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

2 files changed, 94 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
diff --git a/include/private/SkSemaphore.h b/include/private/SkSemaphore.h
new file mode 100644
index 0000000000..b5b5ea5957
--- /dev/null
+++ b/include/private/SkSemaphore.h
@@ -0,0 +1,86 @@
+/*
+ * 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 SkSemaphore_DEFINED
+#define SkSemaphore_DEFINED
+
+#include "SkTypes.h"
+#include "SkAtomics.h"
+#include "../private/SkOncePtr.h"
+
+struct SkBaseSemaphore {
+
+    // Increment the counter by 1.
+    // This is a specialization for supporting SkMutex.
+    void signal() {
+        // Since this fetches the value before the add, 0 indicates that this thread is running and
+        // no threads are waiting, -1 and below means that threads are waiting, but only signal 1
+        // thread to run.
+        if (sk_atomic_fetch_add(&fCount, 1, sk_memory_order_release) < 0) {
+           this->osSignal(1);
+        }
+    }
+
+    // Increment the counter N times.
+    // Generally it's better to call signal(N) instead of signal() N times.
+    void signal(int N);
+
+    // Decrement the counter by 1,
+    // then if the counter is <= 0, sleep this thread until the counter is > 0.
+    void wait() {
+        // Since this fetches the value before the subtract, zero and below means that there are no
+        // resources left, so the thread needs to wait.
+        if (sk_atomic_fetch_sub(&fCount, 1, sk_memory_order_acquire) <= 0) {
+            this->osWait();
+        }
+    }
+
+    struct OSSemaphore;
+
+    void osSignal(int n);
+    void osWait();
+    void deleteSemaphore();
+
+    // This implementation follows the general strategy of
+    //     'A Lightweight Semaphore with Partial Spinning'
+    // found here
+    //     http://preshing.com/20150316/semaphores-are-surprisingly-versatile/
+    // That article (and entire blog) are very much worth reading.
+    //
+    // We wrap an OS-provided semaphore with a user-space atomic counter that
+    // lets us avoid interacting with the OS semaphore unless strictly required:
+    // moving the count from >0 to <=0 or vice-versa, i.e. sleeping or waking threads.
+    int                        fCount;
+    SkBaseOncePtr<OSSemaphore> fOSSemaphore;
+};
+
+/**
+ * SkSemaphore is a fast mostly-user-space semaphore.
+ *
+ * A semaphore is logically an atomic integer with a few special properties:
+ *   - The integer always starts at 0.
+ *   - You can only increment or decrement it, never read or write it.
+ *   - Increment is spelled 'signal()'; decrement is spelled 'wait()'.
+ *   - If a call to wait() decrements the counter to <= 0,
+ *     the calling thread sleeps until another thread signal()s it back above 0.
+ */
+class SkSemaphore : SkNoncopyable {
+public:
+    // Initializes the counter to 0.
+    // (Though all current implementations could start from an arbitrary value.)
+    SkSemaphore();
+    ~SkSemaphore();
+
+    void wait();
+
+    void signal(int n = 1);
+
+private:
+    SkBaseSemaphore fBaseSemaphore;
+};
+
+#endif//SkSemaphore_DEFINED