Improve the `array::map` codegen
The `map` method on arrays [is documented as sometimes performing poorly](https://doc.rust-lang.org/std/primitive.array.html#note-on-performance-and-stack-usage), and after [a question on URLO](https://users.rust-lang.org/t/try-trait-residual-o-trait-and-try-collect-into-array/88510?u=scottmcm) prompted me to take another look at the core [`try_collect_into_array`](7c46fb2111/library/core/src/array/mod.rs (L865-L912)) function, I had some ideas that ended up working better than I'd expected.
There's three main ideas in here, split over three commits:
1. Don't use `array::IntoIter` when we can avoid it, since that seems to not get SRoA'd, meaning that every step writes things like loop counters into the stack unnecessarily
2. Don't return arrays in `Result`s unnecessarily, as that doesn't seem to optimize away even with `unwrap_unchecked` (perhaps because it needs to get moved into a new LLVM type to account for the discriminant)
3. Don't distract LLVM with all the `Option` dances when we know for sure we have enough items (like in `map` and `zip`). This one's a larger commit as to do it I ended up adding a new `pub(crate)` trait, but hopefully those changes are still straight-forward.
(No libs-api changes; everything should be completely implementation-detail-internal.)
It's still not completely fixed -- I think it needs pcwalton's `memcpy` optimizations still (#103830) to get further -- but this seems to go much better than before. And the remaining `memcpy`s are just `transmute`-equivalent (`[T; N] -> ManuallyDrop<[T; N]>` and `[MaybeUninit<T>; N] -> [T; N]`), so hopefully those will be easier to remove with LLVM16 than the previous subobject copies 🤞
r? `@thomcc`
As a simple example, this test
```rust
pub fn long_integer_map(x: [u32; 64]) -> [u32; 64] {
x.map(|x| 13 * x + 7)
}
```
On nightly <https://rust.godbolt.org/z/xK7548TGj> takes `sub rsp, 808`
```llvm
start:
%array.i.i.i.i = alloca [64 x i32], align 4
%_3.sroa.5.i.i.i = alloca [65 x i32], align 4
%_5.i = alloca %"core::iter::adapters::map::Map<core::array::iter::IntoIter<u32, 64>, [closure@/app/example.rs:2:11: 2:14]>", align 8
```
(and yes, that's a 6**5**-element array `alloca` despite 6**4**-element input and output)
But with this PR it's only `sub rsp, 520`
```llvm
start:
%array.i.i.i.i.i.i = alloca [64 x i32], align 4
%array1.i.i.i = alloca %"core::mem::manually_drop::ManuallyDrop<[u32; 64]>", align 4
```
Similarly, the loop it emits on nightly is scalar-only and horrifying
```nasm
.LBB0_1:
mov esi, 64
mov edi, 0
cmp rdx, 64
je .LBB0_3
lea rsi, [rdx + 1]
mov qword ptr [rsp + 784], rsi
mov r8d, dword ptr [rsp + 4*rdx + 528]
mov edi, 1
lea edx, [r8 + 2*r8]
lea r8d, [r8 + 4*rdx]
add r8d, 7
.LBB0_3:
test edi, edi
je .LBB0_11
mov dword ptr [rsp + 4*rcx + 272], r8d
cmp rsi, 64
jne .LBB0_6
xor r8d, r8d
mov edx, 64
test r8d, r8d
jne .LBB0_8
jmp .LBB0_11
.LBB0_6:
lea rdx, [rsi + 1]
mov qword ptr [rsp + 784], rdx
mov edi, dword ptr [rsp + 4*rsi + 528]
mov r8d, 1
lea esi, [rdi + 2*rdi]
lea edi, [rdi + 4*rsi]
add edi, 7
test r8d, r8d
je .LBB0_11
.LBB0_8:
mov dword ptr [rsp + 4*rcx + 276], edi
add rcx, 2
cmp rcx, 64
jne .LBB0_1
```
whereas with this PR it's unrolled and vectorized
```nasm
vpmulld ymm1, ymm0, ymmword ptr [rsp + 64]
vpaddd ymm1, ymm1, ymm2
vmovdqu ymmword ptr [rsp + 328], ymm1
vpmulld ymm1, ymm0, ymmword ptr [rsp + 96]
vpaddd ymm1, ymm1, ymm2
vmovdqu ymmword ptr [rsp + 360], ymm1
```
(though sadly still stack-to-stack)
Refactor iteration logic in the `Flatten` and `FlatMap` iterators
The `Flatten` and `FlatMap` iterators both delegate to `FlattenCompat`:
```rust
struct FlattenCompat<I, U> {
iter: Fuse<I>,
frontiter: Option<U>,
backiter: Option<U>,
}
```
Every individual iterator method that `FlattenCompat` implements needs to carefully manage this state, checking whether the `frontiter` and `backiter` are present, and storing the current iterator appropriately if iteration is aborted. This has led to methods such as `next`, `advance_by`, and `try_fold` all having similar code for managing the iterator's state.
