Improve performance of status on windows

gix-index/src/entry/mode.rs

@@ -46,22 +46,44 @@ impl Mode {
         stat: &crate::fs::Metadata,
         has_symlinks: bool,
         executable_bit: bool,
+    ) -> Option<Change> {
+        self.change_to_match_fs_with_values(
+            stat.is_file(),
+            stat.is_dir(),
+            stat.is_symlink(),
+            stat.is_executable(),
+            has_symlinks,
+            executable_bit,
+        )
+    }
+
+    /// Like [`change_to_match_fs`](Self::change_to_match_fs) but accepts pre-extracted
+    /// file-type and permission bits, for callers that already have them (e.g. cached
+    /// metadata from a batched directory enumeration).
+    pub fn change_to_match_fs_with_values(
+        self,
+        is_file: bool,
+        is_dir: bool,
+        is_symlink: bool,
+        is_executable: bool,
+        has_symlinks: bool,
+        executable_bit: bool,
     ) -> Option<Change> {
         match self {
-            Mode::FILE if !stat.is_file() => (),
-            Mode::SYMLINK if stat.is_symlink() => return None,
-            Mode::SYMLINK if has_symlinks && !stat.is_symlink() => (),
-            Mode::SYMLINK if !has_symlinks && !stat.is_file() => (),
-            Mode::COMMIT | Mode::DIR if !stat.is_dir() => (),
-            Mode::FILE if executable_bit && stat.is_executable() => return Some(Change::ExecutableBit),
-            Mode::FILE_EXECUTABLE if executable_bit && !stat.is_executable() => return Some(Change::ExecutableBit),
+            Mode::FILE if !is_file => (),
+            Mode::SYMLINK if is_symlink => return None,
+            Mode::SYMLINK if has_symlinks && !is_symlink => (),
+            Mode::SYMLINK if !has_symlinks && !is_file => (),
+            Mode::COMMIT | Mode::DIR if !is_dir => (),
+            Mode::FILE if executable_bit && is_executable => return Some(Change::ExecutableBit),
+            Mode::FILE_EXECUTABLE if executable_bit && !is_executable => return Some(Change::ExecutableBit),
             _ => return None,
         }
-        let new_mode = if stat.is_dir() {
+        let new_mode = if is_dir {
             Mode::COMMIT
-        } else if executable_bit && stat.is_executable() {
+        } else if executable_bit && is_executable {
             Mode::FILE_EXECUTABLE
-        } else if has_symlinks && stat.is_symlink() {
+        } else if has_symlinks && is_symlink {
             Mode::SYMLINK
         } else {
             Mode::FILE

gix-status/Cargo.toml

@@ -38,12 +38,21 @@ gix-diff = { version = "^0.62.0", path = "../gix-diff", default-features = false
 thiserror = "2.0.18"
 filetime = "0.2.27"
 bstr = { version = "1.12.0", default-features = false }
+hashbrown = "0.16.0"
 
 document-features = { version = "0.2.0", optional = true }
 
 [target.'cfg(not(target_has_atomic = "64"))'.dependencies]
 portable-atomic = "1"
 
+[target.'cfg(windows)'.dependencies]
+windows-sys = { version = "0.61.1", features = [
+    "Win32_Foundation",
+    "Win32_Storage_FileSystem",
+    # For SECURITY_ATTRIBUTES in CreateFileW (used by metadata_cache::prepare on Windows).
+    "Win32_Security",
+] }
+
 [dev-dependencies]
 gix-hash = { path = "../gix-hash", features = ["sha1"] }
 

gix-status/src/index_as_worktree/function.rs

@@ -21,6 +21,67 @@ use crate::{
     },
     is_dir_to_mode, AtomicU64, SymlinkCheck,
 };
+#[cfg(windows)]
+use crate::metadata_cache::{self, CachedMetadata, MetadataCache};
+
+/// Windows-only union of live `lstat` metadata and pre-cached metadata, so
+/// `compute_status` sees one shape. Other platforms use `gix_index::fs::Metadata`
+/// directly.
+#[cfg(windows)]
+enum FileMetadata<'a> {
+    Live(gix_index::fs::Metadata),
+    Cached(&'a CachedMetadata),
+}
+
+#[cfg(windows)]
+impl FileMetadata<'_> {
+    fn is_dir(&self) -> bool {
+        match self {
+            Self::Live(m) => m.is_dir(),
+            Self::Cached(c) => c.is_dir,
+        }
+    }
+
+    fn is_symlink(&self) -> bool {
+        match self {
+            Self::Live(m) => m.is_symlink(),
+            Self::Cached(c) => c.is_symlink,
+        }
+    }
+
+    fn len(&self) -> u64 {
+        match self {
+            Self::Live(m) => m.len(),
+            Self::Cached(c) => c.size,
+        }
+    }
+
+    fn to_stat(&self) -> Result<gix_index::entry::Stat, std::time::SystemTimeError> {
+        match self {
+            Self::Live(m) => gix_index::entry::Stat::from_fs(m),
+            Self::Cached(c) => Ok(c.to_stat()),
+        }
+    }
+
+    fn mode_change(
+        &self,
+        entry_mode: gix_index::entry::Mode,
+        has_symlinks: bool,
+        executable_bit: bool,
+    ) -> Option<gix_index::entry::mode::Change> {
+        match self {
+            Self::Live(m) => entry_mode.change_to_match_fs(m, has_symlinks, executable_bit),
+            Self::Cached(c) => entry_mode.change_to_match_fs_with_values(
+                !c.is_dir && !c.is_symlink, // is_file: regular file (not dir, not symlink)
+                c.is_dir,
+                c.is_symlink,
+                c.is_executable,
+                has_symlinks,
+                executable_bit,
+            ),
+        }
+    }
+}
 
