> Stephen  knows  this  stuff better than I, but I know that we talked a lot to
> Richard about his experiences with full CPS conversion and I  know  that  the
> notion  of  using  CPS to express the intra-procedural flow was from him.  At
> the time I was arguing for the advantages of CPS (over direct  style)  and  I
> remember  showing  an example from Shiver's thesis where top-level procedures
> were transformed into continuations (because they always were called with the
> same  continuation).   I  think  that Stephen put contification into MLton in
> response to that.  (Although it might have waited until he saw that this  was
> going to be required for nested loops, as in matrix multiplication.)

That all sounds right.  I remember the most annoying problem that we had with
the direct style IL was that we were allocating tuples for code like the
following:

val (x, y) = if ... then (a, b) else (c, d)

This was absolutely killing us in some benchmarks.  We could have gone to some
variant of a direct style IL that allowed a "flat" tuple to be returned in the
branches of the if, but it was so much easier to go to CPS with multi-argument
continuations and do:

fun L (x, y) = ...
if ... then L (a, b) else L (c, d)