path: root/src/parse_tree.cpp

// Programmatic representation of fish code.
#include "config.h"  // IWYU pragma: keep

#include <assert.h>
#include <stdarg.h>
#include <stddef.h>
#include <stdio.h>
#include <wchar.h>
#include <algorithm>
#include <cwchar>
#include <string>
#include <vector>

#include "common.h"
#include "fallback.h" 
#include "parse_constants.h"
#include "parse_productions.h"
#include "parse_tree.h"
#include "proc.h"
#include "tokenizer.h"
#include "wutil.h" // IWYU pragma: keep

// This array provides strings for each symbol in enum parse_token_type_t in parse_constants.h.
const wchar_t *const token_type_map[] = {
    L"token_type_invalid",
    L"symbol_job_list",
    L"symbol_job",
    L"symbol_job_continuation",
    L"symbol_statement",
    L"symbol_block_statement",
    L"symbol_block_header",
    L"symbol_for_header",
    L"symbol_while_header",
    L"symbol_begin_header",
    L"symbol_function_header",
    L"symbol_if_statement",
    L"symbol_if_clause",
    L"symbol_else_clause",
    L"symbol_else_continuation",
    L"symbol_switch_statement",
    L"symbol_case_item_list",
    L"symbol_case_item",
    L"symbol_boolean_statement",
    L"symbol_decorated_statement",
    L"symbol_plain_statement",
    L"symbol_arguments_or_redirections_list",
    L"symbol_argument_or_redirection",
    L"symbol_andor_job_list",
    L"symbol_argument_list",
    L"symbol_freestanding_argument_list",
    L"symbol_argument",
    L"symbol_redirection",
    L"symbol_optional_background",
    L"symbol_end_command",
    L"parse_token_type_string",
    L"parse_token_type_pipe",
    L"parse_token_type_redirection",
    L"parse_token_type_background",
    L"parse_token_type_end",
    L"parse_token_type_terminate",
    L"parse_special_type_parse_error",
    L"parse_special_type_tokenizer_error",
    L"parse_special_type_comment",
};

using namespace parse_productions;

static bool production_is_empty(const production_t *production) {
    return (*production)[0] == token_type_invalid;
}

/// Returns a string description of this parse error.
wcstring parse_error_t::describe_with_prefix(const wcstring &src, const wcstring &prefix,
                                             bool is_interactive, bool skip_caret) const {
    wcstring result = text;
    if (!skip_caret && source_start < src.size() && source_start + source_length <= src.size()) {
        // Locate the beginning of this line of source.
        size_t line_start = 0;

        // Look for a newline prior to source_start. If we don't find one, start at the beginning of
        // the string; otherwise start one past the newline. Note that source_start may itself point
        // at a newline; we want to find the newline before it.
        if (source_start > 0) {
            size_t newline = src.find_last_of(L'\n', source_start - 1);
            if (newline != wcstring::npos) {
                line_start = newline + 1;
            }
        }

        // Look for the newline after the source range. If the source range itself includes a
        // newline, that's the one we want, so start just before the end of the range.
        size_t last_char_in_range =
            (source_length == 0 ? source_start : source_start + source_length - 1);
        size_t line_end = src.find(L'\n', last_char_in_range);
        if (line_end == wcstring::npos) {
            line_end = src.size();
        }

        assert(line_end >= line_start);
        assert(source_start >= line_start);

        // Don't include the caret and line if we're interactive this is the first line, because
        // then it's obvious.
        bool skip_caret = (is_interactive && source_start == 0);

        if (!skip_caret) {
            // Append the line of text.
            if (!result.empty()) {
                result.push_back(L'\n');
            }
            result.append(prefix);
            result.append(src, line_start, line_end - line_start);

            // Append the caret line. The input source may include tabs; for that reason we
            // construct a "caret line" that has tabs in corresponding positions.
            const wcstring line_to_measure =
                prefix + wcstring(src, line_start, source_start - line_start);
            wcstring caret_space_line;
            caret_space_line.reserve(source_start - line_start);
            for (size_t i = 0; i < line_to_measure.size(); i++) {
                wchar_t wc = line_to_measure.at(i);
                if (wc == L'\t') {
                    caret_space_line.push_back(L'\t');
                } else if (wc == L'\n') {
                    // It's possible that the source_start points at a newline itself. In that case,
                    // pretend it's a space. We only expect this to be at the end of the string.
                    caret_space_line.push_back(L' ');
                } else {
                    int width = fish_wcwidth(wc);
                    if (width > 0) {
                        caret_space_line.append(static_cast<size_t>(width), L' ');
                    }
                }
            }
            result.push_back(L'\n');
            result.append(caret_space_line);
            result.push_back(L'^');
        }
    }
    return result;
}

wcstring parse_error_t::describe(const wcstring &src) const {
    return this->describe_with_prefix(src, wcstring(), shell_is_interactive(), false);
}

void parse_error_offset_source_start(parse_error_list_t *errors, size_t amt) {
    assert(errors != NULL);
    if (amt > 0) {
        size_t i, max = errors->size();
        for (i = 0; i < max; i++) {
            parse_error_t *error = &errors->at(i);
            // Preserve the special meaning of -1 as 'unknown'.
            if (error->source_start != SOURCE_LOCATION_UNKNOWN) {
                error->source_start += amt;
            }
        }
    }
}

/// Returns a string description for the given token type.
const wchar_t *token_type_description(parse_token_type_t type) {
    if (type >= 0 && type <= LAST_TOKEN_TYPE) return token_type_map[type];

    // This leaks memory but it should never be run unless we have a bug elsewhere in the code.
    const wcstring d = format_string(L"unknown_token_type_%ld", static_cast<long>(type));
    wchar_t *d2 = new wchar_t[d.size() + 1];
    // cppcheck-suppress memleak
    return std::wcscpy(d2, d.c_str());
}

#define LONGIFY(x) L##x
#define KEYWORD_MAP(x) \
    { parse_keyword_##x, LONGIFY(#x) }
static const struct {
    const parse_keyword_t keyword;
    const wchar_t *const name;
}
keyword_map[] =
{
    // Note that these must be sorted (except for the first), so that we can do binary search.
    KEYWORD_MAP(none),
    KEYWORD_MAP(and),
    KEYWORD_MAP(begin),
    KEYWORD_MAP(builtin),
    KEYWORD_MAP(case),
    KEYWORD_MAP(command),
    KEYWORD_MAP(else),
    KEYWORD_MAP(end),
    KEYWORD_MAP(exec),
    KEYWORD_MAP(for),
    KEYWORD_MAP(function),
    KEYWORD_MAP(if),
    KEYWORD_MAP(in),
    KEYWORD_MAP(not),
    KEYWORD_MAP(or),
    KEYWORD_MAP(switch),
    KEYWORD_MAP(while)
};

const wchar_t *keyword_description(parse_keyword_t type) {
    if (type >= 0 && type <= LAST_KEYWORD) return keyword_map[type].name;

