Skip to content

Audit cleanup - #4

Merged
bgreni merged 70 commits into
mainfrom
audit-cleanup
Sep 24, 2026
Merged

Audit cleanup#4
bgreni merged 70 commits into
mainfrom
audit-cleanup

Conversation

@bgreni

@bgreni bgreni commented Sep 24, 2026

Copy link
Copy Markdown
Owner

No description provided.

bgreni and others added 30 commits September 23, 2026 11:59
@bgreni @claude
cleanup: dedup parser escape handling (\u/\U gate, UTF-8 decode, \c, …
0260f70 
…shorthand classes)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: pcre2 compare reuses bench_compare/gen_pdf table code
85116dd 
bench_compare_pcre2.py carried copies of print_comparison/speedup_bar and
gen_pdf's build_table_data/generate_pdf differing only in labels, widths
and titles. Parametrize the originals (defaults unchanged) and import them.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop _negate_cp, reuse utf8.negate_ranges
7570234 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: @fieldwise_init for the Sbt* comptime structs
d09e0bd 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop dead imports in backtrack.mojo and executor.mojo
acb09d6 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: merge _build_star/_build_plus into _build_loop
d78f79e 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop dead negated param of ASTNode.char_class
2c3afa8 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one anchor check for Pike VM and heap backtracker, drop word…
7250e1a 
…-char forwarders

PikeVM._check_anchor/_is_word_char duplicated _bt_check_anchor; the
_bt_/_sbt_is_word_char forwarders become is_word_byte, and the WORD /
NOT_WORD arms share one body.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop edfa_is_word, WIDE_TABLE_CAP and _ShengState aliases
9ccbb99 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop run_coverage.py --gcov pipeline
7e2f43f 
A one-off cross-check CI never runs (--check-baseline refused it). Removes
parse_gcov_intermediate, the gcov branches in cover_one, the flag and its
guard, the opt/llvm-cov lookup and TestGcovParse. compiler-rt stays: the
default link pulls libclang_rt.osx.a (clang -### shows it), and clang_impl
depends on it anyway; its pixi.toml comment now says so.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop per-site build_bitmap in _parse_escape (parse() builds …
8deefc8 
…them all)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: add explicit flags parameter to Regex
1286022 
Regex[pattern, flags: RegexFlags = RegexFlags()] compiles the flags as a
leading inline group (after any (*UTF8) verbs) through Regex._pat, which
is exactly `pattern` for the default flags, so default instantiations
keep the same memoized NFA and backtracker symbols. Drops the unused
RegexFlags.__and__ and fixes the README Flags section.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: _utf8_trie_fragment builds its own root index list
21bdffd 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one ascii_to_lower helper in constants.mojo
54fd26a 
Replaces _sbt_to_lower, _bt_to_lower and nfa.mojo's _to_lower (same
ASCII A-Z rule; generic over the scalar type so the byte walkers and the
UInt32 codepoint builder share it).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one parametric anchor-continuation walk for the three DFA-la…
8c52cbd 
…ne guards

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: merge LOOKAHEAD/LOOKBEHIND arms in the backtracker and Pike VM
0d43368 
Only the sub-match differs; the negated / keep-restore tail is shared.
The backtracker picks the sub-match with a comptime if, so each state
still elaborates exactly one arm.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one Regex._bt helper for the 14 _sbt_run call sites
027e6e5 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: fold edfa_walk_from/sheng_walk_from and the LF find_end wrap…
d9070e3 
…pers into edfa_match_at/sheng_match_at

