Dmitry's Q1 2024 Update

By Dmitry Arkhipov on

Apr 29, 2024

In the first quarter of 2024 I mostly continued with work on direct serialization, cleanup work in Boost.JSON, and experiments with its JSON parser and serialiser. The experiments have resulted in discovery of a significant performance pessimization in the parser. The fix will soon be merged into the project.

Another experiment that will eventually bring significant performance improvement is replacing JSON’s number serialisation code with usage of the Charconv library, which was recently accepted into Boost. That library has functions for efficient parsing and serialisation of numbers, but Boost.JSON can’t use the parsing code for its default number parsing mode due to several requirements, the main one being streaming parsing. The precise number parsing mode, on the other hand, is already using an earlier embedded version of Charconv. The addition of dependency on Charconv has been delayed, though, due to it not supporting a few old compilers which are supported by JSON. This summer a major user of those compilers (an old Red Hat version) goes EOL, and I decided to wait until that point to deprecate support for those compilers.

One small but important change was replacing internal configuration for endianness and instead relying on Boost.Endian for that. The original reason Boost.JSON had such internal configuration to begin with was the standalone mode, which was removed a while ago. But other changes were of higher priority and thus, I only got around to this now.

One interesting JSON issue fixed was related to conversion of filesystem::path objects (https://github.com/boostorg/json/issues/975). Due to unfortunate decisions made many years ago a path is a sequence of itself. Treating paths as sequences thus results in infinite recursion. The fix added dedicated handling for paths and just in case forbade recursive sequences altogether.