FEX-Emu--FEX/FEXCore/Source/Interface/Core/LookupCache.h

// SPDX-License-Identifier: MIT
#pragma once
#include "Interface/Context/Context.h"
#include <FEXCore/Utils/LogManager.h>
#include <FEXCore/fextl/map.h>
#include <FEXCore/fextl/memory_resource.h>
#include <FEXCore/fextl/robin_map.h>
#include <FEXCore/fextl/vector.h>
#include <FEXCore/fextl/memory_resource.h>

#include <cstdint>
#include <stddef.h>
#include <utility>
#include <mutex>

namespace FEXCore {

struct LookupCacheWriteLockToken {
private:
  // Only constructible by GuestToHostMap
  friend struct GuestToHostMap;
  LookupCacheWriteLockToken(std::mutex& Mutex)
    : Lock {Mutex} {}
  std::lock_guard<std::mutex> Lock;
};

struct GuestToHostMap {
  std::mutex WriteLock;

  [[nodiscard]]
  LookupCacheWriteLockToken AcquireWriteLock() {
    return LookupCacheWriteLockToken {WriteLock};
  }

  struct BlockLinkTag {
    uint64_t GuestDestination;
    FEXCore::Context::ExitFunctionLinkData* HostLink;

    bool operator<(const BlockLinkTag& other) const {
      if (GuestDestination < other.GuestDestination) {
        return true;
      } else if (GuestDestination == other.GuestDestination) {
        return HostLink < other.HostLink;
      } else {
        return false;
      }
    }
  };

  // Use a monotonic buffer resource to allocate both the std::pmr::map and its members.
  // This allows us to quickly clear the block link map by clearing the monotonic allocator.
  // If we had allocated the block link map without the MBR, then clearing the map would require slowly
  // walking each block member and destructing objects.
  //
  // This makes `BlockLinks` look like a raw pointer that could memory leak, but since it is backed by the MBR, it won't.
  fextl::pmr::named_monotonic_page_buffer_resource BlockLinks_mbr;
  using BlockLinksMapType = std::pmr::map<BlockLinkTag, FEXCore::Context::BlockDelinkerFunc>;
  fextl::unique_ptr<std::pmr::polymorphic_allocator<std::byte>> BlockLinks_pma;
  BlockLinksMapType* BlockLinks;

  fextl::robin_map<uint64_t, uint64_t> BlockList;

  fextl::map<uint64_t, fextl::vector<uint64_t>> CodePages;

  GuestToHostMap();

  // Adds to Guest -> Host code mapping
  void AddBlockMapping(uint64_t Address, void* HostCode, const LookupCacheWriteLockToken&) {
    // This may replace an existing mapping
    // NOTE: Generally no previous entry should exist, however there is one exception:
    //       If the backend updates the active thread's CodeBuffer, the new associated LookupCache
    //       may already contain the block address. Since is comparatively rare, we'll just leak
    //       one of the two blocks in this case.
    BlockList[Address] = (uintptr_t)HostCode;
  }

  std::optional<uintptr_t> FindBlock(uint64_t Address, const LookupCacheWriteLockToken&) {
    auto HostCode = BlockList.find(Address);
    if (HostCode == BlockList.end()) {
      return std::nullopt;
    }
    return HostCode->second;
  }

  bool Erase(FEXCore::Core::CpuStateFrame* Frame, uint64_t Address, const LookupCacheWriteLockToken&) {
    // Sever any links to this block
    auto lower = BlockLinks->lower_bound({Address, nullptr});
    auto upper = BlockLinks->upper_bound({Address, reinterpret_cast<FEXCore::Context::ExitFunctionLinkData*>(UINTPTR_MAX)});
    for (auto it = lower; it != upper; it = BlockLinks->erase(it)) {
      it->second(Frame, it->first.HostLink);
    }

