source/binary.cpp

// Copyright (c) 2015 The Khronos Group Inc.
//
// Permission is hereby granted, free of charge, to any person obtaining a
// copy of this software and/or associated documentation files (the
// "Materials"), to deal in the Materials without restriction, including
// without limitation the rights to use, copy, modify, merge, publish,
// distribute, sublicense, and/or sell copies of the Materials, and to
// permit persons to whom the Materials are furnished to do so, subject to
// the following conditions:
//
// The above copyright notice and this permission notice shall be included
// in all copies or substantial portions of the Materials.
//
// MODIFICATIONS TO THIS FILE MAY MEAN IT NO LONGER ACCURATELY REFLECTS
// KHRONOS STANDARDS. THE UNMODIFIED, NORMATIVE VERSIONS OF KHRONOS
// SPECIFICATIONS AND HEADER INFORMATION ARE LOCATED AT
//    https://www.khronos.org/registry/
//
// THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
// IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
// CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
// TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
// MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS.

#include "binary.h"

#include <cassert>
#include <cstring>
#include <limits>
#include <unordered_map>

#include <libspirv/libspirv.h>
#include "assembly_grammar.h"
#include "diagnostic.h"
#include "endian.h"
#include "ext_inst.h"
#include "opcode.h"
#include "operand.h"

spv_result_t spvBinaryHeaderGet(const spv_const_binary binary,
                                const spv_endianness_t endian,
                                spv_header_t* pHeader) {
  if (!binary->code) return SPV_ERROR_INVALID_BINARY;
  if (binary->wordCount < SPV_INDEX_INSTRUCTION)
    return SPV_ERROR_INVALID_BINARY;
  if (!pHeader) return SPV_ERROR_INVALID_POINTER;

  // TODO: Validation checking?
  pHeader->magic = spvFixWord(binary->code[SPV_INDEX_MAGIC_NUMBER], endian);
  pHeader->version = spvFixWord(binary->code[SPV_INDEX_VERSION_NUMBER], endian);
  pHeader->generator =
      spvFixWord(binary->code[SPV_INDEX_GENERATOR_NUMBER], endian);
  pHeader->bound = spvFixWord(binary->code[SPV_INDEX_BOUND], endian);
  pHeader->schema = spvFixWord(binary->code[SPV_INDEX_SCHEMA], endian);
  pHeader->instructions = &binary->code[SPV_INDEX_INSTRUCTION];

  return SPV_SUCCESS;
}

// TODO(dneto): This API is not powerful enough in the case that the
// number and type of operands are not known until partway through parsing
// the operation.  This happens when enum operands might have different number
// of operands, or with extended instructions.
spv_operand_type_t spvBinaryOperandInfo(const uint32_t word,
                                        const uint16_t operandIndex,
                                        const spv_opcode_desc opcodeEntry,
                                        const spv_operand_table operandTable,
                                        spv_operand_desc* pOperandEntry) {
  spv_operand_type_t type;
  if (operandIndex < opcodeEntry->numTypes) {
    // NOTE: Do operand table lookup to set operandEntry if successful
    uint16_t index = operandIndex - 1;
    type = opcodeEntry->operandTypes[index];
    spv_operand_desc entry = nullptr;
    if (!spvOperandTableValueLookup(operandTable, type, word, &entry)) {
      if (SPV_OPERAND_TYPE_NONE != entry->operandTypes[0]) {
        *pOperandEntry = entry;
      }
    }
  } else if (*pOperandEntry) {
    // NOTE: Use specified operand entry operand type for this word
    uint16_t index = operandIndex - opcodeEntry->numTypes;
    type = (*pOperandEntry)->operandTypes[index];
  } else if (SpvOpSwitch == opcodeEntry->opcode) {
    // NOTE: OpSwitch is a special case which expects a list of paired extra
    // operands
    assert(0 &&
           "This case is previously untested, remove this assert and ensure it "
           "is behaving correctly!");
    uint16_t lastIndex = opcodeEntry->numTypes - 1;
    uint16_t index = lastIndex + ((operandIndex - lastIndex) % 2);
    type = opcodeEntry->operandTypes[index];
  } else {
    // NOTE: Default to last operand type in opcode entry
    uint16_t index = opcodeEntry->numTypes - 1;
    type = opcodeEntry->operandTypes[index];
  }
  return type;
}

namespace {

// A SPIR-V binary parser.  A parser instance communicates detailed parse
// results via callbacks.
class Parser {
 public:
  // The user_data value is provided to the callbacks as context.
  Parser(void* user_data, spv_parsed_header_fn_t parsed_header_fn,
         spv_parsed_instruction_fn_t parsed_instruction_fn)
      : user_data_(user_data),
        parsed_header_fn_(parsed_header_fn),
        parsed_instruction_fn_(parsed_instruction_fn) {}

