grit_lib/diff/

changes.rs

//! Staged, unstaged, and working-tree changes with both sides' contents loaded.
//!
//! [`UncommittedDiff`] collects [`DiffEntry`] rows for the index against `HEAD`, the working
//! tree against the index, or the working tree against any tree, reusing one loaded index and
//! one set of [`WorktreeRules`] for every step; [`UncommittedDiff::entries`] picks the comparison
//! from a [`DiffTarget`] and applies pathspecs and `--diff-filter`.
//! [`UncommittedDiff::load_worktree_changes`] and
//! [`load_object_changes`] then turn entries into [`FileChange`]s whose bytes feed structured
//! hunks ([`FileChange::structured`]) or line counts ([`FileChange::line_stat`]).

use std::collections::{BTreeSet, HashSet};
use std::path::Path;
use std::sync::{Arc, Mutex};

use crate::diff::structured::{structured_file_diff, SideBytes, StructuredFileDiff};
use crate::diff::targets::{
    diff_index_to_tree_for_paths, limit_target_entries, stored_target_entries,
};
use crate::diff::tree_content::{gitlink_line, load_new_side, load_old_side};
use crate::diff::{
    apply_rename_detection_to_diff_entries, diff_index_to_tree_with_renames,
    diff_index_to_worktree_with_renames, diff_tree_to_worktree_with_git_dir_and_rules_extras,
    index_mtime_for_diff, line_change_counts, load_worktree_side_for_diff,
    peer_scope_includes_path, read_submodule_head_oid, sort_entries_by_path,
    submodule_porcelain_flags, DiffEntry, DiffIndexToWorktreeOptions, DiffOptions, DiffSide,
    DiffStatus, DiffTarget, DiffTreeToWorktreeExtras, LineStat, RenameDetection,
    RENAME_DETECTION_DISABLED,
};
use crate::error::{Error, Result};
use crate::index::write_tree::cache_tree_extension_valid;
use crate::index::Index;
use crate::objects::odb::Odb;
use crate::objects::{FileMode, ObjectId};
use crate::parallel::{try_par_map, ParallelHashError, ParallelOp};
use crate::repo::Repository;
use crate::state::resolve_head_for_repository;
use crate::worktree::fsmonitor::fsmonitor_enabled;
use crate::worktree::worktree_rules::WorktreeRules;

/// One changed path with both sides' bytes loaded for diffing.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct FileChange {
    /// Display path (new path, or old path for a deletion).
    pub path: String,
    /// Pre-rename path, only when it differs from [`Self::path`].
    pub old_path: Option<String>,
    /// Change kind for this path.
    pub status: DiffStatus,
    /// Old side mode ([`FileMode::ABSENT`] when the path is added).
    pub old_mode: FileMode,
    /// New side mode ([`FileMode::ABSENT`] when the path is deleted).
    pub new_mode: FileMode,
    /// Old side object id (zero when absent).
    pub old_oid: ObjectId,
    /// New side object id (zero when absent or not yet hashed from the working tree).
    pub new_oid: ObjectId,
    /// Old side contents (a `Subproject commit` line for gitlinks).
    pub old_bytes: SideBytes,
    /// New side contents, read from the working tree for working-tree changes.
    pub new_bytes: SideBytes,
    /// Either side is treated as binary; a binary change may carry neither side's bytes.
    pub binary: bool,
    /// The object ids differ (false for a mode-only change).
    pub content_changed: bool,
    /// Rename or copy similarity (0–100), when applicable.
    pub similarity: Option<u32>,
}

impl FileChange {
    fn from_entry(
        entry: DiffEntry,
        old_bytes: SideBytes,
        new_bytes: SideBytes,
        binary: bool,
    ) -> Self {
        let path = entry
            .new_path
            .clone()
            .or_else(|| entry.old_path.clone())
            .unwrap_or_default();
        let similarity = entry.rename_similarity();
        // A deletion's display path is its old path; only a real rename keeps `old_path`.
        let old_path = entry.old_path.filter(|old| *old != path);
        Self {
            path,
            old_path,
            status: entry.status,
            old_mode: entry.old_mode,
            new_mode: entry.new_mode,
            old_oid: entry.old_oid,
            new_oid: entry.new_oid,
            old_bytes,
            new_bytes,
            binary,
            content_changed: entry.old_oid != entry.new_oid,
            similarity,
        }
    }

