This is reenforces the question that bugs me for some time. Is BitC too high-level for OS development?

BitC is C-incomplete. I.e. not every valid C program without assembly snippets can be rewritten as BitC program. And think that the statement "Like C [1 <http://www.coyotos.org/docs/bitc-spec/bitc-spec.html#ansi1999c>], BitC provides full control over data structure representation, which is necessary for high-performance systems programming" should be removed from the specification unless BitC changes. It does not sound true for me right now.

I originally expected it to be of level of C++, C, or Ada. However it is more close to C#, Java or ML now. At least with respect to memory managment. There are implicit memory allocation primitives, but no memory deallocation primitives. Spec never says "garbage collector" however it looks like it is assumed in some form.

I think that there is a choice for memory safety. It can be enforced or it can be proved. As I undertand, BitC has originally chosen path of enfocement. Maybe it is a good point to think about proving branch.

Constantine

BTW while we are on unions I would also like to point that one of addition in Ada2005 is unchecked union type (http://en.wikibooks.org/wiki/Programming:Ada:Types:record#Union). AFAIR they have discovered that it is really hard to interface to other language(s) without them. Possibly BitC will face the same problem later.

Jonathan S. Shapiro wrote:

The union safety problem is especially disturbing for a different
reason: it precludes linked lists (and similar structures) that are
threaded through their respective objects.

Consider a mutable union containing a doubly linked list element. The
union can be mutated as a whole. If the type tag of the union changes
this can create a safety problem because neighboring link elements no
longer point to a link element.

However, there is a higher-level problem here. Imagine that we re-assign
the union to a new value having the same tag. Note that all of the
linked list pointers are now validly typed, but they aren't in a linked
list here. There is no type violation, but there is an invariant
violation.

This raises two problems:

1. Can we avoid the need for such constructs? I'm not sure.
2. If not, do we now need to introduce a notion of construction,
destruction, and overloading of assignment? I would like to avoid this.

We are going to proceed by pushing hard on (1) to see what happens. We
currently see exactly one critical data structure in Coyotos for which
this presents a problem, and there is an alternative implementation.

shap

_______________________________________________
bitc-dev mailing list
[email protected]
http://www.coyotos.org/mailman/listinfo/bitc-dev



_______________________________________________ bitc-dev mailing list [email protected] http://www.coyotos.org/mailman/listinfo/bitc-dev

Reply via email to