#if SPECIAL
#include "main.h"
#endif

static void
Memory_InitTable(global_memory *GlobalMemory, memory *Memory, uint64 Size, memory_table_list TableName, char *Name, uint64 Block_ElementSize = 0) {
    memory_table *Table = &Memory->Slot[TableName];
    Table->Name = Name;
    Table->Address = (ptrsize *)((uint8 *)GlobalMemory->Address + GlobalMemory->CurrentPosition);
    // Table->Address = malloc(Size);
    Table->Size = Size;
    Table->Block_ElementSize = Block_ElementSize;
    GlobalMemory->CurrentPosition += Size;
}

static uint32
Memory_Block_AllocateNew(memory *Memory, memory_table_list TableName)
{
    memory_table *Table = &Memory->Slot[TableName];
    Assert(Table->Block_ElementSize != 0);
    bool32 Empty = 0;
    uint32 Index = 0;
    uint8 *Address_Playhead = (uint8 *)Table->Address;
    while (*Address_Playhead != 0) {
        Address_Playhead += Table->Block_ElementSize;
        Index++;
    }
    Assert((Address_Playhead - (uint8 *)Table->Address) < Table->Size);
    Arbitrary_Zero(Address_Playhead, Table->Block_ElementSize);

    return Index;
}

static void *
Memory_Block_AddressAtIndex(memory *Memory, memory_table_list TableName, uint32 Index, bool32 AssertExists)
{
    memory_table *Table = &Memory->Slot[TableName];
    Assert(Table->Block_ElementSize != 0);
    uint8 *Address = (uint8 *)Table->Address + (Table->Block_ElementSize * Index);
    if (AssertExists)
        Assert(*Address != 0);
    Assert((Address - (uint8 *)Table->Address) < Table->Size);
    return (void *)Address;
}

static uint16
Memory_Block_LazyIndexAtAddress(memory *Memory, memory_table_list TableName, void *Address)
{
    memory_table *Table = &Memory->Slot[TableName];
    return ((uint8 *)Address - (uint8 *)Table->Address) / Table->Block_ElementSize;
}


static void *
Memory_Block_AllocateAddress(memory *Memory, memory_table_list TableName)
{
    uint16 FileIndex = Memory_Block_AllocateNew(Memory, TableName);
    return Memory_Block_AddressAtIndex(Memory, TableName, FileIndex, 0);
}

// IMPORTANT(fox): All block data structs have to start with a uint8 Occupied variable!
static bool32
Block_Loop(memory *Memory, memory_table_list TableName, uint32 TotalCount, int *HasIncremented, int *CurrentCount, int *Index)
{
    for (;;) {
        if (*CurrentCount == TotalCount) {
            return 0;
        }
        if (*HasIncremented) {
            *HasIncremented = 0;
            (*Index)++;
        }
        uint8 *Occupied = (uint8 *)Memory_Block_AddressAtIndex(Memory, TableName, *Index, 0);
        if (*Occupied) {
            *HasIncremented = 1;
            (*CurrentCount)++;
            return 1;
        }
        (*Index)++;
        Assert(*CurrentCount <= TotalCount);
        Assert(*Index <= TotalCount*100); // This can get triggered normally if 100+ items are added and the first 99 in memory are deleted.
    }
    Assert(0);
    return 0;
}

static bool32
Block_Loop(memory *Memory, property_channel *Property, uint32 TotalCount, int *HasIncremented, int *CurrentCount, int *Index)
{
    for (;;) {
        if (*CurrentCount == TotalCount) {
            return 0;
        }
        if (*HasIncremented) {
            *HasIncremented = 0;
            (*Index)++;
        }
        uint8 *Occupied = (uint8 *)Bezier_LookupAddress(Memory, Property->Block_Bezier_Index, *Index, 0);
        if (*Occupied) {
            *HasIncremented = 1;
            (*CurrentCount)++;
            return 1;
        }
        (*Index)++;
        Assert(*CurrentCount <= TotalCount);
        Assert(*Index <= TotalCount*100); // This can get triggered normally if 100+ items are added and the first 99 in memory are deleted.
    }
    Assert(0);
    return 0;
}

static uint32
Memory_Block_PrincipalBitmap_AllocateNew(project_data *File, project_state *State, memory *Memory)
{
    int h = 0, c = 0, i = 0;
    int MaxBlockIndex = -1;
    int PrincipalCount = 0;
    while (Block_Loop(Memory, F_Sources, File->Source_Count, &h, &c, &i)) {
        block_source *Source = (block_source *)Memory_Block_AddressAtIndex(Memory, F_Sources, i);
        if (Source->Type == source_type_principal || Source->Type == source_type_principal_temp)
            PrincipalCount++;
        if (Source->Bitmap_Index > MaxBlockIndex)
            MaxBlockIndex = i;
    }
    if (!PrincipalCount)
        return 0;

    block_source Source = *(block_source *)Memory_Block_AddressAtIndex(Memory, F_Sources, MaxBlockIndex);
    uint32 LastBlock = Source.Bitmap_Index;
    uint32 BlockSize = ((Source.Width * Source.Height * Source.BytesPerPixel) / BitmapBlockSize) + 1;

    uint32 Blocks_Max = Memory->Slot[B_CachedBitmaps].Size / BitmapBlockSize;
    Assert(Blocks_Max > (LastBlock + BlockSize));

    return LastBlock + BlockSize;

