azmq: replace boost.regex with std

[timblechmann]

we can reduce the dependencies onto a boost.regex library by using std::regex and friends.

furthermore we can cache the compiled regex instead of building it every time

[aboseley]

This library is intrinsically coupled to boost, what are the benefits from switching from boost to std::regex. Std::regex is know for poor performance, thou this shouldn't be an issue in a bind call which is rarely a 'fast-path' operation.

[timblechmann]

what are the benefits from switching from boost to std::regex.

main benefit is binary size. shared libs of boost.regex are hundreds of kb, where std::regex is contained in the runtime library already.

Std::regex is know for poor performance

out of curiosity: do you have any reference for this?

[aboseley]

I've seen std::regex mentioned in few blog posts as an example of something that needs improvement but can't be change because it will break the ABI , which is a real shame. I don't think the boost library 'suffers' from the ABI compatibility constraint, so it can fix things like this.

One reference is this, but its light on detail wrt regex https://cor3ntin.github.io/posts/abi/

I've run a quick bench mark this morning for fun. https://github.com/aboseley/benchmark-regex

BM_MatchWithBoost       3786 ns         3785 ns       184163
BM_MatchWithStd        49727 ns        49704 ns        14020

[timblechmann]

out of curiosity: which compiler/stl implementation do you use?

appleclang-11/libcxx seems to be a little slower, but it's nowhere as extreme as what you're seeing

BM_MatchWithBoost       4793 ns         4791 ns       108551
BM_MatchWithStd         5671 ns         5670 ns       117849

[aboseley]

g++ (gcc) 11.1.0 on linux

[timblechmann]

at the risk of comparing apples with oranges, i've tweaked your benchmark program to cache the compiled regular expression (by making them static) ... after all, we don't need to re-compile the regex every time we want to match a string.

in this case std is actually faster than boost (gcc-10/linux)

BM_MatchWithBoost       99.3 ns         99.3 ns      6879716
BM_MatchWithStd         88.8 ns         88.8 ns      7841490

so it seems that constructing the regex fsm from the string is faster in boost, but evaluating the fsm is faster in the stl

[aboseley]

lies, statistics and benchmarks.

switching to libc++ and clang

------------------------------------------------------------
Benchmark                  Time             CPU   Iterations
------------------------------------------------------------
BM_MatchWithBoost       2902 ns         2900 ns       240455
BM_MatchWithStd          997 ns          996 ns       69198

Honestly the speed of this code isn't that much of issue. I can't image the speed of a bind call being an issue

[aboseley]

If nothing else this PR reduces the binary size again a bit further

[aboseley]

Thank-you for the PR