GitoxideLabs/gitoxide gix-status-v0.37.0

gix-status v0.37.0

on GitHub

Performance

• Use safe Rust for the Windows gix-status filesystem cache
  Really just skimmed the unit-tests, but looked carefully at the integration test,
  which was small enough. Overall, much better tests and probably a bugfix as well.

  Replace handwritten Windows directory enumeration with std::fs::read_dir()
  and DirEntry::metadata(). On Windows, entry metadata comes from the directory
  listing without an additional per-entry metadata syscall. This removes unsafe
  FFI, raw-buffer parsing, manual reparse-tag classification, and the crate's
  now-unused windows-sys dependency.

  Keep the compact cached Stat and the existing lazy, per-worker lifetime of
  one status operation. Exact-name misses, unreadable entries, unrepresentable
  filenames, and failed enumerations still use the live metadata path. Missing
  cached entries never become authoritative not-found results. Root validation
  and canonicalisation remain unchanged; this adds no checkout integration or
  cache invalidation.

  Derive file types from standard-library metadata, including file and directory
  symlinks and junctions. Preserve file size and use creation time for index
  ctime. Reuse the index's timestamp conversion so pre-epoch timestamps cause
  a cache miss instead of being silently clamped to the Unix epoch.

  Add isolated Windows fixtures comparing cached and live metadata, including
  nanosecond timestamps, invalid UTF-16 names, links, and junctions. Exercise
  missing entries and failed enumeration after filesystem changes, and compare
  cached and uncached status to verify the live fallback and syscall counts.
  The pre-epoch regression test failed against the original implementation and
  passes with this change.

  Document that enumeration throughput can differ from the previous custom
  64-KiB batching. A release-mode microbenchmark of cache construction and lookups
  used isolated temporary fixtures, warmed runs, and 41 alternating old/new
  samples per layout. Median times in milliseconds were:

  Files	Directories	Previous	Standard library
  10,000	1	13.32	7.16
  10,000	100	114.51	9.93
  1,000	1,000	27.94	31.22

  Repeated measurements confirmed about a 12% slowdown for one-file directories,
  while the denser layouts improved substantially. These are cache microbenchmarks,
  not end-to-end status timings or a guarantee of equivalent performance.

• avoid unrelated attribute and ignore reads
  Went through the tests quickly, but refactored them for read-only usage.

  Repeated Tix status refreshes spend time opening pattern files that the active
  operation does not need. Tracked content checks previously loaded ignores,
  and untracked directory walks loaded attributes through a combined stack.

  Separate the consumed status cache into an ignore stack for directory walking
  and an attribute stack for content filters, pathspecs, and rewrite detection.
  Reuse the existing lazy Repository::pathspec() setup so ordinary pathspecs
  do not eagerly load attributes or ignores.

  A regression uses unreadable symlink loops to detect unnecessary reads while
  checking that attribute-constrained pathspecs and untracked ignores still work.
  It fails before the change and passes after it. All 31 selected gix status
  tests and all 35 gix-status tests pass.

Bug Fixes

• match Git's filter path so Git LFS extensions don't show false modifications
  Status passed the absolute worktree path to convert_to_git(), so filters saw
  it as %f or pathname where Git passes the relative path. Filters whose
  output depends on the path then reported phantom modifications, for example Git
  LFS with an extension that uses %f.

Commit Statistics

• 10 commits contributed to the release over the course of 14 calendar days.
• 14 days passed between releases.
• 3 commits were understood as conventional.
• 0 issues like '(#ID)' were seen in commit messages

Commit Details