_xor_b32 s4, s8, s6
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s4, v0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s8, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv i32 %x, %y
+  store i32 %div, i32 addrspace(1)* %out0
+  %rem = srem i32 %x, %y
+  store i32 %rem, i32 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_i64(i64 addrspace(1)* %out0, i64 addrspace(1)* %out1, i64 %x, i64 %y) {
+; GFX8-LABEL: sdivrem_i64:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx8 s[4:11], s[4:5], 0x0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_ashr_i32 s12, s11, 31
+; GFX8-NEXT:    s_ashr_i32 s2, s9, 31
+; GFX8-NEXT:    s_add_u32 s0, s8, s2
+; GFX8-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX8-NEXT:    s_and_b32 s1, s1, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX8-NEXT:    s_addc_u32 s1, s9, s2
+; GFX8-NEXT:    s_add_u32 s8, s10, s12
+; GFX8-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX8-NEXT:    s_and_b32 s3, s3, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX8-NEXT:    s_mov_b32 s13, s12
+; GFX8-NEXT:    s_addc_u32 s9, s11, s12
+; GFX8-NEXT:    s_xor_b64 s[8:9], s[8:9], s[12:13]
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX8-NEXT:    s_mov_b32 s3, s2
+; GFX8-NEXT:    s_xor_b64 s[10:11], s[0:1], s[2:3]
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX8-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_sub_u32 s14, 0, s8
+; GFX8-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX8-NEXT:    s_and_b32 s0, s0, 1
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX8-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX8-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX8-NEXT:    s_subb_u32 s15, 0, s9
+; GFX8-NEXT:    v_mul_lo_u32 v3, s15, v0
+; GFX8-NEXT:    v_mul_lo_u32 v2, s14, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s14, v0
+; GFX8-NEXT:    v_mul_lo_u32 v4, s14, v0
+; GFX8-NEXT:    v_mov_b32_e32 v6, s9
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, v4
+; GFX8-NEXT:    v_mul_lo_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v7, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_mul_hi_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v7, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v7, v5
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v4, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v5, v4
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v4
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_addc_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v4, s15, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s14, v3
+; GFX8-NEXT:    v_mul_hi_u32 v8, s14, v0
+; GFX8-NEXT:    v_mul_lo_u32 v7, s14, v0
+; GFX8-NEXT:    v_add_u32_e64 v1, s[0:1], v1, v2
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v5
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v8
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX8-NEXT:    v_mul_lo_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX8-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v8, v2
+; GFX8-NEXT:    v_mul_hi_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v7, s[0:1], v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v7, v5
+; GFX8-NEXT:    v_add_u32_e64 v3, s[0:1], v3, v4
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v2, s11, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s10, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s10, v0
+; GFX8-NEXT:    v_mul_hi_u32 v0, s11, v0
+; GFX8-NEXT:    v_mov_b32_e32 v4, s11
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v5, s11, v1
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_mul_hi_u32 v3, s10, v1
+; GFX8-NEXT:    v_mul_hi_u32 v1, s11, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v5, v0
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v2
+; GFX8-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX8-NEXT:    v_mul_hi_u32 v7, s8, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s8, v0
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s10, v5
+; GFX8-NEXT:    v_subb_u32_e64 v4, s[0:1], v4, v2, vcc
+; GFX8-NEXT:    v_sub_u32_e64 v2, s[0:1], s11, v2
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v4
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, v5, v7, s[0:1]
+; GFX8-NEXT:    v_subrev_u32_e32 v7, vcc, s8, v3
+; GFX8-NEXT:    v_subbrev_u32_e64 v8, s[0:1], 0, v2, vcc
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], 1, v0
+; GFX8-NEXT:    v_addc_u32_e64 v10, s[0:1], 0, v1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v8
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v7
+; GFX8-NEXT:    v_subrev_u32_e32 v6, vcc, s8, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, v11, v12, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v12, s[0:1], 1, v9
+; GFX8-NEXT:    v_subbrev_u32_e32 v2, vcc, 0, v2, vcc
+; GFX8-NEXT:    v_addc_u32_e64 v13, s[0:1], 0, v10, s[0:1]
+; GFX8-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v11
+; GFX8-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v8, v2, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, v3, v5, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, v4, v2, s[0:1]
+; GFX8-NEXT:    s_xor_b64 s[0:1], s[2:3], s[12:13]
+; GFX8-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX8-NEXT:    v_xor_b32_e32 v1, s1, v1
+; GFX8-NEXT:    v_mov_b32_e32 v4, s1
+; GFX8-NEXT:    v_subrev_u32_e32 v0, vcc, s0, v0
+; GFX8-NEXT:    v_subb_u32_e32 v1, vcc, v1, v4, vcc
+; GFX8-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX8-NEXT:    v_xor_b32_e32 v4, s2, v2
+; GFX8-NEXT:    v_mov_b32_e32 v5, s2
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s2, v3
+; GFX8-NEXT:    v_subb_u32_e32 v3, vcc, v4, v5, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v4, s4
+; GFX8-NEXT:    v_mov_b32_e32 v5, s5
+; GFX8-NEXT:    flat_store_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dwordx2 v[0:1], v[2:3]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_i64:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[4:11], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_ashr_i32 s12, s11, 31
+; GFX9-NEXT:    s_ashr_i32 s2, s9, 31
+; GFX9-NEXT:    s_add_u32 s0, s8, s2
+; GFX9-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX9-NEXT:    s_and_b32 s1, s1, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX9-NEXT:    s_addc_u32 s1, s9, s2
+; GFX9-NEXT:    s_add_u32 s8, s10, s12
+; GFX9-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX9-NEXT:    s_and_b32 s3, s3, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX9-NEXT:    s_mov_b32 s13, s12
+; GFX9-NEXT:    s_addc_u32 s9, s11, s12
+; GFX9-NEXT:    s_xor_b64 s[8:9], s[8:9], s[12:13]
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX9-NEXT:    s_mov_b32 s3, s2
+; GFX9-NEXT:    s_xor_b64 s[10:11], s[0:1], s[2:3]
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_u32 s14, 0, s8
+; GFX9-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX9-NEXT:    s_and_b32 s0, s0, 1
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX9-NEXT:    s_subb_u32 s15, 0, s9
+; GFX9-NEXT:    v_mul_lo_u32 v3, s15, v0
+; GFX9-NEXT:    v_mul_lo_u32 v2, s14, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s14, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s14, v0
+; GFX9-NEXT:    v_mov_b32_e32 v8, s11
+; GFX9-NEXT:    v_add3_u32 v2, v3, v2, v4
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, v5
+; GFX9-NEXT:    v_mul_lo_u32 v4, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v6, v0, v5
+; GFX9-NEXT:    v_mul_hi_u32 v5, v1, v5
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v6, v1, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v4, v3
+; GFX9-NEXT:    v_mul_hi_u32 v4, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v6, v5
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v5, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v4, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_u32_e32 v5, v6, v5
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_add3_u32 v2, v5, v4, v2
+; GFX9-NEXT:    v_addc_co_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v4, s15, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s14, v3
+; GFX9-NEXT:    v_mul_hi_u32 v6, s14, v0
+; GFX9-NEXT:    v_mul_lo_u32 v7, s14, v0
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v2
+; GFX9-NEXT:    v_add3_u32 v4, v4, v5, v6
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX9-NEXT:    v_mul_lo_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX9-NEXT:    v_add_u32_e32 v2, v6, v2
+; GFX9-NEXT:    v_mul_hi_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_add_u32_e32 v6, v7, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add3_u32 v3, v6, v4, v3
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v2, s11, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s10, v1
+; GFX9-NEXT:    v_mul_hi_u32 v5, s10, v0
+; GFX9-NEXT:    v_mul_hi_u32 v0, s11, v0
+; GFX9-NEXT:    v_mov_b32_e32 v4, s9
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v5
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v5, s11, v1
+; GFX9-NEXT:    v_add_u32_e32 v2, v3, v2
+; GFX9-NEXT:    v_mul_hi_u32 v3, s10, v1
+; GFX9-NEXT:    v_mul_hi_u32 v1, s11, v1
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v5, v0
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v5, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX9-NEXT:    v_mul_hi_u32 v5, s8, v0
+; GFX9-NEXT:    v_mul_lo_u32 v7, s8, v0
+; GFX9-NEXT:    v_mov_b32_e32 v6, 0
+; GFX9-NEXT:    v_add3_u32 v2, v2, v3, v5
+; GFX9-NEXT:    v_sub_co_u32_e32 v3, vcc, s10, v7
+; GFX9-NEXT:    v_subb_co_u32_e64 v5, s[0:1], v8, v2, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v5
+; GFX9-NEXT:    v_sub_u32_e32 v2, s11, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v5
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v4, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s[0:1]
+; GFX9-NEXT:    v_subrev_co_u32_e32 v8, vcc, s8, v3
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v9, s[0:1], 0, v2, vcc
+; GFX9-NEXT:    v_add_co_u32_e64 v10, s[0:1], 1, v0
+; GFX9-NEXT:    v_addc_co_u32_e64 v11, s[0:1], 0, v1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v9
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v4, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v8
+; GFX9-NEXT:    v_subrev_co_u32_e32 v4, vcc, s8, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v9
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, v12, v13, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v13, s[0:1], 1, v10
+; GFX9-NEXT:    v_subbrev_co_u32_e32 v2, vcc, 0, v2, vcc
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; GFX9-NEXT:    v_addc_co_u32_e64 v14, s[0:1], 0, v11, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v7
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v9, v2, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v8, v4, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, v3, v4, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, v5, v2, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v0, v0, v10, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v1, v1, v11, s[0:1]
+; GFX9-NEXT:    s_xor_b64 s[0:1], s[2:3], s[12:13]
+; GFX9-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s1, v1
+; GFX9-NEXT:    v_mov_b32_e32 v4, s1
+; GFX9-NEXT:    v_subrev_co_u32_e32 v0, vcc, s0, v0
+; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v4, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX9-NEXT:    v_xor_b32_e32 v4, s2, v2
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_subrev_co_u32_e32 v2, vcc, s2, v3
+; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v4, v5, vcc
+; GFX9-NEXT:    global_store_dwordx2 v6, v[0:1], s[4:5]
+; GFX9-NEXT:    global_store_dwordx2 v6, v[2:3], s[6:7]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx8 s[4:11], s[4:5], 0x0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_ashr_i32 s2, s9, 31
+; GFX10-NEXT:    s_ashr_i32 s12, s11, 31
+; GFX10-NEXT:    s_add_u32 s0, s8, s2
+; GFX10-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX10-NEXT:    s_mov_b32 s13, s12
+; GFX10-NEXT:    s_and_b32 s1, s1, 1
+; GFX10-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX10-NEXT:    s_addc_u32 s1, s9, s2
+; GFX10-NEXT:    s_add_u32 s8, s10, s12
+; GFX10-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX10-NEXT:    s_and_b32 s3, s3, 1
+; GFX10-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX10-NEXT:    s_mov_b32 s3, s2
+; GFX10-NEXT:    s_addc_u32 s9, s11, s12
+; GFX10-NEXT:    s_xor_b64 s[10:11], s[0:1], s[2:3]
+; GFX10-NEXT:    s_xor_b64 s[8:9], s[8:9], s[12:13]
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX10-NEXT:    s_sub_u32 s1, 0, s8
+; GFX10-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX10-NEXT:    s_and_b32 s0, s0, 1
+; GFX10-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX10-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX10-NEXT:    s_subb_u32 s14, 0, s9
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX10-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX10-NEXT:    v_mul_lo_u32 v2, s1, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s14, v0
+; GFX10-NEXT:    v_mul_hi_u32 v4, s1, v0
+; GFX10-NEXT:    v_mul_lo_u32 v5, s1, v0
+; GFX10-NEXT:    v_add3_u32 v2, v3, v2, v4
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, v5
+; GFX10-NEXT:    v_mul_hi_u32 v6, v1, v5
+; GFX10-NEXT:    v_mul_hi_u32 v5, v0, v5
+; GFX10-NEXT:    v_mul_lo_u32 v4, v0, v2
+; GFX10-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX10-NEXT:    v_mul_hi_u32 v8, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX10-NEXT:    v_add_co_u32 v3, s0, v3, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v7, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v3, s0, v3, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v6, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v3, v4, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, v7, v6
+; GFX10-NEXT:    v_add_co_u32 v3, s0, v5, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v3
+; GFX10-NEXT:    v_add3_u32 v2, v4, v5, v2
+; GFX10-NEXT:    v_mul_lo_u32 v4, s14, v0
+; GFX10-NEXT:    v_mul_hi_u32 v5, s1, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s0, v1, v2, vcc_lo
+; GFX10-NEXT:    v_mul_lo_u32 v7, s1, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v2
+; GFX10-NEXT:    v_mul_lo_u32 v6, s1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v8, v3, v7
+; GFX10-NEXT:    v_add3_u32 v4, v4, v6, v5
+; GFX10-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX10-NEXT:    v_mul_hi_u32 v7, v0, v7
+; GFX10-NEXT:    v_mul_lo_u32 v6, v0, v4
+; GFX10-NEXT:    v_mul_lo_u32 v9, v3, v4
+; GFX10-NEXT:    v_mul_hi_u32 v10, v0, v4
+; GFX10-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v8, s0, v9, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v7, s0, v8, v10
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, v6, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, v9, v8
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v7, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add3_u32 v2, v4, v6, v3
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, v1, v2, vcc_lo
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v5
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT:    v_mul_lo_u32 v2, s11, v0
+; GFX10-NEXT:    v_mul_hi_u32 v4, s11, v0
+; GFX10-NEXT:    v_mul_hi_u32 v0, s10, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s10, v1
+; GFX10-NEXT:    v_mul_lo_u32 v5, s11, v1
+; GFX10-NEXT:    v_mul_hi_u32 v6, s10, v1
+; GFX10-NEXT:    v_mul_hi_u32 v1, s11, v1
+; GFX10-NEXT:    v_add_co_u32 v2, s0, v2, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v4, s0, v5, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v0, s0, v2, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v2, s0, v4, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v3, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v3, v5, v4
+; GFX10-NEXT:    v_add_co_u32 v0, s0, v2, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; GFX10-NEXT:    v_mul_lo_u32 v5, s8, v0
+; GFX10-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX10-NEXT:    v_mul_hi_u32 v3, s8, v0
+; GFX10-NEXT:    v_mul_lo_u32 v4, s8, v1
+; GFX10-NEXT:    v_add3_u32 v2, v2, v4, v3
+; GFX10-NEXT:    v_add_co_u32 v3, vcc_lo, v0, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v5, vcc_lo, s10, v5
+; GFX10-NEXT:    v_sub_nc_u32_e32 v6, s11, v2
+; GFX10-NEXT:    v_sub_co_ci_u32_e64 v2, s0, s11, v2, vcc_lo
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v6, vcc_lo, s9, v6, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s8, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v8, vcc_lo, v5, s8
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v9, s0, 0, v6, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s9, v2
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v6, vcc_lo, s9, v6, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s8, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s9, v9
+; GFX10-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s0
+; GFX10-NEXT:    v_add_co_u32 v13, s0, v3, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v14, s0, 0, v4, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s9, v9
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, v12, v11, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s9, v2
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v10, v7, s0
+; GFX10-NEXT:    v_sub_co_u32 v10, s0, v8, s8
+; GFX10-NEXT:    s_xor_b64 s[8:9], s[2:3], s[12:13]
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v6, s0, 0, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v13, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, v7
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v14, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v8, v10, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v3, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v5, v7, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s0
+; GFX10-NEXT:    v_mov_b32_e32 v4, 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, s8, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s9, v1
+; GFX10-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX10-NEXT:    v_xor_b32_e32 v5, s2, v2
+; GFX10-NEXT:    v_sub_co_u32 v0, vcc_lo, v0, s8
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v1, vcc_lo, s9, v1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v2, vcc_lo, v3, s2
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v3, vcc_lo, s2, v5, vcc_lo
+; GFX10-NEXT:    global_store_dwordx2 v4, v[0:1], s[4:5]
+; GFX10-NEXT:    global_store_dwordx2 v4, v[2:3], s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv i64 %x, %y
+  store i64 %div, i64 addrspace(1)* %out0
+  %rem = srem i64 %x, %y
+  store i64 %rem, i64 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_v2i32(<2 x i32> addrspace(1)* %out0, <2 x i32> addrspace(1)* %out1, <2 x i32> %x, <2 x i32> %y) {
+; GFX8-LABEL: sdivrem_v2i32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x18
+; GFX8-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x10
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX8-NEXT:    s_add_i32 s0, s0, s8
+; GFX8-NEXT:    s_xor_b32 s9, s0, s8
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX8-NEXT:    s_ashr_i32 s11, s1, 31
+; GFX8-NEXT:    s_add_i32 s0, s1, s11
+; GFX8-NEXT:    s_sub_i32 s1, 0, s9
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_xor_b32 s12, s0, s11
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v2, s12
+; GFX8-NEXT:    s_ashr_i32 s10, s2, 31
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    s_add_i32 s0, s2, s10
+; GFX8-NEXT:    s_xor_b32 s0, s0, s10
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_ashr_i32 s2, s3, 31
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v2
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_lo_u32 v2, v0, s9
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s0, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s9, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s9, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s9, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s9, v2
+; GFX8-NEXT:    s_sub_i32 s0, 0, s12
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v3, s0, v1
+; GFX8-NEXT:    s_add_i32 s1, s3, s2
+; GFX8-NEXT:    s_xor_b32 s1, s1, s2
+; GFX8-NEXT:    s_xor_b32 s0, s10, s8
+; GFX8-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX8-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX8-NEXT:    v_xor_b32_e32 v2, s10, v2
+; GFX8-NEXT:    v_subrev_u32_e32 v0, vcc, s0, v0
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v3
+; GFX8-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s10, v2
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, s12
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s1, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s12, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s12, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s12, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s12, v3
+; GFX8-NEXT:    s_xor_b32 s0, s2, s11
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_xor_b32_e32 v1, s0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v4, s4
+; GFX8-NEXT:    v_subrev_u32_e32 v1, vcc, s0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v5, s5
+; GFX8-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX8-NEXT:    flat_store_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s2, v3
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dwordx2 v[0:1], v[2:3]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_v2i32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x18
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_ashr_i32 s10, s6, 31
+; GFX9-NEXT:    s_add_i32 s0, s6, s10
+; GFX9-NEXT:    s_xor_b32 s6, s0, s10
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x10
+; GFX9-NEXT:    s_ashr_i32 s5, s7, 31
+; GFX9-NEXT:    s_add_i32 s7, s7, s5
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_xor_b32 s7, s7, s5
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s7
+; GFX9-NEXT:    s_sub_i32 s11, 0, s6
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_ashr_i32 s4, s8, 31
+; GFX9-NEXT:    s_add_i32 s8, s8, s4
+; GFX9-NEXT:    v_mul_lo_u32 v2, s11, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    s_xor_b32 s8, s8, s4
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX9-NEXT:    s_sub_i32 s12, 0, s7
+; GFX9-NEXT:    s_ashr_i32 s11, s9, 31
+; GFX9-NEXT:    s_add_i32 s9, s9, s11
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v0, s8, v0
+; GFX9-NEXT:    v_mul_lo_u32 v2, s12, v1
+; GFX9-NEXT:    s_xor_b32 s9, s9, s11
+; GFX9-NEXT:    v_mul_lo_u32 v3, v0, s6
+; GFX9-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v3, s8, v3
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v2
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s6, v3
+; GFX9-NEXT:    v_mul_hi_u32 v1, s9, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s6, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, s7
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_xor_b32_e32 v2, s4, v2
+; GFX9-NEXT:    s_xor_b32 s6, s4, s10
+; GFX9-NEXT:    v_sub_u32_e32 v3, s9, v3
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s7, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX9-NEXT:    v_subrev_u32_e32 v2, s4, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s7, v3
+; GFX9-NEXT:    s_xor_b32 s4, s11, s5
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s6, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_xor_b32_e32 v3, s11, v3
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s6, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s4, v1
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s11, v3
+; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
+; GFX9-NEXT:    global_store_dwordx2 v4, v[2:3], s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_v2i32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x18
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_ashr_i32 s2, s0, 31
+; GFX10-NEXT:    s_ashr_i32 s3, s1, 31
+; GFX10-NEXT:    s_add_i32 s0, s0, s2
+; GFX10-NEXT:    s_add_i32 s1, s1, s3
+; GFX10-NEXT:    s_xor_b32 s8, s0, s2
+; GFX10-NEXT:    s_xor_b32 s9, s1, s3
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s9
+; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX10-NEXT:    s_sub_i32 s6, 0, s8
+; GFX10-NEXT:    s_sub_i32 s7, 0, s9
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_ashr_i32 s10, s0, 31
+; GFX10-NEXT:    s_ashr_i32 s11, s1, 31
+; GFX10-NEXT:    s_add_i32 s0, s0, s10
+; GFX10-NEXT:    v_mul_lo_u32 v2, s6, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s7, v1
+; GFX10-NEXT:    s_add_i32 s1, s1, s11
+; GFX10-NEXT:    s_xor_b32 s0, s0, s10
+; GFX10-NEXT:    s_xor_b32 s1, s1, s11
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, v0, s8
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, s9
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v1
+; GFX10-NEXT:    v_sub_nc_u32_e32 v2, s0, v2
+; GFX10-NEXT:    v_sub_nc_u32_e32 v3, s1, v3
+; GFX10-NEXT:    s_xor_b32 s1, s10, s2
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s8, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s8, v2
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s9, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s9, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s8, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v1
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s9, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s8, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s9, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s0
+; GFX10-NEXT:    v_mov_b32_e32 v4, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX10-NEXT:    s_xor_b32 s0, s11, s3
