Fast path memory ordering is inconsistent between libcpp and libc

[gbaraldi]

libc uses a relaxed load

llvm-project/libc/src/__support/threads/linux/callonce.h

Lines 26 to 28 in 95820ca

	LIBC_INLINE bool callonce_fastpath(CallOnceFlag *flag) {
	return flag->load(cpp::MemoryOrder::RELAXED) == FINISH;
	}

while libcpp uses acquire

llvm-project/libcxx/include/__mutex/once_flag.h

Line 127 in 95820ca

if (__libcpp_acquire_load(&__flag.__state_) != once_flag::_Complete) {

A relaxed load has no ordering constraints so if one thread calls pthread_once and another goes through the fast path, even if the other thread observes the FINISH value. What guarantees that the side effects are visible from it?

[nickdesaulniers]

cc @SchrodingerZhu we can change llvm-libc to match libc++.