On 11/19/13 5:12 PM, Val Markovic wrote:
I felt cheated a bit TBH. It's such an obvious design mistake IMO;
everyone's going to get this wrong because it's counter-intuitive.
I agree that we should change this, but give us some credit: it's only
an obvious design mistake now that we've gotten to this point. For those
joining us now, we have gone through the following changes starting 3
years ago:
1. `[T]` is always reference counted. Pushing onto a vector was
optimized to push in-place iff the reference count was 1. There is `[mut
T]` as well. There is no overloading and also no slicing, because there
was no region system.
2. `[T]` is always owned and small vector optimized with 4 elements
in place (number arrived at through measurements). There is no `[mut
T]`. There is no slicing because there is no region system.
3. `[T]` is always owned and not small vector optimized. There is
no slicing.
4. We have 4 vector types, `[T, ..N]`, `[T]/&` and `[T]/~` (today
`[T, ..N]`, `&[T]`, `~[T]` respectively), as well as `[T]/@` which is
reference counted. Vector literals infer to the right one. Owned and
managed vectors can be sliced.
5. We have 4 vector types, `[T, ..N]`, `&[T]`, `~[T]`, and `@[T]`.
6. We have 3 vector types, `[T, ..N]`, `&[T]` and `~[T]`.
Now we're talking about:
7. We have 1 built-in vector type: `[T, ..N]`. Slices are
`Slice<T>` lang items. Owned and managed vectors can be built into the
library.
The current status is the product of the slow evolution of a design that
predated the region and slicing system, and what seems obvious now was
not obvious in retrospect.
Patrick
_______________________________________________
Rust-dev mailing list
[email protected]
https://mail.mozilla.org/listinfo/rust-dev
