Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 30 additions & 5 deletions library/core/src/mem/mod.rs
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,7 @@ use crate::clone::TrivialClone;
use crate::cmp::Ordering;
use crate::marker::{Destruct, DiscriminantKind};
use crate::panic::const_assert;
use crate::ub_checks::assert_unsafe_precondition;
use crate::{clone, cmp, fmt, hash, intrinsics, ptr};

mod alignment;
Expand Down Expand Up @@ -1077,6 +1078,18 @@ pub const fn copy<T: Copy>(x: &T) -> T {
///
/// [ub]: ../../reference/behavior-considered-undefined.html
///
/// If you have a raw pointer instead of a reference, you might be looking for
/// [`ptr::read_unaligned`]`::<Dst>(ptr.cast::<Dst>())` instead.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This seems like an odd syntax to recommend, with Dst being duplicated. Why not ptr.cast::<Dst>().read_unaligned()?

///
/// # Safety
///
/// - Requires `size_of_val::<Src>(src) >= size_of::<Dst>()`
/// - The first `size_of::<Dst>()` bytes behind `src` must be *readable*
/// - The first `size_of::<Dst>()` bytes behind `src` must be *valid*

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you make "valid" a link to this so that the term is clearly defined?

/// when interpreted as a `Dst`.
/// - All safety invariants of the `Dst` type must be upheld. (For example,

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"Safety invariants" is not a standard term (yet).

See this for how we word the same concern elsewhere.

/// `{ transmute_copy::<String, String>(&x); x; }` is UB for the double-drop.)
///
/// # Examples
///
/// ```
Expand All @@ -1101,20 +1114,32 @@ pub const fn copy<T: Copy>(x: &T) -> T {
///
/// // The contents of 'foo_array' should not have changed
/// assert_eq!(foo_array, [10]);
///
/// let bytes: &[u8] = &[1, 2, 3, 4, 5, 6, 7];
/// assert_eq!(
/// unsafe { mem::transmute_copy::<[u8], u32>(bytes) },
/// u32::from_ne_bytes(*bytes.first_chunk().unwrap()),
/// );
/// ```
#[inline]
#[must_use]
#[track_caller]
#[stable(feature = "rust1", since = "1.0.0")]
#[rustc_const_stable(feature = "const_transmute_copy", since = "1.74.0")]
pub const unsafe fn transmute_copy<Src, Dst>(src: &Src) -> Dst {
assert!(
size_of::<Src>() >= size_of::<Dst>(),
"cannot transmute_copy if Dst is larger than Src"
pub const unsafe fn transmute_copy<Src: ?Sized, Dst>(src: &Src) -> Dst {
// library UB because it's possible for the `Src` to be only a subset of the allocation
// and thus for a failure to not be immediate language UB
assert_unsafe_precondition!(
check_library_ub,
"cannot transmute_copy if Dst is larger than Src",
(
src_size: usize = size_of_val::<Src>(src),
dst_size: usize = Dst::SIZE,
) => src_size >= dst_size
);

// If Dst has a higher alignment requirement, src might not be suitably aligned.
if align_of::<Dst>() > align_of::<Src>() {
if align_of::<Dst>() > align_of_val::<Src>(src) {

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For unsized Src this comparison might not be optimized out. Is it even worth having both branches then?

@scottmcm scottmcm May 29, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hmm, interesting question. I guess I could condition it on is_val_statically_known...

(We can't specialize on Sized-ness, right?)

EDIT later: Note to self, include some codegen tests to demonstrate one way or the other.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@RalfJung For my own edification, can you say more about why this comparison might not be optimized out in the unsized case?

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Because align_of_val might not be a constant if the unsized tail is dyn Trait. So this compiles to if some_const > value_loaded_from_vtable ... and I don't see how LLVM could optimize this.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah I think that makes sense. What is the alternative you're suggesting? You mention not having both branches, but which branch are you suggesting be eliminated?

(To be clear, I am nearly certain that these questions are coming from a lack of understanding on my part.)

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Originally I meant to suggest that we should just always use read_unaligned to avoid codegen'ing two read paths.

But as @scottmcm mentioned there's an alternative: use is_val_statically_known.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good news, even at low opt-levels LLVM is smart and merges the loads or the memcpys so even with the branch in the rust code the right thing happens 🎉 Test added.

(I still have doc fixes to do, though, so this PR is not ready yet.)

