Dear libstdc++ developers, I developed a novel algorithm called Teju Jagua which finds the shortest representation of a floating-point number. It can be used by std::to_chars which currently uses Ryu.
IMHO, Teju Jagua is much simpler than Ryu and, according to my measurements, for doubles, Teju Jagua is twice faster than Ryu. Similarly, it's simpler and ~27% faster than Dragonbox which, to the best of my knowledge, was the fastest algorithm for this task until now. Teju Jagua can bring another benefit. All the aforementioned algorithms use look-up tables that often raise memory footprint concerns. At least in their reference implementations, one or more tables is provided for each floating-point type. I don't know about the others but for Teju Jagua the table for a type can be also used for any smaller type. Hence, it just needs to provide the table for the largest type (e.g., float128_t) of the platform. I already started an implementation in libstdc++ and will send a subsequent RFC to gather feedback and to give you a better understanding of this work. The first RFC will replace Ryu only for floats. After this is reviewed and if there's still interest from others, I will send subsequent patches to replace Ryu for each one of the other floating-point types. In the end I'll send a patch to remove Ryu. (Note: This concerns only regular Ryu, which finds the shortest representation, Ryu-printf that produces the result for a given precision is out of the scope of this work and will remain.) A reference C implementation (it's a WIP which lacks documentation) is found in [1]. I presented this algorithm in C++ Now 2024, C++ on Sea 2024 and I'll show it again at CppCon 2024. If any libstdc++ developer plans to attend, I'd be more than happy to discuss this work in person at the conference. The video recording of my C++ Now 2024 talk is on YouTube [2]. [1] https://github.com/cassioneri/teju_jagua: [2] https://www.youtube.com/watch?v=w0WrRdW7eqg
