#pragma once

#include <atomic>
#include <cassert>
#include <cstddef>
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <iterator>
#include <utility>
#include <vector>

#if defined(__x86_64__) || defined(_M_X64)
#include <immintrin.h>
#endif

// Wait strategies for controlling thread blocking behavior when no work is
// available
enum class WaitStrategy {
  // Never block - threads busy-wait (spin) when no work available.
  // Stage threads will always use 100% CPU even when idle.
  // Requires dedicated CPU cores to avoid scheduler thrashing.
  // Use when: latency is critical and you have spare cores.
  Never,

  // Block only when all upstream stages are idle (no new work entering
  // pipeline).
  // Downstream threads busy-wait if upstream has work but not for their stage.
  // Eliminates futex notifications between stages, reduces to 0% CPU when idle.
  // Requires dedicated cores to avoid priority inversion when pipeline has
  // work.
  // Use when: high throughput with spare cores and sustained workloads.
  WaitIfUpstreamIdle,

  // Block when individual stages are empty (original behavior).
  // Each stage waits independently on its input sources.
  // Safe for shared CPU environments, works well with variable workloads.
  // Use when: general purpose, shared cores, or unpredictable workloads.
  WaitIfStageEmpty,
};

// Core thread state
struct ThreadState {
  alignas(128) std::atomic<uint32_t> pops{0};
  uint32_t local_pops{0};
  std::vector<uint32_t> last_push_read;
  bool last_stage;
};

// Runtime topology configuration for dynamic pipelines
//
// This class defines a pipeline topology at runtime:
// - Stage and thread calculations done at runtime
// - Flexible configuration: topology can be set via constructor
// - Dynamic arrays with runtime bounds checking
// - Single implementation works for any topology
//
// Example: PipelineTopology({1, 4, 2}) creates:
//   - Stage 0: 1 thread  (index 0)
//   - Stage 1: 4 threads (indices 1-4)
//   - Stage 2: 2 threads (indices 5-6)
//   - Total: 7 threads across 3 stages
struct PipelineTopology {
  const std::vector<int> threads_per_stage;
  const int num_stages;
  const std::vector<int> stage_offsets;
  const int total_threads;

  explicit PipelineTopology(std::vector<int> threads_per_stage_)
      : threads_per_stage(validate_and_move(std::move(threads_per_stage_))),
        num_stages(static_cast<int>(threads_per_stage.size())),
        stage_offsets(build_stage_offsets(threads_per_stage)),
        total_threads(build_total_threads(threads_per_stage)) {}

  // Runtime stage offset calculation
  int stage_offset(int stage) const {
    if (stage < 0 || stage >= num_stages) {
      std::abort(); // Stage index out of bounds
    }
    return stage_offsets[stage];
  }

  // Runtime thread index calculation
  int thread_index(int stage, int thread) const {
    if (stage < 0 || stage >= num_stages) {
      std::abort(); // Stage index out of bounds
    }
    if (thread < 0 || thread >= threads_per_stage[stage]) {
      std::abort(); // Thread index out of bounds
    }
    return stage_offsets[stage] + thread;
  }

  // Runtime previous stage thread count
  int prev_stage_thread_count(int stage) const {
    if (stage < 0 || stage >= num_stages) {
      std::abort(); // Stage index out of bounds
    }
    if (stage == 0) {
      return 1;
    } else {
      return threads_per_stage[stage - 1];
    }
  }

private:
  static std::vector<int> validate_and_move(std::vector<int> threads) {
    if (threads.empty()) {
      std::abort(); // Must specify at least one stage
    }
    for (int count : threads) {
      if (count <= 0) {
        std::abort(); // All stages must have at least one thread
      }
    }
    return threads;
  }

  static std::vector<int>
  build_stage_offsets(const std::vector<int> &threads_per_stage) {
    std::vector<int> offsets(threads_per_stage.size());
    int offset = 0;
    for (size_t i = 0; i < threads_per_stage.size(); ++i) {
      offsets[i] = offset;
      offset += threads_per_stage[i];
    }
    return offsets;
  }

  static int build_total_threads(const std::vector<int> &threads_per_stage) {
    int total = 0;
    for (int count : threads_per_stage) {
      total += count;
    }
    return total;
  }
};

