NVPTXISelDAGToDAG.cpp

Go to the documentation of this file.

//===-- NVPTXISelDAGToDAG.cpp - A dag to dag inst selector for NVPTX ------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file defines an instruction selector for the NVPTX target.
//
//===----------------------------------------------------------------------===//
 
#include "MCTargetDesc/NVPTXBaseInfo.h"
#include "NVPTX.h"
#include "NVPTXISelLowering.h"
#include "NVPTXSelectionDAGInfo.h"
#include "NVPTXTargetMachine.h"
#include "NVPTXUtilities.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/ADT/StringSwitch.h"
#include "llvm/ADT/Twine.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/InlineAsm.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsNVPTX.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Metadata.h"
#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/FormatVariadic.h"
#include "llvm/Support/KnownFPClass.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/TargetParser/AtomicScope.h"
#include <optional>
 
using namespace llvm;
 
#define DEBUG_TYPE "nvptx-isel"
#define PASS_NAME "NVPTX DAG->DAG Pattern Instruction Selection"
 
static cl::opt<bool>
    EnableRsqrtOpt("nvptx-rsqrt-approx-opt", cl::init(true), cl::Hidden,
                   cl::desc("Enable reciprocal sqrt optimization"));
 
// FIXME: This is a WAR to recover lost performance from #155024.
// We still need to investigate the regression and find a more permanent
// solution.
static cl::opt<bool> EnableMADWide("nvptx-mad-wide-opt", cl::init(false),
                                   cl::Hidden,
                                   cl::desc("Enable MAD wide optimization"));
 
namespace {
 
struct NVPTXScopes {
  NVPTXScopes() = default;
  NVPTXScopes(LLVMContext &C, const Triple &T);
  NVPTX::Scope operator[](SyncScope::ID ID) const;
  bool empty() const;
 
private:
  SmallMapVector<SyncScope::ID, NVPTX::Scope, 8> Scopes{};
  LLVMContext *Context = nullptr;
};
 
enum class NVPTXMemCacheHintInstruction { Ld, St, Atom };
 
struct NVPTXMemCacheHintAccess {
  NVPTXMemCacheHintInstruction Instruction;
  NVPTX::AddressSpace AddrSpace;
  unsigned NumElts;
  unsigned EltWidth;
  bool IsVolatile;
};
 
struct NVPTXMemCacheHintOperands {
  SDValue EvictionAndPrefetchHint;
  SDValue CachePolicyReg;
};
 
class NVPTXDAGToDAGISel : public SelectionDAGISel {
  const NVPTXTargetMachine &TM;
 
  NVPTX::DivPrecisionLevel getDivF32Level(const SDNode *N) const;
  bool usePrecSqrtF32(const SDNode *N) const;
  bool useF32FTZ() const;
  bool allowFMA() const;
  bool doRsqrtOpt() const;
  bool doMADWideOpt() const;
 
  NVPTXScopes Scopes{};
 
public:
  NVPTXDAGToDAGISel() = delete;
 
  explicit NVPTXDAGToDAGISel(NVPTXTargetMachine &tm, CodeGenOptLevel OptLevel);
 
  bool runOnMachineFunction(MachineFunction &MF) override;
  const NVPTXSubtarget *Subtarget = nullptr;
 
  bool SelectInlineAsmMemoryOperand(const SDValue &Op,
                                    InlineAsm::ConstraintCode ConstraintID,
                                    std::vector<SDValue> &OutOps) override;
 
private:
// Include the pieces autogenerated from the target description.
#include "NVPTXGenDAGISel.inc"
 
  void Select(SDNode *N) override;
  bool tryIntrinsicChain(SDNode *N);
  bool tryIntrinsicVoid(SDNode *N);
  bool tryLoad(SDNode *N);
  bool tryLoadVector(SDNode *N);
  bool tryLDU(SDNode *N);
  bool tryLDG(MemSDNode *N);
  bool tryStore(SDNode *N);
  bool tryStoreVector(SDNode *N);
  bool tryFence(SDNode *N);
  bool tryBFE(SDNode *N);
  bool tryBF16ArithToFMA(SDNode *N);
  bool SelectSETP_F16X2(SDNode *N);
  bool SelectSETP_BF16X2(SDNode *N);
  bool tryUNPACK_VECTOR(SDNode *N);
  bool tryEXTRACT_VECTOR_ELEMENT(SDNode *N);
  void PostprocessISelDAG() override;
  void SelectV2I64toI128(SDNode *N);
  void SelectI128toV2I64(SDNode *N);
  void SelectTcgen05Ld(SDNode *N, bool hasOffset = false);
  void SelectTcgen05St(SDNode *N, bool hasOffset = false);
  void selectAtomicSwap128(SDNode *N);
 
  inline SDValue getI32Imm(unsigned Imm, const SDLoc &DL) {
    return CurDAG->getTargetConstant(Imm, DL, MVT::i32);
  }
  NVPTX::Ordering getMemOrder(const MemSDNode *N) const;
  NVPTX::Scope getAtomicScope(const MemSDNode *N) const;
 
  bool SelectADDR(SDValue Addr, SDValue &Base, SDValue &Offset);
  bool SelectFAbs(SDValue N, SDValue &Src);
  SDValue getPTXCmpMode(const CondCodeSDNode &CondCode);
  SDValue selectPossiblyImm(SDValue V);
 
  // Returns the encoded eviction/prefetch hint and cache policy register for a
  // memory operation. Hints unsupported by the subtarget or address space are
  // dropped. If L2::cache_hint is active, returns the hint with
  // L2CacheHintBit set and a register containing the 64-bit cache policy
  // value. Otherwise returns NOREG for the policy operand.
  NVPTXMemCacheHintOperands
  getMemCacheHintOperands(const MemSDNode *N, NVPTXMemCacheHintAccess Access,
                          const SDLoc &DL, bool EmitDiagnostics = true);
 
  // Returns the Memory Order and Scope that the PTX memory instruction should
  // use.
  std::pair<NVPTX::Ordering, NVPTX::Scope>
  getMemInstOrderingScope(MemSDNode *N);
 
  NVPTX::Scope getOperationScope(MemSDNode *N, NVPTX::Ordering O) const;
 
public:
  static NVPTX::AddressSpace getAddrSpace(const MemSDNode *N);
};
 
class NVPTXDAGToDAGISelLegacy : public SelectionDAGISelLegacy {
public:
  static char ID;
  explicit NVPTXDAGToDAGISelLegacy(NVPTXTargetMachine &tm,
                                   CodeGenOptLevel OptLevel);
};
 
} // end anonymous namespace
 
/// createNVPTXISelDag - This pass converts a legalized DAG into a
/// NVPTX-specific DAG, ready for instruction scheduling.
FunctionPass *llvm::createNVPTXISelDag(NVPTXTargetMachine &TM,
                                       llvm::CodeGenOptLevel OptLevel) {
  return new NVPTXDAGToDAGISelLegacy(TM, OptLevel);
}
 
NVPTXDAGToDAGISelLegacy::NVPTXDAGToDAGISelLegacy(NVPTXTargetMachine &tm,
                                                 CodeGenOptLevel OptLevel)
    : SelectionDAGISelLegacy(
          ID, std::make_unique<NVPTXDAGToDAGISel>(tm, OptLevel)) {}
 
char NVPTXDAGToDAGISelLegacy::ID = 0;
 
INITIALIZE_PASS(NVPTXDAGToDAGISelLegacy, DEBUG_TYPE, PASS_NAME, false, false)
 
NVPTXISelDAGToDAGPass::NVPTXISelDAGToDAGPass(NVPTXTargetMachine &TM,
                                             CodeGenOptLevel OptLevel)
    : SelectionDAGISelPass(std::make_unique<NVPTXDAGToDAGISel>(TM, OptLevel)) {}
 
NVPTXDAGToDAGISel::NVPTXDAGToDAGISel(NVPTXTargetMachine &tm,
                                     CodeGenOptLevel OptLevel)
    : SelectionDAGISel(tm, OptLevel), TM(tm) {}
 
bool NVPTXDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
  Subtarget = &MF.getSubtarget<NVPTXSubtarget>();
  Scopes = NVPTXScopes(MF.getFunction().getContext(),
                       MF.getTarget().getTargetTriple());
  return SelectionDAGISel::runOnMachineFunction(MF);
}
 
NVPTX::DivPrecisionLevel
NVPTXDAGToDAGISel::getDivF32Level(const SDNode *N) const {
  return Subtarget->getTargetLowering()->getDivF32Level(*MF, *N);
}
 
bool NVPTXDAGToDAGISel::usePrecSqrtF32(const SDNode *N) const {
  return Subtarget->getTargetLowering()->usePrecSqrtF32(N);
}
 
bool NVPTXDAGToDAGISel::useF32FTZ() const {
  return Subtarget->getTargetLowering()->useF32FTZ(*MF);
}
 
bool NVPTXDAGToDAGISel::allowFMA() const {
  const NVPTXTargetLowering *TL = Subtarget->getTargetLowering();
  return TL->allowFMA(*MF, OptLevel);
}
 
bool NVPTXDAGToDAGISel::doRsqrtOpt() const { return EnableRsqrtOpt; }
 
bool NVPTXDAGToDAGISel::doMADWideOpt() const { return EnableMADWide; }
 
/// Select - Select instructions not customized! Used for
/// expanded, promoted and normal instructions.
void NVPTXDAGToDAGISel::Select(SDNode *N) {
 
  if (N->isMachineOpcode()) {
    N->setNodeId(-1);
    return; // Already selected.
  }
 
  switch (N->getOpcode()) {
  case ISD::LOAD:
  case ISD::ATOMIC_LOAD:
  case NVPTXISD::MLoad:
    if (tryLoad(N))
      return;
    break;
  case ISD::STORE:
  case ISD::ATOMIC_STORE:
    if (tryStore(N))
      return;
    break;
  case ISD::ATOMIC_FENCE:
    if (tryFence(N))
      return;
    break;
  case NVPTXISD::UNPACK_VECTOR:
    tryUNPACK_VECTOR(N);
    return;
  case ISD::EXTRACT_VECTOR_ELT:
    if (tryEXTRACT_VECTOR_ELEMENT(N))
      return;
    break;
  case NVPTXISD::SETP_F16X2:
    SelectSETP_F16X2(N);
    return;
  case NVPTXISD::SETP_BF16X2:
    SelectSETP_BF16X2(N);
    return;
  case NVPTXISD::LoadV2:
  case NVPTXISD::LoadV4:
  case NVPTXISD::LoadV8:
    if (tryLoadVector(N))
      return;
    break;
  case NVPTXISD::LDUV2:
  case NVPTXISD::LDUV4:
    if (tryLDU(N))
      return;
    break;
  case NVPTXISD::StoreV2:
  case NVPTXISD::StoreV4:
  case NVPTXISD::StoreV8:
    if (tryStoreVector(N))
      return;
    break;
  case ISD::INTRINSIC_W_CHAIN:
    if (tryIntrinsicChain(N))
      return;
    break;
  case ISD::INTRINSIC_VOID:
    if (tryIntrinsicVoid(N))
      return;
    break;
  case ISD::AND:
  case ISD::SRA:
  case ISD::SRL:
    // Try to select BFE
    if (tryBFE(N))
      return;
    break;
  case ISD::CopyToReg: {
    if (N->getOperand(1).getValueType() == MVT::i128) {
      SelectV2I64toI128(N);
      return;
    }
    break;
  }
  case ISD::CopyFromReg: {
    if (N->getOperand(1).getValueType() == MVT::i128) {
      SelectI128toV2I64(N);
      return;
    }
    break;
  }
  case NVPTXISD::ATOMIC_CMP_SWAP_B128:
  case NVPTXISD::ATOMIC_SWAP_B128:
    selectAtomicSwap128(N);
    return;
  case ISD::FADD:
  case ISD::FMUL:
  case ISD::FSUB:
    if (tryBF16ArithToFMA(N))
      return;
    break;
  default:
    break;
  }
  SelectCode(N);
}
 
