static struct mi_registers
get_process_registers(struct mi_process proc)
{
    struct user_regs_struct regs = { 0 };
    struct iovec iov = { &regs, sizeof(regs) };
    struct mi_registers result = { 0 };
    
    ptrace(PTRACE_GETREGSET, proc.pid, NT_PRSTATUS, &iov);
    
    result.rax = regs.rax;
    result.rip = regs.rip;
    result.rbp = regs.rbp;
    result.rbx = regs.rbx;
    result.rcx = regs.rcx;
    result.rdx = regs.rdx;
    result.rsi = regs.rsi;
    result.rdi = regs.rdi;
    result.rsp = regs.rsp;
    
    result._sys = regs;
    
    return(result);
}

static void
set_process_registers(struct mi_process proc, struct mi_registers regs)
{
    regs._sys.rax = regs.rax;
    regs._sys.rip = regs.rip;
    regs._sys.rbp = regs.rbp;
    regs._sys.rbx = regs.rbx;
    regs._sys.rcx = regs.rcx;
    regs._sys.rdx = regs.rdx;
    regs._sys.rsi = regs.rsi;
    regs._sys.rdi = regs.rdi;
    regs._sys.rsp = regs.rsp;
    
    struct iovec iov = { &regs._sys, sizeof(regs._sys) };
    ptrace(PTRACE_SETREGSET, proc.pid, NT_PRSTATUS, &iov);
}


static int
is_call_instruction(long word)
{
    // It's a call if:
    // first byte == 0xe8
    // first byte == 0xff AND "reg field" (bit 3-5 in next byte) == 2 or 3
    // REX.W + 0xff/3 ????? TODO TODO TODO
    
    u8 first_byte = word & 0xFF;
    if (first_byte == 0xE8) {
        return 1;
    } else if (first_byte == 0xFF) {
        u8 second_byte = word & 0xFF00UL;
        u8 reg = second_byte & 0x38; // 111000 (to extract bits 3-5)
        if (reg == 2 || reg == 3) {
            return 1;
        }
    }
    
    return 0;
}

static long
place_breakpoint_at(struct mi_process proc, u64 address)
{
    long saved_instruction = ptrace(PTRACE_PEEKDATA, proc.pid, address, NULL);
    long int3_byte = 0x000000000000CC;
    long int3_word = (saved_instruction & 0xFFFFFFFFFFFFFF00) | int3_byte;
    
    /* write 0xcc */
    ptrace(PTRACE_POKEDATA, proc.pid, address, int3_word);
    
    return(saved_instruction);
}

static void
run_until_breakpoint_and_restore(struct mi_process proc, u64 address, long saved_instruction)
{
    /* wait till child hits the interrupt */
    ptrace(PTRACE_CONT, proc.pid, 0, 0);
    waitpid(proc.pid, 0, 0);
    
    /* restore original instrucion */
    ptrace(PTRACE_POKEDATA, proc.pid, address, saved_instruction);
    
    /* step back to original instruction */
    struct mi_registers regs = get_process_registers(proc);
    regs.rip -= 1;
    set_process_registers(proc, regs);
}

static u64
get_address_of_subroutine(struct mi_process proc, char *sr)
{
    for (int i = 0; i < proc.debug.func_count; ++i) {
        struct mi_function *func = proc.debug.functions + i;
        if (0 == strcmp(func->name, sr)) {
            return(func->low_pc);
        }
    }
    
    return(0);
}

static struct mi_sourcepoint *
pc_to_sourcepoint(struct mi_process proc, u64 pc, int *comp_unit)
{
    // NOTE: find first point BIGGER that pc, return the sourcepoint just before that
    // TODO: binary search
    
    for (int c = 0; c < proc.debug.cu_count; ++c) {
        struct mi_compunit unit = proc.debug.compilation_units[c];
        if (unit.low_pc <= pc && pc < unit.high_pc) {
            for (int i = 0; i < unit.source.sp_count; ++i) {
                struct mi_sourcepoint *point = unit.source.sp_table + i;
                if (point->pc > pc) {
                    *comp_unit = c;
                    return(unit.source.sp_table + i - 1);
                }
            }
            break;
        }
    }
    
    return(0);
}

static void
print_word_at_address(int child, u64 address)
{
    long word = ptrace(PTRACE_PEEKDATA, child, address, NULL);
    u8 nb[4];
    memcpy(nb, &word, 4);
    printf("word at %#018lx: {%x %x %x %x}\n", address, nb[0], nb[1], nb[2], nb[3]);
}

static struct mi_buffer
read_file_mmap(char *path)
{
    struct mi_buffer result = { 0 };
    
    struct stat s;
    int fd = open(path, O_RDONLY);
    
    if (fd == -1) {
        printf("File not found: %s\n", path);
        DIE("open() failed\n");
    }
    
    fstat(fd, &s);
    
    result.data = mmap(0, s.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
    result.size = s.st_size;
    
    printf("mmap()-ed %ld bytes (%s)\n", result.size, path);
    
    close(fd);
    
    return(result);
}

static void
print_sourcepoint(struct mi_process proc, int comp_unit, struct mi_sourcepoint *sp)
{
    if (!sp) {
        printf("Unknown location\n");
        return;
    }
    
    struct mi_compunit unit = proc.debug.compilation_units[comp_unit];
    
