// Copyright 2018 The Fuchsia Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#include "src/developer/debug/zxdb/expr/expr_parser.h"

#include <lib/syslog/cpp/macros.h>

#include <algorithm>
#include <iterator>

#include "src/developer/debug/zxdb/expr/expr_tokenizer.h"
#include "src/developer/debug/zxdb/expr/find_name.h"
#include "src/developer/debug/zxdb/expr/found_name.h"
#include "src/developer/debug/zxdb/expr/number_parser.h"
#include "src/developer/debug/zxdb/expr/operator_keyword.h"
#include "src/developer/debug/zxdb/expr/parse_special_identifier.h"
#include "src/developer/debug/zxdb/expr/template_type_extractor.h"
#include "src/developer/debug/zxdb/expr/variable_decl.h"
#include "src/developer/debug/zxdb/symbols/array_type.h"
#include "src/developer/debug/zxdb/symbols/collection.h"
#include "src/developer/debug/zxdb/symbols/modified_type.h"
#include "src/developer/debug/zxdb/symbols/symbol_utils.h"
#include "src/lib/fxl/strings/string_printf.h"

// The parser is a Pratt parser. The basic idea there is to have the precedences (and
// associativities) encoded relative to each other and only parse up until you hit something of that
// precedence. There's a dispatch table in kDispatchInfo that describes how each token dispatches if
// it's seen as either a prefix or infix operator, and if it's infix, what its precedence is.
//
// References:
// http://javascript.crockford.com/tdop/tdop.html
// http://journal.stuffwithstuff.com/2011/03/19/pratt-parsers-expression-parsing-made-easy/

namespace zxdb {

namespace {

// An infix operator is one that combines two sides of things and it modifies both, like "a + b"
// ("a" is the "left" and "+" is the token in the params).
//
// Other things are infix like "[" which combines the expression on the left with some expression to
// the right of it.
//
// A prefix operator are binary operators like "!" in C that only apply to the thing on the right
// and don't require anything on the left. Standalone numbers and names are also considered prefix
// since they represent themselves (not requiring anything on the left).
//
// Some things can be both prefix and infix. An example in C is "(" which is prefix when used in
// casts and math expressions: "(a + b)" "a + (b + c)" but infix when used for function calls:
// "foo(bar)".
using PrefixFunc = fxl::RefPtr<ExprNode> (ExprParser::*)(const ExprToken&);
using InfixFunc = fxl::RefPtr<ExprNode> (ExprParser::*)(fxl::RefPtr<ExprNode> left,
                                                        const ExprToken& token);

// Precedence constants used in DispatchInfo. Note that these aren't contiguous. At least need to do
// every-other-one to handle the possible "precedence - 1" that occurs when evaluating
// right-associative operators. We don't want that operation to push the precedence into a
// completely other category, rather, it should only affect comparisons that would otherwise be
// equal.
//
// This should match the C operator precedence for the subset of operations that we support:
//   https://en.cppreference.com/w/cpp/language/operator_precedence
// The commented-out values are ones we don't currently implement.

// clang-format off
constexpr int kPrecedenceComma = 10;                  // ,  (lowest precedence)
constexpr int kPrecedenceAssignment = 20;             // = += -= *= -= /= %= <<= >>= &= ^= |=
constexpr int kPrecedenceLogicalOr = 30;              // ||
constexpr int kPrecedenceLogicalAnd = 40;             // &&
constexpr int kPrecedenceBitwiseOr = 50;              // |
constexpr int kPrecedenceBitwiseXor = 60;             // ^
constexpr int kPrecedenceBitwiseAnd = 70;             // &
constexpr int kPrecedenceEquality = 80;               // == !=
constexpr int kPrecedenceComparison = 90;             // < <= > >=
constexpr int kPrecedenceThreeWayComparison = 100;    // <=>
constexpr int kPrecedenceShift = 110;                 // << >>
constexpr int kPrecedenceAddition = 120;              // + -
constexpr int kPrecedenceMultiplication = 130;        // * / %
//constexpr int kPrecedencePointerToMember = 140;     // .* ->*
constexpr int kPrecedenceRustCast = 140;              // foo as Bar
constexpr int kPrecedenceUnary = 150;                 // ++ -- +a -a ! ~ *a &a
constexpr int kPrecedenceCallAccess = 160;            // () . -> []
//constexpr int kPrecedenceScope = 170;               // ::  (Highest precedence)
//  clang-format on

// When there is a block or block-like construct, any local variables created inside it must be
// freed when that scope exits.
//
// This object is created on the stack before parsing such a block, and at the bottom the caller
// calls GetExitVarCount() to tell what should have happen. If no local variables were created, it
// will return nullopt, but if the local variable stack needs to be popped, it will return the
// number of items that should be remaining. See vm_op.h for more on local variables.
class LocalVarCounter {
 public:
  explicit LocalVarCounter(std::vector<std::string>* local_vars)
      : local_vars_(local_vars),
        init_var_count_(local_vars->size()) {}

  std::optional<uint32_t> GetExitVarCount() const {
    // The scope should not have reduced the local variable count from when it entered.
    FX_DCHECK(init_var_count_ <= local_vars_->size());

    if (local_vars_->size() > init_var_count_) {
      local_vars_->resize(init_var_count_);           // Trim the local variable record.
      return static_cast<uint32_t>(init_var_count_);  // Tell the caller what to do in bytecode.
    }
    return std::nullopt;  // No local vars were created, do nothing.
  }

 private:
  std::vector<std::string>* local_vars_;
  size_t init_var_count_;
};

// Sets a boolean for the duration of the lifetime of the object. Puts it back to the old value when
// it goes out of scope.
class ScopedBoolSetter {
 public:
  ScopedBoolSetter(bool* var, bool value) : var_(var), original_value_(*var_) {
    *var = value;
  }
  ~ScopedBoolSetter() {
    *var_ = original_value_;
  }

 private:
  bool* var_;
  bool original_value_;
};

}  // namespace

struct ExprParser::DispatchInfo {
  PrefixFunc prefix = nullptr;
  InfixFunc infix = nullptr;
  int precedence = 0;  // Only needed when |infix| is set.
};

// The table is more clear without line wrapping.
// clang-format off
ExprParser::DispatchInfo ExprParser::kDispatchInfo[] = {
    // Prefix handler                Infix handler                 Precedence for infix
    {nullptr,                        nullptr,                      -1},                             // kInvalid
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kName
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kSpecialName
    {nullptr,                        nullptr,                      -1},                             // kComment (will be stripped).
    {&ExprParser::LiteralPrefix,     nullptr,                      -1},                             // kInteger
    {&ExprParser::LiteralPrefix,     nullptr,                      -1},                             // kFloat
    {&ExprParser::LiteralPrefix,     nullptr,                      -1},                             // kStringLiteral
    {&ExprParser::LiteralPrefix,     nullptr,                      -1},                             // kCharLiteral
    {nullptr,                        nullptr,                      -1},                             // kRustLifetime
    {&ExprParser::BadToken,          nullptr,                      -1},                             // kCommentBlockEnd
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceEquality},            // kEquality
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceEquality},            // kInequality
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceComparison},          // kLessEqual
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceComparison},          // kGreaterEqual
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceThreeWayComparison},  // kSpaceship
    {nullptr,                        &ExprParser::DotOrArrowInfix, kPrecedenceCallAccess},          // kDot
    {nullptr,                        nullptr,                      -1},                             // kDotStar (unsupported)
    {nullptr,                        nullptr,                      -1},                             // kComma
    {nullptr,                        nullptr,                      -1},                             // kSemicolon
    {&ExprParser::StarPrefix,        &ExprParser::BinaryOpInfix,   kPrecedenceMultiplication},      // kStar
    {&ExprParser::AmpersandPrefix,   &ExprParser::BinaryOpInfix,   kPrecedenceBitwiseAnd},          // kAmpersand
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceLogicalAnd},          // kDoubleAnd
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceBitwiseOr},           // kBitwiseOr
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceLogicalOr},           // kLogicalOr
    {nullptr,                        &ExprParser::DotOrArrowInfix, kPrecedenceCallAccess},          // kArrow
    {nullptr,                        nullptr,                      -1},                             // kArrowStar (unsupported)
    {nullptr,                        &ExprParser::LeftSquareInfix, kPrecedenceCallAccess},          // kLeftSquare
    {nullptr,                        nullptr,                      -1},                             // kRightSquare
    {&ExprParser::LeftParenPrefix,   &ExprParser::LeftParenInfix,  kPrecedenceCallAccess},          // kLeftParen
    {nullptr,                        nullptr,                      -1},                             // kRightParen
    {&ExprParser::LeftBracketPrefix, nullptr,                      -1},                             // kLeftBracket
    {nullptr,                        nullptr,                      -1},                             // kRightBracket
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceComparison},          // kLess
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceComparison},          // kGreater
    {&ExprParser::UnaryPrefix,       &ExprParser::BinaryOpInfix,   kPrecedenceAddition},            // kMinus
    {nullptr,                        nullptr,                      -1},                             // kMinusMinus (unsupported)
    {&ExprParser::UnaryPrefix,       nullptr,                      -1},                             // kBang
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAddition},            // kPlus
    {nullptr,                        nullptr,                      -1},                             // kPlusPlus (unsupported)
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceMultiplication},      // kSlash
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceUnary},               // kAt
    {nullptr,                        nullptr,   -1},                                                // kOctothorpe
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceBitwiseXor},          // kCaret
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceMultiplication},      // kPercent
    {nullptr,                        &ExprParser::QuestionInfix,   kPrecedenceAssignment},          // kQuestion
    {&ExprParser::UnaryPrefix,       nullptr,                      -1},                             // kTilde
    {nullptr,                        nullptr,                      -1},                             // kColon
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kColonColon
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kPlusEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kMinusEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kStarEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kSlashEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kPercentEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kCaretEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kAndEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kOrEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceShift},               // kShiftLeft
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kShiftLeftEquals
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceShift},               // kShiftRight
    {nullptr,                        &ExprParser::BinaryOpInfix,   kPrecedenceAssignment},          // kShiftRightEquals
    {&ExprParser::LiteralPrefix,     nullptr,                      -1},                             // kTrue
    {&ExprParser::LiteralPrefix,     nullptr,                      -1},                             // kFalse
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kConst
    {nullptr,                        nullptr,                      -1},                             // kMut
    {&ExprParser::LetPrefix,         nullptr,                      -1},                             // kLet
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kVolatile
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kRestrict
    {&ExprParser::CastPrefix,        nullptr,                      -1},                             // kReinterpretCast
    {&ExprParser::CastPrefix,        nullptr,                      -1},                             // kStaticCast
    {&ExprParser::SizeofPrefix,      nullptr,                      -1},                             // kSizeof
    {nullptr,                        &ExprParser::RustCastInfix,   kPrecedenceRustCast},            // kAs
    {&ExprParser::IfPrefix,          nullptr,                      -1},                             // kIf
    {nullptr,                        nullptr,                      -1},                             // kElse
    {&ExprParser::ForPrefix,         nullptr,                      -1},                             // kFor
    {&ExprParser::DoPrefix,          nullptr,                      -1},                             // kDo
    {&ExprParser::WhilePrefix,       nullptr,                      -1},                             // kWhile
    {&ExprParser::LoopPrefix,        nullptr,                      -1},                             // kLoop
    {&ExprParser::BreakPrefix,       nullptr,                      -1},                             // kBreak
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kOperator
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kNew
    {&ExprParser::NamePrefix,        nullptr,                      -1},                             // kDelete
};
// clang-format on

