transmute size check is wrong for overaligned newtype

[theemathas]

I tried this code:

use std::mem::transmute;

#[repr(align(16))]
struct Wrap<T>(T);

fn foo<T: ?Sized>(x: Wrap<*const T>) -> *const T {
    unsafe { transmute(x) }
}

fn main() {
    foo(Wrap(std::ptr::null::<i32>()));
    println!("done");
}

The above code shouldn't compile, since Wrap<*const i32> and *const i32 don't have the same size, and transmute is supposed to check the sizes of the two types to be equal at compile time.

Instead, the code compiles and causes a segmentation fault at run time.

The generated assembly is particularly nonsensical:

example::foo::<i32>:
        mov     byte ptr [rax], 1
        ret

Possibly related to #101084. cc @RalfJung @oli-obk

Meta

Reproducible on the playground with version 1.97.0-nightly (2026-04-15 e8e4541ff19649d95afa)

[theemathas]

Possibly relevant code: https://github.com/rust-lang/rust/blob/main/compiler/rustc_middle/src/ty/layout.rs#L472

[zachs18]

The reason for the nonsensical assembly is that the generated MIR has unconditional UB, it's not an issue with codegen:

• std::mem::transmute lowers to CastKind::Transmute at the LowerIntrinsics Mir pass
• CastKind::Transmute where the src and dest have differing sizes is well-formed MIR but UB to execute
• CastKind::Transmute where the src and dest have differing sizes is lowered as bx.unreachable_nonterminator() (src)
• rustc_codegen_llvm's unreachable_nonterminator emits store i1 true, ptr poison, align 1 which is LLVM's "the preferred LLVM incantation" for non-basic-block-terminator unreachable
• store i1 true, ptr poison naturally lowers to movb [rax], 1 without optimizations

[theemathas]

Possibly related: #88290

[RalfJung]

Ah, fun. Looks like the logic for checking if two types have the same size forgets to check for repr(align) when allowing newtypes...

[RalfJung]

I began working on a fix: #155418