 /// Calculates the changes that need to be applied to an `index` to match the state of the `worktree` and makes them
 /// observable in `collector`, along with information produced by `compare` which gets to see blobs that may have changes, and
@@ -63,6 +124,8 @@ pub fn index_as_worktree<'index, T, U, Find, E>(
         stack,
         filter,
         should_interrupt,
+        #[cfg(windows)]
+        metadata_cache,
     }: Context<'_>,
     options: Options,
 ) -> Result<Outcome, Error>
@@ -122,6 +185,8 @@ where
                     path_backing,
                     filter,
                     options,
+                    #[cfg(windows)]
+                    metadata_cache,
 
                     skipped_by_pathspec,
                     skipped_by_entry_flags,
@@ -228,6 +293,10 @@ struct State<'a, 'b> {
     filter: gix_filter::Pipeline,
     path_backing: &'b gix_index::PathStorageRef,
     options: &'a Options,
+    /// Optional pre-populated metadata cache for faster status checks on Windows.
+    /// Cache lookups happen before falling back to per-file syscalls.
+    #[cfg(windows)]
+    metadata_cache: Option<&'a MetadataCache>,
 
     skipped_by_pathspec: &'a AtomicUsize,
     skipped_by_entry_flags: &'a AtomicUsize,
@@ -374,53 +443,80 @@ impl<'index> State<'_, 'index> {
             }
             Err(err) => return Err(Error::Io(err.into())),
         };
-        self.symlink_metadata_calls.fetch_add(1, Ordering::Relaxed);
-        let metadata = match gix_index::fs::Metadata::from_path_no_follow(worktree_path) {
-            Ok(metadata) if metadata.is_dir() => {
-                // index entries are normally only for files/symlinks
-                // if a file turned into a directory it was removed
-                // the only exception here are submodules which are
-                // part of the index despite being directories
-                if entry.mode.is_submodule() {
-                    let status = submodule
-                        .status(entry, rela_path)
-                        .map_err(|err| Error::SubmoduleStatus {
-                            rela_path: rela_path.into(),
-                            source: Box::new(err),
-                        })?;
-                    return Ok(status.map(|status| Change::SubmoduleModification(status).into()));
-                } else {
-                    return Ok(Some(Change::Removed.into()));
+
+        // Acquire metadata. On Windows we consult the metadata cache first and
+        // only fall back to a syscall on miss; on other platforms per-file
+        // `lstat` is already fast, so we just do the syscall directly.
+        #[cfg(windows)]
+        let metadata = if let Some(cached) = self.metadata_cache.and_then(|c| metadata_cache::lookup(c, rela_path)) {
+            FileMetadata::Cached(cached)
+        } else {
+            self.symlink_metadata_calls.fetch_add(1, Ordering::Relaxed);
+            match gix_index::fs::Metadata::from_path_no_follow(worktree_path) {
+                Ok(m) => FileMetadata::Live(m),
+                Err(err) if gix_fs::io_err::is_not_found(err.kind(), err.raw_os_error()) => {
+                    return Ok(Some(Change::Removed.into()))
                 }
+                Err(err) => return Err(Error::Io(err.into())),
             }
-            Ok(metadata) => metadata,
-            Err(err) if gix_fs::io_err::is_not_found(err.kind(), err.raw_os_error()) => {
-                return Ok(Some(Change::Removed.into()))
-            }
-            Err(err) => {
-                return Err(Error::Io(err.into()));
+        };
+        #[cfg(not(windows))]
+        let metadata = {
+            self.symlink_metadata_calls.fetch_add(1, Ordering::Relaxed);
+            match gix_index::fs::Metadata::from_path_no_follow(worktree_path) {
+                Ok(m) => m,
+                Err(err) if gix_fs::io_err::is_not_found(err.kind(), err.raw_os_error()) => {
+                    return Ok(Some(Change::Removed.into()))
+                }
+                Err(err) => return Err(Error::Io(err.into())),
             }
         };
+
+        // Handle directory: index entries are normally only for files/symlinks.
+        // If a file turned into a directory it was removed.
+        // The only exception here are submodules which are part of the index despite being directories.
+        if metadata.is_dir() {
+            if entry.mode.is_submodule() {
+                let status = submodule
+                    .status(entry, rela_path)
+                    .map_err(|err| Error::SubmoduleStatus {
+                        rela_path: rela_path.into(),
+                        source: Box::new(err),
+                    })?;
+                return Ok(status.map(|status| Change::SubmoduleModification(status).into()));
+            } else {
+                return Ok(Some(Change::Removed.into()));
+            }
+        }
+
         if entry.flags.contains(gix_index::entry::Flags::INTENT_TO_ADD) {
             return Ok(Some(EntryStatus::IntentToAdd));
         }
+
+        #[cfg(windows)]
+        let new_stat = metadata.to_stat()?;
+        #[cfg(not(windows))]
         let new_stat = gix_index::entry::Stat::from_fs(&metadata)?;
-        let executable_bit_changed =
-            match entry
-                .mode
-                .change_to_match_fs(&metadata, self.options.fs.symlink, self.options.fs.executable_bit)
-            {
-                Some(gix_index::entry::mode::Change::Type { new_mode }) => {
-                    return Ok(Some(
-                        Change::Type {
-                            worktree_mode: new_mode,
-                        }
-                        .into(),
-                    ))
-                }
-                Some(gix_index::entry::mode::Change::ExecutableBit) => true,
-                None => false,
-            };
+
+        #[cfg(windows)]
+        let mode_change =
+            metadata.mode_change(entry.mode, self.options.fs.symlink, self.options.fs.executable_bit);
+        #[cfg(not(windows))]
+        let mode_change = entry
+            .mode
+            .change_to_match_fs(&metadata, self.options.fs.symlink, self.options.fs.executable_bit);
+        let executable_bit_changed = match mode_change {
+            Some(gix_index::entry::mode::Change::Type { new_mode }) => {
+                return Ok(Some(
+                    Change::Type {
+                        worktree_mode: new_mode,
+                    }
+                    .into(),
+                ))
+            }
+            Some(gix_index::entry::mode::Change::ExecutableBit) => true,
+            None => false,
+        };
 