// Pipeline algorithms - runtime configurable versions
namespace PipelineAlgorithms {

inline uint32_t calculate_safe_len(WaitStrategy wait_strategy,
                                   const PipelineTopology &topology, int stage,
                                   int thread_in_stage,
                                   std::vector<ThreadState> &all_threads,
                                   std::atomic<uint32_t> &pushes,
                                   bool may_block) {
  int thread_idx = topology.thread_index(stage, thread_in_stage);
  auto &thread = all_threads[thread_idx];
  uint32_t safe_len = UINT32_MAX;

  int prev_stage_threads = topology.prev_stage_thread_count(stage);

  // Runtime loop over previous stage threads
  for (int i = 0; i < prev_stage_threads; ++i) {
    std::atomic<uint32_t> &last_push = [&]() -> std::atomic<uint32_t> & {
      if (stage == 0) {
        return pushes;
      } else {
        int prev_thread_idx = topology.thread_index(stage - 1, i);
        return all_threads[prev_thread_idx].pops;
      }
    }();

    if (thread.last_push_read[i] == thread.local_pops) {
      thread.last_push_read[i] = last_push.load(std::memory_order_acquire);
      if (thread.last_push_read[i] == thread.local_pops) {
        if (!may_block) {
          safe_len = 0;
          return safe_len;
        }

        if (wait_strategy == WaitStrategy::Never) {
          // Empty - busy wait
        } else if (wait_strategy == WaitStrategy::WaitIfUpstreamIdle) {
          // We're allowed to spin as long as we eventually go to 0% cpu
          // usage on idle
          uint32_t push;
          bool should_wait = true;
          for (int j = 0; j < 100000; ++j) {
            push = pushes.load(std::memory_order_relaxed);
            if (push != thread.local_pops) {
              should_wait = false;
              break;
            }
#if defined(__x86_64__) || defined(_M_X64)
            _mm_pause();
#endif
          }
          if (should_wait) {
            pushes.wait(push, std::memory_order_relaxed);
          }
        } else { // WaitStrategy::WaitIfStageEmpty
          last_push.wait(thread.last_push_read[i], std::memory_order_relaxed);
        }

        thread.last_push_read[i] = last_push.load(std::memory_order_acquire);
      }
    }
    safe_len = std::min(safe_len, thread.last_push_read[i] - thread.local_pops);
  }

  return safe_len;
}

inline void update_thread_pops(WaitStrategy wait_strategy,
                               const PipelineTopology &topology, int stage,
                               int thread_in_stage,
                               std::vector<ThreadState> &all_threads,
                               uint32_t local_pops) {
  int thread_idx = topology.thread_index(stage, thread_in_stage);
  auto &thread_state = all_threads[thread_idx];

  if (wait_strategy == WaitStrategy::WaitIfStageEmpty) {
    thread_state.pops.store(local_pops, std::memory_order_seq_cst);
    thread_state.pops.notify_all();
  } else if (stage == topology.num_stages - 1) { // last stage
    thread_state.pops.store(local_pops, std::memory_order_seq_cst);
    thread_state.pops.notify_all();
  } else {
    thread_state.pops.store(local_pops, std::memory_order_release);
  }
}

inline int check_producer_capacity(const PipelineTopology &topology,
                                   std::vector<ThreadState> &all_threads,
                                   uint32_t slot, uint32_t size,
                                   uint32_t slot_count, bool block) {
  int last_stage = topology.num_stages - 1;
  int last_stage_offset = topology.stage_offset(last_stage);
  int last_stage_thread_count = topology.threads_per_stage[last_stage];

  for (int i = 0; i < last_stage_thread_count; ++i) {
    auto &thread = all_threads[last_stage_offset + i];
    uint32_t pops = thread.pops.load(std::memory_order_acquire);
    if (slot + size - pops > slot_count) {
      if (!block) {
        return 2; // Cannot proceed
      }
      thread.pops.wait(pops, std::memory_order_relaxed);
      return 1; // Should retry
    }
  }
  return 0; // Can proceed
}
} // namespace PipelineAlgorithms

