Why is zig passing a @Vector(4, i32) on the stack in my local builds?

[fervent slate]

I'm compiling this code locally and emitting the assembly:

pub fn bar(foo: @Vector(4, i32)) @Vector(4, i32) {
    return foo + @as(@Vector(4, i32), .{ 1, 2, 3, 4 });
}```
LLVM generates the following:
```asm
.LCPI1_0:
    .long    1
    .long    2
    .long    3
    .long    4
    .text
    .p2align    4, 0x90
main.bar:
    push    rbp
    mov    rbp, rsp
    vmovdqa    xmm0, xmmword ptr [rcx]
    vpaddd    xmm0, xmm0, xmmword ptr [rip + .LCPI1_0]
    pop    rbp
    ret

However, godbolt (https://godbolt.org/z/v438xWePT) generates the following, which passes the parameter in xmm0 instead of the stack:

.LCPI0_0:
        .long   1
        .long   2
        .long   3
        .long   4
example.bar:
        push    rbp
        mov     rbp, rsp
        vpaddd  xmm0, xmm0, xmmword ptr [rip + .LCPI0_0]
        pop     rbp
        ret

Why is godbolt's output different?
If relevant, here is my build.zig: https://zigbin.io/3b91f9
The build command I run locally is zig build asm -Doptimize=ReleaseFast

[fossil cloud]

what OS are you on?

[fervent slate]

windows

are u srs

[fossil cloud]

https://learn.microsoft.com/en-us/cpp/build/x64-calling-convention

Any argument that doesn't fit in 8 bytes, or isn't 1, 2, 4, or 8 bytes, must be passed by reference.

you can use vectorcall on windows to make it pass the args in xmm0

[fervent slate]

this is so cooked </3

i will shrimply compile for linux instead

wait, i thought the default calling convention for non export functions was whatever deemed to be the fastest?

[fossil cloud]

in my experience it kinda just uses callconv(.c)

tbh it should maybe use vectorcall on windows

or even sysv

since windows calling convention is so bad

sysv on windows would mean you have to do some possibly expensive stack fixup on boundaries between zig functions and extern/export functions