Revert of Modernize SkMutex and SkSemaphore. (patchset #2 id:20001 of https://codereview.chromium.org/1947153002/ )

Reason for revert:
This CL seems to have broken the Linux x64 and Mac
bots, e.g.

https://build.chromium.org/p/chromium/builders/Linux%20x64/builds/19052
https://build.chromium.org/p/chromium/builders/Mac/builds/15151

The error appears to have something to do with new static initializers being
added.

Original issue's description:
> Modernize SkMutex and SkSemaphore.
>
>    - use <atomic>
>    - fuse SkMutex and SkBaseMutex
>    - fuse SkSemaphore and SkBaseSemaphore
>
> Still TODO:
>    - replace SK_DECLARE_STATIC_MUTEX(name) with static SkMutex name
>
> I just didn't want to bother fixing all that up until I know this CL sticks.
>
> BUG=skia:
> GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1947153002
>
> No public API changes.
> TBR=reed@google.com
>
> Committed: https://skia.googlesource.com/skia/+/427c2819d9237d7d7729c59238036cfc73c072ea

TBR=herb@google.com,mtklein@chromium.org,reed@google.com,bsalomon@google.com
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=609340

Review-Url: https://codereview.chromium.org/1945343003

7 files changed, 142 insertions, 78 deletions

diff --git a/include/core/SkPixelRef.h b/include/core/SkPixelRef.h
index d836b2fd7c..90e0de540d 100644
--- a/include/core/SkPixelRef.h
+++ b/include/core/SkPixelRef.h
@@ -342,7 +342,7 @@ protected:
     /** Return the mutex associated with this pixelref. This value is assigned
         in the constructor, and cannot change during the lifetime of the object.
     */
-    SkMutex* mutex() const { return &fMutex; }
+    SkBaseMutex* mutex() const { return &fMutex; }
 
     // only call from constructor. Flags this to always be locked, removing
     // the need to grab the mutex and call onLockPixels/onUnlockPixels.
diff --git a/include/ports/SkFontConfigInterface.h b/include/ports/SkFontConfigInterface.h
index 9f98e356d2..72cad0ad6b 100644
--- a/include/ports/SkFontConfigInterface.h
+++ b/include/ports/SkFontConfigInterface.h
@@ -13,6 +13,7 @@
 #include "SkRefCnt.h"
 #include "SkTypeface.h"
 
+struct SkBaseMutex;
 class SkFontMgr;
 
 /**
diff --git a/include/private/SkMutex.h b/include/private/SkMutex.h
index b9af00f8a8..8c78e1205c 100644
--- a/include/private/SkMutex.h
+++ b/include/private/SkMutex.h
@@ -8,20 +8,33 @@
 #ifndef SkMutex_DEFINED
 #define SkMutex_DEFINED
 
+// This file is not part of the public Skia API.
 #include "../private/SkSemaphore.h"
-#include "../private/SkThreadID.h"
 #include "SkTypes.h"
 
-// TODO: no need for this anymore.
-#define SK_DECLARE_STATIC_MUTEX(name) static SkMutex name;
-
-class SkMutex {
-public:
-    constexpr SkMutex() = default;
-
-    SkMutex(const SkMutex&)            = delete;
-    SkMutex& operator=(const SkMutex&) = delete;
-
+#ifdef SK_DEBUG
+    #include "../private/SkThreadID.h"
+#endif
+
+#define SK_MUTEX_SEMAPHORE_INIT {1, {0}}
+
+#ifdef SK_DEBUG
+    #define SK_BASE_MUTEX_INIT {SK_MUTEX_SEMAPHORE_INIT, 0}
+#else
+    #define SK_BASE_MUTEX_INIT {SK_MUTEX_SEMAPHORE_INIT}
+#endif
+
+// Using POD-style initialization prevents the generation of a static initializer.
+//
+// Without magic statics there are no thread safety guarantees on initialization
+// of local statics (even POD). As a result, it is illegal to use
+// SK_DECLARE_STATIC_MUTEX in a function.
+//
+// Because SkBaseMutex is not a primitive, a static SkBaseMutex cannot be
+// initialized in a class with this macro.
+#define SK_DECLARE_STATIC_MUTEX(name) namespace {} static SkBaseMutex name = SK_BASE_MUTEX_INIT;
+
+struct SkBaseMutex {
     void acquire() {
         fSemaphore.wait();
         SkDEBUGCODE(fOwner = SkGetThreadID();)
@@ -37,9 +50,20 @@ public:
         SkASSERT(fOwner == SkGetThreadID());
     }
 
-private:
-    SkSemaphore fSemaphore{1};
-    SkDEBUGCODE(SkThreadID fOwner{kIllegalThreadID};)
+    SkBaseSemaphore fSemaphore;
+    SkDEBUGCODE(SkThreadID fOwner;)
+};
+
+// This needs to use subclassing instead of encapsulation to make SkAutoMutexAcquire to work.
+class SkMutex : public SkBaseMutex {
+public:
+    SkMutex () {
+        fSemaphore = SK_MUTEX_SEMAPHORE_INIT;
+        SkDEBUGCODE(fOwner = kIllegalThreadID);
+    }
+    ~SkMutex () { fSemaphore.deleteSemaphore(); }
+    SkMutex(const SkMutex&) = delete;
+    SkMutex& operator=(const SkMutex&) = delete;
 };
 
 template <typename Lock>
@@ -92,10 +116,10 @@ private:
     Lock &fLock;
 };
 