    /// Structured line hunks and mode metadata for this change.
    ///
    /// Binary changes get mode metadata and no hunks.
    ///
    /// # Parameters
    ///
    /// - `options` - line-diff settings (context, algorithm, whitespace handling).
    #[must_use]
    pub fn structured(&self, options: DiffOptions) -> StructuredFileDiff {
        if self.binary {
            return StructuredFileDiff {
                old_mode: self.old_mode.octal_if_present(),
                new_mode: self.new_mode.octal_if_present(),
                old_encoding_lossy: self.old_bytes.encoding_lossy,
                new_encoding_lossy: self.new_bytes.encoding_lossy,
                hunks: Vec::new(),
            };
        }
        structured_file_diff(
            self.old_mode,
            self.new_mode,
            &self.old_bytes,
            &self.new_bytes,
            options,
        )
    }

    /// Inserted and deleted line counts, or `None` for a binary change.
    ///
    /// # Parameters
    ///
    /// - `options` - line-diff settings; whitespace modes change what counts as a changed line.
    #[must_use]
    pub fn line_stat(&self, options: &DiffOptions) -> Option<LineStat> {
        if self.binary {
            return None;
        }
        Some(line_change_counts(
            &self.old_bytes.text(),
            &self.new_bytes.text(),
            options,
        ))
    }
}

/// Load both sides of each entry from the object database.
///
/// Use this for index-against-tree and tree-against-tree entries, whose sides are all stored
/// objects. Entries are loaded in parallel per the repository's thread policy; the result keeps
/// the order of `entries`.
///
/// # Parameters
///
/// - `repo` - repository whose object database holds both sides.
/// - `entries` - diff rows to load, e.g. from [`UncommittedDiff::staged_entries`].
///
/// # Errors
///
/// Returns config load failures and object read errors.
pub fn load_object_changes(repo: &Repository, entries: &[DiffEntry]) -> Result<Vec<FileChange>> {
    let config = repo.config()?;
    let par_opts = repo
        .thread_policy()
        .resolve(ParallelOp::DiffContent, config.as_ref());
    let odb = repo.odb();
    let est_bytes = entries.len().saturating_mul(8192);
    try_par_map(entries, par_opts, est_bytes, |entry| {
        load_change(odb, entry.clone(), None)
    })
    .map_err(|ParallelHashError::Task(err)| err)
}

/// Where a working-tree change's new side is read from.
pub(crate) struct WorktreeSide<'a> {
    pub(crate) work_tree: &'a Path,
    pub(crate) index: &'a Index,
    pub(crate) rules: &'a Mutex<WorktreeRules>,
}

/// Load one entry's sides: the old side from the object database, the new side from the
/// working tree when `worktree` is given, otherwise from the object database.
///
/// A binary new side makes the whole change binary, so the old blob is then not read at all.
pub(crate) fn load_change(
    odb: &Odb,
    entry: DiffEntry,
    worktree: Option<&WorktreeSide<'_>>,
) -> Result<FileChange> {
    let (new_bytes, new_binary) = match worktree {
        Some(side) => load_worktree_new_side(odb, &entry, side)?,
        None => load_new_side(odb, &entry, None)?,
    };
    let (old_bytes, old_binary) = if new_binary {
        (SideBytes::from_bytes(Vec::new()), false)
    } else {
        load_old_side(odb, &entry, None)?
    };
    Ok(FileChange::from_entry(
        entry,
        old_bytes,
        new_bytes,
        old_binary || new_binary,
    ))
}

/// Load every entry with [`load_change`], in order.
pub(crate) fn load_changes(
    odb: &Odb,
    entries: Vec<DiffEntry>,
    worktree: Option<&WorktreeSide<'_>>,
) -> Result<Vec<FileChange>> {
    entries
        .into_iter()
        .map(|entry| load_change(odb, entry, worktree))
        .collect()
}

fn load_worktree_new_side(
    odb: &Odb,
    entry: &DiffEntry,
    side: &WorktreeSide<'_>,
) -> Result<(SideBytes, bool)> {
    if entry.new_mode == FileMode::GITLINK {
        let line = worktree_gitlink_line(entry, side.work_tree);
        return Ok((SideBytes::from_bytes(line.into_bytes()), false));
    }
    if entry.new_mode == FileMode::ABSENT {
        return Ok((SideBytes::from_bytes(Vec::new()), false));
    }
    let Some(path) = entry.new_path.as_deref().or(entry.old_path.as_deref()) else {
        return load_new_side(odb, entry, None);
    };
    // Working-tree entries often carry only a hash of the normalized file, and the blob is not
    // in the object database until it is staged, so always read the live file.
    let index_entry = side.index.get(&crate::repo_relpath::storage_bytes(path), 0);
    let rules = side.rules.lock().unwrap_or_else(|e| e.into_inner());
    load_worktree_side_for_diff(
        odb,
        side.work_tree,
        path,
        entry.new_mode.bits(),
        rules.config(),
        index_entry,
        Some(&rules),
        false,
    )
}