Every caller passed the DFA's own start states; the LF wrappers passed
lf.d's, so they were edfa_match_at/sheng_match_at over lf.d.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop the unused InnerLiteral.is_suffix and its bookkeeping
52fd774 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one shared compile_once for the compile-time tools
22fe9ff 
comptime_regex_stages, comptime_pattern_probe, comptime_stages and
compile_dashboard each wrote a source file, timed `mojo build` and checked
the return code. tools/timed_build.py now does that once; each tool keeps
its own flags, cache policy, timeout and error reporting. The pattern
probe stays separate: it builds against the precompiled package and ranks
many patterns, where the stages ladder builds against the source tree.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: build span-only results through Regex._span_result
2430806 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: collapse Pike VM step arms and MATCH scans, drop unused max_pos
6914f28 
One _add_state call for CHAR/ANY/CHARSET, one MATCH scan for the pinned
and unpinned modes (the cut only when unpinned); max_pos was never
passed by any caller.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one eager/Sheng walker pair with a comptime cap tier
716ad27 
_sheng_walk_impl and _sheng_full_match_impl were _edfa_walk_impl and
_edfa_full_match_impl with a shuffle step (and a state-vector resync after
a region skip); cap == 0 keeps the table step. _sheng_step moves to
simd_kernels next to the tbl lookups it wraps.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: set_oracle drops the region-bounded sweep/sweep_som
c8da766 
sweep_ctx superseded them (MULTIPATTERN_PLAN.md: the region-bounded
oracle was unsound for lookahead and anchors). __main__ now runs sweep_ctx;
_pin keeps a leading global flag group ((?i), (?s)) outside its wrapper,
which Python otherwise rejects. sweep_ctx_som stays as the SOM oracle and
takes over sweep_som's doc references. Case-table output is byte-identical,
and sweep_ctx_som matches the old sweep_som on every case.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: share the literal-alternation head expansion and chain walk
c4110e1 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: LazyDFA.search_forward reuses match_at's walk
8a2bb77 
The per-start walk moves into an @always_inline _longest_end that both
match_at and search_forward inline, so neither gains a call. The
unreachable pos > input_len break goes with it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: hyperscan compare imports PAIRS/run_mojo from bench_compare_set
26c92e2 
The same bench_set table parse and the same nine row pairs lived in both
scripts; the haystack size now comes from the row name's _16k/_64k suffix.
A failing `pixi run bench_set` now prints an error and yields no rows
(bench_compare_set's handling) instead of a traceback.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: run_test.py drops the bench skip and two one-line helpers
a9b2d51 
`if "bench" in root` matched the absolute walk root, so a checkout under
any path containing "bench" skipped every test (test/ has no bench dir).
needs_pkg_rebuild (a != wrapper) and prune_results (one dict
comprehension, one caller) are inlined; their unit tests go with them.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one required-byte fast-fail helper, one lazy-DFA except comment
4a0a4dd 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop bench/tooling leftovers and the unused pixi build task
9bbc864 
- bench_compare: run_mojo_static_benchmarks absorbs _run_mojo_task (its
  only caller); comments naming the nonexistent bench_static.mojo now
  name bench.mojo.
- unused: statistics (comptime_stages), TA_CENTER and LIGHT_GREEN
  (gen_pdf).
- pixi.toml: the `build` task (mojo precompile -> emberregex.mojoc) was
  referenced nowhere, duplicated run_test.ensure_package and read like
  `pixi build`. The lock is unaffected (pixi lock --dry-run --check).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
bgreni and others added 28 commits September 23, 2026 12:32
@bgreni @claude
cleanup: build_lf_dfa returns EagerDFA; drop LFDFA and prev_ids
af95e79 
LFDFA.valid always equalled d.valid (_edfa_finish sets it) and prev_ids
only fed one structural test; the look-behind states it appended to
`starts` were only renumbered, never read by the finish, so the table is
unchanged. That test is deleted: the invariant lives in the builder's
look-behind lane, not in the table, and test_both_class_atom_before_anchor
/ test_differential_both_class_atoms pin the same fix (bd077fe) through the
directly built LF table against the Pike VM. build_lf_dfa's minimize hook
had no other user.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: stale table-name docstrings (edfa_table_arr, rdfa, onepass_t…
5d1b4b0 
…able_arr)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: format
548eff9 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni
Merge branch 'worktree-agent-aa1ac8a7ab22437f9' into audit-cleanup
69d9980
@bgreni
Merge branch 'worktree-agent-a62627dafaec0f0cc' into audit-cleanup
5168a89
@bgreni @claude
cleanup: pick the LF hash-lookup lane slices with a comptime for
80c9fc5 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni
Merge branch 'worktree-agent-ae71b72a41b0856fa' into audit-cleanup
13ed966
@bgreni @claude
cleanup: share the determinizer setup through _FlatNFA and small helpers
5f5d38e 
_flatten_nfa's 13 arguments become a _FlatNFA built in place (its
constructor also runs _byte_classes); _class_ranges, _word_classes,
_eol_ok_bits and _bs_members replace the per-builder copies of the
rep_lo/rep_hi, word-class, pending-EOL and bitset->member-list loops.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: fix stale docstrings (_edfa_finish extra starts, PROBE_RANKS)
25a1897 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: mojo format (pre-existing drift in engine.mojo slots=Array c…
3c888ed 
…alls)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni
Merge branch 'worktree-agent-a29db89f4ef5859f9' into audit-cleanup
c9c259d 
# Conflicts:
#	emberregex/engine.mojo
#	emberregex/onepass.mojo
#	emberregex/static_dfa.mojo
#	emberregex/static_lfdfa.mojo
#	test/test_leftmost_first_dfa.mojo
#	test/test_word_boundary_dfa.mojo
@bgreni @claude
cleanup: one set Pike scan with a comptime som flag
9bfa043 
set_pike_scan and set_pike_som_scan share _pike_scan[som]. The SOM
lane's per-thread / per-report starts move out of the recursive
closure: every state one closure adds carries the caller's start, so
the parallel lists are resized after _set_add_state returns. The plain
instantiation keeps the unchanged _set_add_state and _flush_reports
(asm-diffed against the base: identical save one cmp/b.ne ->
eor/tbnz pair).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: share the Teddy front end between litset and Rose
1ba8c47 
Rose's _rose_pos_masks re-derived _litset_pos_masks from the flat meta
pools and rose_scan copied litset_scan's chunk loop. The masks now come
from litset_masks(Self._rose.lit) at decl level, and both scans run
teddy_front_end with their verify step as an inlined closure over an
immutable view of the input. The lane probe's teddy/rose functions and
constant pools are asm-identical to the previous build.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one bitnfa step shared by block and stream scans
52cfc5d 
bitnfa_scan and bitnfa_stream_chunk repeated the per-byte gather /
consume / exception table / shift / seed body; both now call the
@always_inline _bn_consume and _bn_advance helpers (block mode still
does not call the out-of-line stream function). Lane-probe asm and
constant pools are identical to the previous build.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop the unreachable shared-entry reverse start scan
171e497 
build_union_subset_nfa and splice_nfa give every pattern a fresh entry
state (each fragment start is a state its own build appended), so the
dup/-2 detection in _start_id_map and the exact _start_ids scan it
selected could only be reached by hand-editing pattern_starts, which is
all test_reverse_dfa_shared_entry_uses_exact_start_scan did.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop dead template params from _rose_walk / _rose_confirm
199fcce 
_rose_walk never read meta, lits or bcls, and _rose_confirm took lits
and bcls only to pass them on. Lane-probe asm unchanged.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: one report sort+dedup pair next to SetMatch
30f6565 
set_literal's _sort_reports + inline dedup and set_rose's sort_reports
+ dedup_reports were the same insertion sort (identical predicate) and
the same compaction; both lanes now call the Rose pair, moved to
set_pike.mojo. The moved bodies are asm-identical; the Teddy lane now
calls them out of line instead of inlining its private copy.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: set_semantics dead param, any_ext reuse, dead guard
8f5e45b 
needs_som never read its flags param; has_semantics re-inlined
any_ext; sem_table_arr's `b + SEM_MIN_LEN >= n` break could not fire
(its only caller sizes n = sem_table_len = SEM_STRIDE * num_patterns).
apply_semantics / apply_semantics_spans stay separate: SetMatch has no
start field, so one generic needs a trait or per-branch rebinds, which
is no shorter than the two 10-line loops.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop set-lane alias decls and one-off table converters
60b57d3 
- _SBN_* / _stream_bn / _can_stream only renamed _BN_* / _bitnfa /
  _bitnfa.valid; set_stream.mojo and test_set_phase6 read those.
- rebind[NFA](self._nfa) was a no-op now that _nfa is a plain NFA.
- mdfa_table_str / rdfa_table_str / rose_table_str were pure
  table_bytes wrappers, rose_flags_arr / ac_cls_arr pure int_arr
  copies; the decl-level sites (and the bench/tests that used
  mdfa_table_str) call table_bytes / int_arr directly.