  // Parses the specified binary SPIR-V module, issuing callbacks on a parsed
  // header and for each parsed instruction.  Returns SPV_SUCCESS on success.
  // Otherwise returns an error code and issues a diagnostic.
  spv_result_t parse(const uint32_t* words, size_t num_words,
                     spv_diagnostic* diagnostic);

 private:
  // All remaining methods work on the current module parse state.

  // Like the parse method, but works on the current module parse state.
  spv_result_t parseModule();

  // Parses an instruction at the current position of the binary.  Assumes
  // the header has been parsed, the endian has been set, and the word index is
  // still in range.  Advances the parsing position past the instruction, and
  // updates other parsing state for the current module.
  // On success, returns SPV_SUCCESS and issues the parsed-instruction callback.
  // On failure, returns an error code and issues a diagnostic.
  spv_result_t parseInstruction();

  // Parses an instruction operand with the given type.
  // May update the expected_operands parameter, and the scalar members of the
  // inst parameter. On success, returns SPV_SUCCESS, advances past the
  // operand, and pushes a new entry on to the operands vector.  Otherwise
  // returns an error code and issues a diagnostic.
  spv_result_t parseOperand(spv_parsed_instruction_t* inst,
                            const spv_operand_type_t type,
                            std::vector<spv_parsed_operand_t>* operands,
                            spv_operand_pattern_t* expected_operands);

  // Records the numeric type for an operand according to the type information
  // associated with the given non-zero type Id.  This can fail if the type Id
  // is not a type Id, or if the type Id does not reference a scalar numeric
  // type.  On success, return SPV_SUCCESS and populates the num_words,
  // number_kind, and number_bit_width fields of parsed_operand.
  spv_result_t setNumericTypeInfoForType(spv_parsed_operand_t* parsed_operand,
                                         uint32_t type_id);

  // Records the number type for an instruction if that instruction generates
  // a type.  For types that aren't scalar numbers, record something with
  // number kind SPV_NUMBER_NONE.
  void recordNumberType(const spv_parsed_instruction_t* inst);

  // Returns a diagnostic stream object initialized with current position in
  // the input stream, and for the given error code. Any data written to the
  // returned object will be propagated to the current parse's diagnostic
  // object.
  DiagnosticStream diagnostic(spv_result_t error) {
    return DiagnosticStream({0, 0, _.word_index}, _.diagnostic, error);
  }

  // Returns a diagnostic stream object with the default parse error code.
  DiagnosticStream diagnostic() {
    // The default failure for parsing is invalid binary.
    return diagnostic(SPV_ERROR_INVALID_BINARY);
  }

  // Returns the endian-corrected word at the current position.
  uint32_t peek() const { return peekAt(_.word_index); }

  // Returns the endian-corrected word at the given position.
  uint32_t peekAt(size_t index) const {
    assert(index < _.num_words);
    return spvFixWord(_.words[index], _.endian);
  }

  // Data members

  const libspirv::AssemblyGrammar grammar_;        // SPIR-V syntax utility.
  void* const user_data_;                          // Context for the callbacks
  const spv_parsed_header_fn_t parsed_header_fn_;  // Parsed header callback
  const spv_parsed_instruction_fn_t
      parsed_instruction_fn_;  // Parsed instruction callback

  // Describes the format of a typed literal number.
  struct NumberType {
    spv_number_kind_t type;
    uint32_t bit_width;
  };

  // The state used to parse a single SPIR-V binary module.
  struct State {
    State(const uint32_t* words_arg, size_t num_words_arg,
          spv_diagnostic* diagnostic_arg)
        : words(words_arg),
          num_words(num_words_arg),
          diagnostic(diagnostic_arg),
          word_index(0),
          endian() {}
    State() : State(0, 0, nullptr) {}
    const uint32_t* words;       // Words in the binary SPIR-V module.
    size_t num_words;            // Number of words in the module.
    spv_diagnostic* diagnostic;  // Where diagnostics go.
    size_t word_index;           // The current position in words.
    spv_endianness_t endian;     // The endianness of the binary.

