I tripped over this in an earlier version of ANTLR 3 in which sempreds were broken. The template code is "correct" in the Java version; I can't remember how this gets broken, but you need to keep on looking.
--Loring ________________________________ From: Sam Harwell <[email protected]> To: [email protected] Sent: Tuesday, March 3, 2009 4:06:16 PM Subject: [antlr-dev] cyclicDFAState template incompatible with unbuffered tree node streams Taken from Java.stg, we end up with a mismatched mark/rewind, which causes the temporary stream buffer (FastQueue) to flush during a predict operation that encounters predicates. /** A state in a cyclic DFA; it's a special state and part of a big switch on * state. */ cyclicDFAState(decisionNumber,stateNumber,edges,needErrorClause,semPredState) ::= << int LA<decisionNumber>_<stateNumber> = input.LA(1);<\n> <if(semPredState)> <! get next lookahead symbol to test edges, then rewind !> int index<decisionNumber>_<stateNumber> = input.index(); input.rewind();<\n> <endif> s = -1; <edges; separator="\nelse "> <if(semPredState)> <! return input cursor to state before we rewound !> input.seek(index<decisionNumber>_<stateNumber>);<\n> <endif> if ( s>=0 ) return s; break; >> For now I noticed it “works” to simply add a call input.mark() right after the call to input.rewind(), but I don’t think that’s the proper solution. Sam
_______________________________________________ antlr-dev mailing list [email protected] http://www.antlr.org/mailman/listinfo/antlr-dev
