/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp

Bug Summary

File:	llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp
Warning:	line 369, column 25 Called C++ object pointer is null
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name AMDGPUCallLowering.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/lib/Target/AMDGPU -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/lib/Target/AMDGPU -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/include -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/lib/Target/AMDGPU -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0=. -ferror-limit 19 -fvisibility hidden -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-08-28-193554-24367-1 -x c++ /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp
1//===-- llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp - Call lowering -----===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8///
9/// \file
10/// This file implements the lowering of LLVM calls to machine code calls for
11/// GlobalISel.
12///
13//===----------------------------------------------------------------------===//

15#include "AMDGPUCallLowering.h"
16#include "AMDGPU.h"
17#include "AMDGPULegalizerInfo.h"
18#include "AMDGPUTargetMachine.h"
19#include "SIMachineFunctionInfo.h"
20#include "SIRegisterInfo.h"
21#include "llvm/CodeGen/Analysis.h"
22#include "llvm/CodeGen/FunctionLoweringInfo.h"
23#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
24#include "llvm/IR/IntrinsicsAMDGPU.h"

26#define DEBUG_TYPE"amdgpu-call-lowering" "amdgpu-call-lowering"

28using namespace llvm;

30namespace {

32/// Wrapper around extendRegister to ensure we extend to a full 32-bit register.
33static Register extendRegisterMin32(CallLowering::ValueHandler &Handler,
                                  Register ValVReg, CCValAssign &VA) {
if (VA.getLocVT().getSizeInBits() < 32) {
  // 16-bit types are reported as legal for 32-bit registers. We need to
  // extend and do a 32-bit copy to avoid the verifier complaining about it.
  return Handler.MIRBuilder.buildAnyExt(LLT::scalar(32), ValVReg).getReg(0);
}

return Handler.extendRegister(ValVReg, VA);
42}

44struct AMDGPUOutgoingValueHandler : public CallLowering::OutgoingValueHandler {
AMDGPUOutgoingValueHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI,
                           MachineInstrBuilder MIB)
    : OutgoingValueHandler(B, MRI), MIB(MIB) {}

MachineInstrBuilder MIB;

Register getStackAddress(uint64_t Size, int64_t Offset,
                         MachinePointerInfo &MPO,
                         ISD::ArgFlagsTy Flags) override {
  llvm_unreachable("not implemented")::llvm::llvm_unreachable_internal("not implemented", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 54);
}

void assignValueToAddress(Register ValVReg, Register Addr, LLT MemTy,
                          MachinePointerInfo &MPO, CCValAssign &VA) override {
  llvm_unreachable("not implemented")::llvm::llvm_unreachable_internal("not implemented", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 59);
}

void assignValueToReg(Register ValVReg, Register PhysReg,
                      CCValAssign &VA) override {
  Register ExtReg = extendRegisterMin32(*this, ValVReg, VA);

  // If this is a scalar return, insert a readfirstlane just in case the value
  // ends up in a VGPR.
  // FIXME: Assert this is a shader return.
  const SIRegisterInfo *TRI
    = static_cast<const SIRegisterInfo *>(MRI.getTargetRegisterInfo());
  if (TRI->isSGPRReg(MRI, PhysReg)) {
    auto ToSGPR = MIRBuilder.buildIntrinsic(Intrinsic::amdgcn_readfirstlane,
                                            {MRI.getType(ExtReg)}, false)
      .addReg(ExtReg);
    ExtReg = ToSGPR.getReg(0);
  }

  MIRBuilder.buildCopy(PhysReg, ExtReg);
  MIB.addUse(PhysReg, RegState::Implicit);
}
81};

83struct AMDGPUIncomingArgHandler : public CallLowering::IncomingValueHandler {
uint64_t StackUsed = 0;

AMDGPUIncomingArgHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI)
    : IncomingValueHandler(B, MRI) {}

Register getStackAddress(uint64_t Size, int64_t Offset,
                         MachinePointerInfo &MPO,
                         ISD::ArgFlagsTy Flags) override {
  auto &MFI = MIRBuilder.getMF().getFrameInfo();

  // Byval is assumed to be writable memory, but other stack passed arguments
  // are not.
  const bool IsImmutable = !Flags.isByVal();
  int FI = MFI.CreateFixedObject(Size, Offset, IsImmutable);
  MPO = MachinePointerInfo::getFixedStack(MIRBuilder.getMF(), FI);
  auto AddrReg = MIRBuilder.buildFrameIndex(
      LLT::pointer(AMDGPUAS::PRIVATE_ADDRESS, 32), FI);
  StackUsed = std::max(StackUsed, Size + Offset);
  return AddrReg.getReg(0);
}

void assignValueToReg(Register ValVReg, Register PhysReg,
                      CCValAssign &VA) override {
  markPhysRegUsed(PhysReg);

  if (VA.getLocVT().getSizeInBits() < 32) {
    // 16-bit types are reported as legal for 32-bit registers. We need to do
    // a 32-bit copy, and truncate to avoid the verifier complaining about it.
    auto Copy = MIRBuilder.buildCopy(LLT::scalar(32), PhysReg);

    // If we have signext/zeroext, it applies to the whole 32-bit register
    // before truncation.
    auto Extended =
        buildExtensionHint(VA, Copy.getReg(0), LLT(VA.getLocVT()));
    MIRBuilder.buildTrunc(ValVReg, Extended);
    return;
  }

  IncomingValueHandler::assignValueToReg(ValVReg, PhysReg, VA);
}

void assignValueToAddress(Register ValVReg, Register Addr, LLT MemTy,
                          MachinePointerInfo &MPO, CCValAssign &VA) override {
  MachineFunction &MF = MIRBuilder.getMF();

  auto MMO = MF.getMachineMemOperand(
      MPO, MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant, MemTy,
      inferAlignFromPtrInfo(MF, MPO));
  MIRBuilder.buildLoad(ValVReg, Addr, *MMO);
}

/// How the physical register gets marked varies between formal
/// parameters (it's a basic-block live-in), and a call instruction
/// (it's an implicit-def of the BL).
virtual void markPhysRegUsed(unsigned PhysReg) = 0;
139};

141struct FormalArgHandler : public AMDGPUIncomingArgHandler {
FormalArgHandler(MachineIRBuilder &B, MachineRegisterInfo &MRI)
    : AMDGPUIncomingArgHandler(B, MRI) {}

void markPhysRegUsed(unsigned PhysReg) override {
  MIRBuilder.getMBB().addLiveIn(PhysReg);
}
148};

150struct CallReturnHandler : public AMDGPUIncomingArgHandler {
CallReturnHandler(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
                  MachineInstrBuilder MIB)
    : AMDGPUIncomingArgHandler(MIRBuilder, MRI), MIB(MIB) {}

void markPhysRegUsed(unsigned PhysReg) override {
  MIB.addDef(PhysReg, RegState::Implicit);
}

MachineInstrBuilder MIB;
160};

162struct AMDGPUOutgoingArgHandler : public AMDGPUOutgoingValueHandler {
/// For tail calls, the byte offset of the call's argument area from the
/// callee's. Unused elsewhere.
int FPDiff;

// Cache the SP register vreg if we need it more than once in this call site.
Register SPReg;

bool IsTailCall;

AMDGPUOutgoingArgHandler(MachineIRBuilder &MIRBuilder,
                         MachineRegisterInfo &MRI, MachineInstrBuilder MIB,
                         bool IsTailCall = false, int FPDiff = 0)
    : AMDGPUOutgoingValueHandler(MIRBuilder, MRI, MIB), FPDiff(FPDiff),
      IsTailCall(IsTailCall) {}

