Look at the disassemble of it. I think the first example is the surprising bit. SBCL does TCO, and the the tail-calling version should always be faster. Maybe S

Look at the disassemble of it. I think the first example is the surprising bit. SBCL does TCO, and the the tail-calling version should always be faster. Maybe SBCL is inlining and unrolling the first version for small numbers.