* parse `repr(usize)` and `repr(isize)`
This can be used with enums.
* emit packed repr without type suffix
Previously we emitted the packed attribute with a type suffix for the
value `packed(4u32)`. This is not allowed. Instead we should emit
`packed(4)`.
* read bits instead of creating reference
This is required for packed structs where creating a reference to
fields is not allowed. We can make a copy of the bits because the bits
are `AnyBitPattern` which implies `Copy`.
* emit generics on implied traits
* `impl Zeroable for Atomic*`
fixes#74
* added feature flag for `zeroable_atomics`
Added documentation for some feature flags on the `Zeroable` trait.
Also fixed some invalid references in doc-comments.
* Implement CheckedBitPattern for core::num::NonZero*.
* Add tests for CheckedBitPattern.
* Fix stacked borrows violation.
`<&mut T>::as_ptr` gives a `*const T` which is read-only. `<&mut T>::as_mut_ptr` is necessary to get a `*mut T`.
* MSRV fix.
* Add layout checks for CheckedBitPattern for NonZero*.
(Until Rust guarantees the layout of NonZero[int] matches [int].)
Also adds a test that will fail if any NonZero[int] has a different size or alignment than [int].
Currently, `checked::pod_read_unaligned` requires `T: AnyBitPattern`. (This was likely just a copy-paste error from when the function was introduced in #91.)
There should be no current UB nor semver breakage, since this only relaxes the bound (any type this previously worked with was already sound and will continue to work).
`checked::try_pod_read_unaligned has the correct `T: CheckedBitPattern` bound.
* Implement `ByteEq` and `ByteHash` derives
This adds the derives `ByteEq` and `ByteHash` that can be used as an
alternative to the `Eq` / `PartialEq` and `Hash` derives from the
standard library. The difference is that these variants use `bytemuck`
to convert their values to byte slices before comparing / hashing them.
This allows the comparisons to turn into a simple `memcmp` / `bcmp` (or
completely inlined as a few vector instructions) and allows hashers to
process all bytes at once, possibly allowing for some vector operations
as well.
Here's a quick comparison of the generated assembly:
* Address review comments
Apparently the latest release of `bytemuck` doesn't compile on some of
the targets that it used to because it uses `Arc`. `Arc` is not a type
that exists on every target, because not every targets supports atomics.
It seems like a copy-paste error has happened between `try_from_bytes`/`try_from_bytes_mut` and `try_pod_read_unaligned`, causing `internal::try_pod_read_unaligned` to try to read a &T::Bits instead of a T::Bits, usually failing with a `SizeMismatch` error. In the worst case, this allows UB in safe code by having a type allowing any bit pattern and being pointer-sized.
* Add `allocation::{try_,}cast_{arc,rc}`, and add `{wrap,peel}_{arc,rc}` to `TransparentWrapperAlloc`.
* Avoid intermediate slice reference in `try_cast_slice_{arc,rc}`.
* remove `unsafe` block; run `cargo +nightly fmt` (ignoring files I didn't modify)
* Make `cast_rc` (etc) have the same bounds as `cast_mut`, due to the existence of `Rc::get_mut_unchecked`.