+; GFX10-NEXT:    v_xor_b32_e32 v0, s1, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s0, v1
+; GFX10-NEXT:    v_xor_b32_e32 v2, s10, v2
+; GFX10-NEXT:    v_xor_b32_e32 v3, s11, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s1, v0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v2, s10, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s11, v3
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dwordx2 v4, v[0:1], s[4:5]
+; GFX10-NEXT:    global_store_dwordx2 v4, v[2:3], s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv <2 x i32> %x, %y
+  store <2 x i32> %div, <2 x i32> addrspace(1)* %out0
+  %rem = srem <2 x i32> %x, %y
+  store <2 x i32> %rem, <2 x i32> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_v4i32(<4 x i32> addrspace(1)* %out0, <4 x i32> addrspace(1)* %out1, <4 x i32> %x, <4 x i32> %y) {
+; GFX8-LABEL: sdivrem_v4i32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x20
+; GFX8-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x0
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x10
+; GFX8-NEXT:    v_mov_b32_e32 v3, 0x4f7ffffe
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_ashr_i32 s12, s0, 31
+; GFX8-NEXT:    s_add_i32 s0, s0, s12
+; GFX8-NEXT:    s_xor_b32 s13, s0, s12
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s13
+; GFX8-NEXT:    s_ashr_i32 s15, s1, 31
+; GFX8-NEXT:    s_add_i32 s0, s1, s15
+; GFX8-NEXT:    s_sub_i32 s1, 0, s13
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_xor_b32 s16, s0, s15
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v2, s16
+; GFX8-NEXT:    s_ashr_i32 s14, s4, 31
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    s_add_i32 s0, s4, s14
+; GFX8-NEXT:    s_xor_b32 s0, s0, s14
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_ashr_i32 s4, s5, 31
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, v2, v3
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_lo_u32 v2, v0, s13
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v0
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s0, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s13, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s13, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s13, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s13, v2
+; GFX8-NEXT:    s_sub_i32 s0, 0, s16
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v4, s0, v1
+; GFX8-NEXT:    s_add_i32 s1, s5, s4
+; GFX8-NEXT:    s_xor_b32 s1, s1, s4
+; GFX8-NEXT:    s_xor_b32 s0, s14, s12
+; GFX8-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GFX8-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX8-NEXT:    v_xor_b32_e32 v2, s14, v2
+; GFX8-NEXT:    s_ashr_i32 s5, s2, 31
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v4
+; GFX8-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX8-NEXT:    v_subrev_u32_e32 v0, vcc, s0, v0
+; GFX8-NEXT:    s_add_i32 s0, s2, s5
+; GFX8-NEXT:    v_subrev_u32_e32 v4, vcc, s14, v2
+; GFX8-NEXT:    v_mul_lo_u32 v5, v1, s16
+; GFX8-NEXT:    s_xor_b32 s2, s0, s5
+; GFX8-NEXT:    s_ashr_i32 s12, s6, 31
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s1, v5
+; GFX8-NEXT:    v_add_u32_e32 v5, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s16, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v5, s2
+; GFX8-NEXT:    v_subrev_u32_e64 v6, s[0:1], s16, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 1, v1
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v5, v5
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s16, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v6, s[0:1], s16, v2
+; GFX8-NEXT:    v_mul_f32_e32 v5, v5, v3
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GFX8-NEXT:    s_sub_i32 s0, 0, s2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
+; GFX8-NEXT:    s_add_i32 s1, s6, s12
+; GFX8-NEXT:    v_mul_lo_u32 v6, s0, v5
+; GFX8-NEXT:    s_xor_b32 s1, s1, s12
+; GFX8-NEXT:    s_xor_b32 s0, s4, s15
+; GFX8-NEXT:    v_xor_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_mul_hi_u32 v6, v5, v6
+; GFX8-NEXT:    v_xor_b32_e32 v1, s0, v1
+; GFX8-NEXT:    v_subrev_u32_e32 v1, vcc, s0, v1
+; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v5, v6
+; GFX8-NEXT:    v_mul_hi_u32 v6, s1, v5
+; GFX8-NEXT:    v_subrev_u32_e32 v5, vcc, s4, v2
+; GFX8-NEXT:    s_ashr_i32 s4, s3, 31
+; GFX8-NEXT:    s_add_i32 s0, s3, s4
+; GFX8-NEXT:    v_mul_lo_u32 v7, v6, s2
+; GFX8-NEXT:    s_xor_b32 s3, s0, s4
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s1, v7
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, 1, v6
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v7, s3
+; GFX8-NEXT:    v_subrev_u32_e64 v8, s[0:1], s2, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v8, vcc
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 1, v6
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v7, v7
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v8, vcc
+; GFX8-NEXT:    v_mul_f32_e32 v3, v7, v3
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GFX8-NEXT:    v_subrev_u32_e64 v7, s[0:1], s2, v2
+; GFX8-NEXT:    s_sub_i32 s0, 0, s3
+; GFX8-NEXT:    v_cndmask_b32_e32 v7, v2, v7, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v2, s0, v3
+; GFX8-NEXT:    s_ashr_i32 s2, s7, 31
+; GFX8-NEXT:    s_add_i32 s1, s7, s2
+; GFX8-NEXT:    s_xor_b32 s1, s1, s2
+; GFX8-NEXT:    v_mul_hi_u32 v2, v3, v2
+; GFX8-NEXT:    s_xor_b32 s0, s12, s5
+; GFX8-NEXT:    v_xor_b32_e32 v6, s0, v6
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_mul_hi_u32 v3, s1, v2
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s0, v6
+; GFX8-NEXT:    v_xor_b32_e32 v6, s12, v7
+; GFX8-NEXT:    v_subrev_u32_e32 v6, vcc, s12, v6
+; GFX8-NEXT:    v_mul_lo_u32 v7, v3, s3
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 1, v3
+; GFX8-NEXT:    v_sub_u32_e32 v7, vcc, s1, v7
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v7
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v8, s[0:1], s3, v7
+; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 1, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v7
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v8, s[0:1], s3, v7
+; GFX8-NEXT:    s_xor_b32 s0, s2, s4
+; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
+; GFX8-NEXT:    v_xor_b32_e32 v3, s0, v3
+; GFX8-NEXT:    v_mov_b32_e32 v8, s8
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s0, v3
+; GFX8-NEXT:    v_mov_b32_e32 v9, s9
+; GFX8-NEXT:    v_xor_b32_e32 v7, s2, v7
+; GFX8-NEXT:    flat_store_dwordx4 v[8:9], v[0:3]
+; GFX8-NEXT:    v_subrev_u32_e32 v7, vcc, s2, v7
+; GFX8-NEXT:    v_mov_b32_e32 v0, s10
+; GFX8-NEXT:    v_mov_b32_e32 v1, s11
+; GFX8-NEXT:    flat_store_dwordx4 v[0:1], v[4:7]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_v4i32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x20
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0x4f7ffffe
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_ashr_i32 s6, s12, 31
+; GFX9-NEXT:    s_add_i32 s0, s12, s6
+; GFX9-NEXT:    s_xor_b32 s7, s0, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x10
+; GFX9-NEXT:    s_ashr_i32 s4, s13, 31
+; GFX9-NEXT:    s_add_i32 s5, s13, s4
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_i32 s12, 0, s7
+; GFX9-NEXT:    s_xor_b32 s5, s5, s4
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s5
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_i32 s13, 0, s5
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_lo_u32 v3, s12, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_ashr_i32 s12, s8, 31
+; GFX9-NEXT:    s_add_i32 s8, s8, s12
+; GFX9-NEXT:    s_xor_b32 s8, s8, s12
+; GFX9-NEXT:    v_mul_hi_u32 v3, v0, v3
+; GFX9-NEXT:    v_mul_f32_e32 v1, v1, v2
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    s_xor_b32 s6, s12, s6
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v3
+; GFX9-NEXT:    v_mul_hi_u32 v0, s8, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s13, v1
+; GFX9-NEXT:    s_ashr_i32 s13, s9, 31
+; GFX9-NEXT:    s_add_i32 s9, s9, s13
+; GFX9-NEXT:    v_mul_lo_u32 v4, v0, s7
+; GFX9-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX9-NEXT:    v_add_u32_e32 v5, 1, v0
+; GFX9-NEXT:    s_xor_b32 s4, s13, s4
+; GFX9-NEXT:    v_sub_u32_e32 v4, s8, v4
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v4
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v5, s7, v4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
+; GFX9-NEXT:    v_add_u32_e32 v5, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v4
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v5, s7, v4
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v3
+; GFX9-NEXT:    s_xor_b32 s7, s9, s13
+; GFX9-NEXT:    v_mul_hi_u32 v1, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s6, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s6, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, v1, s5
+; GFX9-NEXT:    v_xor_b32_e32 v3, s12, v3
+; GFX9-NEXT:    s_ashr_i32 s6, s14, 31
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s12, v3
+; GFX9-NEXT:    v_sub_u32_e32 v3, s7, v5
+; GFX9-NEXT:    s_add_i32 s7, s14, s6
+; GFX9-NEXT:    s_xor_b32 s7, s7, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v5, s7
+; GFX9-NEXT:    v_add_u32_e32 v6, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s5, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v5, v5
+; GFX9-NEXT:    v_subrev_u32_e32 v6, s5, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v6, vcc
+; GFX9-NEXT:    v_add_u32_e32 v6, 1, v1
+; GFX9-NEXT:    v_mul_f32_e32 v5, v5, v2
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s5, v3
+; GFX9-NEXT:    s_sub_i32 s8, 0, s7
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v6, s8, v5
+; GFX9-NEXT:    v_xor_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s4, v1
+; GFX9-NEXT:    s_ashr_i32 s4, s15, 31
+; GFX9-NEXT:    s_add_i32 s9, s15, s4
+; GFX9-NEXT:    v_mul_hi_u32 v6, v5, v6
+; GFX9-NEXT:    s_xor_b32 s9, s9, s4
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v8, s9
+; GFX9-NEXT:    v_subrev_u32_e32 v7, s5, v3
+; GFX9-NEXT:    s_ashr_i32 s5, s10, 31
+; GFX9-NEXT:    s_add_i32 s8, s10, s5
+; GFX9-NEXT:    s_xor_b32 s8, s8, s5
+; GFX9-NEXT:    v_add_u32_e32 v5, v5, v6
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v8, v8
+; GFX9-NEXT:    v_mul_hi_u32 v6, s8, v5
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v3, s13, v3
+; GFX9-NEXT:    v_mul_f32_e32 v2, v8, v2
+; GFX9-NEXT:    v_mul_lo_u32 v7, v6, s7
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX9-NEXT:    v_subrev_u32_e32 v5, s13, v3
+; GFX9-NEXT:    s_xor_b32 s6, s5, s6
+; GFX9-NEXT:    v_sub_u32_e32 v3, s8, v7
+; GFX9-NEXT:    s_sub_i32 s8, 0, s9
+; GFX9-NEXT:    v_mul_lo_u32 v8, s8, v2
+; GFX9-NEXT:    v_add_u32_e32 v7, 1, v6
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v7, s7, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
+; GFX9-NEXT:    v_mul_hi_u32 v8, v2, v8
+; GFX9-NEXT:    v_add_u32_e32 v7, 1, v6
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v7, s7, v3
+; GFX9-NEXT:    s_ashr_i32 s7, s11, 31
+; GFX9-NEXT:    s_add_i32 s8, s11, s7
+; GFX9-NEXT:    s_xor_b32 s8, s8, s7
+; GFX9-NEXT:    v_add_u32_e32 v2, v2, v8
+; GFX9-NEXT:    v_mul_hi_u32 v8, s8, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v3, s5, v3
+; GFX9-NEXT:    v_xor_b32_e32 v2, s6, v6
+; GFX9-NEXT:    v_mul_lo_u32 v7, v8, s9
+; GFX9-NEXT:    v_subrev_u32_e32 v6, s5, v3
+; GFX9-NEXT:    s_xor_b32 s4, s7, s4
+; GFX9-NEXT:    v_subrev_u32_e32 v2, s6, v2
+; GFX9-NEXT:    v_sub_u32_e32 v3, s8, v7
+; GFX9-NEXT:    v_add_u32_e32 v7, 1, v8
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s9, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v8, v7, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v8, s9, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc
+; GFX9-NEXT:    v_add_u32_e32 v8, 1, v7
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s9, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v8, s9, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v3, v8, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v3, s4, v7
+; GFX9-NEXT:    v_xor_b32_e32 v7, s7, v8
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s4, v3
+; GFX9-NEXT:    v_mov_b32_e32 v8, 0
+; GFX9-NEXT:    v_subrev_u32_e32 v7, s7, v7
+; GFX9-NEXT:    global_store_dwordx4 v8, v[0:3], s[0:1]
+; GFX9-NEXT:    global_store_dwordx4 v8, v[4:7], s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_v4i32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_clause 0x1
+; GFX10-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10
+; GFX10-NEXT:    v_mov_b32_e32 v4, 0x4f7ffffe
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_ashr_i32 s12, s8, 31
+; GFX10-NEXT:    s_ashr_i32 s14, s10, 31
+; GFX10-NEXT:    s_add_i32 s6, s8, s12
+; GFX10-NEXT:    s_add_i32 s8, s10, s14
+; GFX10-NEXT:    s_xor_b32 s10, s6, s12
+; GFX10-NEXT:    s_ashr_i32 s13, s9, 31
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s10
+; GFX10-NEXT:    s_ashr_i32 s15, s11, 31
+; GFX10-NEXT:    s_add_i32 s7, s9, s13
+; GFX10-NEXT:    s_add_i32 s9, s11, s15
+; GFX10-NEXT:    s_xor_b32 s11, s7, s13
+; GFX10-NEXT:    s_xor_b32 s8, s8, s14
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s11
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v2, s8
+; GFX10-NEXT:    s_xor_b32 s9, s9, s15
+; GFX10-NEXT:    s_sub_i32 s6, 0, s10
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v3, s9
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; GFX10-NEXT:    s_sub_i32 s7, 0, s11
+; GFX10-NEXT:    s_sub_i32 s19, 0, s8
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v3, v3
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    s_ashr_i32 s16, s0, 31
+; GFX10-NEXT:    s_ashr_i32 s17, s1, 31
+; GFX10-NEXT:    s_add_i32 s0, s0, s16
+; GFX10-NEXT:    s_ashr_i32 s18, s2, 31
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, v1, v4
+; GFX10-NEXT:    v_mul_f32_e32 v2, v2, v4
+; GFX10-NEXT:    s_xor_b32 s0, s0, s16
+; GFX10-NEXT:    v_mul_f32_e32 v3, v3, v4
+; GFX10-NEXT:    v_mul_lo_u32 v4, s6, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX10-NEXT:    s_sub_i32 s6, 0, s9
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GFX10-NEXT:    s_add_i32 s1, s1, s17
+; GFX10-NEXT:    v_mul_lo_u32 v5, s7, v1
+; GFX10-NEXT:    v_mul_lo_u32 v6, s19, v2
+; GFX10-NEXT:    v_mul_hi_u32 v4, v0, v4
+; GFX10-NEXT:    v_mul_lo_u32 v7, s6, v3
+; GFX10-NEXT:    s_add_i32 s2, s2, s18
+; GFX10-NEXT:    s_ashr_i32 s19, s3, 31
+; GFX10-NEXT:    s_xor_b32 s1, s1, s17
+; GFX10-NEXT:    s_xor_b32 s2, s2, s18
+; GFX10-NEXT:    v_mul_hi_u32 v5, v1, v5
+; GFX10-NEXT:    v_mul_hi_u32 v6, v2, v6
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v4
+; GFX10-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX10-NEXT:    s_add_i32 s3, s3, s19
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    s_xor_b32 s3, s3, s19
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, v2, v6
+; GFX10-NEXT:    s_xor_b32 s12, s16, s12
+; GFX10-NEXT:    v_add_nc_u32_e32 v3, v3, v7
+; GFX10-NEXT:    s_xor_b32 s13, s17, s13
+; GFX10-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX10-NEXT:    v_mul_hi_u32 v2, s2, v2
+; GFX10-NEXT:    v_mul_lo_u32 v4, v0, s10
+; GFX10-NEXT:    v_mul_hi_u32 v3, s3, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v8, 1, v0
+; GFX10-NEXT:    s_xor_b32 s14, s18, s14
+; GFX10-NEXT:    v_mul_lo_u32 v5, v1, s11
+; GFX10-NEXT:    v_mul_lo_u32 v6, v2, s8
+; GFX10-NEXT:    v_sub_nc_u32_e32 v4, s0, v4
+; GFX10-NEXT:    v_mul_lo_u32 v7, v3, s9
+; GFX10-NEXT:    v_add_nc_u32_e32 v9, 1, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v10, 1, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v11, 1, v3
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s10, v4
+; GFX10-NEXT:    v_sub_nc_u32_e32 v5, s1, v5
+; GFX10-NEXT:    v_sub_nc_u32_e32 v6, s2, v6
+; GFX10-NEXT:    v_sub_nc_u32_e32 v7, s3, v7
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v8, s10, v4
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s11, v5
+; GFX10-NEXT:    v_cmp_le_u32_e64 s1, s8, v6
+; GFX10-NEXT:    v_cmp_le_u32_e64 s2, s9, v7
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v9, s0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v9, s11, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v10, s1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v10, s8, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v11, s2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v11, s9, v7
+; GFX10-NEXT:    v_add_nc_u32_e32 v8, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s10, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, v5, v9, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v6, v10, s1
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v7, v11, s2
+; GFX10-NEXT:    v_add_nc_u32_e32 v9, 1, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v8, s10, v4
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s11, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v10, 1, v2
+; GFX10-NEXT:    v_cmp_le_u32_e64 s1, s8, v6
+; GFX10-NEXT:    v_add_nc_u32_e32 v11, 1, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s9, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v9, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v10, s1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v9, s11, v5
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v10, s8, v6
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v12, s9, v7
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
+; GFX10-NEXT:    v_xor_b32_e32 v0, s12, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, v5, v9, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v6, v10, s1
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v12, vcc_lo
+; GFX10-NEXT:    s_xor_b32 s0, s19, s15
+; GFX10-NEXT:    v_xor_b32_e32 v1, s13, v1
+; GFX10-NEXT:    v_xor_b32_e32 v2, s14, v2
+; GFX10-NEXT:    v_xor_b32_e32 v3, s0, v3
+; GFX10-NEXT:    v_xor_b32_e32 v4, s16, v4
+; GFX10-NEXT:    v_xor_b32_e32 v5, s17, v5
+; GFX10-NEXT:    v_xor_b32_e32 v6, s18, v6
+; GFX10-NEXT:    v_xor_b32_e32 v7, s19, v7
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s12, v0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s13, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v2, s14, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v8, 0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v4, s16, v4
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v5, s17, v5
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s18, v6
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s19, v7
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dwordx4 v8, v[0:3], s[4:5]
+; GFX10-NEXT:    global_store_dwordx4 v8, v[4:7], s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv <4 x i32> %x, %y
+  store <4 x i32> %div, <4 x i32> addrspace(1)* %out0
+  %rem = srem <4 x i32> %x, %y
+  store <4 x i32> %rem, <4 x i32> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_v2i64(<2 x i64> addrspace(1)* %out0, <2 x i64> addrspace(1)* %out1, <2 x i64> %x, <2 x i64> %y) {
+; GFX8-LABEL: sdivrem_v2i64:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x10
+; GFX8-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_ashr_i32 s2, s13, 31
+; GFX8-NEXT:    s_ashr_i32 s6, s9, 31
+; GFX8-NEXT:    s_add_u32 s0, s12, s2
+; GFX8-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX8-NEXT:    s_and_b32 s1, s1, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX8-NEXT:    s_addc_u32 s1, s13, s2
+; GFX8-NEXT:    s_add_u32 s8, s8, s6
+; GFX8-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX8-NEXT:    s_and_b32 s3, s3, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX8-NEXT:    s_mov_b32 s7, s6
+; GFX8-NEXT:    s_addc_u32 s9, s9, s6
+; GFX8-NEXT:    s_xor_b64 s[8:9], s[8:9], s[6:7]
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX8-NEXT:    s_mov_b32 s3, s2
+; GFX8-NEXT:    s_xor_b64 s[12:13], s[0:1], s[2:3]
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX8-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_sub_u32 s16, 0, s8
+; GFX8-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX8-NEXT:    s_and_b32 s0, s0, 1
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX8-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX8-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX8-NEXT:    s_subb_u32 s17, 0, s9
+; GFX8-NEXT:    v_mul_lo_u32 v3, s17, v0
+; GFX8-NEXT:    v_mul_lo_u32 v2, s16, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s16, v0
+; GFX8-NEXT:    v_mul_lo_u32 v4, s16, v0
+; GFX8-NEXT:    v_mov_b32_e32 v6, s9
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, v4
+; GFX8-NEXT:    v_mul_lo_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v7, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_mul_hi_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v7, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v7, v5
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v4, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v5, v4
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v4
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_addc_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v4, s17, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s16, v3
+; GFX8-NEXT:    v_mul_hi_u32 v8, s16, v0
+; GFX8-NEXT:    v_mul_lo_u32 v7, s16, v0
+; GFX8-NEXT:    v_add_u32_e64 v1, s[0:1], v1, v2
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v5
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v8
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX8-NEXT:    v_mul_lo_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX8-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v8, v2
+; GFX8-NEXT:    v_mul_hi_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v7, s[0:1], v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v7, v5
+; GFX8-NEXT:    v_add_u32_e64 v3, s[0:1], v3, v4
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v2, s13, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s12, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s12, v0
+; GFX8-NEXT:    v_mul_hi_u32 v0, s13, v0
+; GFX8-NEXT:    v_mov_b32_e32 v4, s13
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v5, s13, v1
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_mul_hi_u32 v3, s12, v1
+; GFX8-NEXT:    v_mul_hi_u32 v1, s13, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v5, v0
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v2
+; GFX8-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX8-NEXT:    v_mul_hi_u32 v7, s8, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s8, v0
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s12, v5
+; GFX8-NEXT:    v_subb_u32_e64 v4, s[0:1], v4, v2, vcc
+; GFX8-NEXT:    v_sub_u32_e64 v2, s[0:1], s13, v2
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v4
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, v5, v7, s[0:1]
+; GFX8-NEXT:    v_subrev_u32_e32 v7, vcc, s8, v3
+; GFX8-NEXT:    v_subbrev_u32_e64 v8, s[0:1], 0, v2, vcc
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], 1, v0
+; GFX8-NEXT:    v_addc_u32_e64 v10, s[0:1], 0, v1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v8
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v7
+; GFX8-NEXT:    v_subrev_u32_e32 v6, vcc, s8, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, v11, v12, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v12, s[0:1], 1, v9
+; GFX8-NEXT:    v_subbrev_u32_e32 v2, vcc, 0, v2, vcc
+; GFX8-NEXT:    v_addc_u32_e64 v13, s[0:1], 0, v10, s[0:1]
+; GFX8-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v11
+; GFX8-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v8, v2, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, v3, v5, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, v4, v2, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s[0:1]
+; GFX8-NEXT:    s_xor_b64 s[0:1], s[2:3], s[6:7]
+; GFX8-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX8-NEXT:    s_ashr_i32 s6, s15, 31
+; GFX8-NEXT:    s_ashr_i32 s8, s11, 31
+; GFX8-NEXT:    v_subrev_u32_e32 v0, vcc, s0, v0
+; GFX8-NEXT:    s_add_u32 s0, s14, s6
+; GFX8-NEXT:    v_xor_b32_e32 v1, s1, v1
+; GFX8-NEXT:    v_mov_b32_e32 v4, s1
+; GFX8-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX8-NEXT:    s_and_b32 s1, s1, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX8-NEXT:    s_addc_u32 s1, s15, s6
+; GFX8-NEXT:    s_add_u32 s10, s10, s8
+; GFX8-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX8-NEXT:    s_and_b32 s3, s3, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX8-NEXT:    s_mov_b32 s9, s8
+; GFX8-NEXT:    s_addc_u32 s11, s11, s8
+; GFX8-NEXT:    s_xor_b64 s[10:11], s[10:11], s[8:9]
+; GFX8-NEXT:    v_subb_u32_e32 v1, vcc, v1, v4, vcc
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v4, s11
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v5, s10
+; GFX8-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX8-NEXT:    v_xor_b32_e32 v2, s2, v2
+; GFX8-NEXT:    v_mul_f32_e32 v4, 0x4f800000, v4
+; GFX8-NEXT:    v_add_f32_e32 v4, v4, v5
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v7, v4
+; GFX8-NEXT:    v_mov_b32_e32 v6, s2
+; GFX8-NEXT:    v_subrev_u32_e32 v4, vcc, s2, v3
+; GFX8-NEXT:    v_subb_u32_e32 v5, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v7
+; GFX8-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GFX8-NEXT:    s_mov_b32 s7, s6
+; GFX8-NEXT:    v_trunc_f32_e32 v3, v3
+; GFX8-NEXT:    v_mul_f32_e32 v6, 0xcf800000, v3
+; GFX8-NEXT:    s_xor_b64 s[2:3], s[0:1], s[6:7]
+; GFX8-NEXT:    v_add_f32_e32 v2, v6, v2
+; GFX8-NEXT:    s_sub_u32 s12, 0, s10
+; GFX8-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GFX8-NEXT:    s_and_b32 s0, s0, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX8-NEXT:    s_subb_u32 s13, 0, s11
+; GFX8-NEXT:    v_mul_lo_u32 v6, s13, v2
+; GFX8-NEXT:    v_mul_lo_u32 v7, s12, v3
+; GFX8-NEXT:    v_mul_hi_u32 v9, s12, v2
+; GFX8-NEXT:    v_mul_lo_u32 v8, s12, v2