#define TCGEN05_LD_OPCODE(SHAPE, NUM)                                          \
  (enablePack ? NVPTX::TCGEN05_LD_##SHAPE##_##NUM##_PACK                       \
              : NVPTX::TCGEN05_LD_##SHAPE##_##NUM)
 
static unsigned getTcgen05LdOpcode(unsigned IID, bool enablePack) {
  switch (IID) {
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x1:
    return TCGEN05_LD_OPCODE(16x64b, x1);
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x2:
    return TCGEN05_LD_OPCODE(16x64b, x2);
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x4:
    return TCGEN05_LD_OPCODE(16x64b, x4);
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x8:
    return TCGEN05_LD_OPCODE(16x64b, x8);
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x16:
    return TCGEN05_LD_OPCODE(16x64b, x16);
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x32:
    return TCGEN05_LD_OPCODE(16x64b, x32);
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x64:
    return TCGEN05_LD_OPCODE(16x64b, x64);
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x128:
    return TCGEN05_LD_OPCODE(16x64b, x128);
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x1:
    return TCGEN05_LD_OPCODE(16x128b, x1);
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x2:
    return TCGEN05_LD_OPCODE(16x128b, x2);
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x4:
    return TCGEN05_LD_OPCODE(16x128b, x4);
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x8:
    return TCGEN05_LD_OPCODE(16x128b, x8);
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x16:
    return TCGEN05_LD_OPCODE(16x128b, x16);
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x32:
    return TCGEN05_LD_OPCODE(16x128b, x32);
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x64:
    return TCGEN05_LD_OPCODE(16x128b, x64);
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x1:
    return TCGEN05_LD_OPCODE(16x256b, x1);
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x2:
    return TCGEN05_LD_OPCODE(16x256b, x2);
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x4:
    return TCGEN05_LD_OPCODE(16x256b, x4);
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x8:
    return TCGEN05_LD_OPCODE(16x256b, x8);
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x16:
    return TCGEN05_LD_OPCODE(16x256b, x16);
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x32:
    return TCGEN05_LD_OPCODE(16x256b, x32);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x1:
    return TCGEN05_LD_OPCODE(16x32bx2, x1);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x2:
    return TCGEN05_LD_OPCODE(16x32bx2, x2);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x4:
    return TCGEN05_LD_OPCODE(16x32bx2, x4);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x8:
    return TCGEN05_LD_OPCODE(16x32bx2, x8);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x16:
    return TCGEN05_LD_OPCODE(16x32bx2, x16);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x32:
    return TCGEN05_LD_OPCODE(16x32bx2, x32);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x64:
    return TCGEN05_LD_OPCODE(16x32bx2, x64);
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x128:
    return TCGEN05_LD_OPCODE(16x32bx2, x128);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x1:
    return TCGEN05_LD_OPCODE(32x32b, x1);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x2:
    return TCGEN05_LD_OPCODE(32x32b, x2);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x4:
    return TCGEN05_LD_OPCODE(32x32b, x4);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x8:
    return TCGEN05_LD_OPCODE(32x32b, x8);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x16:
    return TCGEN05_LD_OPCODE(32x32b, x16);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x32:
    return TCGEN05_LD_OPCODE(32x32b, x32);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x64:
    return TCGEN05_LD_OPCODE(32x32b, x64);
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x128:
    return TCGEN05_LD_OPCODE(32x32b, x128);
  }
  llvm_unreachable("unhandled tcgen05.ld lowering");
}
 
void NVPTXDAGToDAGISel::SelectTcgen05Ld(SDNode *N, bool hasOffset) {
  if (!Subtarget->hasTcgen05InstSupport())
    report_fatal_error(
        "tcgen05.ld is not supported on this architecture variant");
 
  SDLoc DL(N);
  unsigned IID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
 
  if (hasOffset) {
    bool enablePack = cast<ConstantSDNode>(N->getOperand(4))->getZExtValue();
    auto OffsetNode = CurDAG->getTargetConstant(
        cast<ConstantSDNode>(N->getOperand(3))->getZExtValue(), DL, MVT::i32);
    ReplaceNode(N, CurDAG->getMachineNode(
                       getTcgen05LdOpcode(IID, enablePack), DL, N->getVTList(),
                       {N->getOperand(2), OffsetNode, N->getOperand(0)}));
  } else {
    bool enablePack = cast<ConstantSDNode>(N->getOperand(3))->getZExtValue();
    ReplaceNode(N, CurDAG->getMachineNode(
                       getTcgen05LdOpcode(IID, enablePack), DL, N->getVTList(),
                       {N->getOperand(2), N->getOperand(0)}));
  }
}
 
bool NVPTXDAGToDAGISel::tryIntrinsicChain(SDNode *N) {
  unsigned IID = N->getConstantOperandVal(1);
  switch (IID) {
  default:
    return false;
  case Intrinsic::nvvm_ldu_global_f:
  case Intrinsic::nvvm_ldu_global_i:
  case Intrinsic::nvvm_ldu_global_p:
    return tryLDU(N);
 
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x1:
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x2:
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x4:
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x8:
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x16:
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x32:
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x64:
  case Intrinsic::nvvm_tcgen05_ld_16x64b_x128:
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x1:
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x2:
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x4:
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x16:
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x32:
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x64:
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x1:
  case Intrinsic::nvvm_tcgen05_ld_16x128b_x8:
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x2:
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x4:
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x8:
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x16:
  case Intrinsic::nvvm_tcgen05_ld_16x256b_x32:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x1:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x2:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x4:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x8:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x16:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x32:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x64:
  case Intrinsic::nvvm_tcgen05_ld_32x32b_x128: {
    SelectTcgen05Ld(N);
    return true;
  }
 
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x1:
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x2:
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x4:
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x8:
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x16:
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x32:
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x64:
  case Intrinsic::nvvm_tcgen05_ld_16x32bx2_x128: {
    SelectTcgen05Ld(N, /* hasOffset */ true);
    return true;
  }
  }
}
 
// Map ISD:CONDCODE value to appropriate CmpMode expected by
// NVPTXInstPrinter::printCmpMode()
SDValue NVPTXDAGToDAGISel::getPTXCmpMode(const CondCodeSDNode &CondCode) {
  using NVPTX::PTXCmpMode::CmpMode;
  const unsigned PTXCmpMode = [](ISD::CondCode CC) {
    switch (CC) {
    default:
      llvm_unreachable("Unexpected condition code.");
    case ISD::SETOEQ:
    case ISD::SETEQ:
      return CmpMode::EQ;
    case ISD::SETOGT:
    case ISD::SETGT:
      return CmpMode::GT;
    case ISD::SETOGE:
    case ISD::SETGE:
      return CmpMode::GE;
    case ISD::SETOLT:
    case ISD::SETLT:
      return CmpMode::LT;
    case ISD::SETOLE:
    case ISD::SETLE:
      return CmpMode::LE;
    case ISD::SETONE:
    case ISD::SETNE:
      return CmpMode::NE;
    case ISD::SETO:
      return CmpMode::NUM;
    case ISD::SETUO:
      return CmpMode::NotANumber;
    case ISD::SETUEQ:
      return CmpMode::EQU;
    case ISD::SETUGT:
      return CmpMode::GTU;
    case ISD::SETUGE:
      return CmpMode::GEU;
    case ISD::SETULT:
      return CmpMode::LTU;
    case ISD::SETULE:
      return CmpMode::LEU;
    case ISD::SETUNE:
      return CmpMode::NEU;
    }
  }(CondCode.get());
  return CurDAG->getTargetConstant(PTXCmpMode, SDLoc(), MVT::i32);
}
 
bool NVPTXDAGToDAGISel::SelectSETP_F16X2(SDNode *N) {
  SDValue PTXCmpMode = getPTXCmpMode(*cast<CondCodeSDNode>(N->getOperand(2)));
  SDLoc DL(N);
  SDNode *SetP = CurDAG->getMachineNode(
      NVPTX::SETP_f16x2, DL, MVT::i1, MVT::i1,
      {N->getOperand(0), N->getOperand(1), PTXCmpMode,
       CurDAG->getTargetConstant(useF32FTZ() ? 1 : 0, DL, MVT::i1)});
  ReplaceNode(N, SetP);
  return true;
}
 
bool NVPTXDAGToDAGISel::SelectSETP_BF16X2(SDNode *N) {
  SDValue PTXCmpMode = getPTXCmpMode(*cast<CondCodeSDNode>(N->getOperand(2)));
  SDLoc DL(N);
  SDNode *SetP =
      CurDAG->getMachineNode(NVPTX::SETP_bf16x2, DL, MVT::i1, MVT::i1,
                             {N->getOperand(0), N->getOperand(1), PTXCmpMode});
  ReplaceNode(N, SetP);
  return true;
}
 
bool NVPTXDAGToDAGISel::tryUNPACK_VECTOR(SDNode *N) {
  SDValue Vector = N->getOperand(0);
  MVT EltVT = N->getSimpleValueType(0);
 
  MachineSDNode *N2 =
      CurDAG->getMachineNode(NVPTX::I64toV2I32, SDLoc(N), EltVT, EltVT, Vector);
 
  ReplaceNode(N, N2);
  return true;
}
 
// Find all instances of extract_vector_elt that use this v2f16 vector
// and coalesce them into a scattering move instruction.
bool NVPTXDAGToDAGISel::tryEXTRACT_VECTOR_ELEMENT(SDNode *N) {
  SDValue Vector = N->getOperand(0);
 
  MVT VT = Vector.getSimpleValueType();
  if (!(NVPTX::isPackedVectorTy(VT) && VT.getVectorNumElements() == 2))
    return false;
 
  unsigned Opcode;
  if (VT.is32BitVector())
    Opcode = NVPTX::I32toV2I16;
  else if (VT.is64BitVector())
    Opcode = NVPTX::I64toV2I32;
  else
    llvm_unreachable("Unhandled packed type");
 
  // Find and record all uses of this vector that extract element 0 or 1.
  SmallVector<SDNode *, 4> E0, E1;
  for (auto *U : Vector.getNode()->users()) {
    if (U->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      continue;
    if (U->getOperand(0) != Vector)
      continue;
    if (const ConstantSDNode *IdxConst =
            dyn_cast<ConstantSDNode>(U->getOperand(1))) {
      if (IdxConst->getZExtValue() == 0)
        E0.push_back(U);
      else if (IdxConst->getZExtValue() == 1)
        E1.push_back(U);
      else
        llvm_unreachable("Invalid vector index.");
    }
  }
 
  // There's no point scattering f16x2 if we only ever access one
  // element of it.
  if (E0.empty() || E1.empty())
    return false;
 
  // Merge (EltTy extractelt(V, 0), EltTy extractelt(V,1))
  // into EltTy,EltTy Split[EltTy]x2(V)
  MVT EltVT = VT.getVectorElementType();
  SDNode *ScatterOp =
      CurDAG->getMachineNode(Opcode, SDLoc(N), EltVT, EltVT, Vector);
  for (auto *Node : E0)
    ReplaceUses(SDValue(Node, 0), SDValue(ScatterOp, 0));
  for (auto *Node : E1)
    ReplaceUses(SDValue(Node, 0), SDValue(ScatterOp, 1));
 
  return true;
}
 
static std::optional<unsigned> getI16ExtractIndex(const SDNode *N) {
  if (!N->isMachineOpcode() || N->use_empty())
    return std::nullopt;
  switch (N->getMachineOpcode()) {
  default:
    return std::nullopt;
  case NVPTX::CVT_u16_u32:
    if (N->getConstantOperandVal(1) != NVPTX::PTXCvtMode::NONE)
      return std::nullopt;
    [[fallthrough]];
  case NVPTX::I32toI16L:
  case NVPTX::I32toI16L_Sink:
    return 0;
  case NVPTX::I32toI16H:
  case NVPTX::I32toI16H_Sink:
    return 1;
  }
}
 
// Merge scalar truncations and individual vector extracts after selection, when
// bitcasts no longer hide that they extract the same 16-bit halves.
static void mergeI16Unpacks(SelectionDAG &DAG) {
  auto Position = DAG.allnodes_begin();
  // Replacing extracts can CSE downstream nodes, including the next node.
  SelectionDAG::DAGNodeDeletedListener Listener(
      DAG, [&](SDNode *Deleted, SDNode *) {
        if (Position == SelectionDAG::allnodes_iterator(Deleted))
          ++Position;
      });
 
  bool Changed = false;
  while (Position != DAG.allnodes_end()) {
    SDNode *N = &*Position++;
    if (!getI16ExtractIndex(N))
      continue;
    SDValue Word = N->getOperand(0);
    SmallVector<SDValue, 2> Halves[2];
    for (SDNode *User : Word->users()) {
      std::optional<unsigned> Index = getI16ExtractIndex(User);
      if (Index && User->getOperand(0) == Word)
        Halves[*Index].emplace_back(User, 0);
    }
    if (Halves[0].empty() || Halves[1].empty())
      continue;
 
    SDNode *Scatter = DAG.getMachineNode(NVPTX::I32toV2I16, SDLoc(N), MVT::i16,
                                         MVT::i16, Word);
    for (unsigned Index : {0u, 1u})
      for (SDValue V : Halves[Index])
        DAG.ReplaceAllUsesOfValueWith(V, SDValue(Scatter, Index));
    Changed = true;
  }
  if (Changed)
    DAG.RemoveDeadNodes();
}
 
void NVPTXDAGToDAGISel::PostprocessISelDAG() { mergeI16Unpacks(*CurDAG); }
 
NVPTX::AddressSpace NVPTXDAGToDAGISel::getAddrSpace(const MemSDNode *N) {
  auto AS =
      static_cast<NVPTX::AddressSpace>(N->getMemOperand()->getAddrSpace());
  switch (AS) {
  case NVPTX::AddressSpace::Generic:
  case NVPTX::AddressSpace::Global:
  case NVPTX::AddressSpace::Shared:
  case NVPTX::AddressSpace::Const:
  case NVPTX::AddressSpace::Local:
  case NVPTX::AddressSpace::SharedCluster:
  case NVPTX::AddressSpace::EntryParam:
  case NVPTX::AddressSpace::DeviceParam:
    return AS;
  }
  llvm_unreachable("Unexpected address space");
}
 
