Performance

go-ruby-regexp/regexp is the pure-Go library that rbgo binds for Ruby’s regexp. This page records a comparative benchmark of that module against the reference Ruby runtimes, part of the ecosystem-wide per-module parity suite.

What is measured

The same Ruby script — compile a tokenizer pattern and scan it over a body of source text — is run under every runtime. rbgo’s number reflects this pure-Go library doing the work; every other column is that interpreter’s own regexp stdlib. So the comparison is the Ruby-visible operation, apples-to-apples across interpreters. The script prints a deterministic checksum and its output is checked byte-identical to MRI before timing.

• Host: Apple M4 Max, macOS (darwin/arm64). Method: best-of-5 wall time (best, not mean, to suppress scheduler noise); single-shot processes, no warm-up beyond the script’s own loop.
• Runtimes: ruby 4.0.5 +PRISM (MRI, the oracle) and ruby --yjit; jruby 10.1.0.0 (OpenJDK 25); truffleruby 34.0.1 (GraalVM CE Native).
• The benchmark script and harness live in rbgo’s repo under bench/modules/ (regexp.rb + run.sh). Reproduce: RBGO=./rbgo TRUFFLE=truffleruby bash bench/modules/run.sh 5.

Result (best of 5, ms)

rbgo runs on go-ruby-regexp (a pure-Go Onigmo). On this tokenize loop it is ~1.8x MRI’s C Onigmo — near parity for a from-scratch pure-Go engine. TruffleRuby’s native JIT wins this compute-bound string row (340 ms).

Library-level benchmark (Go API vs runtimes + C Onigmo) — 2026-07-03

This section measures the pure-Go library directly, through its Go API — not the rbgo interpreter path recorded above. It isolates the regexp primitive from Ruby-interpreter dispatch and answers the parity question head-on: is the pure-Go engine as fast as the reference runtimes — and as the C Onigmo it reimplements? The same workload, same inputs, same iteration counts run through the Go library, through each reference runtime’s Regexp / StringScanner, and through C Onigmo 6.2.0 (the library MRI links, built from source). Every driver’s per-op output is checked byte-identical to MRI before any timing (corpus + every op cross-checked; the run aborts on mismatch).

The bar is MRI + YJIT. YJIT removes the interpreter dispatch around MRI’s C Onigmo, so MRI + YJIT is the toughest interpreter column and the one the tokenizer-bound StringScanner ops (scan/skip/match?/scan_until) previously lost to by 1.8×–4.2× — a gap localised to this match engine, the target of a pooled anchored-DFA path, a bounds-only match API, an ASCII-class membership bitset, a pooled backtracking VM, and — closing the last op — a fast anchored class-run consumer that settles a bare \s+/\S+/\w+ run at the cursor with one class-bitset bit test per byte, no per-position NFA state.

• Host: Apple M4 Max (Mac16,5, arm64, 16 cores), macOS 26.5.1 (Darwin 25.5.0) — date 2026-07-03.
• Runtimes: Go 1.26.4 · MRI ruby 4.0.5 +PRISM · MRI + YJIT · JRuby 10.1.0.0 (OpenJDK 25) · TruffleRuby 34.0.1 (GraalVM CE Native) · C Onigmo 6.2.0 (k-takata/Onigmo, Ruby syntax + UTF-8).
• Method: each process runs 3 untimed warm-up passes, then 25 timed passes of a fixed inner loop, timed with a monotonic clock; the best pass is reported as ns/op (lower is better). vs MRI / vs YJIT < 1.00× means faster. Interpreter start-up is outside the timed region, so these are operation costs.
• Corpora: the general match/scan/compile ops run over one deterministic ~13 KB ASCII body; the StringScanner tokenizer ops run over a 2 752-byte lexer-shaped body ("foo123 + bar456 - baz789 * qux000 / quux ; " × 64), both byte-identical across all drivers.

Headline — every previously-losing match op now beats MRI + YJIT

These are the tokenizer-bound ops the strscan parity suite flagged as trailing YJIT by 1.8×–4.2×. The pure-Go library now beats YJIT on all of them — the last holdout, skip, goes from 1.4× behind YJIT to ~10× ahead with the anchored class-run consumer, and now beats even the C Onigmo it reimplements.

skip was the honest floor of the previous round (dominated by consuming long \S+/\s+ runs one NFA step per byte). The anchored class-run consumer now recognises that a bare single-char class repeat at the cursor is exactly “how far does the class reach”, and walks it in a tight loop over a 256-bit ASCII membership bitset — one bit test per byte, no per-position NFA frontier to seed, close, or intern. On the alternating-runs skip workload it drops from 136 771 ns to 9 862 ns (13.9× faster than before), clearing MRI (13.3×), YJIT (9.95×), and C Onigmo’s own hand-written inner loop (34 567 ns → 3.5× ahead). A byte ≥ 0x80 under UTF-8 begins a multi-byte code point the byte bitset cannot decide, so the consumer bows out to the general engine there — correct, unaccelerated, and rare on ASCII tokenizer input; results are byte-identical to the general engine at every position.

