rucc_pp/

directive.rs

//! The directive engine: translation phase 4 over one file's preprocessing tokens.
//!
//! Design: `spec/05-preprocessor.md` section 5.4.
//!
//! A directive is a line whose first token is `#`. That is the whole of the recognition rule,
//! and the two halves of it are both load bearing: `#` has to be first on the line, and the
//! line is what the lexer says it is after splices and comments have been resolved, which is
//! why `x /*\n*/ #define F 1` really does define `F`.
//!
//! The part that is easy to get wrong is skipped regions. Inside `#if 0` a line beginning with
//! `#` still has to be recognised well enough to keep the conditional nesting balanced, and it
//! must not be diagnosed for anything else. Real code puts prose, unbalanced quotes and future
//! syntax inside `#if 0`, and a preprocessor that reports errors from there is unusable. So
//! skipping looks at the directive name and nothing else, and only the seven conditional
//! directives mean anything while it is going on.

use std::collections::HashMap;
use std::path::{Path, PathBuf};

use rucc_base::{Interner, Symbol};
use rucc_diag::{Diagnostic, FileId, SourceMapFull, Span};
use rucc_gnu::Kind;
use rucc_lex::{Options, PpToken, PpTokenKind, Punct, TokenFlags, tokenize};
use rucc_session::{Found, IncludeForm};
use rucc_target::TargetInfo;

use crate::cond;
use crate::embed;
use crate::expand::Expander;
use crate::include::{
    Context, Frame, Header, Reader, directory_of, header_from_token, header_from_tokens, spelling,
};
use crate::macros::{Builtin, MacroTable, parse_define};
use crate::predef::{BUILT_IN, COMMAND_LINE, Predef, built_in, command_line};
use crate::token::Tok;

/// Why a file that has already been read does not need reading again.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Guard {
    /// `#pragma once`, so the file is read once however many times it is named.
    Once,
    /// The whole file is wrapped in `#ifndef NAME`, and `NAME` is now defined, so reading it
    /// again would produce nothing at all. This is the multiple-include optimization, and on
    /// a real code base it is the difference between reading a header once and reading it a
    /// few hundred times.
    Macro(Symbol),
}

/// How far through the file the guard shape has been recognised.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Scan {
    /// Nothing has been seen yet, so the next line may open the guard.
    Start,
    /// Inside the conditional the file opened with.
    Inside(Symbol),
    /// The conditional closed and the file has to end here for the shape to hold.
    Closed(Symbol),
    /// Something else was seen, so this file has no guard.
    No,
}

/// One `#if` and everything hanging off it.
#[derive(Debug)]
struct Cond {
    /// Where the `#if` was written, so an unterminated one can point at it.
    span: Span,
    /// Whether tokens in the branch currently open are kept. Already accounts for whether the
    /// enclosing region was live, so [`Preprocessor::live`] only has to look at the top.
    live: bool,
    /// Whether some branch of this chain has been taken. A later `#elif` is not evaluated once
    /// this is set, which is what makes `#elif 1/0` after a taken branch legal.
    taken: bool,
    /// Whether the enclosing region was live.
    enclosing_live: bool,
    /// Whether `#else` has been seen, so a second one is an error.
    seen_else: bool,
}

/// A `#line` directive, as read and as applied to the source map.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct LineDirective {
    /// Where the directive is.
    pub span: Span,
    /// The line number the next line is to be called.
    pub line: u32,
    /// The file name the following lines are to be called, if one was given.
    pub file: Option<Symbol>,
    /// How many tokens had been emitted when this was read, which is where it sits in the
    /// stream.
    ///
    /// A position is not enough to say that. A file included from here is added to the source
    /// map after this file, so its bytes come after every byte of this one, and the token
    /// after the `#include` is at a lower position than the tokens of the header. `-E` has to
    /// write a marker for a `#line` where the directive was written rather than where its
    /// bytes are, and this is what says where that is.
    pub at: usize,
}

/// Translation phase 4 over one file.
///
/// Holds the macro table and the conditional stack, so a single instance processes a whole
/// translation unit and the definitions a header makes are visible after it.
#[derive(Debug, Default)]
pub struct Preprocessor {
    macros: MacroTable,
    expander: Expander,
    diagnostics: Vec<Diagnostic>,
    conds: Vec<Cond>,
    lines: Vec<LineDirective>,
    /// The files currently open, innermost last. Empty between runs.
    stack: Vec<Frame>,
    /// The files a line marker said were entered, innermost last, by the name in force when it
    /// said so. This is the nesting a `2` flag claims to be leaving, and it is kept apart from
    /// `stack` because a marker set describes a nesting the real files never had.
    markers: Vec<String>,
    /// Files that do not need reading again, and why. Keyed by what the file system calls the
    /// file rather than by the name an include used, so that a header reached two ways is one
    /// entry here.
    seen: HashMap<PathBuf, Guard>,
}

impl Preprocessor {
    /// A preprocessor with an empty macro table.
    pub fn new() -> Preprocessor {
        Preprocessor::default()
    }

    /// The macros defined so far.
    pub fn macros(&self) -> &MacroTable {
        &self.macros
    }

    /// The macro table, for the driver to seed with `-D` and the predefined set.
    pub fn macros_mut(&mut self) -> &mut MacroTable {
        &mut self.macros
    }

    /// Everything reported so far.
    pub fn diagnostics(&self) -> &[Diagnostic] {
        &self.diagnostics
    }

    /// Takes the diagnostics, leaving the preprocessor able to carry on.
    pub fn take_diagnostics(&mut self) -> Vec<Diagnostic> {
        std::mem::take(&mut self.diagnostics)
    }

    /// The `#line` directives seen, in the order they appeared.
    ///
    /// Each one is also applied, to the source map, as it is read. This is the record of them
    /// rather than the mechanism: what a caller wants it for is reporting on the directives
    /// themselves, and asking the map is how to find out where anything is.
    pub fn line_directives(&self) -> &[LineDirective] {
        &self.lines
    }

    /// Defines the predefined macro set, and then `-D` and `-U` from the command line.
    ///
    /// Called before [`Preprocessor::run`], because a predefined macro is a macro like any
    /// other by the time the source file is read. The set arrives as two synthetic files
    /// rather than as a list of definitions, so a diagnostic about one of them points at
    /// `<built-in>` or `<command-line>` the way GCC's does, and so that `-dM` has something
    /// to print. The reasoning is in `crate::predef`.
    ///
    /// # Errors
    ///
    /// When the source map has no room left for the two synthetic files.
    pub fn predefine(
        &mut self,
        target: &TargetInfo,
        opts: &Predef,
        cx: &mut Context<'_>,
    ) -> Result<(), SourceMapFull> {
        let names = Names::new(cx.interner);
        let file = self.synthetic(BUILT_IN, built_in(target, opts), cx, &names)?;
        // The macros that cannot be written as a `#define` line, because what they stand for
        // depends on where they are used. They go in after the generated file and before the
        // command line, so that `-U__FILE__` takes one away the way it takes any other away.
        // The origin is the start of `<built-in>`, which is where a warning about redefining
        // one points, and which is the truthful answer to where they came from.
        let start = cx.sources.file(file).start;
        for (spelling, builtin) in Builtin::ALL {
            let name = cx.interner.intern(spelling);
            self.macros.define_builtin(name, builtin, Span::new(start, start));
        }
        let text = command_line(opts);
        if !text.is_empty() {
            self.synthetic(COMMAND_LINE, text, cx, &names)?;
        }
        Ok(())
    }

    /// Reads a file the compiler wrote rather than one the user did.
    fn synthetic(
        &mut self,
        name: &str,
        text: String,
        cx: &mut Context<'_>,
        names: &Names,
    ) -> Result<FileId, SourceMapFull> {
        let file = cx.sources.add(name, text.into_bytes())?;
        let mut out = Vec::new();
        // A frame, so that the guard scan and the include depth see the same shape they see
        // for a real file. There is no directory, because `#include "x.h"` written in a
        // synthetic file has nowhere of its own to look.
        let path = PathBuf::from(name);
        let id = cx.fs.identity(&path);
        self.stack.push(Frame { at: Span::DUMMY, path, id, dir: None, next: 0 });
        self.process(file, &mut out, cx, names);
        self.stack.clear();
        debug_assert!(out.is_empty(), "{name} is directives only and produces no tokens");
        Ok(file)
    }

    /// Runs phase 4 over `file` and everything it includes.
    ///
    /// The result is the tokens that survived the conditionals, with macros expanded. Nothing
    /// is thrown away silently: an unterminated `#if` and a stray `#endif` are both reported.
    pub fn run(&mut self, file: FileId, cx: &mut Context<'_>) -> Vec<Tok> {
        let names = Names::new(cx.interner);
        let mut out = Vec::new();
        let name = cx.sources.file(file).name.clone();
        let dir = directory_of(&name);
        // The file named on the command line was not found through the search path, so an
        // `#include_next` written in it starts at the top rather than partway down.
        let path = PathBuf::from(name);
        let id = cx.fs.identity(&path);
        self.stack.push(Frame { at: Span::DUMMY, path, id, dir, next: 0 });
        self.process(file, &mut out, cx, &names);
        self.stack.clear();
        out
    }

    /// Reads one file, appending what survives to `out`.
    fn process(&mut self, file: FileId, out: &mut Vec<Tok>, cx: &mut Context<'_>, names: &Names) {
        // The bytes are taken out of the map by sharing rather than by borrowing, because the
        // rest of this function needs the map back to add an included file to it.
        let bytes = cx.sources.file(file).shared_bytes();
        let start = cx.sources.file(file).start;
        let mut reader = Reader::new(bytes.as_slice(), start, cx.lex);
        let depth_on_entry = self.conds.len();
        // Consecutive text lines are expanded as one run rather than line by line, because a
        // function-like macro invocation may span lines. It may not span a directive, which is
        // undefined behaviour, so a directive is where the run ends.
        let mut text: Vec<Tok> = Vec::new();
        let mut body: Vec<PpToken> = Vec::new();
        let mut scan = Scan::Start;

        loop {
            let was_live = self.live();
            let first = reader.next(cx.interner);
            if first.is_eof() {
                break;
            }
            if is_directive(first) {
                self.flush(&mut text, out, cx, names);
                body.clear();
                let name_tok = reader.next(cx.interner);
                // The null directive. A line of just `#` is legal and does nothing, and there
                // is a surprising amount of it in real headers as a visual separator.
                if name_tok.is_eof() || name_tok.flags.has(TokenFlags::START_OF_LINE) {
                    reader.put_back(name_tok);
                    continue;
                }
                body.push(name_tok);
                // The header name has to be scanned here or not at all: `<stdio.h>` and a run
                // of comparisons are the same bytes, and once the line has been scanned the
                // other way the difference is gone. Not in a skipped region, because scanning
                // one there can report an unterminated name that nobody asked about.
                if was_live && is_include(ident_of(&name_tok), names) {
                    if let Some(header) = reader.header_name(cx.interner) {
                        body.push(header);
                    }
                }
                reader.line(cx.interner, &mut body);
                let opens =
                    matches!(scan, Scan::Start).then(|| guard_opener(&body, names)).flatten();
                self.directive(&body, first.span, out, cx, names);
                scan = match scan {
                    // The guard has to be the first line of the file and it has to open a
                    // conditional, which is why the depth is checked after the dispatch
                    // rather than the directive name being trusted on its own.
                    Scan::Start => match opens {
                        Some(name) if self.conds.len() == depth_on_entry + 1 => Scan::Inside(name),
                        _ => Scan::No,
                    },
                    Scan::Inside(name) if self.conds.len() == depth_on_entry => Scan::Closed(name),
                    Scan::Inside(name) => Scan::Inside(name),
                    Scan::Closed(_) | Scan::No => Scan::No,
                };
            } else {
                body.clear();
                reader.line(cx.interner, &mut body);
                if self.live() {
                    // A run of text lines is expanded in one go, and a `_Pragma` is a directive
                    // wearing an operator's clothes: `pop_macro` changes what the names after it
                    // mean. So a line that spells one is expanded on its own, or the line after a
                    // pop would go through the expander in the same batch as the line before it
                    // and would still see the definition the pop was there to undo.
                    let operator = ident_of(&first) == Some(names.pragma_op)
                        || body.iter().any(|t| ident_of(t) == Some(names.pragma_op));
                    if operator {
                        self.flush(&mut text, out, cx, names);
                    }
                    text.push(Tok::new(first));
                    text.extend(body.iter().copied().map(Tok::new));
                    if operator {
                        self.flush(&mut text, out, cx, names);
                    }
                }
                // A token outside the guard is a token that would be produced twice.
                if !matches!(scan, Scan::Inside(_)) {
                    scan = Scan::No;
                }
            }
            // What the lexer complained about while reading that line. A skipped region keeps
            // its complaints to itself, for the same reason it keeps its directives to itself.
            let complaints = reader.take_diagnostics();
            if was_live || self.live() {
                self.diagnostics.extend(complaints);
            }
        }
        self.flush(&mut text, out, cx, names);
        self.diagnostics.extend(reader.take_diagnostics());

        // The guard only counts if the macro really did get defined. A file that opens with
        // `#ifndef X` and never defines `X` is a file that has to be read again.
        if let Scan::Closed(name) = scan {
            if self.macros.is_defined(name) {
                if let Some(frame) = self.stack.last() {
                    self.seen.entry(frame.id.clone()).or_insert(Guard::Macro(name));
                }
            }
        }

        // A file may not close a conditional it did not open. GCC reports this at the `#if`,
        // which is the line the user has to go and look at.
        for cond in self.conds.drain(depth_on_entry..) {
            self.diagnostics
                .push(Diagnostic::error("unterminated `#if`", cond.span).with_code("E0330"));
        }
    }

    /// Whether tokens are currently being kept.
    fn live(&self) -> bool {
        self.conds.last().is_none_or(|c| c.live)
    }

    /// Expands a run of text lines and appends it to the output.
    fn flush(
        &mut self,
        text: &mut Vec<Tok>,
        out: &mut Vec<Tok>,
        cx: &mut Context<'_>,
        names: &Names,
    ) {
        if text.is_empty() {
            return;
        }
        let taken = std::mem::take(text);
        let expanded = self.expander.expand_toks(taken, &self.macros, cx.interner, cx.sources);
        self.diagnostics.append(&mut self.expander.take_diagnostics());
        // To GCC and clang the `__has_*` family are builtin macros rather than something the
        // conditional parser knows about, so they answer in ordinary text too. After expansion
        // and not before it, because a macro is allowed to expand to a call of one and because
        // the operand is expanded first, which is what happens on a `#if` line as well.
        let expanded = self.resolve_has(expanded, cx, names, Pass::Text);
        self.pragma_operator(expanded, out, cx.interner, names);
    }

