Teach transmute_{ref,mut}! to handle slice DSTs

TODO:
- Convince ourselves that it's acceptable to drop the `TransmuteFrom:
  SizeCompatible` super-trait bound.
- Write Kani tests for layout computations
- Print useful error message when `try_compute_cast_params` fails during
  static assertion
- Test when post padding doesn't match between types

Makes progress on #1817

Co-authored-by: Jack Wrenn <jswrenn@amazon.com>
gherrit-pr-id: Ib4bc62202e0b3b09d155333b525087f7aa8f02c2

Author: joshlf

src/impls.rs

@@ -179,39 +179,13 @@ safety_comment! {
     });
 }
 
-// SAFETY: `str` and `[u8]` have the same layout [1].
-//
-// [1] Per https://doc.rust-lang.org/1.81.0/reference/type-layout.html#str-layout:
-//
-//   String slices are a UTF-8 representation of characters that have the same
-//   layout as slices of type `[u8]`.
-unsafe impl pointer::SizeEq<str> for [u8] {
-    fn cast_from_raw(s: NonNull<str>) -> NonNull<[u8]> {
-        cast!(s)
-    }
-}
-// SAFETY: See previous safety comment.
-unsafe impl pointer::SizeEq<[u8]> for str {
-    fn cast_from_raw(bytes: NonNull<[u8]>) -> NonNull<str> {
-        cast!(bytes)
-    }
-}
+impl_size_eq!(str, [u8]);
 
 macro_rules! unsafe_impl_try_from_bytes_for_nonzero {
     ($($nonzero:ident[$prim:ty]),*) => {
         $(
             unsafe_impl!(=> TryFromBytes for $nonzero; |n| {
-                unsafe impl pointer::SizeEq<$nonzero> for Unalign<$prim> {
-                    fn cast_from_raw(n: NonNull<$nonzero>) -> NonNull<Unalign<$prim>> {
-                        cast!(n)
-                    }
-                }
-                unsafe impl pointer::SizeEq<Unalign<$prim>> for $nonzero {
-                    fn cast_from_raw(p: NonNull<Unalign<$prim>>) -> NonNull<$nonzero> {
-                        cast!(p)
-                    }
-                }
-
+                impl_size_eq!($nonzero, Unalign<$prim>);
                 let n = n.transmute::<Unalign<$prim>, invariant::Valid, _>();
                 $nonzero::new(n.read_unaligned().into_inner()).is_some()
             });
@@ -430,7 +404,7 @@ mod atomics {
         ($($($tyvar:ident)? => $atomic:ty [$prim:ty]),*) => {
             const _: () = {
                 use core::{cell::UnsafeCell, ptr::NonNull};
-                use crate::pointer::{TransmuteFrom, SizeEq, invariant::Valid};
+                use crate::pointer::{TransmuteFrom, SizeCompat, invariant::Valid};
 
                 $(
                     #[allow(unused_unsafe)] // Force the caller to call this macro inside `safety_comment!`.
@@ -443,36 +417,38 @@ mod atomics {
                     // the same size and bit validity.
                     unsafe impl<$($tyvar)?> TransmuteFrom<$prim, Valid, Valid> for $atomic {}
 
-                    // SAFETY: THe caller promised that `$atomic` and `$prim`
-                    // have the same size.
-                    unsafe impl<$($tyvar)?> SizeEq<$atomic> for $prim {
+                    // SAFETY: The caller promised that `$atomic` and `$prim`
+                    // have the same size. Thus, this cast preserves address,
+                    // referent size, and provenance.
+                    unsafe impl<$($tyvar)?> SizeCompat<$atomic> for $prim {
                         fn cast_from_raw(a: NonNull<$atomic>) -> NonNull<$prim> {
                             cast!(a)
                         }
                     }
-                    // SAFETY: THe caller promised that `$atomic` and `$prim`
-                    // have the same size.
-                    unsafe impl<$($tyvar)?> SizeEq<$prim> for $atomic {
+                    // SAFETY: See previous safety comment.
+                    unsafe impl<$($tyvar)?> SizeCompat<$prim> for $atomic {
                         fn cast_from_raw(p: NonNull<$prim>) -> NonNull<$atomic> {
                             cast!(p)
                         }
                     }
+
                     // SAFETY: The caller promised that `$atomic` and `$prim`
                     // have the same size. `UnsafeCell<T>` has the same size as
-                    // `T` [1].
+                    // `T` [1]. Thus, this cast preserves address, referent
+                    // size, and provenance.
                     //
                     // [1] Per https://doc.rust-lang.org/1.85.0/std/cell/struct.UnsafeCell.html#memory-layout:
                     //
                     //   `UnsafeCell<T>` has the same in-memory representation as
                     //   its inner type `T`. A consequence of this guarantee is that
                     //   it is possible to convert between `T` and `UnsafeCell<T>`.
-                    unsafe impl<$($tyvar)?> SizeEq<$atomic> for UnsafeCell<$prim> {
+                    unsafe impl<$($tyvar)?> SizeCompat<$atomic> for UnsafeCell<$prim> {
                         fn cast_from_raw(a: NonNull<$atomic>) -> NonNull<UnsafeCell<$prim>> {
                             cast!(a)
                         }
                     }
                     // SAFETY: See previous safety comment.
-                    unsafe impl<$($tyvar)?> SizeEq<UnsafeCell<$prim>> for $atomic {
+                    unsafe impl<$($tyvar)?> SizeCompat<UnsafeCell<$prim>> for $atomic {
                         fn cast_from_raw(p: NonNull<UnsafeCell<$prim>>) -> NonNull<$atomic> {
                             cast!(p)
                         }

