[User23]

The customary way to improve GC performance in Common Lisp is to avoid consing. Which, incidentally, is the customary way to write high performance C code too. Allocate your memory in a piece up front and arrange the data in a way that's friendly to your architecture and then do stuff. Nice thing about Common Lisp is that DISASSEMBLE makes it a fairly smooth process to check that the optimizer is actually doing what you want.

And the new SBCL release aids this approach by expanding the cases where the compiler will stack allocate.