WIP: lock-free work stealing deque based on our stack.

app/benchmark_fft/main.cpp

@@ -6,7 +6,7 @@
 #include <complex>
 #include <complex>
 #include <vector>
 #include <vector>
 
 
-static constexpr int CUTOFF = 10;
+static constexpr int CUTOFF = 16;
 static constexpr int NUM_ITERATIONS = 1000;
 static constexpr int NUM_ITERATIONS = 1000;
 static constexpr int INPUT_SIZE = 2064;
 static constexpr int INPUT_SIZE = 2064;
 typedef std::vector<std::complex<double>> complex_vector;
 typedef std::vector<std::complex<double>> complex_vector;

app/invoke_parallel/main.cpp

@@ -2,48 +2,100 @@
 #include <pls/internal/helpers/profiler.h>
 #include <pls/internal/helpers/profiler.h>
 
 
 #include <iostream>
 #include <iostream>
+#include <complex>
+#include <vector>
 
 
-static pls::static_scheduler_memory<8, 2 << 14> my_scheduler_memory;
+static constexpr int CUTOFF = 16;
+static constexpr int NUM_ITERATIONS = 1000;
+static constexpr int INPUT_SIZE = 2064;
+typedef std::vector<std::complex<double>> complex_vector;
 
 
-static constexpr int CUTOFF = 10;
-
-long fib_serial(long n) {
-  if (n == 0) {
-    return 0;
+void divide(complex_vector::iterator data, int n) {
+  complex_vector tmp_odd_elements(n / 2);
+  for (int i = 0; i < n / 2; i++) {
+    tmp_odd_elements[i] = data[i * 2 + 1];
+  }
+  for (int i = 0; i < n / 2; i++) {
+    data[i] = data[i * 2];
   }
   }
-  if (n == 1) {
-    return 1;
+  for (int i = 0; i < n / 2; i++) {
+    data[i + n / 2] = tmp_odd_elements[i];
   }
   }
+}
+
+void combine(complex_vector::iterator data, int n) {
+  for (int i = 0; i < n / 2; i++) {
+    std::complex<double> even = data[i];
+    std::complex<double> odd = data[i + n / 2];
 
 
-  return fib_serial(n - 1) + fib_serial(n - 2);
+    // w is the "twiddle-factor".
+    // this could be cached, but we run the same 'data_structures' algorithm parallel/serial,
+    // so it won't impact the performance comparison.
+    std::complex<double> w = exp(std::complex<double>(0, -2. * M_PI * i / n));
+
+    data[i] = even + w * odd;
+    data[i + n / 2] = even - w * odd;
+  }
 }
 }
 
 
-long fib(long n) {
-  if (n <= CUTOFF) {
-    return fib_serial(n);
+void fft(complex_vector::iterator data, int n) {
+  if (n < 2) {
+    return;
   }
   }
 
 
-  // Actual 'invoke_parallel' logic/code
-  int left, right;
+  PROFILE_WORK_BLOCK("Divide")
+  divide(data, n);
+  PROFILE_END_BLOCK
+  PROFILE_WORK_BLOCK("Invoke Parallel")
+  if (n == CUTOFF) {
+    PROFILE_WORK_BLOCK("FFT Serial")
+    fft(data, n / 2);
+    fft(data + n / 2, n / 2);
+  } else if (n <= CUTOFF) {
+    fft(data, n / 2);
+    fft(data + n / 2, n / 2);
+  } else {
     pls::invoke_parallel(
     pls::invoke_parallel(
-      [&] { left = fib(n - 1); },
-      [&] { right = fib(n - 2); }
+        [&] { fft(data, n / 2); },
+        [&] { fft(data + n / 2, n / 2); }
     );
     );
-  return left + right;
+  }
+  PROFILE_END_BLOCK
+  PROFILE_WORK_BLOCK("Combine")
+  combine(data, n);
+  PROFILE_END_BLOCK
+}
+
+complex_vector prepare_input(int input_size) {
+  std::vector<double> known_frequencies{2, 11, 52, 88, 256};
+  complex_vector data(input_size);
+
+  // Set our input data to match a time series of the known_frequencies.
+  // When applying fft to this time-series we should find these frequencies.
+  for (int i = 0; i < input_size; i++) {
+    data[i] = std::complex<double>(0.0, 0.0);
+    for (auto frequencie : known_frequencies) {
+      data[i] += sin(2 * M_PI * frequencie * i / input_size);
+    }
+  }
+
+  return data;
 }
 }
 
 
 int main() {
 int main() {
   PROFILE_ENABLE
   PROFILE_ENABLE
+  pls::malloc_scheduler_memory my_scheduler_memory{8, 2u << 14};
   pls::scheduler scheduler{&my_scheduler_memory, 8};
   pls::scheduler scheduler{&my_scheduler_memory, 8};
 
