Jon, if C++ memory management overheads are what really bogs down your parsers, you might want to take a page out of Scott McPeak's playbook:
His Elkhound-generated C++ parsers do not free memory nor do they integrate GC. Instead, you just allocate from a dedicated region, copy out the AST after parsing, and kill the entire region in one operation. Maybe that might help. Then again I like Rats!-generated parsers to be faster :) Robert On Nov 23, 2009, at 2:53 PM, Jon Rafkind wrote: > Hello, > I have written a PEG generator in python that is currently capable > of outputing C++ and python. Mostly I use the C++ version so that has > been heavily worked on, the python output works but probably has some > bugs in it. This is part of a game I am working on, you can get the > generator here: > > svn co > https://paintown.svn.sourceforge.net/svnroot/paintown/trunk/src/mugen/parser > > peg.py is the generator, the .peg files are input written in mostly EBNF. > > I plan to generate more languages in the future (see the TODO at the > top). If anyone is interested in this code don't hesitate to ask me > about it. > > Miscellaneous: > 1. Generating an AST with a C++ PEG necessitates using a garbage > collector (details upon request) so I had to litter the code with > annoying pointer saving operations. Screwing up these little details > causes massive bugs later on! > 2. I implemented a few optimizations from the Rats! paper such as chunks. > 3. Rules are parameterized over values instead of other rules, so you > can say something like > foo = "a" "b" bar($1) > bar(x) = ... {{ x }} > > And 'x' will be "a". This lets me implement tail recursion nicely so I > don't run out of stack space. In particular I can implement infix math > expressions that uses up stack space proportional to the depth of > nesting (as opposed to the length of the expression). > 4. The infix calculator example I referenced gets about 800kb/s on my > laptop while an equivalent version using Rats! gets 1.7mb/s. I did as > much optimizing as I could but I guess Java is just better than C++! > Actually the reason the C++ version is so slow is because I do a lot of > stack copying (again the GC in Java is a win here). > > _______________________________________________ > PEG mailing list > PEG@lists.csail.mit.edu > https://lists.csail.mit.edu/mailman/listinfo/peg _______________________________________________ PEG mailing list PEG@lists.csail.mit.edu https://lists.csail.mit.edu/mailman/listinfo/peg