/// The submodule's checked-out commit, with `-dirty` when its tracked content is modified.
fn worktree_gitlink_line(entry: &DiffEntry, work_tree: &Path) -> String {
    let path = entry.new_path.as_deref().or(entry.old_path.as_deref());
    let head = if entry.new_oid != ObjectId::zero() {
        Some(entry.new_oid)
    } else {
        path.and_then(|p| read_submodule_head_oid(&work_tree.join(p)))
    };
    let Some(head) = head.filter(|oid| *oid != ObjectId::zero()) else {
        return String::new();
    };
    let dirty = entry.old_mode == FileMode::GITLINK
        && path.is_some_and(|p| submodule_porcelain_flags(work_tree, p, entry.old_oid).modified);
    if dirty {
        format!("Subproject commit {}-dirty\n", head.to_hex())
    } else {
        gitlink_line(&head)
    }
}

/// Paths to scan before a worktree or index diff when `pathspecs` limit the operation.
///
/// Includes index entries selected by the pathspecs plus blob paths in `base_tree` that match,
/// so a staged deletion (absent from the index but present in `HEAD`) is not dropped.
pub(crate) fn pathspec_restrict_paths_for_scan(
    repo: &Repository,
    index: &Index,
    pathspecs: &[String],
    base_tree: Option<&ObjectId>,
) -> Result<Option<Arc<BTreeSet<String>>>> {
    if pathspecs.is_empty() {
        return Ok(None);
    }
    let base_ctx = repo.pathspec_match_context();
    let mut paths = crate::porcelain::add::restricted_index_paths_for_pathspecs(
        repo, index, pathspecs, base_ctx,
    )?;
    if let Some(tree) = base_tree {
        let attr_matcher = if pathspecs
            .iter()
            .any(|s| crate::pathspec::pathspec_has_attr_magic(s))
        {
            Some(repo.pathspec_matcher(pathspecs)?)
        } else {
            None
        };
        for (path, mode, _) in super::flatten_tree_entries(repo.odb(), tree)? {
            if crate::porcelain::add::blob_path_matches_restrict_pathspecs(
                &path,
                mode.bits(),
                pathspecs,
                base_ctx,
                attr_matcher.as_ref(),
            )? {
                paths.insert(path);
            }
        }
    }
    Ok(Some(Arc::new(paths)))
}

fn rename_threshold(renames: RenameDetection) -> u32 {
    renames
        .threshold_percent()
        .unwrap_or(RENAME_DETECTION_DISABLED)
}

/// Index against `HEAD`'s tree, sorted by path.
///
/// A non-empty `restrict_to_paths` compares only those exact paths, without rename detection.
pub(crate) fn staged_entries_in(
    repo: &Repository,
    index: &Index,
    rename_threshold: u32,
    restrict_to_paths: Option<&BTreeSet<String>>,
) -> Result<Vec<DiffEntry>> {
    let head_tree = resolve_head_for_repository(repo)?.tree(repo.odb())?;
    let mut entries = match restrict_to_paths.filter(|paths| !paths.is_empty()) {
        Some(paths) => diff_index_to_tree_for_paths(repo, index, head_tree.as_ref(), paths)?,
        None => {
            let config = repo.config()?;
            diff_index_to_tree_with_renames(
                repo.odb(),
                index,
                head_tree.as_ref(),
                false,
                rename_threshold,
                config.as_ref(),
                &repo.thread_policy(),
            )?
        }
    };
    sort_entries_by_path(&mut entries);
    Ok(entries)
}

