path: root/disassembler/disassembler_riscv64.cc

/*
 * Copyright (C) 2023 The Android Open Source Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include "disassembler_riscv64.h"

#include "android-base/logging.h"
#include "android-base/stringprintf.h"

#include "base/bit_utils.h"
#include "base/casts.h"

using android::base::StringPrintf;

namespace art {
namespace riscv64 {

class DisassemblerRiscv64::Printer {
 public:
  Printer(DisassemblerRiscv64* disassembler, std::ostream& os)
      : disassembler_(disassembler), os_(os) {}

  void Dump32(const uint8_t* insn);
  void Dump16(const uint8_t* insn);
  void Dump2Byte(const uint8_t* data);
  void DumpByte(const uint8_t* data);

 private:
  // This enumeration should mirror the declarations in runtime/arch/riscv64/registers_riscv64.h.
  // We do not include that file to avoid a dependency on libart.
  enum {
    Zero = 0,
    RA = 1,
    FP  = 8,
    TR  = 9,
  };

  static const char* XRegName(uint32_t regno);
  static const char* FRegName(uint32_t regno);
  static const char* RoundingModeName(uint32_t rm);

  static int32_t Decode32Imm12(uint32_t insn32) {
    uint32_t sign = (insn32 >> 31);
    uint32_t imm12 = (insn32 >> 20);
    return static_cast<int32_t>(imm12) - static_cast<int32_t>(sign << 12);  // Sign-extend.
  }

  static int32_t Decode32StoreOffset(uint32_t insn32) {
    uint32_t bit11 = insn32 >> 31;
    uint32_t bits5_11 = insn32 >> 25;
    uint32_t bits0_4 = (insn32 >> 7) & 0x1fu;
    uint32_t imm = (bits5_11 << 5) + bits0_4;
    return static_cast<int32_t>(imm) - static_cast<int32_t>(bit11 << 12);  // Sign-extend.
  }

  static uint32_t GetRd(uint32_t insn32) { return (insn32 >> 7) & 0x1fu; }
  static uint32_t GetRs1(uint32_t insn32) { return (insn32 >> 15) & 0x1fu; }
  static uint32_t GetRs2(uint32_t insn32) { return (insn32 >> 20) & 0x1fu; }
  static uint32_t GetRs3(uint32_t insn32) { return insn32 >> 27; }
  static uint32_t GetRoundingMode(uint32_t insn32) { return (insn32 >> 12) & 7u; }

  void PrintBranchOffset(int32_t offset);
  void PrintLoadStoreAddress(uint32_t rs1, int32_t offset);

  void Print32Lui(uint32_t insn32);
  void Print32Auipc(const uint8_t* insn, uint32_t insn32);
  void Print32Jal(const uint8_t* insn, uint32_t insn32);
  void Print32Jalr(const uint8_t* insn, uint32_t insn32);
  void Print32BCond(const uint8_t* insn, uint32_t insn32);
  void Print32Load(uint32_t insn32);
  void Print32Store(uint32_t insn32);
  void Print32FLoad(uint32_t insn32);
  void Print32FStore(uint32_t insn32);
  void Print32BinOpImm(uint32_t insn32);
  void Print32BinOp(uint32_t insn32);
  void Print32Atomic(uint32_t insn32);
  void Print32FpOp(uint32_t insn32);
  void Print32FpFma(uint32_t insn32);
  void Print32Zicsr(uint32_t insn32);
  void Print32Fence(uint32_t insn32);

  DisassemblerRiscv64* const disassembler_;
  std::ostream& os_;
};

const char* DisassemblerRiscv64::Printer::XRegName(uint32_t regno) {
  static const char* const kXRegisterNames[] = {
      "zero",
      "ra",
      "sp",
      "gp",
      "tp",
      "t0",
      "t1",
      "t2",
      "fp",  // s0/fp
      "tr",  // s1/tr - ART thread register
      "a0",
      "a1",
      "a2",
      "a3",
      "a4",
      "a5",
      "a6",
      "a7",
      "s2",
      "s3",
      "s4",
      "s5",
      "s6",
      "s7",
      "s8",
      "s9",
      "s10",
      "s11",
      "t3",
      "t4",
      "t5",
      "t6",
  };
  static_assert(std::size(kXRegisterNames) == 32);
  DCHECK_LT(regno, 32u);
  return kXRegisterNames[regno];
}