+; GFX8-NEXT:    v_mov_b32_e32 v10, s11
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v7
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v9
+; GFX8-NEXT:    v_mul_lo_u32 v7, v3, v8
+; GFX8-NEXT:    v_mul_lo_u32 v9, v2, v6
+; GFX8-NEXT:    v_mul_hi_u32 v11, v2, v8
+; GFX8-NEXT:    v_mul_hi_u32 v8, v3, v8
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v9
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v11
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v11, v3, v6
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v9, v7
+; GFX8-NEXT:    v_mul_hi_u32 v9, v2, v6
+; GFX8-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v11, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v9
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v9, vcc, v11, v9
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v8, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v9, v8
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_addc_u32_e64 v7, s[0:1], v3, v6, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v8, s13, v2
+; GFX8-NEXT:    v_mul_lo_u32 v9, s12, v7
+; GFX8-NEXT:    v_mul_hi_u32 v12, s12, v2
+; GFX8-NEXT:    v_mul_lo_u32 v11, s12, v2
+; GFX8-NEXT:    v_add_u32_e64 v3, s[0:1], v3, v6
+; GFX8-NEXT:    v_add_u32_e64 v8, s[0:1], v8, v9
+; GFX8-NEXT:    v_add_u32_e64 v8, s[0:1], v8, v12
+; GFX8-NEXT:    v_mul_lo_u32 v9, v7, v11
+; GFX8-NEXT:    v_mul_lo_u32 v12, v2, v8
+; GFX8-NEXT:    v_mul_hi_u32 v6, v2, v11
+; GFX8-NEXT:    v_mul_hi_u32 v11, v7, v11
+; GFX8-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x0
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], v9, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v6, s[0:1], v9, v6
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX8-NEXT:    v_mul_lo_u32 v9, v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v6, s[0:1], v12, v6
+; GFX8-NEXT:    v_mul_hi_u32 v12, v2, v8
+; GFX8-NEXT:    v_mul_hi_u32 v7, v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], v9, v11
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], v9, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v11, s[0:1], v11, v12
+; GFX8-NEXT:    v_add_u32_e64 v6, s[0:1], v9, v6
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v8, s[0:1], v11, v9
+; GFX8-NEXT:    v_add_u32_e64 v7, s[0:1], v7, v8
+; GFX8-NEXT:    v_addc_u32_e32 v3, vcc, v3, v7, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; GFX8-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v6, s3, v2
+; GFX8-NEXT:    v_mul_lo_u32 v7, s2, v3
+; GFX8-NEXT:    v_mul_hi_u32 v9, s2, v2
+; GFX8-NEXT:    v_mul_hi_u32 v2, s3, v2
+; GFX8-NEXT:    v_mov_b32_e32 v8, s3
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v9
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v9, s3, v3
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v7, v6
+; GFX8-NEXT:    v_mul_hi_u32 v7, s2, v3
+; GFX8-NEXT:    v_mul_hi_u32 v3, s3, v3
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v9, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v9, v7
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v7, v6
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v6
+; GFX8-NEXT:    v_mul_lo_u32 v6, s11, v2
+; GFX8-NEXT:    v_mul_lo_u32 v7, s10, v3
+; GFX8-NEXT:    v_mul_hi_u32 v11, s10, v2
+; GFX8-NEXT:    v_mul_lo_u32 v9, s10, v2
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v7
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v11
+; GFX8-NEXT:    v_sub_u32_e32 v7, vcc, s2, v9
+; GFX8-NEXT:    v_subb_u32_e64 v8, s[0:1], v8, v6, vcc
+; GFX8-NEXT:    v_sub_u32_e64 v6, s[0:1], s3, v6
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v8
+; GFX8-NEXT:    v_subb_u32_e32 v6, vcc, v6, v10, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, v9, v11, s[0:1]
+; GFX8-NEXT:    v_subrev_u32_e32 v11, vcc, s10, v7
+; GFX8-NEXT:    v_subbrev_u32_e64 v12, s[0:1], 0, v6, vcc
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v11
+; GFX8-NEXT:    v_cndmask_b32_e64 v14, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v13, v13, v14, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v14, s[0:1], 1, v2
+; GFX8-NEXT:    v_subb_u32_e32 v6, vcc, v6, v10, vcc
+; GFX8-NEXT:    v_addc_u32_e64 v15, s[0:1], 0, v3, s[0:1]
+; GFX8-NEXT:    v_add_u32_e32 v10, vcc, 1, v14
+; GFX8-NEXT:    v_addc_u32_e32 v16, vcc, 0, v15, vcc
+; GFX8-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v13
+; GFX8-NEXT:    v_subrev_u32_e64 v13, s[0:1], s10, v11
+; GFX8-NEXT:    v_subbrev_u32_e64 v6, s[0:1], 0, v6, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e32 v10, v14, v10, vcc
+; GFX8-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v9
+; GFX8-NEXT:    v_cndmask_b32_e32 v6, v12, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v14, v15, v16, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, v7, v9, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, v2, v10, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, v8, v6, s[0:1]
+; GFX8-NEXT:    s_xor_b64 s[0:1], s[6:7], s[8:9]
+; GFX8-NEXT:    v_xor_b32_e32 v2, s0, v2
+; GFX8-NEXT:    v_xor_b32_e32 v3, s1, v3
+; GFX8-NEXT:    v_mov_b32_e32 v8, s1
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s0, v2
+; GFX8-NEXT:    v_subb_u32_e32 v3, vcc, v3, v8, vcc
+; GFX8-NEXT:    v_xor_b32_e32 v7, s6, v7
+; GFX8-NEXT:    v_xor_b32_e32 v8, s6, v6
+; GFX8-NEXT:    v_mov_b32_e32 v9, s6
+; GFX8-NEXT:    v_subrev_u32_e32 v6, vcc, s6, v7
+; GFX8-NEXT:    v_subb_u32_e32 v7, vcc, v8, v9, vcc
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v8, s12
+; GFX8-NEXT:    v_mov_b32_e32 v9, s13
+; GFX8-NEXT:    flat_store_dwordx4 v[8:9], v[0:3]
+; GFX8-NEXT:    s_nop 0
+; GFX8-NEXT:    v_mov_b32_e32 v0, s14
+; GFX8-NEXT:    v_mov_b32_e32 v1, s15
+; GFX8-NEXT:    flat_store_dwordx4 v[0:1], v[4:7]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_v2i64:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x10
+; GFX9-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_ashr_i32 s2, s13, 31
+; GFX9-NEXT:    s_ashr_i32 s6, s9, 31
+; GFX9-NEXT:    s_add_u32 s0, s12, s2
+; GFX9-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX9-NEXT:    s_and_b32 s1, s1, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX9-NEXT:    s_addc_u32 s1, s13, s2
+; GFX9-NEXT:    s_add_u32 s8, s8, s6
+; GFX9-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX9-NEXT:    s_and_b32 s3, s3, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX9-NEXT:    s_mov_b32 s7, s6
+; GFX9-NEXT:    s_addc_u32 s9, s9, s6
+; GFX9-NEXT:    s_xor_b64 s[8:9], s[8:9], s[6:7]
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX9-NEXT:    s_mov_b32 s3, s2
+; GFX9-NEXT:    s_xor_b64 s[12:13], s[0:1], s[2:3]
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_u32 s16, 0, s8
+; GFX9-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX9-NEXT:    s_and_b32 s0, s0, 1
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX9-NEXT:    s_subb_u32 s17, 0, s9
+; GFX9-NEXT:    v_mul_lo_u32 v3, s17, v0
+; GFX9-NEXT:    v_mul_lo_u32 v2, s16, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s16, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s16, v0
+; GFX9-NEXT:    v_add3_u32 v2, v3, v2, v4
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, v5
+; GFX9-NEXT:    v_mul_lo_u32 v4, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v6, v0, v5
+; GFX9-NEXT:    v_mul_hi_u32 v5, v1, v5
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v6, v1, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v4, v3
+; GFX9-NEXT:    v_mul_hi_u32 v4, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v6, v5
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v5, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v4, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_u32_e32 v5, v6, v5
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_add3_u32 v2, v5, v4, v2
+; GFX9-NEXT:    v_addc_co_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v4, s17, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s16, v3
+; GFX9-NEXT:    v_mul_hi_u32 v6, s16, v0
+; GFX9-NEXT:    v_mul_lo_u32 v7, s16, v0
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v2
+; GFX9-NEXT:    v_add3_u32 v4, v4, v5, v6
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX9-NEXT:    v_mul_lo_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX9-NEXT:    v_add_u32_e32 v2, v6, v2
+; GFX9-NEXT:    v_mul_hi_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_add_u32_e32 v6, v7, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add3_u32 v3, v6, v4, v3
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v2, s13, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s12, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s12, v0
+; GFX9-NEXT:    v_mul_hi_u32 v0, s13, v0
+; GFX9-NEXT:    v_mov_b32_e32 v7, s13
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v4, s13, v1
+; GFX9-NEXT:    v_add_u32_e32 v2, v3, v2
+; GFX9-NEXT:    v_mul_hi_u32 v3, s12, v1
+; GFX9-NEXT:    v_mul_hi_u32 v1, s13, v1
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v4, v0
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v4, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s8, v0
+; GFX9-NEXT:    v_mul_lo_u32 v6, s8, v0
+; GFX9-NEXT:    v_mov_b32_e32 v5, s9
+; GFX9-NEXT:    v_add3_u32 v2, v2, v3, v4
+; GFX9-NEXT:    v_sub_co_u32_e32 v3, vcc, s12, v6
+; GFX9-NEXT:    v_subb_co_u32_e64 v4, s[0:1], v7, v2, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v4
+; GFX9-NEXT:    v_sub_u32_e32 v2, s13, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v4
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v5, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, v6, v7, s[0:1]
+; GFX9-NEXT:    v_subrev_co_u32_e32 v7, vcc, s8, v3
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v8, s[0:1], 0, v2, vcc
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[0:1], 1, v0
+; GFX9-NEXT:    v_addc_co_u32_e64 v10, s[0:1], 0, v1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v8
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v5, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v7
+; GFX9-NEXT:    v_subrev_co_u32_e32 v5, vcc, s8, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, v11, v12, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v12, s[0:1], 1, v9
+; GFX9-NEXT:    v_subbrev_co_u32_e32 v2, vcc, 0, v2, vcc
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v11
+; GFX9-NEXT:    v_addc_co_u32_e64 v13, s[0:1], 0, v10, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc
+; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v6
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v8, v2, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, v3, v5, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, v4, v2, s[0:1]
+; GFX9-NEXT:    s_ashr_i32 s8, s11, 31
+; GFX9-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s[0:1]
+; GFX9-NEXT:    s_xor_b64 s[0:1], s[2:3], s[6:7]
+; GFX9-NEXT:    s_ashr_i32 s6, s15, 31
+; GFX9-NEXT:    s_add_u32 s12, s14, s6
+; GFX9-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX9-NEXT:    s_and_b32 s3, s3, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX9-NEXT:    s_addc_u32 s13, s15, s6
+; GFX9-NEXT:    s_add_u32 s10, s10, s8
+; GFX9-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX9-NEXT:    s_and_b32 s3, s3, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX9-NEXT:    s_mov_b32 s9, s8
+; GFX9-NEXT:    s_addc_u32 s11, s11, s8
+; GFX9-NEXT:    s_xor_b64 s[10:11], s[10:11], s[8:9]
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v4, s11
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v5, s10
+; GFX9-NEXT:    s_mov_b32 s7, s6
+; GFX9-NEXT:    s_xor_b64 s[12:13], s[12:13], s[6:7]
+; GFX9-NEXT:    v_mul_f32_e32 v4, 0x4f800000, v4
+; GFX9-NEXT:    v_add_f32_e32 v4, v4, v5
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; GFX9-NEXT:    s_sub_u32 s3, 0, s10
+; GFX9-NEXT:    v_xor_b32_e32 v1, s1, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, s1
+; GFX9-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; GFX9-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v4
+; GFX9-NEXT:    v_trunc_f32_e32 v6, v6
+; GFX9-NEXT:    v_mul_f32_e32 v7, 0xcf800000, v6
+; GFX9-NEXT:    v_add_f32_e32 v4, v7, v4
+; GFX9-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; GFX9-NEXT:    s_and_b32 s1, s1, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX9-NEXT:    s_subb_u32 s14, 0, s11
+; GFX9-NEXT:    v_mul_lo_u32 v8, s14, v4
+; GFX9-NEXT:    v_mul_lo_u32 v9, s3, v6
+; GFX9-NEXT:    v_mul_hi_u32 v10, s3, v4
+; GFX9-NEXT:    v_mul_lo_u32 v7, s3, v4
+; GFX9-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX9-NEXT:    v_subrev_co_u32_e32 v0, vcc, s0, v0
+; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v5, vcc
+; GFX9-NEXT:    v_add3_u32 v5, v8, v9, v10
+; GFX9-NEXT:    v_mul_lo_u32 v8, v6, v7
+; GFX9-NEXT:    v_mul_lo_u32 v9, v4, v5
+; GFX9-NEXT:    v_mul_hi_u32 v10, v4, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, v6, v7
+; GFX9-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, v8, v9
+; GFX9-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, v8, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v10, v6, v5
+; GFX9-NEXT:    v_add_u32_e32 v8, v9, v8
+; GFX9-NEXT:    v_mul_hi_u32 v9, v4, v5
+; GFX9-NEXT:    v_mul_hi_u32 v5, v6, v5
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v10, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v7, v9
+; GFX9-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v7, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add_u32_e32 v9, v10, v9
+; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v4, v7
+; GFX9-NEXT:    v_add3_u32 v5, v9, v8, v5
+; GFX9-NEXT:    v_addc_co_u32_e64 v7, s[0:1], v6, v5, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v8, s14, v4
+; GFX9-NEXT:    v_mul_lo_u32 v9, s3, v7
+; GFX9-NEXT:    v_mul_hi_u32 v10, s3, v4
+; GFX9-NEXT:    v_mul_lo_u32 v11, s3, v4
+; GFX9-NEXT:    v_add_u32_e32 v5, v6, v5
+; GFX9-NEXT:    v_xor_b32_e32 v2, s2, v2
+; GFX9-NEXT:    v_add3_u32 v8, v8, v9, v10
+; GFX9-NEXT:    v_mul_lo_u32 v9, v7, v11
+; GFX9-NEXT:    v_mul_lo_u32 v10, v4, v8
+; GFX9-NEXT:    v_mul_hi_u32 v6, v4, v11
+; GFX9-NEXT:    v_mul_hi_u32 v11, v7, v11
+; GFX9-NEXT:    v_mov_b32_e32 v12, s2
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[0:1], v9, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v6, s[0:1], v9, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_mul_lo_u32 v9, v7, v8
+; GFX9-NEXT:    v_add_u32_e32 v6, v10, v6
+; GFX9-NEXT:    v_mul_hi_u32 v10, v4, v8
+; GFX9-NEXT:    v_mul_hi_u32 v7, v7, v8
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[0:1], v9, v11
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[0:1], v9, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v6, s[0:1], v9, v6
+; GFX9-NEXT:    v_add_u32_e32 v10, v11, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add3_u32 v7, v10, v8, v7
+; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, v5, v7, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, v4, v6
+; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v5, vcc
+; GFX9-NEXT:    v_subrev_co_u32_e32 v4, vcc, s2, v3
+; GFX9-NEXT:    v_mul_lo_u32 v8, s13, v6
+; GFX9-NEXT:    v_mul_lo_u32 v9, s12, v7
+; GFX9-NEXT:    v_subb_co_u32_e32 v5, vcc, v2, v12, vcc
+; GFX9-NEXT:    v_mul_hi_u32 v2, s12, v6
+; GFX9-NEXT:    v_mul_hi_u32 v6, s13, v6
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v8, v9
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v3, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v3, s13, v7
+; GFX9-NEXT:    v_add_u32_e32 v2, v8, v2
+; GFX9-NEXT:    v_mul_hi_u32 v8, s12, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, s13, v7
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v3, v2
+; GFX9-NEXT:    v_add_u32_e32 v6, v6, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add3_u32 v3, v6, v3, v7
+; GFX9-NEXT:    v_mul_lo_u32 v6, s11, v2
+; GFX9-NEXT:    v_mul_lo_u32 v7, s10, v3
+; GFX9-NEXT:    v_mul_hi_u32 v8, s10, v2
+; GFX9-NEXT:    v_mul_lo_u32 v10, s10, v2
+; GFX9-NEXT:    v_mov_b32_e32 v11, s13
+; GFX9-NEXT:    v_mov_b32_e32 v9, s11
+; GFX9-NEXT:    v_add3_u32 v6, v6, v7, v8
+; GFX9-NEXT:    v_sub_co_u32_e32 v7, vcc, s12, v10
+; GFX9-NEXT:    v_subb_co_u32_e64 v8, s[0:1], v11, v6, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v8
+; GFX9-NEXT:    v_sub_u32_e32 v6, s13, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v8
+; GFX9-NEXT:    v_subb_co_u32_e32 v6, vcc, v6, v9, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, v10, v11, s[0:1]
+; GFX9-NEXT:    v_subrev_co_u32_e32 v11, vcc, s10, v7
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v12, s[0:1], 0, v6, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v12
+; GFX9-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v11
+; GFX9-NEXT:    v_cndmask_b32_e64 v14, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v12
+; GFX9-NEXT:    v_cndmask_b32_e64 v13, v13, v14, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v14, s[0:1], 1, v2
+; GFX9-NEXT:    v_subb_co_u32_e32 v6, vcc, v6, v9, vcc
+; GFX9-NEXT:    v_addc_co_u32_e64 v15, s[0:1], 0, v3, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e32 v9, vcc, 1, v14
+; GFX9-NEXT:    v_addc_co_u32_e32 v16, vcc, 0, v15, vcc
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v13
+; GFX9-NEXT:    v_cndmask_b32_e32 v9, v14, v9, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v14, v15, v16, vcc
+; GFX9-NEXT:    v_subrev_co_u32_e64 v15, s[0:1], s10, v11
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v6, s[0:1], 0, v6, s[0:1]
+; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v10
+; GFX9-NEXT:    v_cndmask_b32_e32 v6, v12, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, v2, v9, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e32 v9, v11, v15, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, v7, v9, s[0:1]
+; GFX9-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x0
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, v8, v6, s[0:1]
+; GFX9-NEXT:    s_xor_b64 s[0:1], s[6:7], s[8:9]
+; GFX9-NEXT:    v_xor_b32_e32 v2, s0, v2
+; GFX9-NEXT:    v_xor_b32_e32 v3, s1, v3
+; GFX9-NEXT:    v_mov_b32_e32 v8, s1
+; GFX9-NEXT:    v_subrev_co_u32_e32 v2, vcc, s0, v2
+; GFX9-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v8, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v7, s6, v7
+; GFX9-NEXT:    v_xor_b32_e32 v8, s6, v6
+; GFX9-NEXT:    v_mov_b32_e32 v13, 0
+; GFX9-NEXT:    v_mov_b32_e32 v9, s6
+; GFX9-NEXT:    v_subrev_co_u32_e32 v6, vcc, s6, v7
+; GFX9-NEXT:    v_subb_co_u32_e32 v7, vcc, v8, v9, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dwordx4 v13, v[0:3], s[12:13]
+; GFX9-NEXT:    global_store_dwordx4 v13, v[4:7], s[14:15]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_v2i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_clause 0x1
+; GFX10-NEXT:    s_load_dwordx4 s[16:19], s[4:5], 0x10
+; GFX10-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x20
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_ashr_i32 s2, s17, 31
+; GFX10-NEXT:    s_ashr_i32 s0, s13, 31
+; GFX10-NEXT:    s_add_u32 s8, s16, s2
+; GFX10-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX10-NEXT:    s_and_b32 s1, s1, 1
+; GFX10-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX10-NEXT:    s_addc_u32 s9, s17, s2
+; GFX10-NEXT:    s_add_u32 s6, s12, s0
+; GFX10-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX10-NEXT:    s_and_b32 s3, s1, 1
+; GFX10-NEXT:    s_mov_b32 s1, s0
+; GFX10-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX10-NEXT:    s_mov_b32 s3, s2
+; GFX10-NEXT:    s_addc_u32 s7, s13, s0
+; GFX10-NEXT:    s_xor_b64 s[8:9], s[8:9], s[2:3]
+; GFX10-NEXT:    s_xor_b64 s[6:7], s[6:7], s[0:1]
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s7
+; GFX10-NEXT:    s_sub_u32 s20, 0, s6
+; GFX10-NEXT:    s_cselect_b32 s10, 1, 0
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX10-NEXT:    s_and_b32 s10, s10, 1
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; GFX10-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX10-NEXT:    s_subb_u32 s21, 0, s7
+; GFX10-NEXT:    s_ashr_i32 s10, s19, 31
+; GFX10-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX10-NEXT:    s_ashr_i32 s12, s15, 31
+; GFX10-NEXT:    s_xor_b64 s[16:17], s[2:3], s[0:1]
+; GFX10-NEXT:    s_add_u32 s0, s18, s10
+; GFX10-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    s_and_b32 s1, s1, 1
+; GFX10-NEXT:    s_mov_b32 s13, s12
+; GFX10-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX10-NEXT:    s_mov_b32 s11, s10
+; GFX10-NEXT:    s_addc_u32 s1, s19, s10
+; GFX10-NEXT:    s_add_u32 s14, s14, s12
+; GFX10-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX10-NEXT:    s_and_b32 s3, s3, 1
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX10-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX10-NEXT:    s_addc_u32 s15, s15, s12
+; GFX10-NEXT:    s_xor_b64 s[18:19], s[0:1], s[10:11]
+; GFX10-NEXT:    s_xor_b64 s[14:15], s[14:15], s[12:13]
+; GFX10-NEXT:    v_mul_f32_e32 v2, 0x2f800000, v0
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s15
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v3, s14
+; GFX10-NEXT:    s_sub_u32 s3, 0, s14
+; GFX10-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX10-NEXT:    v_trunc_f32_e32 v2, v2
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; GFX10-NEXT:    s_and_b32 s0, s0, 1
+; GFX10-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX10-NEXT:    v_add_f32_e32 v1, v1, v3
+; GFX10-NEXT:    v_mul_f32_e32 v3, 0xcf800000, v2
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX10-NEXT:    s_subb_u32 s22, 0, s15
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_add_f32_e32 v0, v3, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s20, v2
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x5f7ffffc, v1
+; GFX10-NEXT:    v_mul_lo_u32 v4, s21, v0
+; GFX10-NEXT:    v_mul_hi_u32 v5, s20, v0
+; GFX10-NEXT:    v_mul_lo_u32 v6, s20, v0
+; GFX10-NEXT:    v_mul_f32_e32 v7, 0x2f800000, v1
+; GFX10-NEXT:    v_add3_u32 v3, v4, v3, v5
+; GFX10-NEXT:    v_trunc_f32_e32 v4, v7
+; GFX10-NEXT:    v_mul_lo_u32 v5, v2, v6
+; GFX10-NEXT:    v_mul_hi_u32 v7, v0, v6
+; GFX10-NEXT:    v_mul_hi_u32 v6, v2, v6
+; GFX10-NEXT:    v_mul_lo_u32 v8, v0, v3
+; GFX10-NEXT:    v_mul_lo_u32 v10, v2, v3
+; GFX10-NEXT:    v_mul_f32_e32 v9, 0xcf800000, v4
+; GFX10-NEXT:    v_mul_hi_u32 v11, v0, v3
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; GFX10-NEXT:    v_mul_hi_u32 v3, v2, v3
+; GFX10-NEXT:    v_add_f32_e32 v1, v9, v1
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v8
+; GFX10-NEXT:    v_mul_lo_u32 v9, s3, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v10, v6
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_mul_lo_u32 v12, s22, v1
+; GFX10-NEXT:    v_mul_hi_u32 v13, s3, v1
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v6, v11
+; GFX10-NEXT:    v_mul_lo_u32 v11, s3, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, v8, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v6, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v7, v10, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add3_u32 v8, v12, v9, v13
+; GFX10-NEXT:    v_mul_lo_u32 v9, v4, v11
+; GFX10-NEXT:    v_mul_hi_u32 v10, v1, v11
+; GFX10-NEXT:    v_mul_hi_u32 v11, v4, v11
+; GFX10-NEXT:    v_add3_u32 v3, v7, v6, v3
+; GFX10-NEXT:    v_mul_lo_u32 v6, v1, v8
+; GFX10-NEXT:    v_mul_lo_u32 v7, v4, v8
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v5
+; GFX10-NEXT:    v_mul_hi_u32 v5, v1, v8
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v12, s0, v2, v3, vcc_lo
+; GFX10-NEXT:    v_mul_hi_u32 v8, v4, v8
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, v2, v3
+; GFX10-NEXT:    v_mul_lo_u32 v14, s21, v0
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v9, v6
+; GFX10-NEXT:    v_mul_hi_u32 v15, s20, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v7, s0, v7, v11
+; GFX10-NEXT:    v_mul_lo_u32 v16, s20, v12
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v6, v10
+; GFX10-NEXT:    v_mul_lo_u32 v13, s20, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v7, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s0
+; GFX10-NEXT:    v_add3_u32 v14, v14, v16, v15
+; GFX10-NEXT:    v_add_nc_u32_e32 v6, v9, v6
+; GFX10-NEXT:    v_mul_lo_u32 v10, v12, v13
+; GFX10-NEXT:    v_add_nc_u32_e32 v7, v11, v7
+; GFX10-NEXT:    v_mul_lo_u32 v11, v0, v14
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v6
+; GFX10-NEXT:    v_mul_hi_u32 v9, v0, v13
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_mul_hi_u32 v13, v12, v13
+; GFX10-NEXT:    v_mul_lo_u32 v15, v12, v14
+; GFX10-NEXT:    v_add_co_u32 v1, s0, v1, v5
+; GFX10-NEXT:    v_mul_hi_u32 v16, v0, v14
+; GFX10-NEXT:    v_add3_u32 v6, v7, v6, v8
+; GFX10-NEXT:    v_add_co_u32 v5, s1, v10, v11
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s1
+; GFX10-NEXT:    v_mul_lo_u32 v3, s3, v1
+; GFX10-NEXT:    v_add_co_u32 v8, s1, v15, v13
+; GFX10-NEXT:    v_mul_lo_u32 v13, s22, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v11, s1, v4, v6, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s1, v5, v9
+; GFX10-NEXT:    v_mul_hi_u32 v15, s3, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v8, s1, v8, v16
+; GFX10-NEXT:    v_mul_lo_u32 v9, s3, v11
+; GFX10-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s1
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, v4, v6
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, v7, v5
+; GFX10-NEXT:    v_mul_hi_u32 v7, v12, v14
+; GFX10-NEXT:    v_mul_lo_u32 v12, v11, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v10, v10, v16
+; GFX10-NEXT:    s_load_dwordx4 s[20:23], s[4:5], 0x0
+; GFX10-NEXT:    v_add_co_u32 v5, s1, v8, v5
+; GFX10-NEXT:    v_add3_u32 v9, v13, v9, v15
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s1
