mbx-cache-cc 0.20.6

//! Dependency-list parsing and input discovery for C and C++ compiles.

use crate::{
    CcActionContext, CcActionInput, CcBypassReason, CcCompilerFamily, MAX_INPUT_BYTES,
    MAX_MANIFEST_ENTRIES, MAX_PREDICTED_INPUTS, normalize_components,
};
use mbx_cache_core::{
    CacheDigest, FileDigestCache, FileDigestResolution, FileDigestScope, FileIdentity,
    FileSnapshot, RecordedFileDigest, digest_file, digest_file_validated,
};
use std::collections::{BTreeMap, BTreeSet};
use std::io::Read;
use std::path::{Path, PathBuf};
use std::time::SystemTime;

/// Marker distinguishing an include-directory name manifest from a file input.
pub const INCLUDE_MANIFEST_PREFIX: &str = "@include-manifest:";

/// Directives whose operands are consumed by the assembler, after the
/// compiler has finished producing its dependency list.
const ASSEMBLER_INPUT_DIRECTIVES: &[&[u8]] = &[b".include", b".incbin", b".sinclude"];

const SCAN_CHUNK_BYTES: usize = 64 * 1024;

/// A parsed GNU-style dependency list.
#[derive(Debug, Clone, PartialEq, Eq, Default)]
pub struct CcDepfile {
    /// Prerequisite files named by the first rule.
    pub files: Vec<PathBuf>,
}

impl CcDepfile {
    /// Read and parse the dependency list the compiler wrote.
    pub fn read(path: &Path) -> Result<Self, CcBypassReason> {
        let contents =
            std::fs::read_to_string(path).map_err(|error| CcBypassReason::DepfileRead {
                path: path.to_path_buf(),
                message: error.to_string(),
            })?;
        Self::parse(&contents)
    }

    /// Read the dependency output emitted by the selected compiler family.
    pub fn read_for(path: &Path, family: CcCompilerFamily) -> Result<Self, CcBypassReason> {
        if family.is_msvc() {
            Self::read_msvc(path)
        } else {
            Self::read(path)
        }
    }

    /// Read MSVC's `/sourceDependencies` JSON output.
    pub fn read_msvc(path: &Path) -> Result<Self, CcBypassReason> {
        let contents = std::fs::read(path).map_err(|error| CcBypassReason::DepfileRead {
            path: path.to_path_buf(),
            message: error.to_string(),
        })?;
        let value: serde_json::Value = serde_json::from_slice(&contents)
            .map_err(|error| CcBypassReason::MalformedDepfile(error.to_string()))?;
        let data = value
            .get("Data")
            .and_then(serde_json::Value::as_object)
            .ok_or_else(|| CcBypassReason::MalformedDepfile("missing Data object".into()))?;
        if data
            .get("ImportedModules")
            .and_then(serde_json::Value::as_array)
            .is_some_and(|modules| !modules.is_empty())
            || data.get("ProvidedModule").is_some_and(|module| {
                !module.is_null() && module.as_str().is_none_or(|s| !s.is_empty())
            })
        {
            return Err(CcBypassReason::MalformedDepfile(
                "C++ module dependencies are not modeled".into(),
            ));
        }
        let includes = data
            .get("Includes")
            .and_then(serde_json::Value::as_array)
            .ok_or_else(|| CcBypassReason::MalformedDepfile("missing Includes array".into()))?;
        let files = includes
            .iter()
            .map(|entry| {
                entry.as_str().map(PathBuf::from).ok_or_else(|| {
                    CcBypassReason::MalformedDepfile("non-string include path".into())
                })
            })
            .collect::<Result<Vec<_>, _>>()?;
        Ok(Self { files })
    }