}

static uint32
Memory_Block_Bitmap_AllocateNew(project_state *State, memory *Memory, cache_entry Entry, uint64 NewSize)
{
    uint32 LastVal = 0;
    uint32 LastBlock = 0;
    uint32 c = 0;
    cache_entry *EntryArray = State->Render.Entry;
    while (EntryArray[c].IsOccupied != 0) {
        if (EntryArray[c].Block_StartIndex > LastBlock) {
            LastBlock = EntryArray[c].Block_StartIndex;
            LastVal = c;
        }
        c++;
    }
    cache_entry LastEntry = EntryArray[LastVal];
    uint32 LastEntry_BlockCount = 0;
    switch (EntryArray[LastVal].Type) {
        case cache_entry_type_comp:
        {
            block_composition Comp = *(block_composition *)Memory_Block_AddressAtIndex(Memory, F_Precomps, LastEntry.TypeInfo);
            uint64 Size = Comp.Width * Comp.Height * Comp.BytesPerPixel;
            LastEntry_BlockCount = (Size / BitmapBlockSize) + 1;
        } break;
        case cache_entry_type_source:
        {
            block_source Source = *(block_source *)Memory_Block_AddressAtIndex(Memory, F_Sources, LastEntry.TypeInfo);
            uint64 Size = Source.Width * Source.Height * Source.BytesPerPixel;
            LastEntry_BlockCount = (Size / BitmapBlockSize) + 1;
        } break;
        case cache_entry_type_assert:
        {
            Assert(0);
        } break;
        default:
        {
            Assert(0);
        } break;
    }

    uint32 Blocks_Max = Memory->Slot[B_CachedBitmaps].Size / BitmapBlockSize;
    Assert(Blocks_Max > LastBlock);
    return LastBlock + LastEntry_BlockCount;

    /*
    uint32 Blocks_Needed = (NewSize / BitmapBlockSize) + 1;
    uint32 Block_Index_Available = 0;
    */

}
static void
Memory_Cache_Purge(project_data *File, project_state *State, memory *Memory, int32 SingleFrame)
{
    cache_entry *EntryArray = State->Render.Entry;
    int c = 0;
    int count = Memory->EntryCount;
    Assert(Memory->EntryCount < 10000);
    while (count != 0) {
        bool32 ExtraCheck = (SingleFrame == -1) ? 1 : EntryArray[c].TypeInfo_Sub == SingleFrame;
        if (EntryArray[c].Type == cache_entry_type_comp &&
            EntryArray[c].TypeInfo == File->PrincipalCompIndex && ExtraCheck) {
            EntryArray[c].IsCached = 0;
        }
        c++;
        count--;
    }
    State->LastCachedFrame = -10000;
    State->CachedFrameCount = 0;
}

static cache_entry *
Memory_Cache_Search(project_state *State, memory *Memory, cache_entry_type Type, uint32 TypeInfo, uint32 TypeInfo_Sub)
{
    cache_entry *EntryArray = State->Render.Entry;
    int c = 0;
    int count = Memory->EntryCount;
    while (count != 0) {
        if (EntryArray[c].Type == Type &&
            EntryArray[c].TypeInfo == TypeInfo &&
            EntryArray[c].TypeInfo_Sub == TypeInfo_Sub) {
            return &EntryArray[c];
        }
        c++;
        count--;
    }
    if (c != 0)
        EntryArray[c].Block_StartIndex = Memory_Block_Bitmap_AllocateNew(State, Memory, EntryArray[c], 0);
    EntryArray[c].IsOccupied = true;
    EntryArray[c].Type = Type;
    EntryArray[c].TypeInfo = TypeInfo;
    EntryArray[c].TypeInfo_Sub = TypeInfo_Sub;
    Memory->EntryCount++;
    return &EntryArray[c];
}

static void *
Memory_Block_Bitmap_AddressAtIndex(memory *Memory, uint32 Index)
{
    memory_table *Table = &Memory->Slot[B_CachedBitmaps];
    uint8 *Address = (uint8 *)Table->Address + Index*BitmapBlockSize;
    return (void *)Address;
}

static void *
Memory_PushScratch(memory *Memory, uint64 Size) {
    memory_table *Table = &Memory->Slot[B_ScratchSpace];
    uint8 *Address = ((uint8 *)Table->Address + Memory->ScratchPos);
    Memory->ScratchPos += Size;
#if DEBUG
    Assert(Memory->ScratchPos > 0);
    Assert(Memory->ScratchPos < Table->Size);
    Debug.ScratchSize[Debug.ScratchState] = Size;
    Debug.ScratchState++;
#endif
    return (void *)Address;
}

