Back to PELs (Ford, 2004) with a small comparison with Church Rosser languages (McNaughton et al., 1988) this time. Both enjoy a linear time membership problem, include some context-sensitive languages and the full class of DCFLs, and are closed under complement.
I had wondered for some time whether all PELs were Church Rosser, but I just found out that the following Gladkij language was not growing context-sensitive and thus not Church Rosser (Buntrock and Otto, 1998):
We can however recognize ℒGladkij with a PEG since it is the intersection of two deterministic context-free languages
ℒ2 = { w'cwcwR | w,w' ∈ {a,b}* }.
Since PELs are closed under intersection and include all the DCFLs, it follows that ℒGladkij is a parsing expression language. The following PEG recognizes ℒGladkij (I'm not too fond of predicates, so I prefer to stick with the ordered choice operator):
W' ← a W' / b W' / ε
W ← a W a / b W b / c
T ← a / b / c
Z ← T Z / T
The expression W' c W recognizes ℒ2, and the expression W c W' recognizes ℒ1. If an input string is in ℒ2, then the expression (W' c W / ε) Z fails, thus the next alternative ε is used and the input string also has to be in ℒ1 in order to be accepted by the grammar. If an input string is not in ℒ2, then (W' c W / ε) Z succeeds and consumes the whole input through Z, therefore the grammar cannot accept the input.
I have currently no clue whether an inclusion of the Church Rosser languages in the class of parsing expression languages would hold.