/* Ppmd.h -- PPMD codec common code
2023-03-05 : Igor Pavlov : Public domain
This code is based on PPMd var.H (2001): Dmitry Shkarin : Public domain */

#ifndef ZIP7_INC_PPMD_H
#define ZIP7_INC_PPMD_H

#include "CpuArch.h"

EXTERN_C_BEGIN

#if defined(MY_CPU_SIZEOF_POINTER) && (MY_CPU_SIZEOF_POINTER == 4)
/*
   PPMD code always uses 32-bit internal fields in PPMD structures to store internal references in main block.
   if (PPMD_32BIT is     defined), the PPMD code stores internal pointers to 32-bit reference fields.
   if (PPMD_32BIT is NOT defined), the PPMD code stores internal UInt32 offsets to reference fields.
   if (pointer size is 64-bit), then (PPMD_32BIT) mode is not allowed,
   if (pointer size is 32-bit), then (PPMD_32BIT) mode is optional,
     and it's allowed to disable PPMD_32BIT mode even if pointer is 32-bit.
   PPMD code works slightly faster in (PPMD_32BIT) mode.
*/
  #define PPMD_32BIT
#endif

#define PPMD_INT_BITS 7
#define PPMD_PERIOD_BITS 7
#define PPMD_BIN_SCALE (1 << (PPMD_INT_BITS + PPMD_PERIOD_BITS))

#define PPMD_GET_MEAN_SPEC(summ, shift, round) (((summ) + (1 << ((shift) - (round)))) >> (shift))
#define PPMD_GET_MEAN(summ) PPMD_GET_MEAN_SPEC((summ), PPMD_PERIOD_BITS, 2)
#define PPMD_UPDATE_PROB_0(prob) ((prob) + (1 << PPMD_INT_BITS) - PPMD_GET_MEAN(prob))
#define PPMD_UPDATE_PROB_1(prob) ((prob) - PPMD_GET_MEAN(prob))

#define PPMD_N1 4
#define PPMD_N2 4
#define PPMD_N3 4
#define PPMD_N4 ((128 + 3 - 1 * PPMD_N1 - 2 * PPMD_N2 - 3 * PPMD_N3) / 4)
#define PPMD_NUM_INDEXES (PPMD_N1 + PPMD_N2 + PPMD_N3 + PPMD_N4)

MY_CPU_pragma_pack_push_1
/* Most compilers works OK here even without #pragma pack(push, 1), but some GCC compilers need it. */

/* SEE-contexts for PPM-contexts with masked symbols */
typedef struct
{
  UInt16 Summ; /* Freq */
  Byte Shift;  /* Speed of Freq change; low Shift is for fast change */
  Byte Count;  /* Count to next change of Shift */
} CPpmd_See;

#define Ppmd_See_UPDATE(p) \
  { if ((p)->Shift < PPMD_PERIOD_BITS && --(p)->Count == 0) \
    { (p)->Summ = (UInt16)((p)->Summ << 1); \
      (p)->Count = (Byte)(3 << (p)->Shift++); }}


typedef struct
{
  Byte Symbol;
  Byte Freq;
  UInt16 Successor_0;
  UInt16 Successor_1;
} CPpmd_State;

typedef struct CPpmd_State2_
{
  Byte Symbol;
  Byte Freq;
} CPpmd_State2;

typedef struct CPpmd_State4_
{
  UInt16 Successor_0;
  UInt16 Successor_1;
} CPpmd_State4;

MY_CPU_pragma_pop

/*
   PPMD code can write full CPpmd_State structure data to CPpmd*_Context
      at (byte offset = 2) instead of some fields of original CPpmd*_Context structure.
   
   If we use pointers to different types, but that point to shared
   memory space, we can have aliasing problem (strict aliasing).
   
   XLC compiler in -O2 mode can change the order of memory write instructions
   in relation to read instructions, if we have use pointers to different types.
   
   To solve that aliasing problem we use combined CPpmd*_Context structure
   with unions that contain the fields from both structures:
   the original CPpmd*_Context and CPpmd_State.
   So we can access the fields from both structures via one pointer,
   and the compiler doesn't change the order of write instructions
   in relation to read instructions.

   If we don't use memory write instructions to shared memory in
   some local code, and we use only reading instructions (read only),
   then probably it's safe to use pointers to different types for reading.
*/
  


#ifdef PPMD_32BIT

  #define Ppmd_Ref_Type(type)   type *
  #define Ppmd_GetRef(p, ptr)   (ptr)
  #define Ppmd_GetPtr(p, ptr)   (ptr)
  #define Ppmd_GetPtr_Type(p, ptr, note_type) (ptr)

#else

  #define Ppmd_Ref_Type(type)   UInt32
  #define Ppmd_GetRef(p, ptr)   ((UInt32)((Byte *)(ptr) - (p)->Base))
  #define Ppmd_GetPtr(p, offs)  ((void *)((p)->Base + (offs)))
  #define Ppmd_GetPtr_Type(p, offs, type) ((type *)Ppmd_GetPtr(p, offs))

#endif // PPMD_32BIT


typedef Ppmd_Ref_Type(CPpmd_State) CPpmd_State_Ref;
typedef Ppmd_Ref_Type(void)        CPpmd_Void_Ref;
typedef Ppmd_Ref_Type(Byte)        CPpmd_Byte_Ref;


/*
#ifdef MY_CPU_LE_UNALIGN
// the unaligned 32-bit access latency can be too large, if the data is not in L1 cache.
#define Ppmd_GET_SUCCESSOR(p) ((CPpmd_Void_Ref)*(const UInt32 *)(const void *)&(p)->Successor_0)
#define Ppmd_SET_SUCCESSOR(p, v) *(UInt32 *)(void *)(void *)&(p)->Successor_0 = (UInt32)(v)

#else
*/

/*
   We can write 16-bit halves to 32-bit (Successor) field in any selected order.
   But the native order is more consistent way.
   So we use the native order, if LE/BE order can be detected here at compile time.
*/

#ifdef MY_CPU_BE

  #define Ppmd_GET_SUCCESSOR(p) \
    ( (CPpmd_Void_Ref) (((UInt32)(p)->Successor_0 << 16) | (p)->Successor_1) )

  #define Ppmd_SET_SUCCESSOR(p, v) { \
    (p)->Successor_0 = (UInt16)(((UInt32)(v) >> 16) /* & 0xFFFF */); \
    (p)->Successor_1 = (UInt16)((UInt32)(v) /* & 0xFFFF */); }

#else

  #define Ppmd_GET_SUCCESSOR(p) \
    ( (CPpmd_Void_Ref) ((p)->Successor_0 | ((UInt32)(p)->Successor_1 << 16)) )

  #define Ppmd_SET_SUCCESSOR(p, v) { \
    (p)->Successor_0 = (UInt16)((UInt32)(v) /* & 0xFFFF */); \
    (p)->Successor_1 = (UInt16)(((UInt32)(v) >> 16) /* & 0xFFFF */); }

#endif

// #endif


#define PPMD_SetAllBitsIn256Bytes(p) \
  { size_t z; for (z = 0; z < 256 / sizeof(p[0]); z += 8) { \
  p[z+7] = p[z+6] = p[z+5] = p[z+4] = p[z+3] = p[z+2] = p[z+1] = p[z+0] = ~(size_t)0; }}

EXTERN_C_END
 
#endif