Add initial support for unsized MaybeUninit wrapper type

This is achieved by adding a `MaybeUninit` associated type to
`KnownLayout`, whose layout is identical to `Self` except that it admits
uninitialized bytes in all positions.

For sized types, this is bound to `mem::MaybeUninit<Self>`. For
potentially unsized structs, we synthesize a doppelganger with the same
`repr`, whose leading fields are wrapped in `mem::MaybeUninit` and whose
trailing field is the `MaybeUninit` associated type of struct's original
trailing field type. This type-level recursion bottoms out at `[T]`,
whose `MaybeUninit` associated type is bound to `[mem::MaybeUninit<T>]`.

Makes progress towards #1797

SKIP_CARGO_SEMVER_CHECKS=1

Author: jswrenn

.github/workflows/ci.yml

@@ -244,6 +244,8 @@ jobs:
 
     steps:
     - uses: actions/checkout@eef61447b9ff4aafe5dcd4e0bbf5d482be7e7871 # v4.2.1
+      with:
+        fetch-depth: 2
 
     - name: Populate cache
       uses: ./.github/actions/cache
@@ -490,7 +492,7 @@ jobs:
 
         if [ "${{ github.event_name }}" == "pull_request" ]; then
           # Invoked from a PR - get the PR body directly
-          MESSAGE="${{ github.event.pull_request.body }}"
+          MESSAGE="$(git log -1 --pretty=%B ${{ github.event.pull_request.head.sha }})"
         else
           # Invoked from the merge queue - get the commit message
           MESSAGE="$(git log -1 --pretty=%B ${{ github.sha }})"

src/impls.rs

@@ -7,6 +7,8 @@
 // This file may not be copied, modified, or distributed except according to
 // those terms.
 
+use core::mem::MaybeUninit as CoreMaybeUninit;
+
 use super::*;
 
 safety_comment! {
@@ -632,14 +634,14 @@ safety_comment! {
     /// SAFETY:
     /// `TryFromBytes` (with no validator), `FromZeros`, `FromBytes`:
     /// `MaybeUninit<T>` has no restrictions on its contents.
-    unsafe_impl!(T => TryFromBytes for MaybeUninit<T>);
-    unsafe_impl!(T => FromZeros for MaybeUninit<T>);
-    unsafe_impl!(T => FromBytes for MaybeUninit<T>);
+    unsafe_impl!(T => TryFromBytes for CoreMaybeUninit<T>);
+    unsafe_impl!(T => FromZeros for CoreMaybeUninit<T>);
+    unsafe_impl!(T => FromBytes for CoreMaybeUninit<T>);
 }
 
-impl_for_transparent_wrapper!(T: Immutable => Immutable for MaybeUninit<T>);
-impl_for_transparent_wrapper!(T: Unaligned => Unaligned for MaybeUninit<T>);
-assert_unaligned!(MaybeUninit<()>, MaybeUninit<u8>);
+impl_for_transparent_wrapper!(T: Immutable => Immutable for CoreMaybeUninit<T>);
+impl_for_transparent_wrapper!(T: Unaligned => Unaligned for CoreMaybeUninit<T>);
+assert_unaligned!(CoreMaybeUninit<()>, CoreMaybeUninit<u8>);
 
 impl_for_transparent_wrapper!(T: ?Sized + Immutable => Immutable for ManuallyDrop<T>);
 impl_for_transparent_wrapper!(T: ?Sized + TryFromBytes => TryFromBytes for ManuallyDrop<T>);
@@ -1257,8 +1259,8 @@ mod tests {
                             ManuallyDrop<UnsafeCell<()>>,
                             ManuallyDrop<[UnsafeCell<u8>]>,
                             ManuallyDrop<[UnsafeCell<bool>]>,
-                            MaybeUninit<NotZerocopy>,
-                            MaybeUninit<UnsafeCell<()>>,
+                            CoreMaybeUninit<NotZerocopy>,
+                            CoreMaybeUninit<UnsafeCell<()>>,
                             Wrapping<UnsafeCell<()>>
                         );
 