    /// Parse a GNU-style dependency list.
    ///
    /// Only the first rule is read. The adapter never passes `-MP`, so a
    /// well-formed file the adapter asked for has exactly one rule, and
    /// anything further is ignored rather than guessed at.
    pub fn parse(contents: &str) -> Result<Self, CcBypassReason> {
        let joined = join_continuations(contents)?;
        let (_, prerequisites) = joined
            .lines()
            .find_map(|line| line.split_once(RULE_SEPARATOR))
            .ok_or_else(|| CcBypassReason::MalformedDepfile("no dependency rule".into()))?;
        let files = split_prerequisites(prerequisites)?;
        Ok(Self { files })
    }
}

const RULE_SEPARATOR: &str = ": ";

impl CcDepfile {
    /// Render the dependency list a caller asked for, the way the driver
    /// writes one: the rule's targets, then every prerequisite on its own
    /// continued line, then, for `-MP`, an empty rule per header so make does
    /// not fail when one is deleted. `source` is the file that was compiled
    /// and gets no phony rule, exactly as the driver leaves it out.
    ///
    /// Escapes are the ones `parse` reads back: a space, a `#`, and a `$` in
    /// a path. A quoted target gets the same treatment; a literal one is
    /// written as given, which is what `-MT` promises.
    pub fn render(
        targets: &[crate::DepfileTarget],
        files: &[PathBuf],
        source: &Path,
        phony_targets: bool,
    ) -> String {
        let mut rendered = String::new();
        for (index, target) in targets.iter().enumerate() {
            if index > 0 {
                rendered.push(' ');
            }
            if target.quoted {
                rendered.push_str(&escape_make_word(&target.name));
            } else {
                rendered.push_str(&target.name);
            }
        }
        rendered.push(':');
        for file in files {
            rendered.push_str(" \\\n ");
            rendered.push_str(&escape_make_word(&file.to_string_lossy()));
        }
        rendered.push('\n');
        if phony_targets {
            for file in files.iter().filter(|file| file.as_path() != source) {
                rendered.push_str(&escape_make_word(&file.to_string_lossy()));
                rendered.push_str(":\n");
            }
        }
        rendered
    }
}

/// Quote a word for make the way the driver does in a dependency list.
fn escape_make_word(word: &str) -> String {
    let mut escaped = String::with_capacity(word.len());
    for character in word.chars() {
        match character {
            ' ' => escaped.push_str("\\ "),
            '#' => escaped.push_str("\\#"),
            '$' => escaped.push_str("$$"),
            other => escaped.push(other),
        }
    }
    escaped
}

/// Join physical lines the compiler wrapped with a trailing backslash.
fn join_continuations(contents: &str) -> Result<String, CcBypassReason> {
    let mut joined = String::with_capacity(contents.len());
    let mut continued = false;
    for line in contents.lines() {
        let trimmed = line.strip_suffix('\r').unwrap_or(line);
        let (text, continues) = match trimmed.strip_suffix('\\') {
            Some(text) => (text, true),
            None => (trimmed, false),
        };
        if continued {
            joined.push(' ');
        }
        joined.push_str(text.trim_end_matches(['\t']));
        if !continues {
            joined.push('\n');
        }
        continued = continues;
    }
    if continued {
        return Err(CcBypassReason::MalformedDepfile(
            "unterminated line continuation".into(),
        ));
    }
    Ok(joined)
}

/// Split a prerequisite list, honoring exactly the escapes make defines.
///
/// Anything else escaped is a spelling this parser does not model, and a
/// mis-parsed prerequisite would silently drop an input from the key.
fn split_prerequisites(value: &str) -> Result<Vec<PathBuf>, CcBypassReason> {
    // Most compiler paths contain no make escapes. Copy each whole UTF-8
    // path once instead of growing a String one character at a time.
    if memchr::memchr2(b'\\', b'$', value.as_bytes()).is_none() {
        return Ok(value
            .split([' ', '\t'])
            .filter(|word| !word.is_empty())
            .map(PathBuf::from)
            .collect());
    }
    split_escaped_prerequisites(value)
}