Lane-probe asm and static data are identical to the previous build.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: drop the dead literal-chain walk budget
5c6f4b0 
extract_literal_chains expands each state at most once (the seen set),
and an expansion pushes at most two successors, so the walk pops at
most 2 * num_states + 1 entries; the 2 * num_states + 8 budget could
never run out. walk_pops stays (test_set_ac pins it).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: stdlib sort for the bitnfa report ids (NEEDS MEASUREMENT)
7fa2351 
_emit_bits hand-rolled an insertion sort over the accepting ids before
its dedup; use sort(ids). Runtime, once per accepting byte on the
bitnfa block and stream lanes: revert if bench_set's bitnfa / stream
rows regress (the ids lists are small and near-sorted, where insertion
sort can win).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: stdlib sort in the set Pike report flush (NEEDS MEASUREMENT)
8805f07 
_flush_reports hand-rolled an insertion sort over the ids reported at a
position; use sort(ids). Runtime, once per position on the tagged Pike
lane (word-boundary sets, residual Pike): revert if bench_set's
word-boundary scan row regresses.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: stdlib sort in the SOM Pike span flush (NEEDS MEASUREMENT)
6e2e118 
_flush_spans insertion-sorted two parallel lists (ids, starts); emit
the spans first and sort that slice of `out` by id with stdlib sort
instead (ids are unique per position, so the order is unambiguous).
Runtime, once per position on the SOM-carrying Pike (scan_som /
scan_spans of word-boundary and reverse-cap-blowup sets): revert if
those rows regress.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: stdlib sort in leftmost_nonoverlapping (NEEDS MEASUREMENT)
a4f3b6e 
Both hand-rolled insertion sorts (per-id start/longest order, then the
cross-id start/id order) become sort(list, cmp). Both orders are total
over their inputs, so stability is moot. Runtime, once per scan_spans
call: revert if the scan_spans rows regress.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: mojo format (pre-existing blank-line drift in static_dfa.mojo)
85f2318 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
cleanup: flags ride into _build_static_nfa as an operand, not through…
ad8394b 
… _pat