@@ -1300,9 +1302,9 @@ mod tests {
                             Option<FnManyArgs>,
                             Option<extern "C" fn()>,
                             Option<ECFnManyArgs>,
-                            MaybeUninit<u8>,
-                            MaybeUninit<NotZerocopy>,
-                            MaybeUninit<UnsafeCell<()>>,
+                            CoreMaybeUninit<u8>,
+                            CoreMaybeUninit<NotZerocopy>,
+                            CoreMaybeUninit<UnsafeCell<()>>,
                             ManuallyDrop<UnsafeCell<()>>,
                             ManuallyDrop<[UnsafeCell<u8>]>,
                             ManuallyDrop<[UnsafeCell<bool>]>,
@@ -1764,9 +1766,9 @@ mod tests {
         assert_impls!(ManuallyDrop<[UnsafeCell<u8>]>: KnownLayout, TryFromBytes, FromZeros, FromBytes, IntoBytes, Unaligned, !Immutable);
         assert_impls!(ManuallyDrop<[UnsafeCell<bool>]>: KnownLayout, TryFromBytes, FromZeros, IntoBytes, Unaligned, !Immutable, !FromBytes);
 
-        assert_impls!(MaybeUninit<u8>: KnownLayout, Immutable, TryFromBytes, FromZeros, FromBytes, Unaligned, !IntoBytes);
-        assert_impls!(MaybeUninit<NotZerocopy>: KnownLayout, TryFromBytes, FromZeros, FromBytes, !Immutable, !IntoBytes, !Unaligned);
-        assert_impls!(MaybeUninit<UnsafeCell<()>>: KnownLayout, TryFromBytes, FromZeros, FromBytes, Unaligned, !Immutable, !IntoBytes);
+        assert_impls!(CoreMaybeUninit<u8>: KnownLayout, Immutable, TryFromBytes, FromZeros, FromBytes, Unaligned, !IntoBytes);
+        assert_impls!(CoreMaybeUninit<NotZerocopy>: KnownLayout, TryFromBytes, FromZeros, FromBytes, !Immutable, !IntoBytes, !Unaligned);
+        assert_impls!(CoreMaybeUninit<UnsafeCell<()>>: KnownLayout, TryFromBytes, FromZeros, FromBytes, Unaligned, !Immutable, !IntoBytes);
 
         assert_impls!(Wrapping<u8>: KnownLayout, Immutable, TryFromBytes, FromZeros, FromBytes, IntoBytes, Unaligned);
         // This test is important because it allows us to test our hand-rolled

src/lib.rs

@@ -354,7 +354,7 @@ use core::{
     fmt::{self, Debug, Display, Formatter},
     hash::Hasher,
     marker::PhantomData,
-    mem::{self, ManuallyDrop, MaybeUninit},
+    mem::{self, ManuallyDrop, MaybeUninit as CoreMaybeUninit},
     num::{
         NonZeroI128, NonZeroI16, NonZeroI32, NonZeroI64, NonZeroI8, NonZeroIsize, NonZeroU128,
         NonZeroU16, NonZeroU32, NonZeroU64, NonZeroU8, NonZeroUsize, Wrapping,
@@ -732,6 +732,15 @@ pub unsafe trait KnownLayout {
     /// This is `()` for sized types and `usize` for slice DSTs.
     type PointerMetadata: PointerMetadata;
 
+    /// A maybe-uninitialized analog of `Self`
+    ///
+    /// # Safety
+    ///
+    /// `Self::LAYOUT` and `Self::MaybeUninit::LAYOUT` are identical.
+    /// `Self::MaybeUninit` admits uninitialized bytes in all positions.
+    #[doc(hidden)]
+    type MaybeUninit: ?Sized + KnownLayout<PointerMetadata = Self::PointerMetadata>;
+
     /// The layout of `Self`.
     ///
     /// # Safety
@@ -864,6 +873,35 @@ unsafe impl<T> KnownLayout for [T] {
 
     type PointerMetadata = usize;
 