+; GFX10-NEXT:    v_mul_hi_u32 v13, v1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v3, v11, v3
+; GFX10-NEXT:    v_mul_lo_u32 v14, v1, v9
+; GFX10-NEXT:    v_add3_u32 v7, v10, v8, v7
+; GFX10-NEXT:    v_mul_lo_u32 v8, v11, v9
+; GFX10-NEXT:    v_mul_hi_u32 v10, v1, v9
+; GFX10-NEXT:    v_mul_hi_u32 v9, v11, v9
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v2, vcc_lo, v2, v7, vcc_lo
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v5
+; GFX10-NEXT:    v_add_co_u32 v7, s1, v12, v14
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s1
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v2, vcc_lo, 0, v2, vcc_lo
+; GFX10-NEXT:    v_add_co_u32 v3, s1, v8, v3
+; GFX10-NEXT:    v_mul_lo_u32 v8, s9, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v7, s1, v7, v13
+; GFX10-NEXT:    v_mul_lo_u32 v14, s8, v2
+; GFX10-NEXT:    v_mul_hi_u32 v12, s8, v0
+; GFX10-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX10-NEXT:    v_mul_lo_u32 v13, s9, v2
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v3, s1, v3, v10
+; GFX10-NEXT:    v_mul_hi_u32 v15, s8, v2
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v8, s1, v8, v14
+; GFX10-NEXT:    v_add_nc_u32_e32 v7, v11, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v0, s1, v13, v0
+; GFX10-NEXT:    v_mul_hi_u32 v2, s9, v2
+; GFX10-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v8, s1, v8, v12
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, v5, v10
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v0, s1, v0, v15
+; GFX10-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s1
+; GFX10-NEXT:    v_add_nc_u32_e32 v8, v14, v8
+; GFX10-NEXT:    v_add_nc_u32_e32 v10, v13, v12
+; GFX10-NEXT:    v_add_co_u32 v0, s1, v0, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v3, s1, v3, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s1
+; GFX10-NEXT:    v_mul_lo_u32 v6, s7, v0
+; GFX10-NEXT:    v_add3_u32 v2, v10, v8, v2
+; GFX10-NEXT:    v_add3_u32 v5, v5, v7, v9
+; GFX10-NEXT:    v_mul_hi_u32 v7, s6, v0
+; GFX10-NEXT:    v_mul_lo_u32 v8, s6, v2
+; GFX10-NEXT:    v_mov_b32_e32 v9, 0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v4, vcc_lo, v4, v5, s0
+; GFX10-NEXT:    v_mul_lo_u32 v5, s6, v0
+; GFX10-NEXT:    v_add_co_u32 v1, vcc_lo, v1, v3
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v4, vcc_lo
+; GFX10-NEXT:    v_add3_u32 v4, v6, v8, v7
+; GFX10-NEXT:    v_mul_lo_u32 v6, s19, v1
+; GFX10-NEXT:    v_mul_hi_u32 v7, s19, v1
+; GFX10-NEXT:    v_sub_co_u32 v5, vcc_lo, s8, v5
+; GFX10-NEXT:    v_mul_lo_u32 v14, s18, v3
+; GFX10-NEXT:    v_sub_nc_u32_e32 v8, s9, v4
+; GFX10-NEXT:    v_sub_co_ci_u32_e64 v4, s0, s9, v4, vcc_lo
+; GFX10-NEXT:    v_mul_lo_u32 v15, s19, v3
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s6, v5
+; GFX10-NEXT:    v_mul_hi_u32 v1, s18, v1
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v8, vcc_lo, s7, v8, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v4
+; GFX10-NEXT:    v_mul_hi_u32 v17, s18, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s0
+; GFX10-NEXT:    v_mul_hi_u32 v3, s19, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v12, vcc_lo, v5, s6
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v13, s0, 0, v8, vcc_lo
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s7, v4
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v8, vcc_lo, s7, v8, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, v11, v10, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s7, v13
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s6, v12
+; GFX10-NEXT:    v_cndmask_b32_e64 v16, 0, -1, s0
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v6, v14
+; GFX10-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v7, s0, v15, v7
+; GFX10-NEXT:    v_add_co_u32 v1, s1, v6, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v7, s0, v7, v17
+; GFX10-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v17, s0, v0, 1
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v14, v1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v18, s0, 0, v2, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s7, v13
+; GFX10-NEXT:    v_add_nc_u32_e32 v6, v6, v15
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, v11, v16, s0
+; GFX10-NEXT:    v_add_co_u32 v7, s0, v7, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v14, s0, v17, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v15, s0, 0, v18, s0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, v11
+; GFX10-NEXT:    v_add3_u32 v3, v6, v1, v3
+; GFX10-NEXT:    v_mul_lo_u32 v10, s15, v7
+; GFX10-NEXT:    v_mul_lo_u32 v16, s14, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v18, v15, s0
+; GFX10-NEXT:    v_mul_lo_u32 v11, s14, v3
+; GFX10-NEXT:    v_mul_hi_u32 v15, s14, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v17, v14, s0
+; GFX10-NEXT:    v_sub_co_u32 v14, s1, v12, s6
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v8, s1, 0, v8, s1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v2, v6, vcc_lo
+; GFX10-NEXT:    v_add3_u32 v6, v10, v11, v15
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v13, v8, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v12, v12, v14, s0
+; GFX10-NEXT:    v_sub_co_u32 v8, s0, s18, v16
+; GFX10-NEXT:    v_xor_b32_e32 v0, s16, v0
+; GFX10-NEXT:    v_sub_co_ci_u32_e64 v10, s1, s19, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v5, v5, v12, vcc_lo
+; GFX10-NEXT:    v_sub_nc_u32_e32 v4, s19, v6
+; GFX10-NEXT:    v_xor_b32_e32 v1, s17, v1
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s15, v10
+; GFX10-NEXT:    v_xor_b32_e32 v2, s2, v2
+; GFX10-NEXT:    v_xor_b32_e32 v5, s2, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc_lo
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v4, vcc_lo, s15, v4, s0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s14, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v12, vcc_lo, v8, s14
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v13, s0, 0, v4, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v0, s0, v0, s16
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v4, vcc_lo, s15, v4, vcc_lo
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v1, s0, s17, v1, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s15, v10
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v6, v11, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s15, v13
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s14, v12
+; GFX10-NEXT:    v_cndmask_b32_e64 v14, 0, -1, s0
+; GFX10-NEXT:    v_add_co_u32 v15, s0, v7, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v16, s0, 0, v3, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s15, v13
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, v11, v14, s0
+; GFX10-NEXT:    v_add_co_u32 v14, s0, v15, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v17, s0, 0, v16, s0
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11
+; GFX10-NEXT:    v_sub_co_u32 v11, s0, v12, s14
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v4, s0, 0, v4, s0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, v6
+; GFX10-NEXT:    v_cndmask_b32_e32 v14, v15, v14, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v15, v16, v17, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v6, v12, v11, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v7, v14, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v15, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v8, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, v10, v4, s0
+; GFX10-NEXT:    v_sub_co_u32 v4, vcc_lo, v5, s2
+; GFX10-NEXT:    s_xor_b64 s[0:1], s[10:11], s[12:13]
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v5, vcc_lo, s2, v2, vcc_lo
+; GFX10-NEXT:    v_xor_b32_e32 v2, s0, v7
+; GFX10-NEXT:    v_xor_b32_e32 v3, s1, v3
+; GFX10-NEXT:    v_xor_b32_e32 v6, s10, v6
+; GFX10-NEXT:    v_xor_b32_e32 v7, s10, v8
+; GFX10-NEXT:    v_sub_co_u32 v2, vcc_lo, v2, s0
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v3, vcc_lo, s1, v3, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v6, vcc_lo, v6, s10
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v7, vcc_lo, s10, v7, vcc_lo
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dwordx4 v9, v[0:3], s[20:21]
+; GFX10-NEXT:    global_store_dwordx4 v9, v[4:7], s[22:23]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv <2 x i64> %x, %y
+  store <2 x i64> %div, <2 x i64> addrspace(1)* %out0
+  %rem = srem <2 x i64> %x, %y
+  store <2 x i64> %rem, <2 x i64> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdiv_i8(i8 addrspace(1)* %out0, i8 addrspace(1)* %out1, i8 %x, i8 %y) {
+; GFX8-LABEL: sdiv_i8:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX8-NEXT:    s_sext_i32_i8 s1, s1
+; GFX8-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX8-NEXT:    s_add_i32 s1, s1, s6
+; GFX8-NEXT:    s_xor_b32 s7, s1, s6
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX8-NEXT:    s_sub_i32 s1, 0, s7
+; GFX8-NEXT:    s_sext_i32_i8 s0, s0
+; GFX8-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_add_i32 s0, s0, s8
+; GFX8-NEXT:    s_xor_b32 s9, s0, s8
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    s_xor_b32 s4, s8, s6
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s9, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s7
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s9, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_xor_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s4, v2
+; GFX8-NEXT:    v_xor_b32_e32 v3, s8, v3
+; GFX8-NEXT:    flat_store_byte v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s8, v3
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_byte v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdiv_i8:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX9-NEXT:    s_sext_i32_i8 s1, s1
+; GFX9-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX9-NEXT:    s_add_i32 s1, s1, s6
+; GFX9-NEXT:    s_xor_b32 s7, s1, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_sub_i32 s1, 0, s7
+; GFX9-NEXT:    s_sext_i32_i8 s0, s0
+; GFX9-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_add_i32 s0, s0, s8
+; GFX9-NEXT:    s_xor_b32 s9, s0, s8
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_xor_b32 s4, s8, s6
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s9, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s4, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s8, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdiv_i8:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX10-NEXT:    s_sext_i32_i8 s0, s0
+; GFX10-NEXT:    s_sext_i32_i8 s1, s1
+; GFX10-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX10-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX10-NEXT:    s_add_i32 s0, s0, s8
+; GFX10-NEXT:    s_add_i32 s1, s1, s6
+; GFX10-NEXT:    s_xor_b32 s0, s0, s8
+; GFX10-NEXT:    s_xor_b32 s7, s1, s6
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX10-NEXT:    s_sub_i32 s1, 0, s7
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    s_xor_b32 s4, s8, s6
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s4, v0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s8, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv i8 %x, %y
+  store i8 %div, i8 addrspace(1)* %out0
+  %rem = srem i8 %x, %y
+  store i8 %rem, i8 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_v2i8(<2 x i8> addrspace(1)* %out0, <2 x i8> addrspace(1)* %out1, <2 x i8> %x, <2 x i8> %y) {
+; GFX8-LABEL: sdivrem_v2i8:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX8-NEXT:    s_sext_i32_i8 s1, s1
+; GFX8-NEXT:    s_ashr_i32 s2, s1, 31
+; GFX8-NEXT:    s_add_i32 s1, s1, s2
+; GFX8-NEXT:    s_xor_b32 s3, s1, s2
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s3
+; GFX8-NEXT:    s_sub_i32 s7, 0, s3
+; GFX8-NEXT:    s_lshr_b32 s1, s0, 24
+; GFX8-NEXT:    s_sext_i32_i8 s6, s0
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_sext_i32_i8 s1, s1
+; GFX8-NEXT:    s_ashr_i32 s8, s6, 31
+; GFX8-NEXT:    s_ashr_i32 s9, s1, 31
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    s_add_i32 s6, s6, s8
+; GFX8-NEXT:    s_add_i32 s1, s1, s9
+; GFX8-NEXT:    s_xor_b32 s11, s1, s9
+; GFX8-NEXT:    v_mul_lo_u32 v1, s7, v0
+; GFX8-NEXT:    s_xor_b32 s10, s6, s8
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v2, s11
+; GFX8-NEXT:    s_lshr_b32 s12, s0, 8
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v0, s10, v0
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v1, v2
+; GFX8-NEXT:    v_mul_lo_u32 v2, v0, s3
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s10, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s3, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s3, v2
+; GFX8-NEXT:    s_sub_i32 s1, 0, s11
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v3, s1, v1
+; GFX8-NEXT:    s_sext_i32_i8 s1, s12
+; GFX8-NEXT:    s_xor_b32 s0, s8, s2
+; GFX8-NEXT:    s_ashr_i32 s2, s1, 31
+; GFX8-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX8-NEXT:    s_add_i32 s1, s1, s2
+; GFX8-NEXT:    s_xor_b32 s1, s1, s2
+; GFX8-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v3
+; GFX8-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX8-NEXT:    v_xor_b32_e32 v2, s8, v2
+; GFX8-NEXT:    v_subrev_u32_e32 v0, vcc, s0, v0
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s8, v2
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, s11
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s1, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s11, v3
+; GFX8-NEXT:    s_xor_b32 s0, s2, s9
+; GFX8-NEXT:    v_xor_b32_e32 v1, s0, v1
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e32 v1, vcc, s0, v1
+; GFX8-NEXT:    s_movk_i32 s0, 0xff
+; GFX8-NEXT:    v_and_b32_e32 v1, s0, v1
+; GFX8-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; GFX8-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX8-NEXT:    v_or_b32_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s4
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s2, v3
+; GFX8-NEXT:    flat_store_short v[0:1], v4
+; GFX8-NEXT:    v_and_b32_e32 v0, s0, v3
+; GFX8-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX8-NEXT:    v_or_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_short v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_v2i8:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s0, s6, 16
+; GFX9-NEXT:    s_sext_i32_i8 s0, s0
+; GFX9-NEXT:    s_ashr_i32 s7, s0, 31
+; GFX9-NEXT:    s_add_i32 s0, s0, s7
+; GFX9-NEXT:    s_xor_b32 s8, s0, s7
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_lshr_b32 s5, s6, 24
+; GFX9-NEXT:    s_sext_i32_i8 s5, s5
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_ashr_i32 s9, s5, 31
+; GFX9-NEXT:    s_add_i32 s5, s5, s9
+; GFX9-NEXT:    s_xor_b32 s5, s5, s9
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_i32 s10, 0, s8
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v2, s5
+; GFX9-NEXT:    s_lshr_b32 s4, s6, 8
+; GFX9-NEXT:    v_mul_lo_u32 v1, s10, v0
+; GFX9-NEXT:    s_sext_i32_i8 s6, s6
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; GFX9-NEXT:    s_ashr_i32 s10, s6, 31
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    s_add_i32 s6, s6, s10
+; GFX9-NEXT:    s_xor_b32 s6, s6, s10
+; GFX9-NEXT:    s_sub_i32 s11, 0, s5
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v2
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s6, v0
+; GFX9-NEXT:    s_sext_i32_i8 s4, s4
+; GFX9-NEXT:    s_xor_b32 s7, s10, s7
+; GFX9-NEXT:    v_mul_lo_u32 v3, s11, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, v0, s8
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX9-NEXT:    v_sub_u32_e32 v2, s6, v2
+; GFX9-NEXT:    s_ashr_i32 s6, s4, 31
+; GFX9-NEXT:    s_add_i32 s4, s4, s6
+; GFX9-NEXT:    s_xor_b32 s4, s4, s6
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v3
+; GFX9-NEXT:    v_mul_hi_u32 v1, s4, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s8, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s8, v2
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, s5
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s8, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s8, v2
+; GFX9-NEXT:    v_sub_u32_e32 v3, s4, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s5, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s5, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s5, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    s_xor_b32 s4, s6, s9
+; GFX9-NEXT:    v_xor_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s5, v3
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s4, v1
+; GFX9-NEXT:    s_movk_i32 s4, 0xff
+; GFX9-NEXT:    v_xor_b32_e32 v0, s7, v0
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_xor_b32_e32 v3, s6, v3
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s7, v0
+; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v3
+; GFX9-NEXT:    v_xor_b32_e32 v2, s10, v2
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_short v1, v0, s[0:1]
+; GFX9-NEXT:    v_and_b32_e32 v0, s4, v3
+; GFX9-NEXT:    v_subrev_u32_e32 v2, s10, v2
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    global_store_short v1, v0, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_v2i8:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s1, s0, 24
+; GFX10-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX10-NEXT:    s_sext_i32_i8 s1, s1
+; GFX10-NEXT:    s_sext_i32_i8 s2, s2
+; GFX10-NEXT:    s_ashr_i32 s3, s1, 31
+; GFX10-NEXT:    s_ashr_i32 s8, s2, 31
+; GFX10-NEXT:    s_add_i32 s1, s1, s3
+; GFX10-NEXT:    s_add_i32 s2, s2, s8
+; GFX10-NEXT:    s_xor_b32 s1, s1, s3
+; GFX10-NEXT:    s_xor_b32 s2, s2, s8
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s1
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s2
+; GFX10-NEXT:    s_sub_i32 s6, 0, s1
+; GFX10-NEXT:    s_sub_i32 s7, 0, s2
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, s6, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s7, v1
+; GFX10-NEXT:    s_sext_i32_i8 s6, s0
+; GFX10-NEXT:    s_lshr_b32 s0, s0, 8
+; GFX10-NEXT:    s_ashr_i32 s9, s6, 31
+; GFX10-NEXT:    s_sext_i32_i8 s0, s0
+; GFX10-NEXT:    s_add_i32 s6, s6, s9
+; GFX10-NEXT:    s_ashr_i32 s10, s0, 31
+; GFX10-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX10-NEXT:    s_add_i32 s0, s0, s10
+; GFX10-NEXT:    s_xor_b32 s6, s6, s9
+; GFX10-NEXT:    s_xor_b32 s0, s0, s10
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s6, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, v0, s1
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, s2
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v1
+; GFX10-NEXT:    v_sub_nc_u32_e32 v2, s0, v2
+; GFX10-NEXT:    v_sub_nc_u32_e32 v3, s6, v3
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s1, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s1, v2
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s2, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s2, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v7, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s1, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s1, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v1
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s2, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s2, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
+; GFX10-NEXT:    s_xor_b32 s1, s10, s3
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v7, s0
+; GFX10-NEXT:    v_xor_b32_e32 v0, s1, v0
+; GFX10-NEXT:    v_xor_b32_e32 v2, s10, v2
+; GFX10-NEXT:    s_xor_b32 s0, s9, s8
+; GFX10-NEXT:    v_xor_b32_e32 v1, s0, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s1, v0
+; GFX10-NEXT:    v_xor_b32_e32 v3, s9, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v2, s10, v2
+; GFX10-NEXT:    s_movk_i32 s1, 0xff
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    v_and_b32_sdwa v0, v0, s1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s9, v3
+; GFX10-NEXT:    v_and_b32_sdwa v2, v2, s1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_short v1, v0, s[4:5]
+; GFX10-NEXT:    global_store_short v1, v2, s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv <2 x i8> %x, %y
+  store <2 x i8> %div, <2 x i8> addrspace(1)* %out0
+  %rem = srem <2 x i8> %x, %y
+  store <2 x i8> %rem, <2 x i8> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdiv_i16(i16 addrspace(1)* %out0, i16 addrspace(1)* %out1, i16 %x, i16 %y) {
+; GFX8-LABEL: sdiv_i16:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX8-NEXT:    s_sext_i32_i16 s1, s1
+; GFX8-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX8-NEXT:    s_add_i32 s1, s1, s6
+; GFX8-NEXT:    s_xor_b32 s7, s1, s6
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX8-NEXT:    s_sub_i32 s1, 0, s7
+; GFX8-NEXT:    s_sext_i32_i16 s0, s0
+; GFX8-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_add_i32 s0, s0, s8
+; GFX8-NEXT:    s_xor_b32 s9, s0, s8
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    s_xor_b32 s4, s8, s6
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s9, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s7
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s9, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_xor_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s4, v2
+; GFX8-NEXT:    v_xor_b32_e32 v3, s8, v3
+; GFX8-NEXT:    flat_store_short v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s8, v3
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_short v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdiv_i16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX9-NEXT:    s_sext_i32_i16 s1, s1
+; GFX9-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX9-NEXT:    s_add_i32 s1, s1, s6
+; GFX9-NEXT:    s_xor_b32 s7, s1, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_sub_i32 s1, 0, s7
+; GFX9-NEXT:    s_sext_i32_i16 s0, s0
+; GFX9-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_add_i32 s0, s0, s8
+; GFX9-NEXT:    s_xor_b32 s9, s0, s8
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_xor_b32 s4, s8, s6
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s9, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s4, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s8, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_short v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_short v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdiv_i16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX10-NEXT:    s_sext_i32_i16 s0, s0
+; GFX10-NEXT:    s_sext_i32_i16 s1, s1
+; GFX10-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX10-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX10-NEXT:    s_add_i32 s0, s0, s8
+; GFX10-NEXT:    s_add_i32 s1, s1, s6
+; GFX10-NEXT:    s_xor_b32 s0, s0, s8
+; GFX10-NEXT:    s_xor_b32 s7, s1, s6
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX10-NEXT:    s_sub_i32 s1, 0, s7
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    s_xor_b32 s4, s8, s6
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s4, v0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s8, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_short v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_short v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv i16 %x, %y
+  store i16 %div, i16 addrspace(1)* %out0
+  %rem = srem i16 %x, %y
+  store i16 %rem, i16 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_v2i16(<2 x i16> addrspace(1)* %out0, <2 x i16> addrspace(1)* %out1, <2 x i16> %x, <2 x i16> %y) {
+; GFX8-LABEL: sdivrem_v2i16:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x14
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_sext_i32_i16 s1, s0
+; GFX8-NEXT:    s_ashr_i32 s2, s1, 31
+; GFX8-NEXT:    s_add_i32 s1, s1, s2
+; GFX8-NEXT:    s_xor_b32 s3, s1, s2
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s3
+; GFX8-NEXT:    s_sub_i32 s7, 0, s3
+; GFX8-NEXT:    s_load_dword s1, s[4:5], 0x10
+; GFX8-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_sext_i32_i16 s0, s0
+; GFX8-NEXT:    s_ashr_i32 s9, s0, 31
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_sext_i32_i16 s6, s1
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    s_ashr_i32 s8, s6, 31
+; GFX8-NEXT:    s_add_i32 s6, s6, s8
+; GFX8-NEXT:    s_add_i32 s0, s0, s9
+; GFX8-NEXT:    v_mul_lo_u32 v1, s7, v0
+; GFX8-NEXT:    s_xor_b32 s10, s6, s8
+; GFX8-NEXT:    s_xor_b32 s11, s0, s9
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v2, s11
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    s_lshr_b32 s12, s1, 16
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v0, s10, v0
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v1, v2
+; GFX8-NEXT:    v_mul_lo_u32 v2, v0, s3
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s10, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s3, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s3, v2
+; GFX8-NEXT:    s_sub_i32 s1, 0, s11
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v3, s1, v1
+; GFX8-NEXT:    s_sext_i32_i16 s1, s12
+; GFX8-NEXT:    s_xor_b32 s0, s8, s2
+; GFX8-NEXT:    s_ashr_i32 s2, s1, 31
+; GFX8-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX8-NEXT:    s_add_i32 s1, s1, s2
+; GFX8-NEXT:    s_xor_b32 s1, s1, s2
+; GFX8-NEXT:    v_xor_b32_e32 v0, s0, v0
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v3
+; GFX8-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX8-NEXT:    v_xor_b32_e32 v2, s8, v2
+; GFX8-NEXT:    v_subrev_u32_e32 v0, vcc, s0, v0
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s8, v2
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, s11
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s1, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s11, v3
+; GFX8-NEXT:    s_xor_b32 s0, s2, s9
+; GFX8-NEXT:    v_xor_b32_e32 v1, s0, v1
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e32 v1, vcc, s0, v1
+; GFX8-NEXT:    s_mov_b32 s0, 0xffff
