Once a target feature is enabled for a function that means that it in general
can't be inlined into other functions which don't have that target feature
enabled. This can cause both safety and LLVM issues if we were to actually
inline it, so `#[inline(always)]` both can't be respected and would be an error
if we did so!
Today LLVM doesn't inline functions with different `#[target_feature]`
annotations, but it turns out that if one is tagged with `#[inline(always)]`
it'll override this and cause scary LLVM error to arise!
This commit fixes this issue by forbidding these two attributes to be used in
conjunction with one another.
cc rust-lang-nursery/stdsimd#404
In `LLVMRustHasFeature()`, rather than using `MCInfo->getFeatureTable()`
that is specific to Rust's LLVM fork, we can use this in LLVM 6:
/// Check whether the subtarget features are enabled/disabled as per
/// the provided string, ignoring all other features.
bool checkFeatures(StringRef FS) const;
Now rustc using external LLVM can also have `target_feature`.
Stabilize TryFrom / TryInto, and tweak impls for integers
Fixes https://github.com/rust-lang/rust/issues/33417 (tracking issue)
----
This adds:
* `impl From<u16> for usize`
* `impl From<i16> for isize`
* `impl From<u8> for isize`
… replacing corresponding `TryFrom<Error=!>` impls. (`TryFrom` still applies through the generic `impl<T, U> TryFrom<U> for T where T: From<U>`.) Their infallibility is supported by the C99 standard which (indirectly) requires pointers to be at least 16 bits.
The remaining `TryFrom` impls that define `type Error = !` all involve `usize` or `isize`. This PR changes them to use `TryFromIntError` instead, since having a return type change based on the target is a portability hazard.
Note: if we make similar assumptions about the *maximum* bit size of pointers (for all targets Rust will ever run on in the future), we could have similar `From` impls converting pointer-sized integers to large fixed-size integers. RISC-V considers the possibility of a 128-bit address space (RV128), which would leave only `impl From<usize> for u128` and `impl From<isize> for u128`. I [found](https://www.cl.cam.ac.uk/research/security/ctsrd/pdfs/20171017a-cheri-poster.pdf) some [things](http://www.csl.sri.com/users/neumann/2012resolve-cheri.pdf) about 256-bit “capabilities”, but I don’t know how relevant that would be to Rust’s `usize` and `isize` types.
I chose conservatively to make no assumption about the future there. Users making their portability decisions and using something like `.try_into().unwrap()`.
----
Since this feature already went through FCP in the tracking issue https://github.com/rust-lang/rust/issues/33417, this PR also proposes **stabilize** the following items:
* The `convert::TryFrom` trait
* The `convert::TryFrom` trait
* `impl<T> TryFrom<&[T]> for &[T; $N]` (for `$N` up to 32)
* `impl<T> TryFrom<&mut [T]> for &mut [T; $N]` (for `$N` up to 32)
* The `array::TryFromSliceError` struct, with impls of `Debug`, `Copy`, `Clone`, and `Error`
* `impl TryFrom<u32> for char`
* The `char::CharTryFromError` struct, with impls of `Copy`, `Clone`, `Debug`, `PartialEq`, `Eq`, `Display`, and `Error`
* Impls of `TryFrom` for all (?) combinations of primitive integer types where `From` isn’t implemented.
* The `num::TryFromIntError` struct, with impls of `Debug`, `Copy`, `Clone`, `Display`, `From<!>`, and `Error`
Some minor remaining questions that I hope can be resolved in this PR:
* Should the impls for error types be unified?
* ~Should `TryFrom` and `TryInto` be in the prelude? `From` and `Into` are.~ (Yes.)
libsyntax: Remove obsolete.rs
This little piece of infra is obsolete (ha-ha) and is unlikely to be used in the future, even if new obsolete syntax appears.
Add is_whitespace and is_alphanumeric to str.