Register getStackAddress(uint64_t Size, int64_t Offset,
                         MachinePointerInfo &MPO,
                         ISD::ArgFlagsTy Flags) override {
  MachineFunction &MF = MIRBuilder.getMF();
  const LLT PtrTy = LLT::pointer(AMDGPUAS::PRIVATE_ADDRESS, 32);
  const LLT S32 = LLT::scalar(32);

  if (IsTailCall) {
    Offset += FPDiff;
    int FI = MF.getFrameInfo().CreateFixedObject(Size, Offset, true);
    auto FIReg = MIRBuilder.buildFrameIndex(PtrTy, FI);
    MPO = MachinePointerInfo::getFixedStack(MF, FI);
    return FIReg.getReg(0);
  }

  const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

  if (!SPReg)
    SPReg = MIRBuilder.buildCopy(PtrTy, MFI->getStackPtrOffsetReg()).getReg(0);

  auto OffsetReg = MIRBuilder.buildConstant(S32, Offset);

  auto AddrReg = MIRBuilder.buildPtrAdd(PtrTy, SPReg, OffsetReg);
  MPO = MachinePointerInfo::getStack(MF, Offset);
  return AddrReg.getReg(0);
}

void assignValueToReg(Register ValVReg, Register PhysReg,
                      CCValAssign &VA) override {
  MIB.addUse(PhysReg, RegState::Implicit);
  Register ExtReg = extendRegisterMin32(*this, ValVReg, VA);
  MIRBuilder.buildCopy(PhysReg, ExtReg);
}

void assignValueToAddress(Register ValVReg, Register Addr, LLT MemTy,
                          MachinePointerInfo &MPO, CCValAssign &VA) override {
  MachineFunction &MF = MIRBuilder.getMF();
  uint64_t LocMemOffset = VA.getLocMemOffset();
  const auto &ST = MF.getSubtarget<GCNSubtarget>();

  auto MMO = MF.getMachineMemOperand(
      MPO, MachineMemOperand::MOStore, MemTy,
      commonAlignment(ST.getStackAlignment(), LocMemOffset));
  MIRBuilder.buildStore(ValVReg, Addr, *MMO);
}

void assignValueToAddress(const CallLowering::ArgInfo &Arg,
                          unsigned ValRegIndex, Register Addr, LLT MemTy,
                          MachinePointerInfo &MPO, CCValAssign &VA) override {
  Register ValVReg = VA.getLocInfo() != CCValAssign::LocInfo::FPExt
                         ? extendRegister(Arg.Regs[ValRegIndex], VA)
                         : Arg.Regs[ValRegIndex];
  assignValueToAddress(ValVReg, Addr, MemTy, MPO, VA);
}
232};
233}

235AMDGPUCallLowering::AMDGPUCallLowering(const AMDGPUTargetLowering &TLI)
: CallLowering(&TLI) {
237}

239// FIXME: Compatability shim
240static ISD::NodeType extOpcodeToISDExtOpcode(unsigned MIOpc) {
switch (MIOpc) {
case TargetOpcode::G_SEXT:
  return ISD::SIGN_EXTEND;
case TargetOpcode::G_ZEXT:
  return ISD::ZERO_EXTEND;
case TargetOpcode::G_ANYEXT:
  return ISD::ANY_EXTEND;
default:
  llvm_unreachable("not an extend opcode")::llvm::llvm_unreachable_internal("not an extend opcode", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 249);
}
251}

253bool AMDGPUCallLowering::canLowerReturn(MachineFunction &MF,
                                      CallingConv::ID CallConv,
                                      SmallVectorImpl<BaseArgInfo> &Outs,
                                      bool IsVarArg) const {
// For shaders. Vector types should be explicitly handled by CC.
if (AMDGPU::isEntryFunctionCC(CallConv))
  return true;

SmallVector<CCValAssign, 16> ArgLocs;
const SITargetLowering &TLI = *getTLI<SITargetLowering>();
CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs,
               MF.getFunction().getContext());

return checkReturn(CCInfo, Outs, TLI.CCAssignFnForReturn(CallConv, IsVarArg));
267}

269/// Lower the return value for the already existing \p Ret. This assumes that
270/// \p B's insertion point is correct.
271bool AMDGPUCallLowering::lowerReturnVal(MachineIRBuilder &B,
                                      const Value *Val, ArrayRef<Register> VRegs,
                                      MachineInstrBuilder &Ret) const {
if (!Val)
  return true;

auto &MF = B.getMF();
const auto &F = MF.getFunction();
const DataLayout &DL = MF.getDataLayout();
MachineRegisterInfo *MRI = B.getMRI();
LLVMContext &Ctx = F.getContext();

CallingConv::ID CC = F.getCallingConv();
const SITargetLowering &TLI = *getTLI<SITargetLowering>();

SmallVector<EVT, 8> SplitEVTs;
ComputeValueVTs(TLI, DL, Val->getType(), SplitEVTs);
assert(VRegs.size() == SplitEVTs.size() &&(static_cast <bool> (VRegs.size() == SplitEVTs.size() &&
 "For each split Type there should be exactly one VReg.") ? void
 (0) : __assert_fail ("VRegs.size() == SplitEVTs.size() && \"For each split Type there should be exactly one VReg.\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 289, __extension__ __PRETTY_FUNCTION__))
       "For each split Type there should be exactly one VReg.")(static_cast <bool> (VRegs.size() == SplitEVTs.size() &&
 "For each split Type there should be exactly one VReg.") ? void
 (0) : __assert_fail ("VRegs.size() == SplitEVTs.size() && \"For each split Type there should be exactly one VReg.\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 289, __extension__ __PRETTY_FUNCTION__));

SmallVector<ArgInfo, 8> SplitRetInfos;

for (unsigned i = 0; i < SplitEVTs.size(); ++i) {
  EVT VT = SplitEVTs[i];
  Register Reg = VRegs[i];
  ArgInfo RetInfo(Reg, VT.getTypeForEVT(Ctx), 0);
  setArgFlags(RetInfo, AttributeList::ReturnIndex, DL, F);

  if (VT.isScalarInteger()) {
    unsigned ExtendOp = TargetOpcode::G_ANYEXT;
    if (RetInfo.Flags[0].isSExt()) {
      assert(RetInfo.Regs.size() == 1 && "expect only simple return values")(static_cast <bool> (RetInfo.Regs.size() == 1 &&
 "expect only simple return values") ? void (0) : __assert_fail
 ("RetInfo.Regs.size() == 1 && \"expect only simple return values\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 302, __extension__ __PRETTY_FUNCTION__));
      ExtendOp = TargetOpcode::G_SEXT;
    } else if (RetInfo.Flags[0].isZExt()) {
      assert(RetInfo.Regs.size() == 1 && "expect only simple return values")(static_cast <bool> (RetInfo.Regs.size() == 1 &&
 "expect only simple return values") ? void (0) : __assert_fail
 ("RetInfo.Regs.size() == 1 && \"expect only simple return values\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 305, __extension__ __PRETTY_FUNCTION__));
      ExtendOp = TargetOpcode::G_ZEXT;
    }

    EVT ExtVT = TLI.getTypeForExtReturn(Ctx, VT,
                                        extOpcodeToISDExtOpcode(ExtendOp));
    if (ExtVT != VT) {
      RetInfo.Ty = ExtVT.getTypeForEVT(Ctx);
      LLT ExtTy = getLLTForType(*RetInfo.Ty, DL);
      Reg = B.buildInstr(ExtendOp, {ExtTy}, {Reg}).getReg(0);
    }
  }

  if (Reg != RetInfo.Regs[0]) {
    RetInfo.Regs[0] = Reg;
    // Reset the arg flags after modifying Reg.
    setArgFlags(RetInfo, AttributeList::ReturnIndex, DL, F);
  }

  splitToValueTypes(RetInfo, SplitRetInfos, DL, CC);
}

CCAssignFn *AssignFn = TLI.CCAssignFnForReturn(CC, F.isVarArg());

OutgoingValueAssigner Assigner(AssignFn);
AMDGPUOutgoingValueHandler RetHandler(B, *MRI, Ret);
return determineAndHandleAssignments(RetHandler, Assigner, SplitRetInfos, B,
                                     CC, F.isVarArg());
333}

335bool AMDGPUCallLowering::lowerReturn(MachineIRBuilder &B, const Value *Val,
                                   ArrayRef<Register> VRegs,
                                   FunctionLoweringInfo &FLI) const {

