// Copyright (c) 2024-2026 The DragonX developers
// Distributed under the GPLv3 software license, see the accompanying
// file COPYING or https://www.gnu.org/licenses/gpl-3.0.en.html
//
// Consensus-equivalence test for the parallel RandomX pre-verification pool. The pool is purely an
// optimization: a block's transient fRandomXVerified flag (set by CRandomXCheck on a real hash
// match) only lets CheckBlockHeader SKIP the inline recompute. So for every block the pool's
// outcome must equal the inline CheckRandomXSolution outcome — `(preVerified || inline) == inline`.
// We exercise a valid solution, a corrupted solution, and confirm the pool never "succeeds" on a
// block the inline check would reject.

#include <gtest/gtest.h>
#include <string>
#include <vector>

#include "arith_uint256.h"
#include "chain.h"
#include "chainparams.h"
#include "pow.h"
#include "primitives/block.h"
#include "RandomX/src/randomx.h"
#include "hush_defs.h"
#include "util.h"
#include <boost/thread.hpp>
#include <memory>

extern int32_t HUSH_LOADINGBLOCKS;
extern bool fCheckpointsEnabled;

namespace {
// Compute the correct RandomX solution for a header using a standalone reference light VM, via the
// SAME key + input helpers the validator uses (so the bytes/key match exactly).
void ReferenceRandomXHash(const CBlockHeader& hdr, const std::string& key, unsigned char out[RANDOMX_HASH_SIZE])
{
    std::vector<unsigned char> in = GetRandomXInput(hdr);
    randomx_flags flags = randomx_get_flags();
    randomx_cache* c = randomx_alloc_cache(flags);
    ASSERT_NE(c, nullptr);
    randomx_init_cache(c, key.data(), key.size());
    randomx_vm* vm = randomx_create_vm(flags, c, nullptr);
    ASSERT_NE(vm, nullptr);
    randomx_calculate_hash(vm, in.data(), in.size(), out);
    randomx_destroy_vm(vm);
    randomx_release_cache(c);
}
} // namespace

TEST(RandomXPreVerify, ConsensusEquivalence)
{
    // Force RandomX validation to actually run at low heights in the test harness.
    uint32_t savedAlgo = ASSETCHAINS_ALGO, savedRx = ASSETCHAINS_RANDOMX;
    int32_t  savedVal  = ASSETCHAINS_RANDOMX_VALIDATION, savedLoad = HUSH_LOADINGBLOCKS;
    bool     savedCkpt = fCheckpointsEnabled;
    ASSETCHAINS_RANDOMX            = 2;                 // a distinct nonzero algo id
    ASSETCHAINS_ALGO              = ASSETCHAINS_RANDOMX;
    ASSETCHAINS_RANDOMX_VALIDATION = 1;                // enforce from height 1
    HUSH_LOADINGBLOCKS            = 0;                  // not in initial-load (else RandomX skipped)
    fCheckpointsEnabled          = false;              // avoid the below-checkpoint skip

    const int32_t height = 10; // < interval+lag -> the chain-params initial key (no chainActive needed)

    CBlockHeader hdr;
    hdr.nVersion             = 4;
    hdr.hashPrevBlock        = uint256S("0x0000000000000000000000000000000000000000000000000000000000000001");
    hdr.hashMerkleRoot       = uint256S("0x0000000000000000000000000000000000000000000000000000000000000002");
    hdr.hashFinalSaplingRoot = uint256S("0x0000000000000000000000000000000000000000000000000000000000000003");
    hdr.nTime                = 1700000000;
    hdr.nBits                = 0x200f0f0f;
    hdr.nNonce               = uint256S("0x0000000000000000000000000000000000000000000000000000000000000004");

    std::string key = GetRandomXKey(height);
    ASSERT_FALSE(key.empty());

    unsigned char good[RANDOMX_HASH_SIZE];
    ReferenceRandomXHash(hdr, key, good);

    // Run the pool path synchronously on this thread (CRandomXCheck creates its own thread_local VM).
    auto poolVerifies = [&](const CBlockHeader& h) -> bool {
        RandomXValidatorPrepareKey(key); // load the shared cache with this key
        bool slot = false;
        CRandomXCheck chk(key, GetRandomXInput(h), h.nSolution.data(), &slot);
        chk();
        return slot;
    };

    // Case 1 — valid solution: both inline and pool accept; equivalence holds.
    hdr.nSolution.assign(good, good + RANDOMX_HASH_SIZE);
    EXPECT_TRUE(CheckRandomXSolution(&hdr, height));
    EXPECT_TRUE(poolVerifies(hdr));
    EXPECT_EQ(poolVerifies(hdr) || CheckRandomXSolution(&hdr, height), CheckRandomXSolution(&hdr, height));

