as an additional note, this still fails:

#![feature(repr_simd, intrinsics)]

extern "rust-intrinsic" {
    fn simd_as<T, U>(x: T) -> U;
}

#[derive(Copy, Clone)]
#[repr(simd)]
struct V<T>([T; 2]);

fn main() {
    unsafe {
        let f = V::<f32>([f32::MIN, f32::MAX]);
        let i: V<i16> = simd_as(f);
        assert_eq!(i.0[0], f.0[0] as i16);
        assert_eq!(i.0[1], f.0[1] as i16);
    }
}

The i vector will just be [0, 0]. It looks like convert_vector, which is used for this conversion internally, does not match the LLVM semantics

https://github.com/rust-lang/gccjit.rs/blob/ac3ffb0b698ea24c9bbf32e3565fd72c0a1e79b7/src/context.rs#L925-L938

just to be clear: this does not use fptoint_sat currently (even though its code looks like it should work on vectors).

The tests/ui/numbers-arithmetic/saturating-float-casts.rs test now succeeds.

In this case, please remove this test from here.

After that, I'll merge.

Thanks.

Thanks for your contribution!