/// Working tree against `index`, sorted by path.
///
/// `restrict_to_paths` limits the scan to those paths and skips the fsmonitor refresh. A file
/// whose size changed (under an identity clean conversion) is reported without being hashed,
/// with a null `new_oid`; loading its side reads the live file anyway.
pub(crate) fn unstaged_entries_in(
    repo: &Repository,
    index: &mut Index,
    work_tree: &Path,
    rules: &Arc<Mutex<WorktreeRules>>,
    rename_threshold: u32,
    restrict_to_paths: Option<Arc<BTreeSet<String>>>,
) -> Result<Vec<DiffEntry>> {
    let config = rules.lock().unwrap_or_else(|e| e.into_inner()).config_arc();
    let refresh_fsmonitor = restrict_to_paths.is_none() && fsmonitor_enabled(config.as_ref());
    let options = DiffIndexToWorktreeOptions {
        repository_git_dir: Some(repo.git_dir().to_path_buf()),
        index_mtime: index_mtime_for_diff(index, None),
        config: Some(config),
        thread_policy: Some(repo.thread_policy()),
        refresh_index_stat_in_pass: false,
        size_change_proves_modified: true,
        refresh_fsmonitor,
        worktree_rules: Some(Arc::clone(rules)),
        command_runner: Some(repo.command_runner()),
        restrict_to_paths,
        skip_inaccessible_paths: true,
        diagnostics: Some(repo.diagnostics()),
        ..DiffIndexToWorktreeOptions::default()
    };
    let (mut entries, _) =
        diff_index_to_worktree_with_renames(repo, index, work_tree, options, rename_threshold)?;
    sort_entries_by_path(&mut entries);
    Ok(entries)
}

/// Working tree (through `index`) against `base_tree`, sorted by path.
///
/// When the index's cache tree matches `base_tree`, this is the cheaper index-against-worktree
/// scan; otherwise the tree is walked against the working tree directly.
pub(crate) fn worktree_entries_in(
    repo: &Repository,
    index: &mut Index,
    work_tree: &Path,
    rules: &Arc<Mutex<WorktreeRules>>,
    base_tree: Option<&ObjectId>,
    rename_threshold: u32,
    restrict_to_paths: Option<Arc<BTreeSet<String>>>,
) -> Result<Vec<DiffEntry>> {
    let index_matches_base = match (base_tree, index.cache_tree_root.as_ref()) {
        (Some(base), Some(cached)) => base == cached,
        (None, None) => true,
        _ => false,
    };
    if index_matches_base && cache_tree_extension_valid(index.cache_tree.as_ref()) {
        return unstaged_entries_in(
            repo,
            index,
            work_tree,
            rules,
            rename_threshold,
            restrict_to_paths,
        );
    }
    let config = repo.config()?;
    let raw = {
        let rules = rules.lock().unwrap_or_else(|e| e.into_inner());
        diff_tree_to_worktree_with_git_dir_and_rules_extras(
            repo.odb(),
            base_tree,
            work_tree,
            repo.git_dir(),
            index,
            Some(config.as_ref()),
            Some(&rules),
            DiffTreeToWorktreeExtras {
                skip_inaccessible_paths: true,
                diagnostics: Some(repo.diagnostics()),
                restrict_to_paths: restrict_to_paths.clone(),
            },
        )?
    };
    let mut entries = apply_rename_detection_to_diff_entries(
        repo.odb(),
        raw,
        Some(work_tree),
        rename_threshold,
        config.as_ref(),
        &repo.thread_policy(),
    )?;
    sort_entries_by_path(&mut entries);
    Ok(entries)
}

/// Staged, unstaged, and working-tree diffs over one loaded index.
///
/// Every method reuses the same [`Index`] and, once a working-tree method needs them, the same
/// [`WorktreeRules`], so collecting entries and then loading their contents reads neither twice.
/// Working-tree scans may refresh stat data on the in-memory index; nothing is written to disk.
///
/// Obtain one from [`Diff::uncommitted`](crate::diff::Diff::uncommitted) or
/// [`UncommittedDiff::new`].
pub struct UncommittedDiff<'repo> {
    repo: &'repo Repository,
    index: Index,
    rules: Option<Arc<Mutex<WorktreeRules>>>,
}

impl<'repo> UncommittedDiff<'repo> {
    /// Diff `repo` against an index the caller already loaded.
    ///
    /// # Parameters
    ///
    /// - `repo` - repository to diff.
    /// - `index` - its index (normally the on-disk index).
    #[must_use]
    pub fn new(repo: &'repo Repository, index: Index) -> Self {
        Self {
            repo,
            index,
            rules: None,
        }
    }

    /// The index every comparison uses.
    #[must_use]
    pub fn index(&self) -> &Index {
        &self.index
    }

    /// Index against `HEAD`'s tree (`grit diff --staged`), sorted by path.
    ///
    /// An unborn `HEAD` compares against the empty tree, so every index entry is added.
    ///
    /// # Parameters
    ///
    /// - `renames` - rename detection between deleted and added paths.
    ///
    /// # Errors
    ///
    /// Returns `HEAD` resolution, config, and object read errors.
    pub fn staged_entries(&self, renames: RenameDetection) -> Result<Vec<DiffEntry>> {
        self.staged_entries_for_paths(renames, None)
    }

    fn staged_entries_for_paths(
        &self,
        renames: RenameDetection,
        restrict_to_paths: Option<Arc<BTreeSet<String>>>,
    ) -> Result<Vec<DiffEntry>> {
        staged_entries_in(
            self.repo,
            &self.index,
            rename_threshold(renames),
            restrict_to_paths.as_deref(),
        )
    }