const char* DisassemblerRiscv64::Printer::FRegName(uint32_t regno) {
  static const char* const kFRegisterNames[] = {
      "ft0",
      "ft1",
      "ft2",
      "ft3",
      "ft4",
      "ft5",
      "ft6",
      "ft7",
      "fs0",
      "fs1",
      "fa0",
      "fa1",
      "fa2",
      "fa3",
      "fa4",
      "fa5",
      "fa6",
      "fa7",
      "fs2",
      "fs3",
      "fs4",
      "fs5",
      "fs6",
      "fs7",
      "fs8",
      "fs9",
      "fs10",
      "fs11",
      "ft8",
      "ft9",
      "ft10",
      "ft11",
  };
  static_assert(std::size(kFRegisterNames) == 32);
  DCHECK_LT(regno, 32u);
  return kFRegisterNames[regno];
}

const char* DisassemblerRiscv64::Printer::RoundingModeName(uint32_t rm) {
  // Note: We do not print the rounding mode for DYN.
  static const char* const kRoundingModeNames[] = {
      ".rne", ".rtz", ".rdn", ".rup", ".rmm", ".<reserved-rm>", ".<reserved-rm>", /*DYN*/ ""
  };
  static_assert(std::size(kRoundingModeNames) == 8);
  DCHECK_LT(rm, 8u);
  return kRoundingModeNames[rm];
}

void DisassemblerRiscv64::Printer::PrintBranchOffset(int32_t offset) {
  os_ << (offset >= 0 ? "+" : "") << offset;
}

void DisassemblerRiscv64::Printer::PrintLoadStoreAddress(uint32_t rs1, int32_t offset) {
  if (offset != 0) {
    os_ << StringPrintf("%d", offset);
  }
  os_ << "(" << XRegName(rs1) << ")";

  if (rs1 == TR && offset >= 0) {
    // Add entrypoint name.
    os_ << " ; ";
    disassembler_->GetDisassemblerOptions()->thread_offset_name_function_(
        os_, dchecked_integral_cast<uint32_t>(offset));
  }
}

void DisassemblerRiscv64::Printer::Print32Lui(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x37u);
  // TODO(riscv64): Should we also print the actual sign-extend value?
  os_ << StringPrintf("lui %s, %u", XRegName(GetRd(insn32)), insn32 >> 12);
}

void DisassemblerRiscv64::Printer::Print32Auipc([[maybe_unused]] const uint8_t* insn,
                                                uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x17u);
  // TODO(riscv64): Should we also print the calculated address?
  os_ << StringPrintf("auipc %s, %u", XRegName(GetRd(insn32)), insn32 >> 12);
}

void DisassemblerRiscv64::Printer::Print32Jal(const uint8_t* insn, uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x6fu);
  // Print an alias if available.
  uint32_t rd = GetRd(insn32);
  os_ << (rd == Zero ? "j " : "jal ");
  if (rd != Zero && rd != RA) {
    os_ << XRegName(rd) << ", ";
  }
  uint32_t bit20 = (insn32 >> 31);
  uint32_t bits1_10 = (insn32 >> 21) & 0x3ffu;
  uint32_t bit11 = (insn32 >> 20) & 1u;
  uint32_t bits12_19 = (insn32 >> 12) & 0xffu;
  uint32_t imm = (bits1_10 << 1) + (bit11 << 11) + (bits12_19 << 12) + (bit20 << 20);
  int32_t offset = static_cast<int32_t>(imm) - static_cast<int32_t>(bit20 << 21);  // Sign-extend.
  PrintBranchOffset(offset);
  os_ << " ; " << disassembler_->FormatInstructionPointer(insn + offset);

  // TODO(riscv64): When we implement shared thunks to reduce AOT slow-path code size,
  // check if this JAL lands at an entrypoint load from TR and, if so, print its name.
}

void DisassemblerRiscv64::Printer::Print32Jalr([[maybe_unused]] const uint8_t* insn,
                                               uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x67u);
  DCHECK_EQ((insn32 >> 12) & 7u, 0u);
  uint32_t rd = GetRd(insn32);
  uint32_t rs1 = GetRs1(insn32);
  int32_t imm12 = Decode32Imm12(insn32);
  // Print shorter macro instruction notation if available.
  if (rd == Zero && rs1 == RA && imm12 == 0) {
    os_ << "ret";
  } else if (rd == Zero && imm12 == 0) {
    os_ << "jr " << XRegName(rs1);
  } else if (rd == RA && imm12 == 0) {
    os_ << "jalr " << XRegName(rs1);
  } else {
    // TODO(riscv64): Should we also print the calculated address if the preceding
    // instruction is AUIPC? (We would need to record the previous instruction.)
    os_ << "jalr " << XRegName(rd) << ", ";
    // Use the same format as llvm-objdump: "rs1" if `imm12` is zero, otherwise "imm12(rs1)".
    if (imm12 == 0) {
      os_ << XRegName(rs1);
    } else {
      os_ << imm12 << "(" << XRegName(rs1) << ")";
    }
  }
}

void DisassemblerRiscv64::Printer::Print32BCond(const uint8_t* insn, uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x63u);
  static const char* const kOpcodes[] = {
      "beq", "bne", nullptr, nullptr, "blt", "bge", "bltu", "bgeu"
  };
  uint32_t funct3 = (insn32 >> 12) & 7u;
  const char* opcode = kOpcodes[funct3];
  if (opcode == nullptr) {
    os_ << "<unknown32>";
    return;
  }

