[PATCH] Replace ecvt() in FCollada with something else

[historic_bruno]

FCollada uses the deprecated POSIX function ​ecvt to convert floats to strings, but it's not available on FreeBSD and it would be nice to have a portable replacement. The result needs to be compatible with the ​COLLADA schema which uses XML's ​xs:double type for floats.

Possible solutions:

• ​std::stringstream, or
• ​snprintf perhaps with the 'g' flag, if it can be massaged into a xs:double-compatible format
• MIT-licensed ecvt replacement: ​http://piumarta.com/software/fcvt/

It's a good idea to avoid ecvt implementations that depend on dtoa, because it's also not available on FreeBSD and there are numerous buggy versions of that code around which apparently ​break with compiler optimizations.

[historic_bruno]

We could do what OpenCOLLADA does, using their implementation of ​dtoa. It's fairly readable and the main benefit is we know it works for OpenCOLLADA.

The problem with sprintf and similar, in my opinion, is they're just not exacting enough to meet the spec.

[Markus]

I just copied the FloatToString and the corresponding append function to a small test program. For me it does not support negative exponents and the format does not conform to the "canonical representation" of the document linked above.

What are these functions actually used for? Only development?

Achieving a "canonical representation" should be possible with some simple math and formating... I will try this WE.

[historic_bruno]

I don't believe we use these functions for anything at the moment, I assume they are for exporting COLLADA documents or something, but if we do in the future, it would be preferable to not have them broken :)

FWIW, it looks like the FreeBSD ​port of 0 A.D. copy-pasted ​some code from a forum, but I don't know if they actually confirmed the license is compatible with FreeBSD, so that's naughty and I also don't know that the code works beyond compiling.

[Markus]

compare some output formattings

[Markus]

FUStringBuilder.hpp exists twice

[historic_bruno]

Interesting, so it looks like the ecvt implementation only keeps 6 digits and anything on the order of 1e-8 or smaller becomes 0. That's not canonical xs:double but it's valid? I guess we could check if Collada importers can also handle negative exponents without trouble.

[Jorma Rebane]

Taking a look at this patch, I think we're really missing the point, which is to print valid numbers (that make sense) into a Collada file, no?

In this case, any value that is ridiculously small (for example, smaller than +/- 0.00001) should be considered as 0.0; For that matter, ridiculous large float values should also be ignored and set to NAN - because they're obviously invalid vertex coordinates.

The goal is not to produce a float print function that works for every universal solution - the goal is to get valid and reasonable output. In this case NAN, +/- 0.00001 to +/i 10000000.0.

In this case, we can also get away with sprintf(buffer, "%d", ...);

[nf]

Hi folks,

I have successfully built fcollada using

​https://code.wildfiregames.com/P186

It uses musl libc (MIT licensed) implementation of ecvt(). Let me know what you think.

[Stan]

In 24323:

Fix FCollada on platforms without ecvt

Based on a patch by: @NF
Tested by: @nephele
Tested on FreeBSD, Windows, Kubuntu, macOS
Fixes: #1325
Differential Revision: ​https://code.wildfiregames.com/D2399