NVPTX::Ordering NVPTXDAGToDAGISel::getMemOrder(const MemSDNode *N) const {
  // No "sem" orderings for SM/PTX versions which do not support memory ordering
  if (!Subtarget->hasMemoryOrdering())
    return NVPTX::Ordering::NotAtomic;
  auto Ordering = N->getMergedOrdering();
  switch (Ordering) {
  case AtomicOrdering::NotAtomic:
    return NVPTX::Ordering::NotAtomic;
  case AtomicOrdering::Unordered:
  case AtomicOrdering::Monotonic:
    return NVPTX::Ordering::Relaxed;
  case AtomicOrdering::Acquire:
    return NVPTX::Ordering::Acquire;
  case AtomicOrdering::Release:
    return NVPTX::Ordering::Release;
  case AtomicOrdering::AcquireRelease:
    return NVPTX::Ordering::AcquireRelease;
  case AtomicOrdering::SequentiallyConsistent:
    return NVPTX::Ordering::SequentiallyConsistent;
  }
  llvm_unreachable("Invalid atomic ordering");
}
 
// Clusters contain exactly 1 block on targets without cluster support.
static NVPTX::Scope resolveScope(NVPTX::Scope S, const NVPTXSubtarget *T) {
  if (S == NVPTX::Scope::Cluster && !T->hasClusters())
    return NVPTX::Scope::Block;
  return S;
}
 
NVPTX::Scope NVPTXDAGToDAGISel::getAtomicScope(const MemSDNode *N) const {
  NVPTX::Scope Scope = resolveScope(Scopes[N->getSyncScopeID()], Subtarget);
  if (!Subtarget->hasAtomScope()) {
    if (Scope == NVPTX::Scope::System)
      CurDAG->getContext()->diagnose(DiagnosticInfoUnsupported(
          CurDAG->getMachineFunction().getFunction(),
          "NVPTX system scope atomics require sm_60 or later",
          N->getDebugLoc()));
    return NVPTX::Scope::DefaultDevice;
  }
  return Scope;
}
 
namespace {
 
static NVPTX::Ordering getInstructionOrdering(MemSDNode *N,
                                              const NVPTXSubtarget *Subtarget) {
  AtomicOrdering Ordering = N->getSuccessOrdering();
  auto CodeAddrSpace = NVPTXDAGToDAGISel::getAddrSpace(N);
 
  bool HasMemoryOrdering = Subtarget->hasMemoryOrdering();
  bool HasRelaxedMMIO = Subtarget->hasRelaxedMMIO();
 
  // clang-format off
 
  // Lowering for Load/Store Operations (note: AcquireRelease Loads or Stores error).
  // Note: uses of Relaxed in the Atomic column of this table refer
  // to LLVM AtomicOrdering::Monotonic.
  //
  // | Atomic  | Volatile | Statespace         | PTX sm_60- | PTX sm_70+                   |
  // |---------|----------|--------------------|------------|------------------------------|
  // | No      | No       | All                | plain      | .weak                        |
  // | No      | Yes      | Generic,Shared,    | .volatile  | .volatile                    |
  // |         |          | Global [0]         |            |                              |
  // | No      | Yes      | Local (PTX 9.0-)   | plain [1]  | .weak [1]                    |
  // | No      | Yes      | Local (PTX 9.1+)   | .volatile  | .volatile                    |
  // | No      | Yes      | Const,Param        | plain [1]  | .weak [1]                    |
  // | Unorder | Yes/No   | All                | == Relaxed | == Relaxed                   |
  // | Relaxed | No       | Generic,Shared,    | .volatile  | <atomic sem>                 |
  // |         |          | Global [0]         |            |                              |
  // | Other   | No       | Generic,Shared,    | Error [2]  | <atomic sem>                 |
  // |         |          | Global [0]         |            |                              |
  // | Yes     | No       | Local,Const,Param  | plain [1]  | .weak [1]                    |
  // | Relaxed | Yes      | Generic,Shared [0] | .volatile  | .volatile                    |
  // | Relaxed | Yes      | Global [0]         | .volatile  | .mmio.relaxed.sys (PTX 8.2+) |
  // |         |          |                    |            |  or .volatile (PTX 8.1-)     |
  // | Yes     | Yes      | Local (PTX 9.0-)   | plain [1]  | .weak [1]                    |
  // | Yes     | Yes      | Local (PTX 9.1+)   | .volatile  | .volatile                    |
  // | Relaxed | Yes      | Const,Param        | plain [1]  | .weak [1]                    |
  // | Other   | Yes      | Generic, Shared,   | Error [2]  | <atomic sem> [3]             |
  // |         |          | / Global [0]       |            |                              |
 
  // Lowering of CUDA C++ SequentiallyConsistent Operations and Fences to PTX
  // by following the ABI proven sound in:
  //   Lustig et al, A Formal Analysis of the NVIDIA PTX Memory Consistency Model, ASPLOS’19.
  //   https://dl.acm.org/doi/pdf/10.1145/3297858.3304043
  //
  // | CUDA C++ Atomic Operation or Atomic Fence            | PTX Atomic Operation or Fence |
  // |------------------------------------------------------|-------------------------------|
  // | cuda::atomic_thread_fence                            | fence.sc.<scope>;             |
  // |   (memory_order_seq_cst, cuda::thread_scope_<scope>) |                               |
  // |------------------------------------------------------|-------------------------------|
  // | cuda::atomic_load                                    | fence.sc.<scope>;             |
  // |   (memory_order_seq_cst, cuda::thread_scope_<scope>) | ld.acquire.<scope>;           |
  // |------------------------------------------------------|-------------------------------|  
  // | cuda::atomic_store                                   | fence.sc.<scope>;             |
  // |   (memory_order_seq_cst, cuda::thread_scope_<scope>) | st.release.<scope>;           |
  // |------------------------------------------------------|-------------------------------|
  // | cuda::atomic_fetch_<op>                              | fence.sc.<scope>;             |
  // |   (memory_order_seq_cst, cuda::thread_scope_<scope>) | atom.acq_rel.<scope>;         |
 
  // clang-format on
 
  // [0]: volatile and atomics are only supported on global or shared
  //      memory locations, accessed via generic/shared/global pointers.
  //      PTX 9.1 adds volatile support on local ld/st.
  //      MMIO is only supported on global memory locations,
  //      accessed via generic/global pointers.
  // TODO: Implement MMIO access via generic pointer to global.
  //       Currently implemented for global pointers only.
 
  // [1]: Lowering volatile/atomic operations to non-volatile/non-atomic
  //      PTX instructions fails to preserve their C++ side-effects.
  //
  //      Example (https://github.com/llvm/llvm-project/issues/62057):
  //
  //          void example() {
  //              std::atomic<bool> True = true;
  //              while (True.load(std::memory_order_relaxed));
  //          }
  //
  //      A C++ program that calls "example" is well-defined: the infinite loop
  //      performs an atomic operation. By lowering volatile/atomics to
  //      "weak" memory operations, we are transforming the above into:
  //
  //          void undefined_behavior() {
  //              bool True = true;
  //              while (True);
  //          }
  //
  //      which exhibits undefined behavior in both C++ and PTX.
  //
  //      Calling "example" in CUDA C++ compiled for sm_60- exhibits undefined
  //      behavior due to lack of Independent Forward Progress. Lowering these
  //      to weak memory operations in sm_60- is therefore fine.
  //
  //      TODO: Where direct volatile or atomic operations are unsupported,
  //      preserve the side-effect using the weak memory instruction and
  //      another instruction, such as a dead dummy volatile load.
 
  if (CodeAddrSpace == NVPTX::AddressSpace::Local) {
    // Local memory is private to a thread. Drop atomic ordering but preserve
    // volatile accesses where supported.
    return Subtarget->hasLocalVolatile() && N->isVolatile()
               ? NVPTX::Ordering::Volatile
               : NVPTX::Ordering::NotAtomic;
  }
 
  if (CodeAddrSpace == NVPTX::AddressSpace::Const ||
      CodeAddrSpace == NVPTX::AddressSpace::EntryParam ||
      CodeAddrSpace == NVPTX::AddressSpace::DeviceParam) {
    return NVPTX::Ordering::NotAtomic;
  }
 
  // [2]: Atomics with Ordering different than Unordered or Relaxed are not
  //      supported on sm_60 and older; this includes volatile atomics.
  if (!(Ordering == AtomicOrdering::NotAtomic ||
        Ordering == AtomicOrdering::Unordered ||
        Ordering == AtomicOrdering::Monotonic) &&
      !HasMemoryOrdering) {
    report_fatal_error(
        formatv("PTX does not support \"atomic\" for orderings different than"
                "\"NotAtomic\" or \"Monotonic\" for sm_60 or older, but order "
                "is: \"{}\".",
                toIRString(Ordering)));
  }
 
  // [3]: TODO: these should eventually use .mmio<.atomic sem>; for now we drop
  // the volatile semantics and preserve the atomic ones.
 
  // Apart from local volatile accesses handled above, PTX atomics and volatile
  // operations are only available in generic, global, or shared memory.
  bool AddrSupportsVolatileOrAtomic =
      (CodeAddrSpace == NVPTX::AddressSpace::Generic ||
       CodeAddrSpace == NVPTX::AddressSpace::Global ||
       CodeAddrSpace == NVPTX::AddressSpace::Shared ||
       CodeAddrSpace == NVPTX::AddressSpace::SharedCluster);
  if (!AddrSupportsVolatileOrAtomic)
    return NVPTX::Ordering::NotAtomic;
 
  bool UseRelaxedMMIO =
      HasRelaxedMMIO && CodeAddrSpace == NVPTX::AddressSpace::Global;
 
  switch (Ordering) {
  case AtomicOrdering::NotAtomic:
    return N->isVolatile() ? NVPTX::Ordering::Volatile
                           : NVPTX::Ordering::NotAtomic;
  case AtomicOrdering::Unordered:
    // We lower unordered in the exact same way as 'monotonic' to respect
    // LLVM IR atomicity requirements.
  case AtomicOrdering::Monotonic:
    if (N->isVolatile())
      return UseRelaxedMMIO ? NVPTX::Ordering::RelaxedMMIO
                            : NVPTX::Ordering::Volatile;
    else
      return HasMemoryOrdering ? NVPTX::Ordering::Relaxed
                               : NVPTX::Ordering::Volatile;
  // case AtomicOrdering::Consume: // If LLVM ever provides this, lower it to
  // Acquire.
  case AtomicOrdering::Acquire:
    if (!N->readMem())
      report_fatal_error(
          formatv("PTX only supports Acquire Ordering on reads: {}",
                  N->getOperationName()));
    return NVPTX::Ordering::Acquire;
  case AtomicOrdering::Release:
    if (!N->writeMem())
      report_fatal_error(
          formatv("PTX only supports Release Ordering on writes: {}",
                  N->getOperationName()));
    return NVPTX::Ordering::Release;
  case AtomicOrdering::AcquireRelease: {
    report_fatal_error(
        formatv("NVPTX does not support AcquireRelease Ordering on "
                "read-modify-write "
                "yet and PTX does not support it on loads or stores: {}",
                N->getOperationName()));
  }
  case AtomicOrdering::SequentiallyConsistent: {
    // LLVM-IR SequentiallyConsistent atomics map to a two-instruction PTX
    // sequence including a "fence.sc.sco" and the memory instruction with an
    // Ordering that differs from "sc": acq, rel, or acq_rel, depending on
    // whether the memory operation is a read, write, or read-modify-write.
    //
    // This sets the ordering of the fence to SequentiallyConsistent, and
    // sets the corresponding ordering for the instruction.
    llvm_unreachable("seq_cst ordering should not reach this point, "
                     "is emulated in AtomicExpandPass");
  }
  }
  report_fatal_error(
      formatv("NVPTX backend does not support AtomicOrdering \"{}\" yet.",
              toIRString(Ordering)));
}
 
} // namespace
 
NVPTX::Scope NVPTXDAGToDAGISel::getOperationScope(MemSDNode *N,
                                                  NVPTX::Ordering O) const {
  switch (O) {
  case NVPTX::Ordering::NotAtomic:
  case NVPTX::Ordering::Volatile: // Non-atomic volatile operations
    // NVPTX uses Thread scope as the scope of non-atomic operations.
    return NVPTX::Scope::Thread;
  case NVPTX::Ordering::RelaxedMMIO:
    // RelaxedMMIO operations are always system scope.
    // If a RelaxedMMIO order was generated from an atomic volatile operation
    // with a smaller thread scope, we bump it here to system scope.
    return NVPTX::Scope::System;
  case NVPTX::Ordering::Relaxed:
  case NVPTX::Ordering::Acquire:
  case NVPTX::Ordering::Release:
  case NVPTX::Ordering::AcquireRelease:
  case NVPTX::Ordering::SequentiallyConsistent:
    auto S = Scopes[N->getSyncScopeID()];
 
