Returning multiple values puts multiple non-GC'd values on the stack. This is fast, but capturing and using those values is somewhat inconvenient for the programmer. Returning a cons creates one other cons in the heap that will need to be GC'd for each additional value returned.
In theory we could do this with arc2c. However, I'm not 100% sure this is necessary with a "really good" optimizing compiler.
We could defer destructuring of arguments in arc2c to as late as possible, so that we could do some amount of optimizing away 'cons cells when the cells themselves are used only for returning multiple values. Which is arguably difficult, since each stage in arc2c expects arguments to be in undestructured form, i.e. (let (foo bar) niaw ...) => (let tmp niaw (with (foo (car tmp) bar (cadr tmp)))). If arguments are kept in non-destructured form, we would need to modify the way that function parameters are stored (to handle (o foo (expression))) and check each stage in the compiler.
Edit: When optimizing raymyers' treeparse, I actually transformed parsers to CPS form in order to return multiple values without all the construction and deconstruction of 'cons cells, which helped reduce some of the overhead.