    // Case 2 — corrupted solution: both reject; the pool must NOT set verified.
    {
        CBlockHeader bad = hdr;
        bad.nSolution[0] ^= 0xff;
        EXPECT_FALSE(CheckRandomXSolution(&bad, height));
        EXPECT_FALSE(poolVerifies(bad));
        EXPECT_EQ(poolVerifies(bad) || CheckRandomXSolution(&bad, height), CheckRandomXSolution(&bad, height));
    }

    // Case 3 — a verified flag on the block lets CheckBlockHeader skip, but verified is only ever set
    // by a real hash match, so it can never mask an invalid block. (Pool returns false for the bad
    // block above, so its fRandomXVerified stays false and the inline path rejects it at connect.)

    ASSETCHAINS_ALGO = savedAlgo; ASSETCHAINS_RANDOMX = savedRx;
    ASSETCHAINS_RANDOMX_VALIDATION = savedVal; HUSH_LOADINGBLOCKS = savedLoad;
    fCheckpointsEnabled = savedCkpt;
}

// A/B: serial inline verification (single VM) vs the parallel pool (worker threads). Directly
// measures the speedup the pool delivers. We don't care about validity here (mismatched solutions
// still cost a full hash), only wall-clock. parallel must beat serial whenever >1 core is used.
TEST(RandomXPreVerify, ParallelSpeedup)
{
    uint32_t savedAlgo = ASSETCHAINS_ALGO, savedRx = ASSETCHAINS_RANDOMX;
    int32_t  savedVal  = ASSETCHAINS_RANDOMX_VALIDATION, savedLoad = HUSH_LOADINGBLOCKS;
    bool     savedCkpt = fCheckpointsEnabled;
    ASSETCHAINS_RANDOMX = 2; ASSETCHAINS_ALGO = ASSETCHAINS_RANDOMX;
    ASSETCHAINS_RANDOMX_VALIDATION = 1; HUSH_LOADINGBLOCKS = 0; fCheckpointsEnabled = false;

    const int32_t height = 10;
    std::string key = GetRandomXKey(height);
    ASSERT_FALSE(key.empty());
    ASSERT_TRUE(RandomXValidatorPrepareKey(key));

    const int M = 16; // blocks to verify in the window
    std::vector<CBlockHeader> hdrs(M);
    for (int i = 0; i < M; i++) {
        hdrs[i].nVersion = 4;
        hdrs[i].nTime    = 1700000000 + i;
        hdrs[i].nBits    = 0x200f0f0f;
        hdrs[i].nNonce   = ArithToUint256(arith_uint256(i + 1)); // distinct inputs
        hdrs[i].nSolution.assign(RANDOMX_HASH_SIZE, 0);          // arbitrary; we time the hash
    }

    // Serial baseline: inline single-VM verification (each call hashes, then mismatches -> false).
    int64_t t0 = GetTimeMicros();
    for (int i = 0; i < M; i++) CheckRandomXSolution(&hdrs[i], height);
    int64_t serialUs = GetTimeMicros() - t0;

    // Parallel: spawn K-1 workers + the master (this thread) joining via Wait().
    int K = std::min(8, std::max(2, (int)boost::thread::hardware_concurrency()));
    boost::thread_group workers;
    for (int i = 0; i < K - 1; i++) workers.create_thread(&ThreadRandomXVerify);

    std::unique_ptr<bool[]> slots(new bool[M]());
    std::vector<CRandomXCheck> checks;
    checks.reserve(M);
    for (int i = 0; i < M; i++)
        checks.push_back(CRandomXCheck(key, GetRandomXInput(hdrs[i]), hdrs[i].nSolution.data(), &slots[i]));

    int64_t t1 = GetTimeMicros();
    {
        CCheckQueueControl<CRandomXCheck> control(&rxCheckQueue);
        control.Add(checks);
        control.Wait();
    }
    int64_t parallelUs = GetTimeMicros() - t1;

    workers.interrupt_all();
    workers.join_all();

    printf("[ RandomX A/B ] %d blocks: serial(1 VM)=%ldms, parallel(%d threads)=%ldms, speedup=%.1fx\n",
           M, (long)(serialUs / 1000), K, (long)(parallelUs / 1000),
           (double)serialUs / (double)std::max<int64_t>(1, parallelUs));

    EXPECT_LT(parallelUs, serialUs); // parallel must be faster than serial on a multi-core box

    ASSETCHAINS_ALGO = savedAlgo; ASSETCHAINS_RANDOMX = savedRx;
    ASSETCHAINS_RANDOMX_VALIDATION = savedVal; HUSH_LOADINGBLOCKS = savedLoad;
    fCheckpointsEnabled = savedCkpt;
}