// Multi-stage lock-free pipeline for inter-thread communication
// with runtime-configurable topology and wait strategy.
//
// Overview:
// - Items flow from producers through multiple processing stages (stage 0 ->
// stage 1 -> ... -> final stage)
// - Each stage can have multiple worker threads processing items in parallel
// - Uses a shared ring buffer with atomic counters for lock-free coordination
// - Supports batch processing for efficiency
// - Runtime-configurable topology and wait strategy via constructor parameters
//
// Architecture:
// - Producers: External threads that add items to the pipeline via push()
// - Stages: Processing stages numbered 0, 1, 2, ... that consume items via
// acquire(stage, thread)
// - Items flow: Producers -> Stage 0 -> Stage 1 -> ... -> Final Stage
//
// Usage Pattern:
//   ThreadPipeline<Item> pipeline(WaitStrategy::WaitIfStageEmpty, {1, 4, 2},
//   lgSlotCount);
//
//   // Producer threads (add items for stage 0 to consume):
//   auto guard = pipeline.push(batchSize, /*block=*/true);
//   for (auto& item : guard.batch) {
//     // Initialize item data
//   }
//   // Guard destructor publishes batch to stage 0 consumers
//
//   // Stage worker threads (process items and pass to next stage):
//   auto guard = pipeline.acquire(stage, thread, maxBatch, /*may_block=*/true);
//   for (auto& item : guard.batch) {
//     // Process item
//   }
//   // Guard destructor marks items as consumed and available to next stage
//
// Multi-Thread Stage Processing:
// When a stage has multiple threads (e.g., {1, 1, 1, 2} = 2 threads in stage
// 3):
//
// OVERLAPPING BATCHES - EACH THREAD SEES EVERY ENTRY:
// - Multiple threads in the same stage get OVERLAPPING batches from the ring
// buffer
// - Thread 0: calls acquire(3, 0) - gets batch from ring positions 100-110
// - Thread 1: calls acquire(3, 1) - gets batch from ring positions 100-110
// (SAME)
// - Both threads see the same entries and must coordinate processing
//
// PARTITIONING STRATEGIES:
// Choose your partitioning approach based on your use case:
//
// 1. Ring buffer position-based partitioning:
//   for (auto it = batch.begin(); it != batch.end(); ++it) {
//     if (it.index() % 2 != thread_index) continue; // Skip entries for other
//     threads process(*it); // Process only entries assigned to this thread
//   }
//
// 2. Entry content-based partitioning:
//   for (auto& item : guard.batch) {
//     if (hash(item.connection_id) % 2 != thread_index) continue;
//     process(item); // Process based on entry properties
//   }
//
// 3. Process all entries (when each thread does different work):
//   for (auto& item : guard.batch) {
//     process(item); // Both threads process all items, but differently
//   }
//
// Common Partitioning Patterns:
// - Position-based: it.index() % num_threads == thread_index
// - Hash-based: hash(item.key) % num_threads == thread_index
// - Type-based: item.type == MY_THREAD_TYPE
// - Load balancing: assign work based on thread load
// - All entries: each thread processes all items but performs different
// operations
//
// Note: it.index() returns the position in the ring buffer (0 to buffer_size-1)
//
// Memory Model:
// - Ring buffer size must be power of 2 for efficient masking
// - Actual ring slots accessed via: index & (slotCount - 1)
// - 128-byte aligned atomics prevent false sharing between CPU cache lines
//
// Thread Safety:
// - Fully lock-free using atomic operations with acquire/release memory
// ordering
// - Uses C++20 atomic wait/notify for efficient blocking when no work available
// - RAII guards ensure proper cleanup even with exceptions
template <class T> struct ThreadPipeline {
  // Constructor
  // wait_strategy: blocking behavior when no work is available
  // threads_per_stage: number of threads in each stage (e.g., {1, 4, 2})
  // lgSlotCount: log2 of ring buffer size (e.g., 10 -> 1024 slots)
  // Note: Producer threads are external to the pipeline and not counted in
  // threads_per_stage
  explicit ThreadPipeline(WaitStrategy wait_strategy,
                          std::vector<int> threads_per_stage, int lgSlotCount)
      : wait_strategy_(wait_strategy), topology_(std::move(threads_per_stage)),
        slot_count(1 << lgSlotCount), slot_count_mask(slot_count - 1),
        ring(slot_count), all_threads(topology_.total_threads) {
    // Otherwise we can't tell the difference between full and empty.
    assert(!(slot_count_mask & 0x80000000));
    initialize_all_threads();
  }