    /// Dispatches one directive. `body` is the line after the `#`.
    fn directive(
        &mut self,
        body: &[PpToken],
        hash: Span,
        out: &mut Vec<Tok>,
        cx: &mut Context<'_>,
        names: &Names,
    ) {
        let Some(first) = body.first().copied() else {
            return;
        };
        let name = ident_of(&first);
        let rest = &body[1..];

        // Conditionals are handled whether or not the region is live, because the nesting has
        // to stay balanced through a skipped block.
        if name == Some(names.r#if) {
            let value = self.live() && self.eval(rest, hash, cx, names);
            self.open(hash, value);
            return;
        }
        if name == Some(names.ifdef) || name == Some(names.ifndef) {
            let want = name == Some(names.ifdef);
            let value = self.live() && self.defined_check(rest, hash, want, names);
            self.open(hash, value);
            return;
        }
        if name == Some(names.elif) || name == Some(names.elifdef) || name == Some(names.elifndef) {
            self.elif(name, rest, hash, cx, names);
            return;
        }
        if name == Some(names.r#else) {
            self.branch_else(rest, hash);
            return;
        }
        if name == Some(names.endif) {
            self.endif(rest, hash);
            return;
        }
        if !self.live() {
            // Everything else inside a skipped region is text, not a directive. `#error` in
            // the branch that was not taken must not fire, and `# 42 "f.c"` from another
            // preprocessor must not be diagnosed.
            return;
        }

        // A `#` and a number is a GNU line marker rather than a directive whose name happens to
        // be missing, and it is what `-E` output is full of, so it is answered before anything
        // asks what the directive is called.
        if name.is_none() && decimal(&first, cx.interner).is_some() {
            self.line_marker(body, hash, out.len(), cx);
            return;
        }

        let interner = &mut *cx.interner;
        if name == Some(names.define) {
            let (def, diagnostics) = parse_define(rest, interner);
            self.diagnostics.extend(diagnostics);
            if let Some(def) = def {
                if let Some(problem) = self.macros.define(def, interner) {
                    self.diagnostics.push(problem);
                }
            }
        } else if name == Some(names.undef) {
            self.undef(rest, hash, interner);
        } else if name == Some(names.error) || name == Some(names.warning) {
            self.message(rest, hash, name == Some(names.error), interner);
        } else if name == Some(names.line) {
            self.line(rest, hash, out.len(), cx);
        } else if name == Some(names.pragma) {
            // `#pragma once` is answered here and does not reach the output, because it is a
            // question about the file rather than something a later phase can act on.
            // Everything else is passed through unchanged, which is what `-E` has to print
            // and what a later phase looking for `#pragma pack` will read. Inventing an
            // internal representation now, with no consumer, would only be a thing to
            // migrate later.
            if rest.len() == 1 && ident_of(&rest[0]) == Some(names.once) {
                self.pragma_once(rest[0].span);
            } else if !self.macro_stack_pragma(rest, hash, interner, names) {
                self.pass_through(body, hash, out);
            }
        } else if name == Some(names.include) || name == Some(names.include_next) {
            self.include(rest, hash, name == Some(names.include_next), out, cx, names);
        } else if name == Some(names.embed) {
            self.embed(rest, hash, out, cx);
        } else {
            self.diagnostics.push(
                Diagnostic::error("invalid preprocessing directive", first.span).with_code("E0332"),
            );
        }
    }

    /// Answers `#pragma push_macro("X")` and `#pragma pop_macro("X")`, or says it is not one.
    ///
    /// These are the two pragmas that act on the macro table, so this phase is the only one that
    /// can answer them, and like `#pragma once` they do not reach the output: gcc consumes them
    /// and a later phase given one could not do anything with it. That is what clang's
    /// `__clang_cuda_complex_builtins.h` needs, which pushes `__DEVICE__`, redefines it for the
    /// file and pops it at the end.
    ///
    /// The `GCC` namespaced spelling is deliberately not accepted, because gcc does not accept
    /// it either: `#pragma GCC push_macro("X")` is passed through and does nothing, and matching
    /// that matters more than the spelling looking symmetric with the pragmas that do take it.
    fn macro_stack_pragma(
        &mut self,
        rest: &[PpToken],
        at: Span,
        interner: &mut Interner,
        names: &Names,
    ) -> bool {
        let which = match rest.first().and_then(ident_of) {
            Some(name) if name == names.push_macro => names.push_macro,
            Some(name) if name == names.pop_macro => names.pop_macro,
            _ => return false,
        };
        let word = if which == names.push_macro { "push_macro" } else { "pop_macro" };
        // Once the word is recognised the line is one of these whatever follows it, so a line
        // that is not the shape is an error rather than something to pass through. gcc says the
        // same thing, and warns about anything after the closing parenthesis the way it warns
        // about anything after any other directive.
        let [_, open, text, close, extra @ ..] = rest else {
            self.invalid_pragma(word, at);
            return true;
        };
        if open.punct() != Some(Punct::LParen)
            || text.kind != PpTokenKind::StringLit
            || close.punct() != Some(Punct::RParen)
        {
            self.invalid_pragma(word, at);
            return true;
        }
        self.extra_tokens(extra, "#pragma");
        // A string that does not spell one identifier names no macro, and gcc neither complains
        // about it nor does anything with it. `push_macro("a b")` is quietly nothing, which is
        // worth matching rather than improving on: a header that has one is a header that has
        // been building against gcc for years.
        let Some(name) = identifier_in(*text, interner) else {
            return true;
        };
        if which == names.push_macro {
            self.macros.push_macro(name);
        } else {
            self.macros.pop_macro(name);
        }
        true
    }

    fn invalid_pragma(&mut self, word: &str, at: Span) {
        self.diagnostics.push(
            Diagnostic::error(format!("invalid `#pragma {word}` directive"), at).with_code("E0672"),
        );
    }

    /// Records that the file currently being read asked to be read only once.
    fn pragma_once(&mut self, at: Span) {
        // In the main file this is worth saying something about, since the file the user named
        // is not one anything includes and the line usually means the user thought it was a
        // header. It is still applied, because a file that includes itself is exactly where the
        // line does work in a main file, and GCC both warns and applies it.
        if self.stack.len() <= 1 {
            self.diagnostics.push(
                Diagnostic::warning("`#pragma once` in the main file", at).with_code("W0332"),
            );
        }
        if let Some(frame) = self.stack.last() {
            self.seen.insert(frame.id.clone(), Guard::Once);
        }
    }

    /// Whether a file has already given everything it has to give.
    fn skip(&self, id: &Path) -> bool {
        match self.seen.get(id) {
            Some(Guard::Once) => true,
            Some(Guard::Macro(name)) => self.macros.is_defined(*name),
            None => false,
        }
    }

    /// Copies a directive line into the output, `#` included.
    fn pass_through(&mut self, body: &[PpToken], hash: Span, out: &mut Vec<Tok>) {
        let _ = self;
        out.push(Tok::synthetic(
            PpTokenKind::Punct(Punct::Hash),
            None,
            TokenFlags::START_OF_LINE,
            hash,
        ));
        // The space between the hash and the word comes off, so that a directive written
        // `#  pragma` inside a nest of conditionals, which is how glibc indents them, prints
        // back as `#pragma`. gcc does the same, and the rest of the line keeps the spacing it
        // was written with.
        for (at, token) in body.iter().copied().enumerate() {
            let mut token = Tok::new(token);
            if at == 0 {
                token.flags = token.flags.without(TokenFlags::LEADING_SPACE);
            }
            out.push(token);
        }
    }

    /// Resolves an `#include` or `#include_next` and reads what it names.
    fn include(
        &mut self,
        rest: &[PpToken],
        hash: Span,
        is_next: bool,
        out: &mut Vec<Tok>,
        cx: &mut Context<'_>,
        names: &Names,
    ) {
        let Some(header) = self.header_of(rest, hash, cx) else {
            return;
        };
        let (form, relative_to, from) = self.where_to_look(&header, is_next, cx);
        let found = cx.search.resolve(cx.fs, &header.name, form, relative_to.as_deref(), from);
        let Some(found) = found else {
            let tried = cx.search.tried(&header.name, form, relative_to.as_deref(), from);
            // Two ways to have looked nowhere. An absolute name is opened and not searched
            // for, and a search path with nothing on it has nowhere to look. Saying the
            // first when it was the second sends the reader after a path that is not there.
            let where_looked = if tried.is_empty() && Path::new(&header.name).is_absolute() {
                "the name is an absolute path, so the search path was not used".to_owned()
            } else if tried.is_empty() {
                "the include search path is empty".to_owned()
            } else {
                let list: Vec<String> =
                    tried.iter().map(|d| d.to_string_lossy().into_owned()).collect();
                format!("searched: {}", list.join(", "))
            };
            self.diagnostics.push(
                Diagnostic::error(format!("`{}` file not found", header.name), hash)
                    .with_code("E0341")
                    .note(where_looked, hash),
            );
            return;
        };
        // The multiple-include optimization. A file wrapped in an include guard whose macro
        // is now defined, or one that asked for `#pragma once`, would produce nothing, so it
        // is not opened at all. On a real code base this is the difference between reading a
        // header once and reading it a few hundred times.
        let id = cx.fs.identity(&found.path);
        if self.skip(&id) {
            return;
        }
        if self.stack.len() >= cx.max_include_depth as usize {
            let mut diagnostic =
                Diagnostic::error("`#include` nested too deeply", hash).with_code("E0342").note(
                    "a header that includes itself with no include guard is the usual cause",
                    hash,
                );
            if let Some(outer) = self.stack.first().filter(|f| !f.at.is_dummy()) {
                diagnostic = diagnostic.note("the outermost include is here", outer.at);
            }
            self.diagnostics.push(diagnostic);
            return;
        }
        let added = cx.sources.add_shared(found.name.clone(), found.bytes.clone(), Some(hash));
        let file = match added {
            Ok(file) => file,
            Err(full) => {
                self.diagnostics.push(Diagnostic::error(full.to_string(), hash).with_code("E0344"));
                return;
            }
        };
        self.stack.push(Frame {
            at: hash,
            dir: found.path.parent().map(Path::to_path_buf),
            id,
            path: found.path,
            next: found.next,
        });
        self.process(file, out, cx, names);
        self.stack.pop();
    }

    /// Reads an `#embed` and puts the bytes of what it names into the output.
    fn embed(&mut self, rest: &[PpToken], hash: Span, out: &mut Vec<Tok>, cx: &mut Context<'_>) {
        let Some((header, params)) = self.embed_line(rest, hash, cx) else {
            return;
        };
        let Some(found) = self.find(&header, false, cx) else {
            self.diagnostics.push(
                Diagnostic::error(format!("`{}` resource not found", header.name), hash)
                    .with_code("E0341")
                    .note("an `#embed` resource is looked for on the include path", hash),
            );
            return;
        };
        // The bytes are not added to the source map. Nothing will ever point a diagnostic
        // into the middle of a PNG, and adding a few megabytes of binary to the map so that
        // it can be sliced for a caret line nobody will print is the kind of cost that only
        // shows up on the projects this directive exists for.
        embed::tokens(found.bytes.as_slice(), &params, hash, cx.interner, out);
    }

    /// Splits an `#embed` line into the resource it names and the parameters after it.
    fn embed_line(
        &mut self,
        rest: &[PpToken],
        hash: Span,
        cx: &mut Context<'_>,
    ) -> Option<(Header, embed::Params)> {
        if rest.is_empty() {
            self.bad_header(hash);
            return None;
        }
        let line: Vec<Tok> = rest.iter().copied().map(Tok::new).collect();
        // A name the lexer already made a header name of is not expanded, exactly as with
        // `#include`. A computed one has the whole line expanded, parameters included, which
        // is a compromise: the end of the name cannot be found without expanding, and the
        // parameter names would have to be found before expanding to protect them. A macro
        // called `limit` in scope at an `#embed` is not a thing worth splitting the pass for.
        let line = if line[0].kind == PpTokenKind::HeaderName {
            line
        } else {
            let expanded = self.expander.expand_toks(line, &self.macros, cx.interner, cx.sources);
            self.diagnostics.append(&mut self.expander.take_diagnostics());
            expanded
        };
        let Some(used) = embed::header_length(&line) else {
            self.bad_header(line.first().map_or(hash, |t| t.report_span()));
            return None;
        };
        let header = if line[0].kind == PpTokenKind::HeaderName {
            header_from_token(spelling(line[0], cx.interner))
        } else {
            let spellings: Vec<&str> =
                line[..used].iter().map(|t| spelling(*t, cx.interner)).collect();
            header_from_tokens(&spellings)
        };
        let Some(header) = header else {
            self.bad_header(line[0].report_span());
            return None;
        };
        let params = self.embed_params(&line[used..], hash, cx)?;
        Some((header, params))
    }

    /// The parameter list of an `#embed`, or of the `__has_embed` that asks the same question.
    fn embed_params(
        &mut self,
        line: &[Tok],
        at: Span,
        cx: &mut Context<'_>,
    ) -> Option<embed::Params> {
        let Preprocessor { expander, macros, diagnostics, .. } = self;
        let sources = &mut *cx.sources;
        let mut expand = |toks: Vec<Tok>, interner: &mut Interner| {
            expander.expand_toks(toks, macros, interner, sources)
        };
        let params = embed::parse(line, at, cx.interner, diagnostics, &mut expand);
        self.diagnostics.append(&mut self.expander.take_diagnostics());
        params
    }

    /// Where a header written in the file being read is looked for.
    ///
    /// `#include_next` continues from the directory after the one the current file came from,
    /// which is what glibc and the kernel use to wrap a system header with one of the same
    /// name. It never looks next to the current file, because that directory is not on the
    /// path and there would be nothing to continue past.
    ///
    /// `__has_include` has to ask the same question the directive would, so both go through
    /// here. A header that answers yes and then fails to be found is the one outcome that
    /// would make the operator useless.
    fn where_to_look(
        &self,
        header: &Header,
        is_next: bool,
        cx: &Context<'_>,
    ) -> (IncludeForm, Option<PathBuf>, usize) {
        let form = if header.angled { IncludeForm::Angled } else { IncludeForm::Quoted };
        let frame = self.stack.last();
        let from = if is_next {
            frame.map_or(0, |f| f.next).max(cx.search.start(form))
        } else {
            cx.search.start(form)
        };
        let relative_to = if is_next { None } else { frame.and_then(|f| f.dir.clone()) };
        (form, relative_to, from)
    }

    /// Whether a header is there, which is all `__has_include` asks.
    fn find(&self, header: &Header, is_next: bool, cx: &Context<'_>) -> Option<Found> {
        let (form, relative_to, from) = self.where_to_look(header, is_next, cx);
        cx.search.resolve(cx.fs, &header.name, form, relative_to.as_deref(), from)
    }