src/layout.rs

@@ -92,6 +92,136 @@ pub(crate) enum MetadataCastError {
     Size,
 }
 
+/// Parameters necessary to perform an infallible reference cast on slice DSTs.
+///
+/// # Safety
+///
+/// Given `Src: KnownLayout` and `Dst: KnownLayout`, it is only possible to
+/// produce `CastParams` via
+/// `Src::LAYOUT.try_compute_cast_params(&Dst::LAYOUT)`, and that method will
+/// only return `Some` if it is possible to perform an infallible reference cast
+/// from `&Src` to `&Dst`.
+pub(crate) struct CastParams {
+    /// Given `sl: TrailingSliceLayout` (source layout) and `dl:
+    /// TrailingSliceLayout` (destination layout), this is computed as:
+    ///
+    /// `(sl.offset - dl.offset) / dl.elem_size`
+    ///
+    /// INVARIANT: `CastParams` can only be constructed if `sl.offset -
+    /// dl.offset` is an integer multiple of `dl.elem_size`.
+    offset_delta_elems: usize,
+    /// `sl.elem_size / dl.elem_size`
+    ///
+    /// INVARIANT: `CastParams` can only be constructed if `sl.elem_size` is an
+    /// integer multiple of `dl.elem_size`, and if `dl.elem_size > 0`.
+    elem_multiple: usize,
+}
+
+impl CastParams {
+    /// Given the metadata from `src: &Src`, computes the metadata required in
+    /// order for `dst: &Dst` with the same address to reference the same number
+    /// of bytes as `src`.
+    ///
+    /// # Safety
+    ///
+    /// If `Src: KnownLayout` and `Dst: KnownLayout`, and if `self` was computed
+    /// as `Src::LAYOUT.try_compute_cast_params(&Dst::LAYOUT)`, then the
+    /// following is guaranteed:
+    ///
+    /// Let `src_meta` be the pointer metadata from a `src: &Src`, and compute
+    /// `let dst_meta = self.compute_cast(src_meta)`. If `dst: &Dst` is
+    /// constructed using the address of `src` and using `dst_meta` as its
+    /// pointer metadata, then `dst` will address the same bytes as `src`.
+    ///
+    /// TODO: Mention that post-padding may not be preserved.
+    ///
+    /// # Panics
+    ///
+    /// If the safety preconditions do not hold, then `compute_cast` may panic
+    /// due to arithmetic overflow. Note that, as `compute_cast` is a safe
+    /// function, failing to uphold the safety preconditions cannot cause
+    /// immediate undefined behavior.
+    pub(crate) const fn compute_cast(self, src_meta: usize) -> usize {
+        // PANICS: This function is permitted to panic if its safety
+        // preconditions are not upheld.
+        //
+        // SAFETY/PANICS: If `Src: KnownLayout`, `Dst: KnownLayout`, and `self`
+        // was computed as `Src::LAYOUT.try_compute_cast_params(&Dst::LAYOUT)`,
+        // then `self` is a witness to the infallibility of casts from `&Src` to
+        // `&Dst`, and the internal invariants hold with respect to `Src` and
+        // `Dst`. Given `sl: Src::LAYOUT` and `dl: Dst::LAYOUT`, these
+        // invariants guarantee that:
+        //
+        // (1) `self.offset_delta_elems = (sl.offset - dl.offset) /
+        //     dl.elem_size`
+        //
+        // (2) `self.elem_multiple = sl.elem_size / dl.elem_size` where
+        //     `sl.elem_size` is an integer multiple of `dl.elem_size`
+        //
+        // If the safety preconditions are upheld, then the caller has promised
+        // that `src_meta` is the pointer metadata from some `src: &Src`. By
+        // invariant on `&Src`, `src` cannot address more than `isize::MAX`
+        // bytes [1]. Thus:
+        //
+        // (3)  `sl.offset + (sl.elem_size * src_meta) = src_len <= isize::MAX`
+        //
+        // (4)  `self.offset_delta_elems * dl.elem_size = sl.offset - dl.offset`
+        //      (rearranging (1))
+        //
+        // (5)  `sl.offset = (self.offset_delta_elems * dl.elem_size) +
+        //      dl.offset`
+        //
+        // (6)  `(self.offset_delta_elems * dl.elem_size) + dl.offset +
+        //      (sl.elem_size * src_meta) = src_len <= isize::MAX` (substituting
+        //      (5) into (3))
+        //
+        // (7)  `sl.elem_size = self.elem_multiple * dl.elem_size` (rearranging
+        //      (2))
+        //
+        // (8)  `(self.offset_delta_elems * dl.elem_size) + dl.offset +
+        //      (self.elem_multiple * dl.elem_size * src_meta) = src_len <=
+        //      isize::MAX` (substituting (7) into (6))
+        //
+        // (9)  `(self.offset_delta_elems * dl.elem_size) + (self.elem_multiple
+        //      * dl.elem_size * src_meta) = src_len - dl.offset <= isize::MAX -
+        //      dl.offset`
+        //
+        // (10) `self.offset_delta_elems + (self.elem_multiple * src_meta) =
+        //      (src_len - dl.offset) / dl.elem_size <= (isize::MAX - dl.offset)
+        //      / dl.elem_size`
+        //
+        // The left-hand side of (10) is the expression we compute and return
+        // from this method. Since `dl.offset >= 0` and `dl.elem_size > 1`, the
+        // right-hand side is not greater than `usize::MAX`, and thus this
+        // computation will not overflow, and thus it will not panic.
+        //
+        // We further need to prove that, if the caller takes this resulting
+        // value as `dst_meta` and uses it to synthesize a `dst &Dst`, this will
+        // be address the same number of bytes as `src`. The number of bytes is
+        // given by:
+        //
+        // (11) `dst_len = dl.offset + (dl.elem_size * dst_meta)`
+        //
+        // (12) `dst_len = dl.offset + (dl.elem_size * (self.offset_delta_elems
+        //      + (self.elem_multiple * src_meta)))` (substituing this method's
+        //      return value)
+        //
+        // (13) `dst_len = dl.offset + (dl.elem_size * ((src_len - dl.offset) /
+        //      dl.elem_size))` (substituting (10))
+        //
+        // (14) `dst_len = dl.offset + (src_len - dl.offset)`
+        //
+        // (15) `dst_len = src_len`
+        //
+        // [1] TODO
+        //
+        // TODO: Clarify that this algorithm is agnostic to post-padding, and
+        // justify that this is acceptable.
+        #[allow(clippy::arithmetic_side_effects)]
+        return self.offset_delta_elems + (self.elem_multiple * src_meta);
+    }
+}
+
 impl DstLayout {
     /// The minimum possible alignment of a type.
     const MIN_ALIGN: NonZeroUsize = match NonZeroUsize::new(1) {
@@ -202,6 +332,48 @@ impl DstLayout {
         }
     }
 