  // Print shorter macro instruction notation if available.
  uint32_t rs1 = GetRs1(insn32);
  uint32_t rs2 = GetRs2(insn32);
  if (rs2 == Zero) {
    os_ << opcode << "z " << XRegName(rs1);
  } else if (rs1 == Zero && (funct3 == 4u || funct3 == 5u)) {
    // blt zero, rs2, offset ... bgtz rs2, offset
    // bge zero, rs2, offset ... blez rs2, offset
    os_ << (funct3 == 4u ? "bgtz " : "blez ") << XRegName(rs2);
  } else {
    os_ << opcode << " " << XRegName(rs1) << ", " << XRegName(rs2);
  }
  os_ << ", ";

  uint32_t bit12 = insn32 >> 31;
  uint32_t bits5_10 = (insn32 >> 25) & 0x3fu;
  uint32_t bits1_4 = (insn32 >> 8) & 0xfu;
  uint32_t bit11 = (insn32 >> 7) & 1u;
  uint32_t imm = (bit12 << 12) + (bit11 << 11) + (bits5_10 << 5) + (bits1_4 << 1);
  int32_t offset = static_cast<int32_t>(imm) - static_cast<int32_t>(bit12 << 13);  // Sign-extend.
  PrintBranchOffset(offset);
  os_ << " ; " << disassembler_->FormatInstructionPointer(insn + offset);
}

void DisassemblerRiscv64::Printer::Print32Load(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x03u);
  static const char* const kOpcodes[] = {
      "lb", "lh", "lw", "ld", "lbu", "lhu", "lwu", nullptr
  };
  uint32_t funct3 = (insn32 >> 12) & 7u;
  const char* opcode = kOpcodes[funct3];
  if (opcode == nullptr) {
    os_ << "<unknown32>";
    return;
  }

  os_ << opcode << " " << XRegName(GetRd(insn32)) << ", ";
  PrintLoadStoreAddress(GetRs1(insn32), Decode32Imm12(insn32));

  // TODO(riscv64): If previous instruction is AUIPC for current `rs1` and we load
  // from the range specified by assembler options, print the loaded literal.
}

void DisassemblerRiscv64::Printer::Print32Store(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x23u);
  static const char* const kOpcodes[] = {
      "sb", "sh", "sw", "sd", nullptr, nullptr, nullptr, nullptr
  };
  uint32_t funct3 = (insn32 >> 12) & 7u;
  const char* opcode = kOpcodes[funct3];
  if (opcode == nullptr) {
    os_ << "<unknown32>";
    return;
  }

  os_ << opcode << " " << XRegName(GetRs2(insn32)) << ", ";
  PrintLoadStoreAddress(GetRs1(insn32), Decode32StoreOffset(insn32));
}

void DisassemblerRiscv64::Printer::Print32FLoad(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x07u);
  static const char* const kOpcodes[] = {
      nullptr, nullptr, "flw", "fld", nullptr, nullptr, nullptr, nullptr
  };
  uint32_t funct3 = (insn32 >> 12) & 7u;
  const char* opcode = kOpcodes[funct3];
  if (opcode == nullptr) {
    os_ << "<unknown32>";
    return;
  }

  os_ << opcode << " " << FRegName(GetRd(insn32)) << ", ";
  PrintLoadStoreAddress(GetRs1(insn32), Decode32Imm12(insn32));

  // TODO(riscv64): If previous instruction is AUIPC for current `rs1` and we load
  // from the range specified by assembler options, print the loaded literal.
}

void DisassemblerRiscv64::Printer::Print32FStore(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x27u);
  static const char* const kOpcodes[] = {
      nullptr, nullptr, "fsw", "fsd", nullptr, nullptr, nullptr, nullptr
  };
  uint32_t funct3 = (insn32 >> 12) & 7u;
  const char* opcode = kOpcodes[funct3];
  if (opcode == nullptr) {
    os_ << "<unknown32>";
    return;
  }

  os_ << opcode << " " << FRegName(GetRs2(insn32)) << ", ";
  PrintLoadStoreAddress(GetRs1(insn32), Decode32StoreOffset(insn32));
}

void DisassemblerRiscv64::Printer::Print32BinOpImm(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x77u, 0x13u);  // Note: Bit 0x8 selects narrow binop.
  bool narrow = (insn32 & 0x8u) != 0u;
  uint32_t funct3 = (insn32 >> 12) & 7u;
  uint32_t rd = GetRd(insn32);
  uint32_t rs1 = GetRs1(insn32);
  int32_t imm = Decode32Imm12(insn32);