    S = resolveScope(S, Subtarget);
 
    // If operation is volatile, then its scope is system.
    return N->isVolatile() ? NVPTX::Scope::System : S;
  }
  llvm_unreachable("unhandled ordering");
}
 
static bool canLowerToLDG(const MemSDNode &N, const NVPTXSubtarget &Subtarget,
                          NVPTX::AddressSpace CodeAddrSpace) {
  // We use ldg (i.e. ld.global.nc) for invariant loads from the global address
  // space.
  return Subtarget.hasLDG() && CodeAddrSpace == NVPTX::AddressSpace::Global &&
         N.isInvariant();
}
 
static unsigned int getFenceOp(NVPTX::Ordering O, NVPTX::Scope S,
                               NVPTXSubtarget const *T) {
  S = resolveScope(S, T);
 
  // Fall back to .acq_rel if .acquire, .release is not supported.
  if (!T->hasSplitAcquireAndReleaseFences() &&
      (O == NVPTX::Ordering::Acquire || O == NVPTX::Ordering::Release))
    O = NVPTX::Ordering::AcquireRelease;
 
  switch (O) {
  case NVPTX::Ordering::Acquire:
    switch (S) {
    case NVPTX::Scope::System:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acquire_sys
                                    : NVPTX::INT_MEMBAR_SYS;
    case NVPTX::Scope::Block:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acquire_cta
                                    : NVPTX::INT_MEMBAR_CTA;
    case NVPTX::Scope::Cluster:
      return NVPTX::atomic_thread_fence_acquire_cluster;
    case NVPTX::Scope::Device:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acquire_gpu
                                    : NVPTX::INT_MEMBAR_GL;
    case NVPTX::Scope::Thread:
    case NVPTX::Scope::DefaultDevice:
      report_fatal_error(
          formatv("Unsupported scope \"{}\" for acquire/release/acq_rel fence.",
                  ScopeToString(S)));
    }
    break;
  case NVPTX::Ordering::Release:
    switch (S) {
    case NVPTX::Scope::System:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_release_sys
                                    : NVPTX::INT_MEMBAR_SYS;
    case NVPTX::Scope::Block:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_release_cta
                                    : NVPTX::INT_MEMBAR_CTA;
    case NVPTX::Scope::Cluster:
      return NVPTX::atomic_thread_fence_release_cluster;
    case NVPTX::Scope::Device:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_release_gpu
                                    : NVPTX::INT_MEMBAR_GL;
    case NVPTX::Scope::Thread:
    case NVPTX::Scope::DefaultDevice:
      report_fatal_error(
          formatv("Unsupported scope \"{}\" for acquire/release/acq_rel fence.",
                  ScopeToString(S)));
    }
    break;
  case NVPTX::Ordering::AcquireRelease: {
    switch (S) {
    case NVPTX::Scope::System:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acq_rel_sys
                                    : NVPTX::INT_MEMBAR_SYS;
    case NVPTX::Scope::Block:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acq_rel_cta
                                    : NVPTX::INT_MEMBAR_CTA;
    case NVPTX::Scope::Cluster:
      return NVPTX::atomic_thread_fence_acq_rel_cluster;
    case NVPTX::Scope::Device:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acq_rel_gpu
                                    : NVPTX::INT_MEMBAR_GL;
    case NVPTX::Scope::Thread:
    case NVPTX::Scope::DefaultDevice:
      report_fatal_error(
          formatv("Unsupported scope \"{}\" for acquire/release/acq_rel fence.",
                  ScopeToString(S)));
    }
    break;
  }
  case NVPTX::Ordering::SequentiallyConsistent: {
    switch (S) {
    case NVPTX::Scope::System:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_seq_cst_sys
                                    : NVPTX::INT_MEMBAR_SYS;
    case NVPTX::Scope::Block:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_seq_cst_cta
                                    : NVPTX::INT_MEMBAR_CTA;
    case NVPTX::Scope::Cluster:
      return NVPTX::atomic_thread_fence_seq_cst_cluster;
    case NVPTX::Scope::Device:
      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_seq_cst_gpu
                                    : NVPTX::INT_MEMBAR_GL;
    case NVPTX::Scope::Thread:
    case NVPTX::Scope::DefaultDevice:
      report_fatal_error(formatv("Unsupported scope \"{}\" for seq_cst fence.",
                                 ScopeToString(S)));
    }
    break;
  }
  case NVPTX::Ordering::NotAtomic:
  case NVPTX::Ordering::Relaxed:
  case NVPTX::Ordering::Volatile:
  case NVPTX::Ordering::RelaxedMMIO:
    report_fatal_error(
        formatv("Unsupported \"{}\" ordering and \"{}\" scope for fence.",
                OrderingToString(O), ScopeToString(S)));
  }
  llvm_unreachable("unhandled ordering");
}
 
// Returns Memory Order and Scope of a memory instruction.
std::pair<NVPTX::Ordering, NVPTX::Scope>
NVPTXDAGToDAGISel::getMemInstOrderingScope(MemSDNode *N) {
  auto InstructionOrdering = getInstructionOrdering(N, Subtarget);
  auto Scope = getOperationScope(N, InstructionOrdering);
 
  // Singlethread scope has no inter-thread synchronization requirements, so
  // the atomic operation is lowered as plain and the fence is skipped.
  // NotAtomic and Volatile operations naturally have Thread scope and must
  // preserve their ordering.
  if (Scope == NVPTX::Scope::Thread &&
      InstructionOrdering != NVPTX::Ordering::NotAtomic &&
      InstructionOrdering != NVPTX::Ordering::Volatile)
    return {NVPTX::Ordering::NotAtomic, Scope};
 
  return {InstructionOrdering, Scope};
}
 
// Helper function template to reduce amount of boilerplate code for
// opcode selection.
static std::optional<unsigned>
pickOpcodeForVT(MVT::SimpleValueType VT, std::optional<unsigned> Opcode_i16,
                std::optional<unsigned> Opcode_i32,
                std::optional<unsigned> Opcode_i64) {
  switch (VT) {
  case MVT::f16:
  case MVT::i16:
  case MVT::bf16:
    return Opcode_i16;
  case MVT::v2f16:
  case MVT::v2bf16:
  case MVT::v2i16:
  case MVT::v4i8:
  case MVT::i32:
  case MVT::f32:
    return Opcode_i32;
  case MVT::v2f32:
  case MVT::v2i32:
  case MVT::i64:
  case MVT::f64:
    return Opcode_i64;
  default:
    return std::nullopt;
  }
}
 
static inline bool isAddLike(const SDValue V) {
  return V.getOpcode() == ISD::ADD ||
         (V->getOpcode() == ISD::OR && V->getFlags().hasDisjoint());
}
 
static SDValue stripAssertAlign(SDValue N) {
  if (N.getOpcode() == ISD::AssertAlign)
    N = N.getOperand(0);
  return N;
}
 
// selectBaseADDR - Match a dag node which will serve as the base address for an
// ADDR operand pair.
static SDValue selectBaseADDR(SDValue N, SelectionDAG *DAG) {
  N = stripAssertAlign(N);
  if (const auto *GA = dyn_cast<GlobalAddressSDNode>(N))
    return DAG->getTargetGlobalAddress(GA->getGlobal(), SDLoc(N),
                                       GA->getValueType(0), GA->getOffset(),
                                       GA->getTargetFlags());
  if (const auto *ES = dyn_cast<ExternalSymbolSDNode>(N))
    return DAG->getTargetExternalSymbol(ES->getSymbol(), ES->getValueType(0),
                                        ES->getTargetFlags());
  if (const auto *FIN = dyn_cast<FrameIndexSDNode>(N))
    return DAG->getTargetFrameIndex(FIN->getIndex(), FIN->getValueType(0));
  if (N.getOpcode() == NVPTXISD::Symbol)
    return N.getOperand(0);
 
  return N;
}
 
static SDValue accumulateOffset(SDValue &Addr, SDLoc DL, SelectionDAG *DAG) {
  Addr = stripAssertAlign(Addr);
  APInt AccumulatedOffset(64u, 0);
  while (isAddLike(Addr)) {
    const auto *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1));
    if (!CN)
      break;
 
    const APInt CI = CN->getAPIntValue().sext(64);
    if (!(CI + AccumulatedOffset).isSignedIntN(32))
      break;
 
    AccumulatedOffset += CI;
    Addr = stripAssertAlign(Addr->getOperand(0));
  }
  return DAG->getSignedTargetConstant(AccumulatedOffset.getSExtValue(), DL,
                                      MVT::i32);
}
 
static std::pair<SDValue, SDValue> selectADDR(SDValue Addr, SelectionDAG *DAG) {
  SDValue Offset = accumulateOffset(Addr, SDLoc(Addr), DAG);
  SDValue Base = selectBaseADDR(Addr, DAG);
  return {Base, Offset};
}
 
// Select a pair of operands which represent a valid PTX address, this could be
// one of the following things:
//  - [var] - Offset is simply set to 0
//  - [reg] - Offset is simply set to 0
//  - [reg+immOff]
//  - [var+immOff]
// Note that immOff must fit into a 32-bit signed integer.
bool NVPTXDAGToDAGISel::SelectADDR(SDValue Addr, SDValue &Base,
                                   SDValue &Offset) {
  std::tie(Base, Offset) = selectADDR(Addr, CurDAG);
  return true;
}
 
static void emitInvalidMemCacheHint(LLVMContext &Ctx, const Twine &Msg) {
  Ctx.diagnose(DiagnosticInfoGeneric(
      Twine("invalid NVPTX !mem.cache_hint metadata: ") + Msg, DS_Warning));
}
 
static std::optional<NVPTX::L1Eviction> parseL1Eviction(StringRef Str) {
  return StringSwitch<std::optional<NVPTX::L1Eviction>>(Str)
      .Case("normal", NVPTX::L1Eviction::Normal)
      .Case("unchanged", NVPTX::L1Eviction::Unchanged)
      .Case("first", NVPTX::L1Eviction::First)
      .Case("last", NVPTX::L1Eviction::Last)
      .Case("no_allocate", NVPTX::L1Eviction::NoAllocate)
      .Default(std::nullopt);
}
 
static std::optional<NVPTX::L2Eviction> parseL2Eviction(StringRef Str) {
  return StringSwitch<std::optional<NVPTX::L2Eviction>>(Str)
      .Case("normal", NVPTX::L2Eviction::Normal)
      .Case("first", NVPTX::L2Eviction::First)
      .Case("last", NVPTX::L2Eviction::Last)
      .Default(std::nullopt);
}
 
static std::optional<NVPTX::L2Prefetch> parseL2Prefetch(StringRef Str) {
  return StringSwitch<std::optional<NVPTX::L2Prefetch>>(Str)
      .Case("64B", NVPTX::L2Prefetch::Bytes64)
      .Case("128B", NVPTX::L2Prefetch::Bytes128)
      .Case("256B", NVPTX::L2Prefetch::Bytes256)
      .Default(std::nullopt);
}
 
template <typename T>
static std::optional<T> parseMemCacheHintStringValue(
    LLVMContext &Ctx, StringRef Key, const Metadata *Value,
    std::optional<T> (*Parse)(StringRef), bool EmitDiagnostics) {
  const auto *Val = dyn_cast<MDString>(Value);
  if (!Val) {
    if (EmitDiagnostics)
      emitInvalidMemCacheHint(Ctx,
                              Twine("'") + Key + "' expects a string value");
    return std::nullopt;
  }
 
