diff --git a/src/xenia/base/atomic.h b/src/xenia/base/atomic.h
index 5ea64de82..cd2e3b2f0 100644
--- a/src/xenia/base/atomic.h
+++ b/src/xenia/base/atomic.h
@@ -66,6 +66,14 @@ inline bool atomic_cas(int64_t old_value, int64_t new_value,
              old_value) == old_value;
 }
 
+// Atomic store with release semantics (for releasing locks)
+inline void atomic_store_release(int32_t new_value, volatile int32_t* value) {
+  _InterlockedExchange(reinterpret_cast<volatile long*>(value), new_value);
+}
+inline void atomic_store_release(uint32_t new_value, volatile uint32_t* value) {
+  _InterlockedExchange(reinterpret_cast<volatile long*>(value), new_value);
+}
+
 #elif XE_PLATFORM_LINUX || XE_PLATFORM_MAC
 
 inline int32_t atomic_inc(volatile int32_t* value) {
@@ -85,10 +93,26 @@ inline int32_t atomic_xor(volatile int32_t* value, int32_t nv) {
 }
 
 inline int32_t atomic_exchange(int32_t new_value, volatile int32_t* value) {
-  return __sync_val_compare_and_swap(value, *value, new_value);
+  int32_t old_value;
+  do {
+    old_value = *value;
+  } while (!__sync_bool_compare_and_swap(value, old_value, new_value));
+  return old_value;
 }
 inline int64_t atomic_exchange(int64_t new_value, volatile int64_t* value) {
-  return __sync_val_compare_and_swap(value, *value, new_value);
+  int64_t old_value;
+  do {
+    old_value = *value;
+  } while (!__sync_bool_compare_and_swap(value, old_value, new_value));
+  return old_value;
+}
+
+// Atomic store with release semantics (for releasing locks)
+inline void atomic_store_release(int32_t new_value, volatile int32_t* value) {
+  __atomic_store_n(value, new_value, __ATOMIC_RELEASE);
+}
+inline void atomic_store_release(uint32_t new_value, volatile uint32_t* value) {
+  __atomic_store_n(value, new_value, __ATOMIC_RELEASE);
 }
 
 inline int32_t atomic_exchange_add(int32_t amount, volatile int32_t* value) {
diff --git a/src/xenia/base/threading.h b/src/xenia/base/threading.h
index d2d30998e..b7ec709af 100644
--- a/src/xenia/base/threading.h
+++ b/src/xenia/base/threading.h
@@ -83,8 +83,8 @@ class Fence {
   std::mutex mutex_;
   std::condition_variable cond_;
   // Use the highest bit (sign bit) as the signal flag and the rest to count
-  // waiting threads.
-  volatile state_t_ signal_state_;
+  // waiting threads. Protected by mutex_.
+  state_t_ signal_state_;
 };
 
 // Returns the total number of logical processors in the host system.
diff --git a/src/xenia/base/threading_posix.cc b/src/xenia/base/threading_posix.cc
index b768c6872..1b2c988c1 100644
--- a/src/xenia/base/threading_posix.cc
+++ b/src/xenia/base/threading_posix.cc
@@ -546,7 +546,7 @@ class PosixCondition<Timer> final : public PosixConditionBase {
   }
   std::weak_ptr<TimerQueueWaitItem> wait_item_;
   std::function<void()> callback_;
-  volatile bool signal_;
+  bool signal_;  // Protected by mutex_
   const bool manual_reset_;
 };
 
@@ -820,6 +820,7 @@ class PosixCondition<Thread> final : public PosixConditionBase {
     bool is_current_thread = pthread_self() == thread_;
 
     {
+      std::unique_lock lock(state_mutex_);
       if (out_previous_suspend_count) {
         *out_previous_suspend_count = suspend_count_;
       }
@@ -908,8 +909,8 @@ class PosixCondition<Thread> final : public PosixConditionBase {
   pthread_t thread_;
   bool signaled_;
   int exit_code_;
-  volatile State state_;
-  volatile uint32_t suspend_count_;
+  State state_;             // Protected by state_mutex_
+  uint32_t suspend_count_;  // Protected by state_mutex_
   mutable std::mutex state_mutex_;
   mutable std::mutex callback_mutex_;
   mutable std::condition_variable state_signal_;
diff --git a/src/xenia/kernel/xboxkrnl/xboxkrnl_threading.cc b/src/xenia/kernel/xboxkrnl/xboxkrnl_threading.cc
index f284f88bc..56c0b735d 100644
--- a/src/xenia/kernel/xboxkrnl/xboxkrnl_threading.cc
+++ b/src/xenia/kernel/xboxkrnl/xboxkrnl_threading.cc
@@ -1167,8 +1167,8 @@ DECLARE_XBOXKRNL_EXPORT3(KfAcquireSpinLock, kThreading, kImplemented, kBlocking,
 void xeKeKfReleaseSpinLock(PPCContext* ctx, X_KSPINLOCK* lock,
                            uint32_t old_irql, bool change_irql) {
   assert_true(lock->prcb_of_owner == static_cast<uint32_t>(ctx->r[13]));
-  // Unlock.
-  lock->prcb_of_owner.value = 0;
+  // Unlock with release semantics to ensure all prior writes are visible.
+  xe::atomic_store_release(0u, &lock->prcb_of_owner.value);
 
   if (change_irql) {
     // Unlock.