Just wondering... #449

I was just wondering why you chose to use Boost.Spirit classic (instead of boost::spirit::qi) for parsing JSON.

Probably not something that should be filed as a bug...though I'd assume there was a reason that no one will know any longer as satoshi isnt here to comment.

Yeah, my bad. It's unlikely, but since spirit classic is no longer being actively developed there may still be bugs hiding there that won't be addressed very quickly if ever. I mean, parsing JSON isn't really the most intensive parsing that I can think of, but spirit v2 would certainly reduce the amount of code needed here and you could get rid of all these nasty semantic actions and just parse directly into a more useful data structure. I'm not doubting the correctness of this parser, mostly because spirit classic parsers are almost too ugly to look at. I'll code up a JSON parser (EDIT: using spirit v2) this weekend sometime and do a performance comparison with this parser. That's the only potential reason I can think of for using spirit classic (since spirit v2 does a lot of default constructing of attributes by default). I'll send my results if you're interested. Also, I'm somewhat new to github, so advice on where comments like this one should go in the future? Should I email the maintainer directly?

Thanks for the response!

We aren't the ones using spirit v1, we use an JSON Spirit from http://www.codeproject.com/KB/recipes/JSON_Spirit.aspx You'll have to ask that guy. (though we do do some small modifications and are on an old version of that library)

On Wed, Aug 3, 2011 at 9:37 AM, TheBlueMatt reply@reply.github.com wrote:

We aren't the ones using spirit v1, we use an JSON Spirit from http://www.codeproject.com/KB/recipes/JSON_Spirit.aspx You'll have to ask that guy.

Ah yes, I see now. The message board there has the answers straight away:

""" I did look a spirit.qi around 18 months ago. I replaced the spirit::classic implementation with one using spirit::qi. The resultant code was very neat, no semantic actions were needed as an AST was generated automatically. However I found using Qi problematic:

1. Very steep learning curve.
2. Compilations took a very long time. Moreover compilation errors were of course obtuse due to the heavily templated code. I found that the only way I could often discover the cause of a compilation error was to comment out sections of code to isolate the error to a few lines. This was very time consuming due to the long compilation times.
3. The resultant code was actually a bit slower than the original classic Spirit. Probably someone knowing Qi well could have speeded it up considerably, but I am afraid that didn't have the time or inclination owing to 1 and 2 above.
4. I couldn't compile all my test code in one go because the compiler reached some internal limit. (JSON Spirit allows multiple Value types, Value, mValue etc. I could compile the test code for a single type but not all four at once.)

I have also created a version of JSON Spirit in which I hand wrote a recursive descent parser. This version is actually the fastest. I cannot remember how much faster, but it was by a certain percentage not by some factor. Parsing JSON with a recursive descent parser is particularly easy as you can tell what type the next token is just by looking at the next character.

Regards John """

Thanks,

Greg

Reply to this email directly or view it on GitHub: #449 (comment)