xmd

# from xmddecode.dll (sh_sounds_explorer_1.7)

# some parts (*) where reordered to simplify logic
sub_10006F50
 arg_0           = dword ptr  4
 arg_4           = dword ptr  8
 arg_8           = dword ptr  0Ch
 arg_C           = dword ptr  10h
 arg_10          = dword ptr  14h
                                    # esp: pointer table?
* push    ebp                       # (save external value)
* push    edi                       # (save external value)
* push    ebx                       # (save external value)
* push    esi                       # (save external value)
* xor     ebp, ebp                  # (scale = 0)

  mov     eax, [esp+arg_C]          # load p_hist1
  mov     edx, [esp+arg_10]         # load p_hist2
* movsx   ecx, word ptr [eax]       # hist1 = (int16)*p_hist1
* movsx   eax, word ptr [edx]       # hist2 = (int16)*p_hist2

                                    # (frame hist2/1 is already consumed)
  mov     edi, [esp+8+arg_4]        # load p_frame
  mov     bp, [edi]                 # scale = (uint16)*p_frame
* shl     ebp, 0Eh                  # scale <<= 14

  mov     edx, [esp+8+arg_8]        # load bytes_left
  test    edx, edx                  # check if bytes_left == 0
  jle     loc_10007007              # end if true

* add     edi, 2                    # p_frame += 2 //skip scale
  mov     esi, [esp+10h+arg_0]      # load p_buffer
  mov     [esp+10h+arg_C], edx      # save bytes_left


 loc_10006F85:                      # (loop location)

  ****** nibble1 ******
  mov     dl, [edi]                 # byte = (uint8)*p_frame
  and     edx, 0Fh                  # byte &= 0xF //get lower nibble
  test    dl, 8                     # check sign
  jz      short loc_10006F92        # JUMP: not signed
  or      edx, 0FFFFFFF0h           # NOJM: signed, sign extend
  loc_10006F92:
  mov     ebx, ebp                  # (copy scale)
  add     esi, 2                    # p_buffer += 2
  imul    ebx, edx                  # sample *= scale
  mov     edx, ecx                  # (copy hist1)
  imul    edx, 7298h                # hist1 *= 0x7298 //coef1
  add     ebx, edx                  # sample += hist1

  lea     edx, [eax+eax*4]          # hist2b = hist2  + hist2*4
  lea     edx, [eax+edx*8]          # hist2b = hist2  + hist2b*8
  lea     edx, [edx+edx*4]          # hist2b = hist2b + hist2b*4
  lea     eax, [eax+edx*4]          # hist2  = hist2  + hist2b*4
  shl     eax, 4                    # hist2 <<= 4
                                    # AKA: (hist2 + (((hist2*5 + hist2)*8)*5)*4) << 4
                                    # AKA: (hist2 + ((5*8+1)*5*4+1)) << 4 = hist2 * 0x335*10
                                    # hist2 *= 0x3350 //coef2

  sub     ebx, eax                  # sample -= hist2
  sar     ebx, 0Eh                  # sample >>= 14
  mov     eax, ebx                  # (copy sample)
  mov     [esi-2], ax               # *p_buffer-2 = (int16)sample

* xor     edx, edx                  # byte = 0 //clean upper bits

  ****** nibble2 ******
  mov     dl, [edi]                 # byte = *p_frame
  shr     edx, 4                    # byte >> 4 //get upper nibble
  test    dl, 8                     # (check sign)
  jz      short loc_10006FCD        # JUMP: not signed
  or      edx, 0FFFFFFF0h           # NOJM: signed, sign extend
  loc_10006FCD:
  mov     ebx, ebp                  # (copy scale)
  add     esi, 2                    # p_buffer += 2
  imul    ebx, edx                  # sample *= scale
  mov     edx, eax                  # (copy hist1 = sample)
  imul    edx, 7298h                # hist1 *= 0x7298
  add     ebx, edx                  # sample += hist1

  lea     edx, [ecx+ecx*4]          # hist2b = hist2  + hist2*4
  lea     edx, [ecx+edx*8]          # hist2b = hist2  + hist2b*8
  lea     edx, [edx+edx*4]          # hist2b = hist2b + hist2b*4
  lea     ecx, [ecx+edx*4]          # hist2  = hist2  + hist2b*4
                                    # AKA: hist2 + (((hist2*5 + hist2)*8)*5)*4
                                    # AKA: hist2 * ((5*8+1)*5*4+1) = hist2 * 0x335
  shl     ecx, 4                    # hist2 <<= 4
                                    # AKA: hist2 * 0x3350

  sub     ebx, ecx                  # sample -= hist2
  sar     ebx, 0Eh                  # sample >>= 14
  mov     ecx, ebx                  # (copy sample)
  mov     [esi-2], cx               # *p_buffer-2 = (int16)sample

* inc     edi                       # p_frame += 1
* mov     edx, [esp+10h+arg_C]      # load bytes_left
  dec     edx                       # bytes_left--
  mov     [esp+10h+arg_C], edx      # save bytes_left
  jnz     short loc_10006F85        # JUMP (process next byte)

***
 loc_10007007:
* pop     esi                       # (restore external value)
* pop     ebx                       # (restore external value)
  pop     edi                       # (restore external value)
  pop     ebp                       # (restore external value)
  retn
  endp