Routing Regex.nfa through the _pat ternary put that expression into every
lane method's type (via _num_slots) and cost ~20% compile time on
bench.mojo. nfa = _build_static_nfa(pattern, flags.value) keeps the
default-flags instantiations' expression the same shape as before;
_pat stays only at the backtracker call sites, where it measured free.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
Revert "cleanup: stdlib sort in the set Pike report flush (NEEDS MEAS…
d6186cb 
…UREMENT)"

This reverts commit 8805f07.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@bgreni @claude
test: cover apply_flags at runtime
239c742 
apply_flags only ran in the comptime interpreter, so the coverage gate
counted its 26 runtime lines as uncovered (98.65% < 99.01% baseline).
Three runtime asserts on the existing flags test bring it back to 99.03%.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@github-actions

github-actions Bot commented Sep 24, 2026

Copy link
Copy Markdown

Benchmark comparison

Advisory: this check reports but never fails the build. Hosted-runner noise reaches +200% on rows that changed nothing, so treat a flag as a prompt to measure locally, not as a verdict.

Runner: macos-14

070fe5c → 239c742 · 5 paired rounds, alternating order, warmup discarded · 126 shared rows

✅ No row regressed more than 20% on both statistics above the 0.02 ms noise floor.

Geometric mean of the paired-round medians: -0.1%.