    // Maps a result ID to its type ID.  By convention:
    //  - a result ID that is a type definition maps to itself.
    //  - a result ID without a type maps to 0.  (E.g. for OpLabel)
    std::unordered_map<uint32_t, uint32_t> id_to_type_id;
    // Maps a type ID to its number type description.
    std::unordered_map<uint32_t, NumberType> type_id_to_number_type_info;
    // Maps an ExtInstImport id to the extended instruction type.
    std::unordered_map<uint32_t, spv_ext_inst_type_t>
        import_id_to_ext_inst_type;
  } _;
};

spv_result_t Parser::parse(const uint32_t* words, size_t num_words,
                           spv_diagnostic* diagnostic_arg) {
  _ = State(words, num_words, diagnostic_arg);

  const spv_result_t result = parseModule();

  // Clear the module state.  The tables might be big.
  _ = State();

  return result;
}

spv_result_t Parser::parseModule() {
  if (!_.words) return diagnostic() << "Missing module.";

  if (_.num_words < SPV_INDEX_INSTRUCTION)
    return diagnostic() << "Module has incomplete header: only " << _.num_words
                        << " words instead of " << SPV_INDEX_INSTRUCTION;

  // Check the magic number and detect the module's endianness.
  spv_const_binary_t binary{_.words, _.num_words};
  if (spvBinaryEndianness(&binary, &_.endian)) {
    return diagnostic() << "Invalid SPIR-V magic number '" << std::hex
                        << _.words[0] << "'.";
  }

  // Process the header.
  spv_header_t header;
  if (spvBinaryHeaderGet(&binary, _.endian, &header)) {
    // It turns out there is no way to trigger this error since the only
    // failure cases are already handled above, with better messages.
    return diagnostic(SPV_ERROR_INTERNAL)
           << "Internal error: unhandled header parse failure";
  }
  if (parsed_header_fn_) {
    if (auto error = parsed_header_fn_(user_data_, _.endian, header.magic,
                                       header.version, header.generator,
                                       header.bound, header.schema)) {
      return error;
    }
  }

  // Process the instructions.
  _.word_index = SPV_INDEX_INSTRUCTION;
  while (_.word_index < _.num_words)
    if (auto error = parseInstruction()) return error;

  // Running off the end should already have been reported earlier.
  assert(_.word_index == _.num_words);

  return SPV_SUCCESS;
}

spv_result_t Parser::parseInstruction() {
  // The zero values for all members except for opcode are the
  // correct initial values.
  spv_parsed_instruction_t inst = {};
  inst.offset = _.word_index;

  // After a successful parse of the instruction, the inst.operands member
  // will point to this vector's storage.
  // TODO(dneto): If it's too expensive to construct the operands vector for
  // each instruction, then make this a class data member instead, and clear it
  // here.
  std::vector<spv_parsed_operand_t> operands;
  // Most instructions have fewer than 25 logical operands.
  operands.reserve(25);

  assert(_.word_index < _.num_words);
  // Decompose and check the first word.
  uint16_t inst_word_count = 0;
  spvOpcodeSplit(peek(), &inst_word_count, &inst.opcode);
  if (inst_word_count < 1) {
    return diagnostic() << "Invalid instruction word count: "
                        << inst_word_count;
  }
  spv_opcode_desc opcode_desc;
  if (grammar_.lookupOpcode(inst.opcode, &opcode_desc))
    return diagnostic() << "Invalid opcode: " << int(inst.opcode);

  _.word_index++;

  // Maintains the ordered list of expected operand types.
  // For many instructions we only need the {numTypes, operandTypes}
  // entries in opcode_desc.  However, sometimes we need to modify
  // the list as we parse the operands. This occurs when an operand
  // has its own logical operands (such as the LocalSize operand for
  // ExecutionMode), or for extended instructions that may have their
  // own operands depending on the selected extended instruction.
  spv_operand_pattern_t expected_operands(
      opcode_desc->operandTypes,
      opcode_desc->operandTypes + opcode_desc->numTypes);

