`Cow<[T]>` layout forces unnecessary branching

I tried this code (Godbolt):

pub fn test<'a>(cow: &'a std::borrow::Cow<'a, [u8]>) -> &'a [u8] {
    &*cow
}

I expected to see this happen: The generated assembly is branchless. The pointer and length are at the same offset in both variants of the Cow.

Instead, this happened: There is a branch in the generated assembly. The pointer to the slice is at a different offset in the Owned and Borrowed variants of the Cow.

The ideal layout for Cow<[T]> (and Cow<str>) would look like this:

If `T` is not a ZST:
┌──────────┬──────────┬────────────────────────────────────────────────────┐
│ pointer  │ length   │ capacity if `Owned`, or `usize::MAX` if `Borrowed` │
└──────────┴──────────┴────────────────────────────────────────────────────┘

If `T` is a ZST:
┌──────────┬──────────┬────────────────────────────────────────┐
│ pointer  │ length   │ boolean flag for `Owned` vs `Borrowed` │
└──────────┴──────────┴────────────────────────────────────────┘

The present layout looks like this:

`Owned` variant:
┌──────────┬──────────┬──────────┐
│ pointer  │ capacity │ length   │
└──────────┴──────────┴──────────┘

`Borrowed` variant:
┌──────────┬──────────┬──────────┐
│ 0x0      │ pointer  │ length   │
└──────────┴──────────┴──────────┘

Because of this non-optimal layout, every access to the contained pointer requires a branch on the enum variant.

(Fixing #45431 is a prerequisite to fixing this)

Meta

rustc --version --verbose:

rustc 1.75.0-nightly (189d6c71f 2023-11-06)
binary: rustc
commit-hash: 189d6c71f3bb6c52113b5639a80839791974fd22
commit-date: 2023-11-06
host: x86_64-unknown-linux-gnu
release: 1.75.0-nightly
LLVM version: 17.0.4

@rustbot label A-layout I-heavy I-slow T-compiler T-libs

Start by reading prerequisite issue #45431 and running the linked Godbolt example. Use the stated target layouts for non-ZST and ZST types, and branchless access to the contained pointer, as the completion criteria.