    /// Working tree against the index (`grit diff --unstaged`), sorted by path.
    ///
    /// Like Git's `diff-files`, a working-tree side that was not hashed carries a null
    /// `new_oid`; a tracked file whose size changed is reported that way without reading it.
    ///
    /// # Parameters
    ///
    /// - `renames` - rename detection between deleted and added paths.
    ///
    /// # Errors
    ///
    /// Returns [`Error::RequiresWorkTree`] in a bare repository, plus worktree rule, I/O, and
    /// object read errors.
    pub fn unstaged_entries(&mut self, renames: RenameDetection) -> Result<Vec<DiffEntry>> {
        self.unstaged_entries_for_paths(renames, None)
    }

    fn unstaged_entries_for_paths(
        &mut self,
        renames: RenameDetection,
        restrict_to_paths: Option<Arc<BTreeSet<String>>>,
    ) -> Result<Vec<DiffEntry>> {
        let work_tree = self.work_tree()?;
        let rules = self.rules()?;
        unstaged_entries_in(
            self.repo,
            &mut self.index,
            work_tree,
            &rules,
            rename_threshold(renames),
            restrict_to_paths,
        )
    }

    /// Working tree and index together against `base_tree` (`grit diff`, `grit diff --against`),
    /// sorted by path.
    ///
    /// # Parameters
    ///
    /// - `base_tree` - tree to compare with; `None` for the empty tree (an unborn `HEAD`).
    /// - `renames` - rename detection between deleted and added paths.
    ///
    /// # Errors
    ///
    /// Returns [`Error::RequiresWorkTree`] in a bare repository, plus worktree rule, I/O, and
    /// object read errors.
    pub fn worktree_entries(
        &mut self,
        base_tree: Option<&ObjectId>,
        renames: RenameDetection,
    ) -> Result<Vec<DiffEntry>> {
        self.worktree_entries_for_paths(base_tree, renames, None)
    }

    fn worktree_entries_for_paths(
        &mut self,
        base_tree: Option<&ObjectId>,
        renames: RenameDetection,
        restrict_to_paths: Option<Arc<BTreeSet<String>>>,
    ) -> Result<Vec<DiffEntry>> {
        let work_tree = self.work_tree()?;
        let rules = self.rules()?;
        worktree_entries_in(
            self.repo,
            &mut self.index,
            work_tree,
            &rules,
            base_tree,
            rename_threshold(renames),
            restrict_to_paths,
        )
    }

    /// Changed paths for any [`DiffTarget`], limited to `pathspecs` and to the statuses in
    /// `options.filter` (`--diff-filter`), sorted by path. No contents are loaded.
    ///
    /// [`DiffTarget::WorktreeVsHead`] compares with `HEAD`'s tree (the empty tree when unborn).
    /// Commit and tree targets ignore the index, so prefer [`Diff::entries`](crate::diff::Diff::entries)
    /// for them; it does not load one.
    ///
    /// # Parameters
    ///
    /// - `target` - what to compare, e.g. from [`resolve_diff_request`](crate::diff::resolve_diff_request).
    /// - `pathspecs` - repository-relative pathspecs; empty means every path.
    /// - `options` - only `renames` and `filter` are used.
    ///
    /// # Errors
    ///
    /// Returns [`Error::RequiresWorkTree`] for a working-tree target in a bare repository, plus
    /// `HEAD` resolution, pathspec, worktree rule, I/O, and object read errors.
    pub fn entries(
        &mut self,
        target: &DiffTarget,
        pathspecs: &[String],
        options: &DiffOptions,
    ) -> Result<super::core::DiffEntriesResult> {
        let renames = options.renames;
        let result = match target {
            DiffTarget::IndexVsHead => {
                let head_tree = resolve_head_for_repository(self.repo)?.tree(self.repo.odb())?;
                let restrict = pathspec_restrict_paths_for_scan(
                    self.repo,
                    &self.index,
                    pathspecs,
                    head_tree.as_ref(),
                )?;
                super::core::DiffEntriesResult {
                    entries: self.staged_entries_for_paths(renames, restrict)?,
                    inexact_rename_skipped: None,
                }
            }
            DiffTarget::WorktreeVsIndex => {
                let restrict =
                    pathspec_restrict_paths_for_scan(self.repo, &self.index, pathspecs, None)?;
                super::core::DiffEntriesResult {
                    entries: self.unstaged_entries_for_paths(renames, restrict)?,
                    inexact_rename_skipped: None,
                }
            }
            DiffTarget::WorktreeVsHead => {
                let head_tree = resolve_head_for_repository(self.repo)?.tree(self.repo.odb())?;
                let restrict = pathspec_restrict_paths_for_scan(
                    self.repo,
                    &self.index,
                    pathspecs,
                    head_tree.as_ref(),
                )?;
                super::core::DiffEntriesResult {
                    entries: self.worktree_entries_for_paths(
                        head_tree.as_ref(),
                        renames,
                        restrict,
                    )?,
                    inexact_rename_skipped: None,
                }
            }
            DiffTarget::WorktreeVsTree(tree) => {
                let restrict = pathspec_restrict_paths_for_scan(
                    self.repo,
                    &self.index,
                    pathspecs,
                    Some(tree),
                )?;
                super::core::DiffEntriesResult {
                    entries: self.worktree_entries_for_paths(Some(tree), renames, restrict)?,
                    inexact_rename_skipped: None,
                }
            }
            DiffTarget::Commit(_) | DiffTarget::Trees { .. } => {
                stored_target_entries(self.repo, target, pathspecs, options)?.unwrap_or_default()
            }
        };
        limit_target_entries(self.repo, result, pathspecs, options)
    }

