core: Create a thread for each CPU core, keep in lock-step with a barrier.

src/core/core_cpu.h

@@ -4,7 +4,9 @@
 
 #pragma once
 
+#include <condition_variable>
 #include <memory>
+#include <mutex>
 #include <string>
 #include "common/common_types.h"
 
@@ -16,9 +18,32 @@ class Scheduler;
 
 namespace Core {
 
+constexpr unsigned NUM_CPU_CORES{4};
+
+class CpuBarrier {
+public:
+    void Rendezvous() {
+        std::unique_lock<std::mutex> lock(mutex);
+
+        --cores_waiting;
+        if (!cores_waiting) {
+            cores_waiting = NUM_CPU_CORES;
+            condition.notify_all();
+            return;
+        }
+
+        condition.wait(lock);
+    }
+
+private:
+    unsigned cores_waiting{NUM_CPU_CORES};
+    std::mutex mutex;
+    std::condition_variable condition;
+};
+
 class Cpu {
 public:
-    Cpu();
+    Cpu(std::shared_ptr<CpuBarrier> cpu_barrier, size_t core_index);
 
     void RunLoop(bool tight_loop = true);
 
@@ -34,13 +59,19 @@ public:
         return *scheduler;
     }
 
+    bool IsMainCore() const {
+        return core_index == 0;
+    }
+
 private:
     void Reschedule();
 
     std::shared_ptr<ARM_Interface> arm_interface;
+    std::shared_ptr<CpuBarrier> cpu_barrier;
     std::unique_ptr<Kernel::Scheduler> scheduler;
 
     bool reschedule_pending{};
+    size_t core_index;
 };
 
 } // namespace Core