// static
const ExprToken ExprParser::kInvalidToken;

ExprParser::ExprParser(std::vector<ExprToken> tokens, ExprLanguage lang,
                       fxl::RefPtr<EvalContext> eval_context)
    : language_(lang), eval_context_(std::move(eval_context)), tokens_(std::move(tokens)) {
  static_assert(std::size(ExprParser::kDispatchInfo) == static_cast<int>(ExprTokenType::kNumTypes),
                "kDispatchInfo needs updating to match ExprTokenType");

  // Strip comments from the input. This is easier than handling them at read-time and we don't
  // need the flexibility of that information.
  tokens_.erase(
      std::remove_if(tokens_.begin(), tokens_.end(),
                     [](const ExprToken& t) { return t.type() == ExprTokenType::kComment; }),
      tokens_.end());
}

fxl::RefPtr<ExprNode> ExprParser::ParseStandaloneExpression() {
  auto result = ParseExpression(0, EmptyExpression::kReject, StatementEnd::kAny);

  // That should have consumed everything, as we don't support multiple expressions being next to
  // each other (probably the user forgot an operator and wrote something like "foo 5"
  if (!has_error() && !at_end()) {
    SetError(cur_token(), "Unexpected input, did you forget an operator or a semicolon?");
    return nullptr;
  }
  return result;
}

fxl::RefPtr<BlockExprNode> ExprParser::ParseBlock(BlockDelimiter delimiter) {
  if (delimiter == BlockDelimiter::kExplicit) {
    // Expect block to start with a {.
    Consume(ExprTokenType::kLeftBracket, "Expected '{'.");
    if (has_error())
      return nullptr;
  }

  auto block = ParseBlockContents(delimiter == BlockDelimiter::kExplicit ? BlockEnd::kExplicit
                                                                         : BlockEnd::kEndOfInput);
  if (!has_error() && delimiter == BlockDelimiter::kExplicit) {
    // Consume the ending }.
    Consume(ExprTokenType::kRightBracket, "Expected '}'.");
  }
  if (has_error())
    return nullptr;
  return block;
}

// static
Err ExprParser::ParseIdentifier(const std::string& input, Identifier* output) {
  ParsedIdentifier parsed;
  Err err = ParseIdentifier(input, &parsed);
  if (err.has_error())
    return err;
  *output = ToIdentifier(parsed);
  return Err();
}

// static
Err ExprParser::ParseIdentifier(const std::string& input, ParsedIdentifier* output) {
  ExprTokenizer tokenizer(input);
  if (!tokenizer.Tokenize())
    return tokenizer.err();

  ExprParser parser(tokenizer.TakeTokens(), tokenizer.language());
  auto root = parser.ParseStandaloneExpression();
  if (!root)
    return parser.err();

  auto identifier_node = root->AsIdentifier();
  if (!identifier_node)
    return Err("Input did not parse as an identifier.");

  *output = const_cast<IdentifierExprNode*>(identifier_node)->TakeIdentifier();
  return Err();
}

fxl::RefPtr<ExprNode> ExprParser::ParseExpression(int precedence, EmptyExpression empty_expr,
                                                  StatementEnd statement_end,
                                                  bool* had_statement_end) {
  if (had_statement_end)
    *had_statement_end = false;
  if (at_end()) {
    if (empty_expr == EmptyExpression::kReject || statement_end == StatementEnd::kExplicit)
      SetError(ExprToken(), "Expected expression instead of end of input.");
    return nullptr;
  }

  // Here we need ">>" to be treated as a shift operation. When that appears as part of a name and
  // should be separated, the identifier parser will handle it separately.
  ExprToken token = ConsumeWithShiftTokenConversion();
  if (token.type() == ExprTokenType::kSemicolon) {
    if (had_statement_end)
      *had_statement_end = true;  // record that this empty expression had an explicit end.
    if (empty_expr == EmptyExpression::kReject)
      SetError(token, "Empty expression not permitted here.");
    return nullptr;
  }

  PrefixFunc prefix = DispatchForToken(token).prefix;
  if (!prefix) {
    SetError(token, fxl::StringPrintf("Unexpected token '%s'.", token.value().c_str()));
    return nullptr;
  }

  fxl::RefPtr<ExprNode> left = (this->*prefix)(token);
  if (has_error())
    return left;  // In both C++ and Rust, the end of a block is the end of a statement.

  // Checks for the next thing being a semicolon or otherwise the end of a statement.
  // end of the expression.
  while (!at_end() && !LookAhead(ExprTokenType::kSemicolon) && !CountsAsStatementEnd(left.get()) &&
         precedence < CurPrecedenceWithShiftTokenConversion()) {
    const ExprToken& next_token = ConsumeWithShiftTokenConversion();
    InfixFunc infix = DispatchForToken(next_token).infix;
    if (!infix) {
      SetError(next_token, fxl::StringPrintf("Unexpected token '%s'.", next_token.value().c_str()));
      return nullptr;
    }
    left = (this->*infix)(std::move(left), next_token);
    if (has_error())
      return nullptr;
  }

  // Handle end-of-statement requirements.
  switch (statement_end) {
    case StatementEnd::kAny:
    case StatementEnd::kExplicit:
      if (CountsAsStatementEnd(left.get())) {
        if (had_statement_end)
          *had_statement_end = true;
        break;
      } else if (LookAhead(ExprTokenType::kSemicolon)) {
        // Semicolon indicating end of statement, consume it.
        Consume();
        if (had_statement_end)
          *had_statement_end = true;
        break;
      } else if (statement_end == StatementEnd::kExplicit) {
        // Statement end was required but not present.
        SetErrorAtCur("Expected ';'.");
        return nullptr;
      }
      break;
    case StatementEnd::kNone:
      // Nothing to check.
      break;
  }

  return left;
}

// We parse all blocks as expressions. In C++ this is not correct but simplifies the parsing logic.
// During evaluation, C++ blocks will return empty which will throw errors when using blocks in
// places where they can't be used in expressions.
fxl::RefPtr<BlockExprNode> ExprParser::ParseBlockContents(BlockEnd block_end) {
  LocalVarCounter var_counter(&local_vars_);

  // Expect a sequence of semicolon-terminated expressions. In Rust the last expression does not
  // need a semicolon. As we start parsing each new statement, we check whether the previous
  // statement had a separator.
  bool previous_had_statement_end = true;

  constexpr ExprTokenType kBlockEndTokenType = ExprTokenType::kRightBracket;

  std::vector<fxl::RefPtr<ExprNode>> statements;
  while (!at_end() && !LookAhead(kBlockEndTokenType)) {
    if (!previous_had_statement_end) {
      SetError(cur_token(), "Unexpected token, did you forget an operator or a semicolon?");
      return nullptr;
    }

    // Accept any type of statement end and validate later to accommodate Rust.
    auto stmt = ParseExpression(0, EmptyExpression::kAllow, StatementEnd::kAny,
                                &previous_had_statement_end);
    if (has_error())
      return nullptr;
    if (stmt)  // Statement could be empty.
      statements.push_back(std::move(stmt));
  }

  // Don't check the final statement in a block for a statement end. In Rust, this is not necessary
  // because it indicates that the statement is the result of the block (although we do the same
  // thing with the semicolon). In C we allow the same to allow all expressions to be parsed using
  // the block code path. This allows cases like:
  //   print expression1               // No semicolon, but there could have been more statements.
  //   print expression1; expression2  // Semicolon separated statements last ';' is optional.
  // If we wanted to require semicolons, we would check previous_had_statement_end here.