  StringRef ValStr = Val->getString();
  auto Parsed = Parse(ValStr);
  if (!Parsed && EmitDiagnostics)
    emitInvalidMemCacheHint(Ctx, Twine("unknown value '") + ValStr + "' for '" +
                                     Key + "'");
  return Parsed;
}
 
static bool isGlobalOrGeneric(NVPTX::AddressSpace AddrSpace) {
  return AddrSpace == NVPTX::AddressSpace::Global ||
         AddrSpace == NVPTX::AddressSpace::Generic;
}
 
static bool isLdOrSt(NVPTXMemCacheHintAccess Access) {
  return Access.Instruction == NVPTXMemCacheHintInstruction::Ld ||
         Access.Instruction == NVPTXMemCacheHintInstruction::St;
}
 
static bool isL1EvictionSupported(const NVPTXSubtarget &Subtarget,
                                  NVPTX::L1Eviction Eviction,
                                  NVPTXMemCacheHintAccess Access) {
  if (Eviction == NVPTX::L1Eviction::Normal)
    return true;
 
  return isLdOrSt(Access) && !Access.IsVolatile &&
         Subtarget.hasL1EvictionHint();
}
 
static bool isL2PrefetchSupported(const NVPTXSubtarget &Subtarget,
                                  NVPTX::L2Prefetch Prefetch,
                                  NVPTXMemCacheHintAccess Access) {
  switch (Prefetch) {
  case NVPTX::L2Prefetch::None:
    return true;
  case NVPTX::L2Prefetch::Bytes64:
    return Access.Instruction == NVPTXMemCacheHintInstruction::Ld &&
           isGlobalOrGeneric(Access.AddrSpace) && Subtarget.hasL2Prefetch64B();
  case NVPTX::L2Prefetch::Bytes128:
    return Access.Instruction == NVPTXMemCacheHintInstruction::Ld &&
           isGlobalOrGeneric(Access.AddrSpace) && Subtarget.hasL2Prefetch128B();
  case NVPTX::L2Prefetch::Bytes256:
    return Access.Instruction == NVPTXMemCacheHintInstruction::Ld &&
           isGlobalOrGeneric(Access.AddrSpace) && Subtarget.hasL2Prefetch256B();
  }
  llvm_unreachable("Unexpected L2 prefetch hint");
}
 
static bool isL2EvictionSupported(const NVPTXSubtarget &Subtarget,
                                  NVPTX::L2Eviction Eviction,
                                  NVPTXMemCacheHintAccess Access) {
  if (Eviction == NVPTX::L2Eviction::Normal)
    return true;
 
  return isLdOrSt(Access) && !Access.IsVolatile &&
         Subtarget.hasL2EvictionHint() && isGlobalOrGeneric(Access.AddrSpace) &&
         ((Access.NumElts == 8 && Access.EltWidth == 32) ||
          (Access.NumElts == 4 && Access.EltWidth == 64));
}
 
static bool isCachePolicySupported(const NVPTXSubtarget &Subtarget,
                                   NVPTXMemCacheHintAccess Access) {
  return !Access.IsVolatile && isGlobalOrGeneric(Access.AddrSpace) &&
         Subtarget.hasL2CacheHint();
}
 
NVPTXMemCacheHintOperands NVPTXDAGToDAGISel::getMemCacheHintOperands(
    const MemSDNode *N, NVPTXMemCacheHintAccess Access, const SDLoc &DL,
    bool EmitDiagnostics) {
  LLVMContext &Ctx = *CurDAG->getContext();
  const MDNode *Node = N->getMemCacheHint();
  SDValue PolicyReg = CurDAG->getRegister(NVPTX::NoRegister, MVT::i64);
  if (!Node)
    return {getI32Imm(0, DL), PolicyReg};
  if (Node->getNumOperands() == 0) {
    if (EmitDiagnostics)
      emitInvalidMemCacheHint(Ctx, "empty hint node");
    return {getI32Imm(0, DL), PolicyReg};
  }
 
  NVPTX::L1Eviction L1 = NVPTX::L1Eviction::Normal;
  NVPTX::L2Eviction L2 = NVPTX::L2Eviction::Normal;
  NVPTX::L2Prefetch Prefetch = NVPTX::L2Prefetch::None;
  std::optional<uint64_t> CachePolicy;
 
  for (unsigned I = 0; I + 1 < Node->getNumOperands(); I += 2) {
    const auto *Key = cast<MDString>(Node->getOperand(I));
    StringRef KeyStr = Key->getString();
    const Metadata *Value = Node->getOperand(I + 1).get();
 
    if (KeyStr == "nvvm.l1_eviction") {
      auto ParsedL1 = parseMemCacheHintStringValue(
          Ctx, KeyStr, Value, parseL1Eviction, EmitDiagnostics);
      if (ParsedL1 && isL1EvictionSupported(*Subtarget, *ParsedL1, Access))
        L1 = *ParsedL1;
      continue;
    }
 
    if (KeyStr == "nvvm.l2_eviction") {
      auto ParsedL2 = parseMemCacheHintStringValue(
          Ctx, KeyStr, Value, parseL2Eviction, EmitDiagnostics);
      if (ParsedL2 && isL2EvictionSupported(*Subtarget, *ParsedL2, Access))
        L2 = *ParsedL2;
      continue;
    }
 
    if (KeyStr == "nvvm.l2_prefetch_size") {
      auto ParsedPrefetch = parseMemCacheHintStringValue(
          Ctx, KeyStr, Value, parseL2Prefetch, EmitDiagnostics);
      if (ParsedPrefetch &&
          isL2PrefetchSupported(*Subtarget, *ParsedPrefetch, Access))
        Prefetch = *ParsedPrefetch;
      continue;
    }
 
    if (KeyStr == "nvvm.l2_cache_hint") {
      const auto *ValCI = mdconst::dyn_extract<ConstantInt>(Value);
      if (!ValCI) {
        if (EmitDiagnostics)
          emitInvalidMemCacheHint(
              Ctx, "'nvvm.l2_cache_hint' expects an integer value");
      } else if (isCachePolicySupported(*Subtarget, Access)) {
        CachePolicy = ValCI->getZExtValue();
      }
      continue;
    }
 
    if (EmitDiagnostics)
      emitInvalidMemCacheHint(Ctx, Twine("unknown key '") + KeyStr + "'");
  }
 
  unsigned EvictionAndPrefetchHint =
      NVPTX::encodeEvictionAndPrefetchHint(L1, L2, Prefetch);
  if (CachePolicy) {
    SDValue PolicyConst = CurDAG->getTargetConstant(*CachePolicy, DL, MVT::i64);
    PolicyReg = SDValue(
        CurDAG->getMachineNode(NVPTX::MOV_B64_i, DL, MVT::i64, PolicyConst), 0);
    Bitfield::set<NVPTX::L2CacheHintBit>(EvictionAndPrefetchHint, true);
  }
 
  return {getI32Imm(EvictionAndPrefetchHint, DL), PolicyReg};
}
 
bool NVPTXDAGToDAGISel::tryLoad(SDNode *N) {
  MemSDNode *LD = cast<MemSDNode>(N);
  assert(LD->readMem() && "Expected load");
 
  // do not support pre/post inc/dec
  const LoadSDNode *PlainLoad = dyn_cast<LoadSDNode>(LD);
  if (PlainLoad && PlainLoad->isIndexed())
    return false;
 
  // Address Space Setting
  const auto CodeAddrSpace = getAddrSpace(LD);
  if (canLowerToLDG(*LD, *Subtarget, CodeAddrSpace))
    return tryLDG(LD);
 
  SDLoc DL(LD);
  SDValue Chain = N->getOperand(0);
  const auto [Ordering, Scope] = getMemInstOrderingScope(LD);
 
  const unsigned FromTypeWidth = LD->getMemoryVT().getSizeInBits();
 
  // Vector Setting
  const unsigned FromType =
      (PlainLoad && (PlainLoad->getExtensionType() == ISD::SEXTLOAD))
          ? NVPTX::PTXLdStInstCode::Signed
          : NVPTX::PTXLdStInstCode::Untyped;
 
  uint32_t UsedBytesMask;
  switch (N->getOpcode()) {
  case ISD::LOAD:
  case ISD::ATOMIC_LOAD:
    UsedBytesMask = UINT32_MAX;
    break;
  case NVPTXISD::MLoad:
    UsedBytesMask = N->getConstantOperandVal(3);
    break;
  default:
    llvm_unreachable("Unexpected opcode");
  }
 
  assert(isPowerOf2_32(FromTypeWidth) && FromTypeWidth >= 8 &&
         FromTypeWidth <= 128 && "Invalid width for load");
 
  const auto [Base, Offset] = selectADDR(N->getOperand(1), CurDAG);
  const auto [EvictionAndPrefetchHint, PolicyReg] = getMemCacheHintOperands(
      LD,
      {NVPTXMemCacheHintInstruction::Ld, CodeAddrSpace,
       /*NumElts=*/1, /*EltWidth=*/FromTypeWidth, LD->isVolatile()},
      DL);
 
  // Create the machine instruction DAG
  SDValue Ops[] = {getI32Imm(Ordering, DL),
                   getI32Imm(Scope, DL),
                   getI32Imm(CodeAddrSpace, DL),
                   getI32Imm(FromType, DL),
                   getI32Imm(FromTypeWidth, DL),
                   getI32Imm(UsedBytesMask, DL),
                   Base,
                   Offset,
                   EvictionAndPrefetchHint,
                   PolicyReg,
                   Chain};
 
  const MVT::SimpleValueType TargetVT = LD->getSimpleValueType(0).SimpleTy;
  const std::optional<unsigned> Opcode =
      pickOpcodeForVT(TargetVT, NVPTX::LD_i16, NVPTX::LD_i32, NVPTX::LD_i64);
  if (!Opcode)
    return false;
 
  SDNode *NVPTXLD = CurDAG->getMachineNode(*Opcode, DL, LD->getVTList(), Ops);
  if (!NVPTXLD)
    return false;
 
  MachineMemOperand *MemRef = LD->getMemOperand();
  CurDAG->setNodeMemRefs(cast<MachineSDNode>(NVPTXLD), {MemRef});
 
  ReplaceNode(LD, NVPTXLD);
  return true;
}
 
static unsigned getStoreVectorNumElts(SDNode *N) {
  switch (N->getOpcode()) {
  case NVPTXISD::StoreV2:
    return 2;
  case NVPTXISD::StoreV4:
    return 4;
  case NVPTXISD::StoreV8:
    return 8;
  default:
    llvm_unreachable("Unexpected opcode");
  }
}
 
bool NVPTXDAGToDAGISel::tryLoadVector(SDNode *N) {
  MemSDNode *LD = cast<MemSDNode>(N);
 
  // Address Space Setting
  const auto CodeAddrSpace = getAddrSpace(LD);
  if (canLowerToLDG(*LD, *Subtarget, CodeAddrSpace))
    return tryLDG(LD);
 
  const MVT EltVT = LD->getSimpleValueType(0);
  SDLoc DL(LD);
  SDValue Chain = LD->getChain();
  const auto [Ordering, Scope] = getMemInstOrderingScope(LD);
 
  // Type Setting: fromType + fromTypeWidth
  //
  // Sign   : ISD::SEXTLOAD
  // Unsign : ISD::ZEXTLOAD, ISD::NON_EXTLOAD or ISD::EXTLOAD and the
  //          type is integer
  // Float  : ISD::NON_EXTLOAD or ISD::EXTLOAD and the type is float
  // Read at least 8 bits (predicates are stored as 8-bit values)
  // Get the original LoadSDNode::getExtensionType() value
  const unsigned ExtensionType = N->getConstantOperandVal(4);
  const unsigned FromType = (ExtensionType == ISD::SEXTLOAD)
                                ? NVPTX::PTXLdStInstCode::Signed
                                : NVPTX::PTXLdStInstCode::Untyped;
 
  const unsigned FromTypeWidth = getFromTypeWidthForLoad(LD);
  const uint32_t UsedBytesMask = N->getConstantOperandVal(3);
 
  assert(!(EltVT.isVector() && ExtensionType != ISD::NON_EXTLOAD));
 
  const auto [EvictionAndPrefetchHint, PolicyReg] =
      getMemCacheHintOperands(LD,
                              {NVPTXMemCacheHintInstruction::Ld, CodeAddrSpace,
                               /*NumElts=*/LD->getNumValues() - 1,
                               /*EltWidth=*/FromTypeWidth, LD->isVolatile()},
                              DL);
  const auto [Base, Offset] = selectADDR(N->getOperand(1), CurDAG);
  SDValue Ops[] = {getI32Imm(Ordering, DL),
                   getI32Imm(Scope, DL),
                   getI32Imm(CodeAddrSpace, DL),
                   getI32Imm(FromType, DL),
                   getI32Imm(FromTypeWidth, DL),
                   getI32Imm(UsedBytesMask, DL),
                   Base,
                   Offset,
                   EvictionAndPrefetchHint,
                   PolicyReg,
                   Chain};
 
  std::optional<unsigned> Opcode;
  switch (N->getOpcode()) {
  default:
    llvm_unreachable("Unexpected opcode");
  case NVPTXISD::LoadV2:
    Opcode = pickOpcodeForVT(EltVT.SimpleTy, NVPTX::LDV_i16_v2,
                             NVPTX::LDV_i32_v2, NVPTX::LDV_i64_v2);
    break;
  case NVPTXISD::LoadV4:
    Opcode = pickOpcodeForVT(EltVT.SimpleTy, NVPTX::LDV_i16_v4,
                             NVPTX::LDV_i32_v4, NVPTX::LDV_i64_v4);
    break;
  case NVPTXISD::LoadV8:
    Opcode = pickOpcodeForVT(EltVT.SimpleTy, {/* no v8i16 */},
                             NVPTX::LDV_i32_v8, {/* no v8i64 */});
    break;
  }
  if (!Opcode)
    return false;
 