MachineFunction &MF = B.getMF();
MachineRegisterInfo &MRI = MF.getRegInfo();
SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
MFI->setIfReturnsVoid(!Val);
1
Assuming 'Val' is null→

assert(!Val == VRegs.empty() && "Return value without a vreg")(static_cast <bool> (!Val == VRegs.empty() && "Return value without a vreg"
) ? void (0) : __assert_fail ("!Val == VRegs.empty() && \"Return value without a vreg\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 344, __extension__ __PRETTY_FUNCTION__));
2
←
Assuming the condition is true→
3
←
'?' condition is true→

CallingConv::ID CC = B.getMF().getFunction().getCallingConv();
const bool IsShader = AMDGPU::isShader(CC);
const bool IsWaveEnd =
    (IsShader && MFI->returnsVoid()) || AMDGPU::isKernel(CC);
4
←
Assuming 'IsShader' is false→
if (IsWaveEnd4.1
'IsWaveEnd' is false
) {
5
←
Taking false branch→
  B.buildInstr(AMDGPU::S_ENDPGM)
    .addImm(0);
  return true;
}

auto const &ST = MF.getSubtarget<GCNSubtarget>();

unsigned ReturnOpc =
    IsShader5.1
'IsShader' is false
 ? AMDGPU::SI_RETURN_TO_EPILOG : AMDGPU::S_SETPC_B64_return;
6
←
'?' condition is false→

auto Ret = B.buildInstrNoInsert(ReturnOpc);
Register ReturnAddrVReg;
if (ReturnOpc6.1
'ReturnOpc' is equal to S_SETPC_B64_return
 == AMDGPU::S_SETPC_B64_return) {
7
←
Taking true branch→
  ReturnAddrVReg = MRI.createVirtualRegister(&AMDGPU::CCR_SGPR_64RegClass);
  Ret.addUse(ReturnAddrVReg);
}

if (!FLI.CanLowerReturn)
8
←
Assuming field 'CanLowerReturn' is false→
9
←
Taking true branch→
  insertSRetStores(B, Val->getType(), VRegs, FLI.DemoteRegister);
10
←
Called C++ object pointer is null
else if (!lowerReturnVal(B, Val, VRegs, Ret))
  return false;

if (ReturnOpc == AMDGPU::S_SETPC_B64_return) {
  const SIRegisterInfo *TRI = ST.getRegisterInfo();
  Register LiveInReturn = MF.addLiveIn(TRI->getReturnAddressReg(MF),
                                       &AMDGPU::SGPR_64RegClass);
  B.buildCopy(ReturnAddrVReg, LiveInReturn);
}

// TODO: Handle CalleeSavedRegsViaCopy.

B.insertInstr(Ret);
return true;
384}

386void AMDGPUCallLowering::lowerParameterPtr(Register DstReg, MachineIRBuilder &B,
                                         uint64_t Offset) const {
MachineFunction &MF = B.getMF();
const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
MachineRegisterInfo &MRI = MF.getRegInfo();
Register KernArgSegmentPtr =
  MFI->getPreloadedReg(AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
Register KernArgSegmentVReg = MRI.getLiveInVirtReg(KernArgSegmentPtr);

auto OffsetReg = B.buildConstant(LLT::scalar(64), Offset);

B.buildPtrAdd(DstReg, KernArgSegmentVReg, OffsetReg);
398}

400void AMDGPUCallLowering::lowerParameter(MachineIRBuilder &B, ArgInfo &OrigArg,
                                      uint64_t Offset,
                                      Align Alignment) const {
MachineFunction &MF = B.getMF();
const Function &F = MF.getFunction();
const DataLayout &DL = F.getParent()->getDataLayout();
MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS);

LLT PtrTy = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);

SmallVector<ArgInfo, 32> SplitArgs;
SmallVector<uint64_t> FieldOffsets;
splitToValueTypes(OrigArg, SplitArgs, DL, F.getCallingConv(), &FieldOffsets);

unsigned Idx = 0;
for (ArgInfo &SplitArg : SplitArgs) {
  Register PtrReg = B.getMRI()->createGenericVirtualRegister(PtrTy);
  lowerParameterPtr(PtrReg, B, Offset + FieldOffsets[Idx]);

  LLT ArgTy = getLLTForType(*SplitArg.Ty, DL);
  if (SplitArg.Flags[0].isPointer()) {
    // Compensate for losing pointeriness in splitValueTypes.
    LLT PtrTy = LLT::pointer(SplitArg.Flags[0].getPointerAddrSpace(),
                             ArgTy.getScalarSizeInBits());
    ArgTy = ArgTy.isVector() ? LLT::vector(ArgTy.getElementCount(), PtrTy)
                             : PtrTy;
  }

  MachineMemOperand *MMO = MF.getMachineMemOperand(
      PtrInfo,
      MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
          MachineMemOperand::MOInvariant,
      ArgTy, commonAlignment(Alignment, FieldOffsets[Idx]));

  assert(SplitArg.Regs.size() == 1)(static_cast <bool> (SplitArg.Regs.size() == 1) ? void (
0) : __assert_fail ("SplitArg.Regs.size() == 1", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 434, __extension__ __PRETTY_FUNCTION__));

  B.buildLoad(SplitArg.Regs[0], PtrReg, *MMO);
  ++Idx;
}
439}

441// Allocate special inputs passed in user SGPRs.
442static void allocateHSAUserSGPRs(CCState &CCInfo,
                               MachineIRBuilder &B,
                               MachineFunction &MF,
                               const SIRegisterInfo &TRI,
                               SIMachineFunctionInfo &Info) {
// FIXME: How should these inputs interact with inreg / custom SGPR inputs?
if (Info.hasPrivateSegmentBuffer()) {
  Register PrivateSegmentBufferReg = Info.addPrivateSegmentBuffer(TRI);
  MF.addLiveIn(PrivateSegmentBufferReg, &AMDGPU::SGPR_128RegClass);
  CCInfo.AllocateReg(PrivateSegmentBufferReg);
}

if (Info.hasDispatchPtr()) {
  Register DispatchPtrReg = Info.addDispatchPtr(TRI);
  MF.addLiveIn(DispatchPtrReg, &AMDGPU::SGPR_64RegClass);
  CCInfo.AllocateReg(DispatchPtrReg);
}

if (Info.hasQueuePtr()) {
  Register QueuePtrReg = Info.addQueuePtr(TRI);
  MF.addLiveIn(QueuePtrReg, &AMDGPU::SGPR_64RegClass);
  CCInfo.AllocateReg(QueuePtrReg);
}

if (Info.hasKernargSegmentPtr()) {
  MachineRegisterInfo &MRI = MF.getRegInfo();
  Register InputPtrReg = Info.addKernargSegmentPtr(TRI);
  const LLT P4 = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
  Register VReg = MRI.createGenericVirtualRegister(P4);
  MRI.addLiveIn(InputPtrReg, VReg);
  B.getMBB().addLiveIn(InputPtrReg);
  B.buildCopy(VReg, InputPtrReg);
  CCInfo.AllocateReg(InputPtrReg);
}

if (Info.hasDispatchID()) {
  Register DispatchIDReg = Info.addDispatchID(TRI);
  MF.addLiveIn(DispatchIDReg, &AMDGPU::SGPR_64RegClass);
  CCInfo.AllocateReg(DispatchIDReg);
}

if (Info.hasFlatScratchInit()) {
  Register FlatScratchInitReg = Info.addFlatScratchInit(TRI);
  MF.addLiveIn(FlatScratchInitReg, &AMDGPU::SGPR_64RegClass);
  CCInfo.AllocateReg(FlatScratchInitReg);
}

// TODO: Add GridWorkGroupCount user SGPRs when used. For now with HSA we read
// these from the dispatch pointer.
491}