    /// The header name an include directive names, however it spelled it.
    fn header_of(&mut self, rest: &[PpToken], hash: Span, cx: &mut Context<'_>) -> Option<Header> {
        if let Some(first) = rest.first().copied() {
            if first.kind == PpTokenKind::HeaderName {
                let text = first.value.map_or("", |v| cx.interner.resolve(v));
                let header = header_from_token(text);
                if header.is_none() {
                    self.bad_header(first.span);
                }
                self.extra_tokens(&rest[1..], "#include");
                return header;
            }
        }
        // The computed include, `#include MACRO`. The line is macro expanded and then has to
        // look like a header name, which is the one place in the language where the spelling
        // of a token matters after expansion.
        if rest.is_empty() {
            self.bad_header(hash);
            return None;
        }
        let line: Vec<Tok> = rest.iter().copied().map(Tok::new).collect();
        let expanded = self.expander.expand_toks(line, &self.macros, cx.interner, cx.sources);
        self.diagnostics.append(&mut self.expander.take_diagnostics());
        let spellings: Vec<&str> = expanded.iter().map(|t| spelling(*t, cx.interner)).collect();
        let header = header_from_tokens(&spellings);
        if header.is_none() {
            let at = expanded.first().map_or(hash, |t| t.report_span());
            self.bad_header(at);
        }
        header
    }

    /// The diagnostic for a `__has_*` operator whose operand is not an identifier.
    fn bad_operand(&mut self, tok: Tok, at: Span, interner: &Interner) {
        self.diagnostics.push(
            Diagnostic::error(
                format!("expected an identifier as the operand of `{}`", spelling(tok, interner)),
                at,
            )
            .with_code("E0345"),
        );
    }

    fn bad_header(&mut self, at: Span) {
        self.diagnostics.push(
            Diagnostic::error("expected a file name in `<>` or `\"\"`", at).with_code("E0343"),
        );
    }

    /// Pushes a conditional whose first branch is or is not taken.
    fn open(&mut self, span: Span, value: bool) {
        let enclosing_live = self.live();
        self.conds.push(Cond {
            span,
            live: enclosing_live && value,
            taken: value,
            enclosing_live,
            seen_else: false,
        });
    }

    fn elif(
        &mut self,
        name: Option<Symbol>,
        rest: &[PpToken],
        hash: Span,
        cx: &mut Context<'_>,
        names: &Names,
    ) {
        let Some(top) = self.conds.last() else {
            self.stray("elif", hash);
            return;
        };
        if top.seen_else {
            self.diagnostics
                .push(Diagnostic::error("`#elif` after `#else`", hash).with_code("E0333"));
            return;
        }
        // Read what is needed before evaluating, because evaluation borrows the whole
        // preprocessor to report into.
        let (enclosing_live, already_taken) = (top.enclosing_live, top.taken);
        let consider = enclosing_live && !already_taken;
        let value = if !consider {
            false
        } else if name == Some(names.elif) {
            self.eval(rest, hash, cx, names)
        } else {
            self.defined_check(rest, hash, name == Some(names.elifdef), names)
        };
        let top = self.conds.last_mut().expect("checked above and nothing popped");
        top.live = consider && value;
        top.taken = already_taken || value;
    }

    fn branch_else(&mut self, rest: &[PpToken], hash: Span) {
        let Some(top) = self.conds.last_mut() else {
            self.stray("else", hash);
            return;
        };
        if top.seen_else {
            self.diagnostics.push(Diagnostic::error("a second `#else`", hash).with_code("E0333"));
            return;
        }
        top.live = top.enclosing_live && !top.taken;
        top.taken = true;
        top.seen_else = true;
        let enclosing_live = top.enclosing_live;
        if enclosing_live {
            self.extra_tokens(rest, "#else");
        }
    }

    fn endif(&mut self, rest: &[PpToken], hash: Span) {
        if self.conds.pop().is_none() {
            self.stray("endif", hash);
            return;
        }
        if self.live() {
            self.extra_tokens(rest, "#endif");
        }
    }

    fn stray(&mut self, what: &str, hash: Span) {
        self.diagnostics
            .push(Diagnostic::error(format!("`#{what}` without `#if`"), hash).with_code("E0334"));
    }

    /// Warns about tokens after a directive that takes none.
    ///
    /// A warning rather than an error, because `#endif FOO` as a hand written comment is
    /// everywhere in code written before `//` was portable.
    fn extra_tokens(&mut self, rest: &[PpToken], what: &str) {
        if let Some(first) = rest.first() {
            self.diagnostics.push(
                Diagnostic::warning(format!("extra tokens after `{what}`"), first.span)
                    .with_code("W0330"),
            );
        }
    }

    /// Evaluates a `#if` or `#elif` expression.
    fn eval(&mut self, rest: &[PpToken], hash: Span, cx: &mut Context<'_>, names: &Names) -> bool {
        let line: Vec<Tok> = rest.iter().copied().map(Tok::new).collect();
        // `defined X` is resolved before expansion, so that `#if defined FOO` does not depend
        // on what `FOO` expands to. It is resolved again afterwards because a macro that
        // expands to `defined(X)` is undefined behaviour that GCC supports and headers use.
        // It goes first of all because `defined(__has_include)` is a question about the
        // operator rather than a use of it.
        let line = self.resolve_defined(line, cx.interner, names);
        // `__has_include` is resolved before expansion too, and for a stronger reason: its
        // operand is a header name, so expanding `<linux/version.h>` would turn `linux` into
        // `1` on a target where that macro is predefined. The rest of the family take an
        // identifier that GCC does expand, so they wait until afterwards.
        let line = self.resolve_has(line, cx, names, Pass::Headers);
        let line = self.expander.expand_toks(line, &self.macros, cx.interner, cx.sources);
        self.diagnostics.append(&mut self.expander.take_diagnostics());
        let line = self.resolve_defined(line, cx.interner, names);
        let line = self.resolve_has(line, cx, names, Pass::Rest);
        cond::evaluate(&line, cx.interner, &mut self.diagnostics, hash)
    }

    /// Replaces `__has_include(<x.h>)` and the rest of the family with what they answer.
    ///
    /// `pass` says which of the three positions is asking, and each of them answers a
    /// different part of the family. See [`Pass`].
    fn resolve_has(
        &mut self,
        line: Vec<Tok>,
        cx: &mut Context<'_>,
        names: &Names,
        pass: Pass,
    ) -> Vec<Tok> {
        if !line.iter().any(|t| t.ident().is_some_and(|n| names.has.op(n).is_some())) {
            return line;
        }
        let mut out = Vec::with_capacity(line.len());
        let mut at = 0;
        while at < line.len() {
            let tok = line[at];
            let op = tok.ident().and_then(|n| names.has.op(n));
            let Some(op) = op.filter(|op| pass.answers(*op)) else {
                if pass == Pass::Text && op.is_some_and(Op::is_header) {
                    self.outside_a_directive(tok, cx);
                }
                out.push(tok);
                at += 1;
                continue;
            };
            let Some((operand, after)) = arguments(&line, at + 1) else {
                // Reported in the pass after expansion and not in the one before it, because
                // the operator is still there for that pass to find and one mistake is one
                // diagnostic.
                if pass != Pass::Headers {
                    self.diagnostics.push(
                        Diagnostic::error(
                            format!("expected `(` after `{}`", spelling(tok, cx.interner)),
                            tok.report_span(),
                        )
                        .with_code("E0345"),
                    );
                }
                out.push(tok);
                at += 1;
                continue;
            };
            at = after;
            // A number rather than a flag, because `__has_c_attribute` answers with the value
            // the standard gives the attribute and a header compares that against a date.
            let value = self.ask(op, operand, tok, cx);
            let sym = cx.interner.intern(&value.to_string());
            out.push(Tok::synthetic(PpTokenKind::Number, Some(sym), tok.flags, tok.report_span()));
        }
        out
    }

    /// Refuses one of the header operators used in ordinary text.
    ///
    /// Their operand is a header name, and outside a directive the line was scanned as
    /// ordinary tokens, so `<stdio.h>` arrived as a chain of comparisons that no longer says
    /// which of the two it was meant to be. GCC and clang both refuse it for that reason, and
    /// a program that wants the answer in text can put the operator in a `#if` and define a
    /// macro from it, which is what every header that needs one does anyway.
    fn outside_a_directive(&mut self, tok: Tok, cx: &Context<'_>) {
        self.diagnostics.push(
            Diagnostic::error(
                format!(
                    "`{}` used outside of a preprocessing directive",
                    spelling(tok, cx.interner)
                ),
                tok.report_span(),
            )
            .with_code("E0350"),
        );
    }

    /// What one `__has_*` operator answers for one operand.
    fn ask(&mut self, op: Op, operand: &[Tok], tok: Tok, cx: &mut Context<'_>) -> u32 {
        let at = operand.first().map_or(tok.report_span(), |t| t.report_span());
        match op {
            Op::Include | Op::IncludeNext => {
                let spellings: Vec<&str> =
                    operand.iter().map(|t| spelling(*t, cx.interner)).collect();
                let Some(header) = header_from_tokens(&spellings) else {
                    self.bad_header(at);
                    return 0;
                };
                u32::from(self.find(&header, op == Op::IncludeNext, cx).is_some())
            }
            Op::Embed => {
                // Three answers, and the third one is the reason the operator exists. A
                // resource that is present but empty cannot be told from one that is missing
                // by a yes or no, and the two need different code: the empty one still needs
                // its `if_empty` written, the missing one needs a fallback.
                let Some(used) = embed::header_length(operand) else {
                    self.bad_header(at);
                    return 0;
                };
                let header = if operand[0].kind == PpTokenKind::HeaderName {
                    header_from_token(spelling(operand[0], cx.interner))
                } else {
                    let spellings: Vec<&str> =
                        operand[..used].iter().map(|t| spelling(*t, cx.interner)).collect();
                    header_from_tokens(&spellings)
                };
                let Some(header) = header else {
                    self.bad_header(at);
                    return 0;
                };
                // The parameters are read even though only `limit` and `gnu::offset` can
                // change the answer, because a misspelled parameter is the same mistake here
                // as it is on the directive and finding it only on the directive would mean
                // the guard passes and the embed it guards fails.
                let Some(params) = self.embed_params(&operand[used..], at, cx) else {
                    return 0;
                };
                match self.find(&header, false, cx) {
                    None => 0,
                    Some(found) => {
                        let taken = params.taken(found.bytes.as_slice().len() as u64);
                        if taken == 0 { 2 } else { 1 }
                    }
                }
            }
            Op::BuildingModule => {
                if attribute_name(operand, cx.interner).is_none() {
                    self.bad_operand(tok, at, cx.interner);
                }
                // Clang answers this with one only while it is compiling the module named
                // here, and we do not have modules, so the answer is always no. It is
                // recognised rather than left alone because clang's own `stddef.h` asks it
                // inside an `#if`, and an unknown identifier there leaves the parenthesised
                // operand behind as extra tokens, which fails the whole line rather than the
                // one operator.
                0
            }
            Op::Table(kind) => {
                let Some(name) = attribute_name(operand, cx.interner) else {
                    self.bad_operand(tok, at, cx.interner);
                    return 0;
                };
                match kind {
                    Kind::Attribute => rucc_gnu::has_attribute(name),
                    Kind::CAttribute => rucc_gnu::has_c_attribute(name),
                    Kind::Builtin => rucc_gnu::has_builtin(name),
                    Kind::Feature => rucc_gnu::has_feature(name),
                    Kind::Extension => rucc_gnu::has_extension(name),
                }
            }
        }
    }

    /// Replaces `defined X` and `defined(X)` with `1` or `0`.
    fn resolve_defined(
        &mut self,
        line: Vec<Tok>,
        interner: &mut Interner,
        names: &Names,
    ) -> Vec<Tok> {
        if !line.iter().any(|t| t.ident() == Some(names.defined)) {
            return line;
        }
        let mut out = Vec::with_capacity(line.len());
        let mut at = 0;
        while at < line.len() {
            let tok = line[at];
            if tok.ident() != Some(names.defined) {
                out.push(tok);
                at += 1;
                continue;
            }
            let parenthesised = line.get(at + 1).is_some_and(|t| t.is(Punct::LParen));
            let name_at = if parenthesised { at + 2 } else { at + 1 };
            let name = line.get(name_at).and_then(|t| t.ident());
            let Some(name) = name else {
                self.diagnostics.push(
                    Diagnostic::error("`defined` without a macro name", tok.report_span())
                        .with_code("E0335"),
                );
                out.push(tok);
                at += 1;
                continue;
            };
            at = name_at + 1;
            if parenthesised {
                if line.get(at).is_some_and(|t| t.is(Punct::RParen)) {
                    at += 1;
                } else {
                    self.diagnostics.push(
                        Diagnostic::error("expected `)` after `defined`", tok.report_span())
                            .with_code("E0335"),
                    );
                }
            }
            // A header asks `#ifdef __has_include` before using it, because the operator is
            // newer than some of the compilers it has to build under. It is not a macro, but
            // the question being asked is whether the name means something, and it does.
            let value = self.macros.is_defined(name) || names.has.op(name).is_some();
            out.push(number(value, tok.flags, tok.report_span(), interner));
        }
        out
    }

    /// The body of `#ifdef`, `#ifndef`, `#elifdef` and `#elifndef`.
    fn defined_check(
        &mut self,
        rest: &[PpToken],
        hash: Span,
        want_defined: bool,
        names: &Names,
    ) -> bool {
        let Some(name) = rest.first().and_then(ident_of) else {
            self.diagnostics.push(
                Diagnostic::error("expected a macro name", rest.first().map_or(hash, |t| t.span))
                    .with_code("E0336"),
            );
            return false;
        };
        self.extra_tokens(&rest[1..], if want_defined { "#ifdef" } else { "#ifndef" });
        let defined = self.macros.is_defined(name) || names.has.op(name).is_some();
        defined == want_defined
    }

    fn undef(&mut self, rest: &[PpToken], hash: Span, interner: &Interner) {
        let Some(name) = rest.first().and_then(ident_of) else {
            self.diagnostics.push(
                Diagnostic::error("expected a macro name", rest.first().map_or(hash, |t| t.span))
                    .with_code("E0336"),
            );
            return;
        };
        // The standard reserves these and GCC refuses to let them go, because code that
        // undefines `__FILE__` and then uses it is broken in a way that is very hard to see.
        let text = interner.resolve(name);
        if text == "defined" || text.starts_with("__STDC_") {
            self.diagnostics.push(
                Diagnostic::error(format!("`{text}` cannot be undefined"), rest[0].span)
                    .with_code("E0337"),
            );
            return;
        }
        self.macros.undef(name);
        self.extra_tokens(&rest[1..], "#undef");
    }