+    pub(crate) const fn try_compute_cast_params(self, other: &DstLayout) -> Option<CastParams> {
+        let (slf, other) = if let (SizeInfo::SliceDst(slf), SizeInfo::SliceDst(other)) =
+            (self.size_info, other.size_info)
+        {
+            (slf, other)
+        } else {
+            return None;
+        };
+
+        let offset_delta = if let Some(od) = slf.offset.checked_sub(other.offset) {
+            od
+        } else {
+            return None;
+        };
+
+        let delta_mod_other_elem = offset_delta.checked_rem(other.elem_size);
+        let elem_remainder = slf.elem_size.checked_rem(other.elem_size);
+
+        let (delta_mod_other_elem, elem_remainder) =
+            if let (Some(dmoe), Some(em)) = (delta_mod_other_elem, elem_remainder) {
+                (dmoe, em)
+            } else {
+                return None;
+            };
+
+        if delta_mod_other_elem != 0 || slf.elem_size < other.elem_size || elem_remainder != 0 {
+            return None;
+        }
+
+        // PANICS: The preceding `slf.elem_size.checked_rem(other.elem_size)`
+        // could only return `Some` if `other.elem_size > 0`, so this division
+        // will not divide by zero.
+        #[allow(clippy::arithmetic_side_effects)]
+        let offset_delta_elems = offset_delta / other.elem_size;
+        // PANICS: See previous panics comment.
+        #[allow(clippy::arithmetic_side_effects)]
+        let elem_multiple = slf.elem_size / other.elem_size;
+
+        // INVARIANTS: TODO
+        Some(CastParams { offset_delta_elems, elem_multiple })
+    }
+
     /// Like `Layout::extend`, this creates a layout that describes a record
     /// whose layout consists of `self` followed by `next` that includes the
     /// necessary inter-field padding, but not any trailing padding.