Stream: compiler development

Topic: glue field sorting

Dan G Knutson (Oct 18 2024 at 00:27):

Hi! I'm working with Luke on his roc-ray platform, and ran into either a bug or (more likely) a thing I don't get about roc glue.
If I run this command:
roc glue ~/path/to/roc/crates/glue/src/RustGlue.roc glue/out glue/glue.roc
where glue/glue.rochas this in it:

ThingToCheck : {
    timestampMillis : U64,
    frameCount : U64,
    keys : List U64,
    mouseButtons : List U64,
    mousePosX : F32,
    mousePosY : F32,
}

in the output, there's different versions of ThingToCheck as a rust struct for x86, x86_64, wasm32, etc. These different structs have different orders to their fields, which I understand matters because they're repr(C). For example:

#[derive(Clone, Default, Debug, PartialEq, PartialOrd, )]
#[repr(C)]
pub struct ThingToCheck {
    pub frameCount: u64,
    pub keys: roc_std::RocList<u64>,
    pub mouseButtons: roc_std::RocList<u64>,
    pub mousePosX: f32,
    pub mousePosY: f32,
    pub timestampMillis: u64,
}

// x86_64
#[derive(Clone, Default, Debug, PartialEq, PartialOrd, )]
#[repr(C)]
pub struct ThingToCheck {
    pub frameCount: u64,
    pub keys: roc_std::RocList<u64>,
    pub mouseButtons: roc_std::RocList<u64>,
    pub timestampMillis: u64,
    pub mousePosX: f32,
    pub mousePosY: f32,
}

Is that expected, and should our (currently handwritten not generated) host be using cfg(target_arch = "whatever") to handle this kind of thing?

Brendan Hansknecht (Oct 18 2024 at 00:40):

The first one is wasm?

Brendan Hansknecht (Oct 18 2024 at 00:41):

Or I guess x86_32.

Brendan Hansknecht (Oct 18 2024 at 00:41):

Anyway, I think it is correct

Brendan Hansknecht (Oct 18 2024 at 00:41):

Roc sorts by alignment first then by name

Brendan Hansknecht (Oct 18 2024 at 00:42):

On x86_32 and wasm32, u64 has an alignment of 32 bits

Brendan Hansknecht (Oct 18 2024 at 00:42):

So on those targets all fields have the same alignment. Thus just alphabetical order

Brendan Hansknecht (Oct 18 2024 at 00:43):

On x86_64, they have u64 has 64bit alignment. So does List. So they go before the f32s

Dan G Knutson (Oct 18 2024 at 00:45):

I'm not familiar with having to think about alignment at all. Am I right to think that this means using a single handwritten version (without cfg(target_arch)) is probably going to end up broken across different compilation targets? Like the data roc gets out of the bytes passed from rust would just be arbitrarily incorrect?

Dan G Knutson (Oct 18 2024 at 00:48):

For example, given enough variety in fields, if it was sorted correctly for aarch64 it would necessarily be sorted incorrectly for an x86_64 machine.

Luke Boswell (Oct 18 2024 at 00:50):

using a single handwritten version (without cfg(target_arch)) is probably going to end up broken across different compilation targets?

Yes

Luke Boswell (Oct 18 2024 at 00:50):

aarch64 it would necessarily be sorted incorrectly for an x86_64 machine.

I think these are the same though, at least that's why i just hand rolled it

Luke Boswell (Oct 18 2024 at 00:51):

I think at some point we talked about changing roc to just sort by alphabetical. Is there a reason we wouldn't do this?

Dan G Knutson (Oct 18 2024 at 00:51):

ah, gotcha; it only really matters for 32 vs 64, which in practice is likely to mean wasm?

Luke Boswell (Oct 18 2024 at 00:52):

That's my poor mans understanding

Brendan Hansknecht (Oct 18 2024 at 01:05):

Yeah, that is a good poor mans understanding. Sometimes there can be weird exceptions, but that almost always is the difference.

Dan G Knutson (Oct 18 2024 at 01:21):

thanks! that's reassuring.

Last updated: Jul 06 2025 at 12:14 UTC