public Class NFAFactory

From A B build A-e->B (that is, build an epsilon arc from right of A to left of B). As a convenience, return B if A is null or return A if B is null.

Build what amounts to an epsilon transition with an action. The action goes into NFA though it is ignored during analysis. It slows things down a bit, but I must ignore predicates after having seen an action (5-5-2008).

From A|B|..|Z alternative block build o->o-A->o->o (last NFAState is blockEndNFAState pointed to by all alts) | ^ o->o-B->o--| | | ... | | | o->o-Z->o--| So every alternative gets begin NFAState connected by epsilon and every alt right side points at a block end NFAState. There is a new NFAState in the NFAState in the StateCluster for each alt plus one for the end NFAState. Special case: only one alternative: don't make a block with alt begin/end. Special case: if just a list of tokens/chars/sets, then collapse to a single edge'd o-set->o graph. Set alt number (1..n) in the left-Transition NFAState.

From a set ('a'|'b') build o->o-'a'..'b'->o->o (last NFAState is blockEndNFAState pointed to by all alts)

From (A)? build either: o--A->o | ^ o---->| or, if A is a block, just add an empty alt to the end of the block

From (A)+ build |---| (Transition 2 from A.right points at alt 1) v | (follow of loop is Transition 1) o->o-A-o->o Meaning that the last NFAState in A points back to A's left Transition NFAState and we add a new begin/end NFAState. A can be single alternative or multiple. During analysis we'll call the follow link (transition 1) alt n+1 for an n-alt A block.

From (A)* build |---| v | o->o-A-o--o (Transition 2 from block end points at alt 1; follow is Transition 1) | ^ o---------| (optional branch is 2nd alt of optional block containing A+) Meaning that the last (end) NFAState in A points back to A's left side NFAState and we add 3 new NFAStates (the optional branch is built just like an optional subrule). See the Aplus() method for more on the loop back Transition. The new node on right edge is set to RIGHT_EDGE_OF_CLOSURE so we can detect nested (A*)* loops and insert an extra node. Previously, two blocks shared same EOB node. There are 2 or 3 decision points in a A*. If A is not a block (i.e., it only has one alt), then there are two decisions: the optional bypass and then loopback. If A is a block of alts, then there are three decisions: bypass, loopback, and A's decision point. Note that the optional bypass must be outside the loop as (A|B)* is not the same thing as (A|B|)+. This is an accurate NFA representation of the meaning of (A)*, but for generating code, I don't need a DFA for the optional branch by virtue of how I generate code. The exit-loopback-branch decision is sufficient to let me make an appropriate enter, exit, loop determination. See codegen.g

From label A build Graph o-A->o

From char 'c' build StateCluster o-intValue(c)->o

From char 'c' build StateCluster o-intValue(c)->o can include unicode spec likes '$' later. Accepts actual unicode 16-bit now, of course, by default. TODO not supplemental char clean!

add an EOF transition to any rule end NFAState that points to nothing (i.e., for all those rules not invoked by another rule). These are start symbols then. Return the number of grammar entry points; i.e., how many rules are not invoked by another rule (they can only be invoked from outside). These are the start rules.

From an empty alternative build StateCluster o-e->o

Can only complement block of simple alts; can complement build_Set() result, that is. Get set and complement, replace old with complement. public StateCluster build_AlternativeBlockComplement(StateCluster blk) { State s0 = blk.left; IntSet set = getCollapsedBlockAsSet(s0); if ( set!=null ) { // if set is available, then structure known and blk is a set set = nfa.grammar.complement(set); Label label = s0.transition(0).target.transition(0).label; label.setSet(set); } return blk; }

For reference to rule r, build o-e->(r) o where (r) is the start of rule r and the trailing o is not linked to from rule ref state directly (it's done thru the transition(0) RuleClosureTransition. If the rule r is just a list of tokens, it's block will be just a set on an edge o->o->o-set->o->o->o, could inline it rather than doing the rule reference, but i'm not doing this yet as I'm not sure it would help much in the NFA→DFA construction. TODO add to codegen: collapse alt blks that are sets into single matchSet

Build what amounts to an epsilon transition with a semantic predicate action. The pred is a pointer into the AST of the SEMPRED token.

From set build single edge graph o->o-set->o. To conform to what an alt block looks like, must have extra state on left.

For a non-lexer, just build a simple token reference atom. For a lexer, a string is a sequence of char to match. That is, "fog" is treated as 'f' 'o' 'g' not as a single transition in the DFA. Machine== o-'f'->o-'o'->o-'g'->o and has n+1 states for n characters.

Build an atom with all possible values in its label

Build a subrule matching ^(. .*) (any tree or node). Let's use (^(. .+) | .) to be safe.

Optimize an alternative (list of grammar elements). Walk the chain of elements (which can be complicated loop blocks...) and throw away any epsilon transitions used to link up simple elements. This only removes 195 states from the java.g's NFA, but every little bit helps. Perhaps I can improve in the future.