Quoth the C11 standard on atomic_compare_exchange_X_explicit: 'The failure argument shall be no stronger than the success argument.' Knowing this, the fuzzer never generates memory orders that fail this requirement. However, we've found on at least one occasion that we can find fun compiler bugs by making the compiler compile unreachable code that may exhibit UB, so it seems like a good idea to loosen this requirement when building cmpxchg in dead code.
NB: I'm assuming that no compiler errors out when given code like this (Clang certainly doesn't).
Quoth the C11 standard on
atomic_compare_exchange_X_explicit: 'The failure argument shall be no stronger than the success argument.' Knowing this, the fuzzer never generates memory orders that fail this requirement. However, we've found on at least one occasion that we can find fun compiler bugs by making the compiler compile unreachable code that may exhibit UB, so it seems like a good idea to loosen this requirement when building cmpxchg in dead code.NB: I'm assuming that no compiler errors out when given code like this (Clang certainly doesn't).