Returning a double from asm to c

I am looking to take two double variables, pass them to a function, subtract them, return them and print the result.

My example:

#include <stdio.h>
extern double dev(double a, double b);
int main() {
    double a = 10.2;
    double b = 3.6;
    double result = dev(a, b);
    printf("%.2f", result);
    return 0;
}

Asm:

section .text
    global dev
dev:
    movsd xmm0, [rdi]
    movsd xmm1, [rsi]
    subsd xmm0, xmm1
    ret

It compiles, however when I attempt to run the code I get a Segmentation fault (core dumped)

I believe I have identified the problem, I just am not sure how to fix it.

double is a 64 bit value. Whereas xmm is a 128 bit register. It would seem logical (at least to me) I am attempting to return 128 bits, however I see no way of doing it. I tried movsd rax xmm0 but same issue

Solution

In this case, one approach is to look at how the compiler implements a similar function: https://godbolt.org/z/zv8cY3GfG (no optimization) or https://godbolt.org/z/zn8ecKT1f (optimization level "O2").

The first one gives the following (making it apparent that the parameters are passed in registers xmm0 and xmm1):

dev:
        push    rbp
        mov     rbp, rsp
        movsd   QWORD PTR [rbp-8], xmm0
        movsd   QWORD PTR [rbp-16], xmm1
        movsd   xmm0, QWORD PTR [rbp-8]
        subsd   xmm0, QWORD PTR [rbp-16]
        pop     rbp
        ret

The optimized version is as follows:

dev:
        subsd   xmm0, xmm1
        ret

Which illustrates one reason why it perhaps is increasingly uncommon to see embedded assembly: compilers are getting increasingly good at creating optimized assembler code and assembler code is getting more and more optimized towards compilers making it difficult for humans to create good assembler code. The effort and restriction to a specific target CPU is seldomly worth it.