[Sbcl-devel] Re: performance oddity

[Rob M. <ra...@ri...>]

> 
> Unless I'm grossly misunderstanding something, there's no inlining
> going on in your example regardless of which definition of FOO is
> used. The difference between the two cases is that one gets tail-call
> optimized, the other doesn't. 
> 
> Surprisingly the one that doesn't get optimized is the one where the
> compiler has access to more information. For some reason
> FLUSH-FULL-CALL-TAIL-TRANSFER explicitly forbids tail-calls to
> functions when the return convention isn't :UNKNOWN.
> 

This is because in designing the tail-call support in Python I was 
drawing on the Scheme implementation tradition where tail recursion is 
not considered an optional optimization.  All calls that can be tail 
recursive must be tail recursive.  If a function that is returning with 
the unknown convention calls a function with known returns then it must 
suppress tail recursion so that the can receive the values and massage 
them into the unknown convention.

As I recall, SBCL has taken the position that since ANSI says nothing 
about tail recursion then it is unimportant.

   Rob