I have extracted this common logic of iterating the inner iterators with the option to exit early into a `iter_try_fold` method:
```rust
impl<I, U> FlattenCompat<I, U>
where
I: Iterator<Item: IntoIterator<IntoIter = U>>,
{
fn iter_try_fold<Acc, Fold, R>(&mut self, acc: Acc, fold: Fold) -> R
where
Fold: FnMut(Acc, &mut U) -> R,
R: Try<Output = Acc>,
{ ... }
}
```
It passes each of the inner iterators to the given function as long as it keep succeeding. It takes care of managing `FlattenCompat`'s state, so that the actual `Iterator` methods don't need to. The resulting code that makes use of this abstraction is much more straightforward:
```rust
fn next(&mut self) -> Option<U::Item> {
#[inline]
fn next<U: Iterator>((): (), iter: &mut U) -> ControlFlow<U::Item> {
match iter.next() {
None => ControlFlow::CONTINUE,
Some(x) => ControlFlow::Break(x),
}
}
self.iter_try_fold((), next).break_value()
}
```
Note that despite being implemented in terms of `iter_try_fold`, `next` is still able to benefit from `U`'s `next` method. It therefore does not take the performance hit that implementing `next` directly in terms of `Self::try_fold` causes (in some benchmarks).
This PR also adds `iter_try_rfold` which captures the shared logic of `try_rfold` and `advance_back_by`, as well as `iter_fold` and `iter_rfold` for folding without early exits (used by `fold`, `rfold`, `count`, and `last`).
Benchmark results:
```
before after
bench_flat_map_sum 423,255 ns/iter 414,338 ns/iter
bench_flat_map_ref_sum 1,942,139 ns/iter 2,216,643 ns/iter
bench_flat_map_chain_sum 1,616,840 ns/iter 1,246,445 ns/iter
bench_flat_map_chain_ref_sum 4,348,110 ns/iter 3,574,775 ns/iter
bench_flat_map_chain_option_sum 780,037 ns/iter 780,679 ns/iter
bench_flat_map_chain_option_ref_sum 2,056,458 ns/iter 834,932 ns/iter
```
I added the last two benchmarks specifically to demonstrate an extreme case where `FlatMap::next` can benefit from custom internal iteration of the outer iterator, so take it with a grain of salt. We should probably do a perf run to see if the changes to `next` are worth it in practice.
Let `try_collect` take advantage of `try_fold` overrides
No public API changes.
With this change, `try_collect` (#94047) is no longer going through the `impl Iterator for &mut impl Iterator`, and thus will be able to use `try_fold` overrides instead of being forced through `next` for every element.
Here's the test added, to see that it fails before this PR (once a new enough nightly is out): https://play.rust-lang.org/?version=nightly&mode=debug&edition=2021&gist=462f2896f2fed2c238ee63ca1a7e7c56
This might as well go to the same person as my last `try_process` PR (#93572), so
r? ``@yaahc``
Enable conditional checking of values in the Rust codebase
This pull-request enable conditional checking of (well known) values in the Rust codebase.
Well known values were added in https://github.com/rust-lang/rust/pull/94362. All the `target_*` values are taken from all the built-in targets which is why some extra values were needed do be added as they are not (yet ?) defined in any built-in targets.
r? `@Mark-Simulacrum`
The `advance_by(n)` docs state that in the error case `Err(k)` that k is always less than n.
It also states that `advance_by(0)` may return `Err(0)` to indicate an exhausted iterator.
These statements are inconsistent.
Since only one implementation (Skip) actually made use of that I changed it to return Ok(()) in that case too.
While adding some tests I also found a bug in `Take::advance_back_by`.
Due to #20400 the corresponding TrustedLen impls need a helper trait
instead of directly adding `Item = &[T;N]` bounds.
Since TrustedLen is a public trait this in turn means
the helper trait needs to be public. Since it's just a workaround
for a compiler deficit it's marked hidden, unstable and unsafe.
This only works if arrays are passed directly instead of array iterators
because we need to be sure that they have not been advanced before
Flatten does its size calculation.
This was unsound since a panic in a.next_back() would result in the
length not being updated which would then lead to the same element
being revisited in the side-effect preserving code.