  // Print shorter macro instruction notation if available.
  if (funct3 == /*ADDI*/ 0u && imm == 0u) {
    if (narrow) {
      os_ << "sextw " << XRegName(rd) << ", " << XRegName(rs1);
    } else if (rd == Zero && rs1 == Zero) {
      os_ << "nop";  // Only canonical nop. Non-Zero `rd == rs1` nops are printed as "mv".
    } else {
      os_ << "mv " << XRegName(rd) << ", " << XRegName(rs1);
    }
  } else if (!narrow && funct3 == /*XORI*/ 4u && imm == -1) {
    os_ << "not " << XRegName(rd) << ", " << XRegName(rs1);
  } else if (!narrow && funct3 == /*ANDI*/ 7u && imm == 0xff) {
    os_ << "zextb " << XRegName(rd) << ", " << XRegName(rs1);
  } else if (!narrow && funct3 == /*SLTIU*/ 3u && imm == 1) {
    os_ << "seqz " << XRegName(rd) << ", " << XRegName(rs1);
  } else if ((insn32 & 0xfc00707fu) == 0x0800101bu) {
    os_ << "slli.uw " << XRegName(rd) << ", " << XRegName(rs1) << ", " << (imm & 0x3fu);
  } else if ((imm ^ 0x600u) < 3u && funct3 == 1u) {
    static const char* const kBitOpcodes[] = { "clz", "ctz", "cpop" };
    os_ << kBitOpcodes[imm ^ 0x600u] << (narrow ? "w " : " ")
        << XRegName(rd) << ", " << XRegName(rs1);
  } else if ((imm ^ 0x600u) < (narrow ? 32 : 64) && funct3 == 5u) {
    os_ << "rori" << (narrow ? "w " : " ")
        << XRegName(rd) << ", " << XRegName(rs1) << ", " << (imm ^ 0x600u);
  } else if (imm == 0x287u && !narrow && funct3 == 5u) {
    os_ << "orc.b " << XRegName(rd) << ", " << XRegName(rs1);
  } else if (imm == 0x6b8u && !narrow && funct3 == 5u) {
    os_ << "rev8 " << XRegName(rd) << ", " << XRegName(rs1);
  } else {
    bool bad_high_bits = false;
    if (funct3 == /*SLLI*/ 1u || funct3 == /*SRLI/SRAI*/ 5u) {
      imm &= (narrow ? 0x1fu : 0x3fu);
      uint32_t high_bits = insn32 & (narrow ? 0xfe000000u : 0xfc000000u);
      if (high_bits == 0x40000000u && funct3 == /*SRAI*/ 5u) {
        os_ << "srai";
      } else {
        os_ << ((funct3 == /*SRLI*/ 5u) ? "srli" : "slli");
        bad_high_bits = (high_bits != 0u);
      }
    } else if (!narrow || funct3 == /*ADDI*/ 0u) {
      static const char* const kOpcodes[] = {
          "addi", nullptr, "slti", "sltiu", "xori", nullptr, "ori", "andi"
      };
      DCHECK(kOpcodes[funct3] != nullptr);
      os_ << kOpcodes[funct3];
    } else {
      os_ << "<unknown32>";  // There is no SLTIW/SLTIUW/XORIW/ORIW/ANDIW.
      return;
    }
    os_ << (narrow ? "w " : " ") << XRegName(rd) << ", " << XRegName(rs1) << ", " << imm;
    if (bad_high_bits) {
      os_ << " (invalid high bits)";
    }
  }
}

void DisassemblerRiscv64::Printer::Print32BinOp(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x77u, 0x33u);  // Note: Bit 0x8 selects narrow binop.
  bool narrow = (insn32 & 0x8u) != 0u;
  uint32_t funct3 = (insn32 >> 12) & 7u;
  uint32_t rd = GetRd(insn32);
  uint32_t rs1 = GetRs1(insn32);
  uint32_t rs2 = GetRs2(insn32);
  uint32_t high_bits = insn32 & 0xfe000000u;