    /// `#error` and `#warning`. The message is the rest of the line, spelled back.
    fn message(&mut self, rest: &[PpToken], hash: Span, fatal: bool, interner: &Interner) {
        let text = spell_line(rest, interner);
        let span = rest.first().map_or(hash, |t| t.span.to(last_span(rest)));
        let diag = if fatal {
            Diagnostic::error(text, span).with_code("E0338")
        } else {
            Diagnostic::warning(text, span).with_code("W0331")
        };
        self.diagnostics.push(diag);
    }

    /// `#line 42` and `#line 42 "file.c"`.
    ///
    /// The argument is macro expanded first, which is the one place a directive other than
    /// `#if` does that, and which exists because `#line __LINE__ + 1` is real code.
    fn line(&mut self, rest: &[PpToken], hash: Span, at: usize, cx: &mut Context<'_>) {
        let line: Vec<Tok> = rest.iter().copied().map(Tok::new).collect();
        let line = self.expander.expand_toks(line, &self.macros, cx.interner, cx.sources);
        self.diagnostics.append(&mut self.expander.take_diagnostics());
        let interner = &mut *cx.interner;

        let number_text = line
            .first()
            .filter(|t| t.kind == PpTokenKind::Number)
            .and_then(|t| t.value)
            .map(|v| interner.resolve(v));
        let Some(parsed) = number_text.and_then(|t| t.parse::<u64>().ok()) else {
            self.diagnostics.push(
                Diagnostic::error(
                    "`#line` needs a decimal line number",
                    line.first().map_or(hash, |t| t.report_span()),
                )
                .with_code("E0339"),
            );
            return;
        };
        // 2147483647 is the largest line number the standard requires support for, and it is
        // also where every other compiler stops, so matching that keeps diagnostics comparable.
        if parsed == 0 || parsed > 2_147_483_647 {
            self.diagnostics.push(
                Diagnostic::error("`#line` number is out of range", line[0].report_span())
                    .with_code("E0339"),
            );
            return;
        }

        let mut file = None;
        if let Some(second) = line.get(1) {
            if second.kind == PpTokenKind::StringLit {
                file = second.value;
            } else {
                self.diagnostics.push(
                    Diagnostic::error(
                        "`#line` file name must be a string literal",
                        second.report_span(),
                    )
                    .with_code("E0339"),
                );
                return;
            }
        }
        if let Some(extra) = line.get(2) {
            self.diagnostics.push(
                Diagnostic::warning("extra tokens after `#line`", extra.report_span())
                    .with_code("W0330"),
            );
        }
        #[expect(
            clippy::cast_possible_truncation,
            reason = "the range check above keeps this inside i32, let alone u32"
        )]
        let number = parsed as u32;
        self.lines.push(LineDirective { span: hash, line: number, file, at });
        let name = file.map(|v| destringize(cx.interner.resolve(v)));
        cx.sources.set_presumed(hash.lo, number, name);
    }

    /// A GNU line marker: `# 42`, `# 42 "file.c"`, and either of those with flags after it.
    ///
    /// This is the form `-E` writes, so a preprocessed file handed back to the compiler is full
    /// of them, and a compiler that cannot read its own output is not much of a compiler. The
    /// directive is a `#` and a number rather than a `#` and a name, which is why it arrives
    /// here having failed to be anything else.
    ///
    /// It is `#line` with three differences. Nothing is macro expanded, because the tokens came
    /// from a preprocessor rather than from a person. Zero is a line number, since a generator
    /// counting from zero is allowed to say so and `#line 0` is an error only because somebody
    /// wrote it. And there may be flags: `1` for entering a file, `2` for returning from one,
    /// `3` for a system header and `4` for one whose contents are `extern "C"`. The last two
    /// say nothing this phase acts on. The first two are the nesting, and a `2` that does not
    /// name the file it claims to be returning to is ignored with a warning rather than
    /// applied, which is what gcc does and is the only honest answer to a marker set that does
    /// not describe a nesting anything was ever in.
    fn line_marker(&mut self, body: &[PpToken], hash: Span, at: usize, cx: &mut Context<'_>) {
        let Some(number) = decimal(&body[0], cx.interner) else { return };
        let mut rest = &body[1..];
        let mut file = None;
        if let Some(first) = rest.first().filter(|t| t.kind == PpTokenKind::StringLit) {
            file = first.value;
            rest = &rest[1..];
        }

        let (mut entering, mut leaving) = (false, false);
        for flag in rest {
            match decimal(flag, cx.interner) {
                Some(1) => entering = true,
                Some(2) => leaving = true,
                Some(3 | 4) => {}
                _ => {
                    let text = spell_line(std::slice::from_ref(flag), cx.interner);
                    self.diagnostics.push(
                        Diagnostic::error(
                            format!("invalid flag `{text}` in line directive"),
                            flag.span,
                        )
                        .with_code("E0339"),
                    );
                    return;
                }
            }
        }

        let name = file.map(|v| destringize(cx.interner.resolve(v)));
        if leaving {
            if let Some(name) = &name {
                if !self.leave_marker(name) {
                    self.diagnostics.push(
                        Diagnostic::warning(
                            format!("file `{name}` linemarker ignored due to incorrect nesting"),
                            last_span(body),
                        )
                        .with_code("W0330"),
                    );
                    return;
                }
            } else {
                self.markers.pop();
            }
        }
        if entering {
            let here = cx.sources.presumed(hash.lo).map(|loc| loc.name.to_owned());
            self.markers.push(here.unwrap_or_default());
        }

        self.lines.push(LineDirective { span: hash, line: number, file, at });
        cx.sources.set_presumed(hash.lo, number, name);
    }

    /// Unwinds the marker nesting to `name`, saying whether it was in it at all.
    ///
    /// GCC asks whether the file being returned to is the one directly outside, and this asks
    /// whether it is anywhere outside, because a marker set is generated and a generator that
    /// leaves out a return marker is common. Every `-E` that writes markers where its tokens
    /// are rather than where its files change writes such a set, this compiler's own included,
    /// since a header that contributes no tokens between two `#include` lines never gets a
    /// marker of its own. Answering that with a warning on every file would make the warning
    /// noise, and the nesting it describes is still enough to say what a `2` means.
    ///
    /// A name in neither the markers nor the real include stack is the one that is refused.
    /// That is the marker set that describes a nesting nothing was ever in, and gcc refuses it
    /// too, so `# 200 "xyz" 2` written at the top of a file is a warning in both compilers.
    fn leave_marker(&mut self, name: &str) -> bool {
        if let Some(at) = self.markers.iter().rposition(|outer| outer == name) {
            self.markers.truncate(at);
            return true;
        }
        // A marker set may begin partway down a real nesting it did not open, which is what a
        // header full of them looks like when it is included rather than compiled on its own.
        let found = self.stack.iter().rev().skip(1).any(|f| f.path.as_os_str() == name);
        if found {
            self.markers.clear();
        }
        found
    }

    /// Applies the `_Pragma` operator to an expanded run and appends the result.
    ///
    /// `_Pragma("x")` is a pragma written as an expression, which is what makes a pragma
    /// usable from inside a macro. It is handled after expansion because the string it takes
    /// is very often produced by one.
    fn pragma_operator(
        &mut self,
        expanded: Vec<Tok>,
        out: &mut Vec<Tok>,
        interner: &mut Interner,
        names: &Names,
    ) {
        if !expanded.iter().any(|t| t.ident() == Some(names.pragma_op)) {
            out.extend(expanded);
            return;
        }
        let mut at = 0;
        // A pragma is a line, so whatever comes after one has to start a line, even when the
        // source wrote `_Pragma("x") int y;` all on one. Without this the `int` would read as
        // part of the pragma to anything that takes the line as the unit, which is what the
        // phase that turns these into tokens does.
        let mut ends_a_line = false;
        while at < expanded.len() {
            let mut tok = expanded[at];
            if tok.ident() != Some(names.pragma_op) {
                if ends_a_line {
                    tok.flags = tok.flags.with(TokenFlags::START_OF_LINE);
                    ends_a_line = false;
                }
                out.push(tok);
                at += 1;
                continue;
            }
            let open = expanded.get(at + 1).is_some_and(|t| t.is(Punct::LParen));
            let text = expanded.get(at + 2).filter(|t| t.kind == PpTokenKind::StringLit);
            let close = expanded.get(at + 3).is_some_and(|t| t.is(Punct::RParen));
            let (Some(text), true, true) = (text, open, close) else {
                self.diagnostics.push(
                    Diagnostic::error("`_Pragma` takes a single string literal", tok.report_span())
                        .with_code("E0340"),
                );
                out.push(tok);
                at += 1;
                continue;
            };
            let literal = text.value.map(|v| interner.resolve(v)).unwrap_or_default();
            let body = destringize(literal);
            self.emit_pragma(&body, tok, out, interner, names);
            ends_a_line = true;
            at += 4;
        }
    }

    /// Turns destringized `_Pragma` text into the `# pragma ...` tokens a later phase reads.
    fn emit_pragma(
        &mut self,
        body: &str,
        at: Tok,
        out: &mut Vec<Tok>,
        interner: &mut Interner,
        names: &Names,
    ) {
        let span = at.report_span();
        let (tokens, diagnostics) = tokenize(body.as_bytes(), 0, Options::new(), interner);
        // The text came out of a string literal, so a span into it would point at bytes the
        // user cannot see. Every token reports at the `_Pragma` instead.
        self.diagnostics.extend(
            diagnostics
                .into_iter()
                .map(|d| Diagnostic::new(d.severity, d.message, span).with_code("E0340")),
        );
        let tokens: Vec<PpToken> = tokens.into_iter().filter(|t| !t.is_eof()).collect();
        // `_Pragma("push_macro(\"X\")")` is the same pragma written the other way, and the two
        // spellings have to mean the same thing because a macro that wants to save a name has no
        // other way to say it: a `#pragma` line cannot come out of a macro body.
        if self.macro_stack_pragma(&tokens, span, interner, names) {
            return;
        }
        out.push(Tok::synthetic(
            PpTokenKind::Punct(Punct::Hash),
            None,
            TokenFlags::START_OF_LINE,
            span,
        ));
        out.push(Tok::synthetic(PpTokenKind::Ident, Some(names.pragma), TokenFlags::EMPTY, span));
        // The tokens keep the spacing they were written with inside the string, so
        // `_Pragma("pack(push)")` prints back as `pack(push)` rather than `pack ( push )`.
        // Only the first one is forced apart, from the `pragma` before it.
        for (at, t) in tokens.into_iter().enumerate() {
            // Start of line has to come off: the line is the `#pragma` we just emitted, not
            // the inside of the string these came from.
            let spaced = at == 0 || t.flags.has(TokenFlags::LEADING_SPACE);
            let flags = if spaced {
                TokenFlags::EMPTY.with(TokenFlags::LEADING_SPACE)
            } else {
                TokenFlags::EMPTY
            };
            out.push(Tok::synthetic(t.kind, t.value, flags, span));
        }
    }
}

/// The macro a file's opening line guards the whole file with, if the line has that shape.
///
/// `#ifndef NAME` and both spellings of `#if !defined NAME`, which between them are what
/// every header in glibc, musl and the kernel is wrapped in.
fn guard_opener(body: &[PpToken], names: &Names) -> Option<Symbol> {
    let name = ident_of(body.first()?)?;
    let rest = &body[1..];
    if name == names.ifndef {
        let [only] = rest else {
            return None;
        };
        return ident_of(only);
    }
    if name != names.r#if {
        return None;
    }
    let [bang, defined, tail @ ..] = rest else {
        return None;
    };
    if bang.punct() != Some(Punct::Bang) || ident_of(defined) != Some(names.defined) {
        return None;
    }
    match tail {
        [only] => ident_of(only),
        [open, only, close]
            if open.punct() == Some(Punct::LParen) && close.punct() == Some(Punct::RParen) =>
        {
            ident_of(only)
        }
        _ => None,
    }
}

/// Whether a directive name is one that may be followed by a header name.
fn is_include(name: Option<Symbol>, names: &Names) -> bool {
    name == Some(names.include) || name == Some(names.include_next) || name == Some(names.embed)
}

/// Whether this token opens a directive line.
fn is_directive(tok: PpToken) -> bool {
    tok.flags.has(TokenFlags::START_OF_LINE) && tok.punct() == Some(Punct::Hash)
}

fn ident_of(tok: &PpToken) -> Option<Symbol> {
    match tok.kind {
        PpTokenKind::Ident => tok.value,
        _ => None,
    }
}

fn last_span(tokens: &[PpToken]) -> Span {
    tokens.last().map_or(Span::DUMMY, |t| t.span)
}

/// The value of `tok` when it is a plain decimal number a line can be called.
///
/// A preprocessing number is a wider thing than a number: `1.5`, `0x10` and `1f` are all one,
/// and none of them is a line. Nothing but digits is accepted, so `# 1.5` stays what it was
/// before this existed, which is a directive nobody recognises.
fn decimal(tok: &PpToken, interner: &Interner) -> Option<u32> {
    if tok.kind != PpTokenKind::Number {
        return None;
    }
    let text = interner.resolve(tok.value?);
    if text.is_empty() || !text.bytes().all(|b| b.is_ascii_digit()) {
        return None;
    }
    // 2147483647 is the largest line number the standard requires support for, and it is also
    // where every other compiler stops, so matching that keeps diagnostics comparable.
    text.parse::<u32>().ok().filter(|n| *n <= 2_147_483_647)
}

/// A synthetic `1` or `0`.
fn number(value: bool, flags: TokenFlags, span: Span, interner: &mut Interner) -> Tok {
    let sym = interner.intern(if value { "1" } else { "0" });
    Tok::synthetic(PpTokenKind::Number, Some(sym), flags, span)
}

/// Spells a directive's tokens back for an `#error` message.
fn spell_line(tokens: &[PpToken], interner: &Interner) -> String {
    let mut out = String::new();
    for (index, tok) in tokens.iter().enumerate() {
        if index > 0 && tok.flags.has(TokenFlags::LEADING_SPACE) {
            out.push(' ');
        }
        match tok.value {
            Some(sym) => out.push_str(interner.resolve(sym)),
            None => {
                if let Some(p) = tok.punct() {
                    out.push_str(p.as_str());
                }
            }
        }
    }
    out
}

