Improve queuing for concurrent OSR.

src/optimizing-compiler-thread.cc

Index: src/optimizing-compiler-thread.cc
diff --git a/src/optimizing-compiler-thread.cc b/src/optimizing-compiler-thread.cc
index 202e6e5c57e0b2ac2624f369c704344300219e44..acd378e968ca8ef90f73b2b72b7c8d336babb4c4 100644
--- a/src/optimizing-compiler-thread.cc
+++ b/src/optimizing-compiler-thread.cc
@@ -95,10 +95,9 @@ void OptimizingCompilerThread::Run() {
 
 void OptimizingCompilerThread::CompileNext() {
   RecompileJob* job = NULL;
-  bool result = input_queue_.Dequeue(&job);
-  USE(result);
-  ASSERT(result);
-  Barrier_AtomicIncrement(&queue_length_, static_cast<Atomic32>(-1));
+  { LockGuard<Mutex> access_input(&input_mutex_);
+    job = input_queue_.Dequeue();
+  }
 
   // The function may have already been optimized by OSR.  Simply continue.
   RecompileJob::Status status = job->OptimizeGraph();
@@ -131,7 +130,11 @@ static void DisposeRecompileJob(RecompileJob* job,
 
 void OptimizingCompilerThread::FlushInputQueue(bool restore_function_code) {
   RecompileJob* job;
-  while (input_queue_.Dequeue(&job)) {
+  while (true) {
+    { LockGuard<Mutex> access_input(&input_mutex_);
+      if (input_queue_.length() == 0) break;
+      job = input_queue_.Dequeue();
+    }
     // This should not block, since we have one signal on the input queue
     // semaphore corresponding to each element in the input queue.
     input_queue_semaphore_.Wait();
@@ -140,7 +143,6 @@ void OptimizingCompilerThread::FlushInputQueue(bool restore_function_code) {
       DisposeRecompileJob(job, restore_function_code);
     }
   }
-  Release_Store(&queue_length_, static_cast<AtomicWord>(0));
 }
 
 
@@ -157,11 +159,10 @@ void OptimizingCompilerThread::FlushOutputQueue(bool restore_function_code) {
 
 void OptimizingCompilerThread::FlushOsrBuffer(bool restore_function_code) {
   RecompileJob* job;
-  for (int i = 0; i < osr_buffer_size_; i++) {
-    job = osr_buffer_[i];
+  for (int i = 0; i < osr_buffer_.capacity(); i++) {
+    job = osr_buffer_.Remove(i);
     if (job != NULL) DisposeRecompileJob(job, restore_function_code);
   }
-  osr_cursor_ = 0;
 }
 
 
@@ -185,10 +186,9 @@ void OptimizingCompilerThread::Stop() {
   stop_semaphore_.Wait();
 
   if (FLAG_concurrent_recompilation_delay != 0) {
-    // Barrier when loading queue length is not necessary since the write
-    // happens in CompileNext on the same thread.
-    // This is used only for testing.
-    while (NoBarrier_Load(&queue_length_) > 0) CompileNext();
+    // At this point the optimizing compiler threa's event loop has stopped.
+    // There is no need for a mutex when reading input_queue_.
+    while (input_queue_.length() > 0) CompileNext();
     InstallOptimizedFunctions();
   } else {
     FlushInputQueue(false);
@@ -237,7 +237,6 @@ void OptimizingCompilerThread::InstallOptimizedFunctions() {
 void OptimizingCompilerThread::QueueForOptimization(RecompileJob* job) {
   ASSERT(IsQueueAvailable());
   ASSERT(!IsOptimizerThread());
-  Barrier_AtomicIncrement(&queue_length_, static_cast<Atomic32>(1));
   CompilationInfo* info = job->info();
   if (info->is_osr()) {
     if (FLAG_trace_concurrent_recompilation) {
@@ -245,13 +244,18 @@ void OptimizingCompilerThread::QueueForOptimization(RecompileJob* job) {
       info->closure()->PrintName();
       PrintF(" for concurrent on-stack replacement.\n");
     }
-    AddToOsrBuffer(job);
     osr_attempts_++;
     BackEdgeTable::AddStackCheck(info);
+    AddToOsrBuffer(job);
+    { LockGuard<Mutex> access_input(&input_mutex_);
+      input_queue_.Prepend(job);
+    }
   } else {
     info->closure()->MarkInRecompileQueue();
+    { LockGuard<Mutex> access_input(&input_mutex_);
+      input_queue_.Enqueue(job);
+    }
   }
-  input_queue_.Enqueue(job);
   input_queue_semaphore_.Signal();
 }
 