    /// Changes on the other side of the index for the paths in `primary_paths`, which feed
    /// [`hunk_ids_with_duplicate_peers`](crate::diff::hunk_ids_with_duplicate_peers) so hunk ids
    /// stay stable when an identical hunk moves between the staged and unstaged sides.
    ///
    /// For [`DiffSide::Staged`] primaries this loads the unstaged changes on those paths; for
    /// [`DiffSide::Unstaged`] the staged ones. [`DiffSide::Other`] has no peer side. A path
    /// matches when it is the peer's display path or, for a rename, its old path. No
    /// `--diff-filter` applies: ids must not depend on which statuses are shown.
    ///
    /// # Parameters
    ///
    /// - `side` - the side the primary changes came from.
    /// - `pathspecs` - the pathspecs the primary changes were limited to.
    /// - `renames` - rename detection, as for the primary changes.
    /// - `primary_paths` - display paths of the primary changes.
    ///
    /// # Errors
    ///
    /// Returns [`Error::RequiresWorkTree`] for staged primaries in a bare repository, plus the
    /// errors of [`Self::entries`] and of loading the changes.
    pub fn peer_changes<'a>(
        &mut self,
        side: DiffSide,
        pathspecs: &[String],
        renames: RenameDetection,
        primary_paths: impl IntoIterator<Item = &'a str>,
    ) -> Result<Vec<FileChange>> {
        let peer_target = match side {
            DiffSide::Staged => DiffTarget::WorktreeVsIndex,
            DiffSide::Unstaged => DiffTarget::IndexVsHead,
            DiffSide::Other => return Ok(Vec::new()),
        };
        let primary: HashSet<&str> = primary_paths.into_iter().collect();
        if primary.is_empty() {
            return Ok(Vec::new());
        }
        let options = DiffOptions::default().with_renames(renames);
        let entries: Vec<DiffEntry> = self
            .entries(&peer_target, pathspecs, &options)?
            .entries
            .into_iter()
            .filter(|entry| {
                let display = entry
                    .new_path
                    .as_deref()
                    .or(entry.old_path.as_deref())
                    .unwrap_or("");
                peer_scope_includes_path(display, entry.old_path.as_deref(), &primary)
            })
            .collect();
        match side {
            DiffSide::Staged => self.load_worktree_changes(entries),
            DiffSide::Unstaged | DiffSide::Other => load_object_changes(self.repo, &entries),
        }
    }