493bool AMDGPUCallLowering::lowerFormalArgumentsKernel(
  MachineIRBuilder &B, const Function &F,
  ArrayRef<ArrayRef<Register>> VRegs) const {
MachineFunction &MF = B.getMF();
const GCNSubtarget *Subtarget = &MF.getSubtarget<GCNSubtarget>();
MachineRegisterInfo &MRI = MF.getRegInfo();
SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
const SIRegisterInfo *TRI = Subtarget->getRegisterInfo();
const SITargetLowering &TLI = *getTLI<SITargetLowering>();
const DataLayout &DL = F.getParent()->getDataLayout();

Info->allocateModuleLDSGlobal(F.getParent());

SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(F.getCallingConv(), F.isVarArg(), MF, ArgLocs, F.getContext());

allocateHSAUserSGPRs(CCInfo, B, MF, *TRI, *Info);

unsigned i = 0;
const Align KernArgBaseAlign(16);
const unsigned BaseOffset = Subtarget->getExplicitKernelArgOffset(F);
uint64_t ExplicitArgOffset = 0;

// TODO: Align down to dword alignment and extract bits for extending loads.
for (auto &Arg : F.args()) {
  const bool IsByRef = Arg.hasByRefAttr();
  Type *ArgTy = IsByRef ? Arg.getParamByRefType() : Arg.getType();
  unsigned AllocSize = DL.getTypeAllocSize(ArgTy);
  if (AllocSize == 0)
    continue;

  MaybeAlign ABIAlign = IsByRef ? Arg.getParamAlign() : None;
  if (!ABIAlign)
    ABIAlign = DL.getABITypeAlign(ArgTy);

  uint64_t ArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + BaseOffset;
  ExplicitArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + AllocSize;

  if (Arg.use_empty()) {
    ++i;
    continue;
  }

  Align Alignment = commonAlignment(KernArgBaseAlign, ArgOffset);

  if (IsByRef) {
    unsigned ByRefAS = cast<PointerType>(Arg.getType())->getAddressSpace();

    assert(VRegs[i].size() == 1 &&(static_cast <bool> (VRegs[i].size() == 1 && "expected only one register for byval pointers"
) ? void (0) : __assert_fail ("VRegs[i].size() == 1 && \"expected only one register for byval pointers\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 542, __extension__ __PRETTY_FUNCTION__))
           "expected only one register for byval pointers")(static_cast <bool> (VRegs[i].size() == 1 && "expected only one register for byval pointers"
) ? void (0) : __assert_fail ("VRegs[i].size() == 1 && \"expected only one register for byval pointers\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 542, __extension__ __PRETTY_FUNCTION__));
    if (ByRefAS == AMDGPUAS::CONSTANT_ADDRESS) {
      lowerParameterPtr(VRegs[i][0], B, ArgOffset);
    } else {
      const LLT ConstPtrTy = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
      Register PtrReg = MRI.createGenericVirtualRegister(ConstPtrTy);
      lowerParameterPtr(PtrReg, B, ArgOffset);

      B.buildAddrSpaceCast(VRegs[i][0], PtrReg);
    }
  } else {
    ArgInfo OrigArg(VRegs[i], Arg, i);
    const unsigned OrigArgIdx = i + AttributeList::FirstArgIndex;
    setArgFlags(OrigArg, OrigArgIdx, DL, F);
    lowerParameter(B, OrigArg, ArgOffset, Alignment);
  }

  ++i;
}

TLI.allocateSpecialEntryInputVGPRs(CCInfo, MF, *TRI, *Info);
TLI.allocateSystemSGPRs(CCInfo, MF, *Info, F.getCallingConv(), false);
return true;
565}

567bool AMDGPUCallLowering::lowerFormalArguments(
  MachineIRBuilder &B, const Function &F, ArrayRef<ArrayRef<Register>> VRegs,
  FunctionLoweringInfo &FLI) const {
CallingConv::ID CC = F.getCallingConv();

// The infrastructure for normal calling convention lowering is essentially
// useless for kernels. We want to avoid any kind of legalization or argument
// splitting.
if (CC == CallingConv::AMDGPU_KERNEL)
  return lowerFormalArgumentsKernel(B, F, VRegs);

const bool IsGraphics = AMDGPU::isGraphics(CC);
const bool IsEntryFunc = AMDGPU::isEntryFunctionCC(CC);

MachineFunction &MF = B.getMF();
MachineBasicBlock &MBB = B.getMBB();
MachineRegisterInfo &MRI = MF.getRegInfo();
SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
const GCNSubtarget &Subtarget = MF.getSubtarget<GCNSubtarget>();
const SIRegisterInfo *TRI = Subtarget.getRegisterInfo();
const DataLayout &DL = F.getParent()->getDataLayout();

Info->allocateModuleLDSGlobal(F.getParent());

SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CC, F.isVarArg(), MF, ArgLocs, F.getContext());

if (!IsEntryFunc) {
  Register ReturnAddrReg = TRI->getReturnAddressReg(MF);
  Register LiveInReturn = MF.addLiveIn(ReturnAddrReg,
                                       &AMDGPU::SGPR_64RegClass);
  MBB.addLiveIn(ReturnAddrReg);
  B.buildCopy(LiveInReturn, ReturnAddrReg);
}

if (Info->hasImplicitBufferPtr()) {
  Register ImplicitBufferPtrReg = Info->addImplicitBufferPtr(*TRI);
  MF.addLiveIn(ImplicitBufferPtrReg, &AMDGPU::SGPR_64RegClass);
  CCInfo.AllocateReg(ImplicitBufferPtrReg);
}

SmallVector<ArgInfo, 32> SplitArgs;
unsigned Idx = 0;
unsigned PSInputNum = 0;

// Insert the hidden sret parameter if the return value won't fit in the
// return registers.
if (!FLI.CanLowerReturn)
  insertSRetIncomingArgument(F, SplitArgs, FLI.DemoteRegister, MRI, DL);

for (auto &Arg : F.args()) {
  if (DL.getTypeStoreSize(Arg.getType()) == 0)
    continue;

  const bool InReg = Arg.hasAttribute(Attribute::InReg);

  // SGPR arguments to functions not implemented.
  if (!IsGraphics && InReg)
    return false;

  if (Arg.hasAttribute(Attribute::SwiftSelf) ||
      Arg.hasAttribute(Attribute::SwiftError) ||
      Arg.hasAttribute(Attribute::Nest))
    return false;

  if (CC == CallingConv::AMDGPU_PS && !InReg && PSInputNum <= 15) {
    const bool ArgUsed = !Arg.use_empty();
    bool SkipArg = !ArgUsed && !Info->isPSInputAllocated(PSInputNum);

    if (!SkipArg) {
      Info->markPSInputAllocated(PSInputNum);
      if (ArgUsed)
        Info->markPSInputEnabled(PSInputNum);
    }

    ++PSInputNum;

    if (SkipArg) {
      for (int I = 0, E = VRegs[Idx].size(); I != E; ++I)
        B.buildUndef(VRegs[Idx][I]);

      ++Idx;
      continue;
    }
  }

  ArgInfo OrigArg(VRegs[Idx], Arg, Idx);
  const unsigned OrigArgIdx = Idx + AttributeList::FirstArgIndex;
  setArgFlags(OrigArg, OrigArgIdx, DL, F);

  splitToValueTypes(OrigArg, SplitArgs, DL, CC);
  ++Idx;
}

// At least one interpolation mode must be enabled or else the GPU will
// hang.
//
// Check PSInputAddr instead of PSInputEnable. The idea is that if the user
// set PSInputAddr, the user wants to enable some bits after the compilation
// based on run-time states. Since we can't know what the final PSInputEna
// will look like, so we shouldn't do anything here and the user should take
// responsibility for the correct programming.
//
// Otherwise, the following restrictions apply:
// - At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled.
// - If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be
//   enabled too.
if (CC == CallingConv::AMDGPU_PS) {
  if ((Info->getPSInputAddr() & 0x7F) == 0 ||
      ((Info->getPSInputAddr() & 0xF) == 0 &&
       Info->isPSInputAllocated(11))) {
    CCInfo.AllocateReg(AMDGPU::VGPR0);
    CCInfo.AllocateReg(AMDGPU::VGPR1);
    Info->markPSInputAllocated(0);
    Info->markPSInputEnabled(0);
  }