 
-  long result;
+  complex_vector initial_input = prepare_input(INPUT_SIZE);
   scheduler.perform_work([&] {
   scheduler.perform_work([&] {
     PROFILE_MAIN_THREAD
     PROFILE_MAIN_THREAD
     // Call looks just the same, only requirement is
     // Call looks just the same, only requirement is
     // the enclosure in the perform_work lambda.
     // the enclosure in the perform_work lambda.
     for (int i = 0; i < 10; i++) {
     for (int i = 0; i < 10; i++) {
-      result = fib(30);
-      std::cout << "Fib(30)=" << result << std::endl;
+      PROFILE_WORK_BLOCK("Top Level FFT")
+      complex_vector input = initial_input;
+      fft(input.begin(), input.size());
     }
     }
   });
   });
 
 

app/playground/main.cpp

@@ -11,8 +11,5 @@
 #include <pls/internal/helpers/unique_id.h>
 #include <pls/internal/helpers/unique_id.h>
 
 
 int main() {
 int main() {
-  std::cout << pls::internal::scheduling::root_task<void (*)>::create_id().type_.hash_code() << std::endl;
-  std::cout
-      << pls::internal::helpers::unique_id::create<pls::internal::scheduling::root_task<void (*)>>().type_.hash_code()
-      << std::endl;
+
 }
 }

lib/pls/CMakeLists.txt

@@ -20,6 +20,7 @@ add_library(pls STATIC
         include/pls/internal/data_structures/aligned_stack.h src/internal/data_structures/aligned_stack.cpp
         include/pls/internal/data_structures/aligned_stack.h src/internal/data_structures/aligned_stack.cpp
         include/pls/internal/data_structures/aligned_stack_impl.h
         include/pls/internal/data_structures/aligned_stack_impl.h
         include/pls/internal/data_structures/deque.h src/internal/data_structures/deque.cpp
         include/pls/internal/data_structures/deque.h src/internal/data_structures/deque.cpp
+        include/pls/internal/data_structures/work_stealing_deque.h
 
 
         include/pls/internal/helpers/prohibit_new.h
         include/pls/internal/helpers/prohibit_new.h
         include/pls/internal/helpers/profiler.h
         include/pls/internal/helpers/profiler.h

lib/pls/include/pls/internal/base/alignment.h

@@ -19,10 +19,32 @@ struct aligned_wrapper {
 };
 };
 void *allocate_aligned(size_t size);
 void *allocate_aligned(size_t size);
 
 
-std::uintptr_t next_alignment(std::uintptr_t size);
+system_details::pointer_t next_alignment(system_details::pointer_t size);
+system_details::pointer_t previous_alignment(system_details::pointer_t size);
 char *next_alignment(char *pointer);
 char *next_alignment(char *pointer);
 
 
 }
 }
+
+template<typename T>
+struct aligned_aba_pointer {
+  const system_details::pointer_t pointer_;
+
+  explicit aligned_aba_pointer(T *pointer, unsigned int aba = 0) : pointer_{
+      reinterpret_cast<system_details::pointer_t >(pointer) + aba} {}
+
+  T *pointer() const {
+    return reinterpret_cast<T *>(pointer_ & system_details::CACHE_LINE_ADDRESS_USED_BITS);
+  }
+
+  unsigned int aba() const {
+    return pointer_ & system_details::CACHE_LINE_ADDRESS_UNUSED_BITS;
+  }
+
+  aligned_aba_pointer set_aba(unsigned int aba) const {
+    return aligned_aba_pointer(pointer(), aba);
+  }
+};
+
 }
 }
 }
 }
 }
 }

lib/pls/include/pls/internal/base/backoff.h

@@ -14,8 +14,8 @@ namespace internal {
 namespace base {
 namespace base {
 
 
 class backoff {
 class backoff {
-  static constexpr unsigned long INITIAL_SPIN_ITERS = 2u << 2u;
-  static constexpr unsigned long MAX_SPIN_ITERS = 2u << 6u;
+  const unsigned long INITIAL_SPIN_ITERS = 2u << 2u;
+  const unsigned long MAX_SPIN_ITERS = 2u << 6u;
 
 
   unsigned long current_ = INITIAL_SPIN_ITERS;
   unsigned long current_ = INITIAL_SPIN_ITERS;
   std::minstd_rand random_;
   std::minstd_rand random_;

lib/pls/include/pls/internal/base/error_handling.h

@@ -11,5 +11,6 @@
  * (or its inclusion adds too much overhead).
  * (or its inclusion adds too much overhead).
  */
  */
 #define PLS_ERROR(msg) std::cout << msg << std::endl; exit(1);
 #define PLS_ERROR(msg) std::cout << msg << std::endl; exit(1);
+#define PLS_ASSERT(cond, msg) if (!cond) { PLS_ERROR(msg) }
 
