• Forward movement: The value of index() before calling this method is less than the value of index() after calling this method.
• Ordered lookahead: The value of LA(1) before calling this method becomes the value of LA(-1) after calling this method.

This is a one way link. It emanates from a state (usually via a list of transitions) and has a target state.

Since we never have to change the ATN transitions once we construct it, we can fix these transitions as specific classes. The DFA transitions on the other hand need to update the labels as it adds transitions to the states. We'll use the term Edge for the DFA to distinguish them from ATN transitions.

The default implementation returns false .

This event may be reported during SLL prediction in cases where the conflicting SLL configuration set provides sufficient information to determine that the SLL conflict is truly an ambiguity. For example, if none of the ATN configurations in the conflicting SLL configuration set have traversed a global follow transition (i.e. is false for all configurations), then the result of SLL prediction for that input is known to be equivalent to the result of LL prediction for that input.

In some cases, the minimum represented alternative in the conflicting LL configuration set is not equal to the minimum represented alternative in the conflicting SLL configuration set. Grammars and inputs which result in this scenario are unable to use , which in turn means they cannot use the two-stage parsing strategy to improve parsing performance for that input.

In some cases, the unique alternative identified by LL prediction is not equal to the minimum represented alternative in the conflicting SLL configuration set. Grammars and inputs which result in this scenario are unable to use , which in turn means they cannot use the two-stage parsing strategy to improve parsing performance for that input.

Parsing performance in ANTLR 4 is heavily influenced by both static factors (e.g. the form of the rules in the grammar) and dynamic factors (e.g. the choice of input and the state of the DFA cache at the time profiling operations are started). For best results, gather and use aggregate statistics from a large sample of inputs representing the inputs expected in production before using the results to make changes in the grammar.

If DFA caching of SLL transitions is employed by the implementation, ATN computation may cache the computed edge for efficient lookup during future parsing of this decision. Otherwise, the SLL parsing algorithm will use ATN transitions exclusively.

If the ATN simulator implementation does not use DFA caching for SLL transitions, this value will be 0.

Note that this value is not related to whether or not may be used successfully with a particular grammar. If the ambiguity resolution algorithm applied to the SLL conflicts for this decision produce the same result as LL prediction for this decision, would produce the same overall parsing result as .

If DFA caching of LL transitions is employed by the implementation, ATN computation may cache the computed edge for efficient lookup during future parsing of this decision. Otherwise, the LL parsing algorithm will use ATN transitions exclusively.

If the ATN simulator implementation does not use DFA caching for LL transitions, this value will be 0.

The executor tracks position information for position-dependent lexer actions efficiently, ensuring that actions appearing only at the end of the rule do not cause bloating of the created for the lexer.

If speculative is true , this method was called before for the matched character. This method should call before evaluating the predicate to ensure position sensitive values, including , , and , properly reflect the current lexer state. This method should restore input and the simulator to the original state before returning (i.e. undo the actions made by the call to .

This value is the sum of and .

The basic complexity of the adaptive strategy makes it harder to understand. We begin with ATN simulation to build paths in a DFA. Subsequent prediction requests go through the DFA first. If they reach a state without an edge for the current symbol, the algorithm fails over to the ATN simulation to complete the DFA path for the current input (until it finds a conflict state or uniquely predicting state).

All of that is done without using the outer context because we want to create a DFA that is not dependent upon the rule invocation stack when we do a prediction. One DFA works in all contexts. We avoid using context not necessarily because it's slower, although it can be, but because of the DFA caching problem. The closure routine only considers the rule invocation stack created during prediction beginning in the decision rule. For example, if prediction occurs without invoking another rule's ATN, there are no context stacks in the configurations. When lack of context leads to a conflict, we don't know if it's an ambiguity or a weakness in the strong LL(*) parsing strategy (versus full LL(*)).

When SLL yields a configuration set with conflict, we rewind the input and retry the ATN simulation, this time using full outer context without adding to the DFA. Configuration context stacks will be the full invocation stacks from the start rule. If we get a conflict using full context, then we can definitively say we have a true ambiguity for that input sequence. If we don't get a conflict, it implies that the decision is sensitive to the outer context. (It is not context-sensitive in the sense of context-sensitive grammars.)

The next time we reach this DFA state with an SLL conflict, through DFA simulation, we will again retry the ATN simulation using full context mode. This is slow because we can't save the results and have to "interpret" the ATN each time we get that input.

CACHING FULL CONTEXT PREDICTIONS

We could cache results from full context to predicted alternative easily and that saves a lot of time but doesn't work in presence of predicates. The set of visible predicates from the ATN start state changes depending on the context, because closure can fall off the end of a rule. I tried to cache tuples (stack context, semantic context, predicted alt) but it was slower than interpreting and much more complicated. Also required a huge amount of memory. The goal is not to create the world's fastest parser anyway. I'd like to keep this algorithm simple. By launching multiple threads, we can improve the speed of parsing across a large number of files.