  if (Subtarget.isAmdPalOS()) {
    // For isAmdPalOS, the user does not enable some bits after compilation
    // based on run-time states; the register values being generated here are
    // the final ones set in hardware. Therefore we need to apply the
    // workaround to PSInputAddr and PSInputEnable together.  (The case where
    // a bit is set in PSInputAddr but not PSInputEnable is where the frontend
    // set up an input arg for a particular interpolation mode, but nothing
    // uses that input arg. Really we should have an earlier pass that removes
    // such an arg.)
    unsigned PsInputBits = Info->getPSInputAddr() & Info->getPSInputEnable();
    if ((PsInputBits & 0x7F) == 0 ||
        ((PsInputBits & 0xF) == 0 &&
         (PsInputBits >> 11 & 1)))
      Info->markPSInputEnabled(
        countTrailingZeros(Info->getPSInputAddr(), ZB_Undefined));
  }
}

const SITargetLowering &TLI = *getTLI<SITargetLowering>();
CCAssignFn *AssignFn = TLI.CCAssignFnForCall(CC, F.isVarArg());

if (!MBB.empty())
  B.setInstr(*MBB.begin());

if (!IsEntryFunc) {
  // For the fixed ABI, pass workitem IDs in the last argument register.
  if (AMDGPUTargetMachine::EnableFixedFunctionABI)
    TLI.allocateSpecialInputVGPRsFixed(CCInfo, MF, *TRI, *Info);
}

IncomingValueAssigner Assigner(AssignFn);
if (!determineAssignments(Assigner, SplitArgs, CCInfo))
  return false;

FormalArgHandler Handler(B, MRI);
if (!handleAssignments(Handler, SplitArgs, CCInfo, ArgLocs, B))
  return false;

uint64_t StackOffset = Assigner.StackOffset;

if (!IsEntryFunc && !AMDGPUTargetMachine::EnableFixedFunctionABI) {
  // Special inputs come after user arguments.
  TLI.allocateSpecialInputVGPRs(CCInfo, MF, *TRI, *Info);
}

// Start adding system SGPRs.
if (IsEntryFunc) {
  TLI.allocateSystemSGPRs(CCInfo, MF, *Info, CC, IsGraphics);
} else {
  if (!Subtarget.enableFlatScratch())
    CCInfo.AllocateReg(Info->getScratchRSrcReg());
  TLI.allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
}

// When we tail call, we need to check if the callee's arguments will fit on
// the caller's stack. So, whenever we lower formal arguments, we should keep
// track of this information, since we might lower a tail call in this
// function later.
Info->setBytesInStackArgArea(StackOffset);

// Move back to the end of the basic block.
B.setMBB(MBB);

return true;
748}

750bool AMDGPUCallLowering::passSpecialInputs(MachineIRBuilder &MIRBuilder,
                                         CCState &CCInfo,
                                         SmallVectorImpl<std::pair<MCRegister, Register>> &ArgRegs,
                                         CallLoweringInfo &Info) const {
MachineFunction &MF = MIRBuilder.getMF();

const AMDGPUFunctionArgInfo *CalleeArgInfo
  = &AMDGPUArgumentUsageInfo::FixedABIFunctionInfo;

const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
const AMDGPUFunctionArgInfo &CallerArgInfo = MFI->getArgInfo();


// TODO: Unify with private memory register handling. This is complicated by
// the fact that at least in kernels, the input argument is not necessarily
// in the same location as the input.
AMDGPUFunctionArgInfo::PreloadedValue InputRegs[] = {
  AMDGPUFunctionArgInfo::DISPATCH_PTR,
  AMDGPUFunctionArgInfo::QUEUE_PTR,
  AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR,
  AMDGPUFunctionArgInfo::DISPATCH_ID,
  AMDGPUFunctionArgInfo::WORKGROUP_ID_X,
  AMDGPUFunctionArgInfo::WORKGROUP_ID_Y,
  AMDGPUFunctionArgInfo::WORKGROUP_ID_Z
};

MachineRegisterInfo &MRI = MF.getRegInfo();

const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
const AMDGPULegalizerInfo *LI
  = static_cast<const AMDGPULegalizerInfo*>(ST.getLegalizerInfo());

for (auto InputID : InputRegs) {
  const ArgDescriptor *OutgoingArg;
  const TargetRegisterClass *ArgRC;
  LLT ArgTy;

  std::tie(OutgoingArg, ArgRC, ArgTy) =
      CalleeArgInfo->getPreloadedValue(InputID);
  if (!OutgoingArg)
    continue;

  const ArgDescriptor *IncomingArg;
  const TargetRegisterClass *IncomingArgRC;
  std::tie(IncomingArg, IncomingArgRC, ArgTy) =
      CallerArgInfo.getPreloadedValue(InputID);
  assert(IncomingArgRC == ArgRC)(static_cast <bool> (IncomingArgRC == ArgRC) ? void (0)
 : __assert_fail ("IncomingArgRC == ArgRC", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 796, __extension__ __PRETTY_FUNCTION__));

  Register InputReg = MRI.createGenericVirtualRegister(ArgTy);

  if (IncomingArg) {
    LI->loadInputValue(InputReg, MIRBuilder, IncomingArg, ArgRC, ArgTy);
  } else {
    assert(InputID == AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR)(static_cast <bool> (InputID == AMDGPUFunctionArgInfo::
IMPLICIT_ARG_PTR) ? void (0) : __assert_fail ("InputID == AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR"
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 803, __extension__ __PRETTY_FUNCTION__));
    LI->getImplicitArgPtr(InputReg, MRI, MIRBuilder);
  }

  if (OutgoingArg->isRegister()) {
    ArgRegs.emplace_back(OutgoingArg->getRegister(), InputReg);
    if (!CCInfo.AllocateReg(OutgoingArg->getRegister()))
      report_fatal_error("failed to allocate implicit input argument");
  } else {
    LLVM_DEBUG(dbgs() << "Unhandled stack passed implicit input argument\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "Unhandled stack passed implicit input argument\n"
; } } while (false);
    return false;
  }
}

// Pack workitem IDs into a single register or pass it as is if already
// packed.
const ArgDescriptor *OutgoingArg;
const TargetRegisterClass *ArgRC;
LLT ArgTy;

std::tie(OutgoingArg, ArgRC, ArgTy) =
    CalleeArgInfo->getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_X);
if (!OutgoingArg)
  std::tie(OutgoingArg, ArgRC, ArgTy) =
      CalleeArgInfo->getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Y);
if (!OutgoingArg)
  std::tie(OutgoingArg, ArgRC, ArgTy) =
      CalleeArgInfo->getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Z);
if (!OutgoingArg)
  return false;

auto WorkitemIDX =
    CallerArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_X);
auto WorkitemIDY =
    CallerArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Y);
auto WorkitemIDZ =
    CallerArgInfo.getPreloadedValue(AMDGPUFunctionArgInfo::WORKITEM_ID_Z);

const ArgDescriptor *IncomingArgX = std::get<0>(WorkitemIDX);
const ArgDescriptor *IncomingArgY = std::get<0>(WorkitemIDY);
const ArgDescriptor *IncomingArgZ = std::get<0>(WorkitemIDZ);
const LLT S32 = LLT::scalar(32);

// If incoming ids are not packed we need to pack them.
// FIXME: Should consider known workgroup size to eliminate known 0 cases.
Register InputReg;
if (IncomingArgX && !IncomingArgX->isMasked() && CalleeArgInfo->WorkItemIDX) {
  InputReg = MRI.createGenericVirtualRegister(S32);
  LI->loadInputValue(InputReg, MIRBuilder, IncomingArgX,
                     std::get<1>(WorkitemIDX), std::get<2>(WorkitemIDX));
}

if (IncomingArgY && !IncomingArgY->isMasked() && CalleeArgInfo->WorkItemIDY) {
  Register Y = MRI.createGenericVirtualRegister(S32);
  LI->loadInputValue(Y, MIRBuilder, IncomingArgY, std::get<1>(WorkitemIDY),
                     std::get<2>(WorkitemIDY));

