The memory order in the common pointer destructor

Question

I am trying to figure out the most relaxed (and correct) memory order for a common pointer destructor. Now I mean the following:

~shared_ptr() {
   if (p) {
     if (p->cnt.fetch_sub(1, std::memory_order_release) == 1) {
       p->cnt.load(std::memory_order_acquire);
       delete p;
     }
   }
 }

Basically, I think that all the previous ones fetch_sub()should happen before delete p;, and through p->cnt.load(std::memory_order_acquire);I create a release sequence that provides this.

I am new to C ++ memory model and not quite sure. Are my reasoning correct, and the memory order that I set is correct and most relaxed?

Answer 1

In theory , you might have the most efficient code since synchronization is no longer required.

, , / (, ARMv8.3-A). .

, x86_64 fetch_sub(std::memory_order_acq_rel) fetch_sub(std::memory_order_release) .

, , , , :

std::atomic<int> cnt;
int* p;
void optimal_in_therory() {
     if (cnt.fetch_sub(1, std::memory_order_release) == 1) {
       cnt.load(std::memory_order_acquire);
       delete p;
   }
}
void optimal_in_practice_on_x86_64() {
     if (cnt.fetch_sub(1, std::memory_order_acq_rel) == 1) {
       delete p;
   }
}

:

optimal_in_therory():
  lock sub DWORD PTR cnt[rip], 1
  je .L4
  rep ret
.L4:
  mov eax, DWORD PTR cnt[rip]  ;Unnecessary extra load
  mov rdi, QWORD PTR p[rip]
  mov esi, 4
  jmp operator delete(void*, unsigned long)
optimal_in_practice_on_x86_64():
  lock sub DWORD PTR cnt[rip], 1
  je .L7
  rep ret
.L7:
  mov rdi, QWORD PTR p[rip]
  mov esi, 4
  jmp operator delete(void*, unsigned long)

, - Desproges

---

?

, . , :

cnt.load(std::memory_order_acquire);
cnt.load(std::memory_order_acquire);
cnt.load(std::memory_order_acquire);

GCC Clang :

mov eax, DWORD PTR cnt[rip]
mov eax, DWORD PTR cnt[rip]
mov eax, DWORD PTR cnt[rip]

. , - "" "". , volatile , , volatile: " - ". ( ). SO.

, , , , .

atom_thread_fence, Kieth . , , "" , (, , ;)).

, ?

shared_ptr , . Desctructor , "" .

delete p; " " , .

:

[intro.races]/9:

. B, :

• A B [...]

• "sequenced before" .

[intro.races]/10:

A B (, , B A), :

• A B

• B.

, "synchronize with" fetch_sub, delete p, fetch_sub.

[atomics.order]/2:

A, M, B, M , A.

, delete p , , fetch_sub.

[expr.races]/5 fetch_sub ( cnt) fetch_sub, a fetch_sub --, fetch_add (, cnt).

, delete p fetch_sub, , delete p, "". , .