-typedef SkAutoTAcquire<SkMutex> SkAutoMutexAcquire;
+typedef SkAutoTAcquire<SkBaseMutex> SkAutoMutexAcquire;
 #define SkAutoMutexAcquire(...) SK_REQUIRE_LOCAL_VAR(SkAutoMutexAcquire)
 
-typedef SkAutoTExclusive<SkMutex> SkAutoMutexExclusive;
+typedef SkAutoTExclusive<SkBaseMutex> SkAutoMutexExclusive;
 #define SkAutoMutexExclusive(...) SK_REQUIRE_LOCAL_VAR(SkAutoMutexExclusive)
 
 #endif//SkMutex_DEFINED
diff --git a/include/private/SkSemaphore.h b/include/private/SkSemaphore.h
index adee574bb7..cb2f58da4a 100644
--- a/include/private/SkSemaphore.h
+++ b/include/private/SkSemaphore.h
@@ -8,26 +8,43 @@
 #ifndef SkSemaphore_DEFINED
 #define SkSemaphore_DEFINED
 
-#include "../private/SkOnce.h"
 #include "SkTypes.h"
-#include <atomic>
+#include "../private/SkAtomics.h"
+#include "../private/SkOncePtr.h"
 
-class SkSemaphore {
-public:
-    constexpr SkSemaphore(int count = 0)
-        : fCount(count), fOSSemaphore(nullptr) {}
+struct SkBaseSemaphore {
 
-    ~SkSemaphore();
+    // 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 = 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();
+    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();
 
-private:
     // This implementation follows the general strategy of
     //     'A Lightweight Semaphore with Partial Spinning'
     // found here
@@ -37,40 +54,33 @@ private:
     // 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.
-    struct OSSemaphore;
-
-    void osSignal(int n);
-    void osWait();
-
-    std::atomic<int> fCount;
-    SkOnce           fOSSemaphoreOnce;
-    OSSemaphore*     fOSSemaphore;
+    int                        fCount;
+    SkBaseOncePtr<OSSemaphore> fOSSemaphore;
 };
 
-inline void SkSemaphore::signal(int n) {
-    int prev = fCount.fetch_add(n, std::memory_order_release);
+/**
+ * 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();
 
-    // We only want to call the OS semaphore when our logical count crosses
-    // from <= 0 to >0 (when we need to wake sleeping threads).
-    //
-    // This is easiest to think about with specific examples of prev and n.
-    // If n == 5 and prev == -3, there are 3 threads sleeping and we signal
-    // SkTMin(-(-3), 5) == 3 times on the OS semaphore, leaving the count at 2.
-    //
-    // If prev >= 0, no threads are waiting, SkTMin(-prev, n) is always <= 0,
-    // so we don't call the OS semaphore, leaving the count at (prev + n).
-    int toSignal = SkTMin(-prev, n);
-    if (toSignal > 0) {
-        this->osSignal(toSignal);
-    }
-}
+    void wait();
 
-inline void SkSemaphore::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 (fCount.fetch_sub(1, std::memory_order_acquire) <= 0) {
-        this->osWait();
-    }
-}
+    void signal(int n = 1);
+
+private:
+    SkBaseSemaphore fBaseSemaphore;
+};
 
 #endif//SkSemaphore_DEFINED
diff --git a/src/core/SkSemaphore.cpp b/src/core/SkSemaphore.cpp
index 0646b152e8..da422e282f 100644
--- a/src/core/SkSemaphore.cpp
+++ b/src/core/SkSemaphore.cpp
@@ -9,7 +9,7 @@
 
 #if defined(SK_BUILD_FOR_MAC) || defined(SK_BUILD_FOR_IOS)
     #include <mach/mach.h>
-    struct SkSemaphore::OSSemaphore {
+    struct SkBaseSemaphore::OSSemaphore {
         semaphore_t fSemaphore;
 
         OSSemaphore()  {
@@ -21,7 +21,7 @@
         void wait() { semaphore_wait(fSemaphore); }
     };
 #elif defined(SK_BUILD_FOR_WIN32)
-    struct SkSemaphore::OSSemaphore {
+    struct SkBaseSemaphore::OSSemaphore {
         HANDLE fSemaphore;
 
         OSSemaphore()  {
@@ -41,7 +41,7 @@
     // It's important we test for Mach before this.  This code will compile but not work there.
     #include <errno.h>
     #include <semaphore.h>
-    struct SkSemaphore::OSSemaphore {
+    struct SkBaseSemaphore::OSSemaphore {
         sem_t fSemaphore;
 
         OSSemaphore()  { sem_init(&fSemaphore, 0/*cross process?*/, 0/*initial count*/); }
@@ -57,16 +57,43 @@
 