+    // SAFETY: `CoreMaybeUninit<T>::LAYOUT` and `T::LAYOUT` are identical
+    // because `CoreMaybeUninit<T>` has the same size and alignment as `T` [1].
+    // Consequently, `[CoreMaybeUninit<T>]::LAYOUT` and `[T]::LAYOUT` are
+    // identical, because they both lack a fixed-sized prefix and because they
+    // inherit the alignments of their inner element type (which are identical)
+    // [2][3].
+    //
+    // `[CoreMaybeUninit<T>]` admits uninitialized bytes at all positions
+    // because `CoreMaybeUninit<T>` admits uninitialized bytes at all positions
+    // and because the inner elements of `[CoreMaybeUninit<T>]` are laid out
+    // back-to-back [2][3].
+    //
+    // [1] Per https://doc.rust-lang.org/1.81.0/std/mem/union.MaybeUninit.html#layout-1:
+    //
+    //   `MaybeUninit<T>` is guaranteed to have the same size, alignment, and ABI as
+    //   `T`
+    //
+    // [2] Per https://doc.rust-lang.org/1.82.0/reference/type-layout.html#slice-layout:
+    //
+    //   Slices have the same layout as the section of the array they slice.
+    //
+    // [3] Per https://doc.rust-lang.org/1.82.0/reference/type-layout.html#array-layout:
+    //
+    //   An array of `[T; N]` has a size of `size_of::<T>() * N` and the same
+    //   alignment of `T`. Arrays are laid out so that the zero-based `nth`
+    //   element of the array is offset from the start of the array by `n *
+    //   size_of::<T>()` bytes.
+    type MaybeUninit = [CoreMaybeUninit<T>];
+
     const LAYOUT: DstLayout = DstLayout::for_slice::<T>();
 
     // SAFETY: `.cast` preserves address and provenance. The returned pointer
@@ -916,9 +954,11 @@ impl_known_layout!(
     T         => Option<T>,
     T: ?Sized => PhantomData<T>,
     T         => Wrapping<T>,
-    T         => MaybeUninit<T>,
+    T         => CoreMaybeUninit<T>,
     T: ?Sized => *const T,
-    T: ?Sized => *mut T
+    T: ?Sized => *mut T,
+    T: ?Sized => &'_ T,
+    T: ?Sized => &'_ mut T,
 );
 impl_known_layout!(const N: usize, T => [T; N]);
 
@@ -949,6 +989,21 @@ safety_comment! {
     unsafe_impl_known_layout!(T: ?Sized + KnownLayout => #[repr(T)] UnsafeCell<T>);
 }
 