  // Validate the correct end of the block.
  switch (block_end) {
    case BlockEnd::kExplicit:
      if (!LookAhead(kBlockEndTokenType))
        SetErrorAtCur("Expected '}'.");
      break;
    case BlockEnd::kEndOfInput:
      if (!at_end())
        SetErrorAtCur("Unexpected.");
      break;
  }
  if (has_error())
    return nullptr;

  return fxl::MakeRefCounted<BlockExprNode>(std::move(statements), var_counter.GetExitVarCount());
}

ExprParser::ParseNameResult ExprParser::ParseName(bool expand_types) {
  // Grammar we support. Note "identifier" in this context is a single token of type "name" (more
  // like how the C++ spec uses it), while our Identifier class represents a whole name with scopes
  // and templates.
  //
  //   name := type-name | special-identifier | other-identifier
  //
  //   type-name := [ scope-name "::" ] identifier [ "<" template-list ">" ]
  //
  //   special-identifier := "$" [ special-name ] [ "(" special-contents ")" ]
  //
  //   other-identifier := [ <scope-name> "::" ] <identifier>
  //
  //   scope-name := ( namespace-name | type-name )
  //
  // The thing that differentiates type names, namespace names, and other identifiers is the symbol
  // lookup function rather than something in the grammar.
  //
  // The thing this doesn't handle is templatized functions, for example:
  //   auto foo = &MyClass::MyFunc<int>;
  // To handle this we will need the type lookup function to be able to tell us "MyClass::MyFunc" is
  // a thing that has a template so we know to parse the following "<" as part of the name and not
  // as a comparison. Note that when we need to parse function names, there is special handling
  // required for operators.
  //
  // As a special-case, this can also get called with a ~ operator when that begins a name. It gets
  // joined with the following word for a C++ destructor name.

  // The mode of the state machine.
  enum Mode {
    kBegin,       // Initial state with no previous context.
    kColonColon,  // Just saw a "::", expecting a name next.
    kType,        // Identifier is a type.
    kTemplate,    // Identifier is a template, expecting "<" next.
    kNamespace,   // Identifier is a namespace.
    kOtherName,   // Identifier is something other than the above (normally this means a variable).
    kAnything     // Caller can't do symbol lookups, accept anything that makes sense.
  };

  Mode mode = kBegin;
  ParseNameResult result;
  const ExprToken* prev_token = nullptr;

  while (!at_end()) {
    const ExprToken& token = cur_token();
    switch (token.type()) {
      case ExprTokenType::kColonColon: {
        // "::" can only follow nothing, a namespace or type name.
        if (mode != kBegin && mode != kNamespace && mode != kType && mode != kAnything) {
          SetError(token,
                   "Could not identify thing to the left of '::' as a type or "
                   "namespace.");
          return ParseNameResult();
        }

        mode = kColonColon;
        if (result.ident.empty())  // Globally qualified (starts with "::").
          result.ident = ParsedIdentifier(IdentifierQualification::kGlobal);
        result.type = nullptr;  // No longer a type.
        break;
      }

      case ExprTokenType::kLess: {
        // "<" can only come after a template name.
        if (mode == kNamespace || mode == kType) {
          // Generate a nicer error for these cases.
          SetError(token, "Template parameters not valid on this object type.");
          return ParseNameResult();
        }
        if (mode != kTemplate && mode != kAnything) {
          // "<" after anything but a template means the end of the name. In "anything" mode we
          // assume "<" means a template since this is used to parse random identifiers and function
          // names.
          return result;
        }
        if (result.ident.components().back().has_template()) {
          // Got a "<" after a template parameter list was already defined (this will happen in
          // "anything" mode since we don't know what it is for sure). That means this is a
          // comparison operator which will be handled by the outer parser.
          return result;
        }

        prev_token = &Consume();  // Eat the "<".

        // Extract the contents of the template.
        std::vector<std::string> list = ParseTemplateList(ExprTokenType::kGreater);
        if (has_error())
          return ParseNameResult();

        // Ending ">".
        Consume(ExprTokenType::kGreater, "Expected '>' to match.");
        if (has_error())
          return ParseNameResult();

        // Construct a replacement for the last component of the identifier with the template
        // arguments added.
        ParsedIdentifierComponent& back = result.ident.components().back();
        back = ParsedIdentifierComponent(back.name(), std::move(list));

        // The thing we just made is either a type or a name, look it up.
        if (eval_context_) {
          FoundName lookup =
              eval_context_->FindName(FindNameOptions(FindNameOptions::kAllKinds), result.ident);
          switch (lookup.kind()) {
            case FoundName::kType:
              mode = kType;
              result.type = std::move(lookup.type());
              break;
            case FoundName::kNamespace:
            case FoundName::kTemplate:
              // The lookup shouldn't tell us a template name or namespace for something that has
              // template parameters.
              FX_NOTREACHED();
              // Fall through to "other" case for fallback.
              [[fallthrough]];
            case FoundName::kVariable:
            case FoundName::kMemberVariable:
            case FoundName::kFunction:
            case FoundName::kOtherSymbol:
            case FoundName::kNone:
              mode = kOtherName;
              break;
          }
        } else {
          mode = kAnything;
        }
        continue;  // Don't Consume() since we already ate the token.
      }

      case ExprTokenType::kOperator: {
        if (OperatorKeywordResult op_result = ParseOperatorKeyword(tokens_, cur_);
            op_result.success) {
          // Standard C++ "operator+" or similar opeator.
          result.ident.AppendComponent(ParsedIdentifierComponent(op_result.canonical_name));
          mode = kOtherName;
          cur_ = op_result.end_token - 1;  // Point to last token in the operator name.
          break;
        }

        // Here the operator is either invalid or it's a type conversion like
        // "MyClass::operator bool()". We don't currently differentiate between C and C++ so use
        // some heuristics to determine if the user is trying to type a conversion operator function
        // name in C++ or a variable or class member name called "operator" in C. This heuristic
        // expects anything with a "::" before the operator definition.
        //
        // TODO: differentiate between C and C++ in the parser. Then we can remove this condition.
        // C++ would always expect a type name, and C wouldn't have even generated the special
        // "operator" token (it would be a normal name).
        if (mode != kBegin) {
          Consume();  // Eat the "operator" keyword.
          if (auto conv_type = ParseType(nullptr)) {
            // Found a type conversion operator.
            result.ident.AppendComponent(
                ParsedIdentifierComponent("operator " + conv_type->GetFullName()));
            mode = kOtherName;
            break;
          } else {
            // Can't parse as a type. We can't fall through to the name case below since we already
            // advanced arbitrarily forward in the parsing. The error should be already be set.
            return ParseNameResult();
          }
        }

        // Fall-through to regular name parsing to treat "operator" as a normal name as per C.
        [[fallthrough]];
      }

      case ExprTokenType::kName:
      case ExprTokenType::kSpecialName: {
        // Names can only follow nothing or "::".
        if (mode == kType) {
          // This would probably be the case to add support for two-word type names like
          // "unsigned int". For now, consider the type name done.
          return result;
        } else if (mode == kBegin) {
          // Found an identifier name with nothing before it.
          result.ident = ParsedIdentifier(GetIdentifierComponent());
        } else if (mode == kColonColon) {
          result.ident.AppendComponent(GetIdentifierComponent());
        } else {
          // Anything else like "std::vector foo" or "foo bar".
          SetError(token, "Unexpected identifier, did you forget an operator or a semicolon?");
          return ParseNameResult();
        }

        // Decode what adding the name just generated.
        if (std::optional<uint32_t> found_local = GetLocalVariable(result.ident)) {
          // This is a known local variable.
          result.local_var_slot = found_local;
          Consume();  // Eat the name we just saw.
          return result;
        } else if (eval_context_) {
          FoundName lookup =
              eval_context_->FindName(FindNameOptions(FindNameOptions::kAllKinds), result.ident);
          switch (lookup.kind()) {
            case FoundName::kNamespace:
              mode = kNamespace;
              break;
            case FoundName::kTemplate:
              mode = kTemplate;
              break;
            case FoundName::kType:
              mode = kType;
              FX_DCHECK(lookup.type());  // Should be set for "kType" results.