  // Print shorter macro instruction notation if available.
  if (high_bits == 0x40000000u && funct3 == /*SUB*/ 0u && rs1 == Zero) {
    os_ << (narrow ? "negw " : "neg ") << XRegName(rd) << ", " << XRegName(rs2);
  } else if (!narrow && funct3 == /*SLT*/ 2u && rs2 == Zero) {
    os_ << "sltz " << XRegName(rd) << ", " << XRegName(rs1);
  } else if (!narrow && funct3 == /*SLT*/ 2u && rs1 == Zero) {
    os_ << "sgtz " << XRegName(rd) << ", " << XRegName(rs2);
  } else if (!narrow && funct3 == /*SLTU*/ 3u && rs1 == Zero) {
    os_ << "snez " << XRegName(rd) << ", " << XRegName(rs2);
  } else if (narrow && high_bits == 0x08000000u && funct3 == /*ADD.UW*/ 0u && rs2 == Zero) {
    os_ << "zext.w " << XRegName(rd) << ", " << XRegName(rs1);
  } else {
    bool bad_high_bits = false;
    if (high_bits == 0x40000000u && (funct3 == /*SUB*/ 0u || funct3 == /*SRA*/ 5u)) {
      os_ << ((funct3 == /*SUB*/ 0u) ? "sub" : "sra");
    } else if (high_bits == 0x02000000u &&
               (!narrow || (funct3 == /*MUL*/ 0u || funct3 >= /*DIV/DIVU/REM/REMU*/ 4u))) {
      static const char* const kOpcodes[] = {
          "mul", "mulh", "mulhsu", "mulhu", "div", "divu", "rem", "remu"
      };
      os_ << kOpcodes[funct3];
    } else if (high_bits == 0x08000000u && narrow && funct3 == /*ADD.UW*/ 0u) {
      os_ << "add.u";  // "w" is added below.
    } else if (high_bits == 0x20000000u && (funct3 & 1u) == 0u && funct3 != 0u) {
      static const char* const kZbaOpcodes[] = { nullptr, "sh1add", "sh2add", "sh3add" };
      DCHECK(kZbaOpcodes[funct3 >> 1] != nullptr);
      os_ << kZbaOpcodes[funct3 >> 1] << (narrow ? ".u" /* "w" is added below. */ : "");
    } else if (high_bits == 0x40000000u && !narrow && funct3 >= 4u && funct3 != 5u) {
      static const char* const kZbbNegOpcodes[] = { "xnor", nullptr, "orn", "andn" };
      DCHECK(kZbbNegOpcodes[funct3 - 4u] != nullptr);
      os_ << kZbbNegOpcodes[funct3 - 4u];
    } else if (high_bits == 0x0a000000u && !narrow && funct3 >= 4u) {
      static const char* const kZbbMinMaxOpcodes[] = { "min", "minu", "max", "maxu" };
      DCHECK(kZbbMinMaxOpcodes[funct3 - 4u] != nullptr);
      os_ << kZbbMinMaxOpcodes[funct3 - 4u];
    } else if (high_bits == 0x60000000u && (funct3 == /*ROL*/ 1u || funct3 == /*ROL*/ 5u)) {
      os_ << (funct3 == /*ROL*/ 1u ? "rol" : "ror");
    } else if (!narrow || (funct3 == /*ADD*/ 0u || funct3 == /*SLL*/ 1u || funct3 == /*SRL*/ 5u)) {
      static const char* const kOpcodes[] = {
          "add", "sll", "slt", "sltu", "xor", "srl", "or", "and"
      };
      os_ << kOpcodes[funct3];
      bad_high_bits = (high_bits != 0u);
    } else {
      DCHECK(narrow);
      os_ << "<unknown32>";  // Some of the above instructions do not have a narrow version.
      return;
    }
    os_ << (narrow ? "w " : " ") << XRegName(rd) << ", " << XRegName(rs1) << ", " << XRegName(rs2);
    if (bad_high_bits) {
      os_ << " (invalid high bits)";
    }
  }
}

void DisassemblerRiscv64::Printer::Print32Atomic(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x2fu);
  uint32_t funct3 = (insn32 >> 12) & 7u;
  uint32_t funct5 = (insn32 >> 27);
  if ((funct3 != 2u && funct3 != 3u) ||  // There are only 32-bit and 64-bit LR/SC/AMO*.
      (((funct5 & 3u) != 0u) && funct5 >= 4u)) {  // Only multiples of 4, or 1-3.
    os_ << "<unknown32>";
    return;
  }
  static const char* const kMul4Opcodes[] = {
      "amoadd", "amoxor", "amoor", "amoand", "amomin", "amomax", "amominu", "amomaxu"
  };
  static const char* const kOtherOpcodes[] = {
      nullptr, "amoswap", "lr", "sc"
  };
  const char* opcode = ((funct5 & 3u) == 0u) ? kMul4Opcodes[funct5 >> 2] : kOtherOpcodes[funct5];
  DCHECK(opcode != nullptr);
  uint32_t rd = GetRd(insn32);
  uint32_t rs1 = GetRs1(insn32);
  uint32_t rs2 = GetRs2(insn32);
  const char* type = (funct3 == 2u) ? ".w" : ".d";
  const char* aq = (((insn32 >> 26) & 1u) != 0u) ? ".aq" : "";
  const char* rl = (((insn32 >> 25) & 1u) != 0u) ? ".rl" : "";
  os_ << opcode << type << aq << rl << " " << XRegName(rd) << ", " << XRegName(rs1);
  if (funct5 == /*LR*/ 2u) {
    if (rs2 != 0u) {
      os_ << " (bad rs2)";
    }
  } else {
    os_ << ", " << XRegName(rs2);
  }
}

