#include <gtest/gtest.h>
#include "vde/core/performance_tuning.h"
#include "vde/core/point.h"
#include <vector>
#include <string>
#include <thread>
#include <atomic>
#include <chrono>

using namespace vde::core;

// ===========================================================================
//  parallel_task_graph 测试
// ===========================================================================

TEST(PerformanceTuningTest, ParallelTaskGraph_BasicExecution) {
    std::atomic<int> counter{0};

    std::vector<TaskNode> tasks = {
        {0, "task0", [&]{ counter++; }, {}},
        {1, "task1", [&]{ counter++; }, {0}},   // 依赖 task0
        {2, "task2", [&]{ counter++; }, {0}},   // 依赖 task0
        {3, "task3", [&]{ counter++; }, {1,2}}, // 依赖 task1, task2
    };

    parallel_task_graph(tasks, 2);

    EXPECT_EQ(counter.load(), 4);
}

TEST(PerformanceTuningTest, ParallelTaskGraph_DiamondDependency) {
    std::vector<int> execution_order;
    std::mutex mtx;

    auto make_task = [&](int id, std::vector<int> deps) -> TaskNode {
        return {id, "task" + std::to_string(id),
            [&, id]{
                std::lock_guard<std::mutex> lk(mtx);
                execution_order.push_back(id);
            }, deps};
    };

    std::vector<TaskNode> tasks = {
        make_task(0, {}),
        make_task(1, {0}),
        make_task(2, {0}),
        make_task(3, {1, 2}),
        make_task(4, {3}),
    };

    parallel_task_graph(tasks, 4);

    ASSERT_EQ(execution_order.size(), 5u);
    // task0 must be first
    EXPECT_EQ(execution_order[0], 0);
    // task4 must be last
    EXPECT_EQ(execution_order[4], 4);
}

TEST(PerformanceTuningTest, ParallelTaskGraph_CycleDetection) {
    std::vector<TaskNode> tasks = {
        {0, "task0", []{}, {1}},  // 依赖 task1
        {1, "task1", []{}, {0}},  // 依赖 task0 → 循环
    };

    EXPECT_THROW({
        parallel_task_graph(tasks, 2);
    }, std::runtime_error);
}

TEST(PerformanceTuningTest, ParallelTaskGraph_EmptyTasks) {
    std::vector<TaskNode> empty;
    EXPECT_NO_THROW(parallel_task_graph(empty, 1));
}

// ===========================================================================
//  WorkStealingScheduler 测试
// ===========================================================================

TEST(PerformanceTuningTest, WorkStealingScheduler_BasicSubmit) {
    auto& scheduler = work_stealing_scheduler();
    std::atomic<int> counter{0};

    for (int i = 0; i < 100; ++i) {
        scheduler.submit([&]{ counter++; });
    }

    scheduler.wait_all();
    EXPECT_EQ(counter.load(), 100);
}

TEST(PerformanceTuningTest, WorkStealingScheduler_PriorityTasks) {
    auto& scheduler = work_stealing_scheduler();
    std::vector<int> completion_order;
    std::mutex mtx;

    // 提交低优先级任务
    scheduler.submit([&]{
        std::this_thread::sleep_for(std::chrono::milliseconds(50));
        std::lock_guard<std::mutex> lk(mtx);
        completion_order.push_back(0);
    }, 0);

    // 提交高优先级任务（应优先执行）
    scheduler.submit([&]{
        std::lock_guard<std::mutex> lk(mtx);
        completion_order.push_back(1);
    }, 10);

    scheduler.wait_all();
    EXPECT_EQ(completion_order.size(), 2u);
    EXPECT_EQ(completion_order[0], 1) << "High priority should complete first";
}

// ===========================================================================
//  MemoryPoolIntegration 测试
// ===========================================================================

TEST(PerformanceTuningTest, MemoryPool_AllocateDeallocate) {
    auto& pool = memory_pool_integration();
    pool.reset();

    auto* pt = pool.allocate_point3d();
    ASSERT_NE(pt, nullptr);
    *pt = Point3D(1, 2, 3);
    EXPECT_DOUBLE_EQ(pt->x(), 1.0);

    pool.deallocate_point3d(pt);

    auto stats = pool.stats();
    EXPECT_GT(stats.total_allocations + stats.cache_hits, 0u);
    pool.reset();
}

TEST(PerformanceTuningTest, MemoryPool_Reuse) {
    auto& pool = memory_pool_integration();
    pool.reset();
    pool.warm_up(10);

    auto* pt1 = pool.allocate_point3d();
    auto* pt2 = pool.allocate_point3d();

