Guarantee slice representation

Author: cramertj

text/0000-guaranteed-slice-repr.md

@@ -0,0 +1,163 @@
+- Feature Name: guaranteed_slice_repr
+- Start Date: 2025-02-18
+- RFC PR: [rust-lang/rfcs#0000](https://github.com/rust-lang/rfcs/pull/0000)
+- Rust Issue: [rust-lang/rust#0000](https://github.com/rust-lang/rust/issues/0000)
+
+# Summary
+[summary]: #summary
+
+This RFC guarantees the in-memory representation of slice and str references.
+Specifically, `&[T]` is guaranteed to have the same layout as:
+
+```rust
+#[repr(C)]
+struct Slice<T> {
+    data: *const T,
+    len: usize,
+}
+```
+
+The layout of `&str` is the same as that of `&[u8]`, and the layout of
+`&mut str` is the same as that of `&mut [u8]`.
+
+# Motivation
+[motivation]: #motivation
+
+This RFC allows non-Rust (e.g. C or C++) code to read from or write to existing
+slices and to declare slice fields or locals.
+
+For example, guaranteeing the representation of slices allows non-Rust code to
+read from the `data` or `len` fields of `string` in the type below without
+intermediate FFI calls into Rust:
+
+```rust
+#[repr(C)]
+struct HasString {
+    string: &'static str,
+}
+```
+
+Note: prior to this RFC, the type above is not even properly `repr(C)` since the
+size and alignment of slices were not guaranteed. However, the Rust compiler
+accepts `repr(C)` declaration above without warning.
+
+# Guide-level explanation
+[guide-level-explanation]: #guide-level-explanation
+
+Slices are represented with a pointer and length pair. Their in-memory layout is
+the same as a `#[repr(C)]` struct like the following:
+
+```rust
+#[repr(C)]
+struct Slice<T> {
+    data: *const T,
+    len: usize,
+}
+```
+
+The validity requirements for the in-memory type are the same as [those
+documented on `std::slice::from_raw_parts`](https://doc.rust-lang.org/std/slice/fn.from_raw_parts.html).
+Namely:
+
+* `data` must be non-null, [valid] for reads for `len * mem::size_of::<T>()` many bytes,
+  and it must be properly aligned. This means in particular:
+
+    * The entire memory range of this slice must be contained within a single allocated object!
+      Slices can never span across multiple allocated objects. See [below](#incorrect-usage)
+      for an example incorrectly not taking this into account.
+    * `data` must be non-null and aligned even for zero-length slices or slices of ZSTs. One
+      reason for this is that enum layout optimizations may rely on references
+      (including slices of any length) being aligned and non-null to distinguish
+      them from other data. You can obtain a pointer that is usable as `data`
+      for zero-length slices using [`NonNull::dangling()`].
+
+* `data` must point to `len` consecutive properly initialized values of type `T`.
+
+* The memory referenced by the returned slice must not be mutated for the duration
+  of lifetime `'a`, except inside an `UnsafeCell`.
+
+* The total size `len * mem::size_of::<T>()` of the slice must be no larger than `isize::MAX`,
+  and adding that size to `data` must not "wrap around" the address space.
+  See the safety documentation of [`pointer::offset`].
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+## Zero-sized types
+
+One could imagine representing `&[T]` as only `len` for zero-sized `T`.
+This proposal would preclude that choice in favor of a standard representation
+for slices regardless of the underlying type.
+
+Alternatively, we could choose to guarantee that the data pointer is present if
+and only if `size_of::<T> != 0`. This has the possibility of breaking exising
+code which smuggles pointers through the `data` value in `from_raw_parts` /
+`into_raw_parts`.
+
+## Compatibility with C++ `std::span`
+
+The largest drawback of this layout and set of validity requirements is that it
+may preclude `&[T]` from being representationally equivalent to C++'s
+`std::span<T, std::dynamic_extent>`.
+
+* `std::span` does not currently guarantee its layout. In practice, pointer + length
+  is the common representation. This is even observable using `is_layout_compatible`
+  [on MSVC](https://godbolt.org/z/Y8ardrshY), though not
+  [on GCC](https://godbolt.org/z/s4v4xehnG) nor
+  [on Clang](https://godbolt.org/z/qsd1K5oGq). Future changes to guarantee a
+  different layout in the C++ standard (unlikely due to MSVC ABI stabilitiy
+  requirements) could preclude matching the layout with `&[T]`.
+
+* Unlike Rust, `std::span` allows the `data` pointer to be `nullptr`. One
+  possibile workaround for this would be to guarantee that `Option<&[T]>` uses
+  `data: std::ptr::null()` to represent the `None` case, making `std::span<T>`
+  equivalent to `Option<&[T]>` for non-zero-sized types.
+
+* Rust uses a dangling pointer in the representation of zero-length slices.
+  It's unclear whether 
+
+Note that C++ also does not support zero-sized types, so there is no naiive way
+to represent types like `std::span<SomeZeroSizedRustType>`.
+
+## Flexibility
+
+Additionally, guaranteeing layout of Rust-native types limits the compiler's and
+standard library's ability to change and take advantage of new optimization
+opportunities.
+
+# Rationale and alternatives
+[rationale-and-alternatives]: #rationale-and-alternatives
+
+* We could avoid committing to a particular representation for slices.
+
+* We could try to guarantee layout compatibility with a particular target's
+  `std::span` representation, though without standardization this may be
+  impossible. Multiple different C++ stdlib implementations may be used on
+  the same platform and could potentially have different span representations.
+  In practice, current span representations also use ptr+len pairs.
+
+* We could avoid storing a data pointer for zero-sized types. This would result
+  in a more compact representation but would mean that the representation of
+  `&[T]` is dependent on the type of `T`.
+
+# Prior art
+[prior-art]: #prior-art
+
+The layout in this RFC is already documented in
+[the Unsafe Code Guildelines Reference.](https://rust-lang.github.io/unsafe-code-guidelines/layout/pointers.html)
+
+# Unresolved questions
+[unresolved-questions]: #unresolved-questions
+
+* Should `&[T]` include a pointer when `T` is zero-sized?
+
+# Future possibilities
+[future-possibilities]: #future-possibilities
+
+* Consider defining a separate Rust type which is repr-equivalent to the platform's
+  native `std::span<T, std::dynamic_extent>` to allow for easier
+  interoperability with C++ APIs. Unfortunately, the C++ standard does not
+  guarantee the layout of `std::span` (though the representation may be known
+  and fixed on a particular implementation, e.g. libc++/libstdc++/MSVC).
+  Zero-sized types would also not be supported with a naiive implementation of
+  such a type.