Improve/add atomics

- Remove use of non-builtin overloaded `InterlockedExhangeSubtract()` for Windows. Moreover, it was wrongly being used as if it hadn't a 64-bit version instead of being overloaded.
- Implement exchange-if-greater (as a CAS loop).
This commit is contained in:
Pedro J. Estébanez 2017-07-25 01:21:39 +02:00
parent e15a0c5243
commit 27637b9e7f
2 changed files with 50 additions and 8 deletions

View File

@ -78,6 +78,15 @@ static _ALWAYS_INLINE_ T _atomic_add_impl(register T *pw, register T val) {
return *pw;
}
template <class T>
static _ALWAYS_INLINE_ T _atomic_exchange_if_greater_impl(register T *pw, register T val) {
if (val > *pw)
*pw = val;
return *pw;
}
#elif defined(__GNUC__)
/* Implementation for GCC & Clang */
@ -121,6 +130,18 @@ static _ALWAYS_INLINE_ T _atomic_add_impl(register T *pw, register T val) {
return __sync_add_and_fetch(pw, val);
}
template <class T>
static _ALWAYS_INLINE_ T _atomic_exchange_if_greater_impl(register T *pw, register T val) {
while (true) {
T tmp = static_cast<T const volatile &>(*pw);
if (tmp >= val)
return tmp; // already greater, or equal
if (__sync_val_compare_and_swap(pw, tmp, val) == tmp)
return val;
}
}
#elif defined(_MSC_VER)
/* Implementation for MSVC-Windows */
@ -139,6 +160,15 @@ static _ALWAYS_INLINE_ T _atomic_add_impl(register T *pw, register T val) {
return tmp + 1; \
}
#define ATOMIC_EXCHANGE_IF_GREATER_BODY(m_pw, m_val, m_win_type, m_win_cmpxchg, m_cpp_type) \
while (true) { \
m_cpp_type tmp = static_cast<m_cpp_type const volatile &>(*(m_pw)); \
if (tmp >= m_val) \
return tmp; /* already greater, or equal */ \
if (m_win_cmpxchg((m_win_type volatile *)(m_pw), m_val, tmp) == tmp) \
return m_val; \
}
static _ALWAYS_INLINE_ uint32_t _atomic_conditional_increment_impl(register uint32_t *pw) {
ATOMIC_CONDITIONAL_INCREMENT_BODY(pw, LONG, InterlockedCompareExchange, uint32_t)
@ -156,11 +186,7 @@ static _ALWAYS_INLINE_ uint32_t _atomic_increment_impl(register uint32_t *pw) {
static _ALWAYS_INLINE_ uint32_t _atomic_sub_impl(register uint32_t *pw, register uint32_t val) {
#if _WIN32_WINNT >= 0x0601 // Windows 7+
return InterlockedExchangeSubtract(pw, val) - val;
#else
return InterlockedExchangeAdd((LONG volatile *)pw, -(int32_t)val) - val;
#endif
}
static _ALWAYS_INLINE_ uint32_t _atomic_add_impl(register uint32_t *pw, register uint32_t val) {
@ -168,6 +194,11 @@ static _ALWAYS_INLINE_ uint32_t _atomic_add_impl(register uint32_t *pw, register
return InterlockedAdd((LONG volatile *)pw, val);
}
static _ALWAYS_INLINE_ uint32_t _atomic_exchange_if_greater_impl(register uint32_t *pw, register uint32_t val) {
ATOMIC_EXCHANGE_IF_GREATER_BODY(pw, val, LONG, InterlockedCompareExchange, uint32_t)
}
static _ALWAYS_INLINE_ uint64_t _atomic_conditional_increment_impl(register uint64_t *pw) {
ATOMIC_CONDITIONAL_INCREMENT_BODY(pw, LONGLONG, InterlockedCompareExchange64, uint64_t)
@ -185,11 +216,7 @@ static _ALWAYS_INLINE_ uint64_t _atomic_increment_impl(register uint64_t *pw) {
static _ALWAYS_INLINE_ uint64_t _atomic_sub_impl(register uint64_t *pw, register uint64_t val) {
#if _WIN32_WINNT >= 0x0601 && !defined(UWP_ENABLED) // Windows 7+ except UWP
return InterlockedExchangeSubtract64(pw, val) - val;
#else
return InterlockedExchangeAdd64((LONGLONG volatile *)pw, -(int64_t)val) - val;
#endif
}
static _ALWAYS_INLINE_ uint64_t _atomic_add_impl(register uint64_t *pw, register uint64_t val) {
@ -197,6 +224,11 @@ static _ALWAYS_INLINE_ uint64_t _atomic_add_impl(register uint64_t *pw, register
return InterlockedAdd64((LONGLONG volatile *)pw, val);
}
static _ALWAYS_INLINE_ uint64_t _atomic_exchange_if_greater_impl(register uint64_t *pw, register uint64_t val) {
ATOMIC_EXCHANGE_IF_GREATER_BODY(pw, val, LONGLONG, InterlockedCompareExchange64, uint64_t)
}
#else
//no threads supported?
@ -226,6 +258,10 @@ uint32_t atomic_add(register uint32_t *pw, register uint32_t val) {
return _atomic_add_impl(pw, val);
}
uint32_t atomic_exchange_if_greater(register uint32_t *pw, register uint32_t val) {
return _atomic_exchange_if_greater_impl(pw, val);
}
uint64_t atomic_conditional_increment(register uint64_t *counter) {
return _atomic_conditional_increment_impl(counter);
}
@ -245,3 +281,7 @@ uint64_t atomic_sub(register uint64_t *pw, register uint64_t val) {
uint64_t atomic_add(register uint64_t *pw, register uint64_t val) {
return _atomic_add_impl(pw, val);
}
uint64_t atomic_exchange_if_greater(register uint64_t *pw, register uint64_t val) {
return _atomic_exchange_if_greater_impl(pw, val);
}

View File

@ -41,12 +41,14 @@ uint32_t atomic_decrement(register uint32_t *pw);
uint32_t atomic_increment(register uint32_t *pw);
uint32_t atomic_sub(register uint32_t *pw, register uint32_t val);
uint32_t atomic_add(register uint32_t *pw, register uint32_t val);
uint32_t atomic_exchange_if_greater(register uint32_t *pw, register uint32_t val);
uint64_t atomic_conditional_increment(register uint64_t *counter);
uint64_t atomic_decrement(register uint64_t *pw);
uint64_t atomic_increment(register uint64_t *pw);
uint64_t atomic_sub(register uint64_t *pw, register uint64_t val);
uint64_t atomic_add(register uint64_t *pw, register uint64_t val);
uint64_t atomic_exchange_if_greater(register uint64_t *pw, register uint64_t val);
struct SafeRefCount {