即使我有问题，GCC也不会在我的内联asm函数调用周围推送寄存器

Question

I have a function (C) that modifies "ecx" (or any other registers)

int proc(int n) {
    int ret;
    asm volatile ("movl %1, %%ecx\n\t" // mov (n) to ecx
                  "addl $10, %%ecx\n\t" // add (10) to ecx (n)
                  "movl %%ecx, %0" /* ret = n + 10 */
                  : "=r" (ret) : "r" (n) : "ecx");
    return ret;
}

now i want to call this function in another function which that function moves a value in "ecx" before calling "proc" function

int main_proc(int n) {
    asm volatile ("movl     $55, %%ecx" ::: "ecx"); /// mov (55) to ecx
    int ret;
    asm volatile ("call     proc" : "=r" (ret) : "r" (n) : "ecx"); // ecx is modified in proc function and the value of ecx is not 55 anymore even with "ecx" clobber

    asm volatile ("addl     %%ecx, %0" : "=r" (ret));

    return ret;
}

in this function, (55) is moved into "ecx" register and then "proc" function is called (which modifies "ecx"). in this situation, "proc" function Must push "ecx" first and pop it at the end but it's not going to happen !!!! this is the assembly source with (-O3) optimiaztion level

proc:
        movl %edi, %ecx
        addl $10, %ecx
        movl %ecx, %eax
        ret
main_proc:
        movl     $55, %ecx
        call     proc
        addl     %ecx, %eax
        ret

why GCC is not going to use (push) and (pop) for "ecx" register ?? i used "ecx" clobber too !!!!!

Answer 1

You are using inline asm completely wrong. Your input/output constraints need to fully describe the inputs / outputs of each asm statement. To get data between asm statements, you have to hold it in C variables between them.

Also, call isn't safe inside inline asm in general, and specifically in x86-64 code for the System V ABI it steps on the red-zone where gcc might have been keeping things. There's no way to declare a clobber on that. You could use sub $128, %rsp first to skip past the red zone, or you could make calls from pure C like a normal person so the compiler knows about it. (Remember that call pushes a return address.) Your inline asm doesn't even make sense; your proc takes an arg but you didn't do anything in the caller to pass one.

The compiler-generated code in proc could have also destroyed any other call-clobbered registers, so you at least need to declare clobbers on those registers. Or hand-write the whole function in asm so you know what to put in clobbers.

why GCC is not going to use (push) and (pop) for "ecx" register ?? i used "ecx" clobber too !!!!!

An ecx clobber tells GCC that this asm statement destroys whatever GCC had in ECX previously. Using an ECX clobber in two separate inline-asm statements doesn't declare any kind of data dependency between them.

It's not equivalent to declaring a register-asm local variable like
register int foo asm("ecx"); that you use as a "+r" (foo) operand to the first and last asm statement. (Or more simply that you use with a "+c" constraint to make an ordinary variable pick ECX).

From GCC's point of view, your source means only what the constraints + clobbers tell it.

int main_proc(int n) {
    asm volatile ("movl     $55, %%ecx" ::: "ecx");
      // ^^ black box that destroys ECX and produces no outputs
    int ret;
    asm volatile ("call     proc" : "=r" (ret) : "r" (n) : "ecx");
      // ^^ black box that can take `n` in any register, and can produce `ret` in any reg.  And destroys ECX.

    asm volatile ("addl     %%ecx, %0" : "=r" (ret));
     // ^^ black box with no inputs that can produce a new value for `ret` in any register

    return ret;
}

I suspect you wanted the last asm statement to be "+r"(ret) to read/write the C variable ret instead of telling GCC that it was output-only. Because your asm uses it as an input as well as output as the destination of an add .

It might be interesting to add comments like # %%0 = %0 %%1 = %1 inside your 2nd asm statement to see which registers the "=r" and "r" constraints picked. On the Godbolt compiler explorer :

# gcc9.2 -O3 
main_proc:
        movl     $55, %ecx
        call     proc         # %0 = %edi   %1 = %edi
        addl     %ecx, %eax    # "=r" happened to pick EAX,
                      # which happens to still hold the return value from  proc
        ret

That accident of picking EAX as the add destinatino might not happen after this function inlines into something else. or GCC happens to put some compiler-generated instructions between asm statements. ( asm volatile is barrier to compile-time reordering but not not a strog one. It only definitely stops optimizing away entirely).

Remember that inline asm templates are purely text substitution; asking the compiler to fill in an operand into a comment is no different from anywhere else in the template string. (Godbolt strips comment lines by default so sometimes it's handy to tack them onto other instructions, or onto a nop).

As you can see, this is 64-bit code ( n arrives in EDI as per the x86-64 SysV calling convention, like how you built your code), so push %ecx wouldn't be encodeable. push %rcx would be.

Of course if GCC actually wanted to keep a value around past an asm statement with an "ecx" clobber, it would have just used mov %ecx, %edx or whatever other call-clobbered register that wasn't in the clobber list.