  ThreadPipeline(ThreadPipeline const &) = delete;
  ThreadPipeline &operator=(ThreadPipeline const &) = delete;
  ThreadPipeline(ThreadPipeline &&) = delete;
  ThreadPipeline &operator=(ThreadPipeline &&) = delete;

  struct Batch {
    Batch() : ring(), begin_(), end_() {}

    struct Iterator {
      using iterator_category = std::random_access_iterator_tag;
      using difference_type = std::ptrdiff_t;
      using value_type = T;
      using pointer = value_type *;
      using reference = value_type &;

      reference operator*() const {
        return (*ring)[index_ & (ring->size() - 1)];
      }
      pointer operator->() const {
        return &(*ring)[index_ & (ring->size() - 1)];
      }
      Iterator &operator++() {
        ++index_;
        return *this;
      }
      Iterator operator++(int) {
        auto tmp = *this;
        ++(*this);
        return tmp;
      }
      Iterator &operator--() {
        --index_;
        return *this;
      }
      Iterator operator--(int) {
        auto tmp = *this;
        --(*this);
        return tmp;
      }
      Iterator &operator+=(difference_type n) {
        index_ += n;
        return *this;
      }
      Iterator &operator-=(difference_type n) {
        index_ -= n;
        return *this;
      }
      Iterator operator+(difference_type n) const {
        return Iterator(index_ + n, ring);
      }
      Iterator operator-(difference_type n) const {
        return Iterator(index_ - n, ring);
      }
      difference_type operator-(const Iterator &rhs) const {
        assert(ring == rhs.ring);
        return static_cast<difference_type>(index_) -
               static_cast<difference_type>(rhs.index_);
      }
      friend Iterator operator+(difference_type n, const Iterator &iter) {
        return iter + n;
      }
      friend bool operator==(const Iterator &lhs, const Iterator &rhs) {
        assert(lhs.ring == rhs.ring);
        return lhs.index_ == rhs.index_;
      }
      friend bool operator!=(const Iterator &lhs, const Iterator &rhs) {
        assert(lhs.ring == rhs.ring);
        return lhs.index_ != rhs.index_;
      }
      friend bool operator<(const Iterator &lhs, const Iterator &rhs) {
        assert(lhs.ring == rhs.ring);
        return static_cast<int32_t>(lhs.index_ - rhs.index_) < 0;
      }
      friend bool operator<=(const Iterator &lhs, const Iterator &rhs) {
        assert(lhs.ring == rhs.ring);
        return static_cast<int32_t>(lhs.index_ - rhs.index_) <= 0;
      }
      friend bool operator>(const Iterator &lhs, const Iterator &rhs) {
        assert(lhs.ring == rhs.ring);
        return static_cast<int32_t>(lhs.index_ - rhs.index_) > 0;
      }
      friend bool operator>=(const Iterator &lhs, const Iterator &rhs) {
        assert(lhs.ring == rhs.ring);
        return static_cast<int32_t>(lhs.index_ - rhs.index_) >= 0;
      }