    // This leaks memory but it should never be run unless we have a bug elsewhere in the code.
    const wcstring d = format_string(L"unknown_keyword_%ld", static_cast<long>(type));
    wchar_t *d2 = new wchar_t[d.size() + 1];
    // cppcheck-suppress memleak
    return std::wcscpy(d2, d.c_str());
}

static wcstring token_type_user_presentable_description(
    parse_token_type_t type, parse_keyword_t keyword = parse_keyword_none) {
    if (keyword != parse_keyword_none) {
        return format_string(L"keyword '%ls'", keyword_description(keyword));
    }

    switch (type) {
        // Hackish. We only support the following types.
        case symbol_statement: {
            return L"a command";
        }
        case symbol_argument: {
            return L"an argument";
        }
        case parse_token_type_string: {
            return L"a string";
        }
        case parse_token_type_pipe: {
            return L"a pipe";
        }
        case parse_token_type_redirection: {
            return L"a redirection";
        }
        case parse_token_type_background: {
            return L"a '&'";
        }
        case parse_token_type_end: {
            return L"end of the statement";
        }
        case parse_token_type_terminate: {
            return L"end of the input";
        }
        default: { return format_string(L"a %ls", token_type_description(type)); }
    }
}

static wcstring block_type_user_presentable_description(parse_token_type_t type) {
    switch (type) {
        case symbol_for_header: {
            return L"for loop";
        }
        case symbol_while_header: {
            return L"while loop";
        }
        case symbol_function_header: {
            return L"function definition";
        }
        case symbol_begin_header: {
            return L"begin";
        }
        case symbol_if_statement: {
            return L"if statement";
        }
        case symbol_switch_statement: {
            return L"switch statement";
        }
        default: { return token_type_description(type); }
    }
}

/// Returns a string description of the given parse node.
wcstring parse_node_t::describe() const {
    wcstring result = token_type_description(this->type);
    return result;
}

/// Returns a string description of the given parse token.
wcstring parse_token_t::describe() const {
    wcstring result = token_type_description(type);
    if (keyword != parse_keyword_none) {
        append_format(result, L" <%ls>", keyword_description(keyword));
    }
    return result;
}

/// A string description appropriate for presentation to the user.
wcstring parse_token_t::user_presentable_description() const {
    return token_type_user_presentable_description(type, keyword);
}

/// Convert from tokenizer_t's token type to a parse_token_t type.
static inline parse_token_type_t parse_token_type_from_tokenizer_token(
    enum token_type tokenizer_token_type) {
    parse_token_type_t result = token_type_invalid;
    switch (tokenizer_token_type) {
        case TOK_STRING: {
            result = parse_token_type_string;
            break;
        }
        case TOK_PIPE: {
            result = parse_token_type_pipe;
            break;
        }
        case TOK_END: {
            result = parse_token_type_end;
            break;
        }
        case TOK_BACKGROUND: {
            result = parse_token_type_background;
            break;
        }
        case TOK_REDIRECT_OUT:
        case TOK_REDIRECT_APPEND:
        case TOK_REDIRECT_IN:
        case TOK_REDIRECT_FD:
        case TOK_REDIRECT_NOCLOB: {
            result = parse_token_type_redirection;
            break;
        }
        case TOK_ERROR: {
            result = parse_special_type_tokenizer_error;
            break;
        }
        case TOK_COMMENT: {
            result = parse_special_type_comment;
            break;
        }
        default: {
            fprintf(stderr, "Bad token type %d passed to %s\n", (int)tokenizer_token_type,
                    __FUNCTION__);
            assert(0);
            break;
        }
    }
    return result;
}

#if 0
// Disabled for the 2.2.0 release: https://github.com/fish-shell/fish-shell/issues/1809.

/// Helper function for parse_dump_tree().
static void dump_tree_recursive(const parse_node_tree_t &nodes, const wcstring &src,
                                node_offset_t node_idx, size_t indent, wcstring *result,
                                size_t *line, node_offset_t *inout_first_node_not_dumped) {
    assert(node_idx < nodes.size());

    // Update first_node_not_dumped. This takes a bit of explanation. While it's true that a parse
    // tree may be a "forest",  its individual trees are "compact," meaning they are not
    // interleaved. Thus we keep track of the largest node index as we descend a tree. One past the
    // largest is the start of the next tree.
    if (*inout_first_node_not_dumped <= node_idx) {
        *inout_first_node_not_dumped = node_idx + 1;
    }

    const parse_node_t &node = nodes.at(node_idx);

    const size_t spacesPerIndent = 2;

    // Unindent statement lists by 1 to flatten them.
    if (node.type == symbol_job_list || node.type == symbol_arguments_or_redirections_list) {
        if (indent > 0) indent -= 1;
    }

    append_format(*result, L"%2lu - %l2u  ", *line, node_idx);
    result->append(indent * spacesPerIndent, L' ');
    ;
    result->append(node.describe());
    if (node.child_count > 0) {
        append_format(*result, L" <%lu children>", node.child_count);
    }
    if (node.has_comments()) {
        append_format(*result, L" <has_comments>", node.child_count);
    }

    if (node.has_source() && node.type == parse_token_type_string) {
        result->append(L": \"");
        result->append(src, node.source_start, node.source_length);
        result->append(L"\"");
    }

    if (node.type != parse_token_type_string) {
        if (node.has_source()) {
            append_format(*result, L"  [%ld, %ld]", (long)node.source_start,
                          (long)node.source_length);
        } else {
            append_format(*result, L"  [no src]", (long)node.source_start,
                          (long)node.source_length);
        }
    }

    result->push_back(L'\n');
    ++*line;
    for (node_offset_t child_idx = node.child_start;
         child_idx < node.child_start + node.child_count; child_idx++) {
        dump_tree_recursive(nodes, src, child_idx, indent + 1, result, line,
                            inout_first_node_not_dumped);
    }
}

/// Gives a debugging textual description of a parse tree. Note that this supports "parse forests"
/// too. That is, our tree may not really be a tree, but instead a collection of trees.
wcstring parse_dump_tree(const parse_node_tree_t &nodes, const wcstring &src) {
    if (nodes.empty()) return L"(empty!)";

    node_offset_t first_node_not_dumped = 0;
    size_t line = 0;
    wcstring result;
    while (first_node_not_dumped < nodes.size()) {
        if (first_node_not_dumped > 0) {
            result.append(L"---New Tree---\n");
        }
        dump_tree_recursive(nodes, src, first_node_not_dumped, 0, &result, &line,
                            &first_node_not_dumped);
    }
    return result;
}
#endif

/// Struct representing elements of the symbol stack, used in the internal state of the LL parser.
struct parse_stack_element_t {
    enum parse_token_type_t type;
    enum parse_keyword_t keyword;
    node_offset_t node_idx;

    explicit parse_stack_element_t(parse_token_type_t t, node_offset_t idx)
        : type(t), keyword(parse_keyword_none), node_idx(idx) {}

    explicit parse_stack_element_t(production_element_t e, node_offset_t idx)
        : type(production_element_type(e)), keyword(production_element_keyword(e)), node_idx(idx) {}

    wcstring describe(void) const {
        wcstring result = token_type_description(type);
        if (keyword != parse_keyword_none) {
            append_format(result, L" <%ls>", keyword_description(keyword));
        }
        return result;
    }

    /// Returns a name that we can show to the user, e.g. "a command".
    wcstring user_presentable_description(void) const {
        return token_type_user_presentable_description(type, keyword);
    }
};

