org.antlr.analysis

Class NFAContext

java.lang.Object

org.antlr.analysis.NFAContext

public class NFAContext
extends Object

A tree node for tracking the call chains for NFAs that invoke other NFAs. These trees only have to point upwards to their parents so we can walk back up the tree (i.e., pop stuff off the stack). We never walk from stack down down through the children. Each alt predicted in a decision has its own context tree, representing all possible return nodes. The initial stack has EOF ("$") in it. So, for m alternative productions, the lookahead DFA will have m NFAContext trees. To "push" a new context, just do "new NFAContext(context-parent, state)" which will add itself to the parent. The root is NFAContext(null, null). The complete context for an NFA configuration is the set of invoking states on the path from this node thru the parent pointers to the root.

Field Summary

Fields

Modifier and Type	Field and Description
protected int	cachedHashCode Computing the hashCode is very expensive and closureBusy() uses it to track when it's seen a state|ctx before to avoid infinite loops.
NFAState	invokingState The NFA state that invoked another rule's start state is recorded on the rule invocation context stack.
static int	MAX_SAME_RULE_INVOCATIONS_PER_NFA_CONFIG_STACK This is similar to Bermudez's m constant in his LAR(m) where you bound the stack so your states don't explode.
NFAContext	parent

Constructor Summary

Constructors

Constructor and Description
NFAContext(NFAContext parent, NFAState invokingState)

Method Summary

Methods

Modifier and Type	Method and Description
boolean	conflictsWith(NFAContext other) Two contexts conflict() if they are equals() or one is a stack suffix of the other.
boolean	equals(Object o) Two contexts are equals() if both have same call stack; walk upwards to the root.
int	hashCode()
boolean	isEmpty() A context is empty if there is no parent; meaning nobody pushed anything on the call stack.
int	recursionDepthEmanatingFromState(int state) Given an NFA state number, how many times has the NFA-to-DFA conversion pushed that state on the stack? In other words, the NFA state must be a rule invocation state and this method tells you how many times you've been to this state.
protected boolean	suffix(NFAContext other) [$] suffix any context [21 $] suffix [21 12 $] [21 12 $] suffix [21 $] [21 18 $] suffix [21 18 12 9 $] [21 18 12 9 $] suffix [21 18 $] [21 12 $] not suffix [21 9 $] Example "[21 $] suffix [21 12 $]" means: rule r invoked current rule from state 21.
String	toString()

Methods inherited from class java.lang.Object

clone, finalize, getClass, notify, notifyAll, wait, wait, wait

Field Detail

MAX_SAME_RULE_INVOCATIONS_PER_NFA_CONFIG_STACK

public static int MAX_SAME_RULE_INVOCATIONS_PER_NFA_CONFIG_STACK

This is similar to Bermudez's m constant in his LAR(m) where you bound the stack so your states don't explode. The main difference is that I bound only recursion on the stack, not the simple stack size. This looser constraint will let the conversion roam further to find lookahead to resolve a decision. Bermudez's m operates differently as it is his LR stack depth I'm pretty sure it therefore includes all stack symbols. Here I restrict the size of an NFA configuration to be finite because a stack component may mention the same NFA invocation state at most m times. Hence, the number of DFA states will not grow forever. With recursive rules like e : '(' e ')' | INT ; you could chase your tail forever if somebody said "s : e '.' | e ';' ;" This constant prevents new states from being created after a stack gets "too big". Actually (12/14/2007) I realize that this example is trapped by the non-LL(*) detector for recursion in > 1 alt. Here is an example that trips stack overflow: s : a Y | A A A A A X ; // force recursion past m=4 a : A a | Q; If that were: s : a Y | A+ X ; it could loop forever. Imagine doing a depth-first search on the e DFA...as you chase an input sequence you can recurse to same rule such as e above. You'd have a chain of ((((. When you get do some point, you have to give up. The states in the chain will have longer and longer NFA config stacks. Must limit size. max=0 implies you cannot ever jump to another rule during closure. max=1 implies you can make as many calls as you want--you just can't ever visit a state that is on your rule invocation stack. I.e., you cannot ever recurse. max=2 implies you are able to recurse once (i.e., call a rule twice from the same place). This tracks recursion to a rule specific to an invocation site! It does not detect multiple calls to a rule from different rule invocation states. We are guaranteed to terminate because the stack can only grow as big as the number of NFA states * max. I noticed that the Java grammar didn't work with max=1, but did with max=4. Let's set to 4. Recursion is sometimes needed to resolve some fixed lookahead decisions.