+safety_comment! {
+    /// SAFETY:
+    /// - By consequence of the invariant on `T::MaybeUninit` that `T::LAYOUT`
+    ///   and `T::MaybeUninit::LAYOUT` are equal, `T` and `T::MaybeUninit`
+    ///   have the same:
+    ///   - Fixed prefix size
+    ///   - Alignment
+    ///   - (For DSTs) trailing slice element size
+    /// - By consequence of the above, referents `T::MaybeUninit` and `T` have
+    ///   the require the same kind of pointer metadata, and thus it is valid to
+    ///   perform an `as` cast from `*mut T` and `*mut T::MaybeUninit`, and this
+    ///   operation preserves referent size (ie, `size_of_val_raw`).
+    unsafe_impl_known_layout!(T: ?Sized + KnownLayout => #[repr(T::MaybeUninit)] MaybeUninit<T>);
+}
+
 /// Analyzes whether a type is [`FromZeros`].
 ///
 /// This derive analyzes, at compile time, whether the annotated type satisfies
@@ -2550,7 +2605,7 @@ pub unsafe trait TryFromBytes {
     where
         Self: Sized,
     {
-        let candidate = match MaybeUninit::<Self>::read_from_bytes(source) {
+        let candidate = match CoreMaybeUninit::<Self>::read_from_bytes(source) {
             Ok(candidate) => candidate,
             Err(e) => {
                 return Err(TryReadError::Size(e.with_dst()));
@@ -2611,7 +2666,7 @@ pub unsafe trait TryFromBytes {
     where
         Self: Sized,
     {
-        let (candidate, suffix) = match MaybeUninit::<Self>::read_from_prefix(source) {
+        let (candidate, suffix) = match CoreMaybeUninit::<Self>::read_from_prefix(source) {
             Ok(candidate) => candidate,
             Err(e) => {
                 return Err(TryReadError::Size(e.with_dst()));
@@ -2673,7 +2728,7 @@ pub unsafe trait TryFromBytes {
     where
         Self: Sized,
     {
-        let (prefix, candidate) = match MaybeUninit::<Self>::read_from_suffix(source) {
+        let (prefix, candidate) = match CoreMaybeUninit::<Self>::read_from_suffix(source) {
             Ok(candidate) => candidate,
             Err(e) => {
                 return Err(TryReadError::Size(e.with_dst()));
@@ -2746,7 +2801,7 @@ fn swap<T, U>((t, u): (T, U)) -> (U, T) {
 #[inline(always)]
 unsafe fn try_read_from<S, T: TryFromBytes>(
     source: S,
-    mut candidate: MaybeUninit<T>,
+    mut candidate: CoreMaybeUninit<T>,
 ) -> Result<T, TryReadError<S, T>> {
     // We use `from_mut` despite not mutating via `c_ptr` so that we don't need
     // to add a `T: Immutable` bound.
@@ -3035,72 +3090,11 @@ pub unsafe trait FromZeros: TryFromBytes {
     where
         Self: KnownLayout<PointerMetadata = usize>,
     {
-        let size = match count.size_for_metadata(Self::LAYOUT) {
-            Some(size) => size,
-            None => return Err(AllocError),
-        };
-
-        let align = Self::LAYOUT.align.get();
-        // On stable Rust versions <= 1.64.0, `Layout::from_size_align` has a
-        // bug in which sufficiently-large allocations (those which, when
-        // rounded up to the alignment, overflow `isize`) are not rejected,
-        // which can cause undefined behavior. See #64 for details.
-        //
-        // TODO(#67): Once our MSRV is > 1.64.0, remove this assertion.
-        #[allow(clippy::as_conversions)]
-        let max_alloc = (isize::MAX as usize).saturating_sub(align);
-        if size > max_alloc {
-            return Err(AllocError);
-        }
-
-        // TODO(https://github.com/rust-lang/rust/issues/55724): Use
-        // `Layout::repeat` once it's stabilized.
-        let layout = Layout::from_size_align(size, align).or(Err(AllocError))?;
-
-        let ptr = if layout.size() != 0 {
-            // TODO(#429): Add a "SAFETY" comment and remove this `allow`.
-            #[allow(clippy::undocumented_unsafe_blocks)]
-            let ptr = unsafe { alloc::alloc::alloc_zeroed(layout) };
-            match NonNull::new(ptr) {
-                Some(ptr) => ptr,
-                None => return Err(AllocError),
-            }
-        } else {
-            let align = Self::LAYOUT.align.get();
-            // We use `transmute` instead of an `as` cast since Miri (with
-            // strict provenance enabled) notices and complains that an `as`
-            // cast creates a pointer with no provenance. Miri isn't smart
-            // enough to realize that we're only executing this branch when
-            // we're constructing a zero-sized `Box`, which doesn't require
-            // provenance.
-            //
-            // SAFETY: any initialized bit sequence is a bit-valid `*mut u8`.
-            // All bits of a `usize` are initialized.
-            #[allow(clippy::useless_transmute)]
-            let dangling = unsafe { mem::transmute::<usize, *mut u8>(align) };
-            // SAFETY: `dangling` is constructed from `Self::LAYOUT.align`,
-            // which is a `NonZeroUsize`, which is guaranteed to be non-zero.
-            //
-            // `Box<[T]>` does not allocate when `T` is zero-sized or when `len`
-            // is zero, but it does require a non-null dangling pointer for its
-            // allocation.
-            //
-            // TODO(https://github.com/rust-lang/rust/issues/95228): Use
-            // `std::ptr::without_provenance` once it's stable. That may
-            // optimize better. As written, Rust may assume that this consumes
-            // "exposed" provenance, and thus Rust may have to assume that this
-            // may consume provenance from any pointer whose provenance has been
-            // exposed.
-            unsafe { NonNull::new_unchecked(dangling) }
-        };
-
-        let ptr = Self::raw_from_ptr_len(ptr, count);
-
-        // TODO(#429): Add a "SAFETY" comment and remove this `allow`. Make sure
-        // to include a justification that `ptr.as_ptr()` is validly-aligned in
-        // the ZST case (in which we manually construct a dangling pointer).
-        #[allow(clippy::undocumented_unsafe_blocks)]
-        Ok(unsafe { Box::from_raw(ptr.as_ptr()) })
+        // SAFETY: `alloc::alloc::alloc_zeroed` is a valid argument of
+        // `new_box`. The referent of the pointer returned by `alloc_zeroed`
+        // (and, consequently, the `Box` derived from it) is a valid instance of
+        // `Self`, because `Self` is `FromZeros`.
+        unsafe { crate::util::new_box(count, alloc::alloc::alloc_zeroed) }
     }
 