fn split_escaped_prerequisites(value: &str) -> Result<Vec<PathBuf>, CcBypassReason> {
    let mut files = Vec::new();
    let mut current = String::new();
    let mut characters = value.chars().peekable();
    while let Some(character) = characters.next() {
        match character {
            ' ' | '\t' => {
                if !current.is_empty() {
                    files.push(PathBuf::from(std::mem::take(&mut current)));
                }
            }
            '\\' => match characters.next() {
                Some(' ') => current.push(' '),
                Some('#') => current.push('#'),
                Some(other) => {
                    return Err(CcBypassReason::MalformedDepfile(format!(
                        "unmodeled escape \\{other}"
                    )));
                }
                None => {
                    return Err(CcBypassReason::MalformedDepfile(
                        "trailing escape character".into(),
                    ));
                }
            },
            '$' => match characters.next() {
                Some('$') => current.push('$'),
                Some(other) => {
                    return Err(CcBypassReason::MalformedDepfile(format!(
                        "unmodeled variable reference ${other}"
                    )));
                }
                None => {
                    return Err(CcBypassReason::MalformedDepfile(
                        "trailing variable reference".into(),
                    ));
                }
            },
            other => current.push(other),
        }
    }
    if !current.is_empty() {
        files.push(PathBuf::from(current));
    }
    Ok(files)
}

/// A complete, content-addressed compiler input manifest.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct CcDiscoveredInputs {
    working_dir: PathBuf,
    /// Content-addressed inputs, including include-directory manifests.
    pub inputs: Vec<CcActionInput>,
    /// What each file input looked like on disk when its digest was
    /// established, index-aligned with `inputs`; `None` for manifests and
    /// where the filesystem gave nothing to compare against later.
    identities: Vec<Option<FileIdentity>>,
}