/// The single identifier a string literal spells, if that is all it spells.
///
/// The name a `push_macro` saves lives inside a string, so it is destringized and lexed rather
/// than read off a token. Anything that is not exactly one identifier names no macro.
fn identifier_in(text: PpToken, interner: &mut Interner) -> Option<Symbol> {
    let literal = interner.resolve(text.value?).to_string();
    let (tokens, _) = tokenize(destringize(&literal).as_bytes(), 0, Options::new(), interner);
    let mut real = tokens.into_iter().filter(|t| !t.is_eof());
    let first = real.next()?;
    if first.kind != PpTokenKind::Ident || real.next().is_some() {
        return None;
    }
    first.value
}

/// Undoes what `#` would have done, per C23 6.10.10.
///
/// The `L` or `u8` prefix and the quotes come off, then `\"` becomes `"` and `\\` becomes `\`.
/// No other escape is touched, because no other escape was introduced.
fn destringize(literal: &str) -> String {
    let body = literal
        .trim_start_matches(['L', 'u', 'U', '8'])
        .strip_prefix('"')
        .and_then(|s| s.strip_suffix('"'))
        .unwrap_or(literal);
    let mut out = String::with_capacity(body.len());
    let mut chars = body.chars();
    while let Some(c) = chars.next() {
        if c != '\\' {
            out.push(c);
            continue;
        }
        match chars.next() {
            Some('"') => out.push('"'),
            Some('\\') => out.push('\\'),
            Some(other) => {
                out.push('\\');
                out.push(other);
            }
            None => out.push('\\'),
        }
    }
    out
}

/// The parenthesised operand of a `__has_*` operator, and where the line carries on.
///
/// `None` when the next token is not `(`, which is the only shape the operators take. Nesting
/// is counted rather than stopping at the first `)`, so that `__has_include(HEADER(x))` after
/// expansion still finds the end of its own operand.
fn arguments(line: &[Tok], at: usize) -> Option<(&[Tok], usize)> {
    if !line.get(at)?.is(Punct::LParen) {
        return None;
    }
    let mut depth = 1u32;
    let mut end = at + 1;
    while end < line.len() {
        if line[end].is(Punct::LParen) {
            depth += 1;
        } else if line[end].is(Punct::RParen) {
            depth -= 1;
            if depth == 0 {
                return Some((&line[at + 1..end], end + 1));
            }
        }
        end += 1;
    }
    None
}

/// The name `__has_attribute` and its relatives are asked about.
///
/// A bare identifier, or the scoped form `gnu::always_inline` that C23 gives the attributes
/// that came from GCC. The scope is dropped: `__has_c_attribute(gnu::x)` and
/// `__has_attribute(x)` are the same question, and the matrix has one row for it.
fn attribute_name<'i>(operand: &[Tok], interner: &'i Interner) -> Option<&'i str> {
    let name = match operand {
        [one] => one,
        [_, scope, name] if scope.is(Punct::ColonColon) => name,
        _ => return None,
    };
    name.ident().map(|sym| interner.resolve(sym))
}

/// Which of the three sweeps over a line is resolving the `__has_*` operators.
///
/// A `#if` line is swept twice, once either side of macro expansion, because the two halves of
/// the family disagree about whether their operand may be expanded. A text line is swept once,
/// after expansion, and the half whose operand is a header name is refused there rather than
/// answered.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Pass {
    /// Before expansion on a directive line, where only the header operators are answered.
    Headers,
    /// After expansion on a directive line, where everything left over is answered. That
    /// includes a header operator, which reaches here when a macro expanded to one.
    Rest,
    /// After expansion on a text line, where everything but the header operators is answered.
    Text,
}

impl Pass {
    /// Whether this sweep is the one that answers `op`.
    fn answers(self, op: Op) -> bool {
        match self {
            Pass::Headers => op.is_header(),
            Pass::Rest => true,
            Pass::Text => !op.is_header(),
        }
    }
}

/// Which `__has_*` operator a name is, and what answers it.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Op {
    /// `__has_include`, answered by looking for the header.
    Include,
    /// `__has_include_next`, the same question from further down the search path.
    IncludeNext,
    /// `__has_embed`, which answers with three values rather than two because a resource that
    /// exists and is empty is a case the program has to be able to tell apart.
    Embed,
    /// `__building_module`, which is always no because there are no modules.
    BuildingModule,
    /// The rest of the family, answered out of the matrix in `rucc-gnu`.
    Table(Kind),
}

impl Op {
    /// Whether the operand is a header name, which must not be macro expanded.
    fn is_header(self) -> bool {
        matches!(self, Op::Include | Op::IncludeNext | Op::Embed)
    }
}

/// The `__has_*` operators, interned once per file.
///
/// A short array rather than a map: there are nine of them, the comparison is on interned
/// symbols, and it is only reached for a line that mentions one.
struct HasOps {
    ops: [(Symbol, Op); 9],
    /// The lowest and the highest symbol in `ops`.
    ///
    /// Now that text lines are swept too, every identifier in the translation unit is offered
    /// to [`HasOps::op`], so the answer it almost always gives has to be cheap. These nine are
    /// interned before any file is read, so a name out of the source sorts above the range and
    /// one comparison turns it away.
    range: (Symbol, Symbol),
}

impl HasOps {
    fn new(interner: &mut Interner) -> HasOps {
        let ops = [
            (interner.intern("__has_include"), Op::Include),
            (interner.intern("__has_include_next"), Op::IncludeNext),
            (interner.intern("__has_embed"), Op::Embed),
            (interner.intern("__has_attribute"), Op::Table(Kind::Attribute)),
            (interner.intern("__has_c_attribute"), Op::Table(Kind::CAttribute)),
            (interner.intern("__has_builtin"), Op::Table(Kind::Builtin)),
            (interner.intern("__has_feature"), Op::Table(Kind::Feature)),
            (interner.intern("__has_extension"), Op::Table(Kind::Extension)),
            (interner.intern("__building_module"), Op::BuildingModule),
        ];
        let mut range = (ops[0].0, ops[0].0);
        for &(sym, _) in &ops {
            range = (range.0.min(sym), range.1.max(sym));
        }
        HasOps { ops, range }
    }

    /// The operator a name is, if it is one.
    #[inline]
    fn op(&self, name: Symbol) -> Option<Op> {
        if name < self.range.0 || name > self.range.1 {
            return None;
        }
        self.ops.iter().find(|(sym, _)| *sym == name).map(|(_, op)| *op)
    }
}

/// The directive names and the two operators, interned once per file.
///
/// Comparing symbols rather than strings is the point: a directive line is recognised with
/// integer comparisons, and the identifiers were interned during the scan, so there is no
/// string work in the hot path.
struct Names {
    define: Symbol,
    undef: Symbol,
    r#if: Symbol,
    ifdef: Symbol,
    ifndef: Symbol,
    elif: Symbol,
    elifdef: Symbol,
    elifndef: Symbol,
    r#else: Symbol,
    endif: Symbol,
    line: Symbol,
    error: Symbol,
    warning: Symbol,
    pragma: Symbol,
    include: Symbol,
    include_next: Symbol,
    embed: Symbol,
    defined: Symbol,
    once: Symbol,
    push_macro: Symbol,
    pop_macro: Symbol,
    pragma_op: Symbol,
    has: HasOps,
}

impl Names {
    fn new(interner: &mut Interner) -> Names {
        Names {
            define: interner.intern("define"),
            undef: interner.intern("undef"),
            r#if: interner.intern("if"),
            ifdef: interner.intern("ifdef"),
            ifndef: interner.intern("ifndef"),
            elif: interner.intern("elif"),
            elifdef: interner.intern("elifdef"),
            elifndef: interner.intern("elifndef"),
            r#else: interner.intern("else"),
            endif: interner.intern("endif"),
            line: interner.intern("line"),
            error: interner.intern("error"),
            warning: interner.intern("warning"),
            pragma: interner.intern("pragma"),
            include: interner.intern("include"),
            include_next: interner.intern("include_next"),
            embed: interner.intern("embed"),
            defined: interner.intern("defined"),
            once: interner.intern("once"),
            push_macro: interner.intern("push_macro"),
            pop_macro: interner.intern("pop_macro"),
            pragma_op: interner.intern("_Pragma"),
            has: HasOps::new(interner),
        }
    }
}

#[cfg(test)]
mod tests {
    use rucc_diag::{Severity, SourceMap};
    use rucc_session::{MemoryFileSystem, SearchPath};

    use super::*;
    use rucc_session::Std;

    use crate::predef::Timestamp;

    /// A whole file through phase 4, which is what almost every test here wants.
    ///
    /// The main file is always `/main.c`, so a quoted include with no search path set up
    /// finds a header the test put at `/name.h`.
    struct Run {
        interner: Interner,
        sources: SourceMap,
        fs: MemoryFileSystem,
        search: SearchPath,
        pp: Preprocessor,
    }

    impl Run {
        fn new() -> Run {
            Run {
                interner: Interner::new(),
                sources: SourceMap::new(),
                fs: MemoryFileSystem::new(),
                search: SearchPath::new(),
                pp: Preprocessor::new(),
            }
        }

        /// Puts a header where an include can find it.
        fn file(&mut self, path: &str, contents: &str) {
            self.fs.insert(path, contents.as_bytes().to_vec());
        }

        /// Puts a resource where an `#embed` can find it. Bytes rather than text, because the
        /// whole point of the directive is the files that are not text.
        fn bytes(&mut self, path: &str, contents: &[u8]) {
            self.fs.insert(path, contents.to_vec());
        }

        /// Adds a directory to the `-I` part of the search path.
        fn dir(&mut self, path: &str) {
            self.search.push_bracket(path);
        }

        /// Defines the predefined set for a target, as the driver does before reading input.
        fn predefine(&mut self, triple: &str, opts: &Predef) {
            let target = TargetInfo::new(triple.parse().expect("a supported triple"));
            let mut cx =
                Context::new(&mut self.interner, &mut self.sources, &self.fs, &self.search);
            self.pp.predefine(&target, opts, &mut cx).expect("the map has room");
        }

        /// The surviving tokens, spelled with one space wherever they were separated.
        fn go(&mut self, src: &str) -> String {
            self.go_named("/main.c", src)
        }

        /// The surviving tokens themselves, for a test about a flag rather than a spelling.
        fn raw(&mut self, src: &str) -> Vec<Tok> {
            let file = self.sources.add("/main.c", src.as_bytes().to_vec()).expect("room");
            let mut cx =
                Context::new(&mut self.interner, &mut self.sources, &self.fs, &self.search);
            self.pp.run(file, &mut cx)
        }

        /// The same, for a test that cares what the main file is called.
        fn go_named(&mut self, path: &str, src: &str) -> String {
            let file = self.sources.add(path, src.as_bytes().to_vec()).expect("the map has room");
            let out = {
                let mut cx =
                    Context::new(&mut self.interner, &mut self.sources, &self.fs, &self.search);
                self.pp.run(file, &mut cx)
            };
            let mut text = String::new();
            for (at, tok) in out.iter().enumerate() {
                let spaced = tok.flags.has(TokenFlags::LEADING_SPACE)
                    || tok.flags.has(TokenFlags::START_OF_LINE);
                if at > 0 && spaced {
                    text.push(' ');
                }
                match tok.kind {
                    PpTokenKind::Punct(p) => text.push_str(p.as_str()),
                    _ => text.push_str(
                        self.interner.resolve(tok.value.expect("every non-punctuator interns")),
                    ),
                }
            }
            text
        }

        /// How many files were opened, main file included. A header that the guard
        /// optimization skipped never reaches the source map, so this is what says whether
        /// it was really skipped rather than read and thrown away.
        fn files(&self) -> usize {
            self.sources.files().len()
        }

        fn messages(&mut self) -> Vec<String> {
            self.pp.take_diagnostics().into_iter().map(|d| d.message).collect()
        }

        fn severities(&mut self) -> Vec<Severity> {
            self.pp.diagnostics().iter().map(|d| d.severity).collect()
        }
    }

    fn clean(src: &str) -> String {
        let mut run = Run::new();
        let text = run.go(src);
        assert!(run.messages().is_empty(), "expected no diagnostics from {src:?}");
        text
    }

    #[test]
    fn a_taken_branch_is_kept_and_the_other_is_not() {
        assert_eq!(clean("#if 1\nyes\n#else\nno\n#endif\n"), "yes");
        assert_eq!(clean("#if 0\nyes\n#else\nno\n#endif\n"), "no");
    }

    #[test]
    fn ifdef_and_ifndef_ask_the_macro_table() {
        assert_eq!(clean("#define F 1\n#ifdef F\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#ifdef F\nyes\n#endif\n"), "");
        assert_eq!(clean("#ifndef F\nyes\n#endif\n"), "yes");
        // C23 spells the two of them as `#elifdef` and `#elifndef` as well.
        assert_eq!(clean("#define F 1\n#if 0\na\n#elifdef F\nb\n#endif\n"), "b");
        assert_eq!(clean("#if 0\na\n#elifndef F\nb\n#endif\n"), "b");
    }

    #[test]
    fn only_the_first_true_branch_of_a_chain_is_taken() {
        assert_eq!(clean("#if 0\na\n#elif 1\nb\n#elif 1\nc\n#else\nd\n#endif\n"), "b");
        assert_eq!(clean("#if 0\na\n#elif 0\nb\n#else\nc\n#endif\n"), "c");
    }

    #[test]
    fn a_branch_after_one_that_was_taken_is_not_evaluated() {
        // `1/0` in a branch that cannot be reached is legal, and headers rely on it: the
        // guard that made the branch dead is often the thing that made the expression safe.
        assert_eq!(clean("#if 1\na\n#elif 1/0\nb\n#endif\n"), "a");
    }

    #[test]
    fn a_skipped_region_is_not_read_for_anything_but_nesting() {
        // Prose, an unknown directive and a broken `#define` all have to pass silently.
        let src = "#if 0\nthis is not C at all\n#frobnicate\n#define\n#if 1\ninner\n#endif\n#endif\nafter\n";
        assert_eq!(clean(src), "after");
    }

    #[test]
    fn nesting_inside_a_dead_branch_stays_balanced() {
        let src = "#if 0\n#ifdef X\na\n#else\nb\n#endif\n#else\nc\n#endif\n";
        assert_eq!(clean(src), "c");
    }

