Feature #21308: Replacing the Float#to_s (dtoa.c) implementation with a modern algorithm

• Status: Open
• Assignee: -
• Target version: -

[watson1978 (Shizuo Fujita)]

This is a feature request to replace Ruby's Float#to_s implementation with a modern high-performance float-to-string conversion algorithm such as Grisu.

Currently, Float#to_s in Ruby uses the BSD-derived implementation in missing/dtoa.c.
Meanwhile, json gem has adopted a Grisu-based implementation in ruby/json#768, it has been improve the performance of float-to-string conversion dramatically.

It appears to have brought about a 10 times improvement in performance.
However, the improvement is limited to the json gem.

If we can use the same implementation in Float#to_s, it will introduce the improvement in all float-to-string conversion cases.

(In additional, it would be better if an API could be added that could be called directly from the C extension library, it will reduce the overhead of the call.)

[nobu (Nobuyoshi Nakada)]

The interface of fpconv_dtoa() looks quite different from dtoa() in missing/dtoa.c.
Rather grisu2() itself is closer?

[radiospiel (Enrico Thierbach)]

The change in https://github.com/ruby/json/pull/768 was an easy win, because the previous implementation was calling into questionFloat#to_s via rb_funcall. The overhead of doing so is so extreme, that any implementation that would skirt that would show a massive improvement already.

In fact, a first attempt using dtoa.c already showed a 600% or so performance improvement on a clinical example.

I then decided that maybe this would also be a great moment to look at alternative implementations, since the flexibility provided by dtoa.c is not necessary in the context of the json gem. Other motivational factors have been:

• dtoa generates output right to left (IIRC), but, at least for an integration with the json gem, generating left to right saves a memcpy, which would add a 5-10% overhead;
• more modern algorithms are readily available.

I believe integrating a more modern implementation in the ruby runtime certainly is benefitial, but the gains are not as massive as @watson1978 (Shizuo Fujita) hopes to achieve here, unfortunately.

[watson1978 (Shizuo Fujita)]

Looks like we are about to get yet another vendored implementation of the same kind: https://github.com/ohler55/oj/pull/1095 proposes vendoring the same Boost-licensed fpconv (Grisu2) code that the json gem already ships.
I think it is time to address this in Ruby itself rather than gem by gem.