              // Expand typedefs and remove const/volatile. This is needed to expand things like:
              //    static_cast<Foo::Bar>(...)
              // where "Foo" is a typedef for something else. This implementation throws away the
              // user input name and replaces it with the typedef-ed name.
              result.type = eval_context_->GetConcreteType(lookup.type().get());
              result.ident = ToParsedIdentifier(result.type->GetIdentifier());
              result.type = std::move(lookup.type());
              break;
            case FoundName::kVariable:
            case FoundName::kMemberVariable:
            case FoundName::kFunction:
            case FoundName::kOtherSymbol:
            case FoundName::kNone:
              mode = kOtherName;
              break;
          }
        } else {
          mode = kAnything;
        }
        break;
      }

      default: {
        // Any other token type means we're done. The outer parser will figure out what it means.
        if (expand_types && result.type && language_ != ExprLanguage::kRust) {
          // When we found a type, add on any trailing modifiers like "*". Rust doesn't have
          // trailing modifiers, so skip this for Rust.
          result.type = ParseType(std::move(result.type));
        }
        return result;
      }
    }
    prev_token = &Consume();
  }

  // Hit end-of-input.
  switch (mode) {
    case kOtherName:
    case kAnything:
    case kType:
      return result;  // Success cases.
    case kBegin:
      FX_NOTREACHED();
      SetError(ExprToken(), "Unexpected end of input.");
      return ParseNameResult();
    case kColonColon:
      SetError(*prev_token, "Expected name after '::'.");
      return ParseNameResult();
    case kTemplate:
      SetError(*prev_token, "Expected template args after template name.");
      return ParseNameResult();
    case kNamespace:
      SetError(*prev_token, "Expected expression after namespace name.");
      return ParseNameResult();
  }

  FX_NOTREACHED();
  SetError(ExprToken(), "Internal error.");
  return ParseNameResult();
}

ParsedIdentifierComponent ExprParser::GetIdentifierComponent() {
  const ExprToken& token = cur_token();
  FX_DCHECK(!token.value().empty());  // Should not have an empty name token.

  if (token.value()[0] == '$') {
    // Special identifier, need to parse it.
    size_t cur = 0;             // Special name is at the beginning of the token.
    size_t error_location = 0;  // Don't need this.
    SpecialIdentifier si = SpecialIdentifier::kNone;
    std::string special_cont;
    if (Err err = ParseSpecialIdentifier(token.value(), &cur, &si, &special_cont, &error_location);
        err.has_error()) {
      SetError(token, err);
      return ParsedIdentifierComponent();
    }

    return ParsedIdentifierComponent(si, special_cont);
  }

  // Regular identifier, can just return a component for the literal value.
  return ParsedIdentifierComponent(token.value());
}

fxl::RefPtr<Type> ExprParser::ParseType(fxl::RefPtr<Type> optional_base) {
  // For C++, the thing we want to parse is:
  //
  //   cv-qualifier := [ "const" ] [ "volatile" ] [ "restrict" ]
  //
  //   ptr-operator := ( "*" | "&" | "&&" ) cv-qualifier
  //
  //   type-id := cv-qualifier type-name cv-qualifier [ ptr-operator ] *
  //
  // Our logic is much more permissive than C++. This is both because it makes the code simpler, and
  // because certain constructs may be used by other languages. For example, this allows references
  // to references and "int & const" while C++ says you can't apply const to the reference itself
  // (it permits only "const int&" or "int const &" which are the same). It also allows "restrict"
  // to be used in invalid places.
  //
  // For Rust, the thing we want to parse is:
  //
  //   qualifier := "&" | "*" | "mut"
  //
  //   type-id = qualifier* (type-name | array-type | tuple-type)
  //
  //   array-type = "[" type ";" integer "]"
  //
  //   tuple-type = "(" type ("," type)* ")"
  //
  // This is simpler than the C++ case. Rust has some restrictions on how qualifiers appear, and
  // again we are more permissive, mostly because Rust's mutability constraints aren't important
  // to the debugger case, so we can afford to be less particular. References and pointers are all
  // the same in the compiled code, so we interchange them freely.

  fxl::RefPtr<Type> type;
  std::vector<DwarfTag> type_qual;
  size_t pointer_levels = 0;
  if (optional_base) {
    FX_DCHECK(language_ != ExprLanguage::kRust);
    // Type name already known, start parsing after it.
    type = std::move(optional_base);
  } else {
    if (language_ == ExprLanguage::kRust) {
      while (!at_end()) {
        const ExprToken& token = cur_token();
        if (token.type() == ExprTokenType::kStar) {
          pointer_levels++;
        } else if (token.type() == ExprTokenType::kAmpersand) {
          pointer_levels++;
        } else if (token.type() == ExprTokenType::kDoubleAnd) {
          pointer_levels += 2;
        } else if (token.type() == ExprTokenType::kMut) {
          // Ignore mut.
        } else {
          // Done with the ptr-operators.
          break;
        }
        Consume();  // Eat the operator token.
      }
    } else {
      // Read "const", etc. that comes before the type name.
      ConsumeCVQualifier(&type_qual);
      if (has_error())
        return nullptr;
    }

    // Read the type name itself.
    if (at_end()) {
      SetError(ExprToken(), "Expected type name before end of input.");
      return nullptr;
    }
    const ExprToken& first_name_token = cur_token();  // For error blame below.

    if (language_ == ExprLanguage::kRust &&
        (first_name_token.type() == ExprTokenType::kLeftSquare ||
         first_name_token.type() == ExprTokenType::kLeftParen)) {
      if (first_name_token.type() == ExprTokenType::kLeftParen) {
        Consume();

        std::vector<fxl::RefPtr<Type>> members;

        bool expect_type = true;
        while (!at_end() && cur_token().type() != ExprTokenType::kRightParen && expect_type) {
          auto next = ParseType(nullptr);
          if (!next) {
            return nullptr;
          }

          members.push_back(std::move(next));

          expect_type = !at_end() && cur_token().type() == ExprTokenType::kComma;
          if (expect_type) {
            Consume();
          }
        }

        if (at_end()) {
          SetError(ExprToken(), "Expected ')' or ',' before end of input.");
          return nullptr;
        } else if (cur_token().type() != ExprTokenType::kRightParen) {
          SetError(cur_token(), "Expected ')' or ','.");
          return nullptr;
        } else {
          Consume();
        }

        if (members.size() == 1 && !expect_type) {
          // This was just a type in parentheses. Rust appears to handle this in this way so we will
          // too.
          return std::move(members[0]);
        }

        std::string name = "(";

        for (const auto& type : members) {
          if (name.size() > 1) {
            name += ", ";
          }

          name += type->GetAssignedName();
        }

        if (members.size() == 1) {
          name += ",";
        }

        name += ")";

        return MakeRustTuple(name, members);
      } else {
        type = ParseRustArrayType();

        if (!type) {
          return nullptr;
        }
      }
    } else {
      ParseNameResult parse_result = ParseName(false);
      if (has_error())
        return nullptr;
      if (!parse_result.type) {
        SetError(first_name_token,
                 fxl::StringPrintf("Expected a type name but could not find a type named '%s'.",
                                   parse_result.ident.GetFullName().c_str()));
        return nullptr;
      }
      type = std::move(parse_result.type);
    }

    for (size_t i = 0; i < pointer_levels; i++) {
      type = fxl::MakeRefCounted<ModifiedType>(DwarfTag::kPointerType, std::move(type));
    }
  }

  if (language_ == ExprLanguage::kRust) {
    return type;
  }

  // Read "const" etc. that comes after the type name. These apply the same as the ones that come
  // before it so get appended and can't duplicate them.
  ConsumeCVQualifier(&type_qual);
  if (has_error())
    return nullptr;
  type = ApplyQualifiers(std::move(type), type_qual);

  // Parse the ptr-operators that can be present after the type.
  while (!at_end()) {
    // Read the operator.
    const ExprToken& token = cur_token();
    if (token.type() == ExprTokenType::kStar) {
      type = fxl::MakeRefCounted<ModifiedType>(DwarfTag::kPointerType, std::move(type));
    } else if (token.type() == ExprTokenType::kAmpersand) {
      type = fxl::MakeRefCounted<ModifiedType>(DwarfTag::kReferenceType, std::move(type));
    } else if (token.type() == ExprTokenType::kDoubleAnd) {
      type = fxl::MakeRefCounted<ModifiedType>(DwarfTag::kRvalueReferenceType, std::move(type));
    } else {
      // Done with the ptr-operators.
      break;
    }
    Consume();  // Eat the operator token.

