Universal function proxy

blah.S

.section .text, "x", discard, proxy_fn
.balign 16
.globl proxy_fn
.def proxy_fn; .scl 2; .type 32; .endef
proxy_fn:
.seh_proc proxy_fn
    sub $0x68, %rsp
    .seh_stackalloc 0x68
    .seh_endprologue
    mov %rcx, 0x70(%rsp)
    mov %rdx, 0x78(%rsp)
    mov %r8, 0x80(%rsp)
    mov %r9, 0x88(%rsp)
    movups %xmm0, 0x50(%rsp)
    movups %xmm1, 0x40(%rsp)
    movups %xmm2, 0x30(%rsp)
    movups %xmm3, 0x20(%rsp)
    call my_fn
    mov 0x70(%rsp), %rcx
    mov 0x78(%rsp), %rdx
    mov 0x80(%rsp), %r8
    mov 0x88(%rsp), %r9
    movups 0x50(%rsp), %xmm0
    movups 0x40(%rsp), %xmm1
    movups 0x30(%rsp), %xmm2
    movups 0x20(%rsp), %xmm3
    add $0x68, %rsp
    jmp *%rax
.seh_endproc

CMakeLists.txt

cmake_minimum_required(VERSION 3.28)
project(untitled C ASM)

set(CMAKE_CXX_STANDARD 17)

set(CMAKE_ASM_COMPILE_OPTIONS_MSVC_RUNTIME_LIBRARY_MultiThreaded "")
set(CMAKE_ASM_COMPILE_OPTIONS_MSVC_RUNTIME_LIBRARY_MultiThreadedDLL "")
set(CMAKE_ASM_COMPILE_OPTIONS_MSVC_RUNTIME_LIBRARY_MultiThreadedDebug "")
set(CMAKE_ASM_COMPILE_OPTIONS_MSVC_RUNTIME_LIBRARY_MultiThreadedDebugDLL "")

add_executable(untitled main.c blah.S)

main.c

#include <stdio.h>

void original_fn(double a, float b, size_t c, int d) {
    printf("hello from original: %lf, %f, %zu, %d\n", a, b, c, d);
}

void* my_fn() {
    printf("hello from my_fn! %f %f %f\n", 1.0, 2.0, 3.0);
    return (void*)&original_fn;
}

void proxy_fn(double a, float b, size_t c, int d);

int main() {
    proxy_fn(11.0, 12.f, 13, 14);
    return 0;
}