static void
Memory_PopScratch(memory *Memory, uint64 Size) {
    memory_table *Table = &Memory->Slot[B_ScratchSpace];
    Assert(Memory->ScratchPos >= Size);
    Memory->ScratchPos -= Size;
#if DEBUG
    Debug.ScratchState--;
    Assert(Debug.ScratchSize[Debug.ScratchState] == Size);
#endif
}

static void *
Memory_AddressAtOffset(memory *Memory, memory_table_list TableName, uint64 Offset)
{
    memory_table *Table = &Memory->Slot[TableName];
    Assert(Offset < Table->Size);
    return (void *)((uint8 *)Table->Address + Offset);
}

void Memory_Copy(uint8 *Address_Write, uint8 *Address_Read, uint64 Size)
{
    uint64 i = 0;
    while (i < Size) {
        *(Address_Write + i) = *(Address_Read + i);
        i++;
    }
}

void Memory_Fill(uint8 *Address_Write, uint8 *Address_Read, uint64 WriteSize, uint64 ReadSize)
{
    uint64 i = 0;
    while (i < WriteSize) {
        *(Address_Write + i) = *(Address_Read + (i % ReadSize));
        i++;
    }
}

void Arbitrary_Zero(uint8 *Address_Write, uint64 Size)
{
    uint64 i = 0;
#if ARM
#else
    __m256i Zero256 = _mm256_setzero_si256();
    if (Size > 64 && InstructionMode == instruction_mode_avx) {
        uint64 Size_Lane = Size - (Size % 64);
        while (i < Size_Lane) {
            _mm256_storeu_si256((__m256i *)(Address_Write + i), Zero256);
            _mm256_storeu_si256((__m256i *)(Address_Write + i + 32), Zero256);
            i += 64;
        }
    }
#endif
    while (i < Size) {
        *(Address_Write + i) = 0;
        i++;
    }
}

void Arbitrary_SwapData(memory *Memory, uint8 *Address_0, uint8 *Address_1, uint64 Size)
{
    uint8 *Buffer_Scratch = (uint8 *)Memory_PushScratch(Memory, Size);
    Memory_Copy(Buffer_Scratch, Address_0,      Size);
    Memory_Copy(Address_0,      Address_1,      Size);
    Memory_Copy(Address_1,      Buffer_Scratch, Size);
    Memory_PopScratch(Memory, Size);
}


static void
Arbitrary_ShiftData(uint8 *Address_Start, uint8 *Address_End, uint64 ShiftAmount, int32 Direction)
{
    if (Direction > 0) {
        uint8 *AddressPlayhead = Address_End;
        while ((ptrsize)AddressPlayhead >= (ptrsize)Address_Start) {
            *(AddressPlayhead + ShiftAmount) = *AddressPlayhead;
            AddressPlayhead--;
        }
    } else {
        uint8 *AddressPlayhead = Address_Start;
        while ((ptrsize)AddressPlayhead < (ptrsize)Address_End) {
            *(AddressPlayhead - ShiftAmount) = *AddressPlayhead;
            AddressPlayhead++;
        }
    }
}

// TODO(fox): Just use zlib...
extern "C" {
#include "miniz.c"
}

static uint64
Data_Compress(memory *Memory, void *DataSource, uint64 DataSize, void *DataBuffer, uint64 DataBufferSize, int CompressionLevel)
{
    z_stream stream = {};
    stream.next_in = (uint8 *)DataSource;
    stream.avail_in = DataSize;
    stream.next_out = (uint8 *)DataBuffer;
    stream.avail_out = DataBufferSize;

    if (deflateInit(&stream, CompressionLevel) != Z_OK)
    {
        Assert(0);
    }

    int status;
    for ( ; ; )
    {
      status = deflate(&stream, Z_SYNC_FLUSH);

      if (status == Z_STREAM_END || !stream.avail_in)
        break;
      else
          Assert(0);
    }

    uint64 CompressedSize = stream.total_out;

    if (deflateEnd(&stream) != Z_OK)
    {
        Assert(0);
    }

    return CompressedSize;
}

static void
Data_Decompress(memory *Memory, void *CompressedLocation, uint64 CompressedSize, void *BitmapLocation, uint64 ExpectedSize)
{
    z_stream stream = {};
    stream.next_in = (uint8 *)CompressedLocation;
    stream.avail_in = CompressedSize;
    stream.next_out = (uint8 *)BitmapLocation;
    stream.avail_out = 2147483648;

    if (inflateInit(&stream))
    {
        Assert(0);
    }

    int status;
    for ( ; ; )
    {
      status = inflate(&stream, Z_NO_FLUSH);

      if (status == Z_STREAM_END || !stream.avail_in)
        break;
      else
          Assert(0);
    }

    // Assert(stream.total_out == ExpectedSize);

    if (inflateEnd(&stream) != Z_OK)
    {
        Assert(0);
    }
}