impl CcDiscoveredInputs {
    /// Digest every file the compilation read, and summarize the directories it
    /// searched.
    ///
    /// Digesting the files answers "did any input change". The directory
    /// manifests answer the question a dependency list cannot: whether a header
    /// that was *not* read now exists somewhere that would shadow one that was.
    pub fn collect(
        working_dir: &Path,
        files: BTreeSet<PathBuf>,
        directories: BTreeSet<PathBuf>,
        digests: &dyn FileDigestCache,
    ) -> Result<Self, CcBypassReason> {
        if !working_dir.is_absolute() {
            return Err(CcBypassReason::RelativeWorkingDirectory(
                working_dir.to_path_buf(),
            ));
        }
        let directories = minimal_manifest_directories(directories);
        if files.len() + directories.len() > MAX_PREDICTED_INPUTS {
            return Err(CcBypassReason::TooManyInputs);
        }
        let working_dir = normalize_components(working_dir);
        let mut inputs = Vec::with_capacity(files.len() + directories.len());
        let mut total_bytes = 0_u64;
        // Stat everything first so one batched ledger lookup can stand in for
        // rereading headers the session already scanned and hashed. A ledger
        // entry in the cc scope was recorded after the timestamp-macro scan
        // passed, so a hit skips the scan for the same reason it skips the
        // hash: the identity says the contents have not changed since both
        // were established.
        let mut identified = Vec::with_capacity(files.len());
        for path in files {
            let metadata = std::fs::metadata(&path).map_err(|error| CcBypassReason::InputRead {
                path: path.clone(),
                message: error.to_string(),
            })?;
            if !metadata.is_file() {
                return Err(CcBypassReason::InputRead {
                    path,
                    message: "input is not a regular file".into(),
                });
            }
            total_bytes = total_bytes.saturating_add(metadata.len());
            if total_bytes > MAX_INPUT_BYTES {
                return Err(CcBypassReason::TooManyInputs);
            }
            let identity = FileIdentity::for_digest_cache(&path, &metadata).map_err(|error| {
                CcBypassReason::InputRead {
                    path: path.clone(),
                    message: error.to_string(),
                }
            })?;
            identified.push((path, identity));
        }
        let queries = identified
            .iter()
            .filter_map(|(_, identity)| identity.clone())
            .collect::<Vec<_>>();
        let mut recorded = digests
            .resolve(FileDigestScope::CcInput, &queries)
            .into_iter();
        let mut identities = Vec::with_capacity(inputs.capacity());
        let mut fresh = Vec::new();
        for (path, identity) in identified {
            let resolution = identity
                .as_ref()
                .and_then(|_| recorded.next())
                .unwrap_or(FileDigestResolution::Unresolved);
            let cached = match resolution {
                FileDigestResolution::Digest(digest)
                    if identity.as_ref().is_some_and(|identity| {
                        identity.len == digest.size && identity.still_describes().unwrap_or(false)
                    }) =>
                {
                    Some(Ok(digest))
                }
                FileDigestResolution::EmbeddedTimestampMacro
                    if identity
                        .as_ref()
                        .is_some_and(|identity| identity.still_describes().unwrap_or(false)) =>
                {
                    Some(Err(CcBypassReason::EmbeddedTimestampMacro(path.clone())))
                }
                FileDigestResolution::Digest(_)
                | FileDigestResolution::EmbeddedTimestampMacro
                | FileDigestResolution::Unresolved => None,
            };
            let digest = if let Some(cached) = cached {
                identities.push(identity);
                cached?
            } else {
                let observed =
                    digest_file_validated(FileDigestScope::CcInput, &path).map_err(|error| {
                        CcBypassReason::InputRead {
                            path: path.clone(),
                            message: error.to_string(),
                        }
                    })?;
                let (resolution, observed_identity) = if let Some(observed) = observed {
                    let identity = observed.cache_identity.clone();
                    if let FileDigestResolution::Digest(digest) = &observed.resolution
                        && let Some(file) = observed.cache_identity
                        && file.len == digest.size
                    {
                        fresh.push(RecordedFileDigest {
                            file,
                            digest: digest.clone(),
                        });
                    }
                    (observed.resolution, identity)
                } else {
                    let resolution =
                        digest_file(FileDigestScope::CcInput, &path).map_err(|error| {
                            CcBypassReason::InputRead {
                                path: path.clone(),
                                message: error.to_string(),
                            }
                        })?;
                    (resolution, None)
                };
                let FileDigestResolution::Digest(digest) = resolution else {
                    return Err(CcBypassReason::EmbeddedTimestampMacro(path));
                };
                identities.push(observed_identity);
                digest
            };
            inputs.push(CcActionInput { path, digest });
        }
        if !fresh.is_empty() {
            digests.record(FileDigestScope::CcInput, fresh);
        }
        let mut manifest_entries = 0_usize;
        for directory in directories {
            let digest = include_manifest(&directory, &mut manifest_entries)?;
            inputs.push(CcActionInput {
                path: PathBuf::from(format!("{INCLUDE_MANIFEST_PREFIX}{}", directory.display())),
                digest,
            });
            identities.push(None);
        }
        Ok(Self {
            working_dir,
            inputs,
            identities,
        })
    }

    /// File inputs, excluding include-directory manifests.
    pub fn files(&self) -> impl Iterator<Item = &CcActionInput> {
        self.inputs
            .iter()
            .filter(|input| !is_manifest_input(&input.path))
    }

    /// Reject inputs whose modification time overlaps the compiler invocation.
    ///
    /// Contents are hashed after the compiler reports the paths it read. This
    /// timestamp barrier prevents a write that landed during the compile from
    /// being mistaken for the contents that produced the object; `verify`
    /// closes the remaining race after hashing.
    pub fn verify_not_modified_since(&self, started_at: SystemTime) -> Result<(), CcBypassReason> {
        self.verify_not_modified_since_with_snapshots(started_at, &BTreeMap::new())
    }