      uint32_t index() const { return index_ & (ring->size() - 1); }

    private:
      Iterator(uint32_t index, std::vector<T> *const ring)
          : index_(index), ring(ring) {}
      friend struct Batch;
      uint32_t index_;
      std::vector<T> *const ring;
    };

    [[nodiscard]] Iterator begin() { return Iterator(begin_, ring); }
    [[nodiscard]] Iterator end() { return Iterator(end_, ring); }

    [[nodiscard]] size_t size() const { return end_ - begin_; }
    [[nodiscard]] bool empty() const { return end_ == begin_; }
    T &operator[](uint32_t n) {
      return (*ring)[(begin_ + n) & (ring->size() - 1)];
    }

  private:
    friend struct ThreadPipeline;
    Batch(std::vector<T> *const ring, uint32_t begin_, uint32_t end_)
        : ring(ring), begin_(begin_), end_(end_) {}
    std::vector<T> *const ring;
    uint32_t begin_;
    uint32_t end_;
  };

private:
  WaitStrategy wait_strategy_;
  PipelineTopology topology_;

  alignas(128) std::atomic<uint32_t> slots{0};
  alignas(128) std::atomic<uint32_t> pushes{0};
  const uint32_t slot_count;
  const uint32_t slot_count_mask;

  std::vector<T> ring;
  std::vector<ThreadState> all_threads;

  void initialize_all_threads() {
    for (int stage = 0; stage < topology_.num_stages; ++stage) {
      init_stage_threads(stage);
    }
  }

  void init_stage_threads(int stage) {
    int stage_offset = topology_.stage_offset(stage);
    int stage_thread_count = topology_.threads_per_stage[stage];
    int prev_stage_threads = topology_.prev_stage_thread_count(stage);
    bool is_last_stage = (stage == topology_.num_stages - 1);

    for (int thread = 0; thread < stage_thread_count; ++thread) {
      auto &thread_state = all_threads[stage_offset + thread];
      thread_state.last_stage = is_last_stage;
      thread_state.last_push_read = std::vector<uint32_t>(prev_stage_threads);
    }
  }

  Batch acquire_helper(int stage, int thread, uint32_t maxBatch,
                       bool may_block) {
    int thread_idx = topology_.thread_index(stage, thread);
    auto &thread_state = all_threads[thread_idx];

    uint32_t begin = thread_state.local_pops & slot_count_mask;
    uint32_t len = PipelineAlgorithms::calculate_safe_len(
        wait_strategy_, topology_, stage, thread, all_threads, pushes,
        may_block);

    if (maxBatch != 0) {
      len = std::min(len, maxBatch);
    }
    if (len == 0) {
      return Batch{};
    }

    auto result = Batch{&ring, begin, begin + len};
    thread_state.local_pops += len;
    return result;
  }

public:
  struct StageGuard {
    Batch batch;

    ~StageGuard() {
      if (!batch.empty()) {
        PipelineAlgorithms::update_thread_pops(
            pipeline->wait_strategy_, pipeline->topology_, stage, thread,
            pipeline->all_threads, local_pops);
      }
    }

    StageGuard(StageGuard const &) = delete;
    StageGuard &operator=(StageGuard const &) = delete;
    StageGuard(StageGuard &&other) noexcept
        : batch(other.batch), local_pops(other.local_pops), stage(other.stage),
          thread(other.thread),
          pipeline(std::exchange(other.pipeline, nullptr)) {}
    StageGuard &operator=(StageGuard &&other) noexcept {
      batch = other.batch;
      local_pops = other.local_pops;
      stage = other.stage;
      thread = other.thread;
      pipeline = std::exchange(other.pipeline, nullptr);
      return *this;
    }

  private:
    friend struct ThreadPipeline;
    uint32_t local_pops;
    int stage;
    int thread;
    ThreadPipeline *pipeline;

