Missed optimization with slice::get_unchecked followed by <slice as Index>::index at a lower index

[zachs18]

I tried this code:

pub unsafe fn cannot_elide_bounds_check(s: &[u8]) -> u8 {
    let a = *s.get_unchecked(1);
    a + s[0]
}
pub unsafe fn can_elide_bounds_check(s: &[u8]) -> u8 {
    let a = *s.get(1).unwrap_unchecked();
    a + s[0]
}

Compile with -Copt-level=3 on godbolt.org (godbolt link)

I expected to see this happen: Both functions should compile to approximately the same thing, with no bounds check on the s[0] since the optimizer should be able to see that the slice is at least of length 2 due to the previous unchecked indexing.

example::can_elide_bounds_check:
        movzx   eax, byte ptr [rdi]
        add     al, byte ptr [rdi + 1]
        ret

Instead, this happened: cannot_elide_bounds_check contains a bounds check (only can_elide_bounds_check does not contain a bounds check).

example::cannot_elide_bounds_check:
        test    rsi, rsi
        je      .LBB0_2
        movzx   eax, byte ptr [rdi]
        add     al, byte ptr [rdi + 1]
        ret
.LBB0_2:
        push    rax
        lea     rdx, [rip + .L__unnamed_1]
        xor     edi, edi
        xor     esi, esi
        call    qword ptr [rip + core::panicking::panic_bounds_check@GOTPCREL]
        ud2

Meta

rustc stable (1.73.0), nightly (2023-10-17), and 1.58.0 (when Option::unwrap_unchecked was stabilized) all have approximately the same codegen.
rustc --version --verbose:

rustc 1.75.0-nightly (09df6108c 2023-10-17)
binary: rustc
commit-hash: 09df6108c84fdec400043d99d9ee232336fd5a9f
commit-date: 2023-10-17
host: x86_64-unknown-linux-gnu
release: 1.75.0-nightly
LLVM version: 17.0.2

(no backtrace)

@rustbot label +C-optimization -C-bug

[the8472]

unwrap_unchecked contains an assume intrinsic, get_unchecked doesn't. Might be worth adding it to the latter and see how that impacts perf, it's not guaranteed that it would actually improve things since we'd be emitting more IR and llvm would have to work harder.

Commonly it's the case that if one uses get_unchecked one does so for all accesses on the slice, so this seems like an unusual pattern.