Can someone shed some light on this? The trade-off may well be worth it, but I'm only aware of the negative trade-off, not the positive.
If you take a look at the source of klist.h, the preprocessor stuff is a bit hairy, but it basically leaves you with a "compile time function", KLIST_INIT, that takes a datatype and generates, via simple textual substitution, an implementation of a linked list where each node of the list carries a payload of that datatype. klib_test.c includes a simple usage case.
Just to be clear, the preprocessor is generating real functions (inline ones, but functions nonetheless), not simply textually inserting the appropriate list code every time you use one of the macros like kl_next(iter) defined at the end of klist.h.
Check out the GCC docs on macros (and particularly, the section on concatenation) to get a better understanding of how klib implements this. http://gcc.gnu.org/onlinedocs/cpp/Macros.html
As for how "crazy" it is to do this with the preprocessor, it's just the most convenient alternative in C to manually copying and pasting a generic list implementation yourself and editing in the correct type information, so it's a fairly common practice. You'll know code that really abuses the preprocessor when you see it -- I would advise then to run.
I saw some examples in the klib code where macros are used to generate boilerplate code, generating functions.
http://github.com/dmw/caffeine http://software.schmorp.de/pkg/libev.html