 
 #endif //PLS_ERROR_HANDLING_H
 #endif //PLS_ERROR_HANDLING_H

lib/pls/include/pls/internal/base/system_details.h

@@ -18,10 +18,32 @@ namespace base {
  * Currently sane default values for x86.
  * Currently sane default values for x86.
  */
  */
 namespace system_details {
 namespace system_details {
+
+/**
+ * Pointer Types needed for ABA protection mixed into addresses.
+ * pointer_t should be an integer type capable of holding ANY pointer value.
+ */
+using pointer_t = std::uintptr_t;
+constexpr pointer_t ZERO_POINTER = 0;
+constexpr pointer_t MAX_POINTER = ~ZERO_POINTER;
+
+/**
+ * Biggest type that supports atomic CAS operations.
+ * Usually it is sane to assume a pointer can be swapped in a single CAS operation.
+ */
+using cas_integer = pointer_t;
+constexpr cas_integer MIN_CAS_INTEGER = 0;
+constexpr cas_integer MAX_CAS_INTEGER = ~MIN_CAS_INTEGER;
+constexpr cas_integer FIRST_HALF_CAS_INTEGER = MAX_CAS_INTEGER << ((sizeof(cas_integer) / 2) * 8);
+constexpr cas_integer SECOND_HALF_CAS_INTEGER = ~FIRST_HALF_CAS_INTEGER;
+
 /**
 /**
- * Most processors have 64 byte cache lines
+ * Most processors have 64 byte cache lines (last 6 bit of the address are zero at line beginnings).
  */
  */
-constexpr std::uintptr_t CACHE_LINE_SIZE = 64;
+constexpr unsigned int CACHE_LINE_ADDRESS_BITS = 6;
+constexpr pointer_t CACHE_LINE_SIZE = 2u << (CACHE_LINE_ADDRESS_BITS - 1);
+constexpr pointer_t CACHE_LINE_ADDRESS_USED_BITS = MAX_POINTER << CACHE_LINE_ADDRESS_BITS;
+constexpr pointer_t CACHE_LINE_ADDRESS_UNUSED_BITS = ~CACHE_LINE_ADDRESS_USED_BITS;
 
 
 /**
 /**
  * Choose one of the following ways to store thread specific data.
  * Choose one of the following ways to store thread specific data.

lib/pls/include/pls/internal/base/ttas_spin_lock.h

@@ -19,11 +19,10 @@ namespace base {
  */
  */
 class ttas_spin_lock {
 class ttas_spin_lock {
   std::atomic<int> flag_;
   std::atomic<int> flag_;
-  backoff backoff_;
 
 
  public:
  public:
-  ttas_spin_lock() : flag_{0}, backoff_{} {};
-  ttas_spin_lock(const ttas_spin_lock &/*other*/) : flag_{0}, backoff_{} {}
+  ttas_spin_lock() : flag_{0} {};
+  ttas_spin_lock(const ttas_spin_lock &/*other*/) : flag_{0} {}
 
 
   void lock();
   void lock();
   bool try_lock(unsigned int num_tries = 1);
   bool try_lock(unsigned int num_tries = 1);

lib/pls/include/pls/internal/data_structures/aligned_stack.h

@@ -12,6 +12,8 @@ namespace pls {
 namespace internal {
 namespace internal {
 namespace data_structures {
 namespace data_structures {
 
 
+using base::system_details::pointer_t;
+
 /**
 /**
  * Generic stack-like data structure that allows to allocate arbitrary objects in a given memory region.
  * Generic stack-like data structure that allows to allocate arbitrary objects in a given memory region.
  * The objects will be stored aligned in the stack, making the storage cache friendly and very fast
  * The objects will be stored aligned in the stack, making the storage cache friendly and very fast
@@ -26,15 +28,16 @@ namespace data_structures {
  */
  */
 class aligned_stack {
 class aligned_stack {
   // Keep bounds of our memory block
   // Keep bounds of our memory block
-  char *memory_start_;
-  char *memory_end_;
+  pointer_t memory_start_;
+  pointer_t memory_end_;
 
 
   // Current head will always be aligned to cache lines
   // Current head will always be aligned to cache lines
-  char *head_;
+  pointer_t head_;
  public:
  public:
-  typedef char *state;
+  typedef pointer_t state;
 
 
-  aligned_stack() : memory_start_{nullptr}, memory_end_{nullptr}, head_{nullptr} {};
+  aligned_stack() : memory_start_{0}, memory_end_{0}, head_{0} {};
+  aligned_stack(pointer_t memory_region, std::size_t size);
   aligned_stack(char *memory_region, std::size_t size);
   aligned_stack(char *memory_region, std::size_t size);
 
 
   template<typename T>
   template<typename T>
@@ -48,7 +51,6 @@ class aligned_stack {
   void reset_state(state new_state) { head_ = new_state; }
   void reset_state(state new_state) { head_ = new_state; }
 };
 };
 
