Resolve all groups when performing match

[suhothayan]

trafficstars

By default, RE2J does not resolve all the groups when performing a match operation, rather it only resolves groups when Matcher.group(*) is called. This results in input being processed twice. Once during find() and then again during group(*). This increase the latency when Matcher.find() and Matcher.group(*) are called in succession.

This Pull Request attempts to reduce the latency by half by resolving the groups when Matcher.find() is called, such that groups can be served from cache when Matcher.group(*) is called.

We can clearly see from the following benchmark tests that when resolveGroups == true (When the groups are resolved during find()) and when successMatch==ture (When the patterns are matched) for binary data we are seeing 0.609 ms/op vs 1.147 ms/op, providing 47% latency gain. Similarly, when the successMatch==false (When there are no matching patterns) we are only seeing 4% regression (0.332 ms/op vs 0.319 ms/op). When string is used as the input, resolving groups during find() seems to outperform in all cases.

Benchmark                           (binary)  (impl)  (resolveGroups)  (successMatch)  Mode  Cnt  Score   Error  Units
BenchmarkSubMultiMatch.findDomains      true     JDK             true            true  avgt    5  1.636 ± 0.095  ms/op
BenchmarkSubMultiMatch.findDomains      true     JDK             true           false  avgt    5  1.338 ± 0.036  ms/op
BenchmarkSubMultiMatch.findDomains      true     JDK            false            true  avgt    5  1.539 ± 0.052  ms/op
BenchmarkSubMultiMatch.findDomains      true     JDK            false           false  avgt    5  1.336 ± 0.039  ms/op
BenchmarkSubMultiMatch.findDomains      true    RE2J             true            true  avgt    5  0.609 ± 0.023  ms/op
BenchmarkSubMultiMatch.findDomains      true    RE2J             true           false  avgt    5  0.332 ± 0.008  ms/op
BenchmarkSubMultiMatch.findDomains      true    RE2J            false            true  avgt    5  1.147 ± 0.071  ms/op
BenchmarkSubMultiMatch.findDomains      true    RE2J            false           false  avgt    5  0.319 ± 0.007  ms/op
BenchmarkSubMultiMatch.findDomains     false     JDK             true            true  avgt    5  1.430 ± 0.311  ms/op
BenchmarkSubMultiMatch.findDomains     false     JDK             true           false  avgt    5  1.192 ± 0.031  ms/op
BenchmarkSubMultiMatch.findDomains     false     JDK            false            true  avgt    5  1.339 ± 0.064  ms/op
BenchmarkSubMultiMatch.findDomains     false     JDK            false           false  avgt    5  1.196 ± 0.009  ms/op
BenchmarkSubMultiMatch.findDomains     false    RE2J             true            true  avgt    5  0.610 ± 0.022  ms/op
BenchmarkSubMultiMatch.findDomains     false    RE2J             true           false  avgt    5  0.329 ± 0.004  ms/op
BenchmarkSubMultiMatch.findDomains     false    RE2J            false            true  avgt    5  1.100 ± 0.014  ms/op
BenchmarkSubMultiMatch.findDomains     false    RE2J            false           false  avgt    5  0.337 ± 0.039  ms/op

As there are use cases only Matcher.find() being called, and in such cases resolving groups can be an overkill, therefore I have enabled this optimization via a flag so we can tap into this optimization only when it's needed.

Please provide your feedback on the pull request. I'm happy to make necessary changes to get this optimization merged.

[suhothayan]

Any feedback on this pull request? Can we get this merged to the master?