Why are unnecessary atomic loads not optimized awa

问题:

Let's consider this trivial code:

#include <atomic>

std::atomic<int> a;
void f(){
    for(int k=0;k<100;++k)
        a.load(std::memory_order_relaxed);
}

MSVC, Clang and GCC all perform 100 loads of a, while it seems obvious it could have been optimized away. I expected the function f to be a nop (See generated code here)

Actually, I expected this code generation for a volatile atomic:

volatile std::atomic<int> va;
void g(){
    for(int k=0;k<100;++k)
        va.load(std::memory_order_relaxed);
}

Why do compilers not optimize away unnecessary atomic loads?