  while (_.word_index < inst.offset + inst_word_count) {
    const uint16_t inst_word_index = _.word_index - inst.offset;
    if (expected_operands.empty()) {
      return diagnostic() << "Invalid instruction Op" << opcode_desc->name
                          << " starting at word " << inst.offset
                          << ": expected no more operands after "
                          << inst_word_index
                          << " words, but stated word count is "
                          << inst_word_count << ".";
    }

    spv_operand_type_t type = spvTakeFirstMatchableOperand(&expected_operands);

    if (auto error = parseOperand(&inst, type, &operands, &expected_operands))
      return error;
  }

  if (!expected_operands.empty() &&
      !spvOperandIsOptional(expected_operands.front())) {
    return diagnostic() << "End of input reached while decoding Op"
                        << opcode_desc->name << " starting at word "
                        << inst.offset << ": expected more operands after "
                        << inst_word_count << " words.";
  }

  if ((inst.offset + inst_word_count) != _.word_index) {
    return diagnostic() << "Invalid word count: Instruction starting at word "
                        << inst.offset << " says it has " << inst_word_count
                        << " words, but found " << _.word_index - inst.offset
                        << " words instead.";
  }

  recordNumberType(&inst);

  // Must wait until here to set the inst.operands pointer because the vector
  // might be resized while we accumulate itse elements.
  inst.operands = operands.data();
  inst.num_operands = operands.size();

  // Issue the callback.  The callee should know that all the storage in inst
  // is transient, and will disappear immediately afterward.
  if (parsed_instruction_fn_) {
    if (auto error = parsed_instruction_fn_(user_data_, &inst)) return error;
  }

  return SPV_SUCCESS;
}

spv_result_t Parser::parseOperand(spv_parsed_instruction_t* inst,
                                  const spv_operand_type_t type,
                                  std::vector<spv_parsed_operand_t>* operands,
                                  spv_operand_pattern_t* expected_operands) {
  // We'll fill in this result as we go along.
  spv_parsed_operand_t parsed_operand;
  parsed_operand.offset = _.word_index - inst->offset;
  // Most operands occupy one word.  This might be be adjusted later.
  parsed_operand.num_words = 1;
  // The type argument is the one used by the grammar to parse the instruction.
  // But it can exposes internal parser details such as whether an operand is
  // optional or actually represents a variable-length sequence of operands.
  // The resulting type should be adjusted to avoid those internal details.
  // In most cases, the resulting operand type is the same as the grammar type.
  parsed_operand.type = type;

  // Assume non-numeric values.  This will be updated for literal numbers.
  parsed_operand.number_kind = SPV_NUMBER_NONE;
  parsed_operand.number_bit_width = 0;

  const uint32_t word = peek();

  switch (type) {
    case SPV_OPERAND_TYPE_TYPE_ID:
      if (!word) return diagnostic() << "Error: Type Id is 0";
      inst->type_id = word;
      break;

    case SPV_OPERAND_TYPE_RESULT_ID:
      if (!word) return diagnostic() << "Error: Result Id is 0";
      inst->result_id = word;
      // Save the result ID to type ID mapping.
      // In the grammar, type ID always appears before result ID.
      if (_.id_to_type_id.find(inst->result_id) != _.id_to_type_id.end())
        return diagnostic() << "Id " << inst->result_id
                            << " is defined more than once";
      // Record it.
      // A regular value maps to its type.  Some instructions (e.g. OpLabel)
      // have no type Id, and will map to 0.  The result Id for a
      // type-generating instruction (e.g. OpTypeInt) maps to itself.
      _.id_to_type_id[inst->result_id] = spvOpcodeGeneratesType(inst->opcode)
                                             ? inst->result_id
                                             : inst->type_id;
      break;

    case SPV_OPERAND_TYPE_ID:
    case SPV_OPERAND_TYPE_OPTIONAL_ID:
      if (!word) return diagnostic() << "Id is 0";
      parsed_operand.type = SPV_OPERAND_TYPE_ID;

      if (inst->opcode == SpvOpExtInst && parsed_operand.offset == 3) {
        // The current word is the extended instruction set Id.
        // Set the extended instruction set type for the current instruction.
        auto ext_inst_type_iter = _.import_id_to_ext_inst_type.find(word);
        if (ext_inst_type_iter == _.import_id_to_ext_inst_type.end()) {
          return diagnostic()
                 << "OpExtInst set Id " << word
                 << " does not reference an OpExtInstImport result Id";
        }
        inst->ext_inst_type = ext_inst_type_iter->second;
      }
      break;