    /// Reject inputs that changed from snapshots captured before the driver
    /// ran, falling back to the wall-clock barrier for discovered headers.
    pub fn verify_not_modified_since_with_snapshots(
        &self,
        started_at: SystemTime,
        before: &BTreeMap<PathBuf, FileSnapshot>,
    ) -> Result<(), CcBypassReason> {
        for input in self.files() {
            if let Some(previous) = before.get(&input.path)
                && previous.proves_content_change()
            {
                let metadata =
                    std::fs::metadata(&input.path).map_err(|error| CcBypassReason::InputRead {
                        path: input.path.clone(),
                        message: error.to_string(),
                    })?;
                let identity = FileIdentity::for_digest_cache(&input.path, &metadata)
                    .map_err(|error| CcBypassReason::InputRead {
                        path: input.path.clone(),
                        message: error.to_string(),
                    })?
                    .or_else(|| FileIdentity::describe(&input.path, &metadata));
                if previous.matches(identity.as_ref(), &input.digest) {
                    continue;
                }
                return Err(CcBypassReason::InputModifiedDuringCompilation(
                    input.path.clone(),
                ));
            }
            let metadata =
                std::fs::metadata(&input.path).map_err(|error| CcBypassReason::InputRead {
                    path: input.path.clone(),
                    message: error.to_string(),
                })?;
            let modified = metadata
                .modified()
                .map_err(|error| CcBypassReason::InputRead {
                    path: input.path.clone(),
                    message: error.to_string(),
                })?;
            if modified >= started_at {
                return Err(CcBypassReason::InputModifiedDuringCompilation(
                    input.path.clone(),
                ));
            }
        }
        Ok(())
    }

    /// Compatibility form for callers that captured metadata identities.
    pub fn verify_not_modified_since_with_identities(
        &self,
        started_at: SystemTime,
        before: &BTreeMap<PathBuf, FileIdentity>,
    ) -> Result<(), CcBypassReason> {
        let snapshots = before
            .iter()
            .map(|(path, identity)| (path.clone(), identity.clone().into()))
            .collect();
        self.verify_not_modified_since_with_snapshots(started_at, &snapshots)
    }

    /// Confirm every discovered file before publication, degrading a changed
    /// input to a miss rather than storing an object under a stale key.
    ///
    /// A file still wearing the identity `collect` recorded is confirmed by
    /// that stat alone where the identity carries a change time, which cannot
    /// be set from user space and so shows a rewrite that restored the old
    /// modification time. One whose identity moved, that had none, or that a
    /// platform without change times described, is read and hashed again.
    pub fn verify(&self) -> Result<(), CcBypassReason> {
        for (index, input) in self.inputs.iter().enumerate() {
            if is_manifest_input(&input.path) {
                continue;
            }
            let read_error = |error: std::io::Error| CcBypassReason::InputRead {
                path: input.path.clone(),
                message: error.to_string(),
            };
            if let Some(Some(identity)) = self.identities.get(index)
                && identity.can_skip_content_verification()
                && identity.still_describes().map_err(read_error)?
            {
                continue;
            }
            let matches = input.digest.matches_file(&input.path).map_err(|error| {
                CcBypassReason::InputRead {
                    path: input.path.clone(),
                    message: error.to_string(),
                }
            })?;
            if !matches {
                return Err(CcBypassReason::InputChanged(input.path.clone()));
            }
        }
        Ok(())
    }

    /// Merge the manifest into an action context after confirming both use the
    /// same compiler working directory.
    pub fn apply_to(self, context: &mut CcActionContext) -> Result<(), CcBypassReason> {
        if normalize_components(&context.working_dir) != self.working_dir {
            return Err(CcBypassReason::DiscoveryWorkingDirectory);
        }
        context.inputs.extend(self.inputs);
        Ok(())
    }
}

/// Drop include directories already covered by an ancestor's recursive manifest.
///
/// Discovered headers often contribute hundreds of nested parent directories,
/// especially for amalgamated C sources. Keeping both an ancestor and its
/// descendants walks and hashes the same subtree repeatedly, and can exhaust
/// the manifest-entry budget even though the ancestor already names every
/// includable file below it.
fn minimal_manifest_directories(directories: BTreeSet<PathBuf>) -> Vec<PathBuf> {
    let mut directories = directories
        .into_iter()
        .map(|directory| {
            let normalized = manifest_directory_identity(&directory);
            (directory, normalized)
        })
        .collect::<Vec<_>>();
    directories.sort_by(|(left, left_normalized), (right, right_normalized)| {
        left_normalized
            .as_deref()
            .unwrap_or(left)
            .components()
            .count()
            .cmp(
                &right_normalized
                    .as_deref()
                    .unwrap_or(right)
                    .components()
                    .count(),
            )
            .then_with(|| left_normalized.cmp(right_normalized))
            .then_with(|| left.cmp(right))
    });

