What does __asm volatile (“pause” ::: “memory”); do?

Question

I am looking at an open source C++ project which has the following code structure:

while(true) {
  // Do something work

  if(some_condition_becomes_true)
     b         


        

Answer 1

It's _mm_pause() and a compile memory barrier wrapped into one GNU C Extended ASM statement. https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html

asm("" ::: "memory") prevents compile-time reordering of memory operations across it, like C++11 std::atomic_signal_fence(std::memory_order_seq_cst). (not atomic_thread_fence; although on x86 preventing compile-time reordering is sufficient to make it an acquire + release fence because the only run-time reordering that x86 allows is StoreLoad.) See Jeff Preshing's Memory Ordering at Compile Time article.

Making the asm instruction part non-empty also means those asm instructions will run every time the C logically runs that source line (because it's volatile).

pause prevents speculative loads from causing memory-ordering mis-speculation pipeling clears (aka machine nukes). It's useful inside spin loops that are waiting to see a value in memory.

You might find this statement inside a spinloop written without C++11 std::atomic, to tell the compiler it has to re-read the value of a global variable. (Because the "memory" clobber means the compiler has to assume the asm statement might have modified the value of any globally-reachable memory.)

This looks like the context where you found it: some_condition_becomes_true probably includes reading a non-atomic / non-volatile global.

The C++11 equivalent of your loop:

#include <atomic>
#include <immintrin.h>
std::atomic<int> flag;

void wait_for_flag(void) {
    while(flag.load(std::memory_order_seq_cst == 0) {
        _mm_pause();
    }
}

(Not exactly equivalent, because your version has a full compiler barrier while mine only has a seq-cst load, so it's not a full signal-fence. But probably what wasn't needed, and they just used something stronger than necessary to get the effect of volatile).

---

Without the barrier or making flag atomic, the compiler would have optimized it to:

// Do something work

if(some_condition_becomes_true) {
    // empty
} else {

  while(true) {
     // Do something work
     __asm volatile ("pause" ::: );  // no memory clobber
  }
}

i.e. it would hoist the check on some_condition_becomes_true out of the loop and not re-read the global every time.