--- a/src/parallel_execution/main.c	Wed Jan 25 04:14:50 2017 +0900
+++ b/src/parallel_execution/main.c	Wed Jan 25 16:55:53 2017 +0900
@@ -33,7 +33,7 @@
     loopCounter->tree = &createRedBlackTree(context)->Tree;
     loopCounter->i = 0;
     taskManager->taskManager = (union Data*)&createTaskManager(context, cpu_num, 0, 0)->TaskManager;
-    goto meta(context, C_createData1);
+    goto meta(context, C_createTask1);
 }
 
 __code initDataGears_stub(struct Context* context) {
@@ -51,7 +51,8 @@
     /* puts("result"); */
 
     time->next = C_code2;
-    goto meta(context, C_start_time);
+    goto meta(context, C_exit_code);
+    //goto meta(context, C_start_time);
 }
 
 __code code1_stub(struct Context* context) {
@@ -153,9 +154,9 @@
 
         goto meta(context, taskManager->taskManager->TaskManager.spawn);
     }
-
     loopCounter->i = 0;
-    goto meta(context, C_code1);
+    taskManager->next = C_code1;
+    goto meta(context, taskManager->taskManager->TaskManager.shutdown);
 }
 
 __code createTask2_stub(struct Context* context) {
@@ -167,18 +168,6 @@
             array);
 }
 
-__code taskManager(struct Context* context, struct LoopCounter* loopCounter, struct Worker* worker) {
-    loopCounter->i = 0;
-
-    Time *t = &context->data[D_Time]->Time;
-    t->next = C_code2;
-    goto meta(context, C_end_time);
-}
-
-__code taskManager_stub(struct Context* context) {
-    goto taskManager(context, &context->data[D_LoopCounter]->LoopCounter, &context->data[D_Worker]->Worker);
-}
-
 void init(int argc, char** argv) {
     for (int i = 1; argv[i]; ++i) {
         if (strcmp(argv[i], "-cpu") == 0)

--- a/src/parallel_execution/taskManager.c	Wed Jan 25 04:14:50 2017 +0900
+++ b/src/parallel_execution/taskManager.c	Wed Jan 25 16:55:53 2017 +0900
@@ -109,7 +109,7 @@
 __code taskSend(struct Context* context, TaskManagerImpl* taskManager, Queue* queue) {
     queue->queue = (union Data*)taskManager->activeQueue;
     queue->next = C_taskSend1;
-    goto meta(context, taskManager->activeQueue->put);
+    goto meta(context, taskManager->activeQueue->take);
 }
 
 __code taskSend_stub(struct Context* context) {
@@ -131,12 +131,16 @@
     goto taskSend1(context, taskManager, Gearef(context, Queue), Gearef(context, TaskManager)->context, Gearef(context, TaskManager)->next);
 }
 
-__code shutdownTaskManager(struct Context* context, struct LoopCounter* loopCounter, TaskManager* taskManager, TaskManagerImpl* taskManagerImpl) {
+__code shutdownTaskManager(struct Context* context, struct LoopCounter* loopCounter, TaskManager* taskManager, TaskManagerImpl* taskManagerImpl, Queue* queue, enum Code next) {
     int i = loopCounter->i;
     if (taskManager->cpu <= i && i < taskManager->maxCPU) {
+        struct Queue* tasks = taskManagerImpl->workers[i]->tasks;
+        queue->queue = (union Data*)tasks;
+        queue->data = NULL;
+        queue->next = next;
+        goto meta(context, tasks->put);
         pthread_join(taskManagerImpl->workers[i]->worker->CPUWorker.thread, NULL);
         loopCounter->i++;
-
         goto meta(context, C_shutdownTaskManager);
     }
 
@@ -146,5 +150,5 @@
 
 __code shutdownTaskManager_stub(struct Context* context) {
     TaskManagerImpl* taskManagerImpl = (TaskManagerImpl*)GearImpl(context, TaskManager, taskManager);
-    goto shutdownTaskManager(context, Gearef(context, LoopCounter), Gearef(context, TaskManager), taskManagerImpl);
+    goto shutdownTaskManager(context, Gearef(context, LoopCounter), Gearef(context, TaskManager), taskManagerImpl, Gearef(context, Queue), Gearef(context, TaskManager)->next);
 }

--- a/src/parallel_execution/twice.c	Wed Jan 25 04:14:50 2017 +0900
+++ b/src/parallel_execution/twice.c	Wed Jan 25 16:55:53 2017 +0900
@@ -3,9 +3,8 @@
 #include "context.h"
 #include "origin_cs.h"
 
-__code twice(struct Context* context, struct LoopCounter* loopCounter, int index, int prefix, int* array, enum Code next) {
+__code twice(struct Context* context, struct LoopCounter* loopCounter, int index, int prefix, int* array, Context* workerContext) {
     int i = loopCounter->i;
-
     if (i < prefix) {
         array[i+index*prefix] = array[i+index*prefix]*2;
         loopCounter->i++;
@@ -14,11 +13,10 @@
     }
 
     loopCounter->i = 0;
-
-    //stack_pop(context->code_stack, &context->next);
-    goto meta(context, next);
+    goto meta(workerContext, workerContext->next);
 }
 
 __code twice_stub(struct Context* context) {
-    goto twice(context, NULL, 0, 0, NULL, 0);
+    Context* workerContext = context->worker->worker->CPUWorker.context;
+    goto twice(context, Gearef(context, LoopCounter), 0, 0, NULL, workerContext);
 }

--- a/src/parallel_execution/worker.c	Wed Jan 25 04:14:50 2017 +0900
+++ b/src/parallel_execution/worker.c	Wed Jan 25 16:55:53 2017 +0900
@@ -22,31 +22,33 @@
     cpuWorker->context = NEW(struct Context);
     initContext(cpuWorker->context);
     Gearef(cpuWorker->context, Worker)->worker = (union Data*)worker;
-    pthread_cond_wait(&cpuWorker->cond, &cpuWorker->mutex);
     goto meta(cpuWorker->context, C_taskReceiveWorker);
 }
 
 __code taskReceiveWorker(struct Context* context, Worker* worker, Queue* queue) {
     queue->queue = (union Data*)worker->tasks;
-    queue->next = C_getTask1;
-    goto meta(context, queue->take);
+    queue->next = C_getTask;
+    goto meta(context, worker->tasks->take);
 }
 
 __code taskReceiveWorker_stub(struct Context* context) {
+    CPUWorker* cpuWorker = (CPUWorker *)GearImpl(context, Worker, worker);
+    pthread_cond_wait(&cpuWorker->cond, &cpuWorker->mutex);
     goto taskReceiveWorker(context, &Gearef(context, Worker)->worker->Worker, Gearef(context, Queue));
 }
 
-__code getTask1(struct Context* context, Worker* worker, struct Context* task) {
+__code getTask(struct Context* context, Worker* worker, struct Context* task) {
     if (!task)
         return; // end thread
     task->worker = worker;
+    context->next = C_taskReceiveWorker; // set CG after task exec
     goto meta(task, task->next);
 }
 
-__code getTask1_stub(struct Context* context) {
+__code getTask_stub(struct Context* context) {
     Worker* worker = &Gearef(context,Worker)->worker->Worker;
     Context* task = &Gearef(context, Queue)->data->context;
-    goto getTask1(context, worker, task);
+    goto getTask(context, worker, task);
 }
 
 #ifdef USE_CUDA