    // Apply any const-volatile-restrict to the operator.
    std::vector<DwarfTag> qual;
    ConsumeCVQualifier(&qual);
    if (has_error())
      return nullptr;
    type = ApplyQualifiers(std::move(type), qual);
  }

  return type;
}

fxl::RefPtr<ExprNode> ExprParser::ParseAfterType(fxl::RefPtr<TypeExprNode> type) {
  if (language_ == ExprLanguage::kRust)
    return type;  // In Rust types don't start anything.

  // In C/C++ a type name could mean different things.
  //
  // A variable declaration:
  //    int i;
  //    int some_value = 2;
  //    double other_value = some_expression * 7;
  //    char char_value(some_expression - 1);
  //
  // We don't support multiple variable declarations like "int i, j;"
  //
  // We don't currently support functional-style casts like:
  //    int(some_value)
  //    double(23)
  // but this would be the place to add such support in the future.
  if (LookAhead(ExprTokenType::kLeftParen)) {
    // Functional-style cast.
    SetError(Consume(), "Functional-style casts are not currently supported.");
    return nullptr;
  } else if (LookAhead(ExprTokenType::kName)) {
    // Variable declaration.
    const ExprToken& name_token = Consume();
    fxl::RefPtr<ExprNode> init_expr;
    if (LookAhead(ExprTokenType::kLeftParen)) {
      // Paren-style initialization (we don't support function declarations so don't need to check
      // for that here).
      auto left_paren = Consume();

      init_expr = ParseExpression(0);
      if (has_error())
        return nullptr;

      // Closing paren.
      Consume(ExprTokenType::kRightParen, "Expected ')' to match.", left_paren);
      if (has_error())
        return nullptr;
    } else if (LookAhead(ExprTokenType::kEquals)) {
      // "=" initialization.
      Consume();
      init_expr = ParseExpression(0);
      if (has_error())
        return nullptr;
    }
    // If we wanted to support {}-style initization "int a{0};" that would go here.

    // Decode the type information.
    auto decl_or = GetVariableDeclTypeInfo(language_, eval_context_->GetConcreteType(type->type()));
    if (decl_or.has_error()) {
      SetError(name_token, decl_or.err());
      return nullptr;
    }

    uint32_t local_slot = static_cast<uint32_t>(local_vars_.size());
    local_vars_.push_back(name_token.value());

    return fxl::MakeRefCounted<VariableDeclExprNode>(decl_or.take_value(), local_slot, name_token,
                                                     std::move(init_expr));
  }

  // Anything else is nothing special. Return the type token because it could be an argument to
  // sizeof() or something.
  return type;
}

fxl::RefPtr<Type> ExprParser::ParseRustArrayType() {
  FX_DCHECK(!at_end() && cur_token().type() == ExprTokenType::kLeftSquare);
  Consume();

  auto element = ParseType(nullptr);
  if (!element) {
    return nullptr;
  }

  std::optional<size_t> count = std::nullopt;

  if (!at_end() && cur_token().type() == ExprTokenType::kSemicolon) {
    Consume();

    if (at_end()) {
      SetError(ExprToken(), "Expected element count before end of input.");
      return nullptr;
    }

    if (cur_token().type() != ExprTokenType::kInteger) {
      SetError(cur_token(), "Expected element count.");
      return nullptr;
    }

    auto got = StringToNumber(language_, cur_token().value());
    if (got.has_error()) {
      // Should this even be possible?
      SetError(cur_token(), "Could not parse integer: " + got.err().msg());
      return nullptr;
    }

    int64_t value;

    auto promote_err = got.value().PromoteTo64(&value);
    if (promote_err.has_error()) {
      SetError(cur_token(), "Invalid array size: " + promote_err.msg());
      return nullptr;
    }

    if (value < 0) {
      SetError(cur_token(), "Negative array size.");
      return nullptr;
    }

    Consume();

    count = static_cast<size_t>(value);
  }

  if (at_end()) {
    SetError(ExprToken(), "Expected ']' before end of input.");
    return nullptr;
  }

  if (cur_token().type() != ExprTokenType::kRightSquare) {
    SetError(cur_token(), "Expected ']'.");
    return nullptr;
  }

  Consume();

  return fxl::MakeRefCounted<ArrayType>(std::move(element), std::move(count));
}

// A list is any sequence of comma-separated types. We don't parse the types (this is hard) but
// instead skip over them.
std::vector<std::string> ExprParser::ParseTemplateList(ExprTokenType stop_before) {
  std::vector<std::string> result;

  bool first_time = true;
  while (!at_end() && !LookAhead(stop_before)) {
    if (first_time) {
      first_time = false;
    } else {
      // Expect comma to separate items.
      if (LookAhead(ExprTokenType::kComma)) {
        Consume();
      } else {
        SetError(cur_token(), "Expected ',' separating expressions.");
        return {};
      }
    }

    TemplateTypeResult type_result = ExtractTemplateType(tokens_, cur_);
    if (!type_result.success) {
      SetError(tokens_[type_result.unmatched_error_token],
               fxl::StringPrintf("Unmatched '%s'.",
                                 tokens_[type_result.unmatched_error_token].value().c_str()));
      return {};
    } else if (cur_ == type_result.end_token) {
      if (!at_end()) {
        SetError(cur_token(), "Expected template parameter.");
      } else {
        SetError(ExprToken(), "Expected template parameter.");
      }
      return {};
    }
    cur_ = type_result.end_token;
    result.push_back(std::move(type_result.canonical_name));
  }
  return result;
}

// This function is called in contexts where we expect a comma-separated list.  Currently these are
// all known in advance so this simple manual parsing will do. A more general approach would
// implement a comma infix which constructs a new type of ExprNode.
std::vector<fxl::RefPtr<ExprNode>> ExprParser::ParseExpressionList(ExprTokenType stop_before) {
  std::vector<fxl::RefPtr<ExprNode>> result;

  bool first_time = true;
  while (!at_end() && !LookAhead(stop_before)) {
    if (first_time) {
      first_time = false;
    } else {
      // Expect comma to separate items.
      if (LookAhead(ExprTokenType::kComma)) {
        Consume();
      } else {
        SetError(cur_token(), "Expected ',' separating expressions.");
        return {};
      }
    }

    fxl::RefPtr<ExprNode> cur = ParseExpression(kPrecedenceComma);
    if (has_error())
      return {};
    result.push_back(std::move(cur));
  }

  return result;
}

fxl::RefPtr<ExprNode> ExprParser::AmpersandPrefix(const ExprToken& token) {
  fxl::RefPtr<ExprNode> right = ParseExpression(kPrecedenceUnary);
  if (has_error())
    return nullptr;
  return fxl::MakeRefCounted<AddressOfExprNode>(std::move(right));
}

fxl::RefPtr<ExprNode> ExprParser::BadToken(const ExprToken& token) {
  SetError(token, "Unexpected " + token.value());
  return nullptr;
}

fxl::RefPtr<ExprNode> ExprParser::BinaryOpInfix(fxl::RefPtr<ExprNode> left,
                                                const ExprToken& token) {
  const DispatchInfo& dispatch = DispatchForToken(token);

  // Allow empty expressions during parse so we can give a slightly better error message below.
  fxl::RefPtr<ExprNode> right = ParseExpression(dispatch.precedence, EmptyExpression::kAllow);
  if (!has_error() && !right) {
    SetError(token, fxl::StringPrintf("Expected expression after '%s'.", token.value().c_str()));
  }
  if (has_error())
    return nullptr;

  return fxl::MakeRefCounted<BinaryOpExprNode>(std::move(left), token, std::move(right));
}

fxl::RefPtr<ExprNode> ExprParser::DotOrArrowInfix(fxl::RefPtr<ExprNode> left,
                                                  const ExprToken& token) {
  // These are left-associative so use the same precedence as the token. Allow empty expressions
  // during parse so we can give a slightly better error message below.
  fxl::RefPtr<ExprNode> right = ParseExpression(kPrecedenceCallAccess, EmptyExpression::kAllow);

  auto literal = right ? right->AsLiteral() : nullptr;

  // Rust supports tuple structs, which can be addressed like "foo.0"
  if (language_ == ExprLanguage::kRust && literal && token.type() == ExprTokenType::kDot) {
    auto literal_token = literal->token();

    if (literal_token.type() == ExprTokenType::kInteger) {
      auto value = literal_token.value();

      if (value.find_first_not_of("0123456789") == std::string::npos &&
          (value.length() <= 1 || value[0] != '0')) {
        right = fxl::MakeRefCounted<IdentifierExprNode>(value);
      }
    }
  }

  if (!right || !right->AsIdentifier()) {
    SetError(token, fxl::StringPrintf("Expected identifier for right-hand-side of \"%s\".",
                                      token.value().c_str()));
    return nullptr;
  }

