Improve queuing for concurrent OSR.

src/optimizing-compiler-thread.h

Index: src/optimizing-compiler-thread.h
diff --git a/src/optimizing-compiler-thread.h b/src/optimizing-compiler-thread.h
index d98a8b2b7e75ea25ee800ce5975c4da15b847a75..f27eeaeefb75095619d2110a7309c64583d8b35f 100644
--- a/src/optimizing-compiler-thread.h
+++ b/src/optimizing-compiler-thread.h
@@ -53,20 +53,11 @@ class OptimizingCompilerThread : public Thread {
       isolate_(isolate),
       stop_semaphore_(0),
       input_queue_semaphore_(0),

[titzer, 2013/10/10 09:41:57]

The external synchronization on these buffers is starting to be a drag. Now that
we are using mutexes again, can we please encapsulate the synchronization into
the buffers somehow?

-      osr_cursor_(0),
+      input_queue_(FLAG_concurrent_recompilation_queue_length),
+      osr_buffer_(FLAG_concurrent_recompilation_queue_length + 4),
       osr_hits_(0),
       osr_attempts_(0) {
     NoBarrier_Store(&stop_thread_, static_cast<AtomicWord>(CONTINUE));
-    NoBarrier_Store(&queue_length_, static_cast<AtomicWord>(0));
-    if (FLAG_concurrent_osr) {
-      osr_buffer_size_ = FLAG_concurrent_recompilation_queue_length + 4;
-      osr_buffer_ = NewArray<RecompileJob*>(osr_buffer_size_);
-      for (int i = 0; i < osr_buffer_size_; i++) osr_buffer_[i] = NULL;
-    }
-  }
-
-  ~OptimizingCompilerThread() {
-    if (FLAG_concurrent_osr) DeleteArray(osr_buffer_);
   }
 
   void Run();
@@ -81,17 +72,15 @@ class OptimizingCompilerThread : public Thread {
   bool IsQueuedForOSR(JSFunction* function);
 
   inline bool IsQueueAvailable() {
-    // We don't need a barrier since we have a data dependency right
-    // after.
-    Atomic32 current_length = NoBarrier_Load(&queue_length_);
-
-    // This can be queried only from the execution thread.
-    ASSERT(!IsOptimizerThread());
-    // Since only the execution thread increments queue_length_ and
-    // only one thread can run inside an Isolate at one time, a direct
-    // doesn't introduce a race -- queue_length_ may decreased in
-    // meantime, but not increased.
-    return (current_length < FLAG_concurrent_recompilation_queue_length);
+    LockGuard<Mutex> access_input(&input_mutex_);
+    return input_queue_.available();
+  }
+
+  inline void DrainOsrBuffer() {

[titzer, 2013/10/10 09:41:57]

I like the idea of dropping one OSR entry per GC, but this method could use a
better name. "Drain" to me implies that it empties the entire buffer.

+    // Advance cursor of the cyclic buffer to next empty slot or stale OSR job.
+    // Dispose said OSR job in the latter case.  Calling this on every GC
+    // should make sure that we do not hold onto stale jobs indefinitely.
+    AddToOsrBuffer(NULL);
   }
 
 #ifdef DEBUG
@@ -109,9 +98,6 @@ class OptimizingCompilerThread : public Thread {
   // Add a recompilation task for OSR to the cyclic buffer, awaiting OSR entry.
   // Tasks evicted from the cyclic buffer are discarded.
   void AddToOsrBuffer(RecompileJob* compiler);
-  void AdvanceOsrCursor() {
-    osr_cursor_ = (osr_cursor_ + 1) % osr_buffer_size_;
-  }
 
 #ifdef DEBUG
   int thread_id_;
@@ -121,21 +107,16 @@ class OptimizingCompilerThread : public Thread {
   Isolate* isolate_;
   Semaphore stop_semaphore_;
   Semaphore input_queue_semaphore_;
+  Mutex input_mutex_;
 
   // Queue of incoming recompilation tasks (including OSR).
-  UnboundQueue<RecompileJob*> input_queue_;
+  CircularQueue<RecompileJob*> input_queue_;
   // Queue of recompilation tasks ready to be installed (excluding OSR).
   UnboundQueue<RecompileJob*> output_queue_;
   // Cyclic buffer of recompilation tasks for OSR.
-  // TODO(yangguo): This may keep zombie tasks indefinitely, holding on to
-  //                a lot of memory.  Fix this.
-  RecompileJob** osr_buffer_;
-  // Cursor for the cyclic buffer.
-  int osr_cursor_;
-  int osr_buffer_size_;
+  CircularBuffer<RecompileJob*> osr_buffer_;
 
   volatile AtomicWord stop_thread_;
-  volatile Atomic32 queue_length_;
   TimeDelta time_spent_compiling_;
   TimeDelta time_spent_total_;