|
| Pattern | Compile time (ns) | JIT compile time (ns) | JIT compile / compile time |
|---|---|---|---|
| abc | 264 | 7932 | 30.05 |
| a+(b*)[c|d]??e+? | 925 | 17186 | 18.58 |
| \A.*\K[^a-z!]{8,}\Z | 925 | 11501 | 12.43 |
| \b((.)\w{3,}\2{0,3})$ | 1057 | 21747 | 20.57 |
| (?P<cap>\p{N}+|(?(R)a|(?R)))\1 | 1322 | 30869 | 23.35 |
| ^(?![a-zA-Z]*?[[:alpha:]]++)\P{L&} | 1520 | 20425 | 13.44 |
Results
The following patterns are searched on a non-utf8 input and an utf8 input. The utf8 input was shorter so it was repeated 4 times (with a simple memcpy()).The performace is measured on an Intel 2.67GHz with GCC 4.4.5 in both 32 and 64 bit mode. The unit of the runtime values are milliseconds (ms) which is 10-3 second. The PCRE library was compiled with -O3 and -static compiler options to get the best performance form the compiler.
Description of the colums
- Pattern: the pattern string. The light blue lines are caseless, the light yellow lines are caseful matches.
- Int. (ms): interpreter (normal pcre_exec(...)) runtime in milliseconds.
- JIT (ms): JIT compiled machine code runtime in milliseconds.
- Runtime ratio: Interpreter runtime / JIT runtime.
- Runtime save %: Runtime saved by the JIT compiler. Equals to 1 - (JIT runtime / Interpreter runtime)
| Pattern | Int. (ms) | JIT (ms) | Runtime ratio | Runtime save % |
|---|---|---|---|---|
| how to | 60 | 40 | 1.50 | 33.3% |
| walk|get|she|make | 260 | 80 | 3.25 | 69.2% |
| [Ww]alk|[Gg]et|[Ss]he|[Mm]ake | 260 | 100 | 2.60 | 61.5% |
| (?:ba|lu|ro)+(?:r|ck)* | 200 | 60 | 3.33 | 70.0% |
| (?:ba|lu|ro)+(r|ck)* | 220 | 80 | 2.75 | 63.6% |
| \w+our\w* | 1760 | 380 | 4.63 | 78.4% |
| (?:^\w+$)|\b(\W)\1+\b | 1840 | 320 | 5.75 | 82.6% |
| "[A-Z][a-z]*(?:(?:[,\s]|\R)(\s|\R)*[a-z]+){0,4}?[.!]" | 80 | 40 | 2.00 | 50.0% |
| \b(?:(?=\w+ro\w+)(?=\w+th\w+)\w+hr\w+)\b | 1280 | 320 | 4.00 | 75.0% |
| \b(?(?=\w+sh)\w+a|\w+uo)\w+\b | 2020 | 400 | 5.05 | 80.2% |
| \b\WX{0,3}(?:(?:V?I{1,3})|V|IV|IX)\W\b | 780 | 180 | 4.33 | 76.9% |
| ((\w)+?)\.\b(\s)*?\w | 4800 | 560 | 8.57 | 88.3% |
| (.{1,3})(\w*?\1)(?2) | 6780 | 1680 | 4.04 | 75.2% |
| -?-?-?-?-?-?-?-?-?-?-?-----------$ | 640 | 100 | 6.40 | 84.4% |
| Average: | 1498 | 310 | 4.16 | 70.6% |
| Pattern | Int. (ms) | JIT (ms) | Runtime ratio | Runtime save % |
|---|---|---|---|---|
| how to | 60 | 30 | 2.00 | 50.0% |
| walk|get|she|make | 240 | 80 | 3.00 | 66.7% |
| [Ww]alk|[Gg]et|[Ss]he|[Mm]ake | 250 | 90 | 2.78 | 64.0% |
| (?:ba|lu|ro)+(?:r|ck)* | 190 | 80 | 2.38 | 57.9% |
| (?:ba|lu|ro)+(r|ck)* | 200 | 80 | 2.50 | 60.0% |
| \w+our\w* | 1590 | 420 | 3.79 | 73.6% |
| (?:^\w+$)|\b(\W)\1+\b | 1580 | 320 | 4.94 | 79.7% |
| "[A-Z][a-z]*(?:(?:[,\s]|\R)(\s|\R)*[a-z]+){0,4}?[.!]" | 90 | 30 | 3.00 | 66.7% |
| \b(?:(?=\w+ro\w+)(?=\w+th\w+)\w+hr\w+)\b | 1020 | 320 | 3.19 | 68.6% |
| \b(?(?=\w+sh)\w+a|\w+uo)\w+\b | 1710 | 410 | 4.17 | 76.0% |