/// The parser itself, private implementation of class parse_t. This is a hand-coded table-driven LL
/// parser. Most hand-coded LL parsers are recursive descent, but recursive descent parsers are
/// difficult to "pause", unlike table-driven parsers.
class parse_ll_t {
    // Traditional symbol stack of the LL parser.
    std::vector<parse_stack_element_t> symbol_stack;
    // Parser output. This is a parse tree, but stored in an array.
    parse_node_tree_t nodes;
    // Whether we ran into a fatal error, including parse errors or tokenizer errors.
    bool fatal_errored;
    // Whether we should collect error messages or not.
    bool should_generate_error_messages;
    // List of errors we have encountered.
    parse_error_list_t errors;
    // The symbol stack can contain terminal types or symbols. Symbols go on to do productions, but
    // terminal types are just matched against input tokens.
    bool top_node_handle_terminal_types(parse_token_t token);

    void parse_error_unexpected_token(const wchar_t *expected, parse_token_t token);
    void parse_error(parse_token_t token, parse_error_code_t code, const wchar_t *format, ...);
    void parse_error_at_location(size_t location, parse_error_code_t code, const wchar_t *format,
                                 ...);
    void parse_error_failed_production(struct parse_stack_element_t &elem, parse_token_t token);
    void parse_error_unbalancing_token(parse_token_t token);

    // Reports an error for an unclosed block, e.g. 'begin;'. Returns true on success, false on
    // failure (e.g. it is not an unclosed block).
    bool report_error_for_unclosed_block();

    // void dump_stack(void) const;

    /// Get the node corresponding to the top element of the stack.
    parse_node_t &node_for_top_symbol() {
        PARSE_ASSERT(!symbol_stack.empty());
        const parse_stack_element_t &top_symbol = symbol_stack.back();
        PARSE_ASSERT(top_symbol.node_idx != NODE_OFFSET_INVALID);
        PARSE_ASSERT(top_symbol.node_idx < nodes.size());
        return nodes.at(top_symbol.node_idx);
    }

    /// Pop from the top of the symbol stack, then push the given production, updating node counts.
    /// Note that production_t has type "pointer to array" so some care is required.
    inline void symbol_stack_pop_push_production(const production_t *production) {
        bool logit = false;
        if (logit) {
            size_t count = 0;
            fprintf(stderr, "Applying production:\n");
            for (size_t i = 0; i < MAX_SYMBOLS_PER_PRODUCTION; i++) {
                production_element_t elem = (*production)[i];
                if (production_element_is_valid(elem)) {
                    parse_token_type_t type = production_element_type(elem);
                    parse_keyword_t keyword = production_element_keyword(elem);
                    fprintf(stderr, "\t%ls <%ls>\n", token_type_description(type),
                            keyword_description(keyword));
                    count++;
                }
            }
            if (!count) fprintf(stderr, "\t<empty>\n");
        }

        // Get the parent index. But we can't get the parent parse node yet, since it may be made
        // invalid by adding children.
        const node_offset_t parent_node_idx = symbol_stack.back().node_idx;

        // Add the children. Confusingly, we want our nodes to be in forwards order (last token
        // last, so dumps look nice), but the symbols should be reverse order (last token first, so
        // it's lowest on the stack)
        const size_t child_start_big = nodes.size();
        assert(child_start_big < NODE_OFFSET_INVALID);
        node_offset_t child_start = static_cast<node_offset_t>(child_start_big);

        // To avoid constructing multiple nodes, we make a single one that we modify.
        parse_node_t representative_child(token_type_invalid);
        representative_child.parent = parent_node_idx;

        node_offset_t child_count = 0;
        for (size_t i = 0; i < MAX_SYMBOLS_PER_PRODUCTION; i++) {
            production_element_t elem = (*production)[i];
            if (!production_element_is_valid(elem)) {
                break;  // all done, bail out
            }

            // Append the parse node.
            representative_child.type = production_element_type(elem);
            nodes.push_back(representative_child);
            child_count++;
        }

        // Update the parent.
        parse_node_t &parent_node = nodes.at(parent_node_idx);

        // Should have no children yet.
        PARSE_ASSERT(parent_node.child_count == 0);

        // Tell the node about its children.
        parent_node.child_start = child_start;
        parent_node.child_count = child_count;

        // Replace the top of the stack with new stack elements corresponding to our new nodes. Note
        // that these go in reverse order.
        symbol_stack.pop_back();
        symbol_stack.reserve(symbol_stack.size() + child_count);
        node_offset_t idx = child_count;
        while (idx--) {
            production_element_t elem = (*production)[idx];
            PARSE_ASSERT(production_element_is_valid(elem));
            symbol_stack.push_back(parse_stack_element_t(elem, child_start + idx));
        }
    }

   public:
    // Constructor
    explicit parse_ll_t(enum parse_token_type_t goal)
        : fatal_errored(false), should_generate_error_messages(true) {
        this->symbol_stack.reserve(16);
        this->nodes.reserve(64);
        this->reset_symbols_and_nodes(goal);
    }

    // Input
    void accept_tokens(parse_token_t token1, parse_token_t token2);

    /// Report tokenizer errors.
    void report_tokenizer_error(const tok_t &tok);

    /// Indicate if we hit a fatal error.
    bool has_fatal_error(void) const { return this->fatal_errored; }

    /// Indicate whether we want to generate error messages.
    void set_should_generate_error_messages(bool flag) {
        this->should_generate_error_messages = flag;
    }

    /// Clear the parse symbol stack (but not the node tree). Add a node of the given type as the
    /// goal node. This is called from the constructor.
    void reset_symbols(enum parse_token_type_t goal);

    /// Clear the parse symbol stack and the node tree. Add a node of the given type as the goal
    /// node. This is called from the constructor.
    void reset_symbols_and_nodes(enum parse_token_type_t goal);

    /// Once parsing is complete, determine the ranges of intermediate nodes.
    void determine_node_ranges();

    /// Acquire output after parsing. This transfers directly from within self.
    void acquire_output(parse_node_tree_t *output, parse_error_list_t *errors);
};

#if 0
void parse_ll_t::dump_stack(void) const {
    // Walk backwards from the top, looking for parents.
    wcstring_list_t lines;
    if (symbol_stack.empty()) {
        lines.push_back(L"(empty)");
    } else {
        node_offset_t child = symbol_stack.back().node_idx;
        node_offset_t cursor = child;
        lines.push_back(nodes.at(cursor).describe());
        while (cursor--) {
            const parse_node_t &node = nodes.at(cursor);
            if (node.child_start <= child && node.child_start + node.child_count > child) {
                lines.push_back(node.describe());
                child = cursor;
            }
        }
    }

    fprintf(stderr, "Stack dump (%zu elements):\n", symbol_stack.size());
    for (size_t idx = 0; idx < lines.size(); idx++) {
        fprintf(stderr, "    %ls\n", lines.at(idx).c_str());
    }
}
#endif

// Give each node a source range equal to the union of the ranges of its children. Terminal nodes
// already have source ranges (and no children). Since children always appear after their parents,
// we can implement this very simply by walking backwards. We then do a second pass to give empty
// nodes an empty source range (but with a valid offset). We do this by walking forward. If a child
// of a node has an invalid source range, we set it equal to the end of the source range of its
// previous child.
void parse_ll_t::determine_node_ranges(void) {
    size_t idx = nodes.size();
    while (idx--) {
        parse_node_t *parent = &nodes[idx];