 
-
 }
 }
 }
 }
 }
 }

lib/pls/include/pls/internal/data_structures/aligned_stack_impl.h

@@ -9,7 +9,7 @@ namespace data_structures {
 template<typename T>
 template<typename T>
 T *aligned_stack::push(const T &object) {
 T *aligned_stack::push(const T &object) {
   // Copy-Construct
   // Copy-Construct
-  return new((void *) push < T > ())T(object);
+  return new(push < T > ())T(object);
 }
 }
 
 
 template<typename T>
 template<typename T>

lib/pls/include/pls/internal/scheduling/fork_join_task.h

@@ -5,7 +5,7 @@
 #include "pls/internal/helpers/profiler.h"
 #include "pls/internal/helpers/profiler.h"
 
 
 #include "pls/internal/data_structures/aligned_stack.h"
 #include "pls/internal/data_structures/aligned_stack.h"
-#include "pls/internal/data_structures/deque.h"
+#include "pls/internal/data_structures/work_stealing_deque.h"
 
 
 #include "abstract_task.h"
 #include "abstract_task.h"
 #include "thread_state.h"
 #include "thread_state.h"
@@ -15,7 +15,7 @@ namespace internal {
 namespace scheduling {
 namespace scheduling {
 
 
 class fork_join_task;
 class fork_join_task;
-class fork_join_sub_task : public data_structures::deque_item {
+class fork_join_sub_task {
   friend class fork_join_task;
   friend class fork_join_task;
 
 
   // Coordinate finishing of sub_tasks
   // Coordinate finishing of sub_tasks
@@ -25,8 +25,11 @@ class fork_join_sub_task : public data_structures::deque_item {
   // Access to TBB scheduling environment
   // Access to TBB scheduling environment
   fork_join_task *tbb_task_;
   fork_join_task *tbb_task_;
 
 
+  bool executed = false;
+  int executed_at = -1;
+
   // Stack Management (reset stack pointer after wait_for_all() calls)
   // Stack Management (reset stack pointer after wait_for_all() calls)
-  data_structures::aligned_stack::state stack_state_;
+  data_structures::work_stealing_deque<fork_join_sub_task>::state deque_state_;
  protected:
  protected:
   explicit fork_join_sub_task();
   explicit fork_join_sub_task();
   fork_join_sub_task(const fork_join_sub_task &other);
   fork_join_sub_task(const fork_join_sub_task &other);
@@ -37,11 +40,10 @@ class fork_join_sub_task : public data_structures::deque_item {
  public:
  public:
   // Only use them when actually executing this sub_task (only public for simpler API design)
   // Only use them when actually executing this sub_task (only public for simpler API design)
   template<typename T>
   template<typename T>
-  void spawn_child(const T &sub_task);
+  void spawn_child(T &sub_task);
   void wait_for_all();
   void wait_for_all();
 
 
  private:
  private:
-  void spawn_child_internal(fork_join_sub_task *sub_task);
   void execute();
   void execute();
 };
 };
 
 
@@ -50,7 +52,7 @@ class fork_join_lambda_by_reference : public fork_join_sub_task {
   const Function *function_;
   const Function *function_;
 
 
  public:
  public:
-  explicit fork_join_lambda_by_reference(const Function *function) : function_{function} {};
+  explicit fork_join_lambda_by_reference(const Function *function) : fork_join_sub_task{}, function_{function} {};
 
 
  protected:
  protected:
   void execute_internal() override {
   void execute_internal() override {
@@ -63,7 +65,7 @@ class fork_join_lambda_by_value : public fork_join_sub_task {
   const Function function_;
   const Function function_;
 
 
  public:
  public:
-  explicit fork_join_lambda_by_value(const Function &function) : function_{function} {};
+  explicit fork_join_lambda_by_value(const Function &function) : fork_join_sub_task{}, function_{function} {};
 
 
  protected:
  protected:
   void execute_internal() override {
   void execute_internal() override {
@@ -76,10 +78,9 @@ class fork_join_task : public abstract_task {
 
 
   fork_join_sub_task *root_task_;
   fork_join_sub_task *root_task_;
   fork_join_sub_task *currently_executing_;
   fork_join_sub_task *currently_executing_;
-  data_structures::aligned_stack *my_stack_;
 
 
   // Double-Ended Queue management
   // Double-Ended Queue management
-  data_structures::deque<fork_join_sub_task> deque_;
+  data_structures::work_stealing_deque<fork_join_sub_task> deque_;
 