+; GFX8-NEXT:    v_xor_b32_e32 v3, s2, v3
+; GFX8-NEXT:    v_and_b32_e32 v1, s0, v1
+; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s2, v3
+; GFX8-NEXT:    v_or_b32_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    v_and_b32_e32 v0, s0, v3
+; GFX8-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT:    v_or_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s4
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    flat_store_dword v[0:1], v4
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_v2i16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x14
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_sext_i32_i16 s0, s6
+; GFX9-NEXT:    s_ashr_i32 s7, s0, 31
+; GFX9-NEXT:    s_add_i32 s0, s0, s7
+; GFX9-NEXT:    s_xor_b32 s8, s0, s7
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_load_dword s9, s[4:5], 0x10
+; GFX9-NEXT:    s_lshr_b32 s4, s6, 16
+; GFX9-NEXT:    s_sext_i32_i16 s4, s4
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_i32 s10, 0, s8
+; GFX9-NEXT:    s_ashr_i32 s6, s4, 31
+; GFX9-NEXT:    s_add_i32 s4, s4, s6
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    s_xor_b32 s4, s4, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v2, s4
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_sext_i32_i16 s5, s9
+; GFX9-NEXT:    v_mul_lo_u32 v1, s10, v0
+; GFX9-NEXT:    s_ashr_i32 s10, s5, 31
+; GFX9-NEXT:    s_add_i32 s5, s5, s10
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    s_xor_b32 s5, s5, s10
+; GFX9-NEXT:    s_sub_i32 s11, 0, s4
+; GFX9-NEXT:    s_lshr_b32 s9, s9, 16
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s5, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v2
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    s_xor_b32 s7, s10, s7
+; GFX9-NEXT:    v_mul_lo_u32 v2, v0, s8
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s11, v1
+; GFX9-NEXT:    v_sub_u32_e32 v2, s5, v2
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s8, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s8, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s8, v2
+; GFX9-NEXT:    s_sext_i32_i16 s5, s9
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s8, v2
+; GFX9-NEXT:    s_ashr_i32 s8, s5, 31
+; GFX9-NEXT:    s_add_i32 s5, s5, s8
+; GFX9-NEXT:    s_xor_b32 s5, s5, s8
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v3
+; GFX9-NEXT:    v_mul_hi_u32 v1, s5, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s7, v0
+; GFX9-NEXT:    v_xor_b32_e32 v2, s10, v2
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, s4
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s7, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v2, s10, v2
+; GFX9-NEXT:    v_sub_u32_e32 v3, s5, v3
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s4, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s4, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s4, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s4, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    s_xor_b32 s4, s8, s6
+; GFX9-NEXT:    v_xor_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_xor_b32_e32 v3, s8, v3
+; GFX9-NEXT:    v_sub_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0xffff
+; GFX9-NEXT:    v_sub_u32_sdwa v3, v3, s8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-NEXT:    v_and_or_b32 v0, v0, v4, v1
+; GFX9-NEXT:    v_and_or_b32 v1, v2, v4, v3
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_v2i16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x14
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX10-NEXT:    s_sext_i32_i16 s0, s0
+; GFX10-NEXT:    s_sext_i32_i16 s1, s1
+; GFX10-NEXT:    s_ashr_i32 s2, s0, 31
+; GFX10-NEXT:    s_ashr_i32 s3, s1, 31
+; GFX10-NEXT:    s_add_i32 s0, s0, s2
+; GFX10-NEXT:    s_add_i32 s1, s1, s3
+; GFX10-NEXT:    s_xor_b32 s8, s0, s2
+; GFX10-NEXT:    s_xor_b32 s1, s1, s3
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s1
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_sub_i32 s6, 0, s8
+; GFX10-NEXT:    s_sub_i32 s7, 0, s1
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, s6, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s7, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_sext_i32_i16 s6, s0
+; GFX10-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX10-NEXT:    s_ashr_i32 s9, s6, 31
+; GFX10-NEXT:    s_sext_i32_i16 s0, s0
+; GFX10-NEXT:    s_add_i32 s6, s6, s9
+; GFX10-NEXT:    s_ashr_i32 s10, s0, 31
+; GFX10-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX10-NEXT:    s_add_i32 s0, s0, s10
+; GFX10-NEXT:    s_xor_b32 s6, s6, s9
+; GFX10-NEXT:    s_xor_b32 s0, s0, s10
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v0, s6, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s0, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, v0, s8
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, s1
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v1
+; GFX10-NEXT:    v_sub_nc_u32_e32 v2, s6, v2
+; GFX10-NEXT:    v_sub_nc_u32_e32 v3, s0, v3
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s8, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s8, v2
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s1, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s1, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s8, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v1
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s1, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s1, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s8, v2
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s0
+; GFX10-NEXT:    s_xor_b32 s1, s9, s2
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s0
+; GFX10-NEXT:    s_xor_b32 s0, s10, s3
+; GFX10-NEXT:    v_xor_b32_e32 v0, s1, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s0, v1
+; GFX10-NEXT:    v_xor_b32_e32 v3, s10, v3
+; GFX10-NEXT:    v_xor_b32_e32 v2, s9, v2
+; GFX10-NEXT:    v_mov_b32_e32 v4, 0xffff
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s1, v0
+; GFX10-NEXT:    v_sub_nc_u32_sdwa v1, v1, s0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT:    v_sub_nc_u32_sdwa v3, v3, s10 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v2, s9, v2
+; GFX10-NEXT:    v_and_or_b32 v0, v0, v4, v1
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-NEXT:    v_and_or_b32 v2, v2, v4, v3
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dword v1, v0, s[4:5]
+; GFX10-NEXT:    global_store_dword v1, v2, s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv <2 x i16> %x, %y
+  store <2 x i16> %div, <2 x i16> addrspace(1)* %out0
+  %rem = srem <2 x i16> %x, %y
+  store <2 x i16> %rem, <2 x i16> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_i3(i3 addrspace(1)* %out0, i3 addrspace(1)* %out1, i3 %x, i3 %y) {
+; GFX8-LABEL: sdivrem_i3:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX8-NEXT:    s_bfe_i32 s1, s1, 0x30000
+; GFX8-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX8-NEXT:    s_add_i32 s1, s1, s6
+; GFX8-NEXT:    s_xor_b32 s7, s1, s6
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX8-NEXT:    s_sub_i32 s1, 0, s7
+; GFX8-NEXT:    s_bfe_i32 s0, s0, 0x30000
+; GFX8-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_add_i32 s0, s0, s8
+; GFX8-NEXT:    s_xor_b32 s9, s0, s8
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    s_xor_b32 s4, s8, s6
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s9, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s7
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s9, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_xor_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s4, v2
+; GFX8-NEXT:    v_xor_b32_e32 v3, s8, v3
+; GFX8-NEXT:    flat_store_byte v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s8, v3
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_byte v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_i3:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX9-NEXT:    s_bfe_i32 s1, s1, 0x30000
+; GFX9-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX9-NEXT:    s_add_i32 s1, s1, s6
+; GFX9-NEXT:    s_xor_b32 s7, s1, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_sub_i32 s1, 0, s7
+; GFX9-NEXT:    s_bfe_i32 s0, s0, 0x30000
+; GFX9-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_add_i32 s0, s0, s8
+; GFX9-NEXT:    s_xor_b32 s9, s0, s8
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_xor_b32 s4, s8, s6
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s9, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s4, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s8, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_i3:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX10-NEXT:    s_bfe_i32 s0, s0, 0x30000
+; GFX10-NEXT:    s_bfe_i32 s1, s1, 0x30000
+; GFX10-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX10-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX10-NEXT:    s_add_i32 s0, s0, s8
+; GFX10-NEXT:    s_add_i32 s1, s1, s6
+; GFX10-NEXT:    s_xor_b32 s0, s0, s8
+; GFX10-NEXT:    s_xor_b32 s7, s1, s6
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX10-NEXT:    s_sub_i32 s1, 0, s7
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    s_xor_b32 s4, s8, s6
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s4, v0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s8, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv i3 %x, %y
+  store i3 %div, i3 addrspace(1)* %out0
+  %rem = srem i3 %x, %y
+  store i3 %rem, i3 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @sdivrem_i27(i27 addrspace(1)* %out0, i27 addrspace(1)* %out1, i27 %x, i27 %y) {
+; GFX8-LABEL: sdivrem_i27:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_bfe_i32 s1, s1, 0x1b0000
+; GFX8-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX8-NEXT:    s_add_i32 s1, s1, s6
+; GFX8-NEXT:    s_xor_b32 s7, s1, s6
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX8-NEXT:    s_sub_i32 s1, 0, s7
+; GFX8-NEXT:    s_bfe_i32 s0, s0, 0x1b0000
+; GFX8-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_add_i32 s0, s0, s8
+; GFX8-NEXT:    s_xor_b32 s9, s0, s8
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    s_xor_b32 s4, s8, s6
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s9, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s7
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s9, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_xor_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e32 v2, vcc, s4, v2
+; GFX8-NEXT:    v_xor_b32_e32 v3, s8, v3
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_subrev_u32_e32 v3, vcc, s8, v3
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_dword v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: sdivrem_i27:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_bfe_i32 s1, s1, 0x1b0000
+; GFX9-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX9-NEXT:    s_add_i32 s1, s1, s6
+; GFX9-NEXT:    s_xor_b32 s7, s1, s6
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_sub_i32 s1, 0, s7
+; GFX9-NEXT:    s_bfe_i32 s0, s0, 0x1b0000
+; GFX9-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_add_i32 s0, s0, s8
+; GFX9-NEXT:    s_xor_b32 s9, s0, s8
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_xor_b32 s4, s8, s6
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s9, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v0, s4, v0
+; GFX9-NEXT:    v_subrev_u32_e32 v1, s8, v1
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sdivrem_i27:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_bfe_i32 s1, s1, 0x1b0000
+; GFX10-NEXT:    s_bfe_i32 s0, s0, 0x1b0000
+; GFX10-NEXT:    s_ashr_i32 s6, s1, 31
+; GFX10-NEXT:    s_ashr_i32 s8, s0, 31
+; GFX10-NEXT:    s_add_i32 s1, s1, s6
+; GFX10-NEXT:    s_add_i32 s0, s0, s8
+; GFX10-NEXT:    s_xor_b32 s7, s1, s6
+; GFX10-NEXT:    s_xor_b32 s0, s0, s8
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX10-NEXT:    s_sub_i32 s1, 0, s7
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    s_xor_b32 s4, s8, s6
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX10-NEXT:    v_xor_b32_e32 v1, s8, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v0, s4, v0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, s8, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = sdiv i27 %x, %y
+  store i27 %div, i27 addrspace(1)* %out0
+  %rem = srem i27 %x, %y
+  store i27 %rem, i27 addrspace(1)* %out1
+  ret void
+}

llvm/test/CodeGen/AMDGPU/GlobalISel/udivrem.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck --check-prefix=GFX8 %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -amdgpu-bypass-slow-div=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10 %s
+
+define amdgpu_kernel void @udivrem_i32(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 %x, i32 %y) {
+; GFX8-LABEL: udivrem_i32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX8-NEXT:    s_sub_i32 s0, 0, s7
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s0, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s6, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s7
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s6, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s7, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s7, v3
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_dword v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_i32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_sub_i32 s0, 0, s7
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s0, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s6, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s6, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s7, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_i32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX10-NEXT:    s_sub_i32 s0, 0, s7
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s0, v0
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s6, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s7
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s6, v1
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s7, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s7, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv i32 %x, %y
+  store i32 %div, i32 addrspace(1)* %out0
+  %rem = urem i32 %x, %y
+  store i32 %rem, i32 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_i64(i64 addrspace(1)* %out0, i64 addrspace(1)* %out1, i64 %x, i64 %y) {
+; GFX8-LABEL: udivrem_i64:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx8 s[4:11], s[4:5], 0x0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s11
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v1, s10
+; GFX8-NEXT:    s_sub_u32 s2, 0, s10
+; GFX8-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX8-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_and_b32 s0, s0, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX8-NEXT:    s_subb_u32 s3, 0, s11
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX8-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX8-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_mov_b32_e32 v6, s11
+; GFX8-NEXT:    v_mul_lo_u32 v3, s3, v0
+; GFX8-NEXT:    v_mul_lo_u32 v2, s2, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s2, v0
+; GFX8-NEXT:    v_mul_lo_u32 v4, s2, v0
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, v4
+; GFX8-NEXT:    v_mul_lo_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v7, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_mul_hi_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v7, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v7, v5
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v4, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v5, v4
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v4
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_addc_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v4, s3, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s2, v3
+; GFX8-NEXT:    v_mul_hi_u32 v8, s2, v0
+; GFX8-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GFX8-NEXT:    v_add_u32_e64 v1, s[0:1], v1, v2
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v5
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v8
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX8-NEXT:    v_mul_lo_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX8-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v8, v2
+; GFX8-NEXT:    v_mul_hi_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v7, s[0:1], v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v7, v5
+; GFX8-NEXT:    v_add_u32_e64 v3, s[0:1], v3, v4
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s8, v0
+; GFX8-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX8-NEXT:    v_mov_b32_e32 v4, s9
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v5, s9, v1
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_mul_hi_u32 v3, s8, v1
+; GFX8-NEXT:    v_mul_hi_u32 v1, s9, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v5, v0
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v2
+; GFX8-NEXT:    v_mul_lo_u32 v2, s11, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s10, v1
+; GFX8-NEXT:    v_mul_hi_u32 v7, s10, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s10, v0
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s8, v5
+; GFX8-NEXT:    v_subb_u32_e64 v4, s[0:1], v4, v2, vcc
+; GFX8-NEXT:    v_sub_u32_e64 v2, s[0:1], s9, v2
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v4
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, v5, v7, s[0:1]
+; GFX8-NEXT:    v_subrev_u32_e32 v7, vcc, s10, v3
+; GFX8-NEXT:    v_subbrev_u32_e64 v8, s[0:1], 0, v2, vcc
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], 1, v0
+; GFX8-NEXT:    v_addc_u32_e64 v10, s[0:1], 0, v1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v8
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v7
+; GFX8-NEXT:    v_subrev_u32_e32 v6, vcc, s10, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, v11, v12, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v12, s[0:1], 1, v9
+; GFX8-NEXT:    v_subbrev_u32_e32 v2, vcc, 0, v2, vcc
+; GFX8-NEXT:    v_addc_u32_e64 v13, s[0:1], 0, v10, s[0:1]
+; GFX8-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v11
+; GFX8-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v5
+; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v6, v8, v2, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, v3, v5, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, v4, v6, s[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v4, s4
+; GFX8-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v5, s5
+; GFX8-NEXT:    flat_store_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dwordx2 v[0:1], v[2:3]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_i64:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[4:11], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s11
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s10
+; GFX9-NEXT:    s_sub_u32 s2, 0, s10
+; GFX9-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_and_b32 s0, s0, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX9-NEXT:    s_subb_u32 s3, 0, s11
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_lo_u32 v3, s3, v0
+; GFX9-NEXT:    v_mul_lo_u32 v2, s2, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s2, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s2, v0
+; GFX9-NEXT:    v_add3_u32 v2, v3, v2, v4
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, v5
+; GFX9-NEXT:    v_mul_lo_u32 v6, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v4, v0, v5
+; GFX9-NEXT:    v_mul_hi_u32 v5, v1, v5
+; GFX9-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX9-NEXT:    v_mul_hi_u32 v8, v0, v2
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v7, v5
+; GFX9-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v4, v8
+; GFX9-NEXT:    v_add_u32_e32 v3, v6, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v4, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_u32_e32 v5, v5, v6
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_add3_u32 v2, v5, v4, v2
+; GFX9-NEXT:    v_addc_co_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v4, s3, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s2, v3
+; GFX9-NEXT:    v_mul_hi_u32 v6, s2, v0
+; GFX9-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v2
+; GFX9-NEXT:    v_mov_b32_e32 v8, s9
+; GFX9-NEXT:    v_add3_u32 v4, v4, v5, v6
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX9-NEXT:    v_mul_lo_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX9-NEXT:    v_add_u32_e32 v2, v6, v2
+; GFX9-NEXT:    v_mul_hi_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_add_u32_e32 v6, v7, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add3_u32 v3, v6, v4, v3
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX9-NEXT:    v_mul_hi_u32 v5, s8, v0
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    v_mov_b32_e32 v4, s11
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v5
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v5, s9, v1
+; GFX9-NEXT:    v_add_u32_e32 v2, v3, v2
+; GFX9-NEXT:    v_mul_hi_u32 v3, s8, v1
+; GFX9-NEXT:    v_mul_hi_u32 v1, s9, v1
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v5, v0
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v5, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, s11, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s10, v1
+; GFX9-NEXT:    v_mul_hi_u32 v5, s10, v0
+; GFX9-NEXT:    v_mul_lo_u32 v7, s10, v0
+; GFX9-NEXT:    v_mov_b32_e32 v6, 0
+; GFX9-NEXT:    v_add3_u32 v2, v2, v3, v5
+; GFX9-NEXT:    v_sub_co_u32_e32 v3, vcc, s8, v7
+; GFX9-NEXT:    v_subb_co_u32_e64 v5, s[0:1], v8, v2, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v5
+; GFX9-NEXT:    v_sub_u32_e32 v2, s9, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v5
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v4, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s[0:1]
+; GFX9-NEXT:    v_subrev_co_u32_e32 v8, vcc, s10, v3
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v9, s[0:1], 0, v2, vcc
+; GFX9-NEXT:    v_add_co_u32_e64 v10, s[0:1], 1, v0
+; GFX9-NEXT:    v_addc_co_u32_e64 v11, s[0:1], 0, v1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v9
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v4, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v8
+; GFX9-NEXT:    v_subrev_co_u32_e32 v4, vcc, s10, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v9
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, v12, v13, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v13, s[0:1], 1, v10
+; GFX9-NEXT:    v_subbrev_co_u32_e32 v2, vcc, 0, v2, vcc
+; GFX9-NEXT:    v_addc_co_u32_e64 v14, s[0:1], 0, v11, s[0:1]
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v7
+; GFX9-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v0, v0, v10, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v1, v1, v11, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v8, v4, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v9, v2, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, v3, v4, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, v5, v7, s[0:1]
+; GFX9-NEXT:    global_store_dwordx2 v6, v[0:1], s[4:5]
+; GFX9-NEXT:    global_store_dwordx2 v6, v[2:3], s[6:7]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx8 s[4:11], s[4:5], 0x0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s11
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s10
+; GFX10-NEXT:    s_sub_u32 s1, 0, s10
+; GFX10-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX10-NEXT:    s_and_b32 s0, s0, 1
+; GFX10-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX10-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX10-NEXT:    s_subb_u32 s2, 0, s11
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX10-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX10-NEXT:    v_mul_lo_u32 v2, s1, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s2, v0
+; GFX10-NEXT:    v_mul_hi_u32 v4, s1, v0
+; GFX10-NEXT:    v_mul_lo_u32 v5, s1, v0
+; GFX10-NEXT:    v_add3_u32 v2, v3, v2, v4
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, v5
+; GFX10-NEXT:    v_mul_hi_u32 v6, v1, v5
+; GFX10-NEXT:    v_mul_hi_u32 v5, v0, v5
+; GFX10-NEXT:    v_mul_lo_u32 v4, v0, v2
+; GFX10-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX10-NEXT:    v_mul_hi_u32 v8, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX10-NEXT:    v_add_co_u32 v3, s0, v3, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v7, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v3, s0, v3, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v6, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v3, v4, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, v7, v6
+; GFX10-NEXT:    v_add_co_u32 v3, s0, v5, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v3
+; GFX10-NEXT:    v_add3_u32 v2, v4, v5, v2
+; GFX10-NEXT:    v_mul_lo_u32 v4, s2, v0
+; GFX10-NEXT:    v_mul_hi_u32 v5, s1, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s0, v1, v2, vcc_lo
+; GFX10-NEXT:    v_mul_lo_u32 v7, s1, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v2
+; GFX10-NEXT:    v_mul_lo_u32 v6, s1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v8, v3, v7
+; GFX10-NEXT:    v_add3_u32 v4, v4, v6, v5
+; GFX10-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX10-NEXT:    v_mul_hi_u32 v7, v0, v7
+; GFX10-NEXT:    v_mul_lo_u32 v6, v0, v4
+; GFX10-NEXT:    v_mul_lo_u32 v9, v3, v4
+; GFX10-NEXT:    v_mul_hi_u32 v10, v0, v4
+; GFX10-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v8, s0, v9, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v7, s0, v8, v10
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, v6, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, v9, v8