parent

public NFAContext parent

invokingState

public NFAState invokingState

The NFA state that invoked another rule's start state is recorded on the rule invocation context stack.

cachedHashCode

protected int cachedHashCode

Computing the hashCode is very expensive and closureBusy() uses it to track when it's seen a state|ctx before to avoid infinite loops. As we add new contexts, record the hash code as this.invokingState + parent.cachedHashCode. Avoids walking up the tree for every hashCode(). Note that this caching works because a context is a monotonically growing tree of context nodes and nothing on the stack is ever modified...ctx just grows or shrinks.

Constructor Detail

NFAContext

public NFAContext(NFAContext parent,
          NFAState invokingState)

Method Detail

equals

public boolean equals(Object o)

Two contexts are equals() if both have same call stack; walk upwards to the root. Recall that the root sentinel node has no invokingStates and no parent. Note that you may be comparing contexts in different alt trees. The hashCode is now cheap as it's computed once upon each context push on the stack. Use it to make equals() more efficient.

Overrides:

equals in class Object

conflictsWith

public boolean conflictsWith(NFAContext other)

Two contexts conflict() if they are equals() or one is a stack suffix of the other. For example, contexts [21 12 $] and [21 9 $] do not conflict, but [21 $] and [21 12 $] do conflict. Note that I should probably not show the $ in this case. There is a dummy node for each stack that just means empty; $ is a marker that's all. This is used in relation to checking conflicts associated with a single NFA state's configurations within a single DFA state. If there are configurations s and t within a DFA state such that s.state=t.state && s.alt != t.alt && s.ctx conflicts t.ctx then the DFA state predicts more than a single alt--it's nondeterministic. Two contexts conflict if they are the same or if one is a suffix of the other. When comparing contexts, if one context has a stack and the other does not then they should be considered the same context. The only way for an NFA state p to have an empty context and a nonempty context is the case when closure falls off end of rule without a call stack and re-enters the rule with a context. This resolves the issue I discussed with Sriram Srinivasan Feb 28, 2005 about not terminating fast enough upon nondeterminism.

suffix

protected boolean suffix(NFAContext other)

[$] suffix any context [21 $] suffix [21 12 $] [21 12 $] suffix [21 $] [21 18 $] suffix [21 18 12 9 $] [21 18 12 9 $] suffix [21 18 $] [21 12 $] not suffix [21 9 $] Example "[21 $] suffix [21 12 $]" means: rule r invoked current rule from state 21. Rule s invoked rule r from state 12 which then invoked current rule also via state 21. While the context prior to state 21 is different, the fact that both contexts emanate from state 21 implies that they are now going to track perfectly together. Once they converged on state 21, there is no way they can separate. In other words, the prior stack state is not consulted when computing where to go in the closure operation. ?$ and ??$ are considered the same stack. If ? is popped off then $ and ?$ remain; they are now an empty and nonempty context comparison. So, if one stack is a suffix of another, then it will still degenerate to the simple empty stack comparison case.

recursionDepthEmanatingFromState

public int recursionDepthEmanatingFromState(int state)

Given an NFA state number, how many times has the NFA-to-DFA conversion pushed that state on the stack? In other words, the NFA state must be a rule invocation state and this method tells you how many times you've been to this state. If none, then you have not called the target rule from this state before (though another NFA state could have called that target rule). If n=1, then you've been to this state before during this DFA construction and are going to invoke that rule again. Note that many NFA states can invoke rule r, but we ignore recursion unless you hit the same rule invocation state again.

hashCode

public int hashCode()

Overrides:

hashCode in class Object

isEmpty

public boolean isEmpty()

A context is empty if there is no parent; meaning nobody pushed anything on the call stack.

toString

public String toString()

Overrides:

toString in class Object