 
   // Steal Management
   // Steal Management
   fork_join_sub_task *last_stolen_;
   fork_join_sub_task *last_stolen_;
@@ -97,12 +98,21 @@ class fork_join_task : public abstract_task {
 };
 };
 
 
 template<typename T>
 template<typename T>
-void fork_join_sub_task::spawn_child(const T &task) {
+void fork_join_sub_task::spawn_child(T &task) {
   PROFILE_FORK_JOIN_STEALING("spawn_child")
   PROFILE_FORK_JOIN_STEALING("spawn_child")
   static_assert(std::is_base_of<fork_join_sub_task, T>::value, "Only pass fork_join_sub_task subclasses!");
   static_assert(std::is_base_of<fork_join_sub_task, T>::value, "Only pass fork_join_sub_task subclasses!");
 
 
-  T *new_task = tbb_task_->my_stack_->push(task);
-  spawn_child_internal(new_task);
+  // Keep our refcount up to date
+  ref_count_++;
+
+  // Assign forced values
+  task.parent_ = this;
+  task.tbb_task_ = tbb_task_;
+  task.deque_state_ = tbb_task_->deque_.save_state();
+
+  // Push on our deque
+  const T const_task = task;
+  tbb_task_->deque_.push_tail(const_task);
 }
 }
 
 
 }
 }

lib/pls/src/internal/base/alignment.cpp

@@ -10,8 +10,8 @@ void *allocate_aligned(size_t size) {
   return aligned_alloc(system_details::CACHE_LINE_SIZE, size);
   return aligned_alloc(system_details::CACHE_LINE_SIZE, size);
 }
 }
 
 
-std::uintptr_t next_alignment(std::uintptr_t size) {
-  std::uintptr_t miss_alignment = size % base::system_details::CACHE_LINE_SIZE;
+system_details::pointer_t next_alignment(system_details::pointer_t size) {
+  system_details::pointer_t miss_alignment = size % base::system_details::CACHE_LINE_SIZE;
   if (miss_alignment == 0) {
   if (miss_alignment == 0) {
     return size;
     return size;
   } else {
   } else {
@@ -19,8 +19,17 @@ std::uintptr_t next_alignment(std::uintptr_t size) {
   }
   }
 }
 }
 
 
+system_details::pointer_t previous_alignment(system_details::pointer_t size) {
+  system_details::pointer_t miss_alignment = size % base::system_details::CACHE_LINE_SIZE;
+  if (miss_alignment == 0) {
+    return size;
+  } else {
+    return size - miss_alignment;
+  }
+}
+
 char *next_alignment(char *pointer) {
 char *next_alignment(char *pointer) {
-  return reinterpret_cast<char *>(next_alignment(reinterpret_cast<std::uintptr_t >(pointer)));
+  return reinterpret_cast<char *>(next_alignment(reinterpret_cast<system_details::pointer_t >(pointer)));
 }
 }
 
 
 }
 }

lib/pls/src/internal/base/swmr_spin_lock.cpp

@@ -23,22 +23,22 @@ bool swmr_spin_lock::reader_try_lock() {
 
 
 void swmr_spin_lock::reader_unlock() {
 void swmr_spin_lock::reader_unlock() {
   PROFILE_LOCK("Release Read Lock")
   PROFILE_LOCK("Release Read Lock")
-  readers_.fetch_add(-1, std::memory_order_release);
+  readers_--;
 }
 }
 
 
 void swmr_spin_lock::writer_lock() {
 void swmr_spin_lock::writer_lock() {
   PROFILE_LOCK("Acquire Write Lock")
   PROFILE_LOCK("Acquire Write Lock")
   // Tell the readers that we would like to write
   // Tell the readers that we would like to write
-  write_request_.store(1, std::memory_order_acquire);
+  write_request_ = 1;
 
 
   // Wait for all of them to exit the critical section
   // Wait for all of them to exit the critical section
-  while (readers_.load(std::memory_order_acquire) > 0)
+  while (readers_ > 0)
     system_details::relax_cpu(); // Spin, not expensive as relaxed load
     system_details::relax_cpu(); // Spin, not expensive as relaxed load
 }
 }
 
 
 void swmr_spin_lock::writer_unlock() {
 void swmr_spin_lock::writer_unlock() {
   PROFILE_LOCK("Release Write Lock")
   PROFILE_LOCK("Release Write Lock")
-  write_request_.store(0, std::memory_order_release);
+  write_request_ = 0;
 }
 }
 