All 126 shared rows
row base (ms) head (ms) min median
bench/anchor_word_boundary 0.00063 0.00062 -1.6% +67.2%
bench/static_realworld_ipv4 0.00064 0.00066 +3.1% +12.1%
bench/static_dfa_quantifier_bounded 0.00057 0.00059 +3.5% +8.8%
bench_set/set_mdfa_log_sparse_64k 0.10813 0.11004 +1.8% +7.8%
bench/static_dfa_alternation_8 0.00014 0.00015 +7.1% +7.1%
bench/static_realworld_log_parse 0.00326 0.00336 +3.1% +6.4%
bench_set/set_mdfa_log_dense_16k 0.04091 0.04143 +1.3% +6.3%
bench_set/set_mdfa_log_sparse_16k 0.0272 0.02738 +0.7% +5.3%
bench/static_lookbehind_positive 0.00077 0.00081 +5.2% +5.1%
bench/realworld_phone 0.00061 0.00061 +0.0% +4.9%
bench/realworld_key_value_findall 0.0323 0.03099 -4.1% +4.9%
bench/memo_ambiguous_plus_miss_1500 0.00424 0.00439 +3.5% +4.0%
bench/static_capture_email 0.00163 0.00163 +0.0% +3.5%
bench/replace_named_backref 0.00492 0.00498 +1.2% +3.5%
bench/reverse_suffix_search_64KB 0.58527 0.60928 +4.1% +3.4%
bench/capture_findall_sparse_64KB 0.28892 0.29294 +1.4% +3.3%
bench/counted_repeat_giveback_2KB 0.29548 0.28304 -4.2% +2.8%
bench/replace_50_matches 0.1268 0.12893 +1.7% +2.7%
bench/throughput_class_10KB 0.03825 0.0383 +0.1% +2.5%
bench_set/set_rose_full_sparse_64k 0.01106 0.01108 +0.2% +2.2%
bench/pathological_optional_16 0.00139 0.0014 +0.7% +2.2%
bench/findall_3_matches 0.02037 0.02009 -1.4% +2.1%
bench/static_pathological_optional_8 0.00053 0.00053 +0.0% +1.9%
bench/lf_dfa_class_run_search_20KB 0.16454 0.16099 -2.2% +1.7%
bench/findall_500_dot_matches 0.32584 0.33338 +2.3% +1.6%
bench_set/set_pike_wb_sparse_16k 0.45339 0.4453 -1.8% +1.6%
bench/throughput_literal_100B 0.00064 0.00064 +0.0% +1.6%
bench/static_greedy_tag 0.00064 0.00062 -3.1% +1.5%
bench/realworld_url_parse 0.00237 0.00242 +2.1% +1.3%
bench/alternation_4_search_2KB 0.02516 0.02431 -3.4% +1.2%
bench/pathological_triple_backref 0.00094 0.00094 +0.0% +1.1%
bench/static_nested_quantifier 0.00238 0.00238 +0.0% +0.8%
bench/static_realworld_csv_fields 0.03156 0.03158 +0.1% +0.7%
bench/static_email_search_2KB 0.01196 0.01201 +0.4% +0.7%
bench/positional_group_email 0.00162 0.00156 -3.7% +0.6%
bench/static_ignorecase_search_2KB 0.00844 0.00849 +0.6% +0.6%
bench_set/set_rose_log_sparse_64k 0.01046 0.01009 -3.5% +0.6%
bench_set/set_bitnfa_log_dense_16k 0.04083 0.03926 -3.8% +0.5%
bench/pathological_nested_quantifier_miss 0.83102 0.87041 +4.7% +0.5%
bench/static_nested_groups 0.00245 0.00246 +0.4% +0.4%
bench_set/set_rose_log_sparse_16k 0.00277 0.00266 -4.0% +0.4%
bench_set/set_bitnfa_blowup_dense_16k 0.02411 0.02339 -3.0% +0.2%
bench_set/set_rose_anchors_dense_16k 0.00917 0.0088 -4.0% +0.2%
bench/alternation_16 0.00015 0.00015 +0.0% +0.0%
bench/alternation_16_miss 9e-05 9e-05 +0.0% +0.0%
bench/anchor_bol 0.0002 0.00021 +5.0% +0.0%
bench/anchor_bol_miss_10KB 7e-05 7e-05 +0.0% +0.0%
bench/anchor_eol 0.00067 0.00067 +0.0% +0.0%
bench/anchor_word_boundary_miss 0.00105 0.00105 +0.0% +0.0%
bench/dotall_multiline_body 0.00028 0.00026 -7.1% +0.0%
bench/neg_lookahead 0.00054 0.00053 -1.9% +0.0%
bench/neg_lookbehind 0.00044 0.00042 -4.5% +0.0%
bench/password_validation_lookahead 0.00232 0.00224 -3.4% +0.0%
bench/pathological_dotstar_miss_5K 0.0001 0.0001 +0.0% +0.0%
bench/realworld_hex_color 0.00034 0.00034 +0.0% +0.0%
bench/simd_literal_match 8e-05 8e-05 +0.0% +0.0%
bench/static_bol_alternation_miss_10KB 0.00021 0.0002 -4.8% +0.0%
bench/static_dfa_char_class_26 0.00127 0.00126 -0.8% +0.0%
bench/static_dfa_literal_match 0.00037 0.00037 +0.0% +0.0%
bench/static_ignorecase 0.0008 0.00081 +1.2% +0.0%
bench/static_lookahead_positive 0.00048 0.00048 +0.0% +0.0%
bench/static_search_short_11B 0.00066 0.00064 -3.0% +0.0%
bench_set/set_teddy8_sparse_16k 0.00179 0.00184 +2.8% +0.0%
bench/realworld_html_tag_findall 0.03733 0.03605 -3.4% -0.0%
