wsq.hpp

#pragma once
 
#include <bit>
 
#include "../utility/macros.hpp"
#include "../utility/traits.hpp"

 
#ifndef TF_DEFAULT_BOUNDED_TASK_QUEUE_LOG_SIZE 
  #define TF_DEFAULT_BOUNDED_TASK_QUEUE_LOG_SIZE 8
#endif
 
#ifndef TF_DEFAULT_UNBOUNDED_TASK_QUEUE_LOG_SIZE 
  #define TF_DEFAULT_UNBOUNDED_TASK_QUEUE_LOG_SIZE 10
#endif
 
namespace tf {
 
// ----------------------------------------------------------------------------
// Work-stealing queue steal protocol sentinels
//
// These free functions define the two sentinel values used by steal operations
// across all work-stealing queue types (BoundedWSQ, UnboundedWSQ). They encode
// the result of a steal attempt into the return value itself, avoiding any
// out-parameter or separate status type.
//
// For pointer types T:
//   wsq_empty_value<T>()     = nullptr  — queue was genuinely empty
//   wsq_contended_value<T>() = 0x1      — queue had work but CAS was lost to
//                                         another thief; caller should retry
//
// The sentinel 0x1 is safe because any real object pointer is aligned to at
// least alignof(T) >= 1, and the OS never maps address 0x1. For void* there
// is no pointee alignment to check, but the same reasoning applies — no
// allocator ever returns address 0x1.
//
// For non-pointer types T, both return std::nullopt since sentinel encoding
// is not possible without a dedicated out-of-band value.
//
// Both queue classes expose these as static member functions (empty_value,
// contended_value) that delegate here, so callers can use either form.
// ----------------------------------------------------------------------------

template <typename T>
constexpr auto wsq_empty_value() {
  if constexpr (std::is_pointer_v<T>) {
    return T{nullptr};
  } else {
    return std::optional<T>{std::nullopt};
  }
}

template <typename T>
auto wsq_contended_value() {
  if constexpr (std::is_pointer_v<T>) {
    return reinterpret_cast<T>(uintptr_t{1});
  } else {
    return std::optional<T>{std::nullopt};
  }
}
 
// ----------------------------------------------------------------------------
// Unbounded Work-stealing Queue (WSQ)
// ----------------------------------------------------------------------------
 

template <typename T>
class UnboundedWSQ {
 
  struct Array {
 
    size_t C;
    size_t M;
    std::atomic<T>* S;
 
    explicit Array(size_t c) :
      C {c},
      M {c-1},
      S {new std::atomic<T>[C]} {
    }
 
    ~Array() {
      delete [] S;
    }
 
    size_t capacity() const noexcept {
      return C;
    }
 
    void push(int64_t i, T o) noexcept {
      S[i & M].store(o, std::memory_order_relaxed);
    }
 
    T pop(int64_t i) noexcept {
      return S[i & M].load(std::memory_order_relaxed);
    }
 
    Array* resize(int64_t b, int64_t t) {
      Array* ptr = new Array {2*C};
      for(int64_t i=t; i!=b; ++i) {
        ptr->push(i, pop(i));
      }
      return ptr;
    }
 
    Array* resize(int64_t b, int64_t t, size_t N) {
      // assert(N>0);
      Array* ptr = new Array {std::bit_ceil(C + N)};
      for(int64_t i=t; i!=b; ++i) {
        ptr->push(i, pop(i));
      }
      return ptr;
    }
 
  };
 
  alignas(TF_CACHELINE_SIZE) std::atomic<int64_t> _top;
  alignas(TF_CACHELINE_SIZE) std::atomic<int64_t> _bottom;
  std::atomic<Array*> _array;
  std::vector<Array*> _garbage;
 
  public:
  
  using value_type = std::conditional_t<std::is_pointer_v<T>, T, std::optional<T>>;

  explicit UnboundedWSQ(int64_t LogSize = TF_DEFAULT_UNBOUNDED_TASK_QUEUE_LOG_SIZE);

  ~UnboundedWSQ();

  bool empty() const noexcept;

  size_t size() const noexcept;

  size_t capacity() const noexcept;
  
  void push(T item);
  
  template <typename I>
  void bulk_push(I& first, size_t N);

  value_type pop();

  value_type steal();

  value_type steal_with_feedback();
  
  static constexpr auto empty_value()     { return wsq_empty_value<T>();     }

  static auto contended_value()           { return wsq_contended_value<T>(); }
 
  private:
 