        // Skip nodes that already have a source range. These are terminal nodes.
        if (parent->source_start != SOURCE_OFFSET_INVALID) continue;

        // Ok, this node needs a source range. Get all of its children, and then set its range.
        source_offset_t min_start = SOURCE_OFFSET_INVALID,
                        max_end = 0;  // note SOURCE_OFFSET_INVALID is huge
        for (node_offset_t i = 0; i < parent->child_count; i++) {
            const parse_node_t &child = nodes.at(parent->child_offset(i));
            if (child.has_source()) {
                min_start = std::min(min_start, child.source_start);
                max_end = std::max(max_end, child.source_start + child.source_length);
            }
        }

        if (min_start != SOURCE_OFFSET_INVALID) {
            assert(max_end >= min_start);
            parent->source_start = min_start;
            parent->source_length = max_end - min_start;
        }
    }

    // Forward pass.
    size_t size = nodes.size();
    for (idx = 0; idx < size; idx++) {
        // Since we populate the source range based on the sibling node, it's simpler to walk over
        // the children of each node. We keep a running "child_source_cursor" which is meant to be
        // the end of the child's source range. It's initially set to the beginning of the parent'
        // source range.
        parse_node_t *parent = &nodes[idx];
        // If the parent doesn't have a valid source range, then none of its children will either;
        // skip it entirely.
        if (parent->source_start == SOURCE_OFFSET_INVALID) {
            continue;
        }
        source_offset_t child_source_cursor = parent->source_start;
        for (size_t child_idx = 0; child_idx < parent->child_count; child_idx++) {
            parse_node_t *child = &nodes[parent->child_start + child_idx];
            if (child->source_start == SOURCE_OFFSET_INVALID) {
                child->source_start = child_source_cursor;
            }
            child_source_cursor = child->source_start + child->source_length;
        }
    }
}

void parse_ll_t::acquire_output(parse_node_tree_t *output, parse_error_list_t *errors) {
    if (output != NULL) {
        output->swap(this->nodes);
    }
    this->nodes.clear();

    if (errors != NULL) {
        errors->swap(this->errors);
    }
    this->errors.clear();
    this->symbol_stack.clear();
}

void parse_ll_t::parse_error(parse_token_t token, parse_error_code_t code, const wchar_t *fmt,
                             ...) {
    this->fatal_errored = true;
    if (this->should_generate_error_messages) {
        // this->dump_stack();
        parse_error_t err;

        va_list va;
        va_start(va, fmt);
        err.text = vformat_string(fmt, va);
        err.code = code;
        va_end(va);

        err.source_start = token.source_start;
        err.source_length = token.source_length;
        this->errors.push_back(err);
    }
}

void parse_ll_t::parse_error_at_location(size_t source_location, parse_error_code_t code,
                                         const wchar_t *fmt, ...) {
    this->fatal_errored = true;
    if (this->should_generate_error_messages) {
        // this->dump_stack();
        parse_error_t err;

        va_list va;
        va_start(va, fmt);
        err.text = vformat_string(fmt, va);
        err.code = code;
        va_end(va);

        err.source_start = source_location;
        err.source_length = 0;
        this->errors.push_back(err);
    }
}

// Unbalancing token. This includes 'else' or 'case' or 'end' outside of the appropriate block
// This essentially duplicates some logic from resolving the production for symbol_statement_list -
// yuck.
void parse_ll_t::parse_error_unbalancing_token(parse_token_t token) {
    this->fatal_errored = true;
    if (this->should_generate_error_messages) {
        switch (token.keyword) {
            case parse_keyword_end: {
                this->parse_error(token, parse_error_unbalancing_end, L"'end' outside of a block");
                break;
            }
            case parse_keyword_else: {
                this->parse_error(token, parse_error_unbalancing_else,
                                  L"'else' builtin not inside of if block");
                break;
            }
            case parse_keyword_case: {
                this->parse_error(token, parse_error_unbalancing_case,
                                  L"'case' builtin not inside of switch block");
                break;
            }
            default: {
                // At the moment, this case should only be hit if you parse a
                // freestanding_argument_list. For example, 'complete -c foo -a 'one & three'.
                // Hackish error message for that case.
                if (!symbol_stack.empty() &&
                    symbol_stack.back().type == symbol_freestanding_argument_list) {
                    this->parse_error(
                        token, parse_error_generic, L"Expected %ls, but found %ls",
                        token_type_user_presentable_description(symbol_argument).c_str(),
                        token.user_presentable_description().c_str());
                } else {
                    this->parse_error(token, parse_error_generic, L"Did not expect %ls",
                                      token.user_presentable_description().c_str());
                }
                break;
            }
        }
    }
}

/// This is a 'generic' parse error when we can't match the top of the stack element.
void parse_ll_t::parse_error_failed_production(struct parse_stack_element_t &stack_elem,
                                               parse_token_t token) {
    fatal_errored = true;
    if (this->should_generate_error_messages) {
        bool done = false;

        // Check for ||.
        if (token.type == parse_token_type_pipe && token.source_start > 0) {
            // Here we wanted a statement and instead got a pipe. See if this is a double pipe: foo
            // || bar. If so, we have a special error message.
            const parse_node_t *prev_pipe = nodes.find_node_matching_source_location(
                parse_token_type_pipe, token.source_start - 1, NULL);
            if (prev_pipe != NULL) {
                // The pipe of the previous job abuts our current token. So we have ||.
                this->parse_error(token, parse_error_double_pipe, ERROR_BAD_OR);
                done = true;
            }
        }

        // Check for &&.
        if (!done && token.type == parse_token_type_background && token.source_start > 0) {
            // Check to see if there was a previous token_background.
            const parse_node_t *prev_background = nodes.find_node_matching_source_location(
                parse_token_type_background, token.source_start - 1, NULL);
            if (prev_background != NULL) {
                // We have &&.
                this->parse_error(token, parse_error_double_background, ERROR_BAD_AND);
                done = true;
            }
        }

        if (!done) {
            const wcstring expected = stack_elem.user_presentable_description();
            this->parse_error_unexpected_token(expected.c_str(), token);
        }
    }
}

void parse_ll_t::report_tokenizer_error(const tok_t &tok) {
    parse_error_code_t parse_error_code;
    switch (tok.error) {
        case TOK_UNTERMINATED_QUOTE: {
            parse_error_code = parse_error_tokenizer_unterminated_quote;
            break;
        }
        case TOK_UNTERMINATED_SUBSHELL: {
            parse_error_code = parse_error_tokenizer_unterminated_subshell;
            break;
        }
        case TOK_UNTERMINATED_SLICE: {
            parse_error_code = parse_error_tokenizer_unterminated_slice;
            break;
        }
        case TOK_UNTERMINATED_ESCAPE: {
            parse_error_code = parse_error_tokenizer_unterminated_escape;
            break;
        }
        case TOK_OTHER:
        default: {
            parse_error_code = parse_error_tokenizer_other;
            break;
        }
    }
    this->parse_error_at_location(tok.offset + tok.error_offset, parse_error_code, L"%ls",
                                  tok.text.c_str());
}

void parse_ll_t::parse_error_unexpected_token(const wchar_t *expected, parse_token_t token) {
    fatal_errored = true;
    if (this->should_generate_error_messages) {
        this->parse_error(token, parse_error_generic, L"Expected %ls, but instead found %ls",
                          expected, token.user_presentable_description().c_str());
    }
}

void parse_ll_t::reset_symbols(enum parse_token_type_t goal) {
    // Add a new goal node, and then reset our symbol list to point at it.
    node_offset_t where = static_cast<node_offset_t>(nodes.size());
    nodes.push_back(parse_node_t(goal));

    symbol_stack.clear();
    symbol_stack.push_back(parse_stack_element_t(goal, where));  // goal token
    this->fatal_errored = false;
}