There is no strict ordering between the amount of input used by SLL vs LL, which makes it really hard to build a cache for full context. Let's say that we have input A B C that leads to an SLL conflict with full context X. That implies that using X we might only use A B but we could also use A B C D to resolve conflict. Input A B C D could predict alternative 1 in one position in the input and A B C E could predict alternative 2 in another position in input. The conflicting SLL configurations could still be non-unique in the full context prediction, which would lead us to requiring more input than the original A B C. To make a prediction cache work, we have to track the exact input used during the previous prediction. That amounts to a cache that maps X to a specific DFA for that context.

Something should be done for left-recursive expression predictions. They are likely LL(1) + pred eval. Easier to do the whole SLL unless error and retry with full LL thing Sam does.

AVOIDING FULL CONTEXT PREDICTION

We avoid doing full context retry when the outer context is empty, we did not dip into the outer context by falling off the end of the decision state rule, or when we force SLL mode.

As an example of the not dip into outer context case, consider as super constructor calls versus function calls. One grammar might look like this:

            ctorBody
            : '{' superCall? stat* '}'
            ;
            

Or, you might see something like

            stat
            : superCall ';'
            | expression ';'
            | ...
            ;
            

In both cases I believe that no closure operations will dip into the outer context. In the first case ctorBody in the worst case will stop at the '}'. In the 2nd case it should stop at the ';'. Both cases should stay within the entry rule and not dip into the outer context.

PREDICATES

Predicates are always evaluated if present in either SLL or LL both. SLL and LL simulation deals with predicates differently. SLL collects predicates as it performs closure operations like ANTLR v3 did. It delays predicate evaluation until it reaches and accept state. This allows us to cache the SLL ATN simulation whereas, if we had evaluated predicates on-the-fly during closure, the DFA state configuration sets would be different and we couldn't build up a suitable DFA.

When building a DFA accept state during ATN simulation, we evaluate any predicates and return the sole semantically valid alternative. If there is more than 1 alternative, we report an ambiguity. If there are 0 alternatives, we throw an exception. Alternatives without predicates act like they have true predicates. The simple way to think about it is to strip away all alternatives with false predicates and choose the minimum alternative that remains.

When we start in the DFA and reach an accept state that's predicated, we test those and return the minimum semantically viable alternative. If no alternatives are viable, we throw an exception.

During full LL ATN simulation, closure always evaluates predicates and on-the-fly. This is crucial to reducing the configuration set size during closure. It hits a landmine when parsing with the Java grammar, for example, without this on-the-fly evaluation.

SHARING DFA

All instances of the same parser share the same decision DFAs through a static field. Each instance gets its own ATN simulator but they share the same field. They also share a object that makes sure that all objects are shared among the DFA states. This makes a big size difference.

THREAD SAFETY

The locks on the field when it adds a new DFA object to that array. locks on the DFA for the current decision when setting the field. locks on the DFA for the current decision when looking up a DFA state to see if it already exists. We must make sure that all requests to add DFA states that are equivalent result in the same shared DFA object. This is because lots of threads will be trying to update the DFA at once. The method also locks inside the DFA lock but this time on the shared context cache when it rebuilds the configurations' objects using cached subgraphs/nodes. No other locking occurs, even during DFA simulation. This is safe as long as we can guarantee that all threads referencing s.edge[t] get the same physical target , or null . Once into the DFA, the DFA simulation does not reference the map. It follows the field to new targets. The DFA simulator will either find to be null , to be non- null and dfa.edges[t] null, or dfa.edges[t] to be non-null. The method could be racing to set the field but in either case the DFA simulator works; if null , and requests ATN simulation. It could also race trying to get dfa.edges[t] , but either way it will work because it's not doing a test and set operation.

Starting with SLL then failing to combined SLL/LL (Two-Stage Parsing)

Sam pointed out that if SLL does not give a syntax error, then there is no point in doing full LL, which is slower. We only have to try LL if we get a syntax error. For maximum speed, Sam starts the parser set to pure SLL mode with the :

            parser.
            getInterpreter()
            .
            
            (
            
            )
            ;
            parser.
            
            (new
            
            ());
            

If it does not get a syntax error, then we're done. If it does get a syntax error, we need to retry with the combined SLL/LL strategy.

The reason this works is as follows. If there are no SLL conflicts, then the grammar is SLL (at least for that input set). If there is an SLL conflict, the full LL analysis must yield a set of viable alternatives which is a subset of the alternatives reported by SLL. If the LL set is a singleton, then the grammar is LL but not SLL. If the LL set is the same size as the SLL set, the decision is SLL. If the LL set has size > 1, then that decision is truly ambiguous on the current input. If the LL set is smaller, then the SLL conflict resolution might choose an alternative that the full LL would rule out as a possibility based upon better context information. If that's the case, then the SLL parse will definitely get an error because the full LL analysis says it's not viable. If SLL conflict resolution chooses an alternative within the LL set, them both SLL and LL would choose the same alternative because they both choose the minimum of multiple conflicting alternatives.

Let's say we have a set of SLL conflicting alternatives 1, 2, 3}} and a smaller LL set called s. If s is 2, 3}}, then SLL parsing will get an error because SLL will pursue alternative 1. If s is 1, 2}} or 1, 3}} then both SLL and LL will choose the same alternative because alternative one is the minimum of either set. If s is 2}} or 3}} then SLL will get a syntax error. If s is 1}} then SLL will succeed.