void DisassemblerRiscv64::Printer::Print32FpOp(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x4fu);
  uint32_t rd = GetRd(insn32);
  uint32_t rs1 = GetRs1(insn32);
  uint32_t rs2 = GetRs2(insn32);  // Sometimes used to to differentiate opcodes.
  uint32_t rm = GetRoundingMode(insn32);  // Sometimes used to to differentiate opcodes.
  uint32_t funct7 = insn32 >> 25;
  const char* type = ((funct7 & 1u) != 0u) ? ".d" : ".s";
  if ((funct7 & 2u) != 0u) {
    os_ << "<unknown32>";  // Note: This includes the "H" and "Q" extensions.
    return;
  }
  switch (funct7 >> 2) {
    case 0u:
    case 1u:
    case 2u:
    case 3u: {
      static const char* const kOpcodes[] = { "fadd", "fsub", "fmul", "fdiv" };
      os_ << kOpcodes[funct7 >> 2] << type << RoundingModeName(rm) << " "
          << FRegName(rd) << ", " << FRegName(rs1) << ", " << FRegName(rs2);
      return;
    }
    case 4u: {  // FSGN*
      // Print shorter macro instruction notation if available.
      static const char* const kOpcodes[] = { "fsgnj", "fsgnjn", "fsgnjx" };
      if (rm < std::size(kOpcodes)) {
        if (rs1 == rs2) {
          static const char* const kAltOpcodes[] = { "fmv", "fneg", "fabs" };
          static_assert(std::size(kOpcodes) == std::size(kAltOpcodes));
          os_ << kAltOpcodes[rm] << type << " " << FRegName(rd) << ", " << FRegName(rs1);
        } else {
          os_ << kOpcodes[rm] << type << " "
              << FRegName(rd) << ", " << FRegName(rs1) << ", " << FRegName(rs2);
        }
        return;
      }
      break;
    }
    case 5u: {  // FMIN/FMAX
      static const char* const kOpcodes[] = { "fmin", "fmax" };
      if (rm < std::size(kOpcodes)) {
        os_ << kOpcodes[rm] << type << " "
            << FRegName(rd) << ", " << FRegName(rs1) << ", " << FRegName(rs2);
        return;
      }
      break;
    }
    case 0x8u:  // FCVT between FP numbers.
      if ((rs2 ^ 1u) == (funct7 & 1u)) {
        os_ << ((rs2 != 0u) ? "fcvt.s.d" : "fcvt.d.s") << RoundingModeName(rm) << " "
            << FRegName(rd) << ", " << FRegName(rs1);
      }
      break;
    case 0xbu:
      if (rs2 == 0u) {
        os_ << "fsqrt" << type << RoundingModeName(rm) << " "
            << FRegName(rd) << ", " << FRegName(rs1);
        return;
      }
      break;
    case 0x14u: {  // FLE/FLT/FEQ
      static const char* const kOpcodes[] = { "fle", "flt", "feq" };
      if (rm < std::size(kOpcodes)) {
        os_ << kOpcodes[rm] << type << " "
            << XRegName(rd) << ", " << FRegName(rs1) << ", " << FRegName(rs2);
        return;
      }
      break;
    }
    case 0x18u: {  // FCVT from floating point numbers to integers
      static const char* const kIntTypes[] = { "w", "wu", "l", "lu" };
      if (rs2 < std::size(kIntTypes)) {
        os_ << "fcvt." << kIntTypes[rs2] << type << RoundingModeName(rm) << " "
            << XRegName(rd) << ", " << FRegName(rs1);
        return;
      }
      break;
    }
    case 0x1au: {  // FCVT from integers to floating point numbers
      static const char* const kIntTypes[] = { "w", "wu", "l", "lu" };
      if (rs2 < std::size(kIntTypes)) {
        os_ << "fcvt" << type << "." << kIntTypes[rs2] << RoundingModeName(rm) << " "
            << FRegName(rd) << ", " << XRegName(rs1);
        return;
      }
      break;
    }
    case 0x1cu:  // FMV from FPR to GPR, or FCLASS
      if (rs2 == 0u && rm == 0u) {
        os_ << (((funct7 & 1u) != 0u) ? "fmv.x.d " : "fmv.x.w ")
            << XRegName(rd) << ", " << FRegName(rs1);
        return;
      } else if (rs2 == 0u && rm == 1u) {
        os_ << "fclass" << type << " " << XRegName(rd) << ", " << FRegName(rs1);
        return;
      }
      break;
    case 0x1eu:  // FMV from GPR to FPR
      if (rs2 == 0u && rm == 0u) {
        os_ << (((funct7 & 1u) != 0u) ? "fmv.d.x " : "fmv.w.x ")
            << FRegName(rd) << ", " << XRegName(rs1);
        return;
      }
      break;
    default:
      break;
  }
  os_ << "<unknown32>";
}

