On Mon, Dec 13, 2010 at 09:58:27AM -0500, Yitzhak Mandelbaum wrote:
> One small question: could you expand on your last comment:
> 
> On Dec 13, 2010, at 8:06 AM, Jerome Vouillon wrote:
> 
> > <snip>
> > 
> > Ocamljs optimizes tail recursion, but this comes at a large
> > performance cost.
> 
> Do you mean all tail-calls come a large cost, or only those outside
> of plain tail-recursion?  Either way, could you give us some more
> intuition as to why this happens, and why js_of_ocaml doesn't suffer
> from the same problem (assuming it applies to tail-recursion)?

I meant tail calls, indeed.  Tail recursion (when a function calls
itself recursively in tail position) can be optimized efficiently by
wrapping the function body inside a loop.  This is what Js_of_ocaml
does.

For optimizing tail calls in general, you need to use trampolines.
Instead of calling a function in tail position, you return this
function and its arguments.  Then, a piece of code called a trampoline
takes care of invoking repeately the functions it receives this way
until a final result is returned.  This is expansive.  You have to
make sure that this piece of code is there whenever a function is
invoked in tail position.  Also, the JIT compilers cannot optimize the
trampoline, as the functions it will have to call, and their number of
arguments, are unknown.

-- Jerome