This demonstrates how to use re2, a regular expression library written in C++, through FFI based on WebAssembly with wazero.
Note: this is intended to be used as a demonstration for my talk at GopherCon 2022.
Here's a benchmark result in lib_test.go where we compare the standard regexp package with re2 via FFI with wazero:
$ benchstat stdlib.txt re2.txt
name old time/op new time/op delta
RegexpMatch/Hard/not_match/16B-10 3.39ns ± 1% 506.54ns ± 6% +14839.54% (p=0.008 n=5+5)
RegexpMatch/Hard/not_match/1KB-10 20.9µs ± 1% 0.6µs ± 3% -97.25% (p=0.008 n=5+5)
RegexpMatch/Hard/not_match/1MB-10 26.2ms ± 2% 0.0ms ±21% -99.93% (p=0.008 n=5+5)
RegexpMatch/Hard/not_match/32MB-10 835ms ± 1% 1ms ± 1% -99.89% (p=0.008 n=5+5)
RegexpMatch/Hard/match/16B-10 443ns ± 1% 525ns ± 1% +18.55% (p=0.008 n=5+5)
RegexpMatch/Hard/match/1KB-10 9.54µs ± 1% 4.05µs ± 0% -57.56% (p=0.008 n=5+5)
RegexpMatch/Hard/match/1MB-10 60.3ms ± 1% 3.6ms ± 1% -93.98% (p=0.008 n=5+5)
RegexpMatch/Hard/match/32MB-10 1.92s ± 0% 0.12s ± 0% -93.94% (p=0.008 n=5+5)
RegexpMatch/Hard1/not_match/16B-10 1.80µs ± 3% 0.52µs ± 1% -71.09% (p=0.008 n=5+5)
RegexpMatch/Hard1/not_match/1KB-10 107µs ± 1% 4µs ± 0% -96.22% (p=0.008 n=5+5)
RegexpMatch/Hard1/not_match/1MB-10 129ms ± 1% 4ms ± 1% -97.17% (p=0.008 n=5+5)
RegexpMatch/Hard1/not_match/32MB-10 4.10s ± 1% 0.12s ± 0% -97.16% (p=0.008 n=5+5)
RegexpMatch/Hard1/match/16B-10 304ns ± 2% 528ns ± 1% +73.71% (p=0.008 n=5+5)
RegexpMatch/Hard1/match/1KB-10 9.18µs ± 5% 4.06µs ± 1% -55.78% (p=0.008 n=5+5)
RegexpMatch/Hard1/match/1MB-10 167ms ± 2% 4ms ± 2% -97.83% (p=0.008 n=5+5)
RegexpMatch/Hard1/match/32MB-10 5.32s ± 1% 0.12s ± 1% -97.81% (p=0.008 n=5+5)
The benchmark cases are borrowed from the Go official test cases, and this indicates that the larger the input text is, the better wazero+re2 FFI performs. The reason why re2 FFI under-performs with the small text is that: it has to copy the text into the Wasm's linear memory region so that the re2 can reference there message in Wasm, which comes with the allocation each time. Therefore, the allocation can be dominant in such a case, and makes huge difference there.
Note: test cases are chosen somewhat artificially. This doesn't necessary recommend and state that wazero+re2 performs better than the standard Go library. This is just a demo on the potential of FFI with Wasm!