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v7, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add3_u32 v2, v4, v6, v3
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, v1, v2, vcc_lo
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v5
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX10-NEXT:    v_mul_hi_u32 v4, s9, v0
+; GFX10-NEXT:    v_mul_hi_u32 v0, s8, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX10-NEXT:    v_mul_lo_u32 v5, s9, v1
+; GFX10-NEXT:    v_mul_hi_u32 v6, s8, v1
+; GFX10-NEXT:    v_mul_hi_u32 v1, s9, v1
+; GFX10-NEXT:    v_add_co_u32 v2, s0, v2, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v4, s0, v5, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v0, s0, v2, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v2, s0, v4, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v3, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v3, v5, v4
+; GFX10-NEXT:    v_add_co_u32 v0, s0, v2, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; GFX10-NEXT:    v_mul_lo_u32 v5, s10, v0
+; GFX10-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, s11, v0
+; GFX10-NEXT:    v_mul_hi_u32 v3, s10, v0
+; GFX10-NEXT:    v_mul_lo_u32 v4, s10, v1
+; GFX10-NEXT:    v_add3_u32 v2, v2, v4, v3
+; GFX10-NEXT:    v_add_co_u32 v3, vcc_lo, v0, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v5, vcc_lo, s8, v5
+; GFX10-NEXT:    v_sub_nc_u32_e32 v6, s9, v2
+; GFX10-NEXT:    v_sub_co_ci_u32_e64 v7, s0, s9, v2, vcc_lo
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v2, vcc_lo, s11, v6, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s10, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v8, vcc_lo, v5, s10
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v9, s0, 0, v2, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s11, v7
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v2, vcc_lo, s11, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s10, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s11, v9
+; GFX10-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s0
+; GFX10-NEXT:    v_add_co_u32 v13, s0, v3, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v14, s0, 0, v4, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s11, v9
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, v12, v11, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s11, v7
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v11
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v10, v6, s0
+; GFX10-NEXT:    v_sub_co_u32 v10, s0, v8, s10
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v2, s0, 0, v2, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v13, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, v6
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v14, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v9, v2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v9, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v3, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v5, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v7, v8, s0
+; GFX10-NEXT:    global_store_dwordx2 v9, v[0:1], s[4:5]
+; GFX10-NEXT:    global_store_dwordx2 v9, v[2:3], s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv i64 %x, %y
+  store i64 %div, i64 addrspace(1)* %out0
+  %rem = urem i64 %x, %y
+  store i64 %rem, i64 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_v2i32(<2 x i32> addrspace(1)* %out0, <2 x i32> addrspace(1)* %out1, <2 x i32> %x, <2 x i32> %y) {
+; GFX8-LABEL: udivrem_v2i32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x18
+; GFX8-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x0
+; GFX8-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v1, s3
+; GFX8-NEXT:    s_sub_i32 s0, 0, s2
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_lo_u32 v2, s0, v0
+; GFX8-NEXT:    s_sub_i32 s0, 0, s3
+; GFX8-NEXT:    v_mul_lo_u32 v3, s0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v0, s4, v0
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v3
+; GFX8-NEXT:    v_mul_hi_u32 v1, s5, v1
+; GFX8-NEXT:    v_mul_lo_u32 v2, v0, s2
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_mul_lo_u32 v4, v1, s3
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s4, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s2, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s2, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s5, v4
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v4, s8
+; GFX8-NEXT:    v_mov_b32_e32 v5, s9
+; GFX8-NEXT:    flat_store_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v0, s10
+; GFX8-NEXT:    v_mov_b32_e32 v1, s11
+; GFX8-NEXT:    flat_store_dwordx2 v[0:1], v[2:3]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_v2i32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x18
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s3
+; GFX9-NEXT:    s_sub_i32 s0, 0, s2
+; GFX9-NEXT:    s_sub_i32 s1, 0, s3
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, s0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s1, v1
+; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v2
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v3
+; GFX9-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, v0, s2
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, s3
+; GFX9-NEXT:    v_add_u32_e32 v5, 1, v1
+; GFX9-NEXT:    v_sub_u32_e32 v2, s0, v2
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GFX9-NEXT:    v_sub_u32_e32 v3, s1, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s2, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s3, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v1, v1, v5, s[0:1]
+; GFX9-NEXT:    v_subrev_u32_e32 v5, s3, v3
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s2, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s2, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, v3, v5, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s3, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
+; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[4:5]
+; GFX9-NEXT:    global_store_dwordx2 v4, v[2:3], s[6:7]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_v2i32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x18
+; GFX10-NEXT:    v_mov_b32_e32 v8, 0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s3
+; GFX10-NEXT:    s_sub_i32 s0, 0, s2
+; GFX10-NEXT:    s_sub_i32 s1, 0, s3
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, s1, v1
+; GFX10-NEXT:    s_clause 0x1
+; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX10-NEXT:    s_nop 0
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, v0, s2
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, s3
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v1
+; GFX10-NEXT:    v_sub_nc_u32_e32 v2, s0, v2
+; GFX10-NEXT:    v_sub_nc_u32_e32 v3, s1, v3
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s2, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s2, v2
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s3, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s3, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s2, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v1
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s3, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v6, s2, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s3, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX10-NEXT:    global_store_dwordx2 v8, v[0:1], s[4:5]
+; GFX10-NEXT:    global_store_dwordx2 v8, v[2:3], s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv <2 x i32> %x, %y
+  store <2 x i32> %div, <2 x i32> addrspace(1)* %out0
+  %rem = urem <2 x i32> %x, %y
+  store <2 x i32> %rem, <2 x i32> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_v4i32(<4 x i32> addrspace(1)* %out0, <4 x i32> addrspace(1)* %out1, <4 x i32> %x, <4 x i32> %y) {
+; GFX8-LABEL: udivrem_v4i32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX8-NEXT:    v_mov_b32_e32 v2, 0x4f7ffffe
+; GFX8-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x0
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v1, s9
+; GFX8-NEXT:    s_sub_i32 s0, 0, s8
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v6, s10
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, v1, v2
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_lo_u32 v3, s0, v0
+; GFX8-NEXT:    s_sub_i32 s0, 0, s9
+; GFX8-NEXT:    v_mul_lo_u32 v4, s0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v3, v0, v3
+; GFX8-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_mul_hi_u32 v0, s4, v0
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v4
+; GFX8-NEXT:    v_mul_hi_u32 v1, s5, v1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v0, s8
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, v1, s9
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s4, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s8, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s8, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s8, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s8, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v4, v3, v4, vcc
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s5, v5
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v5, v6
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s9, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; GFX8-NEXT:    v_mul_f32_e32 v5, v5, v2
+; GFX8-NEXT:    v_subrev_u32_e64 v6, s[0:1], s9, v3
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v6, vcc
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s9, v3
+; GFX8-NEXT:    s_sub_i32 s0, 0, s10
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v6, s11
+; GFX8-NEXT:    v_mul_lo_u32 v7, s0, v5
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v6, v6
+; GFX8-NEXT:    v_mul_hi_u32 v7, v5, v7
+; GFX8-NEXT:    v_mul_f32_e32 v2, v6, v2
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v7
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX8-NEXT:    v_mul_hi_u32 v7, s6, v5
+; GFX8-NEXT:    v_subrev_u32_e64 v5, s[0:1], s9, v3
+; GFX8-NEXT:    s_sub_i32 s0, 0, s11
+; GFX8-NEXT:    v_mul_lo_u32 v6, s0, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v5, v3, v5, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v3, v7, s10
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 1, v7
+; GFX8-NEXT:    v_mul_hi_u32 v6, v2, v6
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s6, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s10, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v8, s[0:1], s10, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; GFX8-NEXT:    v_mul_hi_u32 v8, s7, v2
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, 1, v7
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s10, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v7, v2, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v7, v8, s11
+; GFX8-NEXT:    v_subrev_u32_e64 v6, s[0:1], s10, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v6, v3, v6, vcc
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s7, v7
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, 1, v8
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v7, v8, v7, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v8, s[0:1], s11, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v8, v3, v8, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v7
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s11, v8
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v7, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v7, s[0:1], s11, v8
+; GFX8-NEXT:    v_cndmask_b32_e32 v7, v8, v7, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v8, s12
+; GFX8-NEXT:    v_mov_b32_e32 v9, s13
+; GFX8-NEXT:    flat_store_dwordx4 v[8:9], v[0:3]
+; GFX8-NEXT:    s_nop 0
+; GFX8-NEXT:    v_mov_b32_e32 v0, s14
+; GFX8-NEXT:    v_mov_b32_e32 v1, s15
+; GFX8-NEXT:    flat_store_dwordx4 v[0:1], v[4:7]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_v4i32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x20
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0x4f7ffffe
+; GFX9-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s0
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s1
+; GFX9-NEXT:    s_sub_i32 s6, 0, s0
+; GFX9-NEXT:    s_sub_i32 s7, 0, s1
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v5, s2
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, v1, v2
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v5, v5
+; GFX9-NEXT:    v_mul_lo_u32 v3, s6, v0
+; GFX9-NEXT:    v_mul_lo_u32 v4, s7, v1
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_hi_u32 v3, v0, v3
+; GFX9-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v3
+; GFX9-NEXT:    v_mul_hi_u32 v0, s8, v0
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v4
+; GFX9-NEXT:    v_mul_hi_u32 v1, s9, v1
+; GFX9-NEXT:    v_mul_f32_e32 v3, v5, v2
+; GFX9-NEXT:    v_mul_lo_u32 v4, v0, s0
+; GFX9-NEXT:    v_add_u32_e32 v6, 1, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, v1, s1
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GFX9-NEXT:    v_sub_u32_e32 v4, s8, v4
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s0, v4
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v6, s0, v4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v4, v6, vcc
+; GFX9-NEXT:    v_add_u32_e32 v6, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s0, v4
+; GFX9-NEXT:    v_sub_u32_e32 v5, s9, v5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v6, s0, v4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v4, v6, vcc
+; GFX9-NEXT:    v_add_u32_e32 v6, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s1, v5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v6, s3
+; GFX9-NEXT:    s_sub_i32 s0, 0, s2
+; GFX9-NEXT:    v_mul_lo_u32 v7, s0, v3
+; GFX9-NEXT:    s_sub_i32 s0, 0, s3
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v6, v6
+; GFX9-NEXT:    v_subrev_u32_e32 v8, s1, v5
+; GFX9-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX9-NEXT:    v_cndmask_b32_e32 v5, v5, v8, vcc
+; GFX9-NEXT:    v_mul_f32_e32 v2, v6, v2
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v3, v7
+; GFX9-NEXT:    v_mul_hi_u32 v3, s10, v3
+; GFX9-NEXT:    v_add_u32_e32 v8, 1, v1
+; GFX9-NEXT:    v_mul_lo_u32 v7, s0, v2
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s1, v5
+; GFX9-NEXT:    v_mul_lo_u32 v6, v3, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc
+; GFX9-NEXT:    v_mul_hi_u32 v7, v2, v7
+; GFX9-NEXT:    v_subrev_u32_e32 v8, s1, v5
+; GFX9-NEXT:    v_sub_u32_e32 v6, s10, v6
+; GFX9-NEXT:    v_cndmask_b32_e32 v5, v5, v8, vcc
+; GFX9-NEXT:    v_add_u32_e32 v2, v2, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, s11, v2
+; GFX9-NEXT:    v_add_u32_e32 v8, 1, v3
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s2, v6
+; GFX9-NEXT:    v_subrev_u32_e32 v2, s2, v6
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v8, v7, s3
+; GFX9-NEXT:    v_cndmask_b32_e32 v6, v6, v2, vcc
+; GFX9-NEXT:    v_add_u32_e32 v2, 1, v3
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s2, v6
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s2, v6
+; GFX9-NEXT:    v_cndmask_b32_e32 v6, v6, v3, vcc
+; GFX9-NEXT:    v_sub_u32_e32 v3, s11, v8
+; GFX9-NEXT:    v_add_u32_e32 v8, 1, v7
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v8, s3, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v3, v8, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v7
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s3, v8
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v7, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v7, s3, v8
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v8, v7, vcc
+; GFX9-NEXT:    v_mov_b32_e32 v8, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dwordx4 v8, v[0:3], s[4:5]
+; GFX9-NEXT:    global_store_dwordx4 v8, v[4:7], s[6:7]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_v4i32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX10-NEXT:    v_mov_b32_e32 v4, 0x4f7ffffe
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x10
+; GFX10-NEXT:    v_mov_b32_e32 v8, 0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s8
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s9
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v2, s10
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v3, s11
+; GFX10-NEXT:    s_sub_i32 s6, 0, s8
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v3, v3
+; GFX10-NEXT:    s_sub_i32 s7, 0, s9
+; GFX10-NEXT:    s_sub_i32 s12, 0, s10
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, v1, v4
+; GFX10-NEXT:    v_mul_f32_e32 v2, v2, v4
+; GFX10-NEXT:    v_mul_f32_e32 v3, v3, v4
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GFX10-NEXT:    v_mul_lo_u32 v4, s6, v0
+; GFX10-NEXT:    v_mul_lo_u32 v5, s7, v1
+; GFX10-NEXT:    v_mul_lo_u32 v6, s12, v2
+; GFX10-NEXT:    s_sub_i32 s6, 0, s11
+; GFX10-NEXT:    v_mul_lo_u32 v7, s6, v3
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    v_mul_hi_u32 v4, v0, v4
+; GFX10-NEXT:    v_mul_hi_u32 v5, v1, v5
+; GFX10-NEXT:    v_mul_hi_u32 v6, v2, v6
+; GFX10-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v4
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, v2, v6
+; GFX10-NEXT:    v_add_nc_u32_e32 v3, v3, v7
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s1, v1
+; GFX10-NEXT:    v_mul_hi_u32 v2, s2, v2
+; GFX10-NEXT:    v_mul_hi_u32 v3, s3, v3
+; GFX10-NEXT:    v_mul_lo_u32 v4, v0, s8
+; GFX10-NEXT:    v_mul_lo_u32 v5, v1, s9
+; GFX10-NEXT:    v_mul_lo_u32 v6, v2, s10
+; GFX10-NEXT:    v_add_nc_u32_e32 v9, 1, v0
+; GFX10-NEXT:    v_mul_lo_u32 v7, v3, s11
+; GFX10-NEXT:    v_add_nc_u32_e32 v10, 1, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v11, 1, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v12, 1, v3
+; GFX10-NEXT:    v_sub_nc_u32_e32 v4, s0, v4
+; GFX10-NEXT:    v_sub_nc_u32_e32 v5, s1, v5
+; GFX10-NEXT:    v_sub_nc_u32_e32 v6, s2, v6
+; GFX10-NEXT:    v_sub_nc_u32_e32 v7, s3, v7
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s8, v4
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s9, v5
+; GFX10-NEXT:    v_cmp_le_u32_e64 s1, s10, v6
+; GFX10-NEXT:    v_cmp_le_u32_e64 s2, s11, v7
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v9, vcc_lo
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v9, s8, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v10, s9, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v11, s1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v11, s10, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v12, s2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v12, s11, v7
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v9, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, v5, v10, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v6, v11, s1
+; GFX10-NEXT:    v_add_nc_u32_e32 v9, 1, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v7, v12, s2
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s8, v4
+; GFX10-NEXT:    v_add_nc_u32_e32 v10, 1, v1
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s9, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v11, 1, v2
+; GFX10-NEXT:    v_cmp_le_u32_e64 s1, s10, v6
+; GFX10-NEXT:    v_add_nc_u32_e32 v12, 1, v3
+; GFX10-NEXT:    v_cmp_le_u32_e64 s2, s11, v7
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v13, s11, v7
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v9, vcc_lo
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v9, s8, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v10, s9, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, v11, s1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v11, s10, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v12, s2
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v9, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, v5, v10, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v7, v13, s2
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v6, v11, s1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dwordx4 v8, v[0:3], s[4:5]
+; GFX10-NEXT:    global_store_dwordx4 v8, v[4:7], s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv <4 x i32> %x, %y
+  store <4 x i32> %div, <4 x i32> addrspace(1)* %out0
+  %rem = urem <4 x i32> %x, %y
+  store <4 x i32> %rem, <4 x i32> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_v2i64(<2 x i64> addrspace(1)* %out0, <2 x i64> addrspace(1)* %out1, <2 x i64> %x, <2 x i64> %y) {
+; GFX8-LABEL: udivrem_v2i64:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX8-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x10
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX8-NEXT:    s_sub_u32 s2, 0, s8
+; GFX8-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX8-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_and_b32 s0, s0, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX8-NEXT:    s_subb_u32 s3, 0, s9
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX8-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX8-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_mov_b32_e32 v6, s9
+; GFX8-NEXT:    v_mul_lo_u32 v3, s3, v0
+; GFX8-NEXT:    v_mul_lo_u32 v2, s2, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s2, v0
+; GFX8-NEXT:    v_mul_lo_u32 v4, s2, v0
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, v4
+; GFX8-NEXT:    v_mul_lo_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v7, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v4, v1, v4
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_mul_hi_u32 v5, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v7, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v7, v5
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v4, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v5, v4
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v4
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_addc_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v4, s3, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s2, v3
+; GFX8-NEXT:    v_mul_hi_u32 v8, s2, v0
+; GFX8-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GFX8-NEXT:    v_add_u32_e64 v1, s[0:1], v1, v2
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v5
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v4, v8
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX8-NEXT:    v_mul_lo_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX8-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX8-NEXT:    s_sub_u32 s2, 0, s10
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX8-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v8, v2
+; GFX8-NEXT:    v_mul_hi_u32 v8, v0, v4
+; GFX8-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v5, s[0:1], v5, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v7, s[0:1], v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v2, s[0:1], v5, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v4, s[0:1], v7, v5
+; GFX8-NEXT:    v_add_u32_e64 v3, s[0:1], v3, v4
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v2, s13, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s12, v1
+; GFX8-NEXT:    v_mul_hi_u32 v5, s12, v0
+; GFX8-NEXT:    v_mul_hi_u32 v0, s13, v0
+; GFX8-NEXT:    v_mov_b32_e32 v4, s13
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v5, s13, v1
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_mul_hi_u32 v3, s12, v1
+; GFX8-NEXT:    v_mul_hi_u32 v1, s13, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v5, v0
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v5, v3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v3, v2
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v2
+; GFX8-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX8-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX8-NEXT:    v_mul_hi_u32 v7, s8, v0
+; GFX8-NEXT:    v_mul_lo_u32 v5, s8, v0
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v3
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s12, v5
+; GFX8-NEXT:    v_subb_u32_e64 v5, s[0:1], v4, v2, vcc
+; GFX8-NEXT:    v_sub_u32_e64 v2, s[0:1], s13, v2
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v5
+; GFX8-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v3
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v5
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v4, v4, v7, s[0:1]
+; GFX8-NEXT:    v_subrev_u32_e32 v7, vcc, s8, v3
+; GFX8-NEXT:    v_subbrev_u32_e64 v8, s[0:1], 0, v2, vcc
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], 1, v0
+; GFX8-NEXT:    v_addc_u32_e64 v10, s[0:1], 0, v1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v8
+; GFX8-NEXT:    v_subb_u32_e32 v2, vcc, v2, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v7
+; GFX8-NEXT:    v_subrev_u32_e32 v6, vcc, s8, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, v11, v12, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v12, s[0:1], 1, v9
+; GFX8-NEXT:    v_subbrev_u32_e32 v2, vcc, 0, v2, vcc
+; GFX8-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v11
+; GFX8-NEXT:    v_addc_u32_e64 v13, s[0:1], 0, v10, s[0:1]
+; GFX8-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v4
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v4, s11
+; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[0:1]
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v9, s10
+; GFX8-NEXT:    v_mul_f32_e32 v4, 0x4f800000, v4
+; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v8, v2, vcc
+; GFX8-NEXT:    v_add_f32_e32 v4, v4, v9
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v7, v4
+; GFX8-NEXT:    v_cndmask_b32_e64 v5, v5, v2, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v4, v3, v6, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX8-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v7
+; GFX8-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; GFX8-NEXT:    v_trunc_f32_e32 v3, v3
+; GFX8-NEXT:    v_mul_f32_e32 v6, 0xcf800000, v3
