[antlr-interest] Actions and non-LL(*) rules ??

[Marcin Rzeźnicki]

2010/5/3 Marcin Rzeźnicki <marcin.rzeznicki at gmail.com>:
> 2010/5/3 Terence Parr <parrt at cs.usfca.edu>:
>> it still won't execute the {{...}}} before the pred is evaluated during
>> prediction. it must not eval disambiguating preds after actions in any
>> circumstance.
>> Ter
>
> I see. So that poses yet another problem with recursive tree-like
> definitions. If I got production like
>
> s returns[Tree tree]
> :
> a1=a {$tree = $a1.tree; }  (DELIMITER a2=a {$tree = new Tree($tree,$a2.tree);})*
>
> AND there is an ambiguity in whether to enter the loop or leave it (it
> creates predicate under the hood) then the first action hides that
> predicate. Then the solution is to explicitly check in the second
> action whether current iteration is the first one, which is pretty
> ugly and imposes performance hit as this means that this test is
> repeated on every loop roll, or unroll the loop manually one time
> which is ugly too :-/ I suppose there is no syntax which marks an
> action as predicate-safe? :-)
>
All this is harder than it sounds, I see - as lot of problems emerge
from automatically generated ANTLR predicates. In presence of this
discovery my question is: how can you deal with actions and implicit
predicates? I cannot "move" the predicate to the left obviously, and
this "hiding" rule seems to be really constraining in this case.

-- 
Greetings
Marcin Rzeźnicki