    #[test]
    fn defined_works_in_both_spellings_and_before_expansion() {
        assert_eq!(clean("#define F 0\n#if defined F\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#define F 0\n#if defined(F)\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if defined(F)\nyes\n#endif\n"), "");
        // `F` expands to 0, but `defined F` is answered before that happens, which is the
        // whole reason `defined` is resolved in a pass of its own.
        assert_eq!(clean("#define F 0\n#if defined F && !F\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn an_identifier_that_survived_expansion_is_zero() {
        assert_eq!(clean("#if NOT_DEFINED_ANYWHERE\nyes\n#else\nno\n#endif\n"), "no");
        assert_eq!(clean("#if !NOT_DEFINED_ANYWHERE\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn short_circuiting_keeps_a_guarded_expression_safe() {
        // The reason `&&` has to short circuit rather than merely produce the right answer:
        // the right hand side divides by zero when the guard is false.
        assert_eq!(clean("#if defined(F) && 1/F\nyes\n#else\nno\n#endif\n"), "no");
        assert_eq!(clean("#if 1 ? 2 : 1/0\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn the_operators_have_the_precedence_they_do_in_c() {
        assert_eq!(clean("#if 1 + 2 * 3 == 7\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if (1 + 2) * 3 == 9\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if 1 << 4 == 16\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if -8 / 3 == -2\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if (0xff & 0x0f) == 15\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn an_unsigned_operand_makes_the_whole_comparison_unsigned() {
        // The rule that catches everyone out in C catches them out here too, and a
        // preprocessor that quietly disagreed with the compiler would be worse than one that
        // is merely surprising.
        assert_eq!(clean("#if -1 < 0u\nyes\n#else\nno\n#endif\n"), "no");
        assert_eq!(clean("#if -1 < 0\nyes\n#else\nno\n#endif\n"), "yes");
    }

    #[test]
    fn character_constants_evaluate() {
        assert_eq!(clean("#if 'A' == 65\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if '\\n' == 10\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn a_macro_is_expanded_before_the_expression_is_evaluated() {
        assert_eq!(clean("#define V 3\n#if V > 2\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#define M(a) ((a) * 2)\n#if M(3) == 6\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn an_invocation_may_span_lines_within_a_run_of_text() {
        assert_eq!(clean("#define M(a, b) a + b\nM(1,\n2)\n"), "1 + 2");
    }

    #[test]
    fn undef_removes_a_definition() {
        assert_eq!(clean("#define F 1\n#undef F\n#ifdef F\nyes\n#else\nno\n#endif\n"), "no");
        // Undefining something that was never defined is not an error, and configure scripts
        // emit it constantly.
        assert_eq!(clean("#undef NEVER_DEFINED\nok\n"), "ok");
    }

    #[test]
    fn some_names_cannot_be_undefined() {
        let mut run = Run::new();
        run.go("#undef defined\n");
        assert_eq!(run.messages(), vec!["`defined` cannot be undefined".to_owned()]);
    }

    #[test]
    fn error_reports_the_rest_of_the_line() {
        let mut run = Run::new();
        run.go("#if 0\n#error not this one\n#else\n#error unsupported target\n#endif\n");
        assert_eq!(run.messages(), vec!["unsupported target".to_owned()]);
    }

    #[test]
    fn warning_is_a_warning() {
        let mut run = Run::new();
        run.go("#warning this is fine\n");
        assert_eq!(run.severities(), vec![Severity::Warning]);
        assert_eq!(run.messages(), vec!["this is fine".to_owned()]);
    }

    #[test]
    fn an_unterminated_conditional_is_reported() {
        let mut run = Run::new();
        assert_eq!(run.go("#if 1\nyes\n"), "yes");
        assert_eq!(run.messages(), vec!["unterminated `#if`".to_owned()]);
    }

    #[test]
    fn a_conditional_without_an_if_is_reported() {
        let mut run = Run::new();
        run.go("#endif\n");
        assert_eq!(run.messages(), vec!["`#endif` without `#if`".to_owned()]);

        let mut run = Run::new();
        run.go("#if 1\n#else\n#else\n#endif\n");
        assert_eq!(run.messages(), vec!["a second `#else`".to_owned()]);

        let mut run = Run::new();
        run.go("#if 1\n#else\n#elif 1\n#endif\n");
        assert_eq!(run.messages(), vec!["`#elif` after `#else`".to_owned()]);
    }

    #[test]
    fn tokens_after_endif_are_a_warning_rather_than_an_error() {
        // `#endif FOO` as a hand written comment predates `//` being portable and there is a
        // great deal of it about. Refusing to compile it would be correct and useless.
        let mut run = Run::new();
        assert_eq!(run.go("#if 1\nyes\n#endif FOO\n"), "yes");
        assert_eq!(run.severities(), vec![Severity::Warning]);
        assert_eq!(run.messages(), vec!["extra tokens after `#endif`".to_owned()]);
    }

    #[test]
    fn the_null_directive_does_nothing() {
        assert_eq!(clean("#\na\n#\nb\n"), "a b");
    }

    #[test]
    fn an_unknown_directive_is_an_error_when_the_region_is_live() {
        let mut run = Run::new();
        run.go("#frobnicate\n");
        assert_eq!(run.messages(), vec!["invalid preprocessing directive".to_owned()]);
    }

    #[test]
    fn line_is_recorded_for_the_source_map() {
        let mut run = Run::new();
        run.go("#line 42 \"other.c\"\n");
        assert!(run.messages().is_empty());
        let recorded = run.pp.line_directives();
        assert_eq!(recorded.len(), 1);
        assert_eq!(recorded[0].line, 42);
        let file = recorded[0].file.expect("a file name was given");
        assert_eq!(run.interner.resolve(file), "\"other.c\"");
    }

    #[test]
    fn line_moves_what_line_and_file_the_lines_after_it_are_on() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("#line 1000\n__LINE__ __FILE__\n__LINE__\n"), "1000 \"/main.c\" 1001");
    }

    #[test]
    fn a_line_marker_moves_the_lines_after_it_the_way_line_does() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("# 200 \"xyz\"\n__FILE__ __LINE__\n"), "\"xyz\" 200");
    }

    #[test]
    fn a_line_marker_with_no_name_leaves_the_name_alone() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("# 20\n__FILE__ __LINE__\n"), "\"/main.c\" 20");
    }

    #[test]
    fn a_line_marker_may_say_line_zero() {
        // `#line 0` is an error and this is not, because a marker is written by a program and a
        // program counting from zero is allowed to say so.
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("# 0 \"xyz\"\n__LINE__\n"), "0");
    }

    #[test]
    fn entering_and_returning_are_a_nesting_the_marker_flags_keep() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        let text =
            run.go("# 200 \"xyz\" 1\n__FILE__\n# 5 \"/main.c\" 2\n__FILE__ __LINE__\n# 9 3 4\n");
        assert_eq!(text, "\"xyz\" \"/main.c\" 5");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn returning_to_a_file_nothing_was_ever_in_is_ignored() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("# 200 \"xyz\" 2 3\n__FILE__ __LINE__\n"), "\"/main.c\" 2");
        assert_eq!(
            run.messages(),
            vec!["file `xyz` linemarker ignored due to incorrect nesting".to_owned()]
        );
    }

    #[test]
    fn a_flag_that_is_not_one_of_the_four_is_an_error() {
        let mut run = Run::new();
        run.go("# 20 \"a\" 7\n");
        assert_eq!(run.messages(), vec!["invalid flag `7` in line directive".to_owned()]);
    }

    #[test]
    fn a_hash_and_something_that_is_not_a_line_number_is_still_an_unknown_directive() {
        // A preprocessing number is a wider thing than a number, and `1.5` is one of them.
        let mut run = Run::new();
        run.go("# 1.5 \"a\"\n");
        assert_eq!(run.messages(), vec!["invalid preprocessing directive".to_owned()]);
    }

    #[test]
    fn a_name_on_the_directive_is_the_name_from_there_on() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("#line 7 \"gen.y\"\n__FILE__ __LINE__\n"), "\"gen.y\" 7");
    }

    #[test]
    fn a_directive_with_no_name_keeps_the_one_already_in_force() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("#line 7 \"gen.y\"\n#line 20\n__FILE__ __LINE__\n"), "\"gen.y\" 20");
    }

    #[test]
    fn the_number_is_expanded_first_because_line_plus_one_is_real_code() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("#define WHERE 300\n#line WHERE\n__LINE__\n"), "300");
    }

    #[test]
    fn a_directive_in_a_header_does_not_move_the_file_that_included_it() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        run.file("/h.h", "#line 500\n__LINE__\n");
        run.dir("/");
        assert_eq!(run.go("#include <h.h>\n__LINE__\n"), "500 2");
    }

    #[test]
    fn extra_tokens_after_the_file_name_are_a_warning_and_not_an_error() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("#line 7 \"gen.y\" and more\n__LINE__\n"), "7");
        assert_eq!(run.messages(), vec!["extra tokens after `#line`".to_owned()]);
    }

    #[test]
    fn a_line_number_out_of_range_is_refused() {
        let mut run = Run::new();
        run.go("#line 0\n");
        assert_eq!(run.messages(), vec!["`#line` number is out of range".to_owned()]);

        let mut run = Run::new();
        run.go("#line notanumber\n");
        assert_eq!(run.messages(), vec!["`#line` needs a decimal line number".to_owned()]);
    }

    #[test]
    fn a_pragma_passes_through_unchanged() {
        assert_eq!(clean("#pragma pack(1)\nint x;\n"), "#pragma pack(1) int x;");
    }

    /// glibc indents a directive inside a nest of conditionals, one space per level, so
    /// `regex.h` writes `# pragma GCC diagnostic push`. gcc prints it back with the space
    /// gone, and a header preprocessed two ways that differ only there is a difference
    /// somebody has to read before deciding it does not matter.
    #[test]
    fn the_space_between_the_hash_and_the_word_comes_off_a_pragma_that_is_indented() {
        assert_eq!(clean("#if 1\n# pragma pack(1)\n#endif\n"), "#pragma pack(1)");
        assert_eq!(clean("#  pragma  pack( 1 )\n"), "#pragma pack( 1 )", "the rest is kept");
    }

    #[test]
    fn the_pragma_operator_becomes_a_pragma() {
        assert_eq!(
            clean("_Pragma(\"GCC visibility push(default)\")\nint x;\n"),
            "#pragma GCC visibility push(default) int x;"
        );
    }

    #[test]
    fn the_pragma_operator_works_from_inside_a_macro() {
        // This is the entire reason `_Pragma` exists: a `#pragma` cannot be written in a macro
        // body, so a header that wants to wrap one has no other option.
        let src = "#define PUSH _Pragma(\"pack(push)\")\nPUSH\nint x;\n";
        assert_eq!(clean(src), "#pragma pack(push) int x;");
    }

    /// A pragma is a line even when it was written as an expression, so whatever follows one
    /// has to start a line. The phase that turns these back into a record takes the line as
    /// its unit, and without this the `int` would be read as part of the pragma.
    #[test]
    fn what_follows_a_pragma_operator_starts_a_line() {
        let mut run = Run::new();
        let out = run.raw("int x; _Pragma(\"pack(1)\") int y;\n");
        let starts: Vec<_> =
            out.iter().map(|tok| tok.flags.has(TokenFlags::START_OF_LINE)).collect();
        // `int x ;` then the six the pragma became, then `int y ;`. Only the first `int` was
        // at the start of a line in the source, and the second one is now.
        assert_eq!(
            starts,
            vec![true, false, false, true, false, false, false, false, false, true, false, false]
        );
    }

    #[test]
    fn a_pragma_operator_that_is_not_given_a_string_is_reported() {
        let mut run = Run::new();
        run.go("_Pragma(x)\n");
        assert_eq!(run.messages(), vec!["`_Pragma` takes a single string literal".to_owned()]);
    }

    #[test]
    fn an_include_reads_the_file_it_names() {
        let mut run = Run::new();
        run.file("/dir/one.h", "int from_the_header;\n");
        run.dir("/dir");
        assert_eq!(run.go("#include <one.h>\nint after;\n"), "int from_the_header; int after;");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn a_quoted_include_looks_next_to_the_including_file_first() {
        let mut run = Run::new();
        run.file("/local.h", "beside\n");
        run.file("/dir/local.h", "on the path\n");
        run.dir("/dir");
        assert_eq!(run.go("#include \"local.h\"\n"), "beside");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn an_angled_include_does_not_look_next_to_the_including_file() {
        let mut run = Run::new();
        run.file("/local.h", "beside\n");
        run.file("/dir/local.h", "on the path\n");
        run.dir("/dir");
        assert_eq!(run.go("#include <local.h>\n"), "on the path");
    }

    #[test]
    fn a_macro_defined_in_a_header_is_visible_after_the_include() {
        let mut run = Run::new();
        run.file("/dir/defs.h", "#define N 42\n");
        run.dir("/dir");
        assert_eq!(run.go("#include <defs.h>\nint a = N;\n"), "int a = 42;");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn an_include_guard_keeps_the_second_read_empty() {
        let mut run = Run::new();
        run.file("/dir/g.h", "#ifndef G\n#define G\nonce\n#endif\n");
        run.dir("/dir");
        assert_eq!(run.go("#include <g.h>\n#include <g.h>\n"), "once");
        assert!(run.messages().is_empty());
        assert_eq!(run.files(), 2, "the second include is not opened at all");
    }

    #[test]
    fn the_other_spelling_of_a_guard_is_recognised_too() {
        for guard in ["#if !defined(G)", "#if !defined G"] {
            let mut run = Run::new();
            run.file("/dir/g.h", &format!("{guard}\n#define G\nonce\n#endif\n"));
            run.dir("/dir");
            assert_eq!(run.go("#include <g.h>\n#include <g.h>\n"), "once");
            assert_eq!(run.files(), 2, "{guard} should be a guard");
        }
    }

    #[test]
    fn a_conditional_that_is_not_a_guard_does_not_skip_anything() {
        // Nothing defines the macro, so the second read is not the same as the first and the
        // file has to be opened again.
        let mut run = Run::new();
        run.file("/dir/g.h", "#ifndef G\ntwice\n#endif\n");
        run.dir("/dir");
        assert_eq!(run.go("#include <g.h>\n#include <g.h>\n"), "twice twice");
        assert_eq!(run.files(), 3);
    }

    #[test]
    fn a_token_outside_the_guard_stops_it_being_a_guard() {
        let mut run = Run::new();
        run.file("/dir/g.h", "#ifndef G\n#define G\n#endif\nalways\n");
        run.dir("/dir");
        assert_eq!(run.go("#include <g.h>\n#include <g.h>\n"), "always always");
        assert_eq!(run.files(), 3);
    }

    #[test]
    fn pragma_once_skips_the_second_read_and_does_not_reach_the_output() {
        let mut run = Run::new();
        run.file("/dir/o.h", "#pragma once\nonce\n");
        run.dir("/dir");
        assert_eq!(run.go("#include <o.h>\n#include <o.h>\n"), "once");
        assert!(run.messages().is_empty());
        assert_eq!(run.files(), 2);
    }

