[antlr-interest] comments on suggestion to reduce tree constr uction code size

[mzukowski at bco.com]

OK, I buy all that.  

I think someone else recommended separate methods to use when guessing v.
when not.  Would be interesting to see how much faster that would be than
just eliminating guards.  Would be significantly bigger code size.
Shouldn't be hard to add though.

Monty

> -----Original Message-----
> From: Terence Parr [mailto:parrt at jguru.com]
> Sent: Wednesday, December 26, 2001 10:50 AM
> To: antlr-interest at yahoogroups.com
> Subject: Re: [antlr-interest] comments on suggestion to reduce tree
> constr uction code size
> 
> 
> 
> On Wednesday, December 26, 2001, at 07:16  AM, 
> mzukowski at bco.com wrote:
> 
> >> Parsing all of ANTLR code itself without gates is 17sec
> >> on my G4 450Mhz OS X box versus the regular 19sec
> >
> > I'm scratching my head as to why the version without gates 
> ran faster 
> > than
> > the one with gates?  Are the if()s taking more time than 
> just creating 
> > the
> > AST node?
> 
> That's what I thought at first until I realized that the gates are 
> always being executed while the wasted ast creation is only 
> done while 
> "guessing", which by design is done infrequently.  Somehow, 
> removing the 
> gates wins over the infrequent but expensive memory allocation.  
> Probably depends a lot on the garbage collector's efficiency at 
> reallocating objects of a common.
> 
> > To make this scheme work, you would have to make sure that your 
> > initialize()
> > methods in your node class were side-effect free.  Folks with 
> > heterogenous
> 
> Very true.  But that makes sense I'd say.  We could make it a rule :)
> 
> > trees might be more inclined to do table adds in their initialize(),
> > perhaps?  I suppose that the guessing state could be passed to the
> > initializer....or for that matter to the factory which 
> could then just 
> > not
> > build a tree, right?
> 
> Sure...ASTFactory.create() could also just ask the parser 
> (have to add a 
> ref to the parser in there though).  That would be like 
> executing all of 
> the gates again though.
> 
> > I like that better.  astFactory.create() returns nullAST if 
> currently
> > guessing.  Will addASTChild and friends deal with that ok?  
> Will a java
> 
> Hm...probably not.
> 
> > optimizer cache inputState.guessing==0 for those tests?  Probably 
> > not--can't
> 
> maybe not across method calls.
> 
> > guarantee another thread won't muck with it.  The generated 
> code could 
> > have
> > a boolean in the method, updated only upon entry to the 
> method and on 
> > every
> > guessing++ and guessing--.
> 
> I'm thinking we can get away with saying that creation of nodes (even 
> hetero nodes) must be side effect free.
> 
> Ter
> --
> Chief Scientist & Co-founder, http://www.jguru.com
> Creator, ANTLR Parser Generator: http://www.antlr.org
> 
> 
>  
> 
> Your use of Yahoo! Groups is subject to 
> http://docs.yahoo.com/info/terms/ 
> 
> 
> 

 

Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/