    let mut minimal = Vec::<(PathBuf, Option<PathBuf>)>::new();
    for (directory, normalized) in directories {
        if !minimal
            .iter()
            .any(|(_, ancestor)| match (ancestor, &normalized) {
                (Some(ancestor), Some(normalized)) => manifest_covers(ancestor, normalized),
                _ => false,
            })
        {
            minimal.push((directory, normalized));
        }
    }
    minimal
        .into_iter()
        .map(|(directory, _)| directory)
        .collect()
}

/// Lexical parent removal is safe only when the removed directory exists and
/// is not a symlink: `link/..` can name a completely different physical tree.
/// An unprovable path keeps its own manifest instead of joining another root.
fn manifest_directory_identity(directory: &Path) -> Option<PathBuf> {
    let mut normalized = PathBuf::new();
    for component in directory.components() {
        match component {
            std::path::Component::CurDir => {}
            std::path::Component::ParentDir => {
                let metadata = std::fs::symlink_metadata(&normalized).ok()?;
                if !metadata.is_dir() || metadata.file_type().is_symlink() {
                    return None;
                }
                normalized.pop();
            }
            component => normalized.push(component.as_os_str()),
        }
    }
    Some(normalized)
}

/// Whether walking `ancestor` recursively is guaranteed to visit `descendant`.
///
/// Component-aware normalization rejects a lexical prefix that escapes through
/// `..`. Directory symlinks need an explicit check because `read_dir` follows
/// the directory it starts at but the recursive walk deliberately does not
/// follow symlink entries beneath it.
fn manifest_covers(ancestor: &Path, descendant: &Path) -> bool {
    let Ok(relative) = descendant.strip_prefix(ancestor) else {
        return false;
    };
    if relative.as_os_str().is_empty() {
        return true;
    }
    let mut current = ancestor.to_path_buf();
    for component in relative.components() {
        current.push(component);
        let Ok(metadata) = std::fs::symlink_metadata(&current) else {
            return false;
        };
        if !metadata.is_dir() || metadata.file_type().is_symlink() {
            return false;
        }
    }
    true
}

fn is_manifest_input(path: &Path) -> bool {
    path.to_str()
        .is_some_and(|path| path.starts_with(INCLUDE_MANIFEST_PREFIX))
}

/// Digest the includable names in each directory, reading no file contents.
///
/// Taken once before the compiler runs and again before publishing, this is
/// what detects a header that appeared in a search directory *while* the
/// compilation was in flight. The manifest recorded in the key is the one from
/// after the compile, and without this check that later state would be claimed
/// as the state the compiler saw.
pub fn manifest_snapshot(
    directories: &BTreeSet<PathBuf>,
) -> Result<BTreeMap<PathBuf, CacheDigest>, CcBypassReason> {
    let mut budget = 0_usize;
    minimal_manifest_directories(directories.iter().cloned().collect())
        .into_iter()
        .map(|directory| {
            include_manifest(&directory, &mut budget).map(|digest| (directory, digest))
        })
        .collect()
}

/// Extensions a file must carry to be a plausible `#include` target.
///
/// An extensionless name also qualifies: C++ standard headers are spelled that
/// way and projects ship their own.
///
/// `gch` and `pch` are here because a precompiled header answers an `#include`
/// without being named by one. GCC prefers `foo.h.gch` over `foo.h` on its own,
/// with nothing on the command line to say so, which is precisely the
/// substitution these manifests exist to notice -- and the one case the
/// adapter's explicit precompiled-header bypass cannot see.
const INCLUDABLE_EXTENSIONS: &[&str] = &[
    "c", "c++", "cc", "cpp", "cxx", "def", "gch", "h", "h++", "hh", "hpp", "hxx", "inc", "inl",
    "ipp", "pch", "s", "tcc",
];

