GitoxideLabs/gitoxide gix-utils-v0.5.0

gix-utils v0.5.0

on GitHub

Bug Fixes

• avoid redundant Unicode precomposition work

  We now preserve macOS conversion failures per filename, but splitting
  every path adds work even for ASCII. In the initial benchmark, ASCII paths
  took about 2.6-2.8 times as long as before. The existing is_nfc()
  check can also perform a full NFC conversion before our compatibility
  composer performs its own decomposition and composition.

  Use is_nfc_quick() and detect non-BMP characters through the same iterator.
  Only scan its unvisited suffix if quick-check stops early. ASCII strings
  return immediately; ASCII paths also avoid UTF-8 validation and splitting.
  Process BMP-only paths together, reserving component splitting for paths
  with conversion failures. Specialize filename and path handling at compile
  time to keep the path fallback out of the filename loop.

  Skip combining-class and composition lookups for ASCII starters, use byte
  lengths for buffer capacities instead of counting characters, and reserve
  the output string before encoding it. Keep noncanonical mark order, Hangul
  composition, component-local fallback, and the original Cow when unchanged.

  The capacity estimate can reserve more space than the previous character
  count. Inconclusive NFC checks can now use a temporary buffer even if the
  result is unchanged; the benchmark explicitly covers that case. There are
  no new runtime dependencies; Criterion is already used in this workspace.

  Add a Criterion benchmark covering filenames and paths with ASCII, NFC,
  NFD, Hangul, noncanonical marks, and mixed emoji/decomposed components.

  Case	Original HEAD median (ns)	This commit median (ns)
  filename/ascii	6.46	2.55
  filename/emoji_first	63.54	3.17
  filename/emoji_last	57.95	15.80
  filename/hangul_nfc	12.20	11.88
  filename/hangul_nfd	141.87	70.28
  filename/nfc	16.60	15.78
  filename/nfc_quick_check_maybe	70.32	32.75
  filename/nfc_with_mark	11.03	11.89
  filename/nfd	303.23	70.96
  filename/noncanonical_marks	46.98	32.58
  filename/several_nfd	139.12	58.56
  path/ascii	34.71	2.69
  path/ascii_long	98.55	4.23
  path/mixed	403.60	134.15
  path/nfc	50.75	26.13
  path/nfd	378.76	234.88
  path/reported	106.27	37.88
  path/several_nfd	343.42	101.15

  The NFC string with an uncomposable mark costs about 0.9 ns more; the
  other measured cases improved. These are normalization microbenchmarks,
  not measurements of complete status operations.

  To compare subsequent changes:

  cargo bench -p gix-utils --bench precompose -- --save-baseline before --sample-size 100 --warm-up-time 1 --measurement-time 2 --noplot
  # Apply the change being measured, then:
  cargo bench -p gix-utils --bench precompose -- --baseline before --sample-size 100 --warm-up-time 1 --measurement-time 2 --noplot

• preserve filenames rejected by macOS precomposition

  The original Starship report states that a tracked filename
  containing both an emoji and a decomposed umlaut is reported
  as untracked by gix status.

  Git's macOS UTF-8-MAC conversion rejects characters outside the Basic
  Multilingual Plane. When conversion fails, Git keeps the entire original
  input, including any decomposed umlauts before or after the rejected
  character. Our precomposition instead changed those umlauts, preventing
  the directory walk from matching Git's index entries.

  Preserve inputs containing non-BMP characters in the shared precomposition
  helper. Apply that rule separately to each filesystem-path component,
  preserving path separators and borrowing unchanged paths. Reference packing
  and lookup also use component-wise precomposition, consistent with loose
  reference iteration.

  Explicit file roots in the directory walk use the same path helper, so
  their emitted spelling agrees with ordinary traversal beneath emoji
  directories. Extend the existing root-precomposition test to cover an
  emoji parent and an emoji filename, with precomposition enabled and disabled.

  An emoji outside a reference name must not suppress that name's
  precomposition: otherwise loose and packed versions of an umlaut branch
  can be emitted twice. Cover that review finding, packing and lookup below
  an emoji component, both emoji/umlaut orderings, another supplementary-plane
  character, the reported Teaching/🎥 Überwachung im digitalen Zeitalter.md
  filename, and BMP-only controls.

  Restore the requested reference spelling after lookup instead of
  decomposing the entire returned name. Whole-name decomposition could alter
  an untouched emoji/umlaut component and make the returned name unusable
  for another lookup or deletion. Cover mixed composition, loose and packed
  references, full names, and worktree-qualified lookup round trips.
  Keep canonical pseudo-ref names when precomposition changes their
  classification, including worktree-qualified pseudo-refs.

  Both the helper regression and macOS status regression failed before the
  fix and now pass, as do the reference regressions added for the review
  finding. Validated the gix-utils, gix-fs, gix-dir, gix-status, and
  gix-ref suites, the gix status tests with default features and with
  --no-default-features --features basic,extras,comfort, targeted Clippy
  with warnings denied, and cargo fmt --all -- --check.

  Git reference: Apple Git 2.54.0 with core.precomposeUnicode=true leaves
  the repository clean. /usr/bin/iconv -f UTF-8-MAC -t UTF-8 rejects these
  non-BMP characters. compat/precompose_utf8.c::precompose_utf8_readdir()
  and precompose_string_if_needed() at
  d38352cd43ab9745686d697872408bc3249a153f retain the original input on
  conversion failure.

  Refs starship/starship#6881

Commit Statistics

• 6 commits contributed to the release over the course of 13 calendar days.
• 13 days passed between releases.
• 2 commits were understood as conventional.
• 0 issues like '(#ID)' were seen in commit messages

Commit Details