+; GFX8-NEXT:    v_add_f32_e32 v2, v6, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s[0:1]
+; GFX8-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GFX8-NEXT:    s_and_b32 s0, s0, 1
+; GFX8-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX8-NEXT:    s_subb_u32 s3, 0, s11
+; GFX8-NEXT:    v_mul_lo_u32 v6, s3, v2
+; GFX8-NEXT:    v_mul_lo_u32 v7, s2, v3
+; GFX8-NEXT:    v_mul_hi_u32 v9, s2, v2
+; GFX8-NEXT:    v_mul_lo_u32 v8, s2, v2
+; GFX8-NEXT:    v_mov_b32_e32 v10, s11
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v7
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v9
+; GFX8-NEXT:    v_mul_lo_u32 v7, v3, v8
+; GFX8-NEXT:    v_mul_lo_u32 v9, v2, v6
+; GFX8-NEXT:    v_mul_hi_u32 v11, v2, v8
+; GFX8-NEXT:    v_mul_hi_u32 v8, v3, v8
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v9
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v11
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v11, v3, v6
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v9, v7
+; GFX8-NEXT:    v_mul_hi_u32 v9, v2, v6
+; GFX8-NEXT:    v_mul_hi_u32 v6, v3, v6
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v11, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v9
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v9, vcc, v11, v9
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v8, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v9, v8
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_addc_u32_e64 v7, s[0:1], v3, v6, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v8, s3, v2
+; GFX8-NEXT:    v_mul_lo_u32 v9, s2, v7
+; GFX8-NEXT:    v_mul_hi_u32 v12, s2, v2
+; GFX8-NEXT:    v_mul_lo_u32 v11, s2, v2
+; GFX8-NEXT:    v_add_u32_e64 v3, s[0:1], v3, v6
+; GFX8-NEXT:    v_add_u32_e64 v8, s[0:1], v8, v9
+; GFX8-NEXT:    v_add_u32_e64 v8, s[0:1], v8, v12
+; GFX8-NEXT:    v_mul_lo_u32 v9, v7, v11
+; GFX8-NEXT:    v_mul_lo_u32 v12, v2, v8
+; GFX8-NEXT:    v_mul_hi_u32 v6, v2, v11
+; GFX8-NEXT:    v_mul_hi_u32 v11, v7, v11
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], v9, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v6, s[0:1], v9, v6
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX8-NEXT:    v_mul_lo_u32 v9, v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v6, s[0:1], v12, v6
+; GFX8-NEXT:    v_mul_hi_u32 v12, v2, v8
+; GFX8-NEXT:    v_mul_hi_u32 v7, v7, v8
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], v9, v11
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v9, s[0:1], v9, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v11, s[0:1], v11, v12
+; GFX8-NEXT:    v_add_u32_e64 v6, s[0:1], v9, v6
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v8, s[0:1], v11, v9
+; GFX8-NEXT:    v_add_u32_e64 v7, s[0:1], v7, v8
+; GFX8-NEXT:    v_addc_u32_e32 v3, vcc, v3, v7, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; GFX8-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v6, s15, v2
+; GFX8-NEXT:    v_mul_lo_u32 v7, s14, v3
+; GFX8-NEXT:    v_mul_hi_u32 v9, s14, v2
+; GFX8-NEXT:    v_mul_hi_u32 v2, s15, v2
+; GFX8-NEXT:    v_mov_b32_e32 v8, s15
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v9
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v9, s15, v3
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v7, v6
+; GFX8-NEXT:    v_mul_hi_u32 v7, s14, v3
+; GFX8-NEXT:    v_mul_hi_u32 v3, s15, v3
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v9, v2
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v9, v7
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v7, v6
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v6
+; GFX8-NEXT:    v_mul_lo_u32 v6, s11, v2
+; GFX8-NEXT:    v_mul_lo_u32 v7, s10, v3
+; GFX8-NEXT:    v_mul_hi_u32 v11, s10, v2
+; GFX8-NEXT:    v_mul_lo_u32 v9, s10, v2
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v7
+; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v11
+; GFX8-NEXT:    v_sub_u32_e32 v7, vcc, s14, v9
+; GFX8-NEXT:    v_subb_u32_e64 v8, s[0:1], v8, v6, vcc
+; GFX8-NEXT:    v_sub_u32_e64 v6, s[0:1], s15, v6
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v8
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v7
+; GFX8-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v8
+; GFX8-NEXT:    v_subb_u32_e32 v6, vcc, v6, v10, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v9, v9, v11, s[0:1]
+; GFX8-NEXT:    v_subrev_u32_e32 v11, vcc, s10, v7
+; GFX8-NEXT:    v_subbrev_u32_e64 v12, s[0:1], 0, v6, vcc
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v11
+; GFX8-NEXT:    v_cndmask_b32_e64 v14, 0, -1, s[0:1]
+; GFX8-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v12
+; GFX8-NEXT:    v_cndmask_b32_e64 v13, v13, v14, s[0:1]
+; GFX8-NEXT:    v_add_u32_e64 v14, s[0:1], 1, v2
+; GFX8-NEXT:    v_subb_u32_e32 v6, vcc, v6, v10, vcc
+; GFX8-NEXT:    v_addc_u32_e64 v15, s[0:1], 0, v3, s[0:1]
+; GFX8-NEXT:    v_add_u32_e32 v10, vcc, 1, v14
+; GFX8-NEXT:    v_addc_u32_e32 v16, vcc, 0, v15, vcc
+; GFX8-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v13
+; GFX8-NEXT:    v_subrev_u32_e64 v13, s[0:1], s10, v11
+; GFX8-NEXT:    v_subbrev_u32_e64 v6, s[0:1], 0, v6, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e32 v10, v14, v10, vcc
+; GFX8-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v9
+; GFX8-NEXT:    v_cndmask_b32_e64 v2, v2, v10, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v10, v12, v6, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v14, v15, v16, vcc
+; GFX8-NEXT:    v_cndmask_b32_e64 v6, v7, v9, s[0:1]
+; GFX8-NEXT:    v_cndmask_b32_e64 v7, v8, v10, s[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v9, s5
+; GFX8-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v8, s4
+; GFX8-NEXT:    flat_store_dwordx4 v[8:9], v[0:3]
+; GFX8-NEXT:    s_nop 0
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dwordx4 v[0:1], v[4:7]
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_v2i64:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX9-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX9-NEXT:    s_sub_u32 s2, 0, s8
+; GFX9-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_and_b32 s0, s0, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX9-NEXT:    s_subb_u32 s3, 0, s9
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_f32_e32 v2, 0xcf800000, v1
+; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v14, s11
+; GFX9-NEXT:    v_mul_lo_u32 v3, s3, v0
+; GFX9-NEXT:    v_mul_lo_u32 v2, s2, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s2, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s2, v0
+; GFX9-NEXT:    v_mul_f32_e32 v14, 0x4f800000, v14
+; GFX9-NEXT:    v_add3_u32 v2, v3, v2, v4
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, v5
+; GFX9-NEXT:    v_mul_lo_u32 v6, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v4, v0, v5
+; GFX9-NEXT:    v_mul_hi_u32 v5, v1, v5
+; GFX9-NEXT:    v_mul_lo_u32 v7, v1, v2
+; GFX9-NEXT:    v_mul_hi_u32 v8, v0, v2
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v7, v5
+; GFX9-NEXT:    v_mul_hi_u32 v2, v1, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, v6, v3
+; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, v4, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v4, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_u32_e32 v5, v5, v6
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_add3_u32 v2, v5, v4, v2
+; GFX9-NEXT:    v_addc_co_u32_e64 v3, s[0:1], v1, v2, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v4, s3, v0
+; GFX9-NEXT:    v_mul_lo_u32 v5, s2, v3
+; GFX9-NEXT:    v_mul_hi_u32 v6, s2, v0
+; GFX9-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v2
+; GFX9-NEXT:    v_add3_u32 v4, v4, v5, v6
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v7
+; GFX9-NEXT:    v_mul_lo_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, v3, v7
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[0:1]
+; GFX9-NEXT:    v_mul_lo_u32 v5, v3, v4
+; GFX9-NEXT:    v_add_u32_e32 v2, v6, v2
+; GFX9-NEXT:    v_mul_hi_u32 v6, v0, v4
+; GFX9-NEXT:    v_mul_hi_u32 v3, v3, v4
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[0:1], v5, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], v5, v2
+; GFX9-NEXT:    v_add_u32_e32 v6, v7, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s[0:1]
+; GFX9-NEXT:    v_add3_u32 v3, v6, v4, v3
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v2, s13, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s12, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s12, v0
+; GFX9-NEXT:    v_mul_hi_u32 v0, s13, v0
+; GFX9-NEXT:    v_mov_b32_e32 v7, s13
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v4, s13, v1
+; GFX9-NEXT:    v_add_u32_e32 v2, v3, v2
+; GFX9-NEXT:    v_mul_hi_u32 v3, s12, v1
+; GFX9-NEXT:    v_mul_hi_u32 v1, s13, v1
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v4, v0
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v4, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, s9, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, s8, v1
+; GFX9-NEXT:    v_mul_hi_u32 v4, s8, v0
+; GFX9-NEXT:    v_mul_lo_u32 v6, s8, v0
+; GFX9-NEXT:    v_mov_b32_e32 v5, s9
+; GFX9-NEXT:    v_add3_u32 v2, v2, v3, v4
+; GFX9-NEXT:    v_sub_co_u32_e32 v3, vcc, s12, v6
+; GFX9-NEXT:    v_subb_co_u32_e64 v6, s[0:1], v7, v2, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v6
+; GFX9-NEXT:    v_sub_u32_e32 v2, s13, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v6
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v5, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, v4, v7, s[0:1]
+; GFX9-NEXT:    v_subrev_co_u32_e32 v7, vcc, s8, v3
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v8, s[0:1], 0, v2, vcc
+; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v5, vcc
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v5, s10
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[0:1], 1, v0
+; GFX9-NEXT:    v_addc_co_u32_e64 v10, s[0:1], 0, v1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v8
+; GFX9-NEXT:    v_add_f32_e32 v5, v14, v5
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v5, v5
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v7
+; GFX9-NEXT:    v_subrev_co_u32_e32 v15, vcc, s8, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, v11, v12, s[0:1]
+; GFX9-NEXT:    v_subbrev_co_u32_e32 v2, vcc, 0, v2, vcc
+; GFX9-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v11
+; GFX9-NEXT:    v_mul_f32_e32 v11, 0x2f800000, v5
+; GFX9-NEXT:    v_add_co_u32_e64 v12, s[0:1], 1, v9
+; GFX9-NEXT:    v_trunc_f32_e32 v11, v11
+; GFX9-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc
+; GFX9-NEXT:    v_mul_f32_e32 v12, 0xcf800000, v11
+; GFX9-NEXT:    v_addc_co_u32_e64 v13, s[0:1], 0, v10, s[0:1]
+; GFX9-NEXT:    v_add_f32_e32 v5, v12, v5
+; GFX9-NEXT:    s_sub_u32 s8, 0, s10
+; GFX9-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v11, v11
+; GFX9-NEXT:    s_and_b32 s0, s0, 1
+; GFX9-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX9-NEXT:    s_subb_u32 s9, 0, s11
+; GFX9-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v13, s9, v5
+; GFX9-NEXT:    v_mul_lo_u32 v14, s8, v11
+; GFX9-NEXT:    v_mul_hi_u32 v16, s8, v5
+; GFX9-NEXT:    v_mul_lo_u32 v12, s8, v5
+; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[0:1]
+; GFX9-NEXT:    v_add3_u32 v4, v13, v14, v16
+; GFX9-NEXT:    v_cndmask_b32_e64 v1, v1, v10, s[0:1]
+; GFX9-NEXT:    v_mul_lo_u32 v9, v11, v12
+; GFX9-NEXT:    v_mul_lo_u32 v10, v5, v4
+; GFX9-NEXT:    v_mul_hi_u32 v13, v5, v12
+; GFX9-NEXT:    v_mul_hi_u32 v12, v11, v12
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[2:3], v9, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[2:3]
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[2:3], v9, v13
+; GFX9-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[2:3]
+; GFX9-NEXT:    v_mul_lo_u32 v13, v11, v4
+; GFX9-NEXT:    v_add_u32_e32 v9, v10, v9
+; GFX9-NEXT:    v_mul_hi_u32 v10, v5, v4
+; GFX9-NEXT:    v_mul_hi_u32 v4, v11, v4
+; GFX9-NEXT:    v_add_co_u32_e64 v12, s[2:3], v13, v12
+; GFX9-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[2:3]
+; GFX9-NEXT:    v_add_co_u32_e64 v10, s[2:3], v12, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[2:3]
+; GFX9-NEXT:    v_add_co_u32_e64 v9, s[2:3], v10, v9
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[2:3]
+; GFX9-NEXT:    v_add_u32_e32 v12, v13, v12
+; GFX9-NEXT:    v_add_co_u32_e64 v5, s[2:3], v5, v9
+; GFX9-NEXT:    v_add3_u32 v10, v12, v10, v4
+; GFX9-NEXT:    v_addc_co_u32_e64 v9, s[6:7], v11, v10, s[2:3]
+; GFX9-NEXT:    v_mul_lo_u32 v12, s9, v5
+; GFX9-NEXT:    v_mul_lo_u32 v13, s8, v9
+; GFX9-NEXT:    v_mul_hi_u32 v14, s8, v5
+; GFX9-NEXT:    v_mul_lo_u32 v15, s8, v5
+; GFX9-NEXT:    v_cndmask_b32_e64 v4, v3, v7, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v8, v2, vcc
+; GFX9-NEXT:    v_add3_u32 v3, v12, v13, v14
+; GFX9-NEXT:    v_mul_lo_u32 v7, v9, v15
+; GFX9-NEXT:    v_mul_lo_u32 v8, v5, v3
+; GFX9-NEXT:    v_add_u32_e32 v10, v11, v10
+; GFX9-NEXT:    v_mul_hi_u32 v11, v5, v15
+; GFX9-NEXT:    v_mul_hi_u32 v12, v9, v15
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v7, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v7, v11
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v11, v9, v3
+; GFX9-NEXT:    v_add_u32_e32 v7, v8, v7
+; GFX9-NEXT:    v_mul_hi_u32 v8, v5, v3
+; GFX9-NEXT:    v_mul_hi_u32 v3, v9, v3
+; GFX9-NEXT:    v_add_co_u32_e32 v11, vcc, v11, v12
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, v11, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v8, v7
+; GFX9-NEXT:    v_add_u32_e32 v11, v12, v11
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add3_u32 v3, v11, v8, v3
+; GFX9-NEXT:    v_addc_co_u32_e64 v3, vcc, v10, v3, s[2:3]
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v5, v7
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v8, s15, v7
+; GFX9-NEXT:    v_mul_lo_u32 v9, s14, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, v6, v2, s[0:1]
+; GFX9-NEXT:    v_mul_hi_u32 v2, s14, v7
+; GFX9-NEXT:    v_mul_hi_u32 v7, s15, v7
+; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, v8, v9
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v6, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v6, s15, v3
+; GFX9-NEXT:    v_add_u32_e32 v2, v8, v2
+; GFX9-NEXT:    v_mul_hi_u32 v8, s14, v3
+; GFX9-NEXT:    v_mul_hi_u32 v3, s15, v3
+; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, v6, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, v6, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v6, v2
+; GFX9-NEXT:    v_add_u32_e32 v7, v7, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_add3_u32 v3, v7, v6, v3
+; GFX9-NEXT:    v_mul_lo_u32 v6, s11, v2
+; GFX9-NEXT:    v_mul_lo_u32 v7, s10, v3
+; GFX9-NEXT:    v_mul_hi_u32 v8, s10, v2
+; GFX9-NEXT:    v_mul_lo_u32 v10, s10, v2
+; GFX9-NEXT:    v_mov_b32_e32 v11, s15
+; GFX9-NEXT:    v_mov_b32_e32 v9, s11
+; GFX9-NEXT:    v_add3_u32 v6, v6, v7, v8
+; GFX9-NEXT:    v_sub_co_u32_e32 v7, vcc, s14, v10
+; GFX9-NEXT:    v_subb_co_u32_e64 v8, s[0:1], v11, v6, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v8
+; GFX9-NEXT:    v_sub_u32_e32 v6, s15, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v8
+; GFX9-NEXT:    v_subb_co_u32_e32 v6, vcc, v6, v9, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v10, v10, v11, s[0:1]
+; GFX9-NEXT:    v_subrev_co_u32_e32 v11, vcc, s10, v7
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v12, s[0:1], 0, v6, vcc
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s11, v12
+; GFX9-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s10, v11
+; GFX9-NEXT:    v_cndmask_b32_e64 v14, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s11, v12
+; GFX9-NEXT:    v_cndmask_b32_e64 v13, v13, v14, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v14, s[0:1], 1, v2
+; GFX9-NEXT:    v_subb_co_u32_e32 v6, vcc, v6, v9, vcc
+; GFX9-NEXT:    v_addc_co_u32_e64 v15, s[0:1], 0, v3, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e32 v9, vcc, 1, v14
+; GFX9-NEXT:    v_addc_co_u32_e32 v16, vcc, 0, v15, vcc
+; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v13
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX9-NEXT:    v_cndmask_b32_e32 v9, v14, v9, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v14, v15, v16, vcc
+; GFX9-NEXT:    v_subrev_co_u32_e64 v15, s[0:1], s10, v11
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v6, s[0:1], 0, v6, s[0:1]
+; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, v2, v9, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v13, 0
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, v3, v14, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e32 v9, v11, v15, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v10, v12, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, v7, v9, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, v8, v10, s[0:1]
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dwordx4 v13, v[0:3], s[4:5]
+; GFX9-NEXT:    global_store_dwordx4 v13, v[4:7], s[6:7]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_v2i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_clause 0x2
+; GFX10-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
+; GFX10-NEXT:    s_load_dwordx4 s[12:15], s[4:5], 0x0
+; GFX10-NEXT:    s_load_dwordx4 s[16:19], s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s9
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v2, s11
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s8
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v3, s10
+; GFX10-NEXT:    s_sub_u32 s2, 0, s8
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; GFX10-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; GFX10-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX10-NEXT:    s_and_b32 s0, s0, 1
+; GFX10-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX10-NEXT:    v_add_f32_e32 v1, v2, v3
+; GFX10-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX10-NEXT:    s_subb_u32 s1, 0, s9
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    s_sub_u32 s3, 0, s10
+; GFX10-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX10-NEXT:    s_and_b32 s0, s0, 1
+; GFX10-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX10-NEXT:    s_subb_u32 s6, 0, s11
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x5f7ffffc, v1
+; GFX10-NEXT:    v_mul_f32_e32 v2, 0x2f800000, v0
+; GFX10-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v1
+; GFX10-NEXT:    v_trunc_f32_e32 v2, v2
+; GFX10-NEXT:    v_trunc_f32_e32 v3, v3
+; GFX10-NEXT:    v_mul_f32_e32 v4, 0xcf800000, v2
+; GFX10-NEXT:    v_mul_f32_e32 v5, 0xcf800000, v3
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; GFX10-NEXT:    v_add_f32_e32 v0, v4, v0
+; GFX10-NEXT:    v_add_f32_e32 v1, v5, v1
+; GFX10-NEXT:    v_mul_lo_u32 v4, s2, v2
+; GFX10-NEXT:    v_mul_lo_u32 v8, s3, v3
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v5, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v6, s2, v0
+; GFX10-NEXT:    v_mul_lo_u32 v9, s6, v1
+; GFX10-NEXT:    v_mul_hi_u32 v10, s3, v1
+; GFX10-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GFX10-NEXT:    v_mul_lo_u32 v11, s3, v1
+; GFX10-NEXT:    v_add3_u32 v4, v5, v4, v6
+; GFX10-NEXT:    v_add3_u32 v8, v9, v8, v10
+; GFX10-NEXT:    v_mul_lo_u32 v5, v2, v7
+; GFX10-NEXT:    v_mul_hi_u32 v6, v0, v7
+; GFX10-NEXT:    v_mul_lo_u32 v12, v0, v4
+; GFX10-NEXT:    v_mul_hi_u32 v7, v2, v7
+; GFX10-NEXT:    v_mul_lo_u32 v13, v2, v4
+; GFX10-NEXT:    v_mul_lo_u32 v9, v3, v11
+; GFX10-NEXT:    v_mul_lo_u32 v15, v1, v8
+; GFX10-NEXT:    v_mul_hi_u32 v10, v1, v11
+; GFX10-NEXT:    v_mul_hi_u32 v11, v3, v11
+; GFX10-NEXT:    v_mul_lo_u32 v16, v3, v8
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v12
+; GFX10-NEXT:    v_mul_hi_u32 v14, v0, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v7, s0, v13, v7
+; GFX10-NEXT:    v_mul_hi_u32 v17, v1, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v9, s0, v9, v15
+; GFX10-NEXT:    v_mul_hi_u32 v4, v2, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v11, s0, v16, v11
+; GFX10-NEXT:    v_mul_hi_u32 v8, v3, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v5, v6
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v7, v14
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v9, s0, v9, v10
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, v12, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v10, s0, v11, v17
+; GFX10-NEXT:    v_add_nc_u32_e32 v7, v13, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v6, v5
+; GFX10-NEXT:    v_add_nc_u32_e32 v9, v15, v9
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v11, v16, v11
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v5
+; GFX10-NEXT:    v_add_co_u32 v9, s0, v10, v9
+; GFX10-NEXT:    v_add3_u32 v4, v7, v6, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s0
+; GFX10-NEXT:    v_mul_lo_u32 v7, s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s0, v2, v4, vcc_lo
+; GFX10-NEXT:    v_add3_u32 v5, v11, v10, v8
+; GFX10-NEXT:    v_add_co_u32 v1, s0, v1, v9
+; GFX10-NEXT:    v_mul_lo_u32 v8, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v9, s2, v0
+; GFX10-NEXT:    v_mul_lo_u32 v11, s2, v6
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v10, s1, v3, v5, s0
+; GFX10-NEXT:    v_mul_lo_u32 v13, s6, v1
+; GFX10-NEXT:    v_mul_hi_u32 v14, s3, v1
+; GFX10-NEXT:    v_mul_lo_u32 v12, s3, v1
+; GFX10-NEXT:    v_mul_lo_u32 v16, v6, v7
+; GFX10-NEXT:    v_mul_lo_u32 v15, s3, v10
+; GFX10-NEXT:    v_mul_hi_u32 v17, v0, v7
+; GFX10-NEXT:    v_add3_u32 v8, v8, v11, v9
+; GFX10-NEXT:    v_mul_hi_u32 v7, v6, v7
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, v2, v4
+; GFX10-NEXT:    v_add_nc_u32_e32 v3, v3, v5
+; GFX10-NEXT:    v_mul_lo_u32 v9, v10, v12
+; GFX10-NEXT:    v_mul_hi_u32 v18, v0, v8
+; GFX10-NEXT:    v_add3_u32 v13, v13, v15, v14
+; GFX10-NEXT:    v_mul_lo_u32 v14, v0, v8
+; GFX10-NEXT:    v_mul_lo_u32 v15, v6, v8
+; GFX10-NEXT:    v_mul_hi_u32 v6, v6, v8
+; GFX10-NEXT:    v_mul_hi_u32 v11, v1, v12
+; GFX10-NEXT:    v_mul_lo_u32 v8, v1, v13
+; GFX10-NEXT:    v_mul_hi_u32 v12, v10, v12
+; GFX10-NEXT:    v_mul_lo_u32 v19, v10, v13
+; GFX10-NEXT:    v_mul_hi_u32 v20, v1, v13
+; GFX10-NEXT:    v_add_co_u32 v14, s1, v16, v14
+; GFX10-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v7, s1, v15, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v8, s1, v9, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v14, s1, v14, v17
+; GFX10-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v12, s1, v19, v12
+; GFX10-NEXT:    v_cndmask_b32_e64 v17, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v7, s1, v7, v18
+; GFX10-NEXT:    v_add_nc_u32_e32 v14, v16, v14
+; GFX10-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v8, s1, v8, v11
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v7, s1, v7, v14
+; GFX10-NEXT:    v_add_nc_u32_e32 v11, v15, v18
+; GFX10-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v12, s1, v12, v20
+; GFX10-NEXT:    v_add_nc_u32_e32 v8, v9, v8
+; GFX10-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s1
+; GFX10-NEXT:    v_add3_u32 v4, v11, v14, v6
+; GFX10-NEXT:    v_mul_hi_u32 v6, v10, v13
+; GFX10-NEXT:    v_mov_b32_e32 v10, 0
+; GFX10-NEXT:    v_add_nc_u32_e32 v9, v17, v15
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v2, vcc_lo, v2, v4, vcc_lo
+; GFX10-NEXT:    v_add_co_u32 v4, s1, v12, v8
+; GFX10-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v7
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s1
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v2, vcc_lo, 0, v2, vcc_lo
+; GFX10-NEXT:    v_add3_u32 v5, v9, v8, v6
+; GFX10-NEXT:    v_mul_lo_u32 v6, s17, v0
+; GFX10-NEXT:    v_mul_lo_u32 v7, s16, v2
+; GFX10-NEXT:    v_mul_hi_u32 v8, s17, v0
+; GFX10-NEXT:    v_mul_hi_u32 v0, s16, v0
+; GFX10-NEXT:    v_mul_lo_u32 v9, s17, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, vcc_lo, v3, v5, s0
+; GFX10-NEXT:    v_mul_hi_u32 v5, s16, v2
+; GFX10-NEXT:    v_mul_hi_u32 v2, s17, v2
+; GFX10-NEXT:    v_add_co_u32 v1, vcc_lo, v1, v4
+; GFX10-NEXT:    v_add_co_u32 v6, s0, v6, v7
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v3, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v8, s0, v9, v8
+; GFX10-NEXT:    v_add_co_u32 v0, s1, v6, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s1
+; GFX10-NEXT:    v_add_co_u32 v5, s0, v8, v5
+; GFX10-NEXT:    v_mul_hi_u32 v12, s18, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v7, v0
+; GFX10-NEXT:    v_mul_hi_u32 v7, s18, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, v6, v8
+; GFX10-NEXT:    v_mul_lo_u32 v6, s19, v1
+; GFX10-NEXT:    v_add_co_u32 v0, s0, v5, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s19, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s0
+; GFX10-NEXT:    v_mul_lo_u32 v8, s9, v0
+; GFX10-NEXT:    v_mul_hi_u32 v9, s8, v0
+; GFX10-NEXT:    v_add3_u32 v2, v4, v5, v2
+; GFX10-NEXT:    v_mul_lo_u32 v4, s18, v3
+; GFX10-NEXT:    v_mul_lo_u32 v5, s19, v3
+; GFX10-NEXT:    v_mul_lo_u32 v13, s8, v0
+; GFX10-NEXT:    v_mul_hi_u32 v3, s19, v3
+; GFX10-NEXT:    v_mul_lo_u32 v11, s8, v2
+; GFX10-NEXT:    v_add_co_u32 v4, s0, v6, v4
+; GFX10-NEXT:    v_add_co_u32 v1, s1, v5, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s1
+; GFX10-NEXT:    v_add3_u32 v5, v8, v11, v9
+; GFX10-NEXT:    v_add_co_u32 v4, s0, v4, v7
+; GFX10-NEXT:    v_sub_co_u32 v8, vcc_lo, s16, v13
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v7, s17, v5
+; GFX10-NEXT:    v_sub_co_ci_u32_e64 v5, s0, s17, v5, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s8, v8
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, v6, v4
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v7, vcc_lo, s9, v7, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s9, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc_lo
+; GFX10-NEXT:    v_sub_co_u32 v13, vcc_lo, v8, s8