/// Reset both symbols and nodes.
void parse_ll_t::reset_symbols_and_nodes(enum parse_token_type_t goal) {
    nodes.clear();
    this->reset_symbols(goal);
}

static bool type_is_terminal_type(parse_token_type_t type) {
    switch (type) {
        case parse_token_type_string:
        case parse_token_type_pipe:
        case parse_token_type_redirection:
        case parse_token_type_background:
        case parse_token_type_end:
        case parse_token_type_terminate: {
            return true;
        }
        default: { return false; }
    }
}

bool parse_ll_t::report_error_for_unclosed_block() {
    bool reported_error = false;
    // Unclosed block, for example, 'while true ; '. We want to show the block node that opened it.
    const parse_node_t &top_node = this->node_for_top_symbol();

    // Hacktastic. We want to point at the source location of the block, but our block doesn't have
    // a source range yet - only the terminal tokens do. So get the block statement corresponding to
    // this end command. In general this block may be of a variety of types: if_statement,
    // switch_statement, etc., each with different node structures. But keep descending the first
    // child and eventually you hit a keyword: begin, if, etc. That's the keyword we care about.
    const parse_node_t *end_command = this->nodes.get_parent(top_node, symbol_end_command);
    const parse_node_t *block_node = end_command ? this->nodes.get_parent(*end_command) : NULL;

    if (block_node && block_node->type == symbol_block_statement) {
        // Get the header.
        block_node = this->nodes.get_child(*block_node, 0, symbol_block_header);
        block_node = this->nodes.get_child(*block_node, 0);  // specific statement
    }
    if (block_node != NULL) {
        // block_node is now an if_statement, switch_statement, for_header, while_header,
        // function_header, or begin_header.
        //
        // Hackish: descend down the first node until we reach the bottom. This will be a keyword
        // node like SWITCH, which will have the source range. Ordinarily the source range would be
        // known by the parent node too, but we haven't completed parsing yet, so we haven't yet
        // propagated source ranges.
        const parse_node_t *cursor = block_node;
        while (cursor->child_count > 0) {
            cursor = this->nodes.get_child(*cursor, 0);
            assert(cursor != NULL);
        }
        if (cursor->source_start != NODE_OFFSET_INVALID) {
            const wcstring node_desc = block_type_user_presentable_description(block_node->type);
            this->parse_error_at_location(cursor->source_start, parse_error_generic,
                                          L"Missing end to balance this %ls", node_desc.c_str());
            reported_error = true;
        }
    }
    return reported_error;
}

bool parse_ll_t::top_node_handle_terminal_types(parse_token_t token) {
    PARSE_ASSERT(!symbol_stack.empty());
    PARSE_ASSERT(token.type >= FIRST_PARSE_TOKEN_TYPE);
    bool handled = false;
    parse_stack_element_t &stack_top = symbol_stack.back();
    if (type_is_terminal_type(stack_top.type)) {
        // The top of the stack is terminal. We are going to handle this (because we can't produce
        // from a terminal type).
        handled = true;

        // Now see if we actually matched
        bool matched = false;
        if (stack_top.type == token.type) {
            switch (stack_top.type) {
                case parse_token_type_string: {
                    // We matched if the keywords match, or no keyword was required.
                    matched = (stack_top.keyword == parse_keyword_none ||
                               stack_top.keyword == token.keyword);
                    break;
                }
                default: {
                    // For other types, we only require that the types match.
                    matched = true;
                    break;
                }
            }
        }

        if (matched) {
            // Success. Tell the node that it matched this token, and what its source range is in
            // the parse phase, we only set source ranges for terminal types. We propagate ranges to
            // parent nodes afterwards.
            parse_node_t &node = node_for_top_symbol();
            node.keyword = token.keyword;
            node.source_start = token.source_start;
            node.source_length = token.source_length;
        } else {
            // Failure
            if (stack_top.type == parse_token_type_string &&
                token.type == parse_token_type_string) {
                // Keyword failure. We should unify this with the 'matched' computation above.
                assert(stack_top.keyword != parse_keyword_none &&
                       stack_top.keyword != token.keyword);

                // Check to see which keyword we got which was considered wrong.
                switch (token.keyword) {
                    // Some keywords are only valid in certain contexts. If this cascaded all the
                    // way down through the outermost job_list, it was not in a valid context.
                    case parse_keyword_case:
                    case parse_keyword_end:
                    case parse_keyword_else: {
                        this->parse_error_unbalancing_token(token);
                        break;
                    }
                    case parse_keyword_none: {
                        // This is a random other string (not a keyword).
                        const wcstring expected = keyword_description(stack_top.keyword);
                        this->parse_error(token, parse_error_generic, L"Expected keyword '%ls'",
                                          expected.c_str());
                        break;
                    }
                    default: {
                        // Got a real keyword we can report.
                        const wcstring actual = (token.keyword == parse_keyword_none
                                                     ? token.describe()
                                                     : keyword_description(token.keyword));
                        const wcstring expected = keyword_description(stack_top.keyword);
                        this->parse_error(token, parse_error_generic,
                                          L"Expected keyword '%ls', instead got keyword '%ls'",
                                          expected.c_str(), actual.c_str());
                        break;
                    }
                }
            } else if (stack_top.keyword == parse_keyword_end &&
                       token.type == parse_token_type_terminate &&
                       this->report_error_for_unclosed_block()) {
                // Handled by report_error_for_unclosed_block.
            } else {
                const wcstring expected = stack_top.user_presentable_description();
                this->parse_error_unexpected_token(expected.c_str(), token);
            }
        }

        // We handled the token, so pop the symbol stack.
        symbol_stack.pop_back();
    }
    return handled;
}

void parse_ll_t::accept_tokens(parse_token_t token1, parse_token_t token2) {
    bool logit = false;
    if (logit) {
        fprintf(stderr, "Accept token %ls\n", token1.describe().c_str());
    }
    PARSE_ASSERT(token1.type >= FIRST_PARSE_TOKEN_TYPE);

    bool consumed = false;

    // Handle special types specially. Note that these are the only types that can be pushed if the
    // symbol stack is empty.
    if (token1.type == parse_special_type_parse_error ||
        token1.type == parse_special_type_tokenizer_error ||
        token1.type == parse_special_type_comment) {
        // We set the special node's parent to the top of the stack. This means that we have an
        // asymmetric relationship: the special node has a parent (which is the node we were trying
        // to generate when we encountered the special node), but the parent node does not have the
        // special node as a child. This means for example that parents don't have to worry about
        // tracking any comment nodes, but we can still recover the parent from the comment.
        parse_node_t special_node(token1.type);
        special_node.parent = symbol_stack.back().node_idx;
        special_node.source_start = token1.source_start;
        special_node.source_length = token1.source_length;
        nodes.push_back(special_node);
        consumed = true;