@@ -259,14 +263,13 @@ void OptimizingCompilerThread::QueueForOptimization(RecompileJob* job) {
 RecompileJob* OptimizingCompilerThread::FindReadyOSRCandidate(
     Handle<JSFunction> function, uint32_t osr_pc_offset) {
   ASSERT(!IsOptimizerThread());
-  RecompileJob* result = NULL;
-  for (int i = 0; i < osr_buffer_size_; i++) {
-    result = osr_buffer_[i];
+  for (int i = 0; i < osr_buffer_.capacity(); i++) {
+    RecompileJob* result = osr_buffer_.Get(i);
     if (result == NULL) continue;
     if (result->IsWaitingForInstall() &&
         result->info()->HasSameOsrEntry(function, osr_pc_offset)) {
       osr_hits_++;
-      osr_buffer_[i] = NULL;
+      osr_buffer_.Remove(i);
       return result;
     }
   }
@@ -277,10 +280,11 @@ RecompileJob* OptimizingCompilerThread::FindReadyOSRCandidate(
 bool OptimizingCompilerThread::IsQueuedForOSR(Handle<JSFunction> function,
                                               uint32_t osr_pc_offset) {
   ASSERT(!IsOptimizerThread());
-  for (int i = 0; i < osr_buffer_size_; i++) {
-    if (osr_buffer_[i] != NULL &&
-        osr_buffer_[i]->info()->HasSameOsrEntry(function, osr_pc_offset)) {
-      return !osr_buffer_[i]->IsWaitingForInstall();
+  for (int i = 0; i < osr_buffer_.capacity(); i++) {
+    RecompileJob* result = osr_buffer_.Get(i);
+    if (result != NULL &&
+        result->info()->HasSameOsrEntry(function, osr_pc_offset)) {
+      return !result->IsWaitingForInstall();
     }
   }
   return false;
@@ -289,10 +293,10 @@ bool OptimizingCompilerThread::IsQueuedForOSR(Handle<JSFunction> function,
 
 bool OptimizingCompilerThread::IsQueuedForOSR(JSFunction* function) {
   ASSERT(!IsOptimizerThread());
-  for (int i = 0; i < osr_buffer_size_; i++) {
-    if (osr_buffer_[i] != NULL &&
-        *osr_buffer_[i]->info()->closure() == function) {
-      return !osr_buffer_[i]->IsWaitingForInstall();
+  for (int i = 0; i < osr_buffer_.capacity(); i++) {
+    RecompileJob* result = osr_buffer_.Get(i);
+    if (result != NULL && *result->info()->closure() == function) {
+      return !result->IsWaitingForInstall();
     }
   }
   return false;
@@ -301,27 +305,25 @@ bool OptimizingCompilerThread::IsQueuedForOSR(JSFunction* function) {
 
 void OptimizingCompilerThread::AddToOsrBuffer(RecompileJob* job) {
   ASSERT(!IsOptimizerThread());
-  // Store into next empty slot or replace next stale OSR job that's waiting
-  // in vain.  Dispose in the latter case.
-  RecompileJob* stale;
+  // Find the next slot that is empty or has a stale job.
   while (true) {
-    stale = osr_buffer_[osr_cursor_];
-    if (stale == NULL) break;
-    if (stale->IsWaitingForInstall()) {
-      CompilationInfo* info = stale->info();
-      if (FLAG_trace_osr) {
-        PrintF("[COSR - Discarded ");
-        info->closure()->PrintName();
-        PrintF(", AST id %d]\n", info->osr_ast_id().ToInt());
-      }
-      DisposeRecompileJob(stale, false);
-      break;
-    }
-    AdvanceOsrCursor();
+    RecompileJob* stale = osr_buffer_.Current();
+    if (stale == NULL || stale->IsWaitingForInstall()) break;
+    osr_buffer_.AdvanceCursor();
   }
 
-  osr_buffer_[osr_cursor_] = job;
-  AdvanceOsrCursor();
+  // Add to found slot and dispose the evicted job.
+  RecompileJob* evicted = osr_buffer_.Add(job);
+  if (evicted != NULL) {
+    ASSERT(evicted->IsWaitingForInstall());
+    CompilationInfo* info = evicted->info();
+    if (FLAG_trace_osr) {
+      PrintF("[COSR - Discarded ");
+      info->closure()->PrintName();
+      PrintF(", AST id %d]\n", info->osr_ast_id().ToInt());
+    }
+    DisposeRecompileJob(evicted, false);
+  }
 }