Semantic Predicates and Lookahead

[gquerret]

I maintain an ANTLR4 grammar for a language where one rule requires a lot of tokens for disambiguation. I have no control over the language specification so I have to deal with the "broken" syntax. I'd like to add a switch to have two modes, one where the broken syntax would be accepted, the other where it would be rejected. I did that with a semantic predicate, as explained in the ANTLR4 book:

classStatement:
   { acceptBrokenSyntax }? methodStatement1
|  { !acceptBrokenSyntax }? methodStatement2 // Trimmed down version of methodStatement1
| ... // other rules
;

The semantic predicate works perfectly in both cases, but when analyzing the content of Parser#getParseInfo()#getDecisionInfo(), I was surprised to see the SLL_MaxLook value was identical in both cases, as if the parser was trying both rules and then discarding one based on the predicate value. I've done a test with only methodStatement2 and no semantic predicate, resulting in a much lower SLL_MaxLook value and improved performance.

Is this the expected behavior ? Or is there another way of handling such a case, without the lookahead performance penalty ?

Context:

• ANTLR 4.9.3
• Java Target

I can provide an example if required.