optimizing mimetype serialization

[mikea]

kj::strTree results in plenty of heap allocation. Using capacity-bound string buffer to do only one allocation per serialize

before:

----------------------------------------------------------------------
Benchmark                            Time             CPU   Iterations
----------------------------------------------------------------------
Mimetype::ParseAndSerialize       2.18 us         2.18 us       324082
Mimetype::Serialize              0.351 us        0.351 us      1982953

after:

----------------------------------------------------------------------
Benchmark                            Time             CPU   Iterations
----------------------------------------------------------------------
Mimetype::ParseAndSerialize       1.48 us         1.48 us       456517
Mimetype::Serialize              0.114 us        0.114 us      6175365

[mikea]

updated string buffer to be unbounded and to grow on the heap. ptal @kentonv @jasnell

[irvinebroque]

Does new benchmarking tooling we added gives us a way to fail CI if a given benchmark regresses more than a given percentage? (like if something shot up 400%)

[mikea]

Does new benchmarking tooling we added gives us a way to fail CI if a given benchmark regresses more than a given percentage? (like if something shot up 400%)

eventually this will be possible. Right now it is just an instrument to measure a single change impact. I plan to start collecting these on TC for a longer term trend

[ohodson]

nice, lgtm with James suggestions.

[jasnell]

Non-blocking... and just thinking out loud.... In Node.js we have a utility type called MaybeStackBuffer<T> that encapsulates this Maybe-Stack-Maybe-Heap pattern. It's a lot like KJ_STACK_ARRAY but in a form that is easy to define and use as class members like this. It might be worthwhile to adopt a version of that for our use.