#define APHELEIA_SETJMP_NO_MACRO

#include <setjmp.h>

// clang-format off
#if defined(__riscv) && __riscv_xlen == 64
__asm__(
    ".globl setjmp\n"
    "setjmp:\n"
    "sd ra, 0(a0)\n"
    "sd s0, 8(a0)\n"
    "sd s1, 16(a0)\n"
    "sd s2, 24(a0)\n"
    "sd s3, 32(a0)\n"
    "sd s4, 40(a0)\n"
    "sd s5, 48(a0)\n"
    "sd s6, 56(a0)\n"
    "sd s7, 64(a0)\n"
    "sd s8, 72(a0)\n"
    "sd s9, 80(a0)\n"
    "sd s10, 88(a0)\n"
    "sd s11, 96(a0)\n"
    "sd sp, 104(a0)\n"
    "li a0, 0\n"
    "ret\n"
    "\n"
    ".globl longjmp\n"
    "longjmp:\n"
    "mv t0, a0\n"
    "mv t1, a1\n"
    "bnez t1, 0f\n"
    "li t1, 1\n"
    "0:\n"
    "ld ra, 0(t0)\n"
    "ld s0, 8(t0)\n"
    "ld s1, 16(t0)\n"
    "ld s2, 24(t0)\n"
    "ld s3, 32(t0)\n"
    "ld s4, 40(t0)\n"
    "ld s5, 48(t0)\n"
    "ld s6, 56(t0)\n"
    "ld s7, 64(t0)\n"
    "ld s8, 72(t0)\n"
    "ld s9, 80(t0)\n"
    "ld s10, 88(t0)\n"
    "ld s11, 96(t0)\n"
    "ld sp, 104(t0)\n"
    "mv a0, t1\n"
    "jr ra\n"
);
#elif defined(__riscv) && __riscv_xlen == 32
__asm__(
    ".globl setjmp\n"
    "setjmp:\n"
    "sw ra, 0(a0)\n"
    "sw s0, 4(a0)\n"
    "sw s1, 8(a0)\n"
    "sw s2, 12(a0)\n"
    "sw s3, 16(a0)\n"
    "sw s4, 20(a0)\n"
    "sw s5, 24(a0)\n"
    "sw s6, 28(a0)\n"
    "sw s7, 32(a0)\n"
    "sw s8, 36(a0)\n"
    "sw s9, 40(a0)\n"
    "sw s10, 44(a0)\n"
    "sw s11, 48(a0)\n"
    "sw sp, 52(a0)\n"
    "li a0, 0\n"
    "ret\n"
    "\n"
    ".globl longjmp\n"
    "longjmp:\n"
    "mv t0, a0\n"
    "mv t1, a1\n"
    "bnez t1, 0f\n"
    "li t1, 1\n"
    "0:\n"
    "lw ra, 0(t0)\n"
    "lw s0, 4(t0)\n"
    "lw s1, 8(t0)\n"
    "lw s2, 12(t0)\n"
    "lw s3, 16(t0)\n"
    "lw s4, 20(t0)\n"
    "lw s5, 24(t0)\n"
    "lw s6, 28(t0)\n"
    "lw s7, 32(t0)\n"
    "lw s8, 36(t0)\n"
    "lw s9, 40(t0)\n"
    "lw s10, 44(t0)\n"
    "lw s11, 48(t0)\n"
    "lw sp, 52(t0)\n"
    "mv a0, t1\n"
    "jr ra\n"
);
#elif defined(__x86_64__)
__asm__(
    ".globl setjmp\n"
    "setjmp:\n"
    "movq %rbx, 0(%rdi)\n"
    "movq %rbp, 8(%rdi)\n"
    "movq %r12, 16(%rdi)\n"
    "movq %r13, 24(%rdi)\n"
    "movq %r14, 32(%rdi)\n"
    "movq %r15, 40(%rdi)\n"
    "movq (%rsp), %rax\n"
    "movq %rax, 48(%rdi)\n"
    "leaq 8(%rsp), %rax\n"
    "movq %rax, 56(%rdi)\n"
    "xorl %eax, %eax\n"
    "ret\n"
    "\n"
    ".globl longjmp\n"
    "longjmp:\n"
    "movq %rdi, %rdx\n"
    "movl %esi, %eax\n"
    "testl %eax, %eax\n"
    "jnz 0f\n"
    "movl $1, %eax\n"
    "0:\n"
    "movq 0(%rdx), %rbx\n"
    "movq 8(%rdx), %rbp\n"
    "movq 16(%rdx), %r12\n"
    "movq 24(%rdx), %r13\n"
    "movq 32(%rdx), %r14\n"
    "movq 40(%rdx), %r15\n"
    "movq 56(%rdx), %rsp\n"
    "jmp *48(%rdx)\n"
);
#elif defined(__i386__)
__asm__(
    ".globl setjmp\n"
    "setjmp:\n"
    "movl 4(%esp), %edx\n"
    "movl %ebx, 0(%edx)\n"
    "movl %esi, 4(%edx)\n"
    "movl %edi, 8(%edx)\n"
    "movl %ebp, 12(%edx)\n"
    "movl (%esp), %eax\n"
    "movl %eax, 16(%edx)\n"
    "leal 4(%esp), %eax\n"
    "movl %eax, 20(%edx)\n"
    "xorl %eax, %eax\n"
    "ret\n"
    "\n"
    ".globl longjmp\n"
    "longjmp:\n"
    "movl 4(%esp), %edx\n"
    "movl 8(%esp), %eax\n"
    "testl %eax, %eax\n"
    "jnz 0f\n"
    "movl $1, %eax\n"
    "0:\n"
    "movl 0(%edx), %ebx\n"
    "movl 4(%edx), %esi\n"
    "movl 8(%edx), %edi\n"
    "movl 12(%edx), %ebp\n"
    "movl 20(%edx), %esp\n"
    "jmp *16(%edx)\n"
);
#else
#error "Unsupported architecture"
#endif
// clang-format on