  // Use the name from the right-hand-side identifier, we don't need a full expression for that. If
  // we add function calls it will be necessary.
  return fxl::MakeRefCounted<MemberAccessExprNode>(std::move(left), token,
                                                   right->AsIdentifier()->ident());
}

fxl::RefPtr<ExprNode> ExprParser::LeftParenPrefix(const ExprToken& token) {
  // "(" as a prefix is a grouping or cast: "a + (b + c)" or "(Foo)bar" where it doesn't modify the
  // thing on the left. Evaluate the thing inside the () and return it.
  //
  // Currently there's no infix version of "(" which would be something like a function call.
  auto expr = ParseExpression(0);
  if (!has_error() && !expr)
    SetError(token, "Expected expression inside '('.");
  if (!has_error())
    Consume(ExprTokenType::kRightParen, "Expected ')' to match.", token);
  if (has_error())
    return nullptr;

  if (const TypeExprNode* type_expr = expr->AsType()) {
    // Convert "(TypeName)..." into a cast. Note the "-1" here which converts to right-associative.
    // With variable names, () is left-associative in that "(foo)(bar)[baz]" means to execute
    // left-to-right. But when "(foo)" is a C-style cast, this means "(bar)[baz]" is a unit.
    auto cast_expr = ParseExpression(kPrecedenceCallAccess - 1);
    if (has_error())
      return nullptr;

    fxl::RefPtr<TypeExprNode> type_ref = RefPtrTo(type_expr);
    return fxl::MakeRefCounted<CastExprNode>(CastType::kC, std::move(type_ref),
                                             std::move(cast_expr));
  }

  return expr;
}

fxl::RefPtr<ExprNode> ExprParser::LeftBracketPrefix(const ExprToken& token) {
  // We could say to terminate on a bracket. But by saying "none" we have the opportunity to give a
  // better error message below, referencing the opening bracket.
  auto block = ParseBlockContents(BlockEnd::kExplicit);
  if (!block)
    return nullptr;

  // Consume the terminating bracket.
  Consume(ExprTokenType::kRightBracket, "Expected '}' to match.", token);
  if (has_error())
    return nullptr;

  return block;
}

fxl::RefPtr<ExprNode> ExprParser::LeftParenInfix(fxl::RefPtr<ExprNode> left,
                                                 const ExprToken& token) {
  // "(" as an infix is a function call. In this case, the thing on the left identifies what to
  // call (either a bare identifier or an object accessor).
  if (!FunctionCallExprNode::IsValidCall(left)) {
    SetError(token, "Unexpected '('.");
    return nullptr;
  }

  // Read the function parameters.
  std::vector<fxl::RefPtr<ExprNode>> args = ParseExpressionList(ExprTokenType::kRightParen);
  if (has_error())
    return nullptr;
  Consume(ExprTokenType::kRightParen, "Expected ')' to match.", token);
  if (has_error())
    return nullptr;

  return fxl::MakeRefCounted<FunctionCallExprNode>(std::move(left), std::move(args));
}

fxl::RefPtr<ExprNode> ExprParser::LeftSquareInfix(fxl::RefPtr<ExprNode> left,
                                                  const ExprToken& token) {
  auto inner = ParseExpression(0);
  if (!has_error())
    Consume(ExprTokenType::kRightSquare, "Expected ']' to match.", token);
  if (has_error())
    return nullptr;
  return fxl::MakeRefCounted<ArrayAccessExprNode>(std::move(left), std::move(inner));
}

fxl::RefPtr<ExprNode> ExprParser::RustCastInfix(fxl::RefPtr<ExprNode> left,
                                                const ExprToken& token) {
  fxl::RefPtr<Type> type = ParseType(fxl::RefPtr<Type>());
  if (has_error())
    return nullptr;
  return fxl::MakeRefCounted<CastExprNode>(
      CastType::kRust,
      fxl::MakeRefCounted<TypeExprNode>(type, eval_context_->GetConcreteType(type)),
      std::move(left));
}

fxl::RefPtr<ExprNode> ExprParser::QuestionInfix(fxl::RefPtr<ExprNode> left,
                                                const ExprToken& token) {
  if (language_ == ExprLanguage::kRust) {
    // Rust uses '?' for error handling.
    SetError(token, "Rust '?' operators are not supported.");
    return nullptr;
  }

  // "If true" block.
  auto then = ParseExpression(0);
  if (has_error())
    return nullptr;

  // Colon.
  Consume(ExprTokenType::kColon, "Expected ':' for previous '?'.");
  if (has_error())
    return nullptr;

  // Else block.
  auto else_then = ParseExpression(0);
  if (has_error())
    return nullptr;

  ConditionExprNode::Condition cond;
  cond.cond = std::move(left);
  return fxl::MakeRefCounted<ConditionExprNode>(
      std::vector<ConditionExprNode::Pair>{{std::move(cond), std::move(then)}},
      std::move(else_then), std::nullopt);
}

fxl::RefPtr<ExprNode> ExprParser::LiteralPrefix(const ExprToken& token) {
  return fxl::MakeRefCounted<LiteralExprNode>(language_, token);
}

fxl::RefPtr<ExprNode> ExprParser::UnaryPrefix(const ExprToken& token) {
  // Allow empty expressions during parse to give a better error message below.
  auto inner = ParseExpression(kPrecedenceUnary, EmptyExpression::kAllow);
  if (!has_error() && !inner)
    SetError(token, fxl::StringPrintf("Expected expression for '%s'.", token.value().c_str()));
  if (has_error())
    return nullptr;
  return fxl::MakeRefCounted<UnaryOpExprNode>(token, std::move(inner));
}

fxl::RefPtr<ExprNode> ExprParser::NamePrefix(const ExprToken& token) {
  // Handles names and "::" which precedes names. This could be a typename ("int", or
  // "::std::vector<int>"), a variable name ("i", "std::basic_string<char>::npos"), and can also
  // include special names like "$plt(zx_foo_bar)"

  // Back up so the current token is the first component of the name so we can hand-off to the
  // specialized name parser.
  FX_DCHECK(cur_ > 0);
  cur_--;

  if (token.type() == ExprTokenType::kConst || token.type() == ExprTokenType::kVolatile ||
      token.type() == ExprTokenType::kRestrict) {
    // These start a type name, force type parsing mode.
    fxl::RefPtr<Type> type = ParseType(fxl::RefPtr<Type>());
    if (has_error())
      return nullptr;
    return ParseAfterType(
        fxl::MakeRefCounted<TypeExprNode>(type, eval_context_->GetConcreteType(type)));
  }

  // All other names.
  ParseNameResult result = ParseName(true);
  if (has_error())
    return nullptr;

  if (result.type) {
    return ParseAfterType(fxl::MakeRefCounted<TypeExprNode>(
        result.type, eval_context_->GetConcreteType(result.type)));
  }
  if (result.local_var_slot) {
    return fxl::MakeRefCounted<LocalVarExprNode>(*result.local_var_slot);
  }
  return fxl::MakeRefCounted<IdentifierExprNode>(std::move(result.ident));
}

fxl::RefPtr<ExprNode> ExprParser::StarPrefix(const ExprToken& token) {
  fxl::RefPtr<ExprNode> right = ParseExpression(kPrecedenceUnary);
  if (has_error())
    return nullptr;
  return fxl::MakeRefCounted<DereferenceExprNode>(std::move(right));
}

fxl::RefPtr<ExprNode> ExprParser::CastPrefix(const ExprToken& token) {
  CastType cast_type =
      token.type() == ExprTokenType::kReinterpretCast ? CastType::kReinterpret : CastType::kStatic;

  // "<" after reinterpret_cast.
  ExprToken left_angle = Consume(ExprTokenType::kLess, "Expected '< >' after cast.");
  if (has_error())
    return nullptr;

  // Type name.
  fxl::RefPtr<Type> dest_type = ParseType(fxl::RefPtr<Type>());
  if (has_error())
    return nullptr;

  // ">" after type name.
  Consume(ExprTokenType::kGreater, "Expected '>' to match.", left_angle);
  if (has_error())
    return nullptr;

  // "(" containing expression.
  ExprToken left_paren = Consume(ExprTokenType::kLeftParen, "Expected '(' for cast.");
  if (has_error())
    return nullptr;

  // Expression to be casted.
  fxl::RefPtr<ExprNode> expr = ParseExpression(0);
  if (has_error())
    return nullptr;

  // ")" at end.
  Consume(ExprTokenType::kRightParen, "Expected ')' to match.", left_paren);
  if (has_error())
    return nullptr;

  return fxl::MakeRefCounted<CastExprNode>(
      cast_type,
      fxl::MakeRefCounted<TypeExprNode>(dest_type, eval_context_->GetConcreteType(dest_type)),
      std::move(expr));
}

fxl::RefPtr<ExprNode> ExprParser::SizeofPrefix(const ExprToken& token) {
  // "(" containing expression.
  ExprToken left_paren = Consume(ExprTokenType::kLeftParen, "Expected '(' for cast.");
  if (has_error())
    return nullptr;

  // Expression to be sized.
  fxl::RefPtr<ExprNode> expr = ParseExpression(0);
  if (has_error())
    return nullptr;

  // ")" at end.
  Consume(ExprTokenType::kRightParen, "Expected ')' to match.", left_paren);
  if (has_error())
    return nullptr;

  return fxl::MakeRefCounted<SizeofExprNode>(std::move(expr));
}

fxl::RefPtr<ExprNode> ExprParser::IfPrefix(const ExprToken& token) {
  bool require_parens = language_ == ExprLanguage::kC;  // Rust doesn't need parens.

