Performance regression with get_many_mut()

[ChayimFriedman2]

In the process of trying to stabilize get_many_mut(), and while trying to check what will be the costs of a more detailed error type, I came across the following perf regression from stable to beta.

Code

I tried this code:

#![feature(get_many_mut)]

#[inline]
fn get_many_check_valid(indices: &[usize; 2], len: usize) -> Result<(), GetManyMutError> {
    for (i, &idx) in indices.iter().enumerate() {
        if idx >= len {
            return Err(GetManyMutError::OutOfBounds);
        }
        for &idx2 in &indices[..i] {
            if idx == idx2 {
                return Err(GetManyMutError::Overlapping);
            }
        }
    }
    Ok(())
}

#[inline]
pub fn after_all<'a>(
    slice: &'a mut [i32],
    indices: [usize; 2],
) -> Result<[&mut i32; 2], GetManyMutError> {
    get_many_check_valid(&indices, slice.len())?;
    unsafe { Ok(slice.get_many_unchecked_mut(indices)) }
}

#[derive(Debug)]
pub enum GetManyMutError {
    OutOfBounds,
    Overlapping,
}

#[no_mangle]
pub fn after_all_unwrap(data: &mut [i32], indices: [usize; 2]) -> [&mut i32; 2] {
    after_all(data, indices).unwrap()
}

On stable (with RUSTC_BOOTSTRAP=1) this optimizes nicely to the following code (https://godbolt.org/z/zq8EG7df7):

after_all_unwrap:
        mov     r8, qword ptr [rcx]
        xor     eax, eax
        cmp     r8, rdx
        jae     .LBB0_4
        mov     rcx, qword ptr [rcx + 8]
        cmp     rcx, rdx
        jae     .LBB0_4
        cmp     rcx, r8
        je      .LBB0_3
        lea     rax, [rsi + 4*r8]
        lea     rcx, [rsi + 4*rcx]
        mov     qword ptr [rdi], rax
        mov     qword ptr [rdi + 8], rcx
        mov     rax, rdi
        ret
.LBB0_3:
        mov     al, 1
.LBB0_4:
        push    rax
        mov     byte ptr [rsp + 7], al
        lea     rdi, [rip + .L__unnamed_1]
        lea     rcx, [rip + .L__unnamed_2]
        lea     r8, [rip + .L__unnamed_3]
        lea     rdx, [rsp + 7]
        mov     esi, 43
        call    qword ptr [rip + core::result::unwrap_failed::haccb9aaa604e1e21@GOTPCREL]

However, on beta (and nightly) this adds a whole bunch of unnecessary instructions:

after_all_unwrap:
        mov     r8, qword ptr [rcx]
        xor     eax, eax
        cmp     r8, rdx
        jae     .LBB0_4
        mov     rcx, qword ptr [rcx + 8]
        cmp     rcx, rdx
        jae     .LBB0_4
        cmp     rcx, r8
        je      .LBB0_3
        lea     rax, [rsi + 4*r8]
        lea     rcx, [rsi + 4*rcx]
        mov     rdx, rcx
        shr     rdx, 8
        mov     qword ptr [rdi], rax
        mov     byte ptr [rdi + 8], cl
        mov     rax, rcx
        shr     rax, 56
        mov     byte ptr [rdi + 15], al
        shr     rcx, 40
        mov     word ptr [rdi + 13], cx
        mov     dword ptr [rdi + 9], edx
        mov     rax, rdi
        ret
.LBB0_3:
        mov     al, 1
.LBB0_4:
        push    rax
        mov     byte ptr [rsp + 7], al
        lea     rdi, [rip + .L__unnamed_1]
        lea     rcx, [rip + .L__unnamed_2]
        lea     r8, [rip + .L__unnamed_3]
        lea     rdx, [rsp + 7]
        mov     esi, 43
        call    qword ptr [rip + core::result::unwrap_failed::h803c538fb296a320@GOTPCREL]

With cargo-bisect-rustc I bisected it to #126578. CC @scottmcm.

@rustbot label +regression-from-stable-to-beta -regression-untriaged -C-bug +perf-regression +I-heavy +I-slow

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

In #126578 the perf. triaging showed some regression in binary size (though not entirely due to that patchset). Probably this is another side effect.

@rustbot label -I-prioritize +P-medium

[ChayimFriedman2]

The root cause is a bug in LLVM: llvm/llvm-project#101899.