  SDNode *NVPTXLD = CurDAG->getMachineNode(*Opcode, DL, LD->getVTList(), Ops);
 
  MachineMemOperand *MemRef = LD->getMemOperand();
  CurDAG->setNodeMemRefs(cast<MachineSDNode>(NVPTXLD), {MemRef});
 
  ReplaceNode(LD, NVPTXLD);
  return true;
}
 
bool NVPTXDAGToDAGISel::tryLDG(MemSDNode *LD) {
  SDLoc DL(LD);
 
  unsigned ExtensionType;
  uint32_t UsedBytesMask;
  if (const auto *Load = dyn_cast<LoadSDNode>(LD)) {
    ExtensionType = Load->getExtensionType();
    UsedBytesMask = UINT32_MAX;
  } else {
    ExtensionType = LD->getConstantOperandVal(4);
    UsedBytesMask = LD->getConstantOperandVal(3);
  }
  const unsigned FromType = (ExtensionType == ISD::SEXTLOAD)
                                ? NVPTX::PTXLdStInstCode::Signed
                                : NVPTX::PTXLdStInstCode::Untyped;
 
  const unsigned FromTypeWidth = getFromTypeWidthForLoad(LD);
 
  assert(!(LD->getSimpleValueType(0).isVector() &&
           ExtensionType != ISD::NON_EXTLOAD));
 
  const auto [Base, Offset] = selectADDR(LD->getOperand(1), CurDAG);
  const auto [EvictionAndPrefetchHint, PolicyReg] = getMemCacheHintOperands(
      LD,
      {NVPTXMemCacheHintInstruction::Ld, NVPTX::AddressSpace::Global,
       LD->getNumValues() - 1, FromTypeWidth, LD->isVolatile()},
      DL);
  SDValue Ops[] = {getI32Imm(FromType, DL),
                   getI32Imm(FromTypeWidth, DL),
                   getI32Imm(UsedBytesMask, DL),
                   Base,
                   Offset,
                   EvictionAndPrefetchHint,
                   PolicyReg,
                   LD->getChain()};
 
  const MVT::SimpleValueType TargetVT = LD->getSimpleValueType(0).SimpleTy;
  std::optional<unsigned> Opcode;
  switch (LD->getOpcode()) {
  default:
    llvm_unreachable("Unexpected opcode");
  case ISD::LOAD:
    Opcode = pickOpcodeForVT(TargetVT, NVPTX::LD_GLOBAL_NC_i16,
                             NVPTX::LD_GLOBAL_NC_i32, NVPTX::LD_GLOBAL_NC_i64);
    break;
  case NVPTXISD::MLoad:
    Opcode = pickOpcodeForVT(TargetVT, std::nullopt, NVPTX::LD_GLOBAL_NC_i32,
                             NVPTX::LD_GLOBAL_NC_i64);
    break;
  case NVPTXISD::LoadV2:
    Opcode =
        pickOpcodeForVT(TargetVT, NVPTX::LD_GLOBAL_NC_v2i16,
                        NVPTX::LD_GLOBAL_NC_v2i32, NVPTX::LD_GLOBAL_NC_v2i64);
    break;
  case NVPTXISD::LoadV4:
    Opcode =
        pickOpcodeForVT(TargetVT, NVPTX::LD_GLOBAL_NC_v4i16,
                        NVPTX::LD_GLOBAL_NC_v4i32, NVPTX::LD_GLOBAL_NC_v4i64);
    break;
  case NVPTXISD::LoadV8:
    Opcode = pickOpcodeForVT(TargetVT, {/* no v8i16 */},
                             NVPTX::LD_GLOBAL_NC_v8i32, {/* no v8i64 */});
    break;
  }
  if (!Opcode)
    return false;
 
  SDNode *NVPTXLDG = CurDAG->getMachineNode(*Opcode, DL, LD->getVTList(), Ops);
 
  ReplaceNode(LD, NVPTXLDG);
  return true;
}
 
bool NVPTXDAGToDAGISel::tryLDU(SDNode *N) {
  auto *LD = cast<MemSDNode>(N);
 
  SDLoc DL(N);
  const unsigned FromTypeWidth = getFromTypeWidthForLoad(LD);
  const MVT::SimpleValueType TargetVT = LD->getSimpleValueType(0).SimpleTy;
 
  // If this is an LDU intrinsic, the address is the third operand. If its an
  // LDU SD node (from custom vector handling), then its the second operand
  SDValue Addr =
      LD->getOperand(LD->getOpcode() == ISD::INTRINSIC_W_CHAIN ? 2 : 1);
 
  const auto [Base, Offset] = selectADDR(Addr, CurDAG);
  SDValue Ops[] = {getI32Imm(FromTypeWidth, DL), Base, Offset, LD->getChain()};
 
  std::optional<unsigned> Opcode;
  switch (N->getOpcode()) {
  default:
    llvm_unreachable("Unexpected opcode");
  case ISD::INTRINSIC_W_CHAIN:
    Opcode = pickOpcodeForVT(TargetVT, NVPTX::LDU_GLOBAL_i16,
                             NVPTX::LDU_GLOBAL_i32, NVPTX::LDU_GLOBAL_i64);
    break;
  case NVPTXISD::LDUV2:
    Opcode = pickOpcodeForVT(TargetVT, NVPTX::LDU_GLOBAL_v2i16,
                             NVPTX::LDU_GLOBAL_v2i32, NVPTX::LDU_GLOBAL_v2i64);
    break;
  case NVPTXISD::LDUV4:
    Opcode = pickOpcodeForVT(TargetVT, NVPTX::LDU_GLOBAL_v4i16,
                             NVPTX::LDU_GLOBAL_v4i32, {/* no v4i64 */});
    break;
  }
  if (!Opcode)
    return false;
 
  SDNode *NVPTXLDU = CurDAG->getMachineNode(*Opcode, DL, LD->getVTList(), Ops);
 
  ReplaceNode(LD, NVPTXLDU);
  return true;
}
 
bool NVPTXDAGToDAGISel::tryStore(SDNode *N) {
  MemSDNode *ST = cast<MemSDNode>(N);
  assert(ST->writeMem() && "Expected store");
  StoreSDNode *PlainStore = dyn_cast<StoreSDNode>(ST);
  AtomicSDNode *AtomicStore = dyn_cast<AtomicSDNode>(ST);
  assert((PlainStore || AtomicStore) && "Expected store");
 
  // do not support pre/post inc/dec
  if (PlainStore && PlainStore->isIndexed())
    return false;
 
  // Address Space Setting
  const auto CodeAddrSpace = getAddrSpace(ST);
 
  SDLoc DL(ST);
  SDValue Chain = ST->getChain();
  const auto [Ordering, Scope] = getMemInstOrderingScope(ST);
 
  // Vector Setting
  const unsigned ToTypeWidth = ST->getMemoryVT().getSizeInBits();
 
  // Create the machine instruction DAG
  SDValue Value = PlainStore ? PlainStore->getValue() : AtomicStore->getVal();
 
  assert(isPowerOf2_32(ToTypeWidth) && ToTypeWidth >= 8 && ToTypeWidth <= 128 &&
         "Invalid width for store");
 
  const auto [Base, Offset] = selectADDR(ST->getBasePtr(), CurDAG);
 
  // Extract eviction/prefetch hint and cache policy register.
  const auto [EvictionAndPrefetchHint, PolicyReg] = getMemCacheHintOperands(
      ST,
      {NVPTXMemCacheHintInstruction::St, CodeAddrSpace,
       /*NumElts=*/1, /*EltWidth=*/ToTypeWidth, ST->isVolatile()},
      DL);
 
  SDValue Ops[] = {selectPossiblyImm(Value),
                   getI32Imm(Ordering, DL),
                   getI32Imm(Scope, DL),
                   getI32Imm(CodeAddrSpace, DL),
                   getI32Imm(ToTypeWidth, DL),
                   Base,
                   Offset,
                   EvictionAndPrefetchHint,
                   PolicyReg,
                   Chain};
 
  const std::optional<unsigned> Opcode =
      pickOpcodeForVT(Value.getSimpleValueType().SimpleTy, NVPTX::ST_i16,
                      NVPTX::ST_i32, NVPTX::ST_i64);
  if (!Opcode)
    return false;
 
  SDNode *NVPTXST = CurDAG->getMachineNode(*Opcode, DL, MVT::Other, Ops);
 
  if (!NVPTXST)
    return false;
 
  MachineMemOperand *MemRef = ST->getMemOperand();
  CurDAG->setNodeMemRefs(cast<MachineSDNode>(NVPTXST), {MemRef});
  ReplaceNode(ST, NVPTXST);
  return true;
}
 
bool NVPTXDAGToDAGISel::tryStoreVector(SDNode *N) {
  MemSDNode *ST = cast<MemSDNode>(N);
  const unsigned TotalWidth = ST->getMemoryVT().getSizeInBits();
 
  // Address Space Setting
  const auto CodeAddrSpace = getAddrSpace(ST);
  if (CodeAddrSpace == NVPTX::AddressSpace::Const) {
    report_fatal_error("Cannot store to pointer that points to constant "
                       "memory space");
  }
 
  SDLoc DL(ST);
  SDValue Chain = ST->getChain();
  const auto [Ordering, Scope] = getMemInstOrderingScope(ST);
 
  const unsigned NumElts = getStoreVectorNumElts(ST);
 
  SmallVector<SDValue, 16> Ops;
  for (auto &V : ST->ops().slice(1, NumElts))
    Ops.push_back(selectPossiblyImm(V));
  SDValue Addr = N->getOperand(NumElts + 1);
  const unsigned ToTypeWidth = TotalWidth / NumElts;
 
  assert(isPowerOf2_32(ToTypeWidth) && ToTypeWidth >= 8 && ToTypeWidth <= 128 &&
         TotalWidth <= 256 && "Invalid width for store");
 
  // Extract eviction/prefetch hint and cache policy register.
  const auto [EvictionAndPrefetchHint, PolicyReg] = getMemCacheHintOperands(
      ST,
      {NVPTXMemCacheHintInstruction::St, CodeAddrSpace,
       /*NumElts=*/NumElts, /*EltWidth=*/ToTypeWidth, ST->isVolatile()},
      DL);
 
  const auto [Base, Offset] = selectADDR(Addr, CurDAG);
  Ops.append({getI32Imm(Ordering, DL), getI32Imm(Scope, DL),
              getI32Imm(CodeAddrSpace, DL), getI32Imm(ToTypeWidth, DL), Base,
              Offset, EvictionAndPrefetchHint, PolicyReg, Chain});
 
  const MVT::SimpleValueType EltVT =
      ST->getOperand(1).getSimpleValueType().SimpleTy;
  std::optional<unsigned> Opcode;
  switch (ST->getOpcode()) {
  default:
    return false;
  case NVPTXISD::StoreV2:
    Opcode = pickOpcodeForVT(EltVT, NVPTX::STV_i16_v2, NVPTX::STV_i32_v2,
                             NVPTX::STV_i64_v2);
    break;
  case NVPTXISD::StoreV4:
    Opcode = pickOpcodeForVT(EltVT, NVPTX::STV_i16_v4, NVPTX::STV_i32_v4,
                             NVPTX::STV_i64_v4);
    break;
  case NVPTXISD::StoreV8:
    Opcode = pickOpcodeForVT(EltVT, {/* no v8i16 */}, NVPTX::STV_i32_v8,
                             {/* no v8i64 */});
    break;
  }
 
  if (!Opcode)
    return false;
 
  SDNode *NVPTXST = CurDAG->getMachineNode(*Opcode, DL, MVT::Other, Ops);
 
  MachineMemOperand *MemRef = ST->getMemOperand();
  CurDAG->setNodeMemRefs(cast<MachineSDNode>(NVPTXST), {MemRef});
 
  ReplaceNode(ST, NVPTXST);
  return true;
}
 
/// SelectBFE - Look for instruction sequences that can be made more efficient
/// by using the 'bfe' (bit-field extract) PTX instruction
bool NVPTXDAGToDAGISel::tryBFE(SDNode *N) {
  // BFE is not a native instruction starting with sm_70. Let ptxas see the
  // original operations so it can optimize them for the target architecture.
  if (Subtarget->hasFeature(NVPTX::SM70))
    return false;
 
  SDLoc DL(N);
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  SDValue Len;
  SDValue Start;
  SDValue Val;
  bool IsSigned = false;
 
  if (N->getOpcode() == ISD::AND) {
    // Canonicalize the operands
    // We want 'and %val, %mask'
    if (isa<ConstantSDNode>(LHS) && !isa<ConstantSDNode>(RHS)) {
      std::swap(LHS, RHS);
    }
 
    ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(RHS);
    if (!Mask) {
      // We need a constant mask on the RHS of the AND
      return false;
    }
 