  Y = MIRBuilder.buildShl(S32, Y, MIRBuilder.buildConstant(S32, 10)).getReg(0);
  InputReg = InputReg ? MIRBuilder.buildOr(S32, InputReg, Y).getReg(0) : Y;
}

if (IncomingArgZ && !IncomingArgZ->isMasked() && CalleeArgInfo->WorkItemIDZ) {
  Register Z = MRI.createGenericVirtualRegister(S32);
  LI->loadInputValue(Z, MIRBuilder, IncomingArgZ, std::get<1>(WorkitemIDZ),
                     std::get<2>(WorkitemIDZ));

  Z = MIRBuilder.buildShl(S32, Z, MIRBuilder.buildConstant(S32, 20)).getReg(0);
  InputReg = InputReg ? MIRBuilder.buildOr(S32, InputReg, Z).getReg(0) : Z;
}

if (!InputReg) {
  InputReg = MRI.createGenericVirtualRegister(S32);

  // Workitem ids are already packed, any of present incoming arguments will
  // carry all required fields.
  ArgDescriptor IncomingArg = ArgDescriptor::createArg(
    IncomingArgX ? *IncomingArgX :
      IncomingArgY ? *IncomingArgY : *IncomingArgZ, ~0u);
  LI->loadInputValue(InputReg, MIRBuilder, &IncomingArg,
                     &AMDGPU::VGPR_32RegClass, S32);
}

if (OutgoingArg->isRegister()) {
  ArgRegs.emplace_back(OutgoingArg->getRegister(), InputReg);
  if (!CCInfo.AllocateReg(OutgoingArg->getRegister()))
    report_fatal_error("failed to allocate implicit input argument");
} else {
  LLVM_DEBUG(dbgs() << "Unhandled stack passed implicit input argument\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "Unhandled stack passed implicit input argument\n"
; } } while (false);
  return false;
}

return true;
895}

897/// Returns a pair containing the fixed CCAssignFn and the vararg CCAssignFn for
898/// CC.
899static std::pair<CCAssignFn *, CCAssignFn *>
900getAssignFnsForCC(CallingConv::ID CC, const SITargetLowering &TLI) {
return {TLI.CCAssignFnForCall(CC, false), TLI.CCAssignFnForCall(CC, true)};
902}

904static unsigned getCallOpcode(const MachineFunction &CallerF, bool IsIndirect,
                            bool IsTailCall) {
return IsTailCall ? AMDGPU::SI_TCRETURN : AMDGPU::SI_CALL;
907}

909// Add operands to call instruction to track the callee.
910static bool addCallTargetOperands(MachineInstrBuilder &CallInst,
                                MachineIRBuilder &MIRBuilder,
                                AMDGPUCallLowering::CallLoweringInfo &Info) {
if (Info.Callee.isReg()) {
  CallInst.addReg(Info.Callee.getReg());
  CallInst.addImm(0);
} else if (Info.Callee.isGlobal() && Info.Callee.getOffset() == 0) {
  // The call lowering lightly assumed we can directly encode a call target in
  // the instruction, which is not the case. Materialize the address here.
  const GlobalValue *GV = Info.Callee.getGlobal();
  auto Ptr = MIRBuilder.buildGlobalValue(
    LLT::pointer(GV->getAddressSpace(), 64), GV);
  CallInst.addReg(Ptr.getReg(0));
  CallInst.add(Info.Callee);
} else
  return false;

return true;
928}

930bool AMDGPUCallLowering::doCallerAndCalleePassArgsTheSameWay(
  CallLoweringInfo &Info, MachineFunction &MF,
  SmallVectorImpl<ArgInfo> &InArgs) const {
const Function &CallerF = MF.getFunction();
CallingConv::ID CalleeCC = Info.CallConv;
CallingConv::ID CallerCC = CallerF.getCallingConv();

// If the calling conventions match, then everything must be the same.
if (CalleeCC == CallerCC)
  return true;

const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();

// Make sure that the caller and callee preserve all of the same registers.
auto TRI = ST.getRegisterInfo();

const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);
const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC);
if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved))
  return false;

// Check if the caller and callee will handle arguments in the same way.
const SITargetLowering &TLI = *getTLI<SITargetLowering>();
CCAssignFn *CalleeAssignFnFixed;
CCAssignFn *CalleeAssignFnVarArg;
std::tie(CalleeAssignFnFixed, CalleeAssignFnVarArg) =
    getAssignFnsForCC(CalleeCC, TLI);

CCAssignFn *CallerAssignFnFixed;
CCAssignFn *CallerAssignFnVarArg;
std::tie(CallerAssignFnFixed, CallerAssignFnVarArg) =
    getAssignFnsForCC(CallerCC, TLI);

// FIXME: We are not accounting for potential differences in implicitly passed
// inputs, but only the fixed ABI is supported now anyway.
IncomingValueAssigner CalleeAssigner(CalleeAssignFnFixed,
                                     CalleeAssignFnVarArg);
IncomingValueAssigner CallerAssigner(CallerAssignFnFixed,
                                     CallerAssignFnVarArg);
return resultsCompatible(Info, MF, InArgs, CalleeAssigner, CallerAssigner);
970}

972bool AMDGPUCallLowering::areCalleeOutgoingArgsTailCallable(
  CallLoweringInfo &Info, MachineFunction &MF,
  SmallVectorImpl<ArgInfo> &OutArgs) const {
// If there are no outgoing arguments, then we are done.
if (OutArgs.empty())
  return true;

const Function &CallerF = MF.getFunction();
CallingConv::ID CalleeCC = Info.CallConv;
CallingConv::ID CallerCC = CallerF.getCallingConv();
const SITargetLowering &TLI = *getTLI<SITargetLowering>();

CCAssignFn *AssignFnFixed;
CCAssignFn *AssignFnVarArg;
std::tie(AssignFnFixed, AssignFnVarArg) = getAssignFnsForCC(CalleeCC, TLI);

// We have outgoing arguments. Make sure that we can tail call with them.
SmallVector<CCValAssign, 16> OutLocs;
CCState OutInfo(CalleeCC, false, MF, OutLocs, CallerF.getContext());
OutgoingValueAssigner Assigner(AssignFnFixed, AssignFnVarArg);

if (!determineAssignments(Assigner, OutArgs, OutInfo)) {
  LLVM_DEBUG(dbgs() << "... Could not analyze call operands.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Could not analyze call operands.\n"
; } } while (false);
  return false;
}

// Make sure that they can fit on the caller's stack.
const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
if (OutInfo.getNextStackOffset() > FuncInfo->getBytesInStackArgArea()) {
  LLVM_DEBUG(dbgs() << "... Cannot fit call operands on caller's stack.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Cannot fit call operands on caller's stack.\n"
; } } while (false);
  return false;
}

// Verify that the parameters in callee-saved registers match.
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
const SIRegisterInfo *TRI = ST.getRegisterInfo();
const uint32_t *CallerPreservedMask = TRI->getCallPreservedMask(MF, CallerCC);
MachineRegisterInfo &MRI = MF.getRegInfo();
return parametersInCSRMatch(MRI, CallerPreservedMask, OutLocs, OutArgs);
1011}

1013/// Return true if the calling convention is one that we can guarantee TCO for.
1014static bool canGuaranteeTCO(CallingConv::ID CC) {
return CC == CallingConv::Fast;
1016}

1018/// Return true if we might ever do TCO for calls with this calling convention.
1019static bool mayTailCallThisCC(CallingConv::ID CC) {
switch (CC) {
case CallingConv::C:
case CallingConv::AMDGPU_Gfx:
  return true;
default:
  return canGuaranteeTCO(CC);
}
1027}

1029bool AMDGPUCallLowering::isEligibleForTailCallOptimization(
  MachineIRBuilder &B, CallLoweringInfo &Info,
  SmallVectorImpl<ArgInfo> &InArgs, SmallVectorImpl<ArgInfo> &OutArgs) const {
// Must pass all target-independent checks in order to tail call optimize.
if (!Info.IsTailCall)
  return false;