void DisassemblerRiscv64::Printer::Print32FpFma(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x73u, 0x43u);  // Note: Bits 0xc select the FMA opcode.
  uint32_t funct2 = (insn32 >> 25) & 3u;
  if (funct2 >= 2u) {
    os_ << "<unknown32>";  // Note: This includes the "H" and "Q" extensions.
    return;
  }
  static const char* const kOpcodes[] = { "fmadd", "fmsub", "fnmsub", "fnmadd" };
  os_ << kOpcodes[(insn32 >> 2) & 3u] << ((funct2 != 0u) ? ".d" : ".s")
      << RoundingModeName(GetRoundingMode(insn32)) << " "
      << FRegName(GetRd(insn32)) << ", " << FRegName(GetRs1(insn32)) << ", "
      << FRegName(GetRs2(insn32)) << ", " << FRegName(GetRs3(insn32));
}

void DisassemblerRiscv64::Printer::Print32Zicsr(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x73u);
  uint32_t funct3 = (insn32 >> 12) & 7u;
  static const char* const kOpcodes[] = {
      nullptr, "csrrw", "csrrs", "csrrc", nullptr, "csrrwi", "csrrsi", "csrrci"
  };
  const char* opcode = kOpcodes[funct3];
  if (opcode == nullptr) {
    os_ << "<unknown32>";
    return;
  }
  uint32_t rd = GetRd(insn32);
  uint32_t rs1_or_uimm = GetRs1(insn32);
  uint32_t csr = insn32 >> 20;
  // Print shorter macro instruction notation if available.
  if (funct3 == /*CSRRW*/ 1u && rd == 0u && rs1_or_uimm == 0u && csr == 0xc00u) {
    os_ << "unimp";
    return;
  } else if (funct3 == /*CSRRS*/ 2u && rs1_or_uimm == 0u) {
    if (csr == 0xc00u) {
      os_ << "rdcycle " << XRegName(rd);
    } else if (csr == 0xc01u) {
      os_ << "rdtime " << XRegName(rd);
    } else if (csr == 0xc02u) {
      os_ << "rdinstret " << XRegName(rd);
    } else {
      os_ << "csrr " << XRegName(rd) << ", " << csr;
    }
    return;
  }

  if (rd == 0u) {
    static const char* const kAltOpcodes[] = {
        nullptr, "csrw", "csrs", "csrc", nullptr, "csrwi", "csrsi", "csrci"
    };
    DCHECK(kAltOpcodes[funct3] != nullptr);
    os_ << kAltOpcodes[funct3] << " " << csr << ", ";
  } else {
    os_ << opcode << " " << XRegName(rd) << ", " << csr << ", ";
  }
  if (funct3 >= /*CSRRWI/CSRRSI/CSRRCI*/ 4u) {
    os_ << rs1_or_uimm;
  } else {
    os_ << XRegName(rs1_or_uimm);
  }
}

void DisassemblerRiscv64::Printer::Print32Fence(uint32_t insn32) {
  DCHECK_EQ(insn32 & 0x7fu, 0x0fu);
  if ((insn32 & 0xf00fffffu) == 0x0000000fu) {
    auto print_flags = [&](uint32_t flags) {
      if (flags == 0u) {
        os_ << "0";
      } else {
        DCHECK_LT(flags, 0x10u);
        static const char kFlagNames[] = "wroi";
        for (size_t bit : { 3u, 2u, 1u, 0u }) {  // Print in the "iorw" order.
          if ((flags & (1u << bit)) != 0u) {
            os_ << kFlagNames[bit];
          }
        }
      }
    };
    os_ << "fence.";
    print_flags((insn32 >> 24) & 0xfu);
    os_ << ".";
    print_flags((insn32 >> 20) & 0xfu);
  } else if (insn32 == 0x8330000fu) {
    os_ << "fence.tso";
  } else if (insn32 == 0x0000100fu) {
    os_ << "fence.i";
  } else {
    os_ << "<unknown32>";
  }
}