 
 }
 }

lib/pls/src/internal/base/ttas_spin_lock.cpp

@@ -9,7 +9,7 @@ namespace base {
 void ttas_spin_lock::lock() {
 void ttas_spin_lock::lock() {
   PROFILE_LOCK("Acquire Lock")
   PROFILE_LOCK("Acquire Lock")
   int expected = 0;
   int expected = 0;
-  backoff_.reset();
+  backoff backoff_;
 
 
   while (true) {
   while (true) {
     while (flag_.load(std::memory_order_relaxed) == 1)
     while (flag_.load(std::memory_order_relaxed) == 1)
@@ -26,7 +26,7 @@ void ttas_spin_lock::lock() {
 bool ttas_spin_lock::try_lock(unsigned int num_tries) {
 bool ttas_spin_lock::try_lock(unsigned int num_tries) {
   PROFILE_LOCK("Try Acquire Lock")
   PROFILE_LOCK("Try Acquire Lock")
   int expected = 0;
   int expected = 0;
-  backoff_.reset();
+  backoff backoff_;
 
 
   while (true) {
   while (true) {
     while (flag_.load() == 1) {
     while (flag_.load() == 1) {

lib/pls/src/internal/data_structures/aligned_stack.cpp

@@ -5,11 +5,16 @@ namespace pls {
 namespace internal {
 namespace internal {
 namespace data_structures {
 namespace data_structures {
 
 
-aligned_stack::aligned_stack(char *memory_region, const std::size_t size) :
+aligned_stack::aligned_stack(pointer_t memory_region, const std::size_t size) :
     memory_start_{memory_region},
     memory_start_{memory_region},
     memory_end_{memory_region + size},
     memory_end_{memory_region + size},
     head_{base::alignment::next_alignment(memory_start_)} {}
     head_{base::alignment::next_alignment(memory_start_)} {}
 
 
+aligned_stack::aligned_stack(char *memory_region, const std::size_t size) :
+    memory_start_{(pointer_t) memory_region},
+    memory_end_{(pointer_t) memory_region + size},
+    head_{base::alignment::next_alignment(memory_start_)} {}
+
 }
 }
 }
 }
 }
 }

lib/pls/src/internal/data_structures/deque.cpp

@@ -14,11 +14,11 @@ deque_item *deque_internal::pop_head_internal() {
   }
   }
 
 
   deque_item *result = head_;
   deque_item *result = head_;
-  head_ = head_->prev_;
+  head_ = head_->next_;
   if (head_ == nullptr) {
   if (head_ == nullptr) {
     tail_ = nullptr;
     tail_ = nullptr;
   } else {
   } else {
-    head_->next_ = nullptr;
+    head_->prev_ = nullptr;
   }
   }
 
 
   return result;
   return result;
@@ -32,11 +32,11 @@ deque_item *deque_internal::pop_tail_internal() {
   }
   }
 
 
   deque_item *result = tail_;
   deque_item *result = tail_;
-  tail_ = tail_->next_;
+  tail_ = tail_->prev_;
   if (tail_ == nullptr) {
   if (tail_ == nullptr) {
     head_ = nullptr;
     head_ = nullptr;
   } else {
   } else {
-    tail_->prev_ = nullptr;
+    tail_->next_ = nullptr;
   }
   }
 
 
   return result;
   return result;
@@ -46,12 +46,12 @@ void deque_internal::push_tail_internal(deque_item *new_item) {
   std::lock_guard<base::spin_lock> lock{lock_};
   std::lock_guard<base::spin_lock> lock{lock_};
 
 
   if (tail_ != nullptr) {
   if (tail_ != nullptr) {
-    tail_->prev_ = new_item;
+    tail_->next_ = new_item;
   } else {
   } else {
     head_ = new_item;
     head_ = new_item;
   }
   }
-  new_item->next_ = tail_;
-  new_item->prev_ = nullptr;
+  new_item->prev_ = tail_;
+  new_item->next_ = nullptr;
   tail_ = new_item;
   tail_ = new_item;
 }
 }
 
 

lib/pls/src/internal/scheduling/fork_join_task.cpp

@@ -8,22 +8,26 @@ namespace internal {
 namespace scheduling {
 namespace scheduling {
 
 
 fork_join_sub_task::fork_join_sub_task() :
 fork_join_sub_task::fork_join_sub_task() :
-    data_structures::deque_item{},
     ref_count_{0},
     ref_count_{0},
     parent_{nullptr},
     parent_{nullptr},
     tbb_task_{nullptr},
     tbb_task_{nullptr},
-    stack_state_{nullptr} {}
+    deque_state_{0} {}
 
 
 fork_join_sub_task::fork_join_sub_task(const fork_join_sub_task &other) :
 fork_join_sub_task::fork_join_sub_task(const fork_join_sub_task &other) :
-    data_structures::deque_item(other),
     ref_count_{0},
     ref_count_{0},
-    parent_{nullptr},
-    tbb_task_{nullptr},
-    stack_state_{nullptr} {}
+    parent_{other.parent_},
+    tbb_task_{other.tbb_task_},
+    deque_state_{other.deque_state_} {}
 
