Prefer std::shared_mutex on Windows

.gitignore

@@ -5,6 +5,7 @@ CMakeCache.txt
 CMakeFiles
 Makefile
 *.ninja
+\.cache
 \.ninja_*
 Testing
 build

src/nanothread.cpp

@@ -64,7 +64,7 @@ struct Worker {
 
 
 static Pool *pool_default_inst = nullptr;
-static std::mutex pool_default_lock;
+static Lock pool_default_lock;
 static uint32_t cached_core_count = 0;
 
 uint32_t core_count() {
@@ -128,7 +128,7 @@ uint32_t pool_thread_id() {
 }
 
 Pool *pool_default() {
-    std::unique_lock<std::mutex> guard(pool_default_lock);
+    std::unique_lock<Lock> guard(pool_default_lock);
 
     if (!pool_default_inst)
         pool_default_inst = pool_create();
@@ -159,7 +159,7 @@ void pool_destroy(Pool *pool) {
 
 uint32_t pool_size(Pool *pool) {
     if (!pool) {
-        std::unique_lock<std::mutex> guard(pool_default_lock);
+        std::unique_lock<Lock> guard(pool_default_lock);
         pool = pool_default_inst;
     }
 
@@ -171,7 +171,7 @@ uint32_t pool_size(Pool *pool) {
 
 void pool_set_size(Pool *pool, uint32_t size) {
     if (!pool) {
-        std::unique_lock<std::mutex> guard(pool_default_lock);
+        std::unique_lock<Lock> guard(pool_default_lock);
         pool = pool_default_inst;
 
         if (!pool) {

src/queue.cpp

@@ -397,7 +397,7 @@ std::pair<Task *, uint32_t> TaskQueue::pop() {
 }
 
 void TaskQueue::wakeup() {
-    std::unique_lock<std::mutex> guard(sleep_mutex);
+    std::unique_lock<Lock> guard(sleep_mutex);
     uint64_t value = sleep_state.load();
     NT_TRACE("wakeup(): sleep_state := (%u, 0)", (uint32_t) (sleep_state >> 32) + 1);
     sleep_state = (value + high_bit) & high_mask;
@@ -440,7 +440,7 @@ TaskQueue::pop_or_sleep(bool (*stopping_criterion)(void *), void *payload,
         attempts++;
 
         if (may_sleep && attempts >= NANOTHREAD_MAX_ATTEMPTS) {
-            std::unique_lock<std::mutex> guard(sleep_mutex);
+            std::unique_lock<Lock> guard(sleep_mutex);
 
             uint64_t value = ++sleep_state, phase = value & high_mask;
             NT_TRACE("pop_or_sleep(): falling asleep after %.2f milliseconds, "

src/queue.h

@@ -11,14 +11,19 @@
 
 #include <atomic>
 #include <vector>
-#include <mutex>
 #include <condition_variable>
 #include <cstring>
 
 #if defined(_WIN32)
 #  include <windows.h>
+#  include <shared_mutex>
+using Lock = std::shared_mutex; // Prefer (more efficient) shared_mutex on Windows
+#else
+#  include <mutex>
+using Lock = std::mutex;
 #endif
 
+
 struct Pool;
 
 constexpr uint64_t high_bit  = (uint64_t) 0x0000000100000000ull;
@@ -241,7 +246,7 @@ struct TaskQueue {
     std::atomic<uint64_t> sleep_state;
 
     /// Mutex protecting the fields below
-    std::mutex sleep_mutex;
+    Lock sleep_mutex;
 
     /// Condition variable used to manage workers that are asleep
     std::condition_variable sleep_cv;