seedlingattempt/godot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Allow inlining of all parts of safe_refcount

Browse Source 
Differences with this aren't huge but the effort is minimal, in some
workloads gain a couple of percent of performance.
This commit is contained in:
Hein-Pieter van Braam
2017-09-25 00:01:55 +02:00
parent 0a338a28d9
commit 158c3fbf3c
2 changed files with 210 additions and 304 deletions
Download Patch File Download Diff File Expand all files Collapse all files

227
core/safe_refcount.h
View File

@ -36,19 +36,212 @@
#include "platform_config.h"
#include "typedefs.h"
uint32_t atomic_conditional_increment(register uint32_t *counter);
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);
// Atomic functions, these are used for multithread safe reference counters!
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);
#ifdef NO_THREADS
/* Bogus implementation unaware of multiprocessing */
template <class T>
static _ALWAYS_INLINE_ T atomic_conditional_increment(register T *pw) {
if (*pw == 0)
return 0;
(*pw)++;
return *pw;
}
template <class T>
static _ALWAYS_INLINE_ T atomic_decrement(register T *pw) {
(*pw)--;
return *pw;
}
template <class T>
static _ALWAYS_INLINE_ T atomic_increment(register T *pw) {
(*pw)++;
return *pw;
}
template <class T, class V>
static _ALWAYS_INLINE_ T atomic_sub(register T *pw, register V val) {
(*pw) -= val;
return *pw;
}
template <class T, class V>
static _ALWAYS_INLINE_ T atomic_add(register T *pw, register V val) {
(*pw) += val;
return *pw;
}
template <class T, class V>
static _ALWAYS_INLINE_ T atomic_exchange_if_greater(register T *pw, register V val) {
if (val > *pw)
*pw = val;
return *pw;
}
#elif defined(__GNUC__)
/* Implementation for GCC & Clang */
// GCC guarantees atomic intrinsics for sizes of 1, 2, 4 and 8 bytes.
// Clang states it supports GCC atomic builtins.
template <class T>
static _ALWAYS_INLINE_ T atomic_conditional_increment(register T *pw) {
while (true) {
T tmp = static_cast<T const volatile &>(*pw);
if (tmp == 0)
return 0; // if zero, can't add to it anymore
if (__sync_val_compare_and_swap(pw, tmp, tmp + 1) == tmp)
return tmp + 1;
}
}
template <class T>
static _ALWAYS_INLINE_ T atomic_decrement(register T *pw) {
return __sync_sub_and_fetch(pw, 1);
}
template <class T>
static _ALWAYS_INLINE_ T atomic_increment(register T *pw) {
return __sync_add_and_fetch(pw, 1);
}
template <class T, class V>
static _ALWAYS_INLINE_ T atomic_sub(register T *pw, register V val) {
return __sync_sub_and_fetch(pw, val);
}
template <class T, class V>
static _ALWAYS_INLINE_ T atomic_add(register T *pw, register V val) {
return __sync_add_and_fetch(pw, val);
}
template <class T, class V>
static _ALWAYS_INLINE_ T atomic_exchange_if_greater(register T *pw, register V 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 */
// don't pollute my namespace!
#include <windows.h>
#define ATOMIC_CONDITIONAL_INCREMENT_BODY(m_pw, m_win_type, m_win_cmpxchg, m_cpp_type) \
/* try to increment until it actually works */ \
/* taken from boost */ \
while (true) { \
m_cpp_type tmp = static_cast<m_cpp_type const volatile &>(*(m_pw)); \
if (tmp == 0) \
return 0; /* if zero, can't add to it anymore */ \
if (m_win_cmpxchg((m_win_type volatile *)(m_pw), tmp + 1, tmp) == tmp) \
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(register uint32_t *pw) {
ATOMIC_CONDITIONAL_INCREMENT_BODY(pw, LONG, InterlockedCompareExchange, uint32_t)
}
static _ALWAYS_INLINE_ uint32_t atomic_decrement(register uint32_t *pw) {
return InterlockedDecrement((LONG volatile *)pw);
}
static _ALWAYS_INLINE_ uint32_t atomic_increment(register uint32_t *pw) {
return InterlockedIncrement((LONG volatile *)pw);
}
static _ALWAYS_INLINE_ uint32_t atomic_sub(register uint32_t *pw, register uint32_t val) {
return InterlockedExchangeAdd((LONG volatile *)pw, -(int32_t)val) - val;
}
static _ALWAYS_INLINE_ uint32_t atomic_add(register uint32_t *pw, register uint32_t val) {
return InterlockedAdd((LONG volatile *)pw, val);
}
static _ALWAYS_INLINE_ uint32_t atomic_exchange_if_greater(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(register uint64_t *pw) {
ATOMIC_CONDITIONAL_INCREMENT_BODY(pw, LONGLONG, InterlockedCompareExchange64, uint64_t)
}
static _ALWAYS_INLINE_ uint64_t atomic_decrement(register uint64_t *pw) {
return InterlockedDecrement64((LONGLONG volatile *)pw);
}
static _ALWAYS_INLINE_ uint64_t atomic_increment(register uint64_t *pw) {
return InterlockedIncrement64((LONGLONG volatile *)pw);
}
static _ALWAYS_INLINE_ uint64_t atomic_sub(register uint64_t *pw, register uint64_t val) {
return InterlockedExchangeAdd64((LONGLONG volatile *)pw, -(int64_t)val) - val;
}
static _ALWAYS_INLINE_ uint64_t atomic_add(register uint64_t *pw, register uint64_t val) {
return InterlockedAdd64((LONGLONG volatile *)pw, val);
}
static _ALWAYS_INLINE_ uint64_t atomic_exchange_if_greater(register uint64_t *pw, register uint64_t val) {
ATOMIC_EXCHANGE_IF_GREATER_BODY(pw, val, LONGLONG, InterlockedCompareExchange64, uint64_t)
}
#else
//no threads supported?
#error Must provide atomic functions for this platform or compiler!
#endif
struct SafeRefCount {
@ -57,17 +250,17 @@ struct SafeRefCount {
public:
// destroy() is called when weak_count_ drops to zero.
bool ref() { //true on success
_ALWAYS_INLINE_ bool ref() { //true on success
return atomic_conditional_increment(&count) != 0;
}
uint32_t refval() { //true on success
_ALWAYS_INLINE_ uint32_t refval() { //true on success
return atomic_conditional_increment(&count);
}
bool unref() { // true if must be disposed of
_ALWAYS_INLINE_ bool unref() { // true if must be disposed of
if (atomic_decrement(&count) == 0) {
return true;
@ -76,12 +269,12 @@ public:
return false;
}
uint32_t get() const { // nothrow
_ALWAYS_INLINE_ uint32_t get() const { // nothrow
return count;
}
void init(uint32_t p_value = 1) {
_ALWAYS_INLINE_ void init(uint32_t p_value = 1) {
count = p_value;
}