bench/findall_100_matches 0.1716 0.17536 +2.2% -0.1%
bench_set/set_ac_256_dense_16k 0.03128 0.03129 +0.0% -0.1%
bench/capture_search_miss_100KB 18.7057 18.7258 +0.1% -0.2%
bench/simd_literal_search 0.01101 0.01099 -0.2% -0.2%
bench/split_100_parts 0.16293 0.1652 +1.4% -0.3%
bench_set/set_teddy64_dense_16k 0.00802 0.00805 +0.4% -0.5%
bench_set/set_teddy8_dense_16k 0.00576 0.00577 +0.2% -0.5%
bench_set/set_rose_log_dense_16k 0.02544 0.0245 -3.7% -0.5%
bench/realworld_semver 0.00182 0.00183 +0.5% -0.5%
bench_set/set_teddy64_sparse_16k 0.00174 0.0017 -2.3% -0.6%
bench/static_realworld_email 0.00168 0.00167 -0.6% -0.6%
bench/realworld_log_search_1000_lines 0.1038 0.10372 -0.1% -0.6%
bench/lf_dfa_lazy_findall_64KB 9.96844 9.93025 -0.4% -0.8%
bench/sheng64_alt_32_search_2KB 0.32665 0.31473 -3.6% -0.8%
bench/static_url_search_2KB 0.12573 0.12593 +0.2% -0.9%
bench/throughput_nomatch_100KB 0.12832 0.12749 -0.6% -0.9%
bench/reverse_inner_search_64KB 0.19913 0.19658 -1.3% -1.1%
bench/inline_multiline_search 0.0008 0.00079 -1.3% -1.2%
bench/engine_pike_with_capture 0.00064 0.00064 +0.0% -1.5%
bench/static_backref 0.00064 0.00063 -1.6% -1.5%
bench/word_boundary_findall_64KB 32.6917 31.6757 -3.1% -1.6%
bench_set/set_mdfa_anchors_dense_16k 0.0113 0.01076 -4.8% -1.7%
bench/inline_ignorecase 0.00054 0.00054 +0.0% -1.8%
bench/static_html_tag 0.00105 0.00101 -3.8% -1.9%
bench/multiline_bol_findall_100_lines 0.17757 0.17196 -3.2% -1.9%
bench/throughput_literal_1MB 1.34079 1.31123 -2.2% -1.9%
bench_set/set_rose_full_dense_16k 0.02005 0.01988 -0.8% -1.9%
bench_set/set_mdfa_full_dense_16k 0.02688 0.02564 -4.6% -2.1%
bench/named_group_date 0.00047 0.00045 -4.3% -2.1%
bench/static_lazy_tag 0.00091 0.00088 -3.3% -2.2%
bench/throughput_literal_10KB 0.01375 0.01349 -1.9% -2.2%
bench/static_replace_alternation 0.12536 0.12211 -2.6% -2.2%
bench/engine_dfa_no_capture 0.00045 0.00044 -2.2% -2.2%
bench/static_ignorecase_alternation_2KB 0.02204 0.02191 -0.6% -2.2%
bench/onepass_findall_2KB 1.12606 1.07011 -5.0% -2.3%
bench/static_search_medium_1KB 0.00112 0.0011 -1.8% -2.6%
bench/memo_ambiguous_plus_in_span_1500 17.6453 18.0723 +2.4% -2.7%
bench/onepass_match_kv 0.01055 0.01023 -3.0% -2.8%
bench/named_group_email 0.00105 0.00101 -3.8% -2.9%
bench/static_search_miss_10KB 0.01354 0.01287 -4.9% -3.0%
bench/static_replace_with_backref 0.02347 0.02357 +0.4% -3.3%
bench/realworld_ws_normalize 0.01648 0.01637 -0.7% -3.4%
bench/throughput_literal_100KB 0.12844 0.12349 -3.9% -3.6%
bench_set/set_ac_256_sparse_64k 0.0111 0.01091 -1.7% -3.7%
bench/match_single_byte_run_20KB 0.11071 0.10827 -2.2% -3.7%
bench/static_teddy_prefix_search_2KB 0.02367 0.02373 +0.3% -3.9%
bench/multiline_eol_findall_100_lines 0.84706 0.8156 -3.7% -4.1%
bench/static_dotstar_search_1K 0.00664 0.00655 -1.4% -4.4%
bench/static_split_delimiters 0.02994 0.02934 -2.0% -4.4%
bench/counted_repeat_search_2KB 0.5762 0.54496 -5.4% -4.8%
bench/static_email_search_long_tokens 0.00966 0.0095 -1.7% -5.1%
bench/static_search_date_capture 0.00183 0.00175 -4.4% -5.4%
bench_set/set_mdfa_full_sparse_64k 0.10682 0.10315 -3.4% -5.4%
bench/alternation_4 0.00017 0.00017 +0.0% -5.6%
bench/static_search_long_20KB 0.0144 0.01395 -3.1% -5.6%
bench/static_dotstar_1K 0.00344 0.00342 -0.6% -5.8%
bench/static_explicit_case_range 0.0008 0.0008 +0.0% -5.9%
bench/static_replace_numbers 0.02045 0.02003 -2.1% -6.1%
bench/static_findall_numbers 0.03112 0.02913 -6.4% -6.4%
bench/pathological_pike_search_miss_600B 0.0017 0.00169 -0.6% -6.6%
bench/pathological_dotstar_anchored_5K 0.01558 0.01559 +0.1% -7.0%
bench/engine_backtrack_with_backref 0.00083 0.00077 -7.2% -11.2%

@bgreni
bgreni merged commit 1f4debb into main Sep 24, 2026
4 of 6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@bgreni