On Wed 31 Aug 2016 11:03, Andy Wingo <wi...@pobox.com> writes: > On Thu 18 Aug 2016 18:14, Mark H Weaver <m...@netris.org> writes: > >> As I wrote above, the current guile compiler can already do this kind of >> type inference, although it does not currently do this for boxes. >> we can already anticipate having native code generation in the >> next couple of years, and we must keep boxes semantically simple so that >> our future compiler will be able to generate good code for this very >> important fundamental type. > > For what it's worth, I don't see the optimization argument as weighing > very heavily on this discussion. I would rather have fewer fundamental > data types rather than more, in the next two years or so. I see the > mid-term result here being that SRFI-111 boxes are much slower than > variables. > > The highest performance compilation tier we can imagine would include > adaptive optimization, and when it runs you can know that the variables > that a bit of code uses are bound or not. Also in that case we can > reasonably make any call to variable-unset! deoptimize any code that > uses variables, forcing it to reoptimize later. Since variable-unset! > is quite rare this is no big deal I think.
Following up here :) So again I think the optimization argument is not so important; if that were the only consideration then IMO the balance of things would be that we should apply Glenn's patch. There is a semantic consideration as well -- box-ref on a box created by make-box should never throw an exception, and code that uses the SRFI-111 should be able to rely on this. We should probably not introduce a gratuitous incompatibility here. I propose to add a flag to variables indicating that certain variables may not be unset. We can also consider reversing this, in that only variables with the flag can be unset; my understanding is that the only user of variable-unset! is the Elisp language on variables that it creates, so that would be acceptable too. Andy
