Now that the necessary associated types exist for the `IntoIterator` trait this
commit stabilizes the trait as-is as well as all existing implementations.
`IntoIterator` now has an extra associated item:
``` rust
trait IntoIterator {
type Item;
type IntoIter: Iterator<Self=Self::Item>;
}
```
This lets you bind the iterator \"`Item`\" directly when writing generic functions:
``` rust
// hypothetical change, not included in this PR
impl Extend<T> for Vec<T> {
// you can now write
fn extend<I>(&mut self, it: I) where I: IntoIterator<Item=T> { .. }
// instead of
fn extend<I: IntoIterator>(&mut self, it: I) where I::IntoIter: Iterator<Item=T> { .. }
}
```
The downside is that now you have to write an extra associated type in your `IntoIterator` implementations:
``` diff
impl<T> IntoIterator for Vec<T> {
+ type Item = T;
type IntoIter = IntoIter<T>;
fn into_iter(self) -> IntoIter<T> { .. }
}
```
Because this breaks all downstream implementations of `IntoIterator`, this is a [breaking-change]
---
r? @aturon
This PR is an optimization of the `FromIterator` implementation of `Vec`
Benchmark: https://gist.github.com/alexcrichton/03d666159a28a80e7c70
Before:
test macro_repeat1 ... bench: 57 ns/iter (+/- 1)
test macro_repeat2 ... bench: 56 ns/iter (+/- 1)
test map_clone1 ... bench: 828 ns/iter (+/- 13)
test map_clone2 ... bench: 828 ns/iter (+/- 8)
test repeat1 ... bench: 1104 ns/iter (+/- 10)
test repeat2 ... bench: 1106 ns/iter (+/- 11)
After:
test macro_repeat1 ... bench: 75 ns/iter (+/- 21)
test macro_repeat2 ... bench: 59 ns/iter (+/- 31)
test map_clone1 ... bench: 34 ns/iter (+/- 22)
test map_clone2 ... bench: 52 ns/iter (+/- 21)
test repeat1 ... bench: 34 ns/iter (+/- 11)
test repeat2 ... bench: 33 ns/iter (+/- 12)
The idea behind this optimization is to avoid all bounds checks for space
already allocated into the vector. This may involve running the iterator twice,
but the first run of the iterator should be optimizable to a memcpy or memset if
possible.
The same treatment can in theory be applied to `Vec::extend` but the benchmarks
for that currently get *worse* if the change is applied. This appears to be some
LLVM optimizations going awry but it's seems prudent to land at least the
`collect` portion beforehand.
This PR is an optimization of the `FromIterator` implementation of `Vec`
Benchmark: https://gist.github.com/alexcrichton/03d666159a28a80e7c70
Before:
test macro_repeat1 ... bench: 57 ns/iter (+/- 1)
test macro_repeat2 ... bench: 56 ns/iter (+/- 1)
test map_clone1 ... bench: 828 ns/iter (+/- 13)
test map_clone2 ... bench: 828 ns/iter (+/- 8)
test repeat1 ... bench: 1104 ns/iter (+/- 10)
test repeat2 ... bench: 1106 ns/iter (+/- 11)
After:
test macro_repeat1 ... bench: 75 ns/iter (+/- 21)
test macro_repeat2 ... bench: 59 ns/iter (+/- 31)
test map_clone1 ... bench: 34 ns/iter (+/- 22)
test map_clone2 ... bench: 52 ns/iter (+/- 21)
test repeat1 ... bench: 34 ns/iter (+/- 11)
test repeat2 ... bench: 33 ns/iter (+/- 12)
The idea behind this optimization is to avoid all bounds checks for space
already allocated into the vector. This may involve running the iterator twice,
but the first run of the iterator should be optimizable to a memcpy or memset if
possible.
The same treatment can in theory be applied to `Vec::extend` but the benchmarks
for that currently get *worse* if the change is applied. This appears to be some
LLVM optimizations going awry but it's seems prudent to land at least the
`collect` portion beforehand.
Port `core::ptr::Unique` to have `PhantomData`. Add `PhantomData` to
`TypedArena` and `Vec` as well.
As a drive-by, switch `ptr::Unique` from a tuple-struct to a struct
with fields.
This is in preparation for stabilization of the `IntoIterator` trait. All
implementations and references to `Iter` need to be renamed to `IntoIter`.
