Wrong constant used

[grisumbras]

While working on boostorg/json#993 I noticed that JSON benchmark performance when using charconv::from_chars_erange has degraded since the original release (by ~8% for canada.json and ~6% for numbers.json). I've pinned down the degradation to 9b2e067. It seems that the reason is the log2 table, because I tried putting it in a cpp file and performance increased. Unfortunately, this couldn't be put in a PR, because the function using it is constexpr. Maybe you will find a solution.

Anyway, while looking at the commit, I noticed that this line looks wrong: https://github.com/boostorg/charconv/blob/develop/include/boost/charconv/detail/from_chars_integer_impl.hpp#L225. The value should presumably be std::numeric_limits<Integer>::digits. If I understand it correctly, this bug can also cause performance degradation.

[mborland]

You're right about changing digits10 to digits. Is it safe to assume your data is entirely in base10?

[grisumbras]

I actually don't use this function at all (I only use from_chars_erange). It seems the impact is purely indirect, from the code layout changes due to the table.

[mborland]

I padded the bottom of the table so that instead of 37 entries it has 64, and ran benchmarks before and after. Got ~5.5% improvement for from_chars for doubles which is inline with your observation.

[grisumbras]

I can test a PR with JSON's benchmark.

[mborland]

I can test a PR with JSON's benchmark.

See #276

[grisumbras]

To be honest, I don't see any improvement in JSON's benchmarks.

[mborland]

Can you try the linked PR again when you have a chance? I tried a different technique which is used all over the library.

[grisumbras]

Didn't help either. Let's just fix the constant for now.