AyCode.Core/AyCode.Benchmark/Workloads/Scenarios/AcBinaryInMemoryRawByteArrayBenchmark.cs

using AyCode.Core.Serializers.Binaries;
using AyCode.Core.Tests.TestModels;
using System.Runtime.CompilerServices;

namespace AyCode.Core.Benchmarks.Workloads.Scenarios;

/// <summary>
/// Raw <c>byte[]</c> over an in-memory cross-thread handoff — NO transport (no NamedPipe, no Pipe, no
/// Channel<see langword="&lt;T&gt;"/>). Calling thread serialises into a fresh <c>byte[]</c>, hands it to a
/// background consumer task via a single byte[] slot + MRES pair; the consumer deserialises and signals done.
///
/// <para><b>Why this benchmark matters</b>: completes the 2x2 transport × wire-format matrix:</para>
/// <list type="bullet">
///   <item><description><b>NamedPipe + Chunked</b> = <see cref="AcBinaryNamedPipeBenchmark{T}"/></description></item>
///   <item><description><b>NamedPipe + Raw</b> = <see cref="AcBinaryNamedPipeRawByteArrayBenchmark{T}"/></description></item>
///   <item><description><b>In-memory Pipe + Chunked</b> = <see cref="AcBinaryInMemoryPipeBenchmark{T}"/></description></item>
///   <item><description><b>In-memory + Raw</b> = THIS row — apples-to-apples baseline for the in-memory chunked row</description></item>
/// </list>
/// <para>Side-by-side with <see cref="AcBinaryInMemoryPipeBenchmark{T}"/> this isolates the chunked-streaming
/// framework's pure CPU cost, with the same in-memory transport (zero kernel involvement) on both sides.
/// Side-by-side with <see cref="AcBinaryNamedPipeRawByteArrayBenchmark{T}"/> this isolates the kernel-NamedPipe
/// overhead on the raw-byte[] side.</para>
/// </summary>
public sealed class AcBinaryInMemoryRawByteArrayBenchmark<T> : ISerializerBenchmark, IDisposable where T : class
{
    private readonly T _order;
    private readonly AcBinarySerializerOptions _options;
    private readonly byte[] _serialized; // for SerializedSize reporting only

    // Long-lived consumer-task infrastructure (Deserialize on BG thread, signaled per iter).
    // No transport — just a byte[] slot for handoff between calling thread and consumer task.
    private readonly CancellationTokenSource _cts;
    private readonly Task _consumerTask;
    private readonly ManualResetEventSlim _consumeRequest = new(false);
    private readonly ManualResetEventSlim _consumeDone = new(false);
    private byte[]? _pendingBytes;             // calling thread → consumer task handoff slot
    private object? _lastResult;               // captured during VerifyRoundTrip; null in benchmark iters
    private bool _captureResult;
    private bool _disposed;

    public BenchmarkEngine Engine => BenchmarkEngine.AcBinary;
    public BenchmarkIoMode IoMode => BenchmarkIoMode.InMemoryRaw;
    public BenchmarkDispatchMode DispatchMode => _options.UseGeneratedCode ? BenchmarkDispatchMode.SGen : BenchmarkDispatchMode.Runtime;
    public Type OrderType => typeof(T);
    public string OptionsPreset { get; }
    public int SerializedSize => _serialized.Length;
    public long SetupSerializeAllocBytes { get; }
    public long SetupDeserializeAllocBytes { get; }
    public bool IsRoundTripOnly => true;
    public string OptionsDescription => BenchmarkOptions.BuildAcBinary(_options, $", BufferSize={_options.BufferWriterChunkSize}B, Transport=in-memory(raw,2-task)");

    public AcBinaryInMemoryRawByteArrayBenchmark(T order, AcBinarySerializerOptions options, string optionsPreset)
    {
        _order = order;
        _options = options;
        OptionsPreset = optionsPreset;

        _serialized = AcBinarySerializer.Serialize(order, _options);

        // === SERIALIZE-side setup measurement ===
        // Nothing to set up — calling thread allocates byte[] per iter via AcBinarySerializer.Serialize.
        SetupSerializeAllocBytes = 0;

        // === DESERIALIZE-side setup measurement ===
        // 1× background consumer-task + 2× MRES (request / done) + cancellation source.
        GC.Collect(); GC.WaitForPendingFinalizers(); GC.Collect();
        var beforeDes = GC.GetAllocatedBytesForCurrentThread();
        _cts = new CancellationTokenSource();
        _consumerTask = Task.Run(ConsumerLoop);
        var afterDes = GC.GetAllocatedBytesForCurrentThread();
        SetupDeserializeAllocBytes = afterDes - beforeDes;
    }

    // BG consumer: parks on _consumeRequest, picks up the byte[] from _pendingBytes, runs Deserialize<T>(bytes),
    // signals _consumeDone. Direct in-process handoff — no transport syscall, no buffer copy beyond the byte[]
    // reference itself (zero-copy by reference).
    private void ConsumerLoop()
    {
        var ct = _cts.Token;
        try
        {
            while (true)
            {
                _consumeRequest.Wait(ct);
                if (ct.IsCancellationRequested) return;
                _consumeRequest.Reset();

                try
                {
                    var bytes = _pendingBytes;
                    if (bytes != null)
                    {
                        var result = AcBinaryDeserializer.Deserialize<T>(bytes, _options);
                        if (_captureResult) _lastResult = result;
                    }
                }
                catch
                {
                    // Swallow — see ConsumerLoop in NamedPipe variant for rationale.
                }
                finally
                {
                    _consumeDone.Set();
                }
            }
        }
        catch (OperationCanceledException)
        {
            // Cooperative cancel — Dispose path. Swallow.
        }
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public void Serialize()
    {
        // 2-task in-memory pipeline:
        // 1. Calling thread serialises → fresh byte[] (per-iter alloc, matches AcBinaryBenchmark contract).
        // 2. Calling thread parks the byte[] into _pendingBytes and signals consumer task. Consumer task
        //    picks up the reference (zero-copy) and runs Deserialize<T>(bytes).
        // 3. Calling thread waits for _consumeDone (consumer task finished Des).
        //
        // Same architectural limitation as the NamedPipe-raw variant: Des cannot start until full bytes
        // are available. Only the per-iter Ser↔Des thread-handoff overlaps slightly (calling thread starts
        // signalling and waiting while consumer thread takes the byte[]).
        var bytes = AcBinarySerializer.Serialize(_order, _options);

        _pendingBytes = bytes;
        _consumeDone.Reset();
        _consumeRequest.Set();

        _consumeDone.Wait();
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public void Deserialize()
    {
        // No-op: per-iter round-trip is captured in Serialize(). See IsRoundTripOnly contract.
    }

    public bool VerifyRoundTrip()
    {
        _captureResult = true;
        try
        {
            Serialize();
            var result = _lastResult as T;
            return result != null && RoundTripValidator.DeepEqualsViaJson(_order, result);
        }
        finally
        {
            _captureResult = false;
            _lastResult = null;
        }
    }

    public void Dispose()
    {
        if (_disposed) return;
        _disposed = true;

        try { _cts.Cancel(); } catch { /* swallow on teardown */ }
        try { _consumeRequest.Set(); } catch { /* nudge in case consumer Wait is parked */ }
        try { _consumerTask.Wait(TimeSpan.FromSeconds(2)); } catch { /* swallow on teardown */ }

        try { _consumeRequest.Dispose(); } catch { /* swallow on teardown */ }
        try { _consumeDone.Dispose(); } catch { /* swallow on teardown */ }
        try { _cts.Dispose(); } catch { /* swallow on teardown */ }
    }
}