Sketch of SK_ONCE

BUG=
R=bungeman@google.com

Review URL: https://codereview.chromium.org/26563002

git-svn-id: http://skia.googlecode.com/svn/trunk/src@11674 2bbb7eff-a529-9590-31e7-b0007b416f81
diff --git a/core/SkOnce.h b/core/SkOnce.h
new file mode 100644
index 0000000..85921a8
--- /dev/null
+++ b/core/SkOnce.h
@@ -0,0 +1,160 @@
+/*
+ * Copyright 2013 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkOnce_DEFINED
+#define SkOnce_DEFINED
+
+// SkOnce.h defines two macros, DEF_SK_ONCE and SK_ONCE.
+// You can use these macros together to create a threadsafe block of code that
+// runs at most once, no matter how many times you call it.  This is
+// particularly useful for lazy singleton initialization.  E.g.
+//
+// DEF_SK_ONCE(set_up_my_singleton, SingletonType* singleton) {
+//   // Code in this block will run at most once.
+//   *singleton = new Singleton(...);
+// }
+// ...
+// const Singleton& getSingleton() {
+//   static Singleton* singleton = NULL;
+//   // Always call SK_ONCE.  It's very cheap to call after the first time.
+//   SK_ONCE(set_up_my_singleton, singleton);
+//   SkASSERT(NULL != singleton);
+//   return *singleton;
+// }
+//
+// OnceTest.cpp also should serve as another simple example.
+
+#include "SkThread.h"
+#include "SkTypes.h"
+
+
+// Pass a unique name (at least in this scope) for name, and a type and name
+// for arg (as if writing a function declaration).
+// E.g.
+//   DEF_SK_ONCE(my_onetime_setup, int* foo) {
+//     *foo += 5;
+//   }
+#define DEF_SK_ONCE(name, arg)                       \
+    static bool sk_once_##name##_done = false;       \
+    SK_DECLARE_STATIC_MUTEX(sk_once_##name##_mutex); \
+    static void sk_once_##name##_function(arg)
+
+// Call this anywhere you need to guarantee that the corresponding DEF_SK_ONCE
+// block of code has run.  name should match the DEF_SK_ONCE, and here you pass
+// the actual value of the argument.
+// E.g
+//   int foo = 0;
+//   SK_ONCE(my_onetime_setup, &foo);
+//   SkASSERT(5 == foo);
+#define SK_ONCE(name, arg) \
+    sk_once_impl(&sk_once_##name##_done, &sk_once_##name##_mutex, sk_once_##name##_function, arg)
+
+
+//  ----------------------  Implementation details below here. -----------------------------
+
+
+// TODO(bungeman, mtklein): move all these *barrier* functions to SkThread when refactoring lands.
+
+#ifdef SK_BUILD_FOR_WIN
+#include <intrin.h>
+inline static void compiler_barrier() {
+    _ReadWriteBarrier();
+}
+#else
+inline static void compiler_barrier() {
+    asm volatile("" : : : "memory");
+}
+#endif
+
+inline static void full_barrier_on_arm() {
+#ifdef SK_CPU_ARM
+    asm volatile("dmb" : : : "memory");
+#endif
+}
+
+// On every platform, we issue a compiler barrier to prevent it from reordering
+// code.  That's enough for platforms like x86 where release and acquire
+// barriers are no-ops.  On other platforms we may need to be more careful;
+// ARM, in particular, needs real code for both acquire and release.  We use a
+// full barrier, which acts as both, because that the finest precision ARM
+// provides.
+
+inline static void release_barrier() {
+    compiler_barrier();
+    full_barrier_on_arm();
+}
+
+inline static void acquire_barrier() {
+    compiler_barrier();
+    full_barrier_on_arm();
+}
+
+// We've pulled a pretty standard double-checked locking implementation apart
+// into its main fast path and a slow path that's called when we suspect the
+// one-time code hasn't run yet.
+
+// This is the guts of the code, called when we suspect the one-time code hasn't been run yet.
+// This should be rarely called, so we separate it from sk_once_impl and don't mark it as inline.
+// (We don't mind if this is an actual function call, but odds are it'll be inlined anyway.)
+template <typename Arg>
+static void sk_once_impl_slow(bool* done, SkBaseMutex* mutex, void (*once)(Arg), Arg arg) {
+    const SkAutoMutexAcquire lock(*mutex);
+    if (!*done) {
+        once(arg);
+        // Also known as a store-store/load-store barrier, this makes sure that the writes
+        // done before here---in particular, those done by calling once(arg)---are observable
+        // before the writes after the line, *done = true.
+        //
+        // In version control terms this is like saying, "check in the work up
+        // to and including once(arg), then check in *done=true as a subsequent change".
+        //
+        // We'll use this in the fast path to make sure once(arg)'s effects are
+        // observable whenever we observe *done == true.
+        release_barrier();
+        *done = true;
+    }
+}
+
+// We nabbed this code from the dynamic_annotations library, and in their honor
+// we check the same define.  If you find yourself wanting more than just
+// ANNOTATE_BENIGN_RACE, it might make sense to pull that in as a dependency
+// rather than continue to reproduce it here.
+
+#ifdef DYNAMIC_ANNOTATIONS_ENABLED
+// TSAN provides this hook to supress a known-safe apparent race.
+extern "C" {
+void AnnotateBenignRace(const char* file, int line, const volatile void* mem, const char* desc);
+}
+#define ANNOTATE_BENIGN_RACE(mem, desc) AnnotateBenignRace(__FILE__, __LINE__, mem, desc)
+#else
+#define ANNOTATE_BENIGN_RACE(mem, desc)
+#endif
+
+// This is our fast path, called all the time.  We do really want it to be inlined.
+template <typename Arg>
+inline static void sk_once_impl(bool* done, SkBaseMutex* mutex, void (*once)(Arg), Arg arg) {
+    ANNOTATE_BENIGN_RACE(done, "Don't worry TSAN, we're sure this is safe.");
+    if (!*done) {
+        sk_once_impl_slow(done, mutex, once, arg);
+    }
+    // Also known as a load-load/load-store barrier, this acquire barrier makes
+    // sure that anything we read from memory---in particular, memory written by
+    // calling once(arg)---is at least as current as the value we read from done.
+    //
+    // In version control terms, this is a lot like saying "sync up to the
+    // commit where we wrote *done = true".
+    //
+    // The release barrier in sk_once_impl_slow guaranteed that *done = true
+    // happens after once(arg), so by syncing to *done = true here we're
+    // forcing ourselves to also wait until the effects of once(arg) are readble.
+    acquire_barrier();
+}
+
+#undef ANNOTATE_BENIGN_RACE
+
+
+#endif  // SkOnce_DEFINED