    // Extract the mask bits
    uint64_t MaskVal = Mask->getZExtValue();
    if (!isMask_64(MaskVal)) {
      // We *could* handle shifted masks here, but doing so would require an
      // 'and' operation to fix up the low-order bits so we would trade
      // shr+and for bfe+and, which has the same throughput
      return false;
    }
 
    // How many bits are in our mask?
    int64_t NumBits = countr_one(MaskVal);
    Len = CurDAG->getTargetConstant(NumBits, DL, MVT::i32);
 
    if (LHS.getOpcode() == ISD::SRL || LHS.getOpcode() == ISD::SRA) {
      // We have a 'srl/and' pair, extract the effective start bit and length
      Val = LHS.getNode()->getOperand(0);
      Start = LHS.getNode()->getOperand(1);
      ConstantSDNode *StartConst = dyn_cast<ConstantSDNode>(Start);
      if (StartConst) {
        uint64_t StartVal = StartConst->getZExtValue();
        // How many "good" bits do we have left?  "good" is defined here as bits
        // that exist in the original value, not shifted in.
        int64_t GoodBits = Val.getValueSizeInBits() - StartVal;
        if (NumBits > GoodBits) {
          // Do not handle the case where bits have been shifted in. In theory
          // we could handle this, but the cost is likely higher than just
          // emitting the srl/and pair.
          return false;
        }
        Start = CurDAG->getTargetConstant(StartVal, DL, MVT::i32);
      } else {
        // Do not handle the case where the shift amount (can be zero if no srl
        // was found) is not constant. We could handle this case, but it would
        // require run-time logic that would be more expensive than just
        // emitting the srl/and pair.
        return false;
      }
    } else {
      // Do not handle the case where the LHS of the and is not a shift. While
      // it would be trivial to handle this case, it would just transform
      // 'and' -> 'bfe', but 'and' has higher-throughput.
      return false;
    }
  } else if (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) {
    if (LHS->getOpcode() == ISD::AND) {
      ConstantSDNode *ShiftCnst = dyn_cast<ConstantSDNode>(RHS);
      if (!ShiftCnst) {
        // Shift amount must be constant
        return false;
      }
 
      uint64_t ShiftAmt = ShiftCnst->getZExtValue();
 
      SDValue AndLHS = LHS->getOperand(0);
      SDValue AndRHS = LHS->getOperand(1);
 
      // Canonicalize the AND to have the mask on the RHS
      if (isa<ConstantSDNode>(AndLHS)) {
        std::swap(AndLHS, AndRHS);
      }
 
      ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(AndRHS);
      if (!MaskCnst) {
        // Mask must be constant
        return false;
      }
 
      uint64_t MaskVal = MaskCnst->getZExtValue();
      uint64_t NumZeros;
      uint64_t NumBits;
      if (isMask_64(MaskVal)) {
        NumZeros = 0;
        // The number of bits in the result bitfield will be the number of
        // trailing ones (the AND) minus the number of bits we shift off
        NumBits = llvm::countr_one(MaskVal) - ShiftAmt;
      } else if (isShiftedMask_64(MaskVal)) {
        NumZeros = llvm::countr_zero(MaskVal);
        unsigned NumOnes = llvm::countr_one(MaskVal >> NumZeros);
        // The number of bits in the result bitfield will be the number of
        // trailing zeros plus the number of set bits in the mask minus the
        // number of bits we shift off
        NumBits = NumZeros + NumOnes - ShiftAmt;
      } else {
        // This is not a mask we can handle
        return false;
      }
 
      if (ShiftAmt < NumZeros) {
        // Handling this case would require extra logic that would make this
        // transformation non-profitable
        return false;
      }
 
      Val = AndLHS;
      Start = CurDAG->getTargetConstant(ShiftAmt, DL, MVT::i32);
      Len = CurDAG->getTargetConstant(NumBits, DL, MVT::i32);
 
      // If pre-shift AND includes the sign bit in the bitfield, we must use
      // signed BFE to replicate that bit during bitfield extraction. If the
      // sign bit is not part of the mask, unsigned BFE will zero out upper bits
      // of the result
      if (N->getOpcode() == ISD::SRA)
        IsSigned = (ShiftAmt + NumBits) == Val.getValueSizeInBits();
    } else if (LHS->getOpcode() == ISD::SHL) {
      // Here, we have a pattern like:
      //
      // (sra (shl val, NN), MM)
      // or
      // (srl (shl val, NN), MM)
      //
      // If MM >= NN, we can efficiently optimize this with bfe
      Val = LHS->getOperand(0);
 
      SDValue ShlRHS = LHS->getOperand(1);
      ConstantSDNode *ShlCnst = dyn_cast<ConstantSDNode>(ShlRHS);
      if (!ShlCnst) {
        // Shift amount must be constant
        return false;
      }
      uint64_t InnerShiftAmt = ShlCnst->getZExtValue();
 
      SDValue ShrRHS = RHS;
      ConstantSDNode *ShrCnst = dyn_cast<ConstantSDNode>(ShrRHS);
      if (!ShrCnst) {
        // Shift amount must be constant
        return false;
      }
      uint64_t OuterShiftAmt = ShrCnst->getZExtValue();
 
      // To avoid extra codegen and be profitable, we need Outer >= Inner
      if (OuterShiftAmt < InnerShiftAmt) {
        return false;
      }
 
      // If the outer shift is more than the type size, we have no bitfield to
      // extract (since we also check that the inner shift is <= the outer shift
      // then this also implies that the inner shift is < the type size)
      if (OuterShiftAmt >= Val.getValueSizeInBits()) {
        return false;
      }
 
      Start = CurDAG->getTargetConstant(OuterShiftAmt - InnerShiftAmt, DL,
                                        MVT::i32);
      Len = CurDAG->getTargetConstant(Val.getValueSizeInBits() - OuterShiftAmt,
                                      DL, MVT::i32);
 
      if (N->getOpcode() == ISD::SRA) {
        // If we have a arithmetic right shift, we need to use the signed bfe
        // variant
        IsSigned = true;
      }
    } else {
      // No can do...
      return false;
    }
  } else {
    // No can do...
    return false;
  }
  // Only 32-bit BFE has a native SASS implementation.
  if (Val.getValueType() != MVT::i32)
    return false;
 
  unsigned Opc = IsSigned ? NVPTX::BFE_S32 : NVPTX::BFE_U32;
 
  SDValue Ops[] = {selectPossiblyImm(Val), Start, Len};
 
  ReplaceNode(N, CurDAG->getMachineNode(Opc, DL, N->getVTList(), Ops));
  return true;
}
 
// Select bf16/bf16v2 FADD, FSUB, FMUL as fma on targets with only fma
bool NVPTXDAGToDAGISel::tryBF16ArithToFMA(SDNode *N) {
  EVT VT = SDValue(N, 0).getValueType();
  if (VT.getScalarType() != MVT::bf16)
    return false;
 
  const NVPTXSubtarget *STI = TM.getSubtargetImpl();
  if (STI->hasNativeBF16Support(N->getOpcode()))
    return false;
 
  const bool IsVec = VT.isVector();
  assert(!IsVec || VT.getVectorNumElements() == 2);
  SDLoc DL(N);
  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);
  SmallVector<SDValue, 3> Operands;
  auto GetConstant = [&](float Value) -> SDValue {
    // BF16 immediates must be legalized to integer register values
    APFloat APF(Value);
    bool LosesInfo;
    APF.convert(APFloat::BFloat(), APFloat::rmNearestTiesToEven, &LosesInfo);
    assert(!LosesInfo);
    if (IsVec) {
      auto API = APF.bitcastToAPInt();
      API = API.concat(API);
      auto Const = CurDAG->getTargetConstant(API, DL, MVT::i32);
      return SDValue(CurDAG->getMachineNode(NVPTX::MOV_B32_i, DL, VT, Const),
                     0);
    }
    auto Const = CurDAG->getTargetConstantFP(APF, DL, VT);
    return SDValue(CurDAG->getMachineNode(NVPTX::MOV_BF16_i, DL, VT, Const), 0);
  };
 
  switch (N->getOpcode()) {
  case ISD::FADD:
    // add(a, b) -> fma(a, 1.0, b)
    Operands = {N0, GetConstant(1.0), N1};
    break;
  case ISD::FSUB:
    // sub(a, b) -> fma(b, -1.0, a)
    Operands = {N1, GetConstant(-1.0), N0};
    break;
  case ISD::FMUL:
    // mul(a, b) -> fma(a, b, -0.0)
    // NOTE: The identity is -0, not 0, because -0 + 0 == 0 for floats
    Operands = {N0, N1, GetConstant(-0.0)};
    break;
  default:
    llvm_unreachable("Unexpected opcode");
  };
 
  int Opcode = IsVec ? NVPTX::FMA_BF16x2 : NVPTX::FMA_BF16;
  MachineSDNode *FMA = CurDAG->getMachineNode(Opcode, DL, VT, Operands);
  ReplaceNode(N, FMA);
  return true;
}
 
// The min/max .abs modifier also accepts operands already known to have no
// negative values (not even -0). NaN signs are immaterial to these
// instructions.
bool NVPTXDAGToDAGISel::SelectFAbs(SDValue N, SDValue &Src) {
  if (N.getOpcode() == ISD::FABS)
    Src = N.getOperand(0);
  else if (CurDAG->computeKnownFPClass(N, fcNegative).signBitIsZeroOrNaN())
    Src = N;
  else
    return false;
  Src = selectPossiblyImm(Src);
  return true;
}
 
SDValue NVPTXDAGToDAGISel::selectPossiblyImm(SDValue V) {
  if (V.getOpcode() == ISD::BITCAST)
    V = V.getOperand(0);
 
  if (auto *CN = dyn_cast<ConstantSDNode>(V))
    return CurDAG->getTargetConstant(CN->getAPIntValue(), SDLoc(V),
                                     V.getValueType());
  if (auto *CN = dyn_cast<ConstantFPSDNode>(V))
    return CurDAG->getTargetConstantFP(CN->getValueAPF(), SDLoc(V),
                                       V.getValueType());
  return V;
}
 
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
/// inline asm expressions.
bool NVPTXDAGToDAGISel::SelectInlineAsmMemoryOperand(
    const SDValue &Op, InlineAsm::ConstraintCode ConstraintID,
    std::vector<SDValue> &OutOps) {
  switch (ConstraintID) {
  default:
    return true;
  case InlineAsm::ConstraintCode::m: { // memory
    const auto [Base, Offset] = selectADDR(Op, CurDAG);
    OutOps.push_back(Base);
    OutOps.push_back(Offset);
    return false;
  }
  }
  return true;
}
 
void NVPTXDAGToDAGISel::SelectV2I64toI128(SDNode *N) {
  // Lower a CopyToReg with two 64-bit inputs
  // Dst:i128, lo:i64, hi:i64
  //
  // CopyToReg Dst, lo, hi;
  //
  // ==>
  //
  // tmp = V2I64toI128 {lo, hi};
  // CopyToReg Dst, tmp;
  SDValue Dst = N->getOperand(1);
  SDValue Lo = selectPossiblyImm(N->getOperand(2));
  SDValue Hi = selectPossiblyImm(N->getOperand(3));
 
  SDLoc DL(N);
  SDNode *Mov =
      CurDAG->getMachineNode(NVPTX::V2I64toI128, DL, MVT::i128, {Lo, Hi});
 
  SmallVector<SDValue, 4> NewOps(N->getNumOperands() - 1);
  NewOps[0] = N->getOperand(0);
  NewOps[1] = Dst;
  NewOps[2] = SDValue(Mov, 0);
  if (N->getNumOperands() == 5)
    NewOps[3] = N->getOperand(4);
  SDValue NewValue = CurDAG->getNode(ISD::CopyToReg, DL, SmallVector<EVT>(N->values()), NewOps);
 
  ReplaceNode(N, NewValue.getNode());
}
 
void NVPTXDAGToDAGISel::SelectI128toV2I64(SDNode *N) {
  // Lower CopyFromReg from a 128-bit regs to two 64-bit regs
  // Dst:i128, Src:i128
  //
  // {lo, hi} = CopyFromReg Src
  //
  // ==>
  //
  // {lo, hi} = I128toV2I64 Src
  //
  SDValue Ch = N->getOperand(0);
  SDValue Src = N->getOperand(1);
  SDValue Glue = N->getOperand(2);
  SDLoc DL(N);
 
  // Add Glue and Ch to the operands and results to avoid break the execution
  // order
  SDNode *Mov = CurDAG->getMachineNode(
      NVPTX::I128toV2I64, DL,
      {MVT::i64, MVT::i64, Ch.getValueType(), Glue.getValueType()},
      {Src, Ch, Glue});
 
  ReplaceNode(N, Mov);
}
 
bool NVPTXDAGToDAGISel::tryFence(SDNode *N) {
  SDLoc DL(N);
  assert(N->getOpcode() == ISD::ATOMIC_FENCE);
  auto Scope = Scopes[N->getConstantOperandVal(2)];
 