    #[test]
    fn pragma_once_in_the_main_file_is_a_warning_and_is_still_applied() {
        // The warning is about the usual case, a main file that meant to be a header. The
        // line is applied anyway, because the file that includes itself is the case where it
        // does work in a main file, and without it this is an infinite include.
        let mut run = Run::new();
        let src = "#pragma once\n#include <s.c>\nbody\n";
        run.file("/dir/s.c", src);
        run.dir("/dir");
        assert_eq!(run.go_named("/dir/s.c", src), "body");
        assert_eq!(run.severities(), vec![Severity::Warning]);
        assert_eq!(run.messages(), vec!["`#pragma once` in the main file".to_owned()]);
        assert_eq!(run.files(), 1);
    }

    #[test]
    fn pragma_once_holds_across_two_spellings_of_the_one_path() {
        // `-I .` puts a `./` in front of everything it finds, and the file that asked to be
        // read once was named without one. Comparing the text as written would read it twice.
        let mut run = Run::new();
        run.file("dir/s.c", "#pragma once\nbody\n");
        run.dir(".");
        assert_eq!(run.go("#include <dir/s.c>\n#include <dir/s.c>\n"), "body");
        assert!(run.messages().is_empty());
        assert_eq!(run.files(), 2);
    }

    #[test]
    fn any_other_pragma_still_passes_through() {
        assert_eq!(clean("#pragma once_upon_a_time\n"), "#pragma once_upon_a_time");
    }

    /// clang's own `__clang_cuda_complex_builtins.h` opens with this, and a header that pushes a
    /// name, defines it for its own use and pops it at the end is the whole idiom.
    #[test]
    fn push_macro_and_pop_macro_put_a_definition_aside_and_bring_it_back() {
        let src = "#define X 1\n#pragma push_macro(\"X\")\n#undef X\n#define X 2\n                   a X\n#pragma pop_macro(\"X\")\nb X\n";
        assert_eq!(clean(src), "a 2 b 1");
    }

    #[test]
    fn a_name_with_no_definition_pushes_and_pops_the_absence() {
        // The pragma is about the state, and "not defined" is a state. A header that pushes a
        // name it does not know about has to get an undefined name back, not the one it made.
        let src = "#pragma push_macro(\"X\")\n#define X 1\na X\n#pragma pop_macro(\"X\")\nb X\n";
        assert_eq!(clean(src), "a 1 b X");
    }

    #[test]
    fn the_pushes_nest() {
        let src = "#define X 1\n#pragma push_macro(\"X\")\n#undef X\n#define X 2\n                   #pragma push_macro(\"X\")\n#undef X\n#define X 3\n                   a X\n#pragma pop_macro(\"X\")\nb X\n#pragma pop_macro(\"X\")\nc X\n";
        assert_eq!(clean(src), "a 3 b 2 c 1");
    }

    #[test]
    fn a_pop_with_nothing_pushed_says_nothing() {
        // The two are written in pairs across headers that do not know about each other, so a
        // diagnostic here would fire on code that is not wrong. gcc is silent as well.
        assert_eq!(clean("#define X 1\n#pragma pop_macro(\"X\")\nX\n"), "1");
        assert_eq!(clean("#pragma pop_macro(\"Never\")\nx\n"), "x");
    }

    #[test]
    fn the_pragma_operator_spelling_works_and_takes_effect_where_it_is_written() {
        // A `#pragma` cannot come out of a macro body, so a macro that wants to save a name has
        // only this spelling. The lines around it are one run of text to the expander, and the
        // pop has to be answered before the line after it is expanded or that line still sees
        // the definition the pop was there to undo.
        let src = "#define X 1\n_Pragma(\"push_macro(\\\"X\\\")\")\n#undef X\n#define X 2\n                   a X\n_Pragma(\"pop_macro(\\\"X\\\")\")\nb X\n";
        assert_eq!(clean(src), "a 2 b 1");
    }

    #[test]
    fn the_gcc_spelling_is_not_one_of_these_and_passes_through() {
        // `#pragma GCC push_macro("X")` does nothing in gcc and is printed back, unlike the
        // namespaced spellings of the pragmas the compiler proper reads. Answering it here
        // would be a difference from gcc dressed up as a courtesy.
        let src = "#define X 1\n#pragma GCC push_macro(\"X\")\n#undef X\n#define X 2\nX\n";
        assert_eq!(clean(src), "#pragma GCC push_macro(\"X\") 2");
    }

    #[test]
    fn a_push_macro_that_is_not_the_shape_is_an_error() {
        for src in ["#pragma push_macro\n", "#pragma push_macro(X)\n", "#pragma pop_macro()\n"] {
            let mut run = Run::new();
            run.go(src);
            let word = if src.contains("push") { "push" } else { "pop" };
            assert_eq!(
                run.messages(),
                vec![format!("invalid `#pragma {word}_macro` directive")],
                "from {src:?}"
            );
        }
    }

    #[test]
    fn a_string_that_does_not_spell_one_identifier_names_no_macro() {
        // gcc neither complains about these nor does anything with them, and matching that is
        // worth more than improving on it: a header that has one has been building for years.
        assert_eq!(clean("#pragma push_macro(\"a b\")\nx\n"), "x");
        assert_eq!(clean("#pragma push_macro(\"2\")\nx\n"), "x");
    }

    #[test]
    fn what_follows_the_closing_parenthesis_is_the_usual_warning() {
        let mut run = Run::new();
        assert_eq!(run.go("#define X 1\n#pragma push_macro(\"X\") junk\nX\n"), "1");
        assert_eq!(run.severities(), vec![Severity::Warning]);
        assert_eq!(run.messages(), vec!["extra tokens after `#pragma`".to_owned()]);
    }

    #[test]
    fn has_include_answers_from_the_search_path() {
        let mut run = Run::new();
        run.file("/dir/there.h", "");
        run.dir("/dir");
        let src = "#if __has_include(<there.h>)\nyes\n#endif\n\
                   #if __has_include(<gone.h>)\nno\n#endif\n";
        assert_eq!(run.go(src), "yes");
        assert!(run.messages().is_empty(), "a header that is not there is an answer, not an error");
    }

    #[test]
    fn has_include_asks_the_question_the_include_on_the_same_line_would() {
        // The quoted form looks next to the file that wrote it, so the two spellings answer
        // differently about the same header. A `__has_include` that did not agree with the
        // `#include` it guards would be worse than not having one.
        let mut run = Run::new();
        run.file("/beside.h", "");
        let src = "#if __has_include(\"beside.h\")\nquoted\n#endif\n\
                   #if __has_include(<beside.h>)\nangled\n#endif\n";
        assert_eq!(run.go(src), "quoted");
    }

    #[test]
    fn has_include_next_starts_where_include_next_would() {
        let mut run = Run::new();
        run.file("/a/both.h", "#if __has_include_next(<both.h>)\nmore\n#endif\n");
        run.file("/b/both.h", "last\n");
        run.file("/a/only.h", "#if __has_include_next(<only.h>)\nmore\n#endif\n");
        run.dir("/a");
        run.dir("/b");
        assert_eq!(run.go("#include <both.h>\n"), "more");
        assert_eq!(run.go("#include <only.h>\n"), "", "there is nothing after /a to find it in");
    }

    #[test]
    fn the_operand_of_has_include_is_not_macro_expanded() {
        // `linux` is a predefined macro on a Linux target, and `<linux/version.h>` is a real
        // header. Expanding the operand would ask about `<1/version.h>`.
        let mut run = Run::new();
        run.file("/dir/linux/version.h", "");
        run.dir("/dir");
        let src = "#define linux 1\n#if __has_include(<linux/version.h>)\nyes\n#endif\n";
        assert_eq!(run.go(src), "yes");
    }

    #[test]
    fn a_macro_may_expand_to_a_has_include() {
        // Which is why the operators are resolved after expansion as well as before it.
        let mut run = Run::new();
        run.file("/dir/there.h", "");
        run.dir("/dir");
        let src = "#define HAVE __has_include(<there.h>)\n#if HAVE\nyes\n#endif\n";
        assert_eq!(run.go(src), "yes");
    }

    #[test]
    fn defined_says_the_has_operators_are_there() {
        // The shape every header that uses them is written in, because they are newer than
        // some of the compilers it has to build under.
        let src = "#if defined(__has_include) && defined __has_builtin\nyes\n#endif\n";
        assert_eq!(clean(src), "yes");
        assert_eq!(clean("#ifdef __has_attribute\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn has_attribute_answers_out_of_the_matrix() {
        // No attribute is implemented until the parser lands, and the table saying so is the
        // whole point: a yes here would send a header down a path that then fails to compile.
        assert_eq!(clean("#if __has_attribute(packed)\nyes\n#endif\n"), "");
        assert_eq!(clean("#if __has_attribute(no_such_attribute)\nyes\n#endif\n"), "");
        assert_eq!(clean("#if !__has_attribute(packed)\nno\n#endif\n"), "no");
    }

    #[test]
    fn the_scoped_spelling_of_an_attribute_is_the_same_question() {
        // `[[gnu::packed]]` and `__attribute__((packed))` are one attribute, and
        // `__has_c_attribute` answers with the value the standard gives it rather than with
        // one. Both answer zero today because the table says the attribute is unimplemented.
        assert_eq!(clean("#if __has_c_attribute(gnu::packed)\nyes\n#endif\n"), "");
        assert_eq!(clean("#if __has_c_attribute(deprecated)\nyes\n#endif\n"), "");
    }

    #[test]
    fn has_builtin_answers_no_until_the_builtin_is_real() {
        assert_eq!(clean("#if __has_builtin(__builtin_expect)\nyes\n#endif\n"), "");
        assert_eq!(clean("#if __has_builtin(__builtin_nonesuch)\nyes\n#endif\n"), "");
    }

    #[test]
    fn has_feature_and_has_extension_read_the_same_table() {
        // The preprocessor features are the ones that are real today, so they are the ones
        // that answer yes, and `__has_extension` answers yes wherever `__has_feature` does.
        assert_eq!(clean("#if __has_feature(pragma_once)\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if __has_extension(pragma_once)\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if __has_extension(include_next)\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if __has_feature(include_next)\nyes\n#endif\n"), "");
        assert_eq!(clean("#if __has_feature(statement_expressions)\nyes\n#endif\n"), "");
    }

    #[test]
    fn building_module_is_always_no_and_is_recognised_so_that_the_line_parses() {
        // Clang's own stddef.h writes this, and the whole point of knowing the name is that
        // the operand disappears with it. An unknown identifier would leave `(m)` behind and
        // the `#if` would fail to parse rather than answering no.
        assert_eq!(clean("#if __building_module(m)\nyes\n#endif\n"), "");
        assert_eq!(clean("#if !__building_module(m)\nyes\n#endif\n"), "yes");
        assert_eq!(
            clean(
                "#if !defined(offsetof) || (__has_feature(modules) && !__building_module(x))\nyes\n#endif\n"
            ),
            "yes"
        );
        // Defined, the same as the rest of the family: a header asks before it uses one.
        assert_eq!(clean("#ifdef __building_module\nyes\n#endif\n"), "yes");
        assert_eq!(clean("#if defined(__building_module)\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn a_has_operator_without_an_operand_is_reported() {
        let mut run = Run::new();
        run.go("#if __has_include\nyes\n#endif\n");
        assert_eq!(run.messages(), ["expected `(` after `__has_include`"]);
        let mut run = Run::new();
        run.go("#if __has_include(1)\nyes\n#endif\n");
        assert_eq!(run.messages(), ["expected a file name in `<>` or `\"\"`"]);
        let mut run = Run::new();
        run.go("#if __has_attribute(\"packed\")\nyes\n#endif\n");
        assert_eq!(run.messages(), ["expected an identifier as the operand of `__has_attribute`"]);
    }

    #[test]
    fn the_has_operators_answer_in_ordinary_text_too() {
        // GCC and clang both make these builtin macros rather than something only the
        // conditional parser knows, so a program may write one in a declaration. Real headers
        // do: an attribute macro is often written as the answer rather than as a `#if`.
        assert_eq!(clean("f __has_feature(pragma_once)\n"), "f 1");
        assert_eq!(clean("b __has_builtin(__builtin_expect)\n"), "b 0");
        assert_eq!(clean("a __has_attribute(packed)\n"), "a 0");
        assert_eq!(clean("c __has_c_attribute(deprecated)\n"), "c 0");
        assert_eq!(clean("m __building_module(foo)\n"), "m 0");
    }

    #[test]
    fn a_macro_that_expands_to_a_has_operator_is_answered_where_it_is_used() {
        // The awkward half of the same feature. The answer is deferred to wherever the macro
        // lands, so the sweep has to run after expansion and not only before it.
        assert_eq!(clean("#define HAVE __has_feature(pragma_once)\nx HAVE\n"), "x 1");
        assert_eq!(clean("#define HAVE(x) __has_attribute(x)\ny HAVE(packed)\n"), "y 0");
    }

    #[test]
    fn a_has_operator_in_text_still_needs_its_operand() {
        let mut run = Run::new();
        run.go("tail __has_attribute;\n");
        assert_eq!(run.messages(), ["expected `(` after `__has_attribute`"]);
    }

    #[test]
    fn the_header_operators_are_refused_in_ordinary_text() {
        // `<stdio.h>` in a text line was scanned as a run of comparisons, so there is no
        // header name left to ask about. GCC and clang both say the same thing here.
        let mut run = Run::new();
        run.file("/dir/there.h", "");
        run.dir("/dir");
        run.go("a __has_include(<there.h>)\n");
        assert_eq!(run.messages(), ["`__has_include` used outside of a preprocessing directive"]);
        let mut run = Run::new();
        run.go("b __has_include_next(\"x.h\")\n");
        assert_eq!(
            run.messages(),
            ["`__has_include_next` used outside of a preprocessing directive"]
        );
    }

    #[test]
    fn the_predefined_set_is_visible_to_the_source_file() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        let src = "#if defined(__x86_64__) && defined(__linux__) && __SIZEOF_LONG__ == 8\n\
                   yes\n#endif\n";
        assert_eq!(run.go(src), "yes");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_predefined_set_follows_the_target_and_not_the_host() {
        let mut run = Run::new();
        run.predefine("aarch64-unknown-linux-gnu", &Predef::new());
        assert_eq!(
            run.go("#ifdef __x86_64__\nno\n#endif\n#ifdef __aarch64__\nyes\n#endif\n"),
            "yes"
        );
    }

    #[test]
    fn a_predefined_macro_expands_where_it_is_used() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("__SIZE_TYPE__ n;\n"), "long unsigned int n;");
    }