/// Whether a file name could be what an `#include` directive names.
///
/// The manifest exists to notice a file appearing where it would shadow a
/// header that was read. A build writes its own objects, dependency files, and
/// archives into these directories -- often the very directory a generated
/// header lives in -- and none of those can shadow an include. Counting them
/// would make the key depend on how many sibling compilations had finished,
/// which is not a property of this compilation at all.
fn is_includable(name: &str) -> bool {
    match name.rsplit_once('.') {
        Some((stem, extension)) if !stem.is_empty() => INCLUDABLE_EXTENSIONS
            .binary_search(&extension.to_ascii_lowercase().as_str())
            .is_ok(),
        // No extension, or a leading-dot name like `.keep`.
        _ => !name.starts_with('.'),
    }
}

/// Digest the sorted includable file names beneath a directory.
///
/// Names only: the contents of anything actually read are digested as inputs,
/// so this exists purely to notice a file appearing where it could shadow one
/// of them. A directory that does not exist has an empty manifest, which is
/// what makes "the directory was created" a key change rather than an error.
///
/// Every call walks the directory afresh. Directory timestamps advance in
/// coarse ticks on some filesystems, so unchanged metadata cannot prove that no
/// header appeared or vanished; a cached answer would make the snapshots taken
/// before and after a compilation agree even when the compiler saw a different
/// header search result.
fn include_manifest(directory: &Path, budget: &mut usize) -> Result<CacheDigest, CcBypassReason> {
    include_manifest_with(directory, budget, &|path| std::fs::read_dir(path))
}

fn include_manifest_with(
    directory: &Path,
    budget: &mut usize,
    read_dir: &impl Fn(&Path) -> std::io::Result<std::fs::ReadDir>,
) -> Result<CacheDigest, CcBypassReason> {
    let mut names = Vec::new();
    let mut pending = vec![(directory.to_path_buf(), String::new())];
    while let Some((current, prefix)) = pending.pop() {
        let entries = match read_dir(&current) {
            Ok(entries) => entries,
            Err(error) if error.kind() == std::io::ErrorKind::NotFound => continue,
            Err(error) => {
                return Err(CcBypassReason::InputRead {
                    path: current,
                    message: error.to_string(),
                });
            }
        };
        for entry in entries {
            let entry = entry.map_err(|error| CcBypassReason::InputRead {
                path: current.clone(),
                message: error.to_string(),
            })?;
            let name = entry.file_name();
            let Some(name) = name.to_str() else {
                return Err(CcBypassReason::NonUtf8Path(entry.path()));
            };
            let relative = if prefix.is_empty() {
                name.to_string()
            } else {
                format!("{prefix}/{name}")
            };
            let file_type = entry
                .file_type()
                .map_err(|error| CcBypassReason::InputRead {
                    path: entry.path(),
                    message: error.to_string(),
                })?;
            if file_type.is_dir() {
                pending.push((entry.path(), relative));
                continue;
            }
            if !is_includable(name) {
                continue;
            }
            *budget += 1;
            if *budget > MAX_MANIFEST_ENTRIES {
                return Err(CcBypassReason::TooManyInputs);
            }
            names.push(relative);
        }
    }
    names.sort();
    let digest = CacheDigest::blake3(names.join("\n").as_bytes());
    Ok(digest)
}

