regex-benchmarks
This is a benchmark of two regular expression quantifiers:
- The possessive quantifier, the
+in e.g..*+ - The non-greedy quantifier, the
?in e.g..*?
In particular, I'd like to know if using them makes regex matching faster or slower.
Disclaimer: We're testing with Rust's regex crate, which is
loosely based on the RE2 library. The conclusion likely only applies to this
crate and #notallregexlibraries! Different regex libraries may apply vastly
different optimizations and therefore have different costs associated.
The experiment
Three regular expressions are being compared:
'([^']*)''([^']*+)''(.*?)'
against two types of quoted strings with ... being 500 random alphanumeric bytes:
'...'which succeeds matching all three regexes'...which fails matching all three regexes
The crux of these six benchmarks is the expression regex.captures(&input).
As a control experiment, three non-capturing variants are also benchmarked:
'[^']*''[^']*+''.*?'
These are tested both as successful and failing matches. This helps assess whether the performance difference between the three is less than the cost of capturing the result.
The crux of the six control benchmarks is the expression regex.is_match(&input).
Questions
- The possessive quantifier is expectably faster, but how much?
- The non-greedy quantifier is expectably slower, but how much?
- Is the speed difference above or below the cost of capturing groups?
Answers
Condensing the cargo bench results,
- The possessive quantifier is actually insignificantly slower than omitting it.
- The non-greedy quantifier is twice as slow, but still insignificantly slower.
- Capturing groups have approximately a factor 7 bigger impact on the performance of a regular expression than a quantifier
capturing regexes that succeed/normal/500 time: [4.7930 µs 4.7942 µs 4.7956 µs]
capturing regexes that succeed/possessive/500 time: [4.8243 µs 4.8256 µs 4.8269 µs]
capturing regexes that succeed/non-greedy/500 time: [8.7937 µs 8.7961 µs 8.7986 µs]
capturing regexes that fail/normal/500 time: [676.01 ns 676.15 ns 676.29 ns]
capturing regexes that fail/possessive/500 time: [684.92 ns 685.06 ns 685.20 ns]
capturing regexes that fail/non-greedy/500 time: [695.49 ns 695.65 ns 695.82 ns]
non-capturing regexes that succeed/normal/500 time: [679.22 ns 679.41 ns 679.59 ns]
non-capturing regexes that succeed/possessive/500 time: [683.18 ns 683.34 ns 683.51 ns]
non-capturing regexes that succeed/non-greedy/500 time: [685.60 ns 685.78 ns 685.97 ns]
non-capturing regexes that fail/normal/500 time: [703.49 ns 703.69 ns 703.90 ns]
non-capturing regexes that fail/possessive/500 time: [692.07 ns 692.25 ns 692.44 ns]
non-capturing regexes that fail/non-greedy/500 time: [694.48 ns 694.64 ns 694.80 ns]
Conclusions
- Don't bother writing
'[^']*+'instead of'[^']*': Slightly slower and less readable. - If you think
'.*?'is more readable, it's probably not slower at a scale that matters. - Remember to use non-capturing groups if you don't actually need to extract the contents.
Appendix: cargo bench
Running benches/regex.rs (target/release/deps/regex-84f60dd518617320)
capturing regexes that succeed/normal/500
time: [4.7930 µs 4.7942 µs 4.7956 µs]
Found 132 outliers among 10000 measurements (1.32%)
97 (0.97%) high mild
35 (0.35%) high severe
capturing regexes that succeed/possessive/500
time: [4.8243 µs 4.8256 µs 4.8269 µs]
Found 143 outliers among 10000 measurements (1.43%)
141 (1.41%) high mild
2 (0.02%) high severe
capturing regexes that succeed/non-greedy/500
time: [8.7937 µs 8.7961 µs 8.7986 µs]
Found 17 outliers among 10000 measurements (0.17%)
14 (0.14%) high mild
3 (0.03%) high severe
capturing regexes that fail/normal/500
time: [676.01 ns 676.15 ns 676.29 ns]
Found 1912 outliers among 10000 measurements (19.12%)
733 (7.33%) high mild
1179 (11.79%) high severe
capturing regexes that fail/possessive/500
time: [684.92 ns 685.06 ns 685.20 ns]
Found 1172 outliers among 10000 measurements (11.72%)
720 (7.20%) low mild
432 (4.32%) high mild
20 (0.20%) high severe
capturing regexes that fail/non-greedy/500
time: [695.49 ns 695.65 ns 695.82 ns]
Found 13 outliers among 10000 measurements (0.13%)
13 (0.13%) high mild
non-capturing regexes that succeed/normal/500
time: [679.22 ns 679.41 ns 679.59 ns]
Found 4488 outliers among 10000 measurements (44.88%)
2096 (20.96%) low severe
30 (0.30%) low mild
422 (4.22%) high mild
1940 (19.40%) high severe
non-capturing regexes that succeed/possessive/500
time: [683.18 ns 683.34 ns 683.51 ns]
Found 2279 outliers among 10000 measurements (22.79%)
27 (0.27%) low severe
1 (0.01%) low mild
142 (1.42%) high mild
2109 (21.09%) high severe
non-capturing regexes that succeed/non-greedy/500
time: [685.60 ns 685.78 ns 685.97 ns]
Found 137 outliers among 10000 measurements (1.37%)
137 (1.37%) high mild
non-capturing regexes that fail/normal/500
time: [703.49 ns 703.69 ns 703.90 ns]
Found 283 outliers among 10000 measurements (2.83%)
279 (2.79%) high mild
4 (0.04%) high severe
non-capturing regexes that fail/possessive/500
time: [692.07 ns 692.25 ns 692.44 ns]
Found 4801 outliers among 10000 measurements (48.01%)
2380 (23.80%) low severe
21 (0.21%) low mild
654 (6.54%) high mild
1746 (17.46%) high severe
non-capturing regexes that fail/non-greedy/500
time: [694.48 ns 694.64 ns 694.80 ns]
Found 2516 outliers among 10000 measurements (25.16%)
846 (8.46%) low severe
22 (0.22%) low mild
334 (3.34%) high mild
1314 (13.14%) high severe