Re: Getting involved in Bison

Hi Victor!

> Le 20 févr. 2020 à 04:07, Morales Cayuela, Victor (NSB - CN/Hangzhou) <[email protected]> a écrit :
> 
> Hi!
> 
>> It really depends what you'd like to do.  You reported you're fluent in C++.  Something quite interesting would be to enable push-parser in lalr1.cc.  It's more ambitious than the subcomplain task, but it will probably take you less time, because it's more compact, it will impact much less code.
> 
> Ok, seems very interesting! I use C++ (currently C++17) in my daily job as a developer and I always try to keep up with the latest of the standard.
> I will familiarize myself first with push/pull parsers and then I will start with it 😊 I will also take a look at TODO to check if there is already any specification about how it should look like.

I think all the information you need is in the manual, in the push example in C (see examples/c/pushcalc), and in yacc.c's implementation of push.  Actually, Java also has push-parser support, and is probably a better source of inspiration than C.

I recommend that you first build by hand a push parser before trying to get this done by m4.  For instance take examples/c++/calc++, and change the generated parser so that it works as a push parser.  Keep this preciously somewhere to make sure "make" doesn't kill it.  When it works, you can show us a diff between the regular pull parser and the push parser, and when we agree, we can proceed to make it work in m4.

Actually, maybe examples/c++/calc++ is too bloated: first create a very simple example in C++, similar to examples/c/calc.  Start from this one, it is much simpler, and yet sufficient.

Most of the effort, I guess, is finding what local variables must be turned into member variables, to survive across calls to yyparse (amusingly enough coroutines would make this easy, but we can't afford them yet ;-).  But lalr1.java should help.

Note that C++ allows to consider other options than the C parsers.  Since the C++ parser is an object, maybe we can have push and pull use the same object.  I mean, *maybe*, we can *always* move most local variables into member variables, even for regular pull-parsing.  Of course it's technically possible, the "maybe" refers to the runtime cost: *if* benchmarks show that it is not a regression to move to a model with more member variables and less local variables, then the m4 part will be much easier.

Maybe you should start with that: find a means to benchmark two pull parsers: one off-the-shelf, generated by today's lalr1.cc, and another one where the local variables that need to be member variables in push-mode are made member variables.

How does that sound?