Make mutex semaphore based.

This implementation improves performance of SkMutex acquire / release pair from 42ns -> 13 ns.

SkSharedMutex and SkSpinlock have the same performance.

It also removes specialized windows and linux/mac code.

BUG=skia:

Review URL: https://codereview.chromium.org/1359733002

1 files changed, 8 insertions, 9 deletions

diff --git a/include/private/SkOncePtr.h b/include/private/SkOncePtr.h
index ee30d8f8db..a1c028c4af 100644
--- a/include/private/SkOncePtr.h
+++ b/include/private/SkOncePtr.h
@@ -14,7 +14,7 @@
 template <typename T> class SkBaseOncePtr;
 
 // Use this to create a global static pointer that's intialized exactly once when you call get().
-#define SK_DECLARE_STATIC_ONCE_PTR(type, name) namespace {} static SkBaseOncePtr<type> name
+#define SK_DECLARE_STATIC_ONCE_PTR(type, name) namespace {} static SkBaseOncePtr<type> name;
 
 // Use this for a local or member pointer that's initialized exactly once when you call get().
 template <typename T, typename Delete = skstd::default_delete<T>>
@@ -61,17 +61,17 @@ class SkBaseOncePtr {
 public:
     template <typename F>
     T* get(const F& f) const {
-        uintptr_t state = fState.load(sk_memory_order_acquire);
+        uintptr_t state = sk_atomic_load(&fState, sk_memory_order_acquire);
         if (state < 2) {
             if (state == 0) {
                 // It looks like no one has tried to create our pointer yet.
                 // We try to claim that task by atomically swapping our state from '0' to '1'.
-                if (fState.compare_exchange(&state, 1, sk_memory_order_relaxed,
-                                                       sk_memory_order_relaxed)) {
+                if (sk_atomic_compare_exchange(
+                    &fState, &state, (uintptr_t)1, sk_memory_order_relaxed, sk_memory_order_relaxed)) {
                     // We've claimed it.  Create our pointer and store it into fState.
                     state = (uintptr_t)f();
                     SkASSERT(state > 1);
-                    fState.store(state, sk_memory_order_release);
+                    sk_atomic_store(&fState, state, sk_memory_order_release);
                 } else {
                     // Someone else claimed it.
                     // We fall through to the spin loop just below to wait for them to finish.
@@ -82,7 +82,7 @@ public:
                 // State '1' is our busy-but-not-done state.
                 // Some other thread has claimed the job of creating our pointer.
                 // We just need to wait for it to finish.
-                state = fState.load(sk_memory_order_acquire);
+                state = sk_atomic_load(&fState, sk_memory_order_acquire);
             }
 
             // We shouldn't be able to get here without having created our pointer.
@@ -92,16 +92,15 @@ public:
     }
 
     operator T*() const {
-        auto state = fState.load(sk_memory_order_acquire);
+        auto state = sk_atomic_load(&fState, sk_memory_order_acquire);
         return state < 2 ? nullptr : (T*)state;
         // TODO: If state == 1 spin until it's not?
     }
 
-private:
     // fState == 0 --> we have not created our ptr yet
     // fState == 1 --> someone is in the middle of creating our ptr
     // else        --> (T*)fState is our ptr
-    mutable SkAtomic<uintptr_t> fState;
+    mutable uintptr_t fState;
 };
 
 #endif//SkOncePtr_DEFINED