  Array* _resize_array(Array* a, int64_t b, int64_t t);
  Array* _resize_array(Array* a, int64_t b, int64_t t, size_t N);
};
 
// Constructor
template <typename T>
UnboundedWSQ<T>::UnboundedWSQ(int64_t LogSize) {
  _top.store(0, std::memory_order_relaxed);
  _bottom.store(0, std::memory_order_relaxed);
  _array.store(new Array{(size_t{1} << LogSize)}, std::memory_order_relaxed);
  _garbage.reserve(32);
}
 
// Destructor
template <typename T>
UnboundedWSQ<T>::~UnboundedWSQ() {
  for(auto a : _garbage) {
    delete a;
  }
  delete _array.load();
}
 
// Function: empty
template <typename T>
bool UnboundedWSQ<T>::empty() const noexcept {
  int64_t t = _top.load(std::memory_order_relaxed);
  int64_t b = _bottom.load(std::memory_order_relaxed);
  return (b <= t);
}
 
// Function: size
template <typename T>
size_t UnboundedWSQ<T>::size() const noexcept {
  int64_t t = _top.load(std::memory_order_relaxed);
  int64_t b = _bottom.load(std::memory_order_relaxed);
  return static_cast<size_t>(b >= t ? b - t : 0);
}
 
// Function: push
template <typename T>
void UnboundedWSQ<T>::push(T o) {
 
  int64_t b = _bottom.load(std::memory_order_relaxed);
  int64_t t = _top.load(std::memory_order_acquire);
  Array* a = _array.load(std::memory_order_relaxed);
 
  // queue is full with one additional item (b-t+1)
  if (a->capacity() < static_cast<size_t>(b - t + 1)) [[unlikely]] {
    a = _resize_array(a, b, t);
  }
 
  a->push(b, o);
  std::atomic_thread_fence(std::memory_order_release);
 
  // original paper uses relaxed here but tsa complains
  _bottom.store(b + 1, std::memory_order_release);
}
 
// Function: bulk_push
template <typename T>
template <typename I>
void UnboundedWSQ<T>::bulk_push(I& first, size_t N) {
 
  if(N == 0) return;
 
  int64_t b = _bottom.load(std::memory_order_relaxed);
  int64_t t = _top.load(std::memory_order_acquire);
  Array* a = _array.load(std::memory_order_relaxed);
 
  // queue is full with N additional items
  if ( (b - t + N) > a->capacity() ) [[unlikely]] {
    a = _resize_array(a, b, t, N);
  }
 
  for(size_t i=0; i<N; ++i) {
    a->push(b++, *first++);
  }
  std::atomic_thread_fence(std::memory_order_release);
 
  // original paper uses relaxed here but tsa complains
  _bottom.store(b, std::memory_order_release);
}
 
// Function: pop
template <typename T>
typename UnboundedWSQ<T>::value_type 
UnboundedWSQ<T>::pop() {
 
  int64_t b = _bottom.load(std::memory_order_relaxed) - 1;
  Array* a = _array.load(std::memory_order_relaxed);
  _bottom.store(b, std::memory_order_relaxed);
  std::atomic_thread_fence(std::memory_order_seq_cst);
  int64_t t = _top.load(std::memory_order_relaxed);
 
  //T item {nullptr};
  auto item = empty_value();
 
  if(t <= b) {
    item = a->pop(b);
    if(t == b) {
      // the last item just got stolen
      if(!_top.compare_exchange_strong(t, t+1, std::memory_order_seq_cst,
                                               std::memory_order_relaxed)) {
        //item = nullptr;
        item = empty_value();
      }
      _bottom.store(b + 1, std::memory_order_relaxed);
    }
  }
  else {
    _bottom.store(b + 1, std::memory_order_relaxed);
  }
 
  return item;
}
 
// Function: steal
template <typename T>
typename UnboundedWSQ<T>::value_type 
UnboundedWSQ<T>::steal() {
  
  int64_t t = _top.load(std::memory_order_acquire);
  std::atomic_thread_fence(std::memory_order_seq_cst);
  int64_t b = _bottom.load(std::memory_order_acquire);
 