  std::vector<ConditionExprNode::Pair> conditions;
  fxl::RefPtr<ExprNode> else_then;

  // The if statement may allow variable introductions "if (int i = ...)" and this will track.
  LocalVarCounter var_counter(&local_vars_);

  // This is the language-specific way to parse a conditional block.
  std::function<fxl::RefPtr<ExprNode>()> parse_block;
  switch (language_) {
    case ExprLanguage::kRust:
      // Rust requires {}.
      parse_block = [this]() { return ParseBlock(BlockDelimiter::kExplicit); };
      break;
    case ExprLanguage::kC:
      // C allows a standalone statement or a block.
      parse_block = [this]() {
        return ParseExpression(0, EmptyExpression::kAllow, StatementEnd::kExplicit);
      };
      break;
  }

  while (true) {
    // "(" at beginning of condition.
    ExprToken left_paren;
    if (require_parens) {
      left_paren = Consume(ExprTokenType::kLeftParen, "Expected '(' for if.");
      if (has_error())
        return nullptr;
    }

    // Check for Rust "if let".
    ConditionExprNode::Condition condition;
    if (language_ == ExprLanguage::kRust && LookAhead(ExprTokenType::kLet)) {
      // Syntax we support is:
      //    "if" "let" <enum-name> [ "(" ( <var-name> "," )* <var-name> ")" ] "=" <expression>
      Consume();  // Eat the "let".

      ParseNameResult enum_name = ParseName(false);
      if (has_error())
        return nullptr;
      condition.rust_pattern_name = std::move(enum_name.ident);

      if (LookAhead(ExprTokenType::kLeftBracket)) {
        // Struct enums patterns looks like "if let Point{x, y} = pt" which we don't support.
        SetErrorAtCur("Struct pattern syntax is not supported yet, sorry.");
        return nullptr;
      }
      if (LookAhead(ExprTokenType::kLeftParen)) {
        // Rust tuple enum patterns.
        Consume(ExprTokenType::kLeftParen, "Expecting '(' after enumeration name.", cur_token());
        if (has_error())
          return nullptr;

        // Variable name(s) for the pattern.
        //
        // Note: this actually leaves the variable in scope for the "else" cases (it will just have
        // an empty value). It's a bit annoying to implement the scoping properly because variables
        // introduced in C++ are in scope for both the "if" and "else" arms of the condition.
        while (true) {
          const ExprToken& var_token =
              Consume(ExprTokenType::kName, "Expecting variable name.", cur_token());
          if (has_error())
            return nullptr;
          condition.rust_pattern_local_slots.push_back(static_cast<uint32_t>(local_vars_.size()));
          local_vars_.push_back(var_token.value());

          if (!LookAhead(ExprTokenType::kComma))
            break;    // Expecting a comma to separate names, anything else means end of list.
          Consume();  // Eat the comma.
        }

        Consume(ExprTokenType::kRightParen, "Expecting ')' after variable name.", cur_token());
        if (has_error())
          return nullptr;
      }
      Consume(ExprTokenType::kEquals, "Expecting '=' for 'if let' expression.", cur_token());
      if (has_error())
        return nullptr;
    }

    // Conditional expression.
    condition.cond = ParseExpression(0);
    if (has_error())
      return nullptr;

    // ")" at end of condition.
    if (require_parens) {
      Consume(ExprTokenType::kRightParen, "Expected ')' to match.", left_paren);
      if (has_error())
        return nullptr;
    }

    // Block to execute on match.
    conditions.emplace_back(std::move(condition), parse_block());
    if (has_error())
      return nullptr;

    // Can be optionally followed by an "else".
    if (!LookAhead(ExprTokenType::kElse))
      break;  // No "else", done with condition.

    // Handle "else".
    Consume();  // Consume the "else" we already identified.
    if (LookAhead(ExprTokenType::kIf)) {
      // "else if", wrap around to doing a new condition for this arm.
      Consume();  // Consume the "if" we already identified.
      continue;
    } else {
      // Standalone "else", get the expression for it.
      else_then = parse_block();
      if (has_error())
        return nullptr;
      break;
    }
  }

  return fxl::MakeRefCounted<ConditionExprNode>(std::move(conditions), std::move(else_then),
                                                var_counter.GetExitVarCount());
}

fxl::RefPtr<ExprNode> ExprParser::LetPrefix(const ExprToken& token) {
  // "let" <name> [ ":" <type> ] [ "=" <expression> ]
  ExprToken name = Consume(ExprTokenType::kName, "Expecting variable name for 'let' statement.");
  if (has_error())
    return nullptr;

  fxl::RefPtr<Type> type;
  if (LookAhead(ExprTokenType::kColon)) {
    Consume();
    type = ParseType(nullptr);
    if (has_error())
      return nullptr;
  }

  fxl::RefPtr<ExprNode> init_expr;
  if (LookAhead(ExprTokenType::kEquals)) {
    Consume();
    init_expr = ParseExpression(0);
    if (has_error())
      return nullptr;
  }

  // Decode the type information (note the type will be null if unspecified).
  fxl::RefPtr<Type> concrete_type;
  if (type)
    concrete_type = eval_context_->GetConcreteType(type);
  auto decl_or = GetVariableDeclTypeInfo(language_, std::move(concrete_type));
  if (decl_or.has_error()) {
    SetError(name, decl_or.err());
    return nullptr;
  }

  uint32_t local_slot = static_cast<uint32_t>(local_vars_.size());
  local_vars_.push_back(name.value());

  return fxl::MakeRefCounted<VariableDeclExprNode>(decl_or.take_value(), local_slot, name,
                                                   std::move(init_expr));
}

fxl::RefPtr<ExprNode> ExprParser::ForPrefix(const ExprToken& token) {
  // This is called only for C/C++. In Rust, the "for" token is not recognized by the tokenizer
  // because we can't handle Rust for loops (they require iterators). We also don't support C++
  // range-based for loops for the same reason. So the syntax we support is only:
  //
  //   "for" "(" [<expr>] ";" [<expr>] ";" [<expr>] ")" <statement>

  // "(" at beginning.
  ExprToken left_paren = Consume(ExprTokenType::kLeftParen, "Expected '(' for 'for' loop.");
  if (has_error())
    return nullptr;

  // We need to be able to pop any local variables created by the init expression at the bottom of
  // the loop.
  LocalVarCounter var_counter(&local_vars_);

  // Init expression (optional).
  fxl::RefPtr<ExprNode> init = ParseExpression(0, EmptyExpression::kAllow, StatementEnd::kExplicit);
  if (has_error())
    return nullptr;

  // Termination expression (optional).
  fxl::RefPtr<ExprNode> term = ParseExpression(0, EmptyExpression::kAllow, StatementEnd::kExplicit);
  if (has_error())
    return nullptr;

  // Increment expression (optional). Since this expression isn't terminated with a semicolon, we
  // need to filter out the empty case here or else ParseExpression will be confused by the closing
  // paren.
  fxl::RefPtr<ExprNode> incr;
  if (!LookAhead(ExprTokenType::kRightParen)) {
    incr = ParseExpression(0, EmptyExpression::kAllow, StatementEnd::kNone);
    if (has_error())
      return nullptr;
  }

  // ")" at end.
  Consume(ExprTokenType::kRightParen, "Expected ')' to match.", left_paren);
  if (has_error())
    return nullptr;

  // Loop contents.
  fxl::RefPtr<ExprNode> contents;
  {
    // Things inside of here support "break".
    ScopedBoolSetter break_setter(&allow_break_, true);

    contents = ParseExpression(0, EmptyExpression::kAllow, StatementEnd::kExplicit);
    if (has_error())
      return nullptr;
  }

  return fxl::MakeRefCounted<LoopExprNode>(token, std::move(init), var_counter.GetExitVarCount(),
                                           std::move(term), nullptr, std::move(incr),
                                           std::move(contents));
}

fxl::RefPtr<ExprNode> ExprParser::DoPrefix(const ExprToken& token) {
  // Loop contents.
  fxl::RefPtr<ExprNode> contents;
  {
    // Things inside of here support "break".
    ScopedBoolSetter break_setter(&allow_break_, true);

    contents = ParseExpression(0, EmptyExpression::kReject, StatementEnd::kNone);
    if (has_error())
      return nullptr;
  }

  // Terminating "while" statement.
  Consume(ExprTokenType::kWhile, "Expected 'while' for do-while loop.");
  if (has_error())
    return nullptr;

  // Opening "(".
  ExprToken left_paren = Consume(ExprTokenType::kLeftParen, "Expected '(' for 'do-while' loop.");
  if (has_error())
    return nullptr;

  // Condition expression.
  fxl::RefPtr<ExprNode> expr = ParseExpression(0, EmptyExpression::kReject, StatementEnd::kNone);
  if (has_error())
    return nullptr;