        // Mark special flags.
        if (token1.type == parse_special_type_comment) {
            this->node_for_top_symbol().flags |= parse_node_flag_has_comments;
        }

        // Tokenizer errors are fatal.
        if (token1.type == parse_special_type_tokenizer_error) this->fatal_errored = true;
    }

    while (!consumed && !this->fatal_errored) {
        PARSE_ASSERT(!symbol_stack.empty());

        if (top_node_handle_terminal_types(token1)) {
            if (logit) {
                fprintf(stderr, "Consumed token %ls\n", token1.describe().c_str());
            }
            // consumed = true;
            break;
        }

        // top_node_match_token may indicate an error if our stack is empty.
        if (this->fatal_errored) break;

        // Get the production for the top of the stack.
        parse_stack_element_t &stack_elem = symbol_stack.back();
        parse_node_t &node = nodes.at(stack_elem.node_idx);
        parse_node_tag_t tag = 0;
        const production_t *production =
            production_for_token(stack_elem.type, token1, token2, &tag);
        node.tag = tag;
        if (production == NULL) {
            parse_error_failed_production(stack_elem, token1);
            // The above sets fatal_errored, which ends the loop.
        } else {
            bool is_terminate = (token1.type == parse_token_type_terminate);

            // When a job_list encounters something like 'else', it returns an empty production to
            // return control to the outer block. But if it's unbalanced, then we'll end up with an
            // empty stack! So make sure that doesn't happen. This is the primary mechanism by which
            // we detect e.g. unbalanced end. However, if we get a true terminate token, then we
            // allow (expect) this to empty the stack.
            if (symbol_stack.size() == 1 && production_is_empty(production) && !is_terminate) {
                this->parse_error_unbalancing_token(token1);
                break;
            }

            // Manipulate the symbol stack. Note that stack_elem is invalidated by popping the
            // stack.
            symbol_stack_pop_push_production(production);

            // Expect to not have an empty stack, unless this was the terminate type. Note we may
            // not have an empty stack with the terminate type (i.e. incomplete input).
            assert(is_terminate || !symbol_stack.empty());

            if (symbol_stack.empty()) {
                break;
            }
        }
    }
}

// Given an expanded string, returns any keyword it matches.
static parse_keyword_t keyword_with_name(const wchar_t *name) {
    // Binary search on keyword_map. Start at 1 since 0 is keyword_none.
    parse_keyword_t result = parse_keyword_none;
    size_t left = 1, right = sizeof keyword_map / sizeof *keyword_map;
    while (left < right) {
        size_t mid = left + (right - left) / 2;
        int cmp = wcscmp(name, keyword_map[mid].name);
        if (cmp < 0) {
            right = mid;  // name was smaller than mid
        } else if (cmp > 0) {
            left = mid + 1;  // name was larger than mid
        } else {
            result = keyword_map[mid].keyword;  // found it
            break;
        }
    }
    return result;
}

static bool is_keyword_char(wchar_t c) {
    return (c >= L'a' && c <= L'z') || (c >= L'A' && c <= L'Z') || (c >= L'0' && c <= L'9') ||
           c == L'\'' || c == L'"' || c == L'\\' || c == '\n';
}

/// Given a token, returns the keyword it matches, or parse_keyword_none.
static parse_keyword_t keyword_for_token(token_type tok, const wcstring &token) {
    /* Only strings can be keywords */
    if (tok != TOK_STRING) {
        return parse_keyword_none;
    }

    // If tok_txt is clean (which most are), we can compare it directly. Otherwise we have to expand
    // it. We only expand quotes, and we don't want to do expensive expansions like tilde
    // expansions. So we do our own "cleanliness" check; if we find a character not in our allowed
    // set we know it's not a keyword, and if we never find a quote we don't have to expand! Note
    // that this lowercase set could be shrunk to be just the characters that are in keywords.
    parse_keyword_t result = parse_keyword_none;
    bool needs_expand = false, all_chars_valid = true;
    const wchar_t *tok_txt = token.c_str();
    for (size_t i = 0; tok_txt[i] != L'\0'; i++) {
        wchar_t c = tok_txt[i];
        if (!is_keyword_char(c)) {
            all_chars_valid = false;
            break;
        }
        // If we encounter a quote, we need expansion.
        needs_expand = needs_expand || c == L'"' || c == L'\'' || c == L'\\';
    }

    if (all_chars_valid) {
        // Expand if necessary.
        if (!needs_expand) {
            result = keyword_with_name(tok_txt);
        } else {
            wcstring storage;
            if (unescape_string(tok_txt, &storage, 0)) {
                result = keyword_with_name(storage.c_str());
            }
        }
    }
    return result;
}

/// Placeholder invalid token.
static const parse_token_t kInvalidToken = {
    token_type_invalid, parse_keyword_none, false, false, SOURCE_OFFSET_INVALID, 0};

/// Terminal token.
static const parse_token_t kTerminalToken = {
    parse_token_type_terminate, parse_keyword_none, false, false, SOURCE_OFFSET_INVALID, 0};

static inline bool is_help_argument(const wcstring &txt) { return contains(txt, L"-h", L"--help"); }

/// Return a new parse token, advancing the tokenizer.
static inline parse_token_t next_parse_token(tokenizer_t *tok, tok_t *token) {
    if (!tok->next(token)) {
        return kTerminalToken;
    }

    parse_token_t result;

    // Set the type, keyword, and whether there's a dash prefix. Note that this is quite sketchy,
    // because it ignores quotes. This is the historical behavior. For example, `builtin --names`
    // lists builtins, but `builtin "--names"` attempts to run --names as a command. Amazingly as of
    // this writing (10/12/13) nobody seems to have noticed this. Squint at it really hard and it
    // even starts to look like a feature.
    result.type = parse_token_type_from_tokenizer_token(token->type);
    result.keyword = keyword_for_token(token->type, token->text);
    result.has_dash_prefix = !token->text.empty() && token->text.at(0) == L'-';
    result.is_help_argument = result.has_dash_prefix && is_help_argument(token->text);

    // These assertions are totally bogus. Basically our tokenizer works in size_t but we work in
    // uint32_t to save some space. If we have a source file larger than 4 GB, we'll probably just
    // crash.
    assert(token->offset < SOURCE_OFFSET_INVALID);
    result.source_start = (source_offset_t)token->offset;

    assert(token->length <= SOURCE_OFFSET_INVALID);
    result.source_length = (source_offset_t)token->length;

    return result;
}

bool parse_tree_from_string(const wcstring &str, parse_tree_flags_t parse_flags,
                            parse_node_tree_t *output, parse_error_list_t *errors,
                            parse_token_type_t goal) {
    parse_ll_t parser(goal);
    parser.set_should_generate_error_messages(errors != NULL);

    // Construct the tokenizer.
    tok_flags_t tok_options = 0;
    if (parse_flags & parse_flag_include_comments) tok_options |= TOK_SHOW_COMMENTS;

    if (parse_flags & parse_flag_accept_incomplete_tokens) tok_options |= TOK_ACCEPT_UNFINISHED;

    if (parse_flags & parse_flag_show_blank_lines) tok_options |= TOK_SHOW_BLANK_LINES;

    if (errors == NULL) tok_options |= TOK_SQUASH_ERRORS;

    tokenizer_t tok(str.c_str(), tok_options);

