Sketch of SK_ONCE

src/core/SkOnce.h

+#ifndef SkOnce_DEFINED
+#define SkOnce_DEFINED
+
+// TODO: docs
+
+#include "SkThread.h"
+#include "SkTypes.h"
+
+#define DEF_SK_ONCE(name, arg)                     \
+    static bool sk_once##name##ready = false;      \

[bungeman-skia, 2013/10/08 22:34:59]

nit: can we put some '__' before and after name? I just like readable symbols,
even if they're auto generated.

[mtklein, 2013/10/09 15:22:25]

You betcha.  Done.  Took the opportunity to rename ready to done, as it sort of
makes more sense in the do-this-once context.

+    SK_DECLARE_STATIC_MUTEX(sk_once##name##mutex); \
+    static void sk_once##name##function(arg)
+
+
+inline static void compiler_barrier() {

[bungeman-skia, 2013/10/08 22:34:59]

note that on Windows this is _ReadWriteBarrier, not that tsan seems to be there
yet.

[mtklein, 2013/10/09 15:22:25]

Done.

+    asm volatile("" ::: "memory");
+}
+
+inline static void full_barrier_on_arm() {
+#ifdef SK_CPU_ARM
+    asm volatile("dmb" ::: "memory");

[bungeman-skia, 2013/10/08 22:34:59]

note that android has android_memory_barrier and android_compiler_barrier in
cutils/atomic-inline.h (of course, these can only be used with the framework).

[mtklein, 2013/10/09 15:22:25]

Yeah.  Given that ARM doesn't quite imply Android, and especially that Android
definitely doesn't imply ARM, I want to keep this CPU specific for now rather
than going the roundabout way of checking the OS then having that code check the
CPU.  I'm happy to let SkThread handle this once that's suitably refactored.

+#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() {

[bungeman-skia, 2013/10/08 22:34:59]

of course, all of this goes in some future cleaned up SkThread.h :-)

[mtklein, 2013/10/09 15:22:25]

Agreed.  I've added a note to remind us.

+    compiler_barrier();
+    full_barrier_on_arm();
+}
+
+// This should be rarely called, so we separate it from sk_once_impl and don't mark it as inline.
+template <typename Arg>
+static void sk_once_impl_slow(bool* ready, SkBaseMutex* mutex, void (*once)(Arg), Arg arg) {
+    const SkAutoMutexAcquire lock(*mutex);
+    if (!*ready) {
+        once(arg);
+        release_barrier();

[bungeman-skia, 2013/10/08 22:34:59]

I know that we know why these barriers are where they are, but it would be nice
to have a short comment here about what the barrier is doing (ensuring ready is
set to true after once) and maybe a larger comment above with the entire thing,
since we're splitting this into two for the speed. I suppose that's part of the
TODO: docs.

[mtklein, 2013/10/09 15:22:25]

Yeah, that was the TODO.  How's this read to you now?

+        *ready = true;
+    }
+}
+
+// Thread sanitizer provides a hook to stifle a known safe race.
+extern "C" {
+void AnnotateBenignRace(const char* file, int line, const volatile void* mem, const char* desc);
+}
+// We need to define it as a no-op if we're not using TSAN
+#if SK_HAS_COMPILER_FEATURE(thread_sanitizer)

[bungeman-skia, 2013/10/08 22:34:59]

SK_HAS_COMPILER_FEATURE is only available on clang, and newer gccs support tsan.
I think we need an actual flag for this.

[mtklein, 2013/10/09 15:22:25]

Done.

+#  define ANNOTATE_BENIGN_RACE(file, line, mem, desc) AnnotateBenignRace(file, line, mem, desc)
+#else
+#  define ANNOTATE_BENIGN_RACE(file, line, mem, desc)
+#endif
+
+template <typename Arg>
+inline static void sk_once_impl(bool* ready, SkBaseMutex* mutex, void (*once)(Arg), Arg arg) {
+    ANNOTATE_BENIGN_RACE(__FILE__, __LINE__, ready, "Don't worry TSAN, we're sure this is safe.");
+    if (!*ready) {
+        sk_once_impl_slow(ready, mutex, once, arg);
+    }
+    acquire_barrier();

[bungeman-skia, 2013/10/08 22:34:59]

A comment like "Ensure that we see the effects of 'once' as having run" or
something like that.

[mtklein, 2013/10/09 15:22:25]

Done.

+}
+
+#undef ANNOTATE_BENIGN_RACE
+
+#define SK_ONCE(name, arg) \
+    sk_once_impl(&sk_once##name##ready, &sk_once##name##mutex, sk_once##name##function, arg)
+
+#endif  // SkOnce_DEFINED