[breaking-change]
This commit performs a final stabilization pass over the std::fmt module,
marking all necessary APIs as stable. One of the more interesting aspects of
this module is that it exposes a good deal of its runtime representation to the
outside world in order for `format_args!` to be able to construct the format
strings. Instead of hacking the compiler to assume that these items are stable,
this commit instead lays out a story for the stabilization and evolution of
these APIs.
There are three primary details used by the `format_args!` macro:
1. `Arguments` - an opaque package of a "compiled format string". This structure
is passed around and the `write` function is the source of truth for
transforming a compiled format string into a string at runtime. This must be
able to be constructed in stable code.
2. `Argument` - an opaque structure representing an argument to a format string.
This is *almost* a trait object as it's just a pointer/function pair, but due
to the function originating from one of many traits, it's not actually a
trait object. Like `Arguments`, this must be constructed from stable code.
3. `fmt::rt` - this module contains the runtime type definitions primarily for
the `rt::Argument` structure. Whenever an argument is formatted with
nonstandard flags, a corresponding `rt::Argument` is generated describing how
the argument is being formatted. This can be used to construct an
`Arguments`.
The primary interface to `std::fmt` is the `Arguments` structure, and as such
this type name is stabilize as-is today. It is expected for libraries to pass
around an `Arguments` structure to represent a pending formatted computation.
The remaining portions are largely "cruft" which would rather not be stabilized,
but due to the stability checks they must be. As a result, almost all pieces
have been renamed to represent that they are "version 1" of the formatting
representation. The theory is that at a later date if we change the
representation of these types we can add new definitions called "version 2" and
corresponding constructors for `Arguments`.
One of the other remaining large questions about the fmt module were how the
pending I/O reform would affect the signatures of methods in the module. Due to
[RFC 526][rfc], however, the writers of fmt are now incompatible with the
writers of io, so this question has largely been solved. As a result the
interfaces are largely stabilized as-is today.
[rfc]: https://github.com/rust-lang/rfcs/blob/master/text/0526-fmt-text-writer.md
Specifically, the following changes were made:
* The contents of `fmt::rt` were all moved under `fmt::rt::v1`
* `fmt::rt` is stable
* `fmt::rt::v1` is stable
* `Error` is stable
* `Writer` is stable
* `Writer::write_str` is stable
* `Writer::write_fmt` is stable
* `Formatter` is stable
* `Argument` has been renamed to `ArgumentV1` and is stable
* `ArgumentV1::new` is stable
* `ArgumentV1::from_uint` is stable
* `Arguments::new_v1` is stable (renamed from `new`)
* `Arguments::new_v1_formatted` is stable (renamed from `with_placeholders`)
* All formatting traits are now stable, as well as the `fmt` method.
* `fmt::write` is stable
* `fmt::format` is stable
* `Formatter::pad_integral` is stable
* `Formatter::pad` is stable
* `Formatter::write_str` is stable
* `Formatter::write_fmt` is stable
* Some assorted top level items which were only used by `format_args!` were
removed in favor of static functions on `ArgumentV1` as well.
* The formatting-flag-accessing methods remain unstable
Within the contents of the `fmt::rt::v1` module, the following actions were
taken:
* Reexports of all enum variants were removed
* All prefixes on enum variants were removed
* A few miscellaneous enum variants were renamed
* Otherwise all structs, fields, and variants were marked stable.
In addition to these actions in the `std::fmt` module, many implementations of
`Show` and `String` were stabilized as well.
In some other modules:
* `ToString` is now stable
* `ToString::to_string` is now stable
* `Vec` no longer implements `fmt::Writer` (this has moved to `String`)
This is a breaking change due to all of the changes to the `fmt::rt` module, but
this likely will not have much impact on existing programs.
Closes#20661
[breaking-change]
Don't reallocate when capacity is already equal to length
`Vec::shrink_to_fit()` may be called on vectors that are already the
correct length. Calling out to `reallocate()` in this case is a bad idea
because there is no guarantee that `reallocate()` won't allocate a new
buffer anyway, and based on performance seen in external benchmarks, it
seems likely that it is in fact reallocating a new buffer.
Before:
test string::tests::bench_exact_size_shrink_to_fit ... bench: 45 ns/iter (+/- 2)
After:
test string::tests::bench_exact_size_shrink_to_fit ... bench: 26 ns/iter (+/- 1)