 
 void fork_join_sub_task::execute() {
 void fork_join_sub_task::execute() {
   PROFILE_WORK_BLOCK("execute sub_task")
   PROFILE_WORK_BLOCK("execute sub_task")
   tbb_task_->currently_executing_ = this;
   tbb_task_->currently_executing_ = this;
+  if (executed) {
+    int my_id = base::this_thread::state<thread_state>()->id_;
+    PLS_ERROR("Double Execution!")
+  }
+  executed = true;
+  executed_at = base::this_thread::state<thread_state>()->id_;
   execute_internal();
   execute_internal();
   tbb_task_->currently_executing_ = nullptr;
   tbb_task_->currently_executing_ = nullptr;
   PROFILE_END_BLOCK
   PROFILE_END_BLOCK
@@ -34,18 +38,6 @@ void fork_join_sub_task::execute() {
   }
   }
 }
 }
 
 
-void fork_join_sub_task::spawn_child_internal(fork_join_sub_task *sub_task) {
-  // Keep our refcount up to date
-  ref_count_++;
-
-  // Assign forced values
-  sub_task->parent_ = this;
-  sub_task->tbb_task_ = tbb_task_;
-  sub_task->stack_state_ = tbb_task_->my_stack_->save_state();
-
-  tbb_task_->deque_.push_tail(sub_task);
-}
-
 void fork_join_sub_task::wait_for_all() {
 void fork_join_sub_task::wait_for_all() {
   while (ref_count_ > 0) {
   while (ref_count_ > 0) {
     PROFILE_STEALING("get local sub task")
     PROFILE_STEALING("get local sub task")
@@ -54,7 +46,6 @@ void fork_join_sub_task::wait_for_all() {
     if (local_task != nullptr) {
     if (local_task != nullptr) {
       local_task->execute();
       local_task->execute();
     } else {
     } else {
-      while (ref_count_ > 0) {
       // Try to steal work.
       // Try to steal work.
       // External steal will be executed implicitly if success
       // External steal will be executed implicitly if success
       PROFILE_STEALING("steal work")
       PROFILE_STEALING("steal work")
@@ -65,8 +56,7 @@ void fork_join_sub_task::wait_for_all() {
       }
       }
     }
     }
   }
   }
-  }
-  tbb_task_->my_stack_->reset_state(stack_state_);
+  tbb_task_->deque_.release_memory_until(deque_state_);
 }
 }
 
 
 fork_join_sub_task *fork_join_task::get_local_sub_task() {
 fork_join_sub_task *fork_join_task::get_local_sub_task() {
@@ -74,7 +64,9 @@ fork_join_sub_task *fork_join_task::get_local_sub_task() {
 }
 }
 
 
 fork_join_sub_task *fork_join_task::get_stolen_sub_task() {
 fork_join_sub_task *fork_join_task::get_stolen_sub_task() {
-  return deque_.pop_head();
+  auto tmp = deque_.save_state();
+  auto result = deque_.pop_head();
+  return result;
 }
 }
 
 
 bool fork_join_task::internal_stealing(abstract_task *other_task) {
 bool fork_join_task::internal_stealing(abstract_task *other_task) {
@@ -87,7 +79,7 @@ bool fork_join_task::internal_stealing(abstract_task *other_task) {
   } else {
   } else {
     // Make sub-task belong to our fork_join_task instance
     // Make sub-task belong to our fork_join_task instance
     stolen_sub_task->tbb_task_ = this;
     stolen_sub_task->tbb_task_ = this;
-    stolen_sub_task->stack_state_ = my_stack_->save_state();
+    stolen_sub_task->deque_state_ = deque_.save_state();
     // We will execute this next without explicitly moving it onto our stack storage
     // We will execute this next without explicitly moving it onto our stack storage
     last_stolen_ = stolen_sub_task;
     last_stolen_ = stolen_sub_task;
 
 
@@ -114,9 +106,12 @@ void fork_join_task::execute() {
   PROFILE_WORK_BLOCK("execute fork_join_task");
   PROFILE_WORK_BLOCK("execute fork_join_task");
 
 
   // Bind this instance to our OS thread
   // Bind this instance to our OS thread
-  my_stack_ = base::this_thread::state<thread_state>()->task_stack_;
+  // TODO: See if we did this right
+  // my_stack_ = base::this_thread::state<thread_state>()->task_stack_;
+  deque_.reset_base_pointer();
+
   root_task_->tbb_task_ = this;
   root_task_->tbb_task_ = this;
-  root_task_->stack_state_ = my_stack_->save_state();
+  root_task_->deque_state_ = deque_.save_state();
 
 
   // Execute it on our OS thread until its finished
   // Execute it on our OS thread until its finished
   root_task_->execute();
   root_task_->execute();
@@ -124,12 +119,12 @@ void fork_join_task::execute() {
 
 
 fork_join_sub_task *fork_join_task::currently_executing() const { return currently_executing_; }
 fork_join_sub_task *fork_join_task::currently_executing() const { return currently_executing_; }
 
