On Wed, Jan 04, 2006 at 09:48:05AM -0000, Simon Peyton-Jones wrote:

Ah, now I see what you mean. I had not thought of doing that, but it'd make sense, esp if there were lots of 4-byte-non-pointer types. But there are only a couple (Int# and Word32#) so I'm not sure it's worth the effort. But (without thinking about it too hard) it looks like a very sensible idea. The trickiness is in the fact that some type constructors (notably (->)) are somewhat kind-polymorphic, which gives rise to some awkwardness in the type checker. So the details might not be entirely straightforward.

perhaps some sort of post-kind-inference can be done. so the only _user visible_ unboxed kind would be # (not counting unboxed tuples) and then after everything is typechecked in terms of that, a separate inference pass figures out the sizes #4, #8, etc. then a specialized version of each polymorphic routine is created for each kind it is called at. This is opposed to full expansion where polymoprhic routines are specialized at each type rather than each kind. This would solve the exponential code growth problem since there are only a finite number of kinds (up to isomorphism if user defined kinds are allowed) in any implementation. One can do separate compilation by ouputing the functions with the special tag saying the linker should combine identical names, so the same function might get outputted a few times, but only a single copy will end up in the final executable. alternatively, a version for each kind can be generated up front and then the linker can clean up the ones that arn't used. John -- John Meacham - ⑆repetae.net⑆john⑈