Compiletime regression for `fast_transpose` crate

[awxkee]

Building this crate on stable takes 2.86s and on the latest nightly 1m 56s that is ~x100 slower. Tested on both macOS and Windows, results are the same. Previous nightly versions (for example nightly-2025-06-01, nightly-2025-07-01, nightly-2025-08-01) build the project in the expected time.

Code

Create new project, add crate cargo add fast_transpose.

use fast_transpose::{FlipMode, FlopMode};

fn main() {
    let src = vec![0u8; 1];
    let mut dst = vec![0u8; 1];
    fast_transpose::transpose_rgb(&src, 1, &mut dst, 1, 1, 1, FlipMode::Flip, FlopMode::NoFlop)
        .unwrap();
}

And try to build in stable

cargo clean && cargo +stable build --release

and in nightly

cargo clean && cargo +nightly build --release

Then see the difference in building time.

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

[lqd]

@nikic this is another slowness/"hang" from the LLVM 21 update

[nikic]

Stack trace looks similar to llvm/llvm-project#156114.

[nikic]

With llvm/llvm-project#158652 this goes down from 2m 40s with nightly to 12.3s with stage1, which is still larger than 3.9s with stable. But this is an assertion-enabled build without all the usual optimizations, so not sure how much the remaining difference really is.

[apiraino]

Assigning P-critical to keep this under the radar and verifying with a newer rustc release build (discussion on Zulip).

@rustbot label -I-prioritize +P-critical