2 files changed, 40 insertions, 12 deletions

diff --git a/src/core/hle/kernel/svc.cpp b/src/core/hle/kernel/svc.cpp
index 6b2995fe2..7b41c9cfd 100644
--- a/src/core/hle/kernel/svc.cpp
+++ b/src/core/hle/kernel/svc.cpp
@@ -650,12 +650,27 @@ static ResultCode SignalProcessWideKey(VAddr condition_variable_addr, s32 target
 
         ASSERT(thread->condvar_wait_address == condition_variable_addr);
 
-        // If the mutex is not yet acquired, acquire it.
-        u32 mutex_val = Memory::Read32(thread->mutex_wait_address);
+        size_t current_core = Core::System::GetInstance().CurrentCoreIndex();
+
+        auto& monitor = Core::System::GetInstance().Monitor();
+
+        // Atomically read the value of the mutex.
+        u32 mutex_val = 0;
+        do {
+            monitor.SetExclusive(current_core, thread->mutex_wait_address);
+
+            // If the mutex is not yet acquired, acquire it.
+            mutex_val = Memory::Read32(thread->mutex_wait_address);
+
+            if (mutex_val != 0) {
+                monitor.ClearExclusive();
+                break;
+            }
+        } while (!monitor.ExclusiveWrite32(current_core, thread->mutex_wait_address,
+                                           thread->wait_handle));
 
         if (mutex_val == 0) {
             // We were able to acquire the mutex, resume this thread.
-            Memory::Write32(thread->mutex_wait_address, thread->wait_handle);
             ASSERT(thread->status == ThreadStatus::WaitMutex);
             thread->ResumeFromWait();
 
@@ -668,7 +683,19 @@ static ResultCode SignalProcessWideKey(VAddr condition_variable_addr, s32 target
             thread->condvar_wait_address = 0;
             thread->wait_handle = 0;
         } else {
-            // Couldn't acquire the mutex, block the thread.
+            // Atomically signal that the mutex now has a waiting thread.
+            do {
+                monitor.SetExclusive(current_core, thread->mutex_wait_address);
+
+                // Ensure that the mutex value is still what we expect.
+                u32 value = Memory::Read32(thread->mutex_wait_address);
+                // TODO(Subv): When this happens, the kernel just clears the exclusive state and
+                // retries the initial read for this thread.
+                ASSERT_MSG(mutex_val == value, "Unhandled synchronization primitive case");
+            } while (!monitor.ExclusiveWrite32(current_core, thread->mutex_wait_address,
+                                               mutex_val | Mutex::MutexHasWaitersFlag));
+
+            // The mutex is already owned by some other thread, make this thread wait on it.
             Handle owner_handle = static_cast<Handle>(mutex_val & Mutex::MutexOwnerMask);
             auto owner = g_handle_table.Get<Thread>(owner_handle);
             ASSERT(owner);
@@ -676,9 +703,6 @@ static ResultCode SignalProcessWideKey(VAddr condition_variable_addr, s32 target
             thread->status = ThreadStatus::WaitMutex;
             thread->wakeup_callback = nullptr;
 
-            // Signal that the mutex now has a waiting thread.
-            Memory::Write32(thread->mutex_wait_address, mutex_val | Mutex::MutexHasWaitersFlag);
-
             owner->AddMutexWaiter(thread);
 
             Core::System::GetInstance().CpuCore(thread->processor_id).PrepareReschedule();
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index f495b623b..c7e3fb4b1 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -425,6 +425,7 @@ union Instruction {
 
     union {
         BitField<50, 3, u64> component_mask_selector;
+        BitField<0, 8, Register> gpr0;
         BitField<28, 8, Register> gpr28;
 
         bool HasTwoDestinations() const {
@@ -432,13 +433,16 @@ union Instruction {
         }
 
         bool IsComponentEnabled(size_t component) const {
-            static constexpr std::array<size_t, 5> one_dest_mask{0x1, 0x2, 0x4, 0x8, 0x3};
-            static constexpr std::array<size_t, 5> two_dest_mask{0x7, 0xb, 0xd, 0xe, 0xf};
-            const auto& mask{HasTwoDestinations() ? two_dest_mask : one_dest_mask};
+            static constexpr std::array<std::array<u32, 8>, 4> mask_lut{
+                {{},
+                 {0x1, 0x2, 0x4, 0x8, 0x3},
+                 {0x1, 0x2, 0x4, 0x8, 0x3, 0x9, 0xa, 0xc},
+                 {0x7, 0xb, 0xd, 0xe, 0xf}}};
 
-            ASSERT(component_mask_selector < mask.size());
+            size_t index{gpr0.Value() != Register::ZeroIndex ? 1U : 0U};
+            index |= gpr28.Value() != Register::ZeroIndex ? 2 : 0;
 
-            return ((1ull << component) & mask[component_mask_selector]) != 0;
+            return ((1ull << component) & mask_lut[index][component_mask_selector]) != 0;
         }
     } texs;