Basically CCL's compiled code is often as fast as code from SBCL, LispWorks, Allegro CL. But one has to declare much more types of variables and functions explicitly with CCL. The code needs to be littered with type declarations.
A compiler with a (better) type inferencer can propagate the type information it already has and this is often enough for fast code. A good compiler can then also tell the developer where it lacks type information and then you can decide to declare types explicitly. SBCL/CMUCL does that in a very noisy way with lots of information generated by the compiler. ;-)
Performance. If you have (+ x y) and you can infer that x and y are fixints, then you can use a couple inline opcodes, rather than calling GENERIC-+.
EDIT: lispm is right, too, of course. You can look at it either way. If your program required top performance and you were going to tag every type until it ran super fast, then inference buys you the convenience of not needing to do as much. If you weren't going to bother with that, then inference buys you extra performance without all the effort.
Basically CCL's compiled code is often as fast as code from SBCL, LispWorks, Allegro CL. But one has to declare much more types of variables and functions explicitly with CCL. The code needs to be littered with type declarations.
A compiler with a (better) type inferencer can propagate the type information it already has and this is often enough for fast code. A good compiler can then also tell the developer where it lacks type information and then you can decide to declare types explicitly. SBCL/CMUCL does that in a very noisy way with lots of information generated by the compiler. ;-)
See for example: http://home.pipeline.com/~hbaker1/TInference.html