  //T item {nullptr};
  auto item = empty_value();
 
  if(t < b) {
    Array* a = _array.load(std::memory_order_consume);
    item = a->pop(t);
    if(!_top.compare_exchange_strong(t, t+1,
                                     std::memory_order_seq_cst,
                                     std::memory_order_relaxed)) {
      //return nullptr;
      return empty_value();
    }
  }
 
  return item;
}
 
// Function: steal_with_feedback
// Returns a stolen item, contended_value(), or empty_value() — see declaration.
template <typename T>
typename UnboundedWSQ<T>::value_type
UnboundedWSQ<T>::steal_with_feedback() {
 
  int64_t t = _top.load(std::memory_order_acquire);
  std::atomic_thread_fence(std::memory_order_seq_cst);
  int64_t b = _bottom.load(std::memory_order_acquire);
 
  if(t < b) {
    // queue is non-empty: load the candidate item and attempt the CAS
    Array* a = _array.load(std::memory_order_consume);
    auto item = a->pop(t);
    if(!_top.compare_exchange_strong(t, t+1,
                                     std::memory_order_seq_cst,
                                     std::memory_order_relaxed)) {
      // CAS lost to another thief — queue had work but we didn't get it.
      // Return contended_value() so the caller knows to retry this victim.
      return contended_value();
    }
    return item;
  }
 
  // bottom <= top: queue is genuinely empty
  return empty_value();
}
 
// Function: capacity
template <typename T>
size_t UnboundedWSQ<T>::capacity() const noexcept {
  return _array.load(std::memory_order_relaxed)->capacity();
}
 
template <typename T>
typename UnboundedWSQ<T>::Array*
UnboundedWSQ<T>::_resize_array(Array* a, int64_t b, int64_t t) {
  Array* tmp = a->resize(b, t);
  _garbage.push_back(a);
  // Note: the original paper using relaxed causes t-san to complain
  _array.store(tmp, std::memory_order_release);
  return tmp;
}
 
template <typename T>
typename UnboundedWSQ<T>::Array*
UnboundedWSQ<T>::_resize_array(Array* a, int64_t b, int64_t t, size_t N) {
  Array* tmp = a->resize(b, t, N);
  _garbage.push_back(a);
  // Note: the original paper using relaxed causes t-san to complain
  _array.store(tmp, std::memory_order_release);
  return tmp;
}
 
// ----------------------------------------------------------------------------
// Bounded Work-stealing Queue (WSQ)
// ----------------------------------------------------------------------------

template <typename T, size_t LogSize = TF_DEFAULT_BOUNDED_TASK_QUEUE_LOG_SIZE>
class BoundedWSQ {
  
  constexpr static size_t BufferSize = size_t{1} << LogSize;
  constexpr static size_t BufferMask = (BufferSize - 1);
 
  static_assert((BufferSize >= 2) && ((BufferSize & (BufferSize - 1)) == 0));
 
  alignas(TF_CACHELINE_SIZE) std::atomic<int64_t> _top {0};
  alignas(TF_CACHELINE_SIZE) std::atomic<int64_t> _bottom {0};
  alignas(TF_CACHELINE_SIZE) std::atomic<T> _buffer[BufferSize];
 
  public:
  
  using value_type = std::conditional_t<std::is_pointer_v<T>, T, std::optional<T>>;
    
  BoundedWSQ() = default;

  ~BoundedWSQ() = default;
  
  bool empty() const noexcept;
  
  size_t size() const noexcept;

  constexpr size_t capacity() const;
  
  template <typename O>
  bool try_push(O&& item);
  
  template <typename I>
  size_t try_bulk_push(I& first, size_t N);
  
  value_type pop();
  
  value_type steal();

  value_type steal_with_feedback();

  static constexpr auto empty_value()     { return wsq_empty_value<T>();     }

  static auto contended_value()           { return wsq_contended_value<T>(); }
};
 
// Function: empty
template <typename T, size_t LogSize>
bool BoundedWSQ<T, LogSize>::empty() const noexcept {
  int64_t t = _top.load(std::memory_order_relaxed);
  int64_t b = _bottom.load(std::memory_order_relaxed);
  return b <= t;
}
 
// Function: size
template <typename T, size_t LogSize>
size_t BoundedWSQ<T, LogSize>::size() const noexcept {
  int64_t t = _top.load(std::memory_order_relaxed);
  int64_t b = _bottom.load(std::memory_order_relaxed);
  return static_cast<size_t>(b >= t ? b - t : 0);
}
 
// Function: try_push
template <typename T, size_t LogSize>
template <typename O>
bool BoundedWSQ<T, LogSize>::try_push(O&& o) {
 
  int64_t b = _bottom.load(std::memory_order_relaxed);
  int64_t t = _top.load(std::memory_order_acquire);
 