    // Remove from BlockList
    return BlockList.erase(Address) != 0;
  }

  void AddBlockLink(uint64_t GuestDestination, FEXCore::Context::ExitFunctionLinkData* HostLink,
                    const FEXCore::Context::BlockDelinkerFunc& delinker, const LookupCacheWriteLockToken&) {
    BlockLinks->insert({{GuestDestination, HostLink}, delinker});
  }

  bool AddBlockExecutableRange(const fextl::set<uint64_t>& Addresses, uint64_t Start, uint64_t Length, const LookupCacheWriteLockToken&) {
    bool rv = false;

    for (auto CurrentPage = Start >> 12, EndPage = (Start + Length - 1) >> 12; CurrentPage <= EndPage; CurrentPage++) {
      auto& CodePage = CodePages[CurrentPage];
      rv |= CodePage.empty();
      CodePage.insert(CodePage.end(), Addresses.begin(), Addresses.end());
    }

    return rv;
  }

  void ClearCache(const LookupCacheWriteLockToken&);
};

class LookupCache {
public:
  struct LookupCacheEntry {
    uintptr_t HostCode;
    uintptr_t GuestCode;
  };

  LookupCache(FEXCore::Context::ContextImpl* CTX);
  ~LookupCache();

  // Swaps out the underlying GuestToHostMap and clears all associated caches.
  // This interface requires the previous CodeBuffer to be provided despite not using it. This ensures the shared write lock is still valid.
  void ChangeGuestToHostMapping([[maybe_unused]] CPU::CodeBuffer& Prev, GuestToHostMap& NewMap, const LookupCacheWriteLockToken& lk) {
    ClearThreadLocalCaches(lk);
    Shared = &NewMap;
  }

  uintptr_t FindBlock(FEXCore::Core::InternalThreadState* Thread, uint64_t Address) {
    // Try L1, no lock needed
    auto& L1Entry = reinterpret_cast<LookupCacheEntry*>(L1Pointer)[Address & L1PointerMask];
    if (L1Entry.GuestCode == Address) {
      return L1Entry.HostCode;
    }

    // L2 and L3 need to be locked
    uintptr_t HostPtr {};
    {
      auto lk = Shared->AcquireWriteLock();

      if (!DisableL2Cache()) {
        // Try L2
        const auto PageIndex = (Address & (VirtualMemSize - 1)) >> 12;
        const auto PageOffset = Address & (0x0FFF);

        const auto Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
        auto LocalPagePointer = Pointers[PageIndex];

        // Do we a page pointer for this address?
        if (LocalPagePointer) {
          // Find there pointer for the address in the blocks
          auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);

          if (BlockPointers[PageOffset].GuestCode == Address) {
            L1Entry.GuestCode = Address;
            L1Entry.HostCode = BlockPointers[PageOffset].HostCode;
            HostPtr = L1Entry.HostCode;
          }
        }
      }

      if (!HostPtr) {
        // Try L3
        auto HostCode = Shared->FindBlock(Address, lk);
        if (HostCode) {
          CacheBlockMapping(Address, HostCode.value(), lk);
          HostPtr = HostCode.value();
        }
      }
    }

    if (HostPtr && DynamicL1Cache()) {
      UpdateDynamicL1Stats(Thread);
    }

    return HostPtr;
  }

  void UpdateDynamicL1Stats(FEXCore::Core::InternalThreadState* Thread) {
    // If host pointer was found in L2 or L3, then add it to the counter.
    // Keeping track not L1 misses, but specifically L2/L3 hits.
    ++L2L3CacheHits;

    const auto CurrentTime = std::chrono::system_clock::now();
    const auto Period = CurrentTime - LastPeriod;
    if (Period >= SamplePeriod) {
      // If larger than the sample period then check if we need to increase L1 cache size.
      const double AveragePerSecond = static_cast<double>(L2L3CacheHits) /
                                      static_cast<double>(std::chrono::duration_cast<std::chrono::milliseconds>(Period).count()) * 1000.0;

      if (AveragePerSecond >= DynamicL1CacheIncreaseCountHeuristic()) {
        if (CurrentL1Entries < MAX_L1_ENTRIES) {
          CurrentL1Entries <<= 1;
          L1PointerMask = CurrentL1Entries - 1;

          // Update the thread's L1 pointer mask to increase how much cache it uses.
          // Since we're in C-code, this is safe to update here.
          Thread->CurrentFrame->State.L1Mask = GetScaledL1PointerMask();
        }
      } else if (AveragePerSecond < DynamicL1CacheDecreaseCountHeuristic()) {
        if (CurrentL1Entries > MIN_L1_ENTRIES) {
          CurrentL1Entries >>= 1;
          L1PointerMask = CurrentL1Entries - 1;

