https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121148
Bug ID: 121148
Summary: Should use modular arithmetic for _Atomic_word
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: redi at gcc dot gnu.org
Blocks: 71945
Target Milestone: ---
The std::atomic<integral-type>::fetch_xxx members are required to never
overflow:
Remarks: [...] for signed integer types the result is as if the object value
and parameters were converted to their corresponding unsigned types, the
computation performed on those types, and the result converted back to the
signed type.
[Note 2 : There are no undefined results arising from the computation. — end
note]
So std::atomic<int>(INT_MAX).fetch_add(1) gives INT_MIN, not undefined
overflow.
Our atomic helper functions __gnu_cxx::__atomic_add and
__gnu_cxx::__exchange_and_add should behave the same, to reduce surprises and
avoid bugs.
The default implementations are defined in terms of the __atomic_fetch_add
built-in, so they're fine, but the single-threaded fallbacks and some
alternative implementations can overflow.
e.g. in include/ext/atomicity.h:
inline _Atomic_word
__attribute__((__always_inline__))
__exchange_and_add_single(_Atomic_word* __mem, int __val)
{
_Atomic_word __result = *__mem;
*__mem += __val;
return __result;
}
Since _Atomic_word is always a signed type (usually int, but long on 64-bit
sparc), this should use an unsigned type for the arithmetic, e.g.
_Atomic_word __result = *__mem;
#if __cplusplus >= 201103L
std::make_unsigned<_Atomic_word>::type __u;
#else
unsigned long long __u;
#endif
__u = __result;
__u += __val;
*__mem = __u;
return __result;
And similarly for __atomic_add.
And we have this in config/cpu/generic/atomicity_mutex/atomicity.h
_Atomic_word
__attribute__ ((__unused__))
__exchange_and_add(volatile _Atomic_word* __mem, int __val) throw ()
{
__gnu_cxx::__scoped_lock sentry(get_atomic_mutex());
_Atomic_word __result;
__result = *__mem;
*__mem += __val;
return __result;
}
This one should just be:
__gnu_cxx::__scoped_lock sentry(get_atomic_mutex());
return __gnu_cxx::__exchange_and_add_single(__mem, __val);
So that we don't need to repeat the logic.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71945
[Bug 71945] Integer overflow in use counter of shared pointers
