The add method of the ATNConfigSet.java class has code that looks like a bug.

[yososs]

The field used to calculate the HashCode is changed after the Hash container is registered.

antlr4/runtime/Java/src/org/antlr/v4/runtime/atn/ATNConfigSet.java

Line 146 in fcab02c

ATNConfig existing = configLookup.getOrAdd(config);

   public boolean add(
       ATNConfig config,
       DoubleKeyMap<PredictionContext,PredictionContext,PredictionContext> mergeCache)
   {
       if ( readonly ) throw new IllegalStateException("This set is readonly");
       if ( config.semanticContext!=SemanticContext.NONE ) {
           hasSemanticContext = true;
       }

       if (config.getOuterContextDepth() > 0) {
           dipsIntoOuterContext = true;
       }

       ATNConfig existing = configLookup.getOrAdd(config);
       if ( existing==config ) { // we added this new one
           cachedHashCode = -1;
           configs.add(config); // track order here
           return true;
       }

       // a previous (s,i,pi,_), merge with it and save result
       boolean rootIsWildcard = !fullCtx;
       PredictionContext merged =
           PredictionContext.merge(existing.context, config.context, rootIsWildcard, mergeCache);
       // no need to check for existing.context, config.context in cache
       // since only way to create new graphs is "call rule" and here. We
       // cache at both places.
       existing.reachesIntoOuterContext =
           Math.max(existing.reachesIntoOuterContext, config.reachesIntoOuterContext);

       // make sure to preserve the precedence filter suppression during the merge
       if (config.isPrecedenceFilterSuppressed()) {
           existing.setPrecedenceFilterSuppressed(true);
       }

// here !!
       existing.context = merged; // replace context; no need to alt mapping
       return true;
   }