         // We implement racy-git. See racy-git.txt in the git documentation for detailed documentation.
         //

gix-status/src/index_as_worktree/types.rs

@@ -3,6 +3,9 @@ use std::sync::atomic::AtomicBool;
 use bstr::{BStr, BString};
 use gix_index::entry;
 
+#[cfg(windows)]
+use crate::metadata_cache::MetadataCache;
+
 /// The error returned by [index_as_worktree()`](crate::index_as_worktree()).
 #[derive(Debug, thiserror::Error)]
 #[allow(missing_docs)]
@@ -55,6 +58,10 @@ pub struct Context<'a> {
     pub filter: gix_filter::Pipeline,
     /// A flag to query to learn if cancellation is requested.
     pub should_interrupt: &'a AtomicBool,
+    /// Windows-only pre-populated metadata cache. See [`crate::metadata_cache`].
+    /// Transparent: `None`/empty/partial are all correct.
+    #[cfg(windows)]
+    pub metadata_cache: Option<&'a MetadataCache>,
 }
 
 /// Provide additional information collected during the runtime of [`index_as_worktree()`](crate::index_as_worktree()).

gix-status/src/index_as_worktree_with_renames/mod.rs

@@ -153,6 +153,8 @@ pub(super) mod function {
                                 stack,
                                 filter,
                                 should_interrupt: ctx.should_interrupt,
+                                #[cfg(windows)]
+                                metadata_cache: ctx.metadata_cache,
                             },
                             options.tracked_file_modifications,
                         )

gix-status/src/index_as_worktree_with_renames/types.rs

@@ -342,6 +342,11 @@ pub struct Context<'a> {
     pub should_interrupt: &'a AtomicBool,
     /// The context for the directory walk.
     pub dirwalk: DirwalkContext<'a>,
+    /// An optional pre-populated metadata cache for faster status checks on Windows.
+    ///
+    /// See [`crate::index_as_worktree::Context::metadata_cache`] for details.
+    #[cfg(windows)]
+    pub metadata_cache: Option<&'a crate::metadata_cache::MetadataCache>,
 }
 
 /// All information that is required to perform a [dirwalk](gix_dir::walk()).

gix-status/src/lib.rs

@@ -37,6 +37,19 @@ use portable_atomic::AtomicU64;
 pub mod index_as_worktree;
 pub use index_as_worktree::function::index_as_worktree;
 
+/// The metadata cache is a **Windows-only** optimization. Its job is to skip
+/// per-file `lstat` calls by pre-populating stat results via one batched
+/// directory enumeration. That trade only pays off where per-file stat is
+/// expensive (Windows), not on Linux/macOS where `lstat` is sub-microsecond.
+/// A Linux-friendly cache would almost certainly be keyed by *directory* (à la
+/// git's `UNTRACKED_CACHE`) rather than by file path, so forcing this type to
+/// exist there would encourage the wrong abstraction. Keep the two separate;
+/// lift this gate if a cross-platform use case actually appears.
+#[cfg(windows)]
+pub mod metadata_cache;
+#[cfg(windows)]
+pub use metadata_cache::{CachedMetadata, MetadataCache};
+
 #[cfg(feature = "worktree-rewrites")]
 pub mod index_as_worktree_with_renames;
 #[cfg(feature = "worktree-rewrites")]