converting float -8536070.f to string gives "-08536070" about glaze HOT 10 CLOSED

This is strange behavior and it doesn't conform to "the shortest possible representation", which is what the underlying algorithm is supposed to do. Glaze uses a slightly modified version of the code yyjson uses, and looking over the code I would expect this to also show up in yyjson, which is interesting.

Will have to be investigated further. But, the output is valid JSON and correct, so it isn't a major concern.

from glaze.

I do not think it is valid json. Looking at the railroad diagram on https://www.json.org/json-en.html it seems like a leading 0 must be followed by an optional fraction and optional exponent.

Also, it is rejected when deserializing back from json - I discovered this as part of a roundtrip test. So I think this should be labeled bug, not enhancement.

from glaze.

Yup, my bad. You're right this is a bug and broken.

from glaze.

Note: This case works for double, just not float. The original code from yyjson only handles double, so the optimizations for float using double parsing code introduce some incompatibility. Looking into it.

from glaze.

I'm moving to use std::to_chars with special conversions for inf and nan to null for JSON compliance. This should fix any serialization issues with floating point numbers.

from glaze.

Well... the performance for std::to_chars is abysmal at least on MacOS Arm. I'm going to figure out a faster solution, it just might take a bit longer.

from glaze.

I've switched Glaze to use the dragonbox algorithm underneath with new float handling. It solves this particular case and I haven't gotten it to error in my tests.

from glaze.

Unfortunately this still does not handle my roundtrip tests. It failed on the denormalized number 1.40129846e-45f which got serialized to {"value":1.E-45} which is rejected by the glaze deserializer.

from glaze.

It is work in progress, but you can view that particular test here: https://github.com/pauldreik/glaze/blob/fuzz/fuzzing/json_roundtrip_floating.cpp

from glaze.

Thanks for sharing your fuzz testing! Also, I've added better random float and double testing to Glaze in #1190.

#1190 also fixes the denormalized number serialization issue.

from glaze.