 ///////////////////////////////////////////////////////////////////////////////
 
-void SkSemaphore::osSignal(int n) {
-    fOSSemaphoreOnce([this] { fOSSemaphore = new OSSemaphore; });
-    fOSSemaphore->signal(n);
+void SkBaseSemaphore::signal(int n) {
+    SkASSERT(n >= 0);
+
+    // We only want to call the OS semaphore when our logical count crosses
+    // from <= 0 to >0 (when we need to wake sleeping threads).
+    //
+    // This is easiest to think about with specific examples of prev and n.
+    // If n == 5 and prev == -3, there are 3 threads sleeping and we signal
+    // SkTMin(-(-3), 5) == 3 times on the OS semaphore, leaving the count at 2.
+    //
+    // If prev >= 0, no threads are waiting, SkTMin(-prev, n) is always <= 0,
+    // so we don't call the OS semaphore, leaving the count at (prev + n).
+    int prev = sk_atomic_fetch_add(&fCount, n, sk_memory_order_release);
+    int toSignal = SkTMin(-prev, n);
+    if (toSignal > 0) {
+        this->osSignal(toSignal);
+    }
 }
 
-void SkSemaphore::osWait() {
-    fOSSemaphoreOnce([this] { fOSSemaphore = new OSSemaphore; });
-    fOSSemaphore->wait();
+static SkBaseSemaphore::OSSemaphore* semaphore(SkBaseSemaphore* semaphore) {
+    return semaphore->fOSSemaphore.get([](){ return new SkBaseSemaphore::OSSemaphore(); });
 }
 
-SkSemaphore::~SkSemaphore() {
-    delete fOSSemaphore;
+void SkBaseSemaphore::osSignal(int n) { semaphore(this)->signal(n); }
+
+void SkBaseSemaphore::osWait() { semaphore(this)->wait(); }
+
+void SkBaseSemaphore::deleteSemaphore() {
+    delete (OSSemaphore*) fOSSemaphore;
 }
+
+///////////////////////////////////////////////////////////////////////////////
+
+SkSemaphore::SkSemaphore(){ fBaseSemaphore = {0, {0}}; }
+
+SkSemaphore::~SkSemaphore() { fBaseSemaphore.deleteSemaphore(); }
+
+void SkSemaphore::wait() { fBaseSemaphore.wait(); }
+
+void SkSemaphore::signal(int n) {fBaseSemaphore.signal(n); }
diff --git a/src/lazy/SkDiscardableMemoryPool.cpp b/src/lazy/SkDiscardableMemoryPool.cpp
index 2be4c755f1..1f3bcf93e9 100644
--- a/src/lazy/SkDiscardableMemoryPool.cpp
+++ b/src/lazy/SkDiscardableMemoryPool.cpp
@@ -29,7 +29,7 @@ public:
     /**
      *  Without mutex, will be not be thread safe.
      */
-    DiscardableMemoryPool(size_t budget, SkMutex* mutex = nullptr);
+    DiscardableMemoryPool(size_t budget, SkBaseMutex* mutex = nullptr);
     virtual ~DiscardableMemoryPool();
 
     SkDiscardableMemory* create(size_t bytes) override;
@@ -52,9 +52,9 @@ public:
     #endif  // SK_LAZY_CACHE_STATS
 
 private:
-    SkMutex* fMutex;
-    size_t   fBudget;
-    size_t   fUsed;
+    SkBaseMutex* fMutex;
+    size_t       fBudget;
+    size_t       fUsed;
     SkTInternalLList<PoolDiscardableMemory> fList;
 
     /** Function called to free memory if needed */
@@ -128,7 +128,8 @@ void PoolDiscardableMemory::unlock() {
 
 ////////////////////////////////////////////////////////////////////////////////
 
-DiscardableMemoryPool::DiscardableMemoryPool(size_t budget, SkMutex* mutex)
+DiscardableMemoryPool::DiscardableMemoryPool(size_t budget,
+                                             SkBaseMutex* mutex)
     : fMutex(mutex)
     , fBudget(budget)
     , fUsed(0) {
@@ -240,7 +241,7 @@ void DiscardableMemoryPool::dumpPool() {
 
 }  // namespace
 
-SkDiscardableMemoryPool* SkDiscardableMemoryPool::Create(size_t size, SkMutex* mutex) {
+SkDiscardableMemoryPool* SkDiscardableMemoryPool::Create(size_t size, SkBaseMutex* mutex) {
     return new DiscardableMemoryPool(size, mutex);
 }
 
diff --git a/src/lazy/SkDiscardableMemoryPool.h b/src/lazy/SkDiscardableMemoryPool.h
index ad8d796f66..92ba48bcb4 100644
--- a/src/lazy/SkDiscardableMemoryPool.h
+++ b/src/lazy/SkDiscardableMemoryPool.h
@@ -52,7 +52,8 @@ public:
      *  the pool works.
      *  Without mutex, will be not be thread safe.
      */
-    static SkDiscardableMemoryPool* Create(size_t size, SkMutex* mutex = nullptr);
+    static SkDiscardableMemoryPool* Create(
+            size_t size, SkBaseMutex* mutex = nullptr);
 };
 
 /**