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v14, s0, 0, v7, vcc_lo
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s9, v5
+; GFX10-NEXT:    v_cndmask_b32_e64 v9, v11, v9, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s8, v13
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s0
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s9, v14
+; GFX10-NEXT:    v_cndmask_b32_e64 v15, 0, -1, s0
+; GFX10-NEXT:    v_add_co_u32 v1, s0, v1, v12
+; GFX10-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v17, s0, v0, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v18, s0, 0, v2, s0
+; GFX10-NEXT:    v_cmp_eq_u32_e64 s0, s9, v14
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v15, v11, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v11, v16, v12
+; GFX10-NEXT:    v_add_co_u32 v12, s0, v1, v4
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX10-NEXT:    v_add_co_u32 v4, s0, v17, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v15, s0, 0, v18, s0
+; GFX10-NEXT:    v_sub_co_u32 v19, s0, v13, s8
+; GFX10-NEXT:    v_add3_u32 v3, v11, v1, v3
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v1, vcc_lo, s9, v7, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX10-NEXT:    v_mul_lo_u32 v6, s11, v12
+; GFX10-NEXT:    v_mul_hi_u32 v11, s10, v12
+; GFX10-NEXT:    v_mul_lo_u32 v7, s10, v3
+; GFX10-NEXT:    v_mul_lo_u32 v16, s10, v12
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v20, s0, 0, v1, s0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v17, v4, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, v9
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v18, v15, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v14, v14, v20, vcc_lo
+; GFX10-NEXT:    v_add3_u32 v6, v6, v7, v11
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, v1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v2, v4, s0
+; GFX10-NEXT:    v_sub_co_u32 v7, s1, s18, v16
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v13, v19, vcc_lo
+; GFX10-NEXT:    v_sub_nc_u32_e32 v2, s19, v6
+; GFX10-NEXT:    v_sub_co_ci_u32_e64 v9, s2, s19, v6, s1
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, v5, v14, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, v8, v4, s0
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v2, s1, s11, v2, s1
+; GFX10-NEXT:    v_cmp_le_u32_e64 s1, s10, v7
+; GFX10-NEXT:    v_cmp_le_u32_e64 s2, s11, v9
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, s11, v9
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s1
+; GFX10-NEXT:    v_sub_co_u32 v13, s1, v7, s10
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s2
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v15, s2, 0, v2, s1
+; GFX10-NEXT:    v_cndmask_b32_e32 v6, v6, v11, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s11, v15
+; GFX10-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s10, v13
+; GFX10-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc_lo
+; GFX10-NEXT:    v_add_co_u32 v16, vcc_lo, v12, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v17, vcc_lo, 0, v3, vcc_lo
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, s11, v15
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v8, v11, vcc_lo
+; GFX10-NEXT:    v_add_co_u32 v11, vcc_lo, v16, 1
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v18, vcc_lo, 0, v17, vcc_lo
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v2, vcc_lo, s11, v2, s1
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX10-NEXT:    v_sub_co_u32 v8, s1, v13, s10
+; GFX10-NEXT:    v_subrev_co_ci_u32_e64 v2, s1, 0, v2, s1
+; GFX10-NEXT:    v_cndmask_b32_e32 v11, v16, v11, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s1, 0, v6
+; GFX10-NEXT:    v_cndmask_b32_e32 v16, v17, v18, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v6, v13, v8, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v15, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v12, v11, s1
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v16, s1
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v7, v6, s1
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v9, v8, s1
+; GFX10-NEXT:    global_store_dwordx4 v10, v[0:3], s[12:13]
+; GFX10-NEXT:    global_store_dwordx4 v10, v[4:7], s[14:15]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv <2 x i64> %x, %y
+  store <2 x i64> %div, <2 x i64> addrspace(1)* %out0
+  %rem = urem <2 x i64> %x, %y
+  store <2 x i64> %rem, <2 x i64> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udiv_i8(i8 addrspace(1)* %out0, i8 addrspace(1)* %out1, i8 %x, i8 %y) {
+; GFX8-LABEL: udiv_i8:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s1, s[4:5], 0x10
+; GFX8-NEXT:    s_movk_i32 s0, 0xff
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s2, s1, 8
+; GFX8-NEXT:    s_and_b32 s6, s2, s0
+; GFX8-NEXT:    v_cvt_f32_ubyte0_e32 v0, s6
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_sub_i32 s2, 0, s6
+; GFX8-NEXT:    s_and_b32 s7, s1, s0
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s2, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s7, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s6
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s7, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    flat_store_byte v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_byte v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udiv_i8:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s1, s[4:5], 0x10
+; GFX9-NEXT:    s_movk_i32 s0, 0xff
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s2, s1, 8
+; GFX9-NEXT:    s_and_b32 s6, s2, s0
+; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v0, s6
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_i32 s2, 0, s6
+; GFX9-NEXT:    s_and_b32 s7, s1, s0
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s2, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s7, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s7, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udiv_i8:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_movk_i32 s1, 0xff
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s2, s0, 8
+; GFX10-NEXT:    s_and_b32 s0, s0, s1
+; GFX10-NEXT:    s_and_b32 s6, s2, s1
+; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v0, s6
+; GFX10-NEXT:    s_sub_i32 s2, 0, s6
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s2, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv i8 %x, %y
+  store i8 %div, i8 addrspace(1)* %out0
+  %rem = urem i8 %x, %y
+  store i8 %rem, i8 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_v2i8(<2 x i8> addrspace(1)* %out0, <2 x i8> addrspace(1)* %out1, <2 x i8> %x, <2 x i8> %y) {
+; GFX8-LABEL: udivrem_v2i8:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    s_movk_i32 s2, 0xff
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX8-NEXT:    s_and_b32 s3, s1, s2
+; GFX8-NEXT:    v_cvt_f32_ubyte0_e32 v0, s3
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    v_cvt_f32_ubyte3_e32 v1, s0
+; GFX8-NEXT:    s_sub_i32 s1, 0, s3
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    s_lshr_b32 s8, s0, 24
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_lo_u32 v2, s1, v0
+; GFX8-NEXT:    s_and_b32 s1, s0, s2
+; GFX8-NEXT:    s_sub_i32 s9, 0, s8
+; GFX8-NEXT:    v_mul_lo_u32 v3, s9, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX8-NEXT:    s_lshr_b32 s0, s0, 8
+; GFX8-NEXT:    s_and_b32 s9, s0, s2
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v0, s1, v0
+; GFX8-NEXT:    v_mul_hi_u32 v2, v1, v3
+; GFX8-NEXT:    v_mul_lo_u32 v3, v0, s3
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v0
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s1, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v2
+; GFX8-NEXT:    v_mul_hi_u32 v1, s9, v1
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v2, v1, s8
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s9, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s8, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s8, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s8, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_and_b32_e32 v1, s2, v1
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s8, v2
+; GFX8-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_or_b32_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX8-NEXT:    v_mov_b32_e32 v0, s4
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    flat_store_short v[0:1], v4
+; GFX8-NEXT:    v_and_b32_e32 v0, s2, v2
+; GFX8-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX8-NEXT:    v_or_b32_sdwa v2, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_short v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_v2i8:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX9-NEXT:    s_movk_i32 s6, 0xff
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX9-NEXT:    s_and_b32 s7, s1, s6
+; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v0, s7
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    v_cvt_f32_ubyte3_e32 v1, s0
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX9-NEXT:    s_sub_i32 s1, 0, s7
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    s_lshr_b32 s8, s0, 24
+; GFX9-NEXT:    v_mul_lo_u32 v2, s1, v0
+; GFX9-NEXT:    s_sub_i32 s1, 0, s8
+; GFX9-NEXT:    v_mul_lo_u32 v3, s1, v1
+; GFX9-NEXT:    s_and_b32 s9, s0, s6
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX9-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX9-NEXT:    s_and_b32 s10, s1, s6
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v3
+; GFX9-NEXT:    v_mul_hi_u32 v1, s10, v1
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_lo_u32 v2, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, s8
+; GFX9-NEXT:    v_sub_u32_e32 v2, s9, v2
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s7, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s7, v2
+; GFX9-NEXT:    v_sub_u32_e32 v3, s10, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s8, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s8, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s8, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_and_b32_e32 v1, s6, v1
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s8, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
+; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_short v1, v0, s[0:1]
+; GFX9-NEXT:    v_and_b32_e32 v0, s6, v3
+; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
+; GFX9-NEXT:    v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NEXT:    global_store_short v1, v0, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_v2i8:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_movk_i32 s1, 0xff
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_cvt_f32_ubyte3_e32 v0, s0
+; GFX10-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX10-NEXT:    s_lshr_b32 s3, s0, 24
+; GFX10-NEXT:    s_and_b32 s2, s2, s1
+; GFX10-NEXT:    s_sub_i32 s6, 0, s3
+; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v1, s2
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, s6, v0
+; GFX10-NEXT:    s_sub_i32 s6, 0, s2
+; GFX10-NEXT:    v_mul_lo_u32 v3, s6, v1
+; GFX10-NEXT:    s_lshr_b32 s6, s0, 8
+; GFX10-NEXT:    s_and_b32 s0, s0, s1
+; GFX10-NEXT:    s_and_b32 s6, s6, s1
+; GFX10-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v0, s6, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s0, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, v0, s3
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, s2
+; GFX10-NEXT:    v_add_nc_u32_e32 v6, 1, v1
+; GFX10-NEXT:    v_sub_nc_u32_e32 v2, s6, v2
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v3, s0, v3
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s3, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v5, s3, v2
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s2, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v4, s2, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v6, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v5, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s3, v2
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v4, s0
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v4, s3, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v6, 1, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s2, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v5, s2, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
+; GFX10-NEXT:    v_and_b32_sdwa v0, v0, s1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v5, s0
+; GFX10-NEXT:    v_and_b32_sdwa v2, v2, s1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_short v1, v0, s[4:5]
+; GFX10-NEXT:    global_store_short v1, v2, s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv <2 x i8> %x, %y
+  store <2 x i8> %div, <2 x i8> addrspace(1)* %out0
+  %rem = urem <2 x i8> %x, %y
+  store <2 x i8> %rem, <2 x i8> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udiv_i16(i16 addrspace(1)* %out0, i16 addrspace(1)* %out1, i16 %x, i16 %y) {
+; GFX8-LABEL: udiv_i16:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX8-NEXT:    s_sub_i32 s1, 0, s6
+; GFX8-NEXT:    s_and_b32 s7, s0, 0xffff
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s7, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s6
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s7, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    flat_store_short v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_short v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udiv_i16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX9-NEXT:    s_sub_i32 s1, 0, s6
+; GFX9-NEXT:    s_and_b32 s7, s0, 0xffff
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s7, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s7, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_short v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_short v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udiv_i16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX10-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX10-NEXT:    s_sub_i32 s1, 0, s6
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_short v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_short v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv i16 %x, %y
+  store i16 %div, i16 addrspace(1)* %out0
+  %rem = urem i16 %x, %y
+  store i16 %rem, i16 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_v2i16(<2 x i16> addrspace(1)* %out0, <2 x i16> addrspace(1)* %out1, <2 x i16> %x, <2 x i16> %y) {
+; GFX8-LABEL: udivrem_v2i16:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x14
+; GFX8-NEXT:    s_mov_b32 s2, 0xffff
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_and_b32 s6, s0, s2
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX8-NEXT:    s_lshr_b32 s3, s0, 16
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v1, s3
+; GFX8-NEXT:    s_sub_i32 s1, 0, s6
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x0
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX8-NEXT:    v_mul_lo_u32 v2, s1, v0
+; GFX8-NEXT:    s_sub_i32 s1, 0, s3
+; GFX8-NEXT:    s_and_b32 s0, s0, s2
+; GFX8-NEXT:    v_mul_lo_u32 v3, s1, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; GFX8-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX8-NEXT:    v_add_u32_e32 v1, vcc, v1, v3
+; GFX8-NEXT:    v_mul_hi_u32 v1, s4, v1
+; GFX8-NEXT:    v_mul_lo_u32 v2, v0, s6
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_sub_u32_e32 v2, vcc, s0, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v3, s[0:1], s6, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
+; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 1, v0
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX8-NEXT:    v_mul_lo_u32 v3, v1, s3
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s4, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v1
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s3, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s3, v3
+; GFX8-NEXT:    v_and_b32_e32 v1, s2, v1
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX8-NEXT:    v_or_b32_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    v_and_b32_e32 v0, s2, v3
+; GFX8-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT:    v_or_b32_sdwa v2, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    v_mov_b32_e32 v0, s8
+; GFX8-NEXT:    v_mov_b32_e32 v1, s9
+; GFX8-NEXT:    flat_store_dword v[0:1], v4
+; GFX8-NEXT:    v_mov_b32_e32 v0, s10
+; GFX8-NEXT:    v_mov_b32_e32 v1, s11
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_v2i16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s1, s[4:5], 0x14
+; GFX9-NEXT:    s_mov_b32 s0, 0xffff
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_and_b32 s7, s1, s0
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s7
+; GFX9-NEXT:    s_lshr_b32 s6, s1, 16
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s6
+; GFX9-NEXT:    s_sub_i32 s1, 0, s7
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_i32 s2, 0, s6
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, s1, v0
+; GFX9-NEXT:    s_load_dword s1, s[4:5], 0x10
+; GFX9-NEXT:    v_mul_lo_u32 v3, s2, v1
+; GFX9-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_and_b32 s9, s1, s0
+; GFX9-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v2
+; GFX9-NEXT:    v_mul_hi_u32 v0, s9, v0
+; GFX9-NEXT:    s_lshr_b32 s8, s1, 16
+; GFX9-NEXT:    v_add_u32_e32 v1, v1, v3
+; GFX9-NEXT:    v_mul_hi_u32 v1, s8, v1
+; GFX9-NEXT:    v_mul_lo_u32 v2, v0, s7
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_lo_u32 v3, v1, s6
+; GFX9-NEXT:    v_sub_u32_e32 v2, s9, v2
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s7, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s7, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s7, v2
+; GFX9-NEXT:    v_sub_u32_e32 v3, s8, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v5, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s6, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_add_u32_e32 v4, 1, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v4, s6, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0xffff
+; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX9-NEXT:    v_and_or_b32 v0, v0, v4, v1
+; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
+; GFX9-NEXT:    v_and_or_b32 v1, v2, v4, v1
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_v2i16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x14
+; GFX10-NEXT:    s_mov_b32 s2, 0xffff
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX10-NEXT:    s_and_b32 s3, s0, s2
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s1
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v1, s3
+; GFX10-NEXT:    s_sub_i32 s6, 0, s1
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, s6, v0
+; GFX10-NEXT:    s_sub_i32 s6, 0, s3
+; GFX10-NEXT:    v_mul_lo_u32 v3, s6, v1
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX10-NEXT:    s_and_b32 s0, s0, s2
+; GFX10-NEXT:    v_mul_hi_u32 v2, v0, v2
+; GFX10-NEXT:    v_mul_hi_u32 v3, v1, v3
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX10-NEXT:    v_mul_hi_u32 v0, s6, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, s0, v1
+; GFX10-NEXT:    v_mul_lo_u32 v2, v0, s1
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_mul_lo_u32 v3, v1, s3
+; GFX10-NEXT:    v_add_nc_u32_e32 v6, 1, v1
+; GFX10-NEXT:    v_sub_nc_u32_e32 v2, s6, v2
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v3, s0, v3
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s1, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v5, s1, v2
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s3, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s3, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v7, s0
+; GFX10-NEXT:    v_add_nc_u32_e32 v4, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s1, v2
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v5, s1, v2
+; GFX10-NEXT:    v_add_nc_u32_e32 v6, 1, v1
+; GFX10-NEXT:    v_cmp_le_u32_e64 s0, s3, v3
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v7, s3, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v4, 0xffff
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v6, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, v7, s0
+; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT:    v_and_or_b32 v0, v1, v4, v0
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-NEXT:    v_and_or_b32 v2, v3, v4, v2
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dword v1, v0, s[4:5]
+; GFX10-NEXT:    global_store_dword v1, v2, s[6:7]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv <2 x i16> %x, %y
+  store <2 x i16> %div, <2 x i16> addrspace(1)* %out0
+  %rem = urem <2 x i16> %x, %y
+  store <2 x i16> %rem, <2 x i16> addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_i3(i3 addrspace(1)* %out0, i3 addrspace(1)* %out1, i3 %x, i3 %y) {
+; GFX8-LABEL: udivrem_i3:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX8-NEXT:    s_and_b32 s6, s1, 7
+; GFX8-NEXT:    v_cvt_f32_ubyte0_e32 v0, s6
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    s_sub_i32 s1, 0, s6
+; GFX8-NEXT:    s_and_b32 s7, s0, 7
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s7, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s6
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s7, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    flat_store_byte v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_byte v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_i3:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX9-NEXT:    s_and_b32 s6, s1, 7
+; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v0, s6
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    s_sub_i32 s1, 0, s6
+; GFX9-NEXT:    s_and_b32 s7, s0, 7
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s7, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s7, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_i3:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x10
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_lshr_b32 s1, s0, 8
+; GFX10-NEXT:    s_and_b32 s0, s0, 7
+; GFX10-NEXT:    s_and_b32 s6, s1, 7
+; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v0, s6
+; GFX10-NEXT:    s_sub_i32 s1, 0, s6
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_byte v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_byte v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv i3 %x, %y
+  store i3 %div, i3 addrspace(1)* %out0
+  %rem = urem i3 %x, %y
+  store i3 %rem, i3 addrspace(1)* %out1
+  ret void
+}
+
+define amdgpu_kernel void @udivrem_i27(i27 addrspace(1)* %out0, i27 addrspace(1)* %out1, i27 %x, i27 %y) {
+; GFX8-LABEL: udivrem_i27:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX8-NEXT:    s_mov_b32 s2, 0x7ffffff
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    s_and_b32 s6, s1, s2
+; GFX8-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX8-NEXT:    s_sub_i32 s1, 0, s6
+; GFX8-NEXT:    s_and_b32 s7, s0, s2
+; GFX8-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX8-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX8-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX8-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX8-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, v0, v1
+; GFX8-NEXT:    v_mul_hi_u32 v2, s7, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_mul_lo_u32 v3, v2, s6
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_sub_u32_e32 v3, vcc, s7, v3
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 1, v2
+; GFX8-NEXT:    v_cmp_le_u32_e32 vcc, s6, v3
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; GFX8-NEXT:    v_subrev_u32_e64 v4, s[0:1], s6, v3
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    v_mov_b32_e32 v0, s2
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    flat_store_dword v[0:1], v3
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-LABEL: udivrem_i27:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX9-NEXT:    s_mov_b32 s2, 0x7ffffff
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_and_b32 s6, s1, s2
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX9-NEXT:    s_sub_i32 s1, 0, s6
+; GFX9-NEXT:    s_and_b32 s7, s0, s2
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
+; GFX9-NEXT:    v_mul_hi_u32 v0, s7, v0
+; GFX9-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_sub_u32_e32 v1, s7, v1
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    v_add_u32_e32 v3, 1, v0
+; GFX9-NEXT:    v_cmp_le_u32_e32 vcc, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; GFX9-NEXT:    v_subrev_u32_e32 v3, s6, v1
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX9-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: udivrem_i27:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x10
+; GFX10-NEXT:    s_mov_b32 s2, 0x7ffffff
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_and_b32 s6, s1, s2
+; GFX10-NEXT:    s_and_b32 s0, s0, s2
+; GFX10-NEXT:    v_cvt_f32_u32_e32 v0, s6
+; GFX10-NEXT:    s_sub_i32 s1, 0, s6
+; GFX10-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0
+; GFX10-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, s1, v0
+; GFX10-NEXT:    v_mul_hi_u32 v1, v0, v1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, v0, v1
+; GFX10-NEXT:    v_mul_hi_u32 v0, s0, v0
+; GFX10-NEXT:    v_mul_lo_u32 v1, v0, s6
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_sub_nc_u32_e32 v1, s0, v1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    v_add_nc_u32_e32 v2, 1, v0
+; GFX10-NEXT:    v_cmp_le_u32_e32 vcc_lo, s6, v1
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v3, s6, v1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-NEXT:    global_store_dword v2, v1, s[2:3]
+; GFX10-NEXT:    s_endpgm
+  %div = udiv i27 %x, %y
+  store i27 %div, i27 addrspace(1)* %out0
+  %rem = urem i27 %x, %y
+  store i27 %rem, i27 addrspace(1)* %out1
+  ret void
+}