Floating-point casts (sometimes) don't work on AVR

[Patryk27]

This code:

#[atmega_hal::entry]
fn main() -> ! {
    let dp = Peripherals::take().unwrap();
    let pins = pins!(dp);

    let mut uart = Usart0::<MHz16>::new(
        dp.USART0,
        pins.pd0,
        pins.pd1.into_output(),
        115200u32.into_baudrate(),
    );

    let value: f32 = core::hint::black_box(1.0);
    let a = value as i64;

    ufmt::uwriteln!(&mut uart, "A: {:?}", a.to_le_bytes());

    loop {
        //
    }
}

... prints (under simavr):

A: [1, 0, 0, 0, 0, 0, 0, 0]

... but moving the code into a loop:

loop {
    let value: f32 = core::hint::black_box(1.0);
    let a = value as i64;

    ufmt::uwriteln!(&mut uart, "A: {:?}", a.to_le_bytes());
}

... breaks something:

A: [0, 0, 0, 128, 0, 0, 0, 128]

I'll try to investigate what's going on, I just thought I'll create an issue in case someone else stumbles upon this one as well.

[Patryk27]

Heh, ABI mismatch strikes again 😄

In this case it's caused by an intrinsic called __gtsf2() - the one from compiler-builtins takes priority and it's got a bit different ABI than the expected intrinsic from libgcc; lemme do some more tests and we'll probably #[avr_skip] those in compiler-builtins.

[Patryk27]

Yeah; as a hot-fix, dropping this to .cargo/config should do it:

build-std-features = ["compiler-builtins-mangled-names"]

[Patryk27]

(note that we still need to bump compiler-builtins in rustc for the patch to apply here, so it's actually semi-resolved at the moment)