     #[deprecated(since = "0.8.0", note = "renamed to `FromZeros::new_box_zeroed_with_elems`")]
@@ -4562,7 +4556,7 @@ pub unsafe trait FromBytes: FromZeros {
         Self: Sized,
         R: io::Read,
     {
-        let mut buf = MaybeUninit::<Self>::zeroed();
+        let mut buf = CoreMaybeUninit::<Self>::zeroed();
         let ptr = Ptr::from_mut(&mut buf);
         // SAFETY: `buf` consists entirely of initialized, zeroed bytes.
         let ptr = unsafe { ptr.assume_validity::<invariant::Initialized>() };

src/util/macros.rs

@@ -557,6 +557,17 @@ macro_rules! impl_known_layout {
 
                 type PointerMetadata = ();
 
+                // SAFETY: `CoreMaybeUninit<T>::LAYOUT` and `T::LAYOUT` are
+                // identical because `CoreMaybeUninit<T>` has the same size and
+                // alignment as `T` [1], and `CoreMaybeUninit` admits
+                // uninitialized bytes in all positions.
+                //
+                // [1] Per https://doc.rust-lang.org/1.81.0/std/mem/union.MaybeUninit.html#layout-1:
+                //
+                //   `MaybeUninit<T>` is guaranteed to have the same size,
+                //   alignment, and ABI as `T`
+                type MaybeUninit = core::mem::MaybeUninit<Self>;
+
                 const LAYOUT: crate::DstLayout = crate::DstLayout::for_type::<$ty>();
 
                 // SAFETY: `.cast` preserves address and provenance.
@@ -599,6 +610,7 @@ macro_rules! unsafe_impl_known_layout {
                 fn only_derive_is_allowed_to_implement_this_trait() {}
 
                 type PointerMetadata = <$repr as KnownLayout>::PointerMetadata;
+                type MaybeUninit = <$repr as KnownLayout>::MaybeUninit;
 
                 const LAYOUT: DstLayout = <$repr as KnownLayout>::LAYOUT;