skia2/include/private/SkAtomics.h
mtklein bf90520f63 SkPath::fFirstDirection: seq-cst -> relaxed
We landed this originally with lazily-correct sequentially-consistent memory
order.  It turns out that's regressed performance, we think particularly when
recording paths.  We also think there's no need for anything but relaxed memory
order here.

We should see this chart go down if all goes well: https://perf.skia.org/#4329
There are also Chrome performance charts to watch in the linked bug.

BUG=chromium:537700

CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-TSAN-Trybot,Test-Ubuntu-GCC-Golo-GPU-GT610-x86_64-Release-TSAN

No public API changes.
TBR=reed@google.com

Review URL: https://codereview.chromium.org/1393833003
2015-10-07 12:46:43 -07:00

138 lines
4.5 KiB
C++

/*
* Copyright 2015 Google Inc.
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#ifndef SkAtomics_DEFINED
#define SkAtomics_DEFINED
// This file is not part of the public Skia API.
#include "SkTypes.h"
enum sk_memory_order {
sk_memory_order_relaxed,
sk_memory_order_consume,
sk_memory_order_acquire,
sk_memory_order_release,
sk_memory_order_acq_rel,
sk_memory_order_seq_cst,
};
template <typename T>
T sk_atomic_load(const T*, sk_memory_order = sk_memory_order_seq_cst);
template <typename T>
void sk_atomic_store(T*, T, sk_memory_order = sk_memory_order_seq_cst);
template <typename T>
T sk_atomic_fetch_add(T*, T, sk_memory_order = sk_memory_order_seq_cst);
template <typename T>
T sk_atomic_fetch_sub(T*, T, sk_memory_order = sk_memory_order_seq_cst);
template <typename T>
bool sk_atomic_compare_exchange(T*, T* expected, T desired,
sk_memory_order success = sk_memory_order_seq_cst,
sk_memory_order failure = sk_memory_order_seq_cst);
template <typename T>
T sk_atomic_exchange(T*, T, sk_memory_order = sk_memory_order_seq_cst);
// A little wrapper class for small T (think, builtins: int, float, void*) to
// ensure they're always used atomically. This is our stand-in for std::atomic<T>.
// !!! Please _really_ know what you're doing if you change default_memory_order. !!!
template <typename T, sk_memory_order default_memory_order = sk_memory_order_seq_cst>
class SkAtomic : SkNoncopyable {
public:
SkAtomic() {}
explicit SkAtomic(const T& val) : fVal(val) {}
// It is essential we return by value rather than by const&. fVal may change at any time.
T load(sk_memory_order mo = default_memory_order) const {
return sk_atomic_load(&fVal, mo);
}
void store(const T& val, sk_memory_order mo = default_memory_order) {
sk_atomic_store(&fVal, val, mo);
}
// Alias for .load(default_memory_order).
operator T() const {
return this->load();
}
// Alias for .store(v, default_memory_order).
T operator=(const T& v) {
this->store(v);
return v;
}
T fetch_add(const T& val, sk_memory_order mo = default_memory_order) {
return sk_atomic_fetch_add(&fVal, val, mo);
}
T fetch_sub(const T& val, sk_memory_order mo = default_memory_order) {
return sk_atomic_fetch_sub(&fVal, val, mo);
}
bool compare_exchange(T* expected, const T& desired,
sk_memory_order success = default_memory_order,
sk_memory_order failure = default_memory_order) {
return sk_atomic_compare_exchange(&fVal, expected, desired, success, failure);
}
private:
T fVal;
};
// IWYU pragma: begin_exports
#if defined(_MSC_VER)
#include "../ports/SkAtomics_std.h"
#elif !defined(SK_BUILD_FOR_IOS) && defined(__ATOMIC_RELAXED)
#include "../ports/SkAtomics_atomic.h"
#else
#include "../ports/SkAtomics_sync.h"
#endif
// IWYU pragma: end_exports
// From here down we have shims for our old atomics API, to be weaned off of.
// We use the default sequentially-consistent memory order to make things simple
// and to match the practical reality of our old _sync and _win implementations.
inline int32_t sk_atomic_inc(int32_t* ptr) { return sk_atomic_fetch_add(ptr, +1); }
inline int32_t sk_atomic_dec(int32_t* ptr) { return sk_atomic_fetch_add(ptr, -1); }
inline int32_t sk_atomic_add(int32_t* ptr, int32_t v) { return sk_atomic_fetch_add(ptr, v); }
inline int64_t sk_atomic_inc(int64_t* ptr) { return sk_atomic_fetch_add<int64_t>(ptr, +1); }
inline bool sk_atomic_cas(int32_t* ptr, int32_t expected, int32_t desired) {
return sk_atomic_compare_exchange(ptr, &expected, desired);
}
inline void* sk_atomic_cas(void** ptr, void* expected, void* desired) {
(void)sk_atomic_compare_exchange(ptr, &expected, desired);
return expected;
}
inline int32_t sk_atomic_conditional_inc(int32_t* ptr) {
int32_t prev = sk_atomic_load(ptr);
do {
if (0 == prev) {
break;
}
} while(!sk_atomic_compare_exchange(ptr, &prev, prev+1));
return prev;
}
template <typename T>
T sk_acquire_load(T* ptr) { return sk_atomic_load(ptr, sk_memory_order_acquire); }
template <typename T>
void sk_release_store(T* ptr, T val) { sk_atomic_store(ptr, val, sk_memory_order_release); }
inline void sk_membar_acquire__after_atomic_dec() {}
inline void sk_membar_acquire__after_atomic_conditional_inc() {}
#endif//SkAtomics_DEFINED