Audio artefacts

[Uriopass]

This is the official issue for the audio artefacts I had in #1129
I tried to reproduce the problem, however I hit an incredibly weird bug.

When compiling this snippet alone

use rodio::Source;

fn main() {
    let (a, b) = rodio::OutputStream::try_default().unwrap();
    let buf = std::fs::read("music1.ogg").ok().unwrap();

    let s = rodio::Decoder::new(std::io::Cursor::new(buf)).unwrap();
    b.play_raw(s.convert_samples()).unwrap();
    loop {}
}

everything works fine and the music sounds good.

However when I compile the exact same main with the exact same audio file in my big project with lots of dependencies, there are hugs artefacts which look like this in audacity: Top is what is output and bottom is expected.

I tried deleting the target folder and rebuilding and I still got the artefacts.

Then I found that by removing theses directives about optimizing in my workspace Cargo.toml

[profile.dev.build-override]
opt-level = 1

[profile.dev.package."*"]
opt-level = 2

[profile.dev.package.common]
opt-level = 2

[profile.dev.package.egregoria]
opt-level = 0

[profile.dev.package.map_model]
opt-level = 0

[profile.dev.package.native_app] # <- main binary
opt-level = 0

[profile.dev.package.imgui-inspect]
opt-level = 1

[profile.dev.package.imgui-inspect-derive]
opt-level = 1

the audio artefacts were gone !

However, when reenabling

[profile.dev.package."*"]
opt-level = 2

the artefacts came back.

So I started removing dependencies, and I found that, basically, enabling optimizations on rodio using cg_clif gave thoses artefacts..

You can find a repo to reproduce here.
https://github.com/Uriopass/cranelift_bug

EDIT:
Now that I think about it, you still have to dive a bit further, it's not like there's no dependencies. I'll try on cpal examples.

[bjorn3]

So I started removing dependencies, and I found that, basically, enabling optimizations on rodio using cg_clif gave thoses artefacts.

Thanks for the investigation! This must be a miscompilation caused by one of the optimizations then. I will see if I can narrow down which optimization is responsible.

[bjorn3]

By the way, which license is the repro?

[Uriopass]

Interestingly, when I use a simple sin wave as a source, I get normal sounding results! I wonder if it comes from the vorbis decoder, I'll try the mp3 decoder.

[Uriopass]

Added the Unlicense LICENCE :-)

[bjorn3]

Thanks. I won't have to worry about adding a test derived from it to cg_clif then.

[Uriopass]

I confirm that using the mp3 decoder gives normal result. It probably comes from the vorbis decoder used by rodio: lewton.
https://github.com/RustAudio/lewton

[Uriopass]

Adding

[profile.dev.package.lewton]
opt-level = 0

does solve the artefacts issues
and

[profile.dev.package.lewton]
opt-level = 1

still gives artefacts.

Also, generating a sin wave ogg file gives normal result (maybe it compresses well and does not go through a miscompiled zone ?)

[bjorn3]

I am able to reproduce the issue. Disabling my stack2reg optimization fixes it. I already kind of suspected this to be the culprit, as I wrote it with absolutely zero knowledge about how to write this kind of optimization. I will see if I can fix it and otherwise I will just disable it completely.

[bjorn3]

Disabling just the the dead store removal of that optimization fixes it too.

[bjorn3]

I have disabled stack2reg. Maybe I will try to fix it in the future.