// SAFETY: `src` is a reference which is guaranteed to be valid for reads.
// The caller must guarantee that the actual transmutation is safe.
unsafe { ptr::read_unaligned(src as *const Src as *const Dst) }
Expand Down
26 changes: 0 additions & 26 deletions library/coretests/tests/mem.rs
Original file line number Diff line number Diff line change
Expand Up @@ -138,32 +138,6 @@ fn test_transmute_copy_unaligned() {
assert_eq!(0_u64, unsafe { transmute_copy(&u.b) });
}

#[test]
#[cfg(panic = "unwind")]
fn test_transmute_copy_grow_panics() {
use std::panic;

let err = panic::catch_unwind(panic::AssertUnwindSafe(|| unsafe {
let _unused: u64 = transmute_copy(&1_u8);
}));

match err {
Ok(_) => unreachable!(),
Err(payload) => {
payload
.downcast::<&'static str>()
.and_then(|s| {
if *s == "cannot transmute_copy if Dst is larger than Src" {
Ok(s)
} else {
Err(s)
}
})
.unwrap_or_else(|p| panic::resume_unwind(p));
}
}
}

#[test]
#[allow(dead_code)]
fn test_discriminant_send_sync() {
Expand Down
65 changes: 65 additions & 0 deletions tests/codegen-llvm/lib-optimizations/transmute_copy.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,65 @@
//@ compile-flags: -Copt-level=1
//@ only-64bit

#![crate_type = "lib"]

// Check that we don't have runtime alignment checks, just a safe alignment.
// (Rust emits a branch, but LLVM merges the loads from the arms.)

use std::mem::transmute_copy;

// CHECK-LABEL: @transmute_copy_i16_from_i32_slice
// CHECK-SAME: ptr{{.+}}%_0,
// CHECK-SAME: ptr{{.+}}%x.0,
// CHECK-SAME: i64{{.+}}%x.1)
#[no_mangle]
pub unsafe fn transmute_copy_i16_from_i32_slice(x: &[i32]) -> [i16; 7] {
// CHECK: start:
// CHECK-NEXT: @llvm.memcpy
// CHECK-SAME: align 2{{.+}}%_0,
// CHECK-SAME: align 4{{.+}}%x.0,
// CHECK-SAME: i64 14,
// CHECK-NEXT: ret void
transmute_copy(x)
}

// CHECK-LABEL: @transmute_copy_i32_from_i16_slice
// CHECK-SAME: ptr{{.+}}%_0,
// CHECK-SAME: ptr{{.+}}%x.0,
// CHECK-SAME: i64{{.+}}%x.1)
#[no_mangle]
pub unsafe fn transmute_copy_i32_from_i16_slice(x: &[i16]) -> [i32; 3] {
// CHECK: start:
// CHECK-NEXT: @llvm.memcpy
// CHECK-SAME: align 4{{.+}}%_0,
// CHECK-SAME: align 2{{.+}}%x.0,
// CHECK-SAME: i64 12,
// CHECK-NEXT: ret void
transmute_copy(x)
}

// CHECK-LABEL: i32 @transmute_copy_i32_from_dyn
// CHECK-SAME: ptr{{.+}}%x.0,
// CHECK-SAME: ptr{{.+}}%x.1)
#[no_mangle]
pub unsafe fn transmute_copy_i32_from_dyn(x: &dyn std::fmt::Debug) -> i32 {
// CHECK: start:
// CHECK-NEXT: [[TEMP:%.+]] = load i32, ptr %x.0, align 1
// CHECK-NEXT: ret i32 [[TEMP]]
transmute_copy(x)
}

// CHECK-LABEL: @transmute_copy_i16_array_from_dyn
// CHECK-SAME: ptr{{.+}}%_0,
// CHECK-SAME: ptr{{.+}}%x.0,
// CHECK-SAME: ptr{{.+}}%x.1)
#[no_mangle]
pub unsafe fn transmute_copy_i16_array_from_dyn(x: &dyn std::fmt::Debug) -> [i16; 7] {
// CHECK: start:
// CHECK-NEXT: @llvm.memcpy
// CHECK-SAME: align 2{{.+}}%_0,
// CHECK-SAME: align 1{{.+}}%x.0,
// CHECK-SAME: i64 14,
// CHECK-NEXT: ret void
transmute_copy(x)
}
9 changes: 9 additions & 0 deletions tests/ui/precondition-checks/transmute_copy.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
//@ run-crash
//@ compile-flags: -Copt-level=3 -Cdebug-assertions=no -Zub-checks=yes
//@ error-pattern: unsafe precondition(s) violated: cannot transmute_copy if Dst is larger than Src

fn main() {
unsafe {
let _unused: u64 = std::mem::transmute_copy(&1_u8);
}
}
Loading