How to emit vunpck instruction

Help
,
1

Does anyone know if it is possible to have zig emit vunpck[hl][pd|ps] etc. instructions?

Everything i try is being compiled to two permutes and a blend

        push    rbp
        mov     rbp, rsp
        vpermpd ymm0, ymmword ptr [rsi], 96
        vpermpd ymm1, ymmword ptr [rdi], 212
        vblendps        ymm0, ymm1, ymm0, 204
        vmovups ymmword ptr [rdx], ymm0
        pop     rbp
        vzeroupper
        ret

Maybe i seeing thing wrong but thats 3 uops with a latency of 5 wheres a vunpck instruction would be 1 uop with latency 1.

I tried using @“llvm” intrinsics but in there i cannot find unpck instructions for avx2 (only mmx). Are they not implemented or something?

Also in doc/langref/destructuring_vectors.zig there is:

// emulate punpckldq
pub fn unpack(x: @Vector(4, f32), y: @Vector(4, f32)) @Vector(4, f32) {
    const a, const c, _, _ = x;
    const b, const d, _, _ = y;
    return .{ a, b, c, d };
}

but this gives me the aformentioned assembly output.

2

Try to use @shuffle. Not sure it always works, but this compiles to a vunpcklps instruction:

export fn unpack(x: @Vector(4, f32), y: @Vector(4, f32)) @Vector(4, f32) {
    return @shuffle(f32, x, y, [_]i32{ 0, -1, 1, -2 });
}
3

I’am sure i tried that as the first thing! Must have been sleeping at the time or so.
Thanks

(edit)
I see i was testing f64.
Then it still emits permpd’s

image
image781×288 32.3 KB