S -> E E -> E + T | T T -> n
n+nn+nup to the point at which the algorithm determines that the input it invalid.
A rule with the • in it is called an "LR item". Think of them as a goal: E→•T means "I'm trying to read E by reading T", E→E+•T means "I'm trying to read E by reading E+T and I've already read E+". Each transisition represents moving to a subgoal that helps achive the previous goal in a way consistent with the grammar.
This machine we build this way is the non-deterministic "Characteristic Finite State Machine" (CFSM). We can convert to an equivalent deterministic machine, and that is what is usually meant by CFSM.
This procedure is called "Simple LR" parsing, or SLR. You should notice that in our example we can parse any string without any lookahead, therefore this grammar is SLR(0), i.e. SLR parsing with zero lookahead tokens suffices.
S → Ewhich produces a deterministic CSFM that includes the following fragment:
E → E + T | T
T → T * F | F
F → n | (E)
Now, if we end up in the left state, we have some ambiguity: LR item T→T•*F tells us to shift (push a token), while LR item E→T• tells us to reduce T to E. This is a "shift/reduce conflict". With lookahead, however, we can resolve this. If the lookahead symbol is a *, we know we must reduce. Why? Because * can't follow an "E" in this grammar, so we really can't do the reduce and succeed. In contrast, * is exactly what LR item T→T•*F wants to read next. So we shift.
In other scenarios we might have two different LR items in the same state that tell us to reduce but, of course, using different reductions. This is a "reduce/reduce conflict". Sometimes lookahead can disambiguate this kind of conflict, but not in SLR parsers.