MachineFunction &MF = B.getMF();
const Function &CallerF = MF.getFunction();
CallingConv::ID CalleeCC = Info.CallConv;
CallingConv::ID CallerCC = CallerF.getCallingConv();

const SIRegisterInfo *TRI = MF.getSubtarget<GCNSubtarget>().getRegisterInfo();
const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);
// Kernels aren't callable, and don't have a live in return address so it
// doesn't make sense to do a tail call with entry functions.
if (!CallerPreserved)
  return false;

if (!mayTailCallThisCC(CalleeCC)) {
  LLVM_DEBUG(dbgs() << "... Calling convention cannot be tail called.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Calling convention cannot be tail called.\n"
; } } while (false);
  return false;
}

if (any_of(CallerF.args(), [](const Argument &A) {
      return A.hasByValAttr() || A.hasSwiftErrorAttr();
    })) {
  LLVM_DEBUG(dbgs() << "... Cannot tail call from callers with byval "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Cannot tail call from callers with byval "
 "or swifterror arguments\n"; } } while (false)
                       "or swifterror arguments\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Cannot tail call from callers with byval "
 "or swifterror arguments\n"; } } while (false);
  return false;
}

// If we have -tailcallopt, then we're done.
if (MF.getTarget().Options.GuaranteedTailCallOpt)
  return canGuaranteeTCO(CalleeCC) && CalleeCC == CallerF.getCallingConv();

// Verify that the incoming and outgoing arguments from the callee are
// safe to tail call.
if (!doCallerAndCalleePassArgsTheSameWay(Info, MF, InArgs)) {
  LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Caller and callee have incompatible calling conventions.\n"
; } } while (false)
      dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Caller and callee have incompatible calling conventions.\n"
; } } while (false)
      << "... Caller and callee have incompatible calling conventions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Caller and callee have incompatible calling conventions.\n"
; } } while (false);
  return false;
}

if (!areCalleeOutgoingArgsTailCallable(Info, MF, OutArgs))
  return false;

LLVM_DEBUG(dbgs() << "... Call is eligible for tail call optimization.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "... Call is eligible for tail call optimization.\n"
; } } while (false);
return true;
1079}

1081// Insert outgoing implicit arguments for a call, by inserting copies to the
1082// implicit argument registers and adding the necessary implicit uses to the
1083// call instruction.
1084void AMDGPUCallLowering::handleImplicitCallArguments(
  MachineIRBuilder &MIRBuilder, MachineInstrBuilder &CallInst,
  const GCNSubtarget &ST, const SIMachineFunctionInfo &FuncInfo,
  ArrayRef<std::pair<MCRegister, Register>> ImplicitArgRegs) const {
if (!ST.enableFlatScratch()) {
  // Insert copies for the SRD. In the HSA case, this should be an identity
  // copy.
  auto ScratchRSrcReg = MIRBuilder.buildCopy(LLT::fixed_vector(4, 32),
                                             FuncInfo.getScratchRSrcReg());
  MIRBuilder.buildCopy(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, ScratchRSrcReg);
  CallInst.addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Implicit);
}

for (std::pair<MCRegister, Register> ArgReg : ImplicitArgRegs) {
  MIRBuilder.buildCopy((Register)ArgReg.first, ArgReg.second);
  CallInst.addReg(ArgReg.first, RegState::Implicit);
}
1101}

1103bool AMDGPUCallLowering::lowerTailCall(
  MachineIRBuilder &MIRBuilder, CallLoweringInfo &Info,
  SmallVectorImpl<ArgInfo> &OutArgs) const {
MachineFunction &MF = MIRBuilder.getMF();
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
const Function &F = MF.getFunction();
MachineRegisterInfo &MRI = MF.getRegInfo();
const SITargetLowering &TLI = *getTLI<SITargetLowering>();

// True when we're tail calling, but without -tailcallopt.
bool IsSibCall = !MF.getTarget().Options.GuaranteedTailCallOpt;

// Find out which ABI gets to decide where things go.
CallingConv::ID CalleeCC = Info.CallConv;
CCAssignFn *AssignFnFixed;
CCAssignFn *AssignFnVarArg;
std::tie(AssignFnFixed, AssignFnVarArg) = getAssignFnsForCC(CalleeCC, TLI);

MachineInstrBuilder CallSeqStart;
if (!IsSibCall)
  CallSeqStart = MIRBuilder.buildInstr(AMDGPU::ADJCALLSTACKUP);

unsigned Opc = getCallOpcode(MF, Info.Callee.isReg(), true);
auto MIB = MIRBuilder.buildInstrNoInsert(Opc);
if (!addCallTargetOperands(MIB, MIRBuilder, Info))
  return false;

// Byte offset for the tail call. When we are sibcalling, this will always
// be 0.
MIB.addImm(0);

// Tell the call which registers are clobbered.
const SIRegisterInfo *TRI = ST.getRegisterInfo();
const uint32_t *Mask = TRI->getCallPreservedMask(MF, CalleeCC);
MIB.addRegMask(Mask);

// FPDiff is the byte offset of the call's argument area from the callee's.
// Stores to callee stack arguments will be placed in FixedStackSlots offset
// by this amount for a tail call. In a sibling call it must be 0 because the
// caller will deallocate the entire stack and the callee still expects its
// arguments to begin at SP+0.
int FPDiff = 0;

// This will be 0 for sibcalls, potentially nonzero for tail calls produced
// by -tailcallopt. For sibcalls, the memory operands for the call are
// already available in the caller's incoming argument space.
unsigned NumBytes = 0;
if (!IsSibCall) {
  // We aren't sibcalling, so we need to compute FPDiff. We need to do this
  // before handling assignments, because FPDiff must be known for memory
  // arguments.
  unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
  SmallVector<CCValAssign, 16> OutLocs;
  CCState OutInfo(CalleeCC, false, MF, OutLocs, F.getContext());

  // FIXME: Not accounting for callee implicit inputs
  OutgoingValueAssigner CalleeAssigner(AssignFnFixed, AssignFnVarArg);
  if (!determineAssignments(CalleeAssigner, OutArgs, OutInfo))
    return false;

  // The callee will pop the argument stack as a tail call. Thus, we must
  // keep it 16-byte aligned.
  NumBytes = alignTo(OutInfo.getNextStackOffset(), ST.getStackAlignment());

  // FPDiff will be negative if this tail call requires more space than we
  // would automatically have in our incoming argument space. Positive if we
  // actually shrink the stack.
  FPDiff = NumReusableBytes - NumBytes;

  // The stack pointer must be 16-byte aligned at all times it's used for a
  // memory operation, which in practice means at *all* times and in
  // particular across call boundaries. Therefore our own arguments started at
  // a 16-byte aligned SP and the delta applied for the tail call should
  // satisfy the same constraint.
  assert(isAligned(ST.getStackAlignment(), FPDiff) &&(static_cast <bool> (isAligned(ST.getStackAlignment(), FPDiff
) && "unaligned stack on tail call") ? void (0) : __assert_fail
 ("isAligned(ST.getStackAlignment(), FPDiff) && \"unaligned stack on tail call\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 1179, __extension__ __PRETTY_FUNCTION__))
         "unaligned stack on tail call")(static_cast <bool> (isAligned(ST.getStackAlignment(), FPDiff
) && "unaligned stack on tail call") ? void (0) : __assert_fail
 ("isAligned(ST.getStackAlignment(), FPDiff) && \"unaligned stack on tail call\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp"
, 1179, __extension__ __PRETTY_FUNCTION__));
}

SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(Info.CallConv, Info.IsVarArg, MF, ArgLocs, F.getContext());

// We could pass MIB and directly add the implicit uses to the call
// now. However, as an aesthetic choice, place implicit argument operands
// after the ordinary user argument registers.
SmallVector<std::pair<MCRegister, Register>, 12> ImplicitArgRegs;

if (AMDGPUTargetMachine::EnableFixedFunctionABI &&
    Info.CallConv != CallingConv::AMDGPU_Gfx) {
  // With a fixed ABI, allocate fixed registers before user arguments.
  if (!passSpecialInputs(MIRBuilder, CCInfo, ImplicitArgRegs, Info))
    return false;
}

