See for example this quick investigation and look at the number of branches for what's not a very big program https://bnikolic.co.uk/blog/hpc-perf-branchprediction
Only branches that capture exceptional events, e.g. error checks, show 99% hit rate.
Branches that are part of an algorithm are driven by the data you feed to them. For example in a classical binary search, the prediction is right only 50% of the time.
If we have an instruction that will conditionally load one of two pointers from memory, the left or right traversal can be made branch-free.
The only reason we need to switch between two code paths in the binary tree descent is that we have two different pieces of code for loading the left or right pointer to the different offsets used in the load instruction.
If we use a two element array for left and right, the indices [0] and [1] select between them. We make the comparison condition calculate a 0 or 1 result and use it as an index.
Because newlines are relatively common and unpredictable a state machine is likely better. But on a long file with no new lines and no spaces the branching one should be slightly faster. (All reasoning from first principles; I have not done any benchmarking!)
wc2 and wc might be compiled with different flags / use a different strategy for IO, which makes it hard to compare speeds directly. The theoretical speedup of branches is tiny compared potential speedup of changing your compiler flags!
I think the fourth transition represents illegal Unicode. From there it stays in the illegal state until it hits legal UTF-8, then goes back to counting.
To handle illegal chars, it would need a 4th class but also a 5th state, so that's not the reason.
Can it be to replace a 'modulo' operation with an 'and' in the access to table?
By working back from the fact that state 2 is the word count, it must be entered when a word character (class 0) is hit after some space/newline characters. Hence state 0 represents the machine being in the middle of a sequence of whitespace, state 1 is when the machine is in a sequences of newlines, state 2 is the first letter of a word, and state 3 is any subsequent letter of a word. You can then verify that the machine transitions to the correct state from those by checking what it does in case of a word, space, or newline character. But I'm not sure why the table is 4×4. The fourth transition from every state is unreachable since the symbols are all in [0, 2].
As for why the approach is better, I think it's because it avoids branching. If you think about the length of a word, it's quite short compared to a CPU pipeline, so on a branching version of this you're going to be stuck spending most of your execution time on branch miss-prediction.