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

Review-Url: https://codereview.chromium.org/1947153002
diff --git a/include/ports/SkFontConfigInterface.h b/include/ports/SkFontConfigInterface.h
index 72cad0a..9f98e35 100644
--- a/include/ports/SkFontConfigInterface.h
+++ b/include/ports/SkFontConfigInterface.h
@@ -13,7 +13,6 @@
 #include "SkRefCnt.h"
 #include "SkTypeface.h"
 
-struct SkBaseMutex;
 class SkFontMgr;
 
 /**
diff --git a/include/private/SkMutex.h b/include/private/SkMutex.h
index 8c78e12..3b0e1c4 100644
--- a/include/private/SkMutex.h
+++ b/include/private/SkMutex.h
@@ -8,33 +8,16 @@
 #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"
 
-#ifdef SK_DEBUG
-    #include "../private/SkThreadID.h"
-#endif
+#define SK_DECLARE_STATIC_MUTEX(name) static SkBaseMutex name;
 
-#define SK_MUTEX_SEMAPHORE_INIT {1, {0}}
+class SkBaseMutex {
+public:
+    constexpr SkBaseMutex() = default;
 
-#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();)
@@ -50,20 +33,15 @@
         SkASSERT(fOwner == SkGetThreadID());
     }
 
-    SkBaseSemaphore fSemaphore;
-    SkDEBUGCODE(SkThreadID fOwner;)
+protected:
+    SkBaseSemaphore fSemaphore{1};
+    SkDEBUGCODE(SkThreadID fOwner{kIllegalThreadID};)
 };
 
-// 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;
+    using SkBaseMutex::SkBaseMutex;
+    ~SkMutex() { fSemaphore.cleanup(); }
 };
 
 template <typename Lock>
diff --git a/include/private/SkSemaphore.h b/include/private/SkSemaphore.h
index cb2f58d..3da2b99 100644
--- a/include/private/SkSemaphore.h
+++ b/include/private/SkSemaphore.h
@@ -8,43 +8,27 @@
 #ifndef SkSemaphore_DEFINED
 #define SkSemaphore_DEFINED
 
+#include "../private/SkOnce.h"
 #include "SkTypes.h"
-#include "../private/SkAtomics.h"
-#include "../private/SkOncePtr.h"
+#include <atomic>
 
-struct SkBaseSemaphore {
+class SkBaseSemaphore {
+public:
+    constexpr SkBaseSemaphore(int count = 0)
+        : fCount(count), fOSSemaphore(nullptr) {}
 
-    // 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);
+    // Increment the counter n times.
+    // Generally it's better to call signal(n) instead of signal() n times.
+    void signal(int n = 1);
 
     // 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();
-        }
-    }
+    void wait();
 
-    struct OSSemaphore;
+    // SkBaseSemaphore has no destructor.  Call this to clean it up.
+    void cleanup();
 
-    void osSignal(int n);
-    void osWait();
-    void deleteSemaphore();
-
+private:
     // This implementation follows the general strategy of
     //     'A Lightweight Semaphore with Partial Spinning'
     // found here
@@ -54,33 +38,46 @@
     // 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;
+    struct OSSemaphore;
+
+    void osSignal(int n);
+    void osWait();
+
+    std::atomic<int> fCount;
+    SkOnce           fOSSemaphoreOnce;
+    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 {
+class SkSemaphore : public SkBaseSemaphore {
 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;
+    using SkBaseSemaphore::SkBaseSemaphore;
+    ~SkSemaphore() { this->cleanup(); }
 };
 
+inline void SkBaseSemaphore::signal(int n) {
+    int prev = fCount.fetch_add(n, std::memory_order_release);
+
+    // 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);
+    }
+}
+
+inline void SkBaseSemaphore::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();
+    }
+}
+
 #endif//SkSemaphore_DEFINED
diff --git a/src/core/SkSemaphore.cpp b/src/core/SkSemaphore.cpp
index da422e2..2286c0b 100644
--- a/src/core/SkSemaphore.cpp
+++ b/src/core/SkSemaphore.cpp
@@ -57,43 +57,16 @@
 
 ///////////////////////////////////////////////////////////////////////////////
 
-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 SkBaseSemaphore::osSignal(int n) {
+    fOSSemaphoreOnce([this] { fOSSemaphore = new OSSemaphore; });
+    fOSSemaphore->signal(n);
 }
 
-static SkBaseSemaphore::OSSemaphore* semaphore(SkBaseSemaphore* semaphore) {
-    return semaphore->fOSSemaphore.get([](){ return new SkBaseSemaphore::OSSemaphore(); });
+void SkBaseSemaphore::osWait() {
+    fOSSemaphoreOnce([this] { fOSSemaphore = new OSSemaphore; });
+    fOSSemaphore->wait();
 }
 
-void SkBaseSemaphore::osSignal(int n) { semaphore(this)->signal(n); }
-
-void SkBaseSemaphore::osWait() { semaphore(this)->wait(); }
-
-void SkBaseSemaphore::deleteSemaphore() {
-    delete (OSSemaphore*) fOSSemaphore;
+void SkBaseSemaphore::cleanup() {
+    delete fOSSemaphore;
 }
-
-///////////////////////////////////////////////////////////////////////////////
-
-SkSemaphore::SkSemaphore(){ fBaseSemaphore = {0, {0}}; }
-
-SkSemaphore::~SkSemaphore() { fBaseSemaphore.deleteSemaphore(); }
-
-void SkSemaphore::wait() { fBaseSemaphore.wait(); }
-
-void SkSemaphore::signal(int n) {fBaseSemaphore.signal(n); }