DANGER DANGER. Switching to global critical region.

This changes almost all locks held by guest threads to use a single global
critical region. This emulates the behavior on the PPC of disabling
interrupts (by calls like KeRaiseIrqlToDpcLevel or masking interrupts),
and prevents deadlocks from occuring when threads are suspended or
otherwise blocked.
This has performance implications and a pass is needed to ensure the
locking is as granular as possible. It could also break everything
because it's fundamentally unsound. We'll see.

src/xenia/cpu/backend/x64/x64_code_cache.cc

@@ -129,7 +129,7 @@ void* X64CodeCache::PlaceGuestCode(uint32_t guest_address, void* machine_code,
   uint8_t* code_address = nullptr;
   UnwindReservation unwind_reservation;
   {
-    std::lock_guard<xe::mutex> allocation_lock(allocation_mutex_);
+    auto global_lock = global_critical_region_.Acquire();
 
     low_mark = generated_code_offset_;
 
@@ -192,7 +192,7 @@ uint32_t X64CodeCache::PlaceData(const void* data, size_t length) {
   size_t high_mark;
   uint8_t* data_address = nullptr;
   {
-    std::lock_guard<xe::mutex> allocation_lock(allocation_mutex_);
+    auto global_lock = global_critical_region_.Acquire();
 
     // Reserve code.
     // Always move the code to land on 16b alignment.