> Allegro does, and apparently so does SBCL.

> As has been pointed out, it isn't necessary that formatter turn the
> format control string into "code", whatever that means, but only that a
> function be returned that is more efficient presumably by preprocessing
> the format control.

> In Allegro formatter returns a function, as required, but since the
> translation of a format control string into a function often produces
> a rather large function, the formatter function is a closure capturing
> an intermediate representation that achieves most of the speed
> optimization without incurring the huge code space overhead.

> The ANS doesn't point this out, but it should be obvious that calls to
> format with a constant format control string can have the work done by
> formatter performed by a compiler macro on format.  On the other hand,
> most programs contain lots of format control strings that are used only
> in rare conditions, so it is a little wasteful of space to transform
> unconditionally at compile time.  The behavior in Allegro under default
> optimization is to transform into a lazy conversion of the format
> control.  A format control string in a function is translated and
> cached by the function the first time it is used.

> These optimizations were implemented perhaps 16 years ago, when memory
> was measured in MB rather than GB.  But instruction caches are still
> measured in MB...