OutgoingValueAssigner Assigner(AssignFnFixed, AssignFnVarArg);

if (!determineAssignments(Assigner, OutArgs, CCInfo))
  return false;

// Do the actual argument marshalling.
AMDGPUOutgoingArgHandler Handler(MIRBuilder, MRI, MIB, true, FPDiff);
if (!handleAssignments(Handler, OutArgs, CCInfo, ArgLocs, MIRBuilder))
  return false;

handleImplicitCallArguments(MIRBuilder, MIB, ST, *FuncInfo, ImplicitArgRegs);

// If we have -tailcallopt, we need to adjust the stack. We'll do the call
// sequence start and end here.
if (!IsSibCall) {
  MIB->getOperand(1).setImm(FPDiff);
  CallSeqStart.addImm(NumBytes).addImm(0);
  // End the call sequence *before* emitting the call. Normally, we would
  // tidy the frame up after the call. However, here, we've laid out the
  // parameters so that when SP is reset, they will be in the correct
  // location.
  MIRBuilder.buildInstr(AMDGPU::ADJCALLSTACKDOWN).addImm(NumBytes).addImm(0);
}

// Now we can add the actual call instruction to the correct basic block.
MIRBuilder.insertInstr(MIB);

// If Callee is a reg, since it is used by a target specific
// instruction, it must have a register class matching the
// constraint of that instruction.

// FIXME: We should define regbankselectable call instructions to handle
// divergent call targets.
if (MIB->getOperand(0).isReg()) {
  MIB->getOperand(0).setReg(constrainOperandRegClass(
      MF, *TRI, MRI, *ST.getInstrInfo(), *ST.getRegBankInfo(), *MIB,
      MIB->getDesc(), MIB->getOperand(0), 0));
}

MF.getFrameInfo().setHasTailCall();
Info.LoweredTailCall = true;
return true;
1239}

1241bool AMDGPUCallLowering::lowerCall(MachineIRBuilder &MIRBuilder,
                                 CallLoweringInfo &Info) const {
if (Info.IsVarArg) {
  LLVM_DEBUG(dbgs() << "Variadic functions not implemented\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "Variadic functions not implemented\n"
; } } while (false);
  return false;
}

MachineFunction &MF = MIRBuilder.getMF();
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
const SIRegisterInfo *TRI = ST.getRegisterInfo();

const Function &F = MF.getFunction();
MachineRegisterInfo &MRI = MF.getRegInfo();
const SITargetLowering &TLI = *getTLI<SITargetLowering>();
const DataLayout &DL = F.getParent()->getDataLayout();

if (!AMDGPUTargetMachine::EnableFixedFunctionABI &&
    Info.CallConv != CallingConv::AMDGPU_Gfx) {
  LLVM_DEBUG(dbgs() << "Variable function ABI not implemented\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "Variable function ABI not implemented\n"
; } } while (false);
  return false;
}

SmallVector<ArgInfo, 8> OutArgs;
for (auto &OrigArg : Info.OrigArgs)
  splitToValueTypes(OrigArg, OutArgs, DL, Info.CallConv);

SmallVector<ArgInfo, 8> InArgs;
if (Info.CanLowerReturn && !Info.OrigRet.Ty->isVoidTy())
  splitToValueTypes(Info.OrigRet, InArgs, DL, Info.CallConv);

// If we can lower as a tail call, do that instead.
bool CanTailCallOpt =
    isEligibleForTailCallOptimization(MIRBuilder, Info, InArgs, OutArgs);

// We must emit a tail call if we have musttail.
if (Info.IsMustTailCall && !CanTailCallOpt) {
  LLVM_DEBUG(dbgs() << "Failed to lower musttail call as tail call\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-call-lowering")) { dbgs() << "Failed to lower musttail call as tail call\n"
; } } while (false);
  return false;
}

if (CanTailCallOpt)
  return lowerTailCall(MIRBuilder, Info, OutArgs);

// Find out which ABI gets to decide where things go.
CCAssignFn *AssignFnFixed;
CCAssignFn *AssignFnVarArg;
std::tie(AssignFnFixed, AssignFnVarArg) =
    getAssignFnsForCC(Info.CallConv, TLI);

MIRBuilder.buildInstr(AMDGPU::ADJCALLSTACKUP)
  .addImm(0)
  .addImm(0);

// Create a temporarily-floating call instruction so we can add the implicit
// uses of arg registers.
unsigned Opc = getCallOpcode(MF, Info.Callee.isReg(), false);

auto MIB = MIRBuilder.buildInstrNoInsert(Opc);
MIB.addDef(TRI->getReturnAddressReg(MF));

if (!addCallTargetOperands(MIB, MIRBuilder, Info))
  return false;

// Tell the call which registers are clobbered.
const uint32_t *Mask = TRI->getCallPreservedMask(MF, Info.CallConv);
MIB.addRegMask(Mask);

SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(Info.CallConv, Info.IsVarArg, MF, ArgLocs, F.getContext());

// We could pass MIB and directly add the implicit uses to the call
// now. However, as an aesthetic choice, place implicit argument operands
// after the ordinary user argument registers.
SmallVector<std::pair<MCRegister, Register>, 12> ImplicitArgRegs;

if (AMDGPUTargetMachine::EnableFixedFunctionABI &&
    Info.CallConv != CallingConv::AMDGPU_Gfx) {
  // With a fixed ABI, allocate fixed registers before user arguments.
  if (!passSpecialInputs(MIRBuilder, CCInfo, ImplicitArgRegs, Info))
    return false;
}

// Do the actual argument marshalling.
SmallVector<Register, 8> PhysRegs;

OutgoingValueAssigner Assigner(AssignFnFixed, AssignFnVarArg);
if (!determineAssignments(Assigner, OutArgs, CCInfo))
  return false;

AMDGPUOutgoingArgHandler Handler(MIRBuilder, MRI, MIB, false);
if (!handleAssignments(Handler, OutArgs, CCInfo, ArgLocs, MIRBuilder))
  return false;

const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

handleImplicitCallArguments(MIRBuilder, MIB, ST, *MFI, ImplicitArgRegs);

// Get a count of how many bytes are to be pushed on the stack.
unsigned NumBytes = CCInfo.getNextStackOffset();

// If Callee is a reg, since it is used by a target specific
// instruction, it must have a register class matching the
// constraint of that instruction.

// FIXME: We should define regbankselectable call instructions to handle
// divergent call targets.
if (MIB->getOperand(1).isReg()) {
  MIB->getOperand(1).setReg(constrainOperandRegClass(
      MF, *TRI, MRI, *ST.getInstrInfo(),
      *ST.getRegBankInfo(), *MIB, MIB->getDesc(), MIB->getOperand(1),
      1));
}

// Now we can add the actual call instruction to the correct position.
MIRBuilder.insertInstr(MIB);

// Finally we can copy the returned value back into its virtual-register. In
// symmetry with the arguments, the physical register must be an
// implicit-define of the call instruction.
if (Info.CanLowerReturn && !Info.OrigRet.Ty->isVoidTy()) {
  CCAssignFn *RetAssignFn = TLI.CCAssignFnForReturn(Info.CallConv,
                                                    Info.IsVarArg);
  IncomingValueAssigner Assigner(RetAssignFn);
  CallReturnHandler Handler(MIRBuilder, MRI, MIB);
  if (!determineAndHandleAssignments(Handler, Assigner, InArgs, MIRBuilder,
                                     Info.CallConv, Info.IsVarArg))
    return false;
}

uint64_t CalleePopBytes = NumBytes;

MIRBuilder.buildInstr(AMDGPU::ADJCALLSTACKDOWN)
          .addImm(0)
          .addImm(CalleePopBytes);

if (!Info.CanLowerReturn) {
  insertSRetLoads(MIRBuilder, Info.OrigRet.Ty, Info.OrigRet.Regs,
                  Info.DemoteRegister, Info.DemoteStackIndex);
}

return true;
1382}