  // queue is full with one additional item (b-t+1)
  if(static_cast<size_t>(b - t + 1) > BufferSize) [[unlikely]] {
    return false;
  }
  
  _buffer[b & BufferMask].store(std::forward<O>(o), std::memory_order_relaxed);
 
  std::atomic_thread_fence(std::memory_order_release);
  
  // original paper uses relaxed here but tsa complains
  _bottom.store(b + 1, std::memory_order_release);
 
  return true;
}
 
// Function: try_bulk_push
template <typename T, size_t LogSize>
template <typename I>
size_t BoundedWSQ<T, LogSize>::try_bulk_push(I& first, size_t N) {
 
  if(N == 0) return 0;
 
  int64_t b = _bottom.load(std::memory_order_relaxed);
  int64_t t = _top.load(std::memory_order_acquire);
 
  size_t r = BufferSize - (b - t);  // remaining capacity
  size_t n = std::min(N, r);        // number of pushable elements
 
  if(n > 0) {
    // push n elements into the queue
    for(size_t i=0; i<n; ++i) {
      _buffer[b++ & BufferMask].store(*first++, std::memory_order_relaxed);
    }
    std::atomic_thread_fence(std::memory_order_release);
    // original paper uses relaxed here but tsa complains
    _bottom.store(b, std::memory_order_release);
  }
  
  return n;
}
 
// Function: pop
template <typename T, size_t LogSize>
typename BoundedWSQ<T, LogSize>::value_type 
BoundedWSQ<T, LogSize>::pop() {
 
  int64_t b = _bottom.load(std::memory_order_relaxed) - 1;
  _bottom.store(b, std::memory_order_relaxed);
  std::atomic_thread_fence(std::memory_order_seq_cst);
  int64_t t = _top.load(std::memory_order_relaxed);
 
  //T item {nullptr};
  auto item = empty_value();
 
  if(t <= b) {
    item = _buffer[b & BufferMask].load(std::memory_order_relaxed);
    if(t == b) {
      // the last item just got stolen
      if(!_top.compare_exchange_strong(t, t+1, 
                                       std::memory_order_seq_cst, 
                                       std::memory_order_relaxed)) {
        //item = nullptr;
        item = empty_value();
      }
      _bottom.store(b + 1, std::memory_order_relaxed);
    }
  }
  else {
    _bottom.store(b + 1, std::memory_order_relaxed);
  }
 
  return item;
}
 
// Function: steal
template <typename T, size_t LogSize>
typename BoundedWSQ<T, LogSize>::value_type 
BoundedWSQ<T, LogSize>::steal() {
  int64_t t = _top.load(std::memory_order_acquire);
  std::atomic_thread_fence(std::memory_order_seq_cst);
  int64_t b = _bottom.load(std::memory_order_acquire);
  
  //T item{nullptr};
  auto item = empty_value();
 
  if(t < b) {
    item = _buffer[t & BufferMask].load(std::memory_order_relaxed);
    if(!_top.compare_exchange_strong(t, t+1,
                                     std::memory_order_seq_cst,
                                     std::memory_order_relaxed)) {
      //return nullptr;
      return empty_value();
    }
  }
 
  return item;
}
 
// Function: steal_with_feedback
// Returns a stolen item, contended_value(), or empty_value() — see declaration.
template <typename T, size_t LogSize>
typename BoundedWSQ<T, LogSize>::value_type
BoundedWSQ<T, LogSize>::steal_with_feedback() {
 
  int64_t t = _top.load(std::memory_order_acquire);
  std::atomic_thread_fence(std::memory_order_seq_cst);
  int64_t b = _bottom.load(std::memory_order_acquire);
 
  if(t < b) {
    // queue is non-empty: load the candidate item and attempt the CAS
    auto item = _buffer[t & BufferMask].load(std::memory_order_relaxed);
    if(!_top.compare_exchange_strong(t, t+1,
                                     std::memory_order_seq_cst,
                                     std::memory_order_relaxed)) {
      // CAS lost to another thief — queue had work but we didn't get it.
      // Return contended_value() so the caller knows to retry this victim.
      return contended_value();
    }
    return item;
  }
 
  // bottom <= top: queue is genuinely empty
  return empty_value();
}
 
// Function: capacity
template <typename T, size_t LogSize>
constexpr size_t BoundedWSQ<T, LogSize>::capacity() const {
  return BufferSize;
}
 
 
}  // end of namespace tf -----------------------------------------------------