    StageGuard(Batch batch, uint32_t local_pops, int stage, int thread,
               ThreadPipeline *pipeline)
        : batch(batch), local_pops(local_pops), stage(stage), thread(thread),
          pipeline(batch.empty() ? nullptr : pipeline) {}
  };

  struct ProducerGuard {
    Batch batch;

    ~ProducerGuard() {
      if (tp == nullptr) {
        return;
      }
      for (;;) {
        uint32_t p = tp->pushes.load(std::memory_order_acquire);
        if (p == old_slot) {
          break;
        }
        tp->pushes.wait(p, std::memory_order_relaxed);
      }
      tp->pushes.store(new_slot, std::memory_order_seq_cst);
      tp->pushes.notify_all();
    }

  private:
    friend struct ThreadPipeline;
    ProducerGuard() : batch(), tp() {}
    ProducerGuard(Batch batch, ThreadPipeline *tp, uint32_t old_slot,
                  uint32_t new_slot)
        : batch(batch), tp(tp), old_slot(old_slot), new_slot(new_slot) {}
    ThreadPipeline *const tp;
    uint32_t old_slot;
    uint32_t new_slot;
  };

  // Acquire a batch of items for processing by a consumer thread.
  // stage: which processing stage (0 = first consumer stage after producers)
  // thread: thread ID within the stage (0 to threads_per_stage[stage]-1)
  // maxBatch: maximum items to acquire (0 = no limit)
  // may_block: whether to block waiting for items (false = return empty batch
  // if none available) Returns: StageGuard with batch of items to process
  [[nodiscard]] StageGuard acquire(int stage, int thread, int maxBatch = 0,
                                   bool may_block = true) {
    auto batch = acquire_helper(stage, thread, maxBatch, may_block);

    int thread_idx = topology_.thread_index(stage, thread);
    uint32_t local_pops = all_threads[thread_idx].local_pops;

    return StageGuard{std::move(batch), local_pops, stage, thread, this};
  }

  // Reserve slots in the ring buffer for a producer thread to fill with items.
  // This is used by producer threads to add new items to stage 0 of the
  // pipeline.
  //
  // size: number of slots to reserve (must be > 0 and <= ring buffer capacity)
  // block: if true, blocks when ring buffer is full; if false, returns empty
  // guard Returns: ProducerGuard with exclusive access to reserved slots
  //
  // Usage: Fill items in the returned batch, then let guard destructor publish
  // them. The guard destructor ensures items are published in the correct
  // order.
  //
  // Preconditions:
  //   - size > 0 (must request at least one slot)
  //   - size <= slotCount (cannot request more slots than ring buffer capacity)
  // Violating preconditions results in program termination via abort().
  [[nodiscard]] ProducerGuard push(uint32_t const size, bool block) {
    if (size == 0) {
      std::abort();
    }
    if (size > slot_count) {
      std::abort();
    }

    uint32_t slot;
    uint32_t begin;
    for (;;) {
      slot = slots.load(std::memory_order_relaxed);
      begin = slot & slot_count_mask;

      int capacity_result = PipelineAlgorithms::check_producer_capacity(
          topology_, all_threads, slot, size, slot_count, block);
      if (capacity_result == 1) {
        continue;
      }
      if (capacity_result == 2) {
        return ProducerGuard{};
      }

      if (slots.compare_exchange_weak(slot, slot + size,
                                      std::memory_order_relaxed,
                                      std::memory_order_relaxed)) {
        break;
      }
    }
    return ProducerGuard{Batch{&ring, begin, begin + size}, this, slot,
                         slot + size};
  }
};