codegen regression for bool

[Nugine]

Code

I tried this code:

https://rust.godbolt.org/z/3EW6zoffE

pub fn all_zero(data: &[u64]) -> bool {
    data.iter().copied().fold(true, |acc, x| acc & (x == 0))
}

I expected to see this happen:

.LBB0_3:
        cmp     qword ptr [rdi + rcx], 0
        sete    dl
        and     al, dl
        add     rcx, 8
        cmp     rsi, rcx
        jne     .LBB0_3

Instead, this happened:

.LBB0_3:
        test    al, al
        setne   dl
        cmp     qword ptr [rdi + rcx], 0
        sete    al
        and     al, dl
        add     rcx, 8
        cmp     rsi, rcx
        jne     .LBB0_3

There are two unnecessary instructions:

        test    al, al
        setne   dl

Version it worked on

It most recently worked on: Rust 1.63, Rust beta

Version with regression

rustc --version --verbose:

rustc 1.65.0-nightly (748038961 2022-08-25)
binary: rustc
commit-hash: 7480389611f9d04bd34adf41a2b3029be4eb815e
commit-date: 2022-08-25
host: x86_64-unknown-linux-gnu
release: 1.65.0-nightly
LLVM version: 15.0.0

@rustbot modify labels: +regression-from-stable-to-nightly -regression-untriaged

[Nugine]

Related (maybe):

• Codegen weirdness for sum of count_ones over an array #101060
• Regression: Inlined loops not fully evaluating with const-eable dependencies #101082

[alion02]

The "canonical" code produces identical (and, since it has early exit, much more performant in most cases) assemblies between versions.

pub fn all_zero(data: &[u64]) -> bool {
    data.iter().all(|&x| x == 0)
}

[Nugine]

The "canonical" code produces identical (and, since it has early exit, much more performant in most cases) assemblies between versions.

pub fn all_zero(data: &[u64]) -> bool {
    data.iter().all(|&x| x == 0)
}

But constant-time functions can not early exit.
The "fold bool" variant should also be optimized when compiled with nighly. So I think it's regression.

[nikic]

Upstream issue: llvm/llvm-project#57448

[pnkfelix]

Aside: I think our labels here are a bit too general; while the text for "regression-from-stable-to-nightly" label itself says "Performance or correctness regression from stable to nightly", I think it should probably be revised to be restricted to "correctness or severe performance regression ..."

@rustbot label: -regression-from-stable-to-nightly

[pnkfelix]

I'm putting back the label for now; I want to get more consensus amongst the team about what the labels should be used for and what they denote (and how to get a narrower view when one wants to filter out cases like this)

@rustbot label: +regression-from-stable-to-nightly

[nikic]

Fixed upstream and will be pulled in with the next LLVM major version upgrade. I'm downgrading the priority as there's no point in revisiting this in triage until that time.

[nikic]

Fixed by the LLVM 16 upgrade.