  // Closing ")".
  Consume(ExprTokenType::kRightParen, "Expected ')' to match.", left_paren);
  if (has_error())
    return nullptr;

  return fxl::MakeRefCounted<LoopExprNode>(token, nullptr, std::nullopt, nullptr, std::move(expr),
                                           nullptr, std::move(contents));
}

fxl::RefPtr<ExprNode> ExprParser::WhilePrefix(const ExprToken& token) {
  bool need_cond_parens = language_ == ExprLanguage::kC;

  // Opening "(".
  ExprToken left_paren;
  if (need_cond_parens) {
    left_paren = Consume(ExprTokenType::kLeftParen, "Expected '(' for 'while' loop.");
    if (has_error())
      return nullptr;
  }

  // Condition expression (always required).
  fxl::RefPtr<ExprNode> expr = ParseExpression(0, EmptyExpression::kReject, StatementEnd::kNone);
  if (has_error())
    return nullptr;

  // Closing ")".
  if (need_cond_parens) {
    Consume(ExprTokenType::kRightParen, "Expected ')' to match.", left_paren);
    if (has_error())
      return nullptr;
  }

  // Loop contents.
  fxl::RefPtr<ExprNode> contents;
  {
    // Things inside of here support "break".
    ScopedBoolSetter break_setter(&allow_break_, true);

    if (language_ == ExprLanguage::kRust) {
      // Rust loops require {}.
      contents = ParseBlock(BlockDelimiter::kExplicit);
    } else {
      // C also allows statements without {}.
      contents = ParseExpression(0, EmptyExpression::kAllow, StatementEnd::kExplicit);
    }
    if (has_error())
      return nullptr;
  }

  return fxl::MakeRefCounted<LoopExprNode>(token, nullptr, std::nullopt, std::move(expr), nullptr,
                                           nullptr, std::move(contents));
}

// Rust ininite loop:
//   loop { ... }
fxl::RefPtr<ExprNode> ExprParser::LoopPrefix(const ExprToken& token) {
  // Things inside of here support "break".
  ScopedBoolSetter break_setter(&allow_break_, true);

  fxl::RefPtr<ExprNode> contents = ParseBlock(BlockDelimiter::kExplicit);
  if (has_error())
    return nullptr;
  return fxl::MakeRefCounted<LoopExprNode>(token, nullptr, std::nullopt, nullptr, nullptr, nullptr,
                                           std::move(contents));
}

fxl::RefPtr<ExprNode> ExprParser::BreakPrefix(const ExprToken& token) {
  if (!allow_break_) {
    SetError(token, "Use of 'break' in this context is not allowed.");
    return nullptr;
  }
  return fxl::MakeRefCounted<BreakExprNode>(token);
}

bool ExprParser::LookAhead(ExprTokenType type) const {
  if (at_end())
    return false;
  return cur_token().type() == type;
}

const ExprToken& ExprParser::Consume() {
  if (at_end())
    return kInvalidToken;
  return tokens_[cur_++];
}

const ExprToken& ExprParser::Consume(ExprTokenType type, const char* error_msg,
                                     const ExprToken& error_token) {
  FX_DCHECK(!has_error());  // Should have error-checked before calling.
  if (at_end()) {
    SetError(error_token, std::string(error_msg) + " Hit the end of input instead.");
    return kInvalidToken;
  }

  if (cur_token().type() == type)
    return Consume();

  SetError(error_token.type() == ExprTokenType::kInvalid ? cur_token() : error_token, error_msg);
  return kInvalidToken;
}

ExprToken ExprParser::ConsumeWithShiftTokenConversion() {
  if (IsCurTokenShiftRight()) {
    // Eat both ">" operators and synthesize a shift operator.
    const ExprToken& first = Consume();
    Consume();
    return ExprToken(ExprTokenType::kShiftRight, ">>", first.byte_offset());
  }
  if (IsCurTokenShiftRightEquals()) {
    // Eat both ">" and ">=" operators and synthesize a shift operator.
    const ExprToken& first = Consume();
    Consume();
    return ExprToken(ExprTokenType::kShiftRightEquals, ">>=", first.byte_offset());
  }
  return Consume();
}

void ExprParser::ConsumeCVQualifier(std::vector<DwarfTag>* qual) {
  while (!at_end()) {
    const ExprToken& token = cur_token();

    DwarfTag tag = DwarfTag::kNone;
    if (token.type() == ExprTokenType::kConst) {
      tag = DwarfTag::kConstType;
    } else if (token.type() == ExprTokenType::kVolatile) {
      tag = DwarfTag::kVolatileType;
    } else if (token.type() == ExprTokenType::kRestrict) {
      tag = DwarfTag::kRestrictType;
    } else {
      // Not a qualification token, done.
      return;
    }

    // Can't have duplicates.
    if (std::find(qual->begin(), qual->end(), tag) != qual->end()) {
      SetError(token,
               fxl::StringPrintf("Duplicate '%s' type qualification.", token.value().c_str()));
      return;
    }

    qual->push_back(tag);
    Consume();
  }
}

fxl::RefPtr<Type> ExprParser::ApplyQualifiers(fxl::RefPtr<Type> input,
                                              const std::vector<DwarfTag>& qual) {
  fxl::RefPtr<Type> type = std::move(input);

  // Apply the qualifiers in reverse order so the rightmost one is applied first.
  for (auto iter = qual.rbegin(); iter != qual.rend(); ++iter) {
    type = fxl::MakeRefCounted<ModifiedType>(*iter, std::move(type));
  }
  return type;
}

void ExprParser::SetError(const ExprToken& token, Err err) {
  err_ = std::move(err);
  error_token_ = token;
}

void ExprParser::SetError(const ExprToken& token, std::string msg) {
  err_ = Err(std::move(msg));
  error_token_ = token;
}

void ExprParser::SetErrorAtCur(std::string msg) {
  if (at_end())
    SetError(ExprToken(), msg + " Hit the end of input instead.");
  else
    SetError(cur_token(), msg);
}

bool ExprParser::IsCurTokenShiftRight() const {
  if (tokens_.size() < 2 || cur_ > tokens_.size() - 2)
    return false;  // Not enough room for two tokens.

  if (tokens_[cur_].type() != ExprTokenType::kGreater ||
      tokens_[cur_ + 1].type() != ExprTokenType::kGreater)
    return false;  // Not two ">" in a row.

  // They must also be next to each other with no space.
  return tokens_[cur_].ImmediatelyPrecedes(tokens_[cur_ + 1]);
}

bool ExprParser::IsCurTokenShiftRightEquals() const {
  if (tokens_.size() < 2 || cur_ > tokens_.size() - 2)
    return false;  // Not enough room for two tokens.

  if (tokens_[cur_].type() != ExprTokenType::kGreater ||
      tokens_[cur_ + 1].type() != ExprTokenType::kGreaterEqual)
    return false;  // Not ">" followed by ">=".

  // They must also be next to each other with no space.
  return tokens_[cur_].ImmediatelyPrecedes(tokens_[cur_ + 1]);
}

int ExprParser::CurPrecedenceWithShiftTokenConversion() const {
  if (IsCurTokenShiftRight())
    return kPrecedenceShift;
  if (IsCurTokenShiftRightEquals())
    return kPrecedenceAssignment;
  return DispatchForToken(cur_token()).precedence;
}

std::optional<uint32_t> ExprParser::GetLocalVariable(const ParsedIdentifier& name) const {
  // Only simple one-compomnent non-special identifiers can be local variables.
  if (name.qualification() != IdentifierQualification::kRelative || name.components().size() != 1 ||
      name.components()[0].special() != SpecialIdentifier::kNone) {
    return std::nullopt;
  }
  const std::string& name_str = name.components()[0].name();

  // Since variables are inserted by appending, search backwards to find the most recently
  // declared one of the given name.
  for (int i = static_cast<int>(local_vars_.size()) - 1; i >= 0; i--) {
    if (local_vars_[i] == name_str)
      return static_cast<uint32_t>(i);
  }
  return std::nullopt;
}

bool ExprParser::CountsAsStatementEnd(const ExprNode* node) const {
  // This function does the checking so that, for example:
  //   if (foo) {
  //     ...
  //   }
  // doesn't need a semicolon at the end to indicate the end of a statement.
  switch (language_) {
    case ExprLanguage::kC:
      if (node->AsBlock() || node->AsCondition())
        return true;
      if (const LoopExprNode* loop = node->AsLoop()) {
        // For and while loops don't need semicolons in C, but do-while loops do!
        return loop->token().type() == ExprTokenType::kFor ||
               loop->token().type() == ExprTokenType::kWhile;
      }
      return false;
    case ExprLanguage::kRust:
      return node->AsBlock() || node->AsCondition() || node->AsLoop();
  }
  return false;
}

// static
const ExprParser::DispatchInfo& ExprParser::DispatchForToken(const ExprToken& token) {
  size_t index = static_cast<size_t>(token.type());
  FX_DCHECK(index < std::size(kDispatchInfo));
  return kDispatchInfo[index];
}

}  // namespace zxdb