          // Madvise the entries that we are dropping. Gives the memory back to the OS.
          LookupCacheEntry* FirstZeroL1Entry = &reinterpret_cast<LookupCacheEntry*>(L1Pointer)[CurrentL1Entries];
          size_t ZeroMemorySize = (MAX_L1_ENTRIES - CurrentL1Entries) * sizeof(LookupCacheEntry);
          FEXCore::Allocator::VirtualDontNeed(FirstZeroL1Entry, ZeroMemorySize, false);

          // Update the thread's L1 pointer mask to increase how much cache it uses.
          // Since we're in C-code, this is safe to update here.
          Thread->CurrentFrame->State.L1Mask = GetScaledL1PointerMask();
        }
      }

      // Update Last period to start again.
      LastPeriod = CurrentTime;
      L2L3CacheHits = 0;
    }
  }

  GuestToHostMap* Shared = nullptr;

  // Appends a list of Block {Address} to CodePages [Start, Start + Length)
  // Returns true if new pages are marked as containing code
  bool AddBlockExecutableRange(const fextl::set<uint64_t>& Addresses, uint64_t Start, uint64_t Length) {
    auto lk = Shared->AcquireWriteLock();
    return Shared->AddBlockExecutableRange(Addresses, Start, Length, lk);
  }

  // Adds to Guest -> Host code mapping
  void AddBlockMapping(uint64_t Address, void* HostCode) {
    auto lk = Shared->AcquireWriteLock();

    Shared->AddBlockMapping(Address, HostCode, lk);

    // There is no need to update L1 or L2, they will get updated on first lookup
    // However, adding to L1 here increases performance
    auto& L1Entry = reinterpret_cast<LookupCacheEntry*>(L1Pointer)[Address & L1PointerMask];
    L1Entry.GuestCode = Address;
    L1Entry.HostCode = (uintptr_t)HostCode;
  }

  // NOTE: It's the caller's responsibility to call Erase() for all other
  //       GuestToHostMaps that share the same LookupCache. Otherwise, the
  //       L1/L2 caches will contain stale references to deallocated memory.
  bool Erase(FEXCore::Core::CpuStateFrame* Frame, uint64_t Address, const LookupCacheWriteLockToken& lk) {
    bool ErasedAny = Shared->Erase(Frame, Address, lk);

    // Do L1
    auto& L1Entry = reinterpret_cast<LookupCacheEntry*>(L1Pointer)[Address & L1PointerMask];
    if (L1Entry.GuestCode == Address) {
      L1Entry.GuestCode = 0;
      ErasedAny = true;
      // Leave L1Entry.HostCode as is, so that concurrent lookups won't read a null pointer
      // This is a soft guarantee for cross thread invalidation, as atomics are not used
      // and it hasn't been thoroughly tested
    }

    if (!DisableL2Cache()) {
      // Do full map
      Address = Address & (VirtualMemSize - 1);
      uint64_t PageOffset = Address & (0x0FFF);
      Address >>= 12;

      uintptr_t* Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
      uint64_t LocalPagePointer = Pointers[Address];
      if (!LocalPagePointer) {
        // Page for this code didn't even exist, nothing to do
        return ErasedAny;
      }

      // Page exists, just set the offset to zero
      auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);
      BlockPointers[PageOffset].GuestCode = 0;
      BlockPointers[PageOffset].HostCode = 0;
    }
    return true;
  }

  void AddBlockLink(uint64_t GuestDestination, FEXCore::Context::ExitFunctionLinkData* HostLink,
                    const FEXCore::Context::BlockDelinkerFunc& delinker, const LookupCacheWriteLockToken& lk) {
    Shared->AddBlockLink(GuestDestination, HostLink, delinker, lk);
  }

  void ClearCache(const LookupCacheWriteLockToken&);
  void ClearL2Cache(const LookupCacheWriteLockToken&);
  void ClearThreadLocalCaches(const LookupCacheWriteLockToken&);

  uintptr_t GetL1Pointer() const {
    return L1Pointer;
  }
  uintptr_t GetScaledL1PointerMask() const {
    return L1PointerMask << FEXCore::ilog2(sizeof(LookupCache::LookupCacheEntry));
  }
  uintptr_t GetPagePointer() const {
    return PagePointer;
  }
  uintptr_t GetVirtualMemorySize() const {
    return VirtualMemSize;
  }