The other methods from `UnicodeStr` are already stable inherent
methods on str, but these have not been included.
r? @SimonSapin
Add slice::sort_by_cached_key as a memoised sort_by_key
At present, `slice::sort_by_key` calls its key function twice for each comparison that is made. When the key function is expensive (which can often be the case when `sort_by_key` is chosen over `sort_by`), this can lead to very suboptimal behaviour.
To address this, I've introduced a new slice method, `sort_by_cached_key`, which has identical semantic behaviour to `sort_by_key`, except that it guarantees the key function will only be called once per element.
Where there are `n` elements and the key function is `O(m)`:
- `slice::sort_by_cached_key` time complexity is `O(m n log m n)`, compared to `slice::sort_by_key`'s `O(m n + n log n)`.
- `slice::sort_by_cached_key` space complexity remains at `O(n + m)`. (Technically, it now reserves a slice of size `n`, whereas before it reserved a slice of size `n/2`.)
`slice::sort_unstable_by_key` has not been given an analogue, as it is important that unstable sorts are in-place, which is not a property that is guaranteed here. However, this also means that `slice::sort_unstable_by_key` is likely to be slower than `slice::sort_by_cached_key` when the key function does not have negligible complexity. We might want to explore this trade-off further in the future.
Benchmarks (for a vector of 100 `i32`s):
```
# Lexicographic: `|x| x.to_string()`
test bench_sort_by_key ... bench: 112,638 ns/iter (+/- 19,563)
test bench_sort_by_cached_key ... bench: 15,038 ns/iter (+/- 4,814)
# Identity: `|x| *x`
test bench_sort_by_key ... bench: 1,346 ns/iter (+/- 238)
test bench_sort_by_cached_key ... bench: 1,839 ns/iter (+/- 765)
# Power: `|x| x.pow(31)`
test bench_sort_by_key ... bench: 3,624 ns/iter (+/- 738)
test bench_sort_by_cached_key ... bench: 1,997 ns/iter (+/- 311)
# Abs: `|x| x.abs()`
test bench_sort_by_key ... bench: 1,546 ns/iter (+/- 174)
test bench_sort_by_cached_key ... bench: 1,668 ns/iter (+/- 790)
```
(So it seems functions that are single operations do perform slightly worse with this method, but for pretty much any more complex key, you're better off with this optimisation.)
I've definitely found myself using expensive keys in the past and wishing this optimisation was made (e.g. for https://github.com/rust-lang/rust/pull/47415). This feels like both desirable and expected behaviour, at the small cost of slightly more stack allocation and minute degradation in performance for extremely trivial keys.
Resolves#34447.
Fix implicit closure return type generation for libsyntax
The `lambda` function for constructing closures in libsyntax was explicitly setting the return type to `_`, which resulted in incorrect corresponding syntax (as `|| -> _ x` is not valid, without the enclosing brackets). This meant the generated code, when printed, was invalid.
I also took the opportunity to slightly improve the generated code for the `RustcEncodable::encode` method for unit structs.
Fixes#42213.
implement minmax intrinsics
This adds the `simd_{fmin,fmax}` intrinsics, which do a vertical (lane-wise) `min`/`max` for floating point vectors that's equivalent to Rust's `min`/`max` for `f32`/`f64`.
It might make sense to make `{f32,f64}::{min,max}` use the `minnum` and `minmax` intrinsics as well.
---
~~HELP: I need some help with these. Either I should go to sleep or there must be something that I must be missing. AFAICT I am calling the `maxnum` builder correctly, yet rustc/LLVM seem to insert a call to `llvm.minnum` there instead...~~ EDIT: Rust's LLVM version is too old :/
Simply checking for the presence of `llvm.memset` is too brittle because
this instrinsic can be used for seemingly trivial operations, such as
zero-initializing a `RawVec`.
There are types that implement `Clone` and `Copy` but are not mentioned
in the documentation, because the implementations are provided by the
compiler. They are types of variants that cannot be fully covered by
trait implementations in Rust code, because the language is not
expressive enough.
