clang vs gcc: different code for volatile access

Question

Consider this example:

volatile unsigned int x;
unsigned int y;

void f() {
    x /= 2;
}
void g() {
    y /= 2;
}

When compiled with -Os, clang-6.0 produces on x64 for both f and g the same shrl <offset>(%rip) instruction pattern (See https://godbolt.org/g/hUPprL), while gcc-7.3 produces this (See https://godbolt.org/g/vMcKVV) for f():

 mov 0x200b67(%rip),%eax # 601034 <x>
 shr %eax
 mov %eax,0x200b5f(%rip) # 601034 <x>

Is this just a missed optimization or is there a justification for gcc to reject shrl <offset>(%rip) in case of volatile access? Who is wrong?

Answer 1

This is just a missed optimization by gcc. Both implementations preserve the read from and write to x precisely, and thus are correct.

"Under the hood" operating on a memory operand performs the same loads and stores as the longer implementation.