    case SPV_OPERAND_TYPE_EXECUTION_SCOPE:
    case SPV_OPERAND_TYPE_MEMORY_SEMANTICS:
      if (!word) return diagnostic() << spvOperandTypeStr(type) << " Id is 0";
      break;

    case SPV_OPERAND_TYPE_EXTENSION_INSTRUCTION_NUMBER: {
      assert(SpvOpExtInst == inst->opcode);
      assert(inst->ext_inst_type != SPV_EXT_INST_TYPE_NONE);
      spv_ext_inst_desc ext_inst;
      if (grammar_.lookupExtInst(inst->ext_inst_type, word, &ext_inst))
        return diagnostic() << "Invalid extended instruction number: " << word;
      spvPrependOperandTypes(ext_inst->operandTypes, expected_operands);
    } break;

    case SPV_OPERAND_TYPE_SPEC_CONSTANT_OP_NUMBER: {
      assert(SpvOpSpecConstantOp == inst->opcode);
      if (grammar_.lookupSpecConstantOpcode(SpvOp(word))) {
        return diagnostic() << "Invalid " << spvOperandTypeStr(type) << ": "
                            << word;
      }
      spv_opcode_desc opcode_entry = nullptr;
      if (grammar_.lookupOpcode(SpvOp(word), &opcode_entry)) {
        return diagnostic(SPV_ERROR_INTERNAL)
               << "OpSpecConstant opcode table out of sync";
      }
      // OpSpecConstant opcodes must have a type and result. We've already
      // processed them, so skip them when preparing to parse the other
      // operants for the opcode.
      assert(opcode_entry->hasType);
      assert(opcode_entry->hasResult);
      assert(opcode_entry->numTypes >= 2);
      spvPrependOperandTypes(opcode_entry->operandTypes + 2, expected_operands);
    } break;

    case SPV_OPERAND_TYPE_LITERAL_INTEGER:
    case SPV_OPERAND_TYPE_OPTIONAL_LITERAL_INTEGER:
      // These are regular single-word literal integer operands.
      // Post-parsing validation should check the range of the parsed value.
      parsed_operand.type = SPV_OPERAND_TYPE_LITERAL_INTEGER;
      // It turns out they are always unsigned integers!
      parsed_operand.number_kind = SPV_NUMBER_UNSIGNED_INT;
      parsed_operand.number_bit_width = 32;
      break;

    case SPV_OPERAND_TYPE_TYPED_LITERAL_NUMBER:
    case SPV_OPERAND_TYPE_OPTIONAL_TYPED_LITERAL_INTEGER:
      parsed_operand.type = SPV_OPERAND_TYPE_TYPED_LITERAL_NUMBER;
      if (inst->opcode == SpvOpSwitch) {
        // The literal operands have the same type as the value
        // referenced by the selector Id.
        const uint32_t selector_id = peekAt(inst->offset + 1);
        auto type_id_iter = _.id_to_type_id.find(selector_id);
        if (type_id_iter == _.id_to_type_id.end()) {
          return diagnostic() << "Invalid OpSwitch: selector id " << selector_id
                              << " has no type";
        }
        uint32_t type_id = type_id_iter->second;

        if (selector_id == type_id) {
          // Recall that by convention, a result ID that is a type definition
          // maps to itself.
          return diagnostic() << "Invalid OpSwitch: selector id " << selector_id
                              << " is a type, not a value";
        }
        if (auto error = setNumericTypeInfoForType(&parsed_operand, type_id))
          return error;
        if (parsed_operand.number_kind != SPV_NUMBER_UNSIGNED_INT &&
            parsed_operand.number_kind != SPV_NUMBER_SIGNED_INT) {
          return diagnostic() << "Invalid OpSwitch: selector id " << selector_id
                              << " is not a scalar integer";
        }
      } else {
        assert(inst->opcode == SpvOpConstant ||
               inst->opcode == SpvOpSpecConstant);
        // The literal number type is determined by the type Id for the
        // constant.
        assert(inst->type_id);
        if (auto error =
                setNumericTypeInfoForType(&parsed_operand, inst->type_id))
          return error;
      }
      break;