/// Whether a preprocessor input can make the assembler read another file.
///
/// Searching for the directive text, including in comments and inactive
/// conditional branches, deliberately accepts false positives. Missing a real
/// directive would publish an object whose complete inputs are absent from the
/// key; bypassing an otherwise cacheable object is the safe outcome instead.
pub(crate) fn contains_assembler_input_directive(path: &Path) -> Result<bool, CcBypassReason> {
    let needles = ASSEMBLER_INPUT_DIRECTIVES;
    let file = std::fs::File::open(path).map_err(|error| CcBypassReason::InputRead {
        path: path.to_path_buf(),
        message: error.to_string(),
    })?;
    let longest = needles
        .iter()
        .map(|needle| needle.len())
        .max()
        .unwrap_or_default();
    let mut reader = std::io::BufReader::new(file);
    let mut window = Vec::with_capacity(SCAN_CHUNK_BYTES + longest);
    let mut chunk = vec![0_u8; SCAN_CHUNK_BYTES];
    loop {
        let read = reader
            .read(&mut chunk)
            .map_err(|error| CcBypassReason::InputRead {
                path: path.to_path_buf(),
                message: error.to_string(),
            })?;
        if read == 0 {
            return Ok(false);
        }
        window.extend_from_slice(&chunk[..read]);
        // Every directive begins with a dot and an ASCII letter. Search for
        // that letter in either case, then check the dot and complete token.
        // Searching for the letter also skips runs of dots in assembly labels.
        if needles.iter().any(|needle| {
            memchr::memchr2_iter(needle[1], needle[1].to_ascii_uppercase(), &window).any(|offset| {
                offset.checked_sub(1).is_some_and(|start| {
                    window[start..]
                        .get(..needle.len())
                        .is_some_and(|candidate| candidate.eq_ignore_ascii_case(needle))
                })
            })
        }) {
            return Ok(true);
        }
        let keep = window.len().saturating_sub(longest.saturating_sub(1));
        window.drain(..keep);
    }
}

#[cfg(test)]
#[path = "depfile_tests.rs"]
mod tests;

#[cfg(test)]
mod plain_depfile_tests {
    use super::*;

    #[test]
    fn plain_paths_preserve_whitespace_unicode_and_make_escapes() {
        for input in [
            "",
            " \t ",
            "src/main.c include/header.h",
            "\t路径/é.c  a#b.h\t",
            "a\u{a0}b.h",
            r"a\ b.h a\#b.h a$$b.h",
            r"bad\q.h",
            "bad$",
            "bad\\",
        ] {
            let expected = split_escaped_prerequisites(input);
            let actual = split_prerequisites(input);
            match (actual, expected) {
                (Ok(actual), Ok(expected)) => assert_eq!(actual, expected, "{input}"),
                (Err(actual), Err(expected)) => {
                    assert_eq!(format!("{actual:?}"), format!("{expected:?}"))
                }
                results => panic!("mismatched results for {input}: {results:?}"),
            }
        }
    }

    #[test]
    #[ignore = "manual performance measurement; run in release mode"]
    fn benchmark_plain_depfile() {
        for count in [256, 4096] {
            for escaped in [false, true] {
                let prerequisites = (0..count)
                    .map(|index| {
                        format!(
                            " /workspace/{}/include/library-{index}/header-{index}.h",
                            if escaped {
                                r"some\ directory"
                            } else {
                                "some-directory"
                            }
                        )
                    })
                    .collect::<String>();
                let mut before = Vec::new();
                let mut after = Vec::new();
                for round in 0..15 {
                    for optimized in if round % 2 == 0 {
                        [false, true]
                    } else {
                        [true, false]
                    } {
                        let start = std::time::Instant::now();
                        for _ in 0..50 {
                            let result = if optimized {
                                split_prerequisites(std::hint::black_box(&prerequisites))
                            } else {
                                split_escaped_prerequisites(std::hint::black_box(&prerequisites))
                            };
                            std::hint::black_box(result.unwrap());
                        }
                        let sample = start.elapsed().as_secs_f64() * 1e6 / 50.0;
                        if optimized {
                            after.push(sample);
                        } else {
                            before.push(sample);
                        }
                    }
                }
                before.sort_by(f64::total_cmp);
                after.sort_by(f64::total_cmp);
                eprintln!(
                    "{count} paths, escaped={escaped}: before {:.3} us, after {:.3} us",
                    before[7], after[7]
                );
            }
        }
    }
}