We use stable Rust only.
Please check the minimum version of Rust we use in Cargo.toml
.
If you are proposing a major change to CI or a new iterator adaptor for this crate, then please first file an issue describing your proposal. Usual concerns about new methods.
To pass CI tests successfully, your code must be free of "compiler warnings" and "clippy warnings" and be "rustfmt" formatted.
Note that small PRs are easier to review and therefore are more easily merged.
In general, the code logic should be tested with quickcheck tests in tests/quick.rs
which allow us to test properties about the code with randomly generated inputs.
If it needs the "std" (such as using hashes) then it should be behind the use_std
feature,
or if it requires heap allocation (such as using vectors) then it should be behind the use_alloc
feature.
Otherwise it should be able to run in no_std
context.
This mostly applies to your new module, each import from it, and to your new Itertools
method.
self
, &mut self
or &self
? From #710:
- Take by value when:
- It transfers ownership to another iterator type, such as
filter
,map
... - It consumes the iterator completely, such as
count
,last
,max
...
- It transfers ownership to another iterator type, such as
- Mutably borrow when it consumes only part of the iterator, such as
find
,all
,try_collect
... - Immutably borrow when there is no change, such as
size_hint
.
Iterators are lazy:
- structs of iterator adaptors should have
#[must_use = "iterator adaptors are lazy and do nothing unless consumed"]
; - structs of iterators should have
#[must_use = "iterators are lazy and do nothing unless consumed"]
.
Those behaviors are tested in tests/laziness.rs
.
It might be more performant to specialize some methods. However, each specialization should be thoroughly tested.
Correctly specializing methods can be difficult, and we do not require that you do it on your initial PR.
Most of the time, we want specializations of:
size_hint
: It mostly allows allocation optimizations. When always exact, it also enables to implementExactSizeIterator
. See our private modulesrc/size_hint.rs
for helpers.fold
might make iteration faster than callingnext
repeatedly.count
,last
,nth
as we might be able to avoid iterating on every item withnext
.
Additionally,
for_each
,reduce
,max/min[_by[_key]]
andpartition
all rely onfold
so you should specialize it instead.all
,any
,find
,find_map
,cmp
,partial_cmp
,eq
,ne
,lt
,le
,gt
,ge
andposition
all rely (by default) ontry_fold
which we can not specialize on stable rust, so you might want to wait it stabilizes or specialize each of them.DoubleEndedIterator::{nth_back, rfold, rfind}
: similar reasoning.
An adaptor might use the inner iterator specializations for its own specializations.
They are tested in tests/specializations.rs
and benchmarked in benches/specializations.rs
(build those benchmarks is slow so you might want to temporarily remove the ones you do not want to measure).
The Debug
implementation
All our iterators should implement Debug
.
When one of the field is not debuggable (such as functions), you must not derive Debug
.
Instead, manually implement it and ignore this field in our helper macro debug_fmt_fields
.
4 examples (click to expand)
use std::fmt;
/* ===== Simple derive. ===== */
#[derive(Debug)]
struct Name1<I> {
iter: I,
}
/* ===== With an unclonable field. ===== */
struct Name2<I, F> {
iter: I,
func: F,
}
// No `F: Debug` bound and the field `func` is ignored.
impl<I: fmt::Debug, F> fmt::Debug for Name2<I, F> {
// it defines the `fmt` function from a struct name and the fields you want to debug.
debug_fmt_fields!(Name2, iter);
}
/* ===== With an unclonable field, but another bound to add. ===== */
struct Name3<I: Iterator, F> {
iter: I,
item: Option<I::Item>,
func: F,
}
// Same about `F` and `func`, similar about `I` but we must add the `I::Item: Debug` bound.
impl<I: Iterator + fmt::Debug, F> fmt::Debug for Name3<I, F>
where
I::Item: fmt::Debug,
{
debug_fmt_fields!(Name3, iter, item);
}
/* ===== With an unclonable field for which we can provide some information. ===== */
struct Name4<I, F> {
iter: I,
func: Option<F>,
}
// If ignore a field is not good enough, implement Debug fully manually.
impl<I: fmt::Debug, F> fmt::Debug for Name4<I, F> {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
let func = if self.func.is_some() { "Some(_)" } else { "None" };
f.debug_struct("Name4")
.field("iter", &self.iter)
.field("func", &func)
.finish()
}
}
When/How to implement Clone
All our iterators should implement Clone
when possible.
Note that a mutable reference is never clonable so struct Name<'a, I: 'a> { iter: &'a mut I }
can not implement Clone
.
Derive Clone
on a generic struct adds the bound Clone
on each generic parameter.
It might be an issue in which case you should manually implement it with our helper macro clone_fields
(it defines the clone
function calling clone
on each field) and be careful about the bounds.
When to implement std::iter::FusedIterator
This trait should be implemented by all iterators that always return None
after returning None
once, because it allows to optimize Iterator::fuse()
.
The conditions on which it should be implemented are usually the ones from the Iterator
implementation, eventually refined to ensure it behaves in a fused way.
When to implement ExactSizeIterator
When we are always able to return an exact non-overflowing length.
Therefore, we do not implement it on adaptors that makes the iterator longer as the resulting length could overflow.
One should not override ExactSizeIterator::len
method but rely on an exact Iterator::size_hint
implementation, meaning it returns (length, Some(length))
(unless you could make len
more performant than the default).
The conditions on which it should be implemented are usually the ones from the Iterator
implementation, probably refined to ensure the size hint is exact.
When to implement DoubleEndedIterator
When the iterator structure allows to handle iterating on both fronts simultaneously. The iteration might stop in the middle when both fronts meet.
The conditions on which it should be implemented are usually the ones from the Iterator
implementation, probably refined to ensure we can iterate on both fronts simultaneously.
When to implement itertools::PeekingNext
TODO
This is currently tested in tests/test_std.rs
.
TODO
No guideline about using #[inline]
yet.
In the Rust standard library, it's quite common for fold
to be implemented in terms of try_fold
. But it's not something we do yet because we can not specialize try_fold
methods yet (it uses the unstable Try
).
From #781, the general rule to follow is something like this:
- If you need to completely consume an iterator:
- Use
fold
if you need an owned access to an accumulator. - Use
for_each
otherwise.
- Use
- If you need to partly consume an iterator, the same applies with
try_
versions:- Use
try_fold
if you need an owned access to an accumulator. - Use
try_for_each
otherwise.
- Use