    case SPV_OPERAND_TYPE_LITERAL_STRING:
    case SPV_OPERAND_TYPE_OPTIONAL_LITERAL_STRING: {
      // TODO(dneto): Make and use spvFixupString();
      const char* string =
          reinterpret_cast<const char*>(_.words + _.word_index);
      size_t string_num_words = (strlen(string) / 4) + 1;  // Account for null.
      // Make sure we can record the word count without overflow.
      // We still might have a string that's 64K words, but would still
      // make the instruction too long because of earlier operands.
      // That will be caught later at the end of the instruciton.
      if (string_num_words > std::numeric_limits<uint16_t>::max()) {
        return diagnostic() << "Literal string is longer than "
                            << std::numeric_limits<uint16_t>::max()
                            << " words: " << string_num_words << " words long";
      }
      parsed_operand.num_words = string_num_words;
      parsed_operand.type = SPV_OPERAND_TYPE_LITERAL_STRING;

      if (SpvOpExtInstImport == inst->opcode) {
        // Record the extended instruction type for the ID for this import.
        // There is only one string literal argument to OpExtInstImport,
        // so it's sufficient to guard this just on the opcode.
        const spv_ext_inst_type_t ext_inst_type =
            spvExtInstImportTypeGet(string);
        if (SPV_EXT_INST_TYPE_NONE == ext_inst_type) {
          return diagnostic() << "Invalid extended instruction import '"
                              << string << "'";
        }
        // We must have parsed a valid result ID.  It's a condition
        // of the grammar, and we only accept non-zero result Ids.
        assert(inst->result_id);
        _.import_id_to_ext_inst_type[inst->result_id] = ext_inst_type;
      }
    } break;

    case SPV_OPERAND_TYPE_OPTIONAL_EXECUTION_MODE:
      parsed_operand.type = SPV_OPERAND_TYPE_EXECUTION_MODE;
    // Fall through
    case SPV_OPERAND_TYPE_CAPABILITY:
    case SPV_OPERAND_TYPE_SOURCE_LANGUAGE:
    case SPV_OPERAND_TYPE_EXECUTION_MODEL:
    case SPV_OPERAND_TYPE_ADDRESSING_MODEL:
    case SPV_OPERAND_TYPE_MEMORY_MODEL:
    case SPV_OPERAND_TYPE_EXECUTION_MODE:
    case SPV_OPERAND_TYPE_STORAGE_CLASS:
    case SPV_OPERAND_TYPE_DIMENSIONALITY:
    case SPV_OPERAND_TYPE_SAMPLER_ADDRESSING_MODE:
    case SPV_OPERAND_TYPE_SAMPLER_FILTER_MODE:
    case SPV_OPERAND_TYPE_FP_ROUNDING_MODE:
    case SPV_OPERAND_TYPE_LINKAGE_TYPE:
    case SPV_OPERAND_TYPE_ACCESS_QUALIFIER:
    case SPV_OPERAND_TYPE_FUNCTION_PARAMETER_ATTRIBUTE:
    case SPV_OPERAND_TYPE_DECORATION:
    case SPV_OPERAND_TYPE_BUILT_IN:
    case SPV_OPERAND_TYPE_GROUP_OPERATION:
    case SPV_OPERAND_TYPE_KERNEL_ENQ_FLAGS:
    case SPV_OPERAND_TYPE_KERNEL_PROFILING_INFO: {
      // A single word that is a plain enum value.
      spv_operand_desc entry;
      if (grammar_.lookupOperand(type, word, &entry)) {
        return diagnostic() << "Invalid "
                            << spvOperandTypeStr(parsed_operand.type)
                            << " operand: " << word;
      }
      // Prepare to accept operands to this operand, if needed.
      spvPrependOperandTypes(entry->operandTypes, expected_operands);
    } break;

    case SPV_OPERAND_TYPE_FP_FAST_MATH_MODE:
    case SPV_OPERAND_TYPE_FUNCTION_CONTROL:
    case SPV_OPERAND_TYPE_LOOP_CONTROL:
    case SPV_OPERAND_TYPE_OPTIONAL_IMAGE:
    case SPV_OPERAND_TYPE_OPTIONAL_MEMORY_ACCESS:
    case SPV_OPERAND_TYPE_SELECTION_CONTROL: {
      // This operand is a mask.