  // This needs to be taken before reads or writes to L2, L3, CodePages,
  // and before writes to L1. Concurrent access from a thread that this LookupCache doesn't belong to
  // may only happen during cross thread invalidation (::Erase).
  // All other operations must be done from the owning thread.
  // Some care is taken so that L1 lookups can be done without locks, and even tearing is unlikely to lead to a crash.
  // This approach has not been fully vetted yet.
  // Also note that L1 lookups might be inlined in the JIT Dispatcher and/or block ends.
  auto AcquireWriteLock() {
    return Shared->AcquireWriteLock();
  }

private:
  void CacheBlockMapping(uint64_t Address, uintptr_t HostCode, const LookupCacheWriteLockToken& lk) {
    // Do L1
    auto& L1Entry = reinterpret_cast<LookupCacheEntry*>(L1Pointer)[Address & L1PointerMask];
    L1Entry.GuestCode = Address;
    L1Entry.HostCode = HostCode;

    if (!DisableL2Cache()) {
      // Do ful map
      auto FullAddress = Address;
      Address = Address & (VirtualMemSize - 1);

      uint64_t PageOffset = Address & (0x0FFF);
      Address >>= 12;

      uintptr_t* Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
      uint64_t LocalPagePointer = Pointers[Address];
      if (!LocalPagePointer) {
        // We don't have a page pointer for this address
        // Allocate one now if we can
        uintptr_t NewPageBacking = AllocateBackingForPage();
        if (!NewPageBacking) {
          // Couldn't allocate, clear L2 and retry
          ClearL2Cache(lk);
          CacheBlockMapping(Address, HostCode, lk);
          return;
        }
        Pointers[Address] = NewPageBacking;
        LocalPagePointer = NewPageBacking;
      }

      // Add the new pointer to the page block
      auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);

      // This silently replaces existing mappings
      BlockPointers[PageOffset].GuestCode = FullAddress;
      BlockPointers[PageOffset].HostCode = HostCode;
    }
  }

  uintptr_t AllocateBackingForPage() {
    uintptr_t NewBase = AllocateOffset;
    uintptr_t NewEnd = AllocateOffset + SIZE_PER_PAGE;

    if (NewEnd >= CODE_SIZE) {
      // We ran out of block backing space. Need to clear the block cache and tell the JIT cores to clear their caches as well
      // Tell whatever is calling this that it needs to do it.
      return 0;
    }

    AllocateOffset = NewEnd;
    return PageMemory + NewBase;
  }

  uintptr_t PagePointer;
  uintptr_t PageMemory;
  uintptr_t L1Pointer;
  uintptr_t L1PointerMask;

  size_t TotalCacheSize;

  // Start with 8k entries in L1 to give 128KB of L1 cache to each thread.
  // Max out at 1 million entries to give each thread 16MB of L1 cache maximum.
  constexpr static size_t MIN_L1_ENTRIES = 8 * 1024;        // Must be a power of 2
  constexpr static size_t MAX_L1_ENTRIES = 1 * 1024 * 1024; // Must be a power of 2

  constexpr static size_t CODE_SIZE = 128 * 1024 * 1024;
  constexpr static size_t SIZE_PER_PAGE = FEXCore::Utils::FEX_PAGE_SIZE * sizeof(LookupCacheEntry);
  constexpr static size_t MAX_L1_SIZE = MAX_L1_ENTRIES * sizeof(LookupCacheEntry);

  size_t AllocateOffset {};

  FEXCore::Context::ContextImpl* ctx;
  uint64_t VirtualMemSize {};

  size_t CurrentL1Entries = MIN_L1_ENTRIES;
  uint64_t L2L3CacheHits {};
  std::chrono::time_point<std::chrono::system_clock> LastPeriod {};
  constexpr static std::chrono::seconds SamplePeriod {1};
  FEX_CONFIG_OPT(DynamicL1CacheIncreaseCountHeuristic, DYNAMICL1CACHEINCREASECOUNTHEURISTIC);
  FEX_CONFIG_OPT(DynamicL1CacheDecreaseCountHeuristic, DYNAMICL1CACHEDECREASECOUNTHEURISTIC);

  FEX_CONFIG_OPT(DynamicL1Cache, DYNAMICL1CACHE);
  FEX_CONFIG_OPT(DisableL2Cache, DISABLEL2CACHE);
};
} // namespace FEXCore