[antlr-interest] disambiguating predicates / wrong decision code

[Davis, Alan]

I'm trying to use a disambiguating semantic predicate to differentiate expressions involving different-class identifiers. For example "A1+A2" is an "expr" while "V1+V2" is a "vexpr".  The problem is ANTLR seems to incorrectly predict a valid sentence, resulting in a syntax error, even though there is no warning about an ambiguity or anything.

Here is the grammar:

-----------------------------
grammar vexpr;
options { language = C; }

program : stat+ ;

stat  : vid '=' rhs ';'
      ;

rhs : vid             // handles V2
    | vexpr           // handles V2 + V3
    | expr            // handles(?) A1 + A2 
    ;

vexpr : vid '+' vid
      ;

expr: id
    | id '+' id
    ;

// IDs that start with 'V'
vid : { *(LT(1)->getText(LT(1))->chars) == 'V' }? ID 
    ;   

// IDs that don't start with 'V'
id  : { *(LT(1)->getText(LT(1))->chars) != 'V' }? ID 
    ; 

ID  :   ('a'..'z'|'A'..'Z'|'_') ('a'..'z'|'A'..'Z'|'0'..'9'|'_')* 
    ;

WS  :   (' ' | '\t' | '\r' | '\n')+ { $channel=HIDDEN; } 
    ;  

------------------
(The asymmetry in handling atoms is an artifact of the cutdown; if I move vid from rhs to vexpr this problem goes away).

The parser cannot parse the following input: V1 = A1 + A2 ;

When I look at the generated code for rhs (below), it seems to predict alternative 2 (vexpr) if LA(2) is '+' (8).  But '+' in that position is not sufficient to disambiguate alternative 2 from alternative 3!

This seems like a bug -- or am I missing something.

Here is the decision code (reformatted for brevity):

switch ( LA(1) )
{        
   case ID:
   {         
      if ( (LA(2) == 8) )   // '+'
         alt2=2;      // vexpr

      else if ( (( *(LT(1)->getText(LT(1))->chars) == 'V' )) )
         alt2=1;      // vid 

      else if ( (( *(LT(1)->getText(LT(1))->chars) != 'V' )) )
         alt2=3;      // expr

      else               
      {                  
         CONSTRUCTEX();
         ...      

- Alan Davis
  Texas Instruments