    // We are an LL(2) parser. We pass two tokens at a time. New tokens come in at index 1. Seed our
    // queue with an initial token at index 1.
    parse_token_t queue[2] = {kInvalidToken, kInvalidToken};

    // Loop until we have a terminal token.
    tok_t tokenizer_token;
    for (size_t token_count = 0; queue[0].type != parse_token_type_terminate; token_count++) {
        // Push a new token onto the queue.
        queue[0] = queue[1];
        queue[1] = next_parse_token(&tok, &tokenizer_token);

        // If we are leaving things unterminated, then don't pass parse_token_type_terminate.
        if (queue[0].type == parse_token_type_terminate &&
            (parse_flags & parse_flag_leave_unterminated)) {
            break;
        }

        // Pass these two tokens, unless we're still loading the queue. We know that queue[0] is
        // valid; queue[1] may be invalid.
        if (token_count > 0) {
            parser.accept_tokens(queue[0], queue[1]);
        }

        // Handle tokenizer errors. This is a hack because really the parser should report this for
        // itself; but it has no way of getting the tokenizer message.
        if (queue[1].type == parse_special_type_tokenizer_error) {
            parser.report_tokenizer_error(tokenizer_token);
        }

        // Handle errors.
        if (parser.has_fatal_error()) {
            if (parse_flags & parse_flag_continue_after_error) {
                // Hack. Typically the parse error is due to the first token. However, if it's a
                // tokenizer error, then has_fatal_error was set due to the check above; in that
                // case the second token is what matters.
                size_t error_token_idx = 0;
                if (queue[1].type == parse_special_type_tokenizer_error) {
                    error_token_idx = (queue[1].type == parse_special_type_tokenizer_error ? 1 : 0);
                    token_count = -1;  // so that it will be 0 after incrementing, and our tokenizer
                                       // error will be ignored
                }

                // Mark a special error token, and then keep going.
                const parse_token_t token = {parse_special_type_parse_error,
                                             parse_keyword_none,
                                             false,
                                             false,
                                             queue[error_token_idx].source_start,
                                             queue[error_token_idx].source_length};
                parser.accept_tokens(token, kInvalidToken);
                parser.reset_symbols(goal);
            } else {
                break;  // bail out
            }
        }
    }

    // Teach each node where its source range is.
    parser.determine_node_ranges();

    // Acquire the output from the parser.
    parser.acquire_output(output, errors);

#if 0
    //wcstring result = dump_tree(this->parser->nodes, str);
    //fprintf(stderr, "Tree (%ld nodes):\n%ls", this->parser->nodes.size(), result.c_str());
    fprintf(stderr, "%lu nodes, node size %lu, %lu bytes\n", output->size(), sizeof(parse_node_t), output->size() * sizeof(parse_node_t));
#endif

    // Indicate if we had a fatal error.
    return !parser.has_fatal_error();
}

const parse_node_t *parse_node_tree_t::get_child(const parse_node_t &parent, node_offset_t which,
                                                 parse_token_type_t expected_type) const {
    const parse_node_t *result = NULL;

    // We may get nodes with no children if we had an incomplete parse. Don't consider than an
    // error.
    if (parent.child_count > 0) {
        PARSE_ASSERT(which < parent.child_count);
        node_offset_t child_offset = parent.child_offset(which);
        if (child_offset < this->size()) {
            result = &this->at(child_offset);

            // If we are given an expected type, then the node must be null or that type.
            assert(expected_type == token_type_invalid || expected_type == result->type);
        }
    }

    return result;
}

const parse_node_t &parse_node_tree_t::find_child(const parse_node_t &parent,
                                                  parse_token_type_t type) const {
    for (node_offset_t i = 0; i < parent.child_count; i++) {
        const parse_node_t *child = this->get_child(parent, i);
        if (child->type == type) {
            return *child;
        }
    }
    PARSE_ASSERT(0);
    return *(parse_node_t *)(NULL);  // unreachable
}

const parse_node_t *parse_node_tree_t::get_parent(const parse_node_t &node,
                                                  parse_token_type_t expected_type) const {
    const parse_node_t *result = NULL;
    if (node.parent != NODE_OFFSET_INVALID) {
        PARSE_ASSERT(node.parent < this->size());
        const parse_node_t &parent = this->at(node.parent);
        if (expected_type == token_type_invalid || expected_type == parent.type) {
            // The type matches (or no type was requested).
            result = &parent;
        }
    }
    return result;
}

static void find_nodes_recursive(const parse_node_tree_t &tree, const parse_node_t &parent,
                                 parse_token_type_t type,
                                 parse_node_tree_t::parse_node_list_t *result, size_t max_count) {
    if (result->size() < max_count) {
        if (parent.type == type) result->push_back(&parent);
        for (node_offset_t i = 0; i < parent.child_count; i++) {
            const parse_node_t *child = tree.get_child(parent, i);
            assert(child != NULL);
            find_nodes_recursive(tree, *child, type, result, max_count);
        }
    }
}

parse_node_tree_t::parse_node_list_t parse_node_tree_t::find_nodes(const parse_node_t &parent,
                                                                   parse_token_type_t type,
                                                                   size_t max_count) const {
    parse_node_list_t result;
    find_nodes_recursive(*this, parent, type, &result, max_count);
    return result;
}

/// Return true if the given node has the proposed ancestor as an ancestor (or is itself that
/// ancestor).
static bool node_has_ancestor(const parse_node_tree_t &tree, const parse_node_t &node,
                              const parse_node_t &proposed_ancestor) {
    if (&node == &proposed_ancestor) {
        return true;  // found it
    } else if (node.parent == NODE_OFFSET_INVALID) {
        return false;  // no more parents
    }

    // Recurse to the parent.
    return node_has_ancestor(tree, tree.at(node.parent), proposed_ancestor);
}

const parse_node_t *parse_node_tree_t::find_last_node_of_type(parse_token_type_t type,
                                                              const parse_node_t *parent) const {
    const parse_node_t *result = NULL;
    // Find nodes of the given type in the tree, working backwards.
    size_t idx = this->size();
    while (idx--) {
        const parse_node_t &node = this->at(idx);
        if (node.type == type) {
            // Types match. Check if it has the right parent.
            if (parent == NULL || node_has_ancestor(*this, node, *parent)) {
                // Success
                result = &node;
                break;
            }
        }
    }
    return result;
}

const parse_node_t *parse_node_tree_t::find_node_matching_source_location(
    parse_token_type_t type, size_t source_loc, const parse_node_t *parent) const {
    const parse_node_t *result = NULL;
    // Find nodes of the given type in the tree, working backwards.
    const size_t len = this->size();
    for (size_t idx = 0; idx < len; idx++) {
        const parse_node_t &node = this->at(idx);

        // Types must match.
        if (node.type != type) continue;

        // Must contain source location.
        if (!node.location_in_or_at_end_of_source_range(source_loc)) continue;

        // If a parent is given, it must be an ancestor.
        if (parent != NULL && !node_has_ancestor(*this, node, *parent)) continue;