  // Singlethread fences have no inter-thread synchronization requirements.
  // Note: std::atomic_signal_fence lowers to singlethread LLVM IR fences;
  // this intentionally drops these before emitting PTX.
  if (Scope == NVPTX::Scope::Thread) {
    CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), N->getOperand(0));
    CurDAG->RemoveDeadNode(N);
    return true;
  }
 
  unsigned int FenceOp = getFenceOp(
      NVPTX::Ordering(N->getConstantOperandVal(1)), Scope, Subtarget);
  SDValue Chain = N->getOperand(0);
  SDNode *FenceNode = CurDAG->getMachineNode(FenceOp, DL, MVT::Other, Chain);
  ReplaceNode(N, FenceNode);
  return true;
}
 
NVPTXScopes::NVPTXScopes(LLVMContext &C, const Triple &T) : Context(&C) {
  auto ScopeID = [&](AtomicScope Scope) {
    return C.getOrInsertSyncScopeID(*getAtomicScopeIRString(T, Scope));
  };
  Scopes[ScopeID(AtomicScope::Single)] = NVPTX::Scope::Thread;
  Scopes[ScopeID(AtomicScope::System)] = NVPTX::Scope::System;
  Scopes[ScopeID(AtomicScope::Workgroup)] = NVPTX::Scope::Block;
  Scopes[ScopeID(AtomicScope::Cluster)] = NVPTX::Scope::Cluster;
  Scopes[ScopeID(AtomicScope::Device)] = NVPTX::Scope::Device;
}
 
NVPTX::Scope NVPTXScopes::operator[](SyncScope::ID ID) const {
  if (Scopes.empty())
    llvm_unreachable("NVPTX Scopes must be initialized before calling "
                     "NVPTXScopes::operator[]");
 
  auto S = Scopes.find(ID);
  if (S == Scopes.end()) {
    auto scopeName = Context->getSyncScopeName(ID);
    assert(scopeName.has_value() && "Scope name must exist.");
 
    // Build list of supported syncscopes programmatically
    SmallVector<StringRef> supportedScopes;
    for (const auto &Entry : Scopes) {
      if (auto name = Context->getSyncScopeName(Entry.first))
        supportedScopes.push_back(name->empty() ? "<empty string>" : *name);
    }
 
    reportFatalUsageError(
        formatv("NVPTX backend does not support syncscope \"{0}\" (ID={1}).\n"
                "Supported syncscopes are: {2}.",
                scopeName.value(), int(ID),
                make_range(supportedScopes.begin(), supportedScopes.end())));
  }
  return S->second;
}
 
bool NVPTXScopes::empty() const { return Scopes.size() == 0; }
 
#define TCGEN05_ST_OPCODE(SHAPE, NUM)                                          \
  (enableUnpack ? NVPTX::TCGEN05_ST_##SHAPE##_##NUM##_UNPACK                   \
                : NVPTX::TCGEN05_ST_##SHAPE##_##NUM)
 
static unsigned getTcgen05StOpcode(unsigned IID, bool enableUnpack) {
  switch (IID) {
  case Intrinsic::nvvm_tcgen05_st_16x64b_x1:
    return TCGEN05_ST_OPCODE(16x64b, x1);
  case Intrinsic::nvvm_tcgen05_st_16x64b_x2:
    return TCGEN05_ST_OPCODE(16x64b, x2);
  case Intrinsic::nvvm_tcgen05_st_16x64b_x4:
    return TCGEN05_ST_OPCODE(16x64b, x4);
  case Intrinsic::nvvm_tcgen05_st_16x64b_x8:
    return TCGEN05_ST_OPCODE(16x64b, x8);
  case Intrinsic::nvvm_tcgen05_st_16x64b_x16:
    return TCGEN05_ST_OPCODE(16x64b, x16);
  case Intrinsic::nvvm_tcgen05_st_16x64b_x32:
    return TCGEN05_ST_OPCODE(16x64b, x32);
  case Intrinsic::nvvm_tcgen05_st_16x64b_x64:
    return TCGEN05_ST_OPCODE(16x64b, x64);
  case Intrinsic::nvvm_tcgen05_st_16x64b_x128:
    return TCGEN05_ST_OPCODE(16x64b, x128);
  case Intrinsic::nvvm_tcgen05_st_16x128b_x1:
    return TCGEN05_ST_OPCODE(16x128b, x1);
  case Intrinsic::nvvm_tcgen05_st_16x128b_x2:
    return TCGEN05_ST_OPCODE(16x128b, x2);
  case Intrinsic::nvvm_tcgen05_st_16x128b_x4:
    return TCGEN05_ST_OPCODE(16x128b, x4);
  case Intrinsic::nvvm_tcgen05_st_16x128b_x8:
    return TCGEN05_ST_OPCODE(16x128b, x8);
  case Intrinsic::nvvm_tcgen05_st_16x128b_x16:
    return TCGEN05_ST_OPCODE(16x128b, x16);
  case Intrinsic::nvvm_tcgen05_st_16x128b_x32:
    return TCGEN05_ST_OPCODE(16x128b, x32);
  case Intrinsic::nvvm_tcgen05_st_16x128b_x64:
    return TCGEN05_ST_OPCODE(16x128b, x64);
  case Intrinsic::nvvm_tcgen05_st_16x256b_x1:
    return TCGEN05_ST_OPCODE(16x256b, x1);
  case Intrinsic::nvvm_tcgen05_st_16x256b_x2:
    return TCGEN05_ST_OPCODE(16x256b, x2);
  case Intrinsic::nvvm_tcgen05_st_16x256b_x4:
    return TCGEN05_ST_OPCODE(16x256b, x4);
  case Intrinsic::nvvm_tcgen05_st_16x256b_x8:
    return TCGEN05_ST_OPCODE(16x256b, x8);
  case Intrinsic::nvvm_tcgen05_st_16x256b_x16:
    return TCGEN05_ST_OPCODE(16x256b, x16);
  case Intrinsic::nvvm_tcgen05_st_16x256b_x32:
    return TCGEN05_ST_OPCODE(16x256b, x32);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x1:
    return TCGEN05_ST_OPCODE(16x32bx2, x1);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x2:
    return TCGEN05_ST_OPCODE(16x32bx2, x2);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x4:
    return TCGEN05_ST_OPCODE(16x32bx2, x4);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x8:
    return TCGEN05_ST_OPCODE(16x32bx2, x8);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x16:
    return TCGEN05_ST_OPCODE(16x32bx2, x16);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x32:
    return TCGEN05_ST_OPCODE(16x32bx2, x32);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x64:
    return TCGEN05_ST_OPCODE(16x32bx2, x64);
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x128:
    return TCGEN05_ST_OPCODE(16x32bx2, x128);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x1:
    return TCGEN05_ST_OPCODE(32x32b, x1);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x2:
    return TCGEN05_ST_OPCODE(32x32b, x2);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x4:
    return TCGEN05_ST_OPCODE(32x32b, x4);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x8:
    return TCGEN05_ST_OPCODE(32x32b, x8);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x16:
    return TCGEN05_ST_OPCODE(32x32b, x16);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x32:
    return TCGEN05_ST_OPCODE(32x32b, x32);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x64:
    return TCGEN05_ST_OPCODE(32x32b, x64);
  case Intrinsic::nvvm_tcgen05_st_32x32b_x128:
    return TCGEN05_ST_OPCODE(32x32b, x128);
  }
  llvm_unreachable("unhandled tcgen05.st lowering");
}
 
void NVPTXDAGToDAGISel::SelectTcgen05St(SDNode *N, bool hasOffset) {
  if (!Subtarget->hasTcgen05InstSupport())
    report_fatal_error(
        "tcgen05.st is not supported on this architecture variant");
 
  SDLoc DL(N);
  unsigned IID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
 
  SmallVector<SDValue, 128> Operands = {
      N->getOperand(2) // taddr
  };
 
  if (hasOffset)
    Operands.push_back(CurDAG->getTargetConstant(
        cast<ConstantSDNode>(N->getOperand(3))->getZExtValue(), DL,
        MVT::i32)); // Offset
 
  for (unsigned I = hasOffset ? 4 : 3; I < (N->getNumOperands() - 1); I++)
    Operands.push_back(N->getOperand(I));
 
  bool enableUnpack =
      cast<ConstantSDNode>(N->getOperand(N->getNumOperands() - 1))
          ->getZExtValue();
 
  Operands.push_back(N->getOperand(0)); // Chain
  ReplaceNode(N, CurDAG->getMachineNode(getTcgen05StOpcode(IID, enableUnpack),
                                        DL, N->getVTList(), Operands));
}
 
bool NVPTXDAGToDAGISel::tryIntrinsicVoid(SDNode *N) {
  unsigned IID = N->getConstantOperandVal(1);
  switch (IID) {
  default:
    return false;
  case Intrinsic::nvvm_tcgen05_st_16x64b_x1:
  case Intrinsic::nvvm_tcgen05_st_16x64b_x2:
  case Intrinsic::nvvm_tcgen05_st_16x64b_x4:
  case Intrinsic::nvvm_tcgen05_st_16x64b_x8:
  case Intrinsic::nvvm_tcgen05_st_16x64b_x16:
  case Intrinsic::nvvm_tcgen05_st_16x64b_x32:
  case Intrinsic::nvvm_tcgen05_st_16x64b_x64:
  case Intrinsic::nvvm_tcgen05_st_16x64b_x128:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x1:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x2:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x4:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x8:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x16:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x32:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x64:
  case Intrinsic::nvvm_tcgen05_st_32x32b_x128:
  case Intrinsic::nvvm_tcgen05_st_16x128b_x1:
  case Intrinsic::nvvm_tcgen05_st_16x128b_x2:
  case Intrinsic::nvvm_tcgen05_st_16x128b_x4:
  case Intrinsic::nvvm_tcgen05_st_16x128b_x8:
  case Intrinsic::nvvm_tcgen05_st_16x128b_x16:
  case Intrinsic::nvvm_tcgen05_st_16x128b_x32:
  case Intrinsic::nvvm_tcgen05_st_16x128b_x64:
  case Intrinsic::nvvm_tcgen05_st_16x256b_x1:
  case Intrinsic::nvvm_tcgen05_st_16x256b_x2:
  case Intrinsic::nvvm_tcgen05_st_16x256b_x4:
  case Intrinsic::nvvm_tcgen05_st_16x256b_x8:
  case Intrinsic::nvvm_tcgen05_st_16x256b_x16:
  case Intrinsic::nvvm_tcgen05_st_16x256b_x32: {
    SelectTcgen05St(N);
    return true;
  }
 
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x1:
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x2:
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x4:
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x8:
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x16:
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x32:
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x64:
  case Intrinsic::nvvm_tcgen05_st_16x32bx2_x128: {
    SelectTcgen05St(N, /*  hasOffset */ true);
    return true;
  }
  }
}
 
void NVPTXDAGToDAGISel::selectAtomicSwap128(SDNode *N) {
  MemSDNode *AN = cast<MemSDNode>(N);
  SDLoc dl(N);
 
  const SDValue Chain = N->getOperand(0);
  const auto [Base, Offset] = selectADDR(N->getOperand(1), CurDAG);
  SmallVector<SDValue, 10> Ops{Base, Offset};
  for (const SDValue &Operand : drop_begin(N->op_values(), 2))
    Ops.push_back(selectPossiblyImm(Operand));
  Ops.append({getI32Imm(getMemOrder(AN), dl), getI32Imm(getAtomicScope(AN), dl),
              getI32Imm(getAddrSpace(AN), dl)});
 
  if (N->getOpcode() == NVPTXISD::ATOMIC_SWAP_B128) {
    unsigned EltWidth = AN->getMemoryVT().getFixedSizeInBits();
    NVPTXMemCacheHintAccess Access{NVPTXMemCacheHintInstruction::Atom,
                                   getAddrSpace(AN),
                                   /*NumElts=*/1, EltWidth, AN->isVolatile()};
    const auto [EvictionAndPrefetchHint, CachePolicyReg] =
        getMemCacheHintOperands(AN, Access, dl);
    Ops.push_back(EvictionAndPrefetchHint);
    Ops.push_back(CachePolicyReg);
  }
 
  Ops.push_back(Chain);
 
  assert(N->getOpcode() == NVPTXISD::ATOMIC_CMP_SWAP_B128 ||
         N->getOpcode() == NVPTXISD::ATOMIC_SWAP_B128);
  unsigned Opcode = N->getOpcode() == NVPTXISD::ATOMIC_SWAP_B128
                        ? NVPTX::ATOM_EXCH_B128
                        : NVPTX::ATOM_CAS_B128;
 
  auto *ATOM = CurDAG->getMachineNode(Opcode, dl, N->getVTList(), Ops);
  CurDAG->setNodeMemRefs(ATOM, AN->getMemOperand());
 
  ReplaceNode(N, ATOM);
}