    /// Load entries from [`Self::unstaged_entries`] or [`Self::worktree_entries`]: old sides
    /// from the object database, new sides from the working tree (with gitlinks showing the
    /// submodule's checked-out commit and `-dirty` state).
    ///
    /// # Errors
    ///
    /// Returns [`Error::RequiresWorkTree`] in a bare repository, plus I/O, filter, and object
    /// read errors.
    pub fn load_worktree_changes(&mut self, entries: Vec<DiffEntry>) -> Result<Vec<FileChange>> {
        let work_tree = self.work_tree()?;
        let rules = self.rules()?;
        let side = WorktreeSide {
            work_tree,
            index: &self.index,
            rules: &rules,
        };
        load_changes(self.repo.odb(), entries, Some(&side))
    }

    fn work_tree(&self) -> Result<&'repo Path> {
        self.repo.work_tree().ok_or(Error::RequiresWorkTree)
    }

    fn rules(&mut self) -> Result<Arc<Mutex<WorktreeRules>>> {
        if let Some(rules) = &self.rules {
            return Ok(Arc::clone(rules));
        }
        let rules = Arc::new(Mutex::new(WorktreeRules::from_repository(
            self.repo,
            &self.index,
        )?));
        self.rules = Some(Arc::clone(&rules));
        Ok(rules)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::repo::InitOptions;

    fn entry(status: DiffStatus, old: Option<&str>, new: Option<&str>) -> DiffEntry {
        DiffEntry {
            status,
            old_path: old.map(str::to_owned),
            new_path: new.map(str::to_owned),
            old_mode: FileMode::REGULAR,
            new_mode: FileMode::REGULAR,
            old_oid: ObjectId::zero(),
            new_oid: ObjectId::zero(),
            score: Some(87),
        }
    }

    #[test]
    fn from_entry_keeps_old_path_only_for_real_renames() {
        let bytes = || SideBytes::from_bytes(Vec::new());
        let renamed = FileChange::from_entry(
            entry(DiffStatus::Renamed, Some("a"), Some("b")),
            bytes(),
            bytes(),
            false,
        );
        assert_eq!(renamed.path, "b");
        assert_eq!(renamed.old_path.as_deref(), Some("a"));
        assert_eq!(renamed.similarity, Some(87));

        let deleted = FileChange::from_entry(
            entry(DiffStatus::Deleted, Some("a"), None),
            bytes(),
            bytes(),
            false,
        );
        assert_eq!(deleted.path, "a");
        assert_eq!(deleted.old_path, None);
        assert_eq!(deleted.similarity, None);
    }

    #[test]
    fn binary_change_has_no_hunks_or_line_stat() {
        let mut change = FileChange::from_entry(
            entry(DiffStatus::Modified, Some("a"), Some("a")),
            SideBytes::from_bytes(b"one\n".to_vec()),
            SideBytes::from_bytes(b"two\n".to_vec()),
            true,
        );
        assert!(change.structured(DiffOptions::default()).hunks.is_empty());
        assert_eq!(change.line_stat(&DiffOptions::default()), None);

        change.binary = false;
        assert_eq!(change.structured(DiffOptions::default()).hunks.len(), 1);
        assert_eq!(
            change.line_stat(&DiffOptions::default()),
            Some(LineStat {
                insertions: 1,
                deletions: 1,
            })
        );
    }

    #[test]
    fn worktree_gitlink_line_reads_submodule_head_when_unhashed() {
        let dir = tempfile::tempdir().unwrap_or_else(|e| panic!("tempdir: {e}"));
        let mut gitlink = entry(DiffStatus::Added, None, Some("missing-sub"));
        gitlink.old_mode = FileMode::ABSENT;
        gitlink.new_mode = FileMode::GITLINK;
        assert_eq!(worktree_gitlink_line(&gitlink, dir.path()), "");

        let oid = ObjectId::from_hex("1111111111111111111111111111111111111111")
            .unwrap_or_else(|e| panic!("oid: {e}"));
        gitlink.new_oid = oid;
        assert_eq!(
            worktree_gitlink_line(&gitlink, dir.path()),
            format!("Subproject commit {}\n", oid.to_hex())
        );
    }

    /// Repository whose index tracks `files`, staged from the working tree.
    fn staged_repo(files: &[(&str, &[u8])]) -> (tempfile::TempDir, Repository) {
        let dir = tempfile::tempdir().unwrap_or_else(|e| panic!("tempdir: {e}"));
        let repo = Repository::init(dir.path(), &InitOptions::default())
            .unwrap_or_else(|e| panic!("init: {e}"));
        for (path, bytes) in files {
            std::fs::write(dir.path().join(path), bytes).unwrap_or_else(|e| panic!("write: {e}"));
        }
        crate::porcelain::add::stage(
            &repo,
            &crate::porcelain::add::StageOptions::default(),
            &mut crate::progress::NullProgress,
            &crate::progress::NeverCancel,
        )
        .unwrap_or_else(|e| panic!("stage: {e}"));
        (dir, repo)
    }

    fn unstaged(repo: &Repository) -> Vec<DiffEntry> {
        let index = repo
            .index_handle()
            .load()
            .unwrap_or_else(|e| panic!("index: {e}"));
        UncommittedDiff::new(repo, index)
            .unstaged_entries(RenameDetection::Off)
            .unwrap_or_else(|e| panic!("unstaged: {e}"))
    }