| \b\WX{0,3}(?:(?:V?I{1,3})|V|IV|IX)\W\b | 720 | 170 | 4.24 | 76.4% |
| ((\w)+?)\.\b(\s)*?\w | 4330 | 590 | 7.34 | 86.4% |
| (.{1,3})(\w*?\1)(?2) | 5970 | 1590 | 3.75 | 73.4% |
| -?-?-?-?-?-?-?-?-?-?-?-----------$ | 580 | 100 | 5.80 | 82.8% |
| Average: | 1323 | 307 | 3.78 | 70.1% |
| Pattern | Int. (ms) | JIT (ms) | Runtime ratio | Runtime save % |
|---|---|---|---|---|
| die der | 20 | 20 | 1.00 | 0.0% |
| ist|der|die|und | 100 | 20 | 5.00 | 80.0% |
| \b\w+\b | 240 | 160 | 1.50 | 33.3% |
| (?:da|ge|om)+(?:n|me)* | 80 | 20 | 4.00 | 75.0% |
| \b(?(?=\w+ro)\w+pa|\w+lle)\w+\b | 620 | 300 | 2.07 | 51.6% |
| \b(\W)\1+\b|(^(?=.*kl)(?=.*no).{15,40}$) | 740 | 200 | 3.70 | 73.0% |
| ^.{4,32}(\P{N})\1{2,}.{4,32}(? | 260 | 60 | 4.33 | 76.9% |
| ^(\w{3,})(?!\1).*\h.*\1$ | 5020 | 2500 | 2.01 | 50.2% |
| ((\w{2,8},?(\P{Z}|\R)){1,2}\.\s?)$ | 5780 | 1500 | 3.85 | 74.0% |
| \b\w*?((.){1,3}\w*\2)\w*?(?1) | 2000 | 780 | 2.56 | 61.0% |
| \w*?(b{2,3})\w*?c | 1340 | 420 | 3.19 | 68.7% |
| \P{Lu}\P{L&}{0,12}[\s\-]{1,4}..[\P{L}\P{N}]{4} | 320 | 120 | 2.67 | 62.5% |
| \b(\B([c-h])\B|[a-z]+?(?1)[a-z]) | 1300 | 320 | 4.06 | 75.4% |
| Average: | 1370 | 493 | 3.07 | 60.1% |
| Pattern | Int. (ms) | JIT (ms) | Runtime ratio | Runtime save % |
|---|---|---|---|---|
| die der | 10 | 10 | 1.00 | 0.0% |
| ist|der|die|und | 100 | 40 | 2.50 | 60.0% |
| \b\w+\b | 250 | 120 | 2.08 | 52.0% |
| (?:da|ge|om)+(?:n|me)* | 80 | 30 | 2.67 | 62.5% |
| \b(?(?=\w+ro)\w+pa|\w+lle)\w+\b | 570 | 250 | 2.28 | 56.1% |
| \b(\W)\1+\b|(^(?=.*kl)(?=.*no).{15,40}$) | 690 | 180 | 3.83 | 73.9% |
| ^.{4,32}(\P{N})\1{2,}.{4,32}(? | 220 | 60 | 3.67 | 72.7% |
| ^(\w{3,})(?!\1).*\h.*\1$ | 4410 | 1900 | 2.32 | 56.9% |
| ((\w{2,8},?(\P{Z}|\R)){1,2}\.\s?)$ | 4940 | 1230 | 4.02 | 75.1% |
| \b\w*?((.){1,3}\w*\2)\w*?(?1) | 1720 | 630 | 2.73 | 63.4% |
| \w*?(b{2,3})\w*?c | 1200 | 340 | 3.53 | 71.7% |
| \P{Lu}\P{L&}{0,12}[\s\-]{1,4}..[\P{L}\P{N}]{4} | 300 | 100 | 3.00 | 66.7% |
| \b(\B([c-h])\B|[a-z]+?(?1)[a-z]) | 930 | 300 | 3.10 | 67.7% |
| Average: | 1186 | 399 | 2.83 | 59.9% |
Conclusion
JIT compiling is a powerful technology, which is able to speed up interpreted execution including Java, JavaScript, ActionScript (with NanoJIT) or regular expression engines.Contribution
I would like to encourage all of you to join the PCRE Performance Project, and make PCRE faster than ever. My email address is hzmester (at) freemail (dot) hu.Special thanks
First and foremost, the author thanks Philip Hazel for his continual help and support.Sorted as familiy name in alphabetic order:
| Giuseppe D'Angelo | - | QRegularExpression in Qt 5 and above. | |
| Victor Julien | - | Suricata project (http://www.openinfosecfoundation.org/). | |
| Sheri Pierce | - | Testing and feedback. | |
| Petr Pisar |
| Last modification: 7.4.2012 |