    // NOTE: sourcepoint file indices are 1-based
    if (unit.source.source_files[sp->file - 1].file.data == 0) {
        char *file_filename = unit.source.source_files[sp->file - 1].filename;
        char *file_dirname = unit.source.source_files[sp->file - 1].dir;
        char *comp_dir = unit.source.comp_dir;
        
        char full_path[512] = { 0 };
        snprintf(full_path, 511, "%s/%s/%s", comp_dir, file_dirname, file_filename);
        
        struct mi_buffer file = read_file_mmap(full_path);
        unit.source.source_files[sp->file - 1].file = file;
    }
    
    struct mi_buffer file = unit.source.source_files[sp->file - 1].file;
    
    char *source = (char *) file.data;
    u64 size = file.size;
    int current_line = 1;
    
    for (u32 i = 0; i < size; ++i) {
        if (current_line == sp->line) {
            int len = 0;
            int start = 0;
            int leading = 1;
            
            for (u32 j = i; j < size; ++j) {
                if (leading && (source[j] == '\t' || source[j] == ' ')) {
                    ++start;
                } else {
                    leading = 0;
                }
                
                if (source[j] == '\n') {
                    break;
                }
                
                ++len;
            }
            
            
            printf("%.*s\n", len - start, source + i + start);
            return;
        }
        
        if (source[i] == '\n') {
            current_line++;
        }
    }
}

static struct mi_process
process_create(char *path)
{
    struct mi_process result = { 0 };
    
    /* disable ASLR */
    syscall(SYS_personality, ADDR_NO_RANDOMIZE);
    
    pid_t pid = fork();
    
    if (pid == -1) {
        DIE("fork error\n");
    }
    
    if (pid == 0) {
        int rt = ptrace(PTRACE_TRACEME, 0, 0, 0);
        assert(rt == 0);
        char *args[] = { path, NULL };
        rt = execvp(path, args);
        if (rt == -1) {
            DIE("exevp error\n");
        }
    }
    
    struct mi_buffer file = read_file_mmap(path);
    
    result.elf = file.data;
    result.elf_size = file.size;
    result.pid = pid;
    
    return(result);
}

static void
print_current_instruction(struct mi_process proc)
{
    struct mi_registers regs = get_process_registers(proc);
    
    long wordb = ptrace(PTRACE_PEEKDATA, proc.pid, regs.rip - 4, NULL);
    long word = ptrace(PTRACE_PEEKDATA, proc.pid, regs.rip, NULL);
    long worda = ptrace(PTRACE_PEEKDATA, proc.pid, regs.rip + 4, NULL);
    
    u8 nb[12];
    memcpy(nb, &wordb, 4);
    memcpy(nb + 4, &word, 4);
    memcpy(nb + 8, &worda, 4);
    
    printf("PC = %#018lx: %x %x %x %x [%x] %x %x %x %x %x %x %x\n", regs.rip, 
           nb[0], nb[1], nb[2], nb[3], nb[4], nb[5], nb[6], nb[7],
           nb[8], nb[9], nb[10], nb[11]);
}

static struct mi_function *
get_function_around_pc(struct mi_process proc, u64 pc)
{
    struct mi_debuginfo debug = proc.debug;
    
    for (int c = 0; c < debug.cu_count; ++c) {
        struct mi_compunit *unit = debug.compilation_units + c;
        
        if (unit->low_pc <= pc && pc < unit->high_pc) {
            
            for (int f = 0; f < unit->functions_count; ++f) {
                struct mi_function *func = debug.functions + unit->functions_from + f;
                
                if (func->low_pc <= pc && pc < func->high_pc) {
                    return(func);
                }
            }
            
            break;
        }
    }
    
    return(0);
}

static u64
get_cfa_at_pc(struct mi_process proc, u64 pc)
{
    struct dwarf_fde *fde = eh_frame_find_fde(proc, pc);
    if (!fde) {
        DIE("could not find FDE for pc!\n");
    }
    
    struct dwarf_regset regs = eh_frame_init_registers(proc, fde->cie);
    regs = eh_frame_find_pc(proc, fde, regs, pc);
    u64 cfa = eh_frame_compute_cfa(proc, regs);
    
    return(cfa);
}