 
-fork_join_task::fork_join_task(fork_join_sub_task *root_task, const abstract_task::id &id) :
+fork_join_task::fork_join_task(fork_join_sub_task *root_task,
+                               const abstract_task::id &id) :
     abstract_task{0, id},
     abstract_task{0, id},
     root_task_{root_task},
     root_task_{root_task},
     currently_executing_{nullptr},
     currently_executing_{nullptr},
-    my_stack_{nullptr},
-    deque_{},
+    deque_{base::this_thread::state<thread_state>()->task_stack_},
     last_stolen_{nullptr} {}
     last_stolen_{nullptr} {}
 
 
 }
 }

test/CMakeLists.txt

 add_executable(tests
 add_executable(tests
         main.cpp
         main.cpp
-        base_tests.cpp scheduling_tests.cpp data_structures_test.cpp)
+        data_structures_test.cpp)
 target_link_libraries(tests catch2 pls)
 target_link_libraries(tests catch2 pls)

test/data_structures_test.cpp

@@ -4,6 +4,7 @@
 
 
 #include <pls/internal/data_structures/aligned_stack.h>
 #include <pls/internal/data_structures/aligned_stack.h>
 #include <pls/internal/data_structures/deque.h>
 #include <pls/internal/data_structures/deque.h>
+#include <pls/internal/data_structures/work_stealing_deque.h>
 
 
 #include <vector>
 #include <vector>
 #include <mutex>
 #include <mutex>
@@ -130,3 +131,90 @@ TEST_CASE("deque stores objects correctly", "[internal/data_structures/deque.h]"
     REQUIRE(deque.pop_tail() == &three);
     REQUIRE(deque.pop_tail() == &three);
   }
   }
 }
 }
+
+TEST_CASE("work stealing deque stores objects correctly", "[internal/data_structures/aligned_stack.h]") {
+  constexpr long data_size = 2 << 14;
+  char data[data_size];
+  aligned_stack stack{data, data_size};
+  work_stealing_deque<int> deque{&stack};
+
+  int one = 1, two = 2, three = 3, four = 4;
+
+  SECTION("add and remove items form the tail") {
+    deque.push_tail(one);
+    deque.push_tail(two);
+    deque.push_tail(three);
+
+    REQUIRE(*deque.pop_tail() == three);
+    REQUIRE(*deque.pop_tail() == two);
+    REQUIRE(*deque.pop_tail() == one);
+  }
+
+  SECTION("handles getting empty by popping the tail correctly") {
+    deque.push_tail(one);
+    REQUIRE(*deque.pop_tail() == one);
+
+    deque.push_tail(two);
+    REQUIRE(*deque.pop_tail() == two);
+  }
+
+  SECTION("remove items form the head") {
+    deque.push_tail(one);
+    deque.push_tail(two);
+    deque.push_tail(three);
+
+    REQUIRE(*deque.pop_head() == one);
+    REQUIRE(*deque.pop_head() == two);
+    REQUIRE(*deque.pop_head() == three);
+  }
+
+  SECTION("handles getting empty by popping the head correctly") {
+    deque.push_tail(one);
+    REQUIRE(*deque.pop_head() == one);
+
+    deque.push_tail(two);
+    REQUIRE(*deque.pop_head() == two);
+  }
+
+  SECTION("handles getting empty by popping the head and tail correctly") {
+    deque.push_tail(one);
+    REQUIRE(*deque.pop_tail() == one);
+
+    deque.push_tail(two);
+    REQUIRE(*deque.pop_head() == two);
+
+    deque.push_tail(three);
+    REQUIRE(*deque.pop_tail() == three);
+  }
+
+  SECTION("handles jumps bigger 1 correctly") {
+    deque.push_tail(one);
+    deque.push_tail(two);
+    REQUIRE(*deque.pop_tail() == two);
+
+    deque.push_tail(three);
+    deque.push_tail(four);
+    REQUIRE(*deque.pop_head() == one);
+    REQUIRE(*deque.pop_head() == three);
+    REQUIRE(*deque.pop_head() == four);
+  }
+
+  SECTION("handles stack reset 1 correctly when emptied by tail") {
+    deque.push_tail(one);
+    deque.push_tail(two);
+    auto tmp_result = deque.pop_tail();
+    REQUIRE(*tmp_result == two);
+
+    deque.release_memory_until(tmp_result);
+    REQUIRE(*deque.pop_tail() == one);
+
+    deque.push_tail(three);
+    deque.push_tail(four);
+    REQUIRE(*deque.pop_head() == three);
+    REQUIRE(*deque.pop_tail() == four);
+  }
+
+  SECTION("synces correctly") {
+
+  }
+}