On Wed, Mar 11, 2009 at 12:12 PM, Jonathan S. Shapiro <[email protected]> wrote: > Geoffrey is right about this issue. The reason I don't want to tackle > this right now is because I'ld rather get a lot of stuff right (and > not do this) than not get anything at all, and this requires a degree > of thought that we shouldn't stop to do right now.
Yep. That's why I marked it as a question, not a feature request. > That said, I do think that this is a good idea. > > <snip> > > There is an enormous safety problem here. If the vector body can be in > scope without being fully initialized, we would be force to introduce > a dependent type system that could accomodate disjoint subranges (some > initialized, others not). I'm not aware that *anybody* has managed to > do that. It's definitely a hairball, especially so since BitC does not > (and will not) support multi-dimensional arrays in the type system. > The reason this matters is that it makes describing the initialized > sub-slices moderately complex. I don't think this is necessary to get almost all of the benefit. The cases where this matters usually involve POD data, so you can either pre-initialize the data to zero or allocate it uninitialized. Uninitialized allocation is perfectly safe as long as you don't mind errors resulting in nondeterministic (but still safe) behavior. In any case, your kludgy intuition is in accord with my even kludgier intuition, which is enough to leave me hopeful for the future. Thanks! Geoffrey _______________________________________________ bitc-dev mailing list [email protected] http://www.coyotos.org/mailman/listinfo/bitc-dev