void DisassemblerRiscv64::Printer::Dump32(const uint8_t* insn) {
  uint32_t insn32 = static_cast<uint32_t>(insn[0]) +
                    (static_cast<uint32_t>(insn[1]) << 8) +
                    (static_cast<uint32_t>(insn[2]) << 16) +
                    (static_cast<uint32_t>(insn[3]) << 24);
  CHECK_EQ(insn32 & 3u, 3u);
  os_ << disassembler_->FormatInstructionPointer(insn) << StringPrintf(": %08x\t", insn32);
  switch (insn32 & 0x7fu) {
    case 0x37u:
      Print32Lui(insn32);
      break;
    case 0x17u:
      Print32Auipc(insn, insn32);
      break;
    case 0x6fu:
      Print32Jal(insn, insn32);
      break;
    case 0x67u:
      switch ((insn32 >> 12) & 7u) {  // funct3
        case 0:
          Print32Jalr(insn, insn32);
          break;
        default:
          os_ << "<unknown32>";
          break;
      }
      break;
    case 0x63u:
      Print32BCond(insn, insn32);
      break;
    case 0x03u:
      Print32Load(insn32);
      break;
    case 0x23u:
      Print32Store(insn32);
      break;
    case 0x07u:
      Print32FLoad(insn32);
      break;
    case 0x27u:
      Print32FStore(insn32);
      break;
    case 0x13u:
    case 0x1bu:
      Print32BinOpImm(insn32);
      break;
    case 0x33u:
    case 0x3bu:
      Print32BinOp(insn32);
      break;
    case 0x2fu:
      Print32Atomic(insn32);
      break;
    case 0x53u:
      Print32FpOp(insn32);
      break;
    case 0x43u:
    case 0x47u:
    case 0x4bu:
    case 0x4fu:
      Print32FpFma(insn32);
      break;
    case 0x73u:
      if ((insn32 & 0xffefffffu) == 0x00000073u) {
        os_ << ((insn32 == 0x00000073u) ? "ecall" : "ebreak");
      } else {
        Print32Zicsr(insn32);
      }
      break;
    case 0x0fu:
      Print32Fence(insn32);
      break;
    default:
      // TODO(riscv64): Disassemble more instructions.
      os_ << "<unknown32>";
      break;
  }
  os_ << "\n";
}

void DisassemblerRiscv64::Printer::Dump16(const uint8_t* insn) {
  uint32_t insn16 = static_cast<uint32_t>(insn[0]) + (static_cast<uint32_t>(insn[1]) << 8);
  CHECK_NE(insn16 & 3u, 3u);
  // TODO(riscv64): Disassemble instructions from the "C" extension.
  os_ << disassembler_->FormatInstructionPointer(insn)
      << StringPrintf(": %04x    \t<unknown16>\n", insn16);
}

void DisassemblerRiscv64::Printer::Dump2Byte(const uint8_t* data) {
  uint32_t value = data[0] + (data[1] << 8);
  os_ << disassembler_->FormatInstructionPointer(data)
      << StringPrintf(": %04x    \t.2byte %u\n", value, value);
}

void DisassemblerRiscv64::Printer::DumpByte(const uint8_t* data) {
  uint32_t value = *data;
  os_ << disassembler_->FormatInstructionPointer(data)
      << StringPrintf(": %02x      \t.byte %u\n", value, value);
}

size_t DisassemblerRiscv64::Dump(std::ostream& os, const uint8_t* begin) {
  if (begin < GetDisassemblerOptions()->base_address_ ||
      begin >= GetDisassemblerOptions()->end_address_) {
    return 0u;  // Outside the range.
  }
  Printer printer(this, os);
  if (!IsAligned<2u>(begin) || GetDisassemblerOptions()->end_address_ - begin == 1) {
    printer.DumpByte(begin);
    return 1u;
  }
  if ((*begin & 3u) == 3u) {
    if (GetDisassemblerOptions()->end_address_ - begin >= 4) {
      printer.Dump32(begin);
      return 4u;
    } else {
      printer.Dump2Byte(begin);
      return 2u;
    }
  } else {
    printer.Dump16(begin);
    return 2u;
  }
}

void DisassemblerRiscv64::Dump(std::ostream& os, const uint8_t* begin, const uint8_t* end) {
  Printer printer(this, os);
  const uint8_t* cur = begin;
  if (cur < end && !IsAligned<2u>(cur)) {
    // Unaligned, dump as a `.byte` to get to an aligned address.
    printer.DumpByte(cur);
    cur += 1;
  }
  if (cur >= end) {
    return;
  }
  while (end - cur >= 4) {
    if ((*cur & 3u) == 3u) {
      printer.Dump32(cur);
      cur += 4;
    } else {
      printer.Dump16(cur);
      cur += 2;
    }
  }
  if (end - cur >= 2) {
    if ((*cur & 3u) == 3u) {
      // Not enough data for a 32-bit instruction. Dump as `.2byte`.
      printer.Dump2Byte(cur);
    } else {
      printer.Dump16(cur);
    }
    cur += 2;
  }
  if (end != cur) {
    CHECK_EQ(end - cur, 1);
    printer.DumpByte(cur);
  }
}

}  // namespace riscv64
}  // namespace art