[antlr-interest] Re: nondeterminism warning?

[Xue Yong Zhi]

To understand this warning you have to be very familiar with "linear 
approximate lookahead".

I recommend the following articles:

http://www.antlr.org/doc/glossary.html#Linear_approximate_lookahead
and related entries in antlr's FAQ.
http://seclib.blogspot.com/2005/11/linear-approximate-lookahead.html

In your case, antlr needs to compute exit branch to know when to leave a 
loop, and it will use linear approximate lookahead(instead of full 
LL(K)) to do that. Normally there are several exit paths for a rule, for 
example:
"[ assembly {"
and "[ xxx :"

With linear approximate lookahead, antlr will compress them so you get 
the warning for "[ assembly  :".

You can check the generated code so if it is actually correct(there are 
a great chance it does even with the warning), if so we can shutdown the 
warning. Otherwise, use semantic predict or syntax predict.

Dima wrote:
>>> 
>>>With this part, I become this warning:
>>>    nondeterminism upon k==1:LBRACK k==2:"assembly" between alt 1 and 
>>>exit branch of block
>>> 
>>
>>Please be more clear with your question, which rule causes this 
>>warning(copy the entire antlr output please)?
> 
> 
>>If the warning comes from c_unit, it is because global_attributes and 
>>attribute_section can both start with LBRACK "assembly", and antlr can 
>>not distinguish when k = 2.
>>
> 
> 
> Ok sorry, with k=2 it can't work, but I use k>=3.
> It's true, the warning comes from c_unit.
> With this rules global_attributes must be: 
> 	[ assembly : ...   (for k=3)
> But attributes can be:
> 	[ assembly ] ... or
> 	[ assembly , ... or
> 	[ assembly ( ... or
> 	[ type :     ...
> That's all not the same 
> 
> If I make attributes rule like this:
> 	LBRACK "type" COLON RBRACK
> I get the same warning.
> Antlr output:
> 	test1.g:39: warning:nondeterminism upon
> 	test1.g:39: 	k==1:LBRACK
> 	test1.g:39: 	k==2:"assembly"
> 	test1.g:39: 	k==3:COLON
> 	test1.g:39: 	between alt 1 and exit branch of block
> 
> 
> Thanks!
> 
> Dima
> 
> 


-- 
Xue Yong Zhi
http://seclib.blogspot.com