      // Map an optional operand type to its corresponding concrete type.
      if (type == SPV_OPERAND_TYPE_OPTIONAL_IMAGE)
        parsed_operand.type = SPV_OPERAND_TYPE_IMAGE;
      else if (type == SPV_OPERAND_TYPE_OPTIONAL_MEMORY_ACCESS)
        parsed_operand.type = SPV_OPERAND_TYPE_MEMORY_ACCESS;

      // Check validity of set mask bits. Also prepare for operands for those
      // masks if they have any.  To get operand order correct, scan from
      // MSB to LSB since we can only prepend operands to a pattern.
      // The only case in the grammar where you have more than one mask bit
      // having an operand is for image operands.  See SPIR-V 3.14 Image
      // Operands.
      uint32_t remaining_word = word;
      for (uint32_t mask = (1u << 31); remaining_word; mask >>= 1) {
        if (remaining_word & mask) {
          spv_operand_desc entry;
          if (grammar_.lookupOperand(type, mask, &entry)) {
            return diagnostic()
                   << "Invalid " << spvOperandTypeStr(parsed_operand.type)
                   << " operand: " << word << " has invalid mask component "
                   << mask;
          }
          remaining_word ^= mask;
          spvPrependOperandTypes(entry->operandTypes, expected_operands);
        }
      }
      if (word == 0) {
        // An all-zeroes mask *might* also be valid.
        spv_operand_desc entry;
        if (SPV_SUCCESS == grammar_.lookupOperand(type, 0, &entry)) {
          // Prepare for its operands, if any.
          spvPrependOperandTypes(entry->operandTypes, expected_operands);
        }
      }
    } break;
    default:
      return diagnostic() << "Internal error: Unhandled operand type: " << type;
  }

  assert(int(SPV_OPERAND_TYPE_FIRST_CONCRETE_TYPE) <= int(parsed_operand.type));
  assert(int(SPV_OPERAND_TYPE_LAST_CONCRETE_TYPE) >= int(parsed_operand.type));

  operands->push_back(parsed_operand);

  _.word_index += parsed_operand.num_words;

  return SPV_SUCCESS;
}

spv_result_t Parser::setNumericTypeInfoForType(
    spv_parsed_operand_t* parsed_operand, uint32_t type_id) {
  assert(type_id != 0);
  auto type_info_iter = _.type_id_to_number_type_info.find(type_id);
  if (type_info_iter == _.type_id_to_number_type_info.end()) {
    return diagnostic() << "Type Id " << type_id << " is not a type";
  }
  const NumberType& info = type_info_iter->second;
  if (info.type == SPV_NUMBER_NONE) {
    // This is a valid type, but for something other than a scalar number.
    return diagnostic() << "Type Id " << type_id
                        << " is not a scalar numeric type";
  }

  parsed_operand->number_kind = info.type;
  parsed_operand->number_bit_width = info.bit_width;
  parsed_operand->num_words = (info.bit_width + 31) / 32;  // Round up
  return SPV_SUCCESS;
}

void Parser::recordNumberType(const spv_parsed_instruction_t* inst) {
  if (spvOpcodeGeneratesType(inst->opcode)) {
    NumberType info = {SPV_NUMBER_NONE, 0};
    if (SpvOpTypeInt == inst->opcode) {
      const bool is_signed = peekAt(inst->offset + 3) != 0;
      info.type = is_signed ? SPV_NUMBER_SIGNED_INT : SPV_NUMBER_UNSIGNED_INT;
      info.bit_width = peekAt(inst->offset + 2);
    } else if (SpvOpTypeFloat == inst->opcode) {
      info.type = SPV_NUMBER_FLOATING;
      info.bit_width = peekAt(inst->offset + 2);
    }
    // The *result* Id of a type generating instruction is the type Id.
    _.type_id_to_number_type_info[inst->result_id] = info;
  }
}

}  // anonymous namespace

spv_result_t spvBinaryParse(void* user_data, const uint32_t* code,
                            const size_t num_words,
                            spv_parsed_header_fn_t parsed_header,
                            spv_parsed_instruction_fn_t parsed_instruction,
                            spv_diagnostic* diagnostic) {
  Parser parser(user_data, parsed_header, parsed_instruction);
  return parser.parse(code, num_words, diagnostic);
}

// TODO(dneto): This probably belongs in text.cpp since that's the only place
// that a spv_binary_t value is created.
void spvBinaryDestroy(spv_binary binary) {
  if (!binary) return;
  if (binary->code) {
    delete[] binary->code;
  }
  delete binary;
}