        // Found it.
        result = &node;
        break;
    }
    return result;
}

bool parse_node_tree_t::argument_list_is_root(const parse_node_t &node) const {
    bool result = true;
    assert(node.type == symbol_argument_list || node.type == symbol_arguments_or_redirections_list);
    const parse_node_t *parent = this->get_parent(node);
    if (parent != NULL) {
        // We have a parent - check to make sure it's not another list!
        result = parent->type != symbol_arguments_or_redirections_list &&
                 parent->type != symbol_argument_list;
    }
    return result;
}

enum parse_statement_decoration_t parse_node_tree_t::decoration_for_plain_statement(
    const parse_node_t &node) const {
    assert(node.type == symbol_plain_statement);
    const parse_node_t *decorated_statement = this->get_parent(node, symbol_decorated_statement);
    parse_node_tag_t tag =
        decorated_statement ? decorated_statement->tag : parse_statement_decoration_none;
    return static_cast<parse_statement_decoration_t>(tag);
}

bool parse_node_tree_t::command_for_plain_statement(const parse_node_t &node, const wcstring &src,
                                                    wcstring *out_cmd) const {
    bool result = false;
    assert(node.type == symbol_plain_statement);
    const parse_node_t *cmd_node = this->get_child(node, 0, parse_token_type_string);
    if (cmd_node != NULL && cmd_node->has_source()) {
        out_cmd->assign(src, cmd_node->source_start, cmd_node->source_length);
        result = true;
    } else {
        out_cmd->clear();
    }
    return result;
}

bool parse_node_tree_t::statement_is_in_pipeline(const parse_node_t &node,
                                                 bool include_first) const {
    // Moderately nasty hack! Walk up our ancestor chain and see if we are in a job_continuation.
    // This checks if we are in the second or greater element in a pipeline; if we are the first
    // element we treat this as false. This accepts a few statement types.
    bool result = false;
    const parse_node_t *ancestor = &node;

    // If we're given a plain statement, try to get its decorated statement parent.
    if (ancestor && ancestor->type == symbol_plain_statement)
        ancestor = this->get_parent(*ancestor, symbol_decorated_statement);
    if (ancestor) ancestor = this->get_parent(*ancestor, symbol_statement);
    if (ancestor) ancestor = this->get_parent(*ancestor);

    if (ancestor) {
        if (ancestor->type == symbol_job_continuation) {
            // Second or more in a pipeline.
            result = true;
        } else if (ancestor->type == symbol_job && include_first) {
            // Check to see if we have a job continuation that's not empty.
            const parse_node_t *continuation =
                this->get_child(*ancestor, 1, symbol_job_continuation);
            result = (continuation != NULL && continuation->child_count > 0);
        }
    }

    return result;
}

enum token_type parse_node_tree_t::type_for_redirection(const parse_node_t &redirection_node,
                                                        const wcstring &src, int *out_fd,
                                                        wcstring *out_target) const {
    assert(redirection_node.type == symbol_redirection);
    enum token_type result = TOK_NONE;
    const parse_node_t *redirection_primitive =
        this->get_child(redirection_node, 0, parse_token_type_redirection);  // like 2>
    const parse_node_t *redirection_target =
        this->get_child(redirection_node, 1, parse_token_type_string);  // like &1 or file path

    if (redirection_primitive != NULL && redirection_primitive->has_source()) {
        result = redirection_type_for_string(redirection_primitive->get_source(src), out_fd);
    }
    if (out_target != NULL) {
        *out_target = redirection_target ? redirection_target->get_source(src) : L"";
    }
    return result;
}

const parse_node_t *parse_node_tree_t::header_node_for_block_statement(
    const parse_node_t &node) const {
    const parse_node_t *result = NULL;
    if (node.type == symbol_block_statement) {
        const parse_node_t *block_header = this->get_child(node, 0, symbol_block_header);
        if (block_header != NULL) {
            result = this->get_child(*block_header, 0);
        }
    }
    return result;
}

parse_node_tree_t::parse_node_list_t parse_node_tree_t::specific_statements_for_job(
    const parse_node_t &job) const {
    assert(job.type == symbol_job);
    parse_node_list_t result;

    // Initial statement (non-specific).
    result.push_back(get_child(job, 0, symbol_statement));

    // Our cursor variable. Walk over the list of continuations.
    const parse_node_t *continuation = get_child(job, 1, symbol_job_continuation);
    while (continuation != NULL && continuation->child_count > 0) {
        result.push_back(get_child(*continuation, 1, symbol_statement));
        continuation = get_child(*continuation, 2, symbol_job_continuation);
    }

    // Result now contains a list of statements. But we want a list of specific statements e.g.
    // symbol_switch_statement. So replace them in-place in the vector.
    for (size_t i = 0; i < result.size(); i++) {
        const parse_node_t *statement = result.at(i);
        assert(statement->type == symbol_statement);
        result.at(i) = this->get_child(*statement, 0);
    }

    return result;
}

parse_node_tree_t::parse_node_list_t parse_node_tree_t::comment_nodes_for_node(
    const parse_node_t &parent) const {
    parse_node_list_t result;
    if (parent.has_comments()) {
        // Walk all our nodes, looking for comment nodes that have the given node as a parent.
        for (size_t i = 0; i < this->size(); i++) {
            const parse_node_t &potential_comment = this->at(i);
            if (potential_comment.type == parse_special_type_comment &&
                this->get_parent(potential_comment) == &parent) {
                result.push_back(&potential_comment);
            }
        }
    }
    return result;
}

enum parse_bool_statement_type_t parse_node_tree_t::statement_boolean_type(
    const parse_node_t &node) {
    assert(node.type == symbol_boolean_statement);
    return static_cast<parse_bool_statement_type_t>(node.tag);
}

bool parse_node_tree_t::job_should_be_backgrounded(const parse_node_t &job) const {
    assert(job.type == symbol_job);
    const parse_node_t *opt_background = get_child(job, 2, symbol_optional_background);
    bool result = opt_background != NULL && opt_background->tag == parse_background;
    return result;
}

const parse_node_t *parse_node_tree_t::next_node_in_node_list(
    const parse_node_t &node_list, parse_token_type_t entry_type,
    const parse_node_t **out_list_tail) const {
    parse_token_type_t list_type = node_list.type;

    // Paranoia - it doesn't make sense for a list type to contain itself.
    assert(list_type != entry_type);

    const parse_node_t *list_cursor = &node_list;
    const parse_node_t *list_entry = NULL;

    // Loop while we don't have an item but do have a list. Note that some nodes may contain
    // nothing; e.g. job_list contains blank lines as a production.
    while (list_entry == NULL && list_cursor != NULL) {
        const parse_node_t *next_cursor = NULL;

        // Walk through the children.
        for (node_offset_t i = 0; i < list_cursor->child_count; i++) {
            const parse_node_t *child = this->get_child(*list_cursor, i);
            if (child->type == entry_type) {
                // This is the list entry.
                list_entry = child;
            } else if (child->type == list_type) {
                // This is the next in the list.
                next_cursor = child;
            }
        }
        // Go to the next entry, even if it's NULL.
        list_cursor = next_cursor;
    }

    // Return what we got.
    assert(list_cursor == NULL || list_cursor->type == list_type);
    assert(list_entry == NULL || list_entry->type == entry_type);
    if (out_list_tail != NULL) *out_list_tail = list_cursor;
    return list_entry;
}