StringScanner-style tokenizer ops (many short matches)

The classic lexer workload: anchored match from an advancing cursor, thousands of short matches per pass. The Go driver drives the engine through the bounds-only MatchBoundsAt / MatchBounds primitives (no MatchData allocated); the Ruby column is that runtime’s own StringScanner (C Onigmo); the C column is raw Onigmo via onig_match / onig_search.

scan-tokenize (anchored scan per token, 1280 tokens)

skip (alternating \s+ / \S+ runs)

match? (anchored non-advancing match? at every position, 6016 chars)

scan_until (forward scan_until hopping past each operator)

Match / scan over the fixed corpus

scan-word (\w+ full pass, 2689 matches)

search-email (=~ leftmost search)

match-ipv4 (Regexp#match, early hit + captures)

gsub-space (gsub(/\s+/, "_"), full-pass replace)

Compile (pattern → matcher)

The lazy machine build (parse eagerly, defer the program/DFA/prefilter lowering to first match) makes compilation cheap: a pattern that is compiled but not yet matched pays only the parse. go-ruby-regexp beats every interpreter here, including MRI + YJIT, by 6–120×.

(JRuby / TruffleRuby compile columns are omitted here: their JITs recognise the loop rebuilds the same constant pattern and elide the work, so 24–150 ns measures an elided compile, not real compilation. The Go driver defeats this with a sink; MRI and C Onigmo compile every iteration.)

Reading the numbers

• The match-throughput target is met — every flagged op now beats YJIT. skip (9.95×), match? (16.4×), scan-tokenize (2.95×), scan_until (1.25×), gsub (1.25×) all clear MRI + YJIT. What moved them: the anchored MatchAt fast path runs on the pooled lazy-NFA instead of allocating a fresh backtracking machine per call; a bounds-only API (MatchBoundsAt/MatchBounds) returns a span with zero allocation for skip/match?/scan, matching MRI’s integer-returning StringScanner#skip; an ASCII-class bitset makes class membership an O(1) bit test; the backtracking VM itself is pooled; and — the final lever — a fast anchored class-run consumer walks a bare \s+/\S+/ \w+/[…]+ run at the cursor with one bitset bit test per byte and no per-position NFA state.
• skip is no longer the floor — it is now the biggest win. Recognising that a bare single-char class repeat anchored at the cursor is exactly “how far does the class reach”, the consumer replaces the per-byte NFA stepping with a tight bitset loop: 136 771 ns → 9 862 ns (13.9× faster than before), clearing MRI (13.3×), YJIT (9.95×) and even C Onigmo’s hand-written inner loop (3.5×). Under UTF-8 a byte ≥ 0x80 starts a multi-byte code point the byte bitset cannot decide, so the consumer defers the whole match to the general engine there — correct and byte-identical, just unaccelerated on non-ASCII runs. match? (over [A-Za-z0-9_]+) rides the same fast path and jumps from 0.75× MRI to 0.05×.
• =~ beats MRI and YJIT and C Onigmo (search-email 0.34× MRI, 0.25× of C): the literal/first-byte prefilter jumps straight to the @-anchored match instead of stepping every position. The one match lag is the match ipv4 early-hit micro-case (642 ns vs 184 ns YJIT): the capture-extracting match ends a few bytes in, so the backtracking-VM per-call setup dominates a tiny scan — the residual the C-parity report already names.
• Compile is now a strength, not a cost. The lazy build puts every compile op ahead of MRI + YJIT (6–120×), including \p{L}+, which skips Onigmo’s large per-compile Unicode-property table build.
• C Onigmo column. Raw C is still ahead on scan-tokenize (whose [A-Za-z_]… ident pattern is two atoms, not a single class-run, so it keeps the general engine) and on the full-corpus scans — the hand-asm walls. But on the class-run ops the pure-Go engine now beats raw C: skip 3.5× and match? 3.2× ahead of C Onigmo, because the bitset consumer does per byte what Onigmo’s inner loop cannot beat without the same specialisation. The stated bar is MRI + YJIT (C Onigmo through the interpreter), which the pure-Go engine now clears on every tokenizer op.