    EXPECT_NE(pt1, pt2) << "Should get different pointers";

    pool.deallocate_point3d(pt1);
    pool.deallocate_point3d(pt2);

    auto* pt3 = pool.allocate_point3d();
    // pt3 很可能来自池（pt1 或 pt2）
    EXPECT_TRUE(pt3 == pt1 || pt3 == pt2) << "Should reuse from pool";

    pool.reset();
}

TEST(PerformanceTuningTest, MemoryPool_GenericAllocate) {
    auto& pool = memory_pool_integration();
    pool.reset();
    pool.set_pool_size(64);

    void* buf = pool.allocate(128);
    ASSERT_NE(buf, nullptr);

    pool.deallocate(buf, 128);

    auto stats = pool.stats();
    EXPECT_GE(stats.total_allocations, 1u);
    pool.reset();
}

// ===========================================================================
//  cache_optimization_hints 测试
// ===========================================================================

TEST(PerformanceTuningTest, CacheOptimizationHints_ReturnsValid) {
    auto hints = cache_optimization_hints();

    EXPECT_GT(hints.l1_cache_size, 0u);
    EXPECT_GT(hints.l2_cache_size, 0u);
    EXPECT_GT(hints.cache_line_size, 0u);
    EXPECT_EQ(hints.cache_line_size, 64u) << "Expected 64-byte cache line";
    EXPECT_GE(hints.numa_node_count, 1);
}

TEST(PerformanceTuningTest, Prefetch_Compiles) {
    int value = 42;
    // 预取不应崩溃
    EXPECT_NO_THROW(prefetch(&value));
    EXPECT_NO_THROW(prefetch_write(&value));
}

TEST(PerformanceTuningTest, PaddedAtomic_Alignment) {
    PaddedAtomic<> padded;
    padded.value = 42;
    EXPECT_EQ(padded.value.load(), 42);
    // 验证对齐
    EXPECT_EQ(sizeof(padded), CACHE_LINE_SIZE);
    EXPECT_EQ(alignof(decltype(padded)), CACHE_LINE_SIZE);
}

// ===========================================================================
//  profile_guided_layout 测试
// ===========================================================================

TEST(PerformanceTuningTest, ProfileGuidedLayout_HotColdSplit) {
    std::vector<AccessRecord> records = {
        {"field_a", 1000, 50, 0.0},
        {"field_b", 100,  5,  0.0},
        {"field_c", 10,   1,  0.0},
    };

    auto plan = profile_guided_layout(records, "TestStruct");

    EXPECT_GT(plan.hot_fields.size(), 0u);
    EXPECT_GT(plan.cold_fields.size(), 0u);
    EXPECT_GT(plan.estimated_improvement, 0.0)
        << "Should estimate performance improvement";
}

TEST(PerformanceTuningTest, ProfileGuidedLayout_EmptyRecords) {
    std::vector<AccessRecord> empty;
    auto plan = profile_guided_layout(empty, "Empty");
    EXPECT_EQ(plan.hot_fields.size(), 0u);
    EXPECT_EQ(plan.cold_fields.size(), 0u);
}

// ===========================================================================
//  Point3DSoA 测试
// ===========================================================================

TEST(PerformanceTuningTest, Point3DSoA_ConversionRoundtrip) {
    std::vector<Point3D> original = {
        Point3D(1, 2, 3),
        Point3D(4, 5, 6),
        Point3D(7, 8, 9),
        Point3D(10, 11, 12),
    };

    auto soa = Point3DSoA::from_aos(original);
    EXPECT_EQ(soa.size(), 4u);
    EXPECT_DOUBLE_EQ(soa.x[0], 1.0);
    EXPECT_DOUBLE_EQ(soa.y[1], 5.0);
    EXPECT_DOUBLE_EQ(soa.z[2], 9.0);

    auto restored = soa.to_aos();
    EXPECT_EQ(restored.size(), 4u);
    for (size_t i = 0; i < original.size(); ++i) {
        EXPECT_DOUBLE_EQ(restored[i].x(), original[i].x());
        EXPECT_DOUBLE_EQ(restored[i].y(), original[i].y());
        EXPECT_DOUBLE_EQ(restored[i].z(), original[i].z());
    }
}

TEST(PerformanceTuningTest, Point3DSoA_Empty) {
    std::vector<Point3D> empty;
    auto soa = Point3DSoA::from_aos(empty);
    EXPECT_EQ(soa.size(), 0u);

    auto restored = soa.to_aos();
    EXPECT_EQ(restored.size(), 0u);
}