    #[test]
    fn unstaged_size_change_is_reported_without_hashing() {
        let (dir, repo) = staged_repo(&[("grown.txt", b"one\n"), ("same.txt", b"aaaa\n")]);
        std::fs::write(dir.path().join("grown.txt"), b"one\ntwo\n")
            .unwrap_or_else(|e| panic!("{e}"));
        std::fs::write(dir.path().join("same.txt"), b"bbbb\n").unwrap_or_else(|e| panic!("{e}"));

        let metrics = repo.odb().hot_path_metrics();
        metrics.set_blob_counting(true);
        metrics.reset_blob_content_reads();
        let entries = unstaged(&repo);
        // Only the same-size file is read (to hash it); the size change alone proves the other.
        assert_eq!(metrics.blob_content_reads(), 1);
        assert_eq!(entries.len(), 2);
        let grown = &entries[0];
        assert_eq!(grown.path(), "grown.txt");
        assert_eq!(grown.status, DiffStatus::Modified);
        assert_eq!(grown.new_oid, ObjectId::zero());
        // A same-size edit is only reported once hashing proves the content changed.
        let same = &entries[1];
        assert_eq!(same.path(), "same.txt");
        assert_eq!(same.status, DiffStatus::Modified);

        let index = repo
            .index_handle()
            .load()
            .unwrap_or_else(|e| panic!("index: {e}"));
        let changes = UncommittedDiff::new(&repo, index)
            .load_worktree_changes(entries)
            .unwrap_or_else(|e| panic!("load: {e}"));
        assert!(changes[0].content_changed);
        assert_eq!(
            changes[0].line_stat(&DiffOptions::default()),
            Some(LineStat {
                insertions: 1,
                deletions: 0,
            })
        );
    }

    #[test]
    fn binary_worktree_side_skips_reading_the_old_blob() {
        let (dir, repo) = staged_repo(&[("data.bin", b"old text\n"), ("notes.txt", b"a\n")]);
        std::fs::write(dir.path().join("data.bin"), b"new\0binary")
            .unwrap_or_else(|e| panic!("{e}"));
        std::fs::write(dir.path().join("notes.txt"), b"a\nb\n").unwrap_or_else(|e| panic!("{e}"));

        let entries = unstaged(&repo);
        let index = repo
            .index_handle()
            .load()
            .unwrap_or_else(|e| panic!("index: {e}"));
        let mut diff = UncommittedDiff::new(&repo, index);
        let metrics = repo.odb().hot_path_metrics();
        metrics.set_blob_counting(true);
        metrics.reset_blob_content_reads();
        let binary = diff
            .load_worktree_changes(vec![entries[0].clone()])
            .unwrap_or_else(|e| panic!("load: {e}"));
        assert_eq!(binary[0].path, "data.bin");
        assert!(binary[0].binary);
        assert!(binary[0].content_changed);
        // Only the working-tree file was read.
        assert_eq!(metrics.blob_content_reads(), 1);

        metrics.reset_blob_content_reads();
        let text = diff
            .load_worktree_changes(vec![entries[1].clone()])
            .unwrap_or_else(|e| panic!("load: {e}"));
        assert!(!text[0].binary);
        assert_eq!(text[0].old_bytes.text(), "a\n");
        assert_eq!(metrics.blob_content_reads(), 2);
    }

    #[test]
    fn bare_repository_worktree_methods_require_a_work_tree() {
        let dir = tempfile::tempdir().unwrap_or_else(|e| panic!("tempdir: {e}"));
        let init = InitOptions {
            bare: true,
            ..InitOptions::default()
        };
        let repo = Repository::init(dir.path(), &init).unwrap_or_else(|e| panic!("init: {e}"));
        let mut diff = UncommittedDiff::new(&repo, Index::default());
        assert!(diff
            .staged_entries(RenameDetection::Off)
            .is_ok_and(|e| e.is_empty()));
        assert!(matches!(
            diff.unstaged_entries(RenameDetection::Off),
            Err(Error::RequiresWorkTree)
        ));
        assert!(matches!(
            diff.worktree_entries(None, RenameDetection::Off),
            Err(Error::RequiresWorkTree)
        ));
        assert!(matches!(
            diff.load_worktree_changes(Vec::new()),
            Err(Error::RequiresWorkTree)
        ));
    }
}