    #[test]
    fn a_command_line_define_is_a_definition_like_any_other() {
        let mut opts = Predef::new();
        opts.defines = vec!["FOO".to_owned(), "BAR=3".to_owned()];
        opts.undefines = vec!["__linux__".to_owned()];
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &opts);
        let src = "#if FOO && BAR == 3 && !defined(__linux__)\nyes\n#endif\n";
        assert_eq!(run.go(src), "yes");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_predefined_set_produces_no_tokens_of_its_own() {
        // It is a file of directives, so the output of the compilation is the source file
        // and nothing else. A stray token here would appear at the top of every `-E` run.
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("alone\n"), "alone");
    }

    #[test]
    fn the_predefined_files_are_named_the_way_gcc_names_them() {
        let mut run = Run::new();
        let mut opts = Predef::new();
        opts.defines = vec!["FOO=1".to_owned()];
        run.predefine("x86_64-unknown-linux-gnu", &opts);
        let names: Vec<&str> = run.sources.files().iter().map(|f| f.name.as_str()).collect();
        assert_eq!(names, ["<built-in>", "<command-line>"]);
    }

    #[test]
    fn a_dialect_without_the_gnu_extensions_says_so() {
        let mut opts = Predef::new();
        opts.gnu_extensions = false;
        opts.std = Std::C99;
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &opts);
        let src = "#if defined(__STRICT_ANSI__) && __STDC_VERSION__ == 199901L && !defined(linux)\n\
                   yes\n#endif\n";
        assert_eq!(run.go(src), "yes");
    }

    #[test]
    fn the_date_and_time_are_the_same_for_the_whole_translation_unit() {
        let mut opts = Predef::new();
        opts.timestamp = Timestamp::from_unix(0);
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &opts);
        assert_eq!(run.go("__DATE__ __TIME__\n"), "\"Jan  1 1970\" \"00:00:00\"");
    }

    #[test]
    fn a_has_operator_in_a_dead_branch_is_not_asked_about() {
        // The line is not evaluated at all, so a malformed one inside `#if 0` is text.
        assert_eq!(clean("#if 0\n#if __has_include\n#endif\n#endif\nafter\n"), "after");
    }

    #[test]
    fn a_conditional_may_not_span_an_include() {
        // GCC and Clang both refuse this, and the reason is that a header which opens a
        // conditional it does not close leaves the file that included it in a state nothing
        // downstream can reason about.
        let mut run = Run::new();
        run.file("/dir/open.h", "#if 1\n");
        run.dir("/dir");
        run.go("#include <open.h>\nkept\n#endif\n");
        let messages = run.messages();
        assert_eq!(messages.len(), 2);
        assert!(messages[0].contains("unterminated"));
        assert!(messages[1].contains("without"));
    }

    #[test]
    fn include_next_continues_after_the_directory_the_file_came_from() {
        // The wrapper header trick: `/a` has a `limits.h` that pulls in the real one from
        // `/b`, and the two have the same name on purpose.
        let mut run = Run::new();
        run.file("/a/limits.h", "wrapper\n#include_next <limits.h>\n");
        run.file("/b/limits.h", "real\n");
        run.dir("/a");
        run.dir("/b");
        assert_eq!(run.go("#include <limits.h>\n"), "wrapper real");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn a_computed_include_is_expanded_first() {
        let mut run = Run::new();
        run.file("/dir/sub/thing.h", "computed\n");
        run.dir("/dir");
        let src = "#define HEADER <sub/thing.h>\n#include HEADER\n";
        assert_eq!(run.go(src), "computed");
        assert!(run.messages().is_empty());
        // The string literal form goes through the same path and keeps its delimiters.
        let mut run = Run::new();
        run.file("/dir/sub/thing.h", "computed\n");
        run.dir("/dir");
        assert_eq!(run.go("#define H \"sub/thing.h\"\n#include H\n"), "computed");
    }

    #[test]
    fn a_header_that_is_not_there_says_where_it_looked() {
        let mut run = Run::new();
        run.dir("/dir");
        run.go("#include <nope.h>\n");
        let diagnostics = run.pp.take_diagnostics();
        assert_eq!(diagnostics.len(), 1);
        assert_eq!(diagnostics[0].code, Some("E0341"));
        assert_eq!(diagnostics[0].message, "`nope.h` file not found");
        assert!(diagnostics[0].children[0].message.contains("/dir"));
    }

    #[test]
    fn an_include_that_is_not_a_header_name_is_reported() {
        let mut run = Run::new();
        run.go("#include 3\n");
        let diagnostics = run.pp.take_diagnostics();
        assert_eq!(diagnostics[0].code, Some("E0343"));
    }

    #[test]
    fn a_header_that_includes_itself_stops() {
        let mut run = Run::new();
        run.file("/dir/loop.h", "#include <loop.h>\n");
        run.dir("/dir");
        run.go("#include <loop.h>\n");
        let diagnostics = run.pp.take_diagnostics();
        assert_eq!(diagnostics.len(), 1, "one complaint, not one per level");
        assert_eq!(diagnostics[0].code, Some("E0342"));
    }

    #[test]
    fn an_include_in_a_dead_branch_is_not_read() {
        let mut run = Run::new();
        assert_eq!(run.go("#if 0\n#include <nothing.h>\n#endif\nafter\n"), "after");
        assert!(run.messages().is_empty(), "a skipped include is not resolved");
    }

    #[test]
    fn embed_writes_the_bytes_of_the_resource() {
        let mut run = Run::new();
        run.bytes("/logo.bin", &[0, 1, 127, 128, 255]);
        assert_eq!(run.go("#embed \"logo.bin\"\n"), "0, 1, 127, 128, 255");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn an_embed_is_a_valid_initializer_on_both_sides_of_empty() {
        // The reason `prefix` and `suffix` exist. An empty resource is `if_empty` alone, with
        // neither of them, so the same three lines are a well formed array whether the file
        // has bytes in it or not. Emitting `prefix` and `suffix` around nothing would leave a
        // trailing comma inside the braces and turn an empty file into a syntax error.
        let mut run = Run::new();
        run.bytes("/some.bin", &[7, 8]);
        run.bytes("/none.bin", &[]);
        let line = |name: &str| {
            format!("{{\n#embed \"{name}\" prefix(0xEF,) suffix(,0xFE) if_empty(0)\n}}\n")
        };
        assert_eq!(run.go(&line("some.bin")), "{ 0xEF,7, 8 ,0xFE }");
        assert_eq!(run.go_named("/other.c", &line("none.bin")), "{ 0 }");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_limit_and_the_offset_choose_a_window_of_the_resource() {
        let mut run = Run::new();
        run.bytes("/eight.bin", &[1, 2, 3, 4, 5, 6, 7, 8]);
        assert_eq!(run.go("#embed \"eight.bin\" limit(3)\n"), "1, 2, 3");
        assert_eq!(
            run.go_named("/b.c", "#embed \"eight.bin\" gnu::offset(4) limit(3)\n"),
            "5, 6, 7"
        );
        // A limit of zero is an empty embed, not an unlimited one, and an offset past the end
        // is empty rather than an error.
        assert_eq!(run.go_named("/c.c", "#embed \"eight.bin\" limit(0) if_empty(9)\n"), "9");
        assert_eq!(run.go_named("/d.c", "#embed \"eight.bin\" gnu::offset(99)\n"), "");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_limit_is_a_constant_expression_and_not_just_a_number() {
        // It is the `#if` language, so a macro and arithmetic both work. A header that writes
        // `limit(CHUNK * 2)` is doing the ordinary thing.
        let mut run = Run::new();
        run.bytes("/eight.bin", &[1, 2, 3, 4, 5, 6, 7, 8]);
        assert_eq!(
            run.go("#define CHUNK 2\n#embed \"eight.bin\" limit(CHUNK * 2)\n"),
            "1, 2, 3, 4"
        );
        assert!(run.messages().is_empty());
    }

    #[test]
    fn a_misspelled_embed_parameter_is_refused_rather_than_ignored() {
        // Carrying on without it would produce an array with the wrong contents and no
        // message, which is the worst outcome available.
        let mut run = Run::new();
        run.bytes("/eight.bin", &[1, 2]);
        assert_eq!(run.go("#embed \"eight.bin\" limits(1)\n"), "");
        assert_eq!(run.messages(), vec!["unknown `#embed` parameter `limits`".to_owned()]);
        let mut vendor = Run::new();
        vendor.bytes("/eight.bin", &[1, 2]);
        assert_eq!(vendor.go("#embed \"eight.bin\" clang::offset(1)\n"), "");
        assert_eq!(
            vendor.messages(),
            vec!["unknown `#embed` parameter `clang::offset`".to_owned()]
        );
    }

    #[test]
    fn a_missing_embed_resource_is_reported_as_a_resource() {
        let mut run = Run::new();
        run.go("#embed <nothing.bin>\n");
        assert_eq!(run.messages(), vec!["`nothing.bin` resource not found".to_owned()]);
    }

    #[test]
    fn has_embed_tells_missing_from_present_from_empty() {
        // Three answers, which is the reason the operator is not `__has_include` with a
        // different name. A present but empty resource needs its `if_empty` written and a
        // missing one needs a fallback, and a yes or no cannot tell the two apart.
        let mut run = Run::new();
        run.bytes("/some.bin", &[1]);
        run.bytes("/none.bin", &[]);
        let src = "#if __has_embed(\"none.bin\") == __STDC_EMBED_EMPTY__\nempty\n#endif\n\
                   #if __has_embed(\"some.bin\") == __STDC_EMBED_FOUND__\nfound\n#endif\n\
                   #if __has_embed(\"gone.bin\") == __STDC_EMBED_NOT_FOUND__\ngone\n#endif\n";
        run.predefine("x86_64-unknown-linux-gnu", &Predef::default());
        assert_eq!(run.go(src), "empty found gone");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn has_embed_takes_the_limit_into_account() {
        // The guard has to answer the question the directive it guards will ask. A resource
        // that exists but has nothing left after `limit(0)` is empty to both of them.
        let mut run = Run::new();
        run.bytes("/some.bin", &[1, 2, 3]);
        run.predefine("x86_64-unknown-linux-gnu", &Predef::default());
        let src = "#if __has_embed(\"some.bin\" limit(0)) == __STDC_EMBED_EMPTY__\nempty\n#endif\n";
        assert_eq!(run.go(src), "empty");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn a_directive_may_have_space_before_the_hash_and_after_it() {
        assert_eq!(clean("  #  define F 1\n#ifdef F\nyes\n#endif\n"), "yes");
    }

    #[test]
    fn a_definition_survives_across_a_conditional() {
        assert_eq!(clean("#if 1\n#define F 7\n#endif\nF\n"), "7");
    }

    #[test]
    fn an_empty_if_expression_is_reported() {
        let mut run = Run::new();
        run.go("#if\n#endif\n");
        assert_eq!(run.messages(), vec!["`#if` with no expression".to_owned()]);
    }

    #[test]
    fn the_file_and_the_line_say_where_the_use_is() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("__FILE__ __LINE__\n__LINE__\n"), "\"/main.c\" 1 2");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn a_macro_that_mentions_the_line_answers_with_the_call() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        run.file("/where.h", "#define WHERE __FILE__ __LINE__\n");
        // The point of the whole arrangement. `assert` is this macro, and a version that
        // answered with the header the macro was written in would name a file the user has
        // never opened and a line that means nothing.
        assert_eq!(run.go("#include \"where.h\"\n\n\nWHERE\n"), "\"/main.c\" 4");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_file_name_is_the_file_without_the_directories() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go_named("/deep/down/main.c", "__FILE_NAME__\n"), "\"main.c\"");
    }

    #[test]
    fn a_backslash_in_the_name_is_escaped() {
        let mut run = Run::new();
        run.predefine("x86_64-pc-windows-msvc", &Predef::new());
        // The literal has to mean the path, so the separators are escaped. Getting this wrong
        // turns `\src` into an unknown escape and `\a` into a bell character.
        let text = run.go_named("C:\\src\\main.c", "__FILE__ __FILE_NAME__\n");
        assert_eq!(text, "\"C:\\\\src\\\\main.c\" \"main.c\"");
    }

    #[test]
    fn the_base_file_is_the_one_named_on_the_command_line() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        run.file("/deep.h", "__FILE__ __BASE_FILE__\n");
        assert_eq!(run.go("#include \"deep.h\"\n"), "\"/deep.h\" \"/main.c\"");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_include_level_counts_the_headers_above_it() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        run.file("/one.h", "__INCLUDE_LEVEL__\n#include \"two.h\"\n");
        run.file("/two.h", "__INCLUDE_LEVEL__\n");
        assert_eq!(run.go("__INCLUDE_LEVEL__\n#include \"one.h\"\n"), "0 1 2");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_counter_is_a_different_number_every_time() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("__COUNTER__ __COUNTER__ __COUNTER__\n"), "0 1 2");
    }

    #[test]
    fn the_counter_advances_once_per_argument_rather_than_once_per_use() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        // An argument is expanded once however many times the body names it, so `TWICE`
        // produces the same number twice. That is what GCC does, and the reason for it is
        // that expanding an argument twice would report anything wrong inside it twice.
        assert_eq!(run.go("#define TWICE(x) x x\nTWICE(__COUNTER__) __COUNTER__\n"), "0 0 1");
    }

    #[test]
    fn the_line_is_a_number_an_if_can_use() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("#if __LINE__ == 1 && __INCLUDE_LEVEL__ == 0\nyes\n#endif\n"), "yes");
        assert!(run.messages().is_empty());
    }

    #[test]
    fn the_dynamic_macros_are_defined_like_any_others() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        let src = "#ifdef __FILE__\nyes\n#endif\n#undef __LINE__\n#ifndef __LINE__\ngone\n#endif\n";
        assert_eq!(run.go(src), "yes gone");
        assert!(run.messages().is_empty(), "`#undef` of a builtin is allowed, as it is in GCC");
    }

    #[test]
    fn redefining_a_dynamic_macro_warns_and_points_at_the_built_in_file() {
        let mut run = Run::new();
        run.predefine("x86_64-unknown-linux-gnu", &Predef::new());
        assert_eq!(run.go("#define __FILE__ \"mine.c\"\n__FILE__\n"), "\"mine.c\"");
        let complaints = run.pp.take_diagnostics();
        assert_eq!(complaints.len(), 1);
        assert_eq!(complaints[0].code, Some("W0301"));
        let previous = complaints[0].children.first().expect("a note saying where it was");
        assert_eq!(run.sources.lookup(previous.span.lo).map(|loc| loc.file), {
            let built_in = run.sources.files().iter().find(|f| f.name == BUILT_IN);
            built_in.map(|f| f.id)
        });
    }

    #[test]
    fn destringizing_undoes_what_stringizing_did() {
        assert_eq!(destringize(r#""a \"b\" c""#), r#"a "b" c"#);
        assert_eq!(destringize(r#""a \\ b""#), r"a \ b");
        assert_eq!(destringize(r#"L"wide""#), "wide");
    }
}