unit FastcodeCharPos;

//Version : 0.2 Preliminary version
//Only direct calling supported

interface

function CharPosFastcodeP3(Ch : Char; const Str : AnsiString) : Integer;
function CharPosFastcodeP4(Ch : Char; const Str : AnsiString) : Integer;
function CharPosFastcodePrescott(Ch : Char; const Str : AnsiString) : Integer;
function CharPosFastcodeXP(Ch : Char; const Str : AnsiString) : Integer;
function CharPosFastcodeOpteron(lnpChr: Char; const lcpStr: AnsiString): Integer;
function CharPosFastcodePascal(Ch : Char; const Str : AnsiString) : Integer;
function CharPosFastcodeRTL(Ch : Char; const Str : AnsiString) : Integer;
function CharPosFastcodeBlended(Ch : Char; const Str : AnsiString) : Integer;

implementation

uses
  Windows, SysUtils, elCPUID, elCodeHook;

//Author:            John O'Harrow
//Date:              N/A
//Optimized for:     P3
//Instructionset(s): N/A
//Original Name:     CharPosJOH_SSE

function CharPosFastcodeP3(Ch : Char; const Str : AnsiString) : Integer;
asm
  TEST      EDX, EDX         {Str = NIL?}
  JZ        @@NotFound       {Yes - Jump}
  MOV       ECX, [EDX-4]     {ECX = Length(Str)}
  CMP       ECX, 8
  JG        @@NotSmall
  TEST      ECX, ECX
  JZ        @@NotFound       {Exit if Length = 0}
@@Small:
  CMP       AL, [EDX]
  JZ        @Found1
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+1]
  JZ        @Found2
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+2]
  JZ        @Found3
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+3]
  JZ        @Found4
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+4]
  JZ        @Found5
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+5]
  JZ        @Found6
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+6]
  JZ        @Found7
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+7]
  JZ        @Found8
@@NotFound:
  XOR       EAX, EAX
  RET
@Found1:
  MOV       EAX, 1
  RET
@Found2:
  MOV       EAX, 2
  RET
@Found3:
  MOV       EAX, 3
  RET
@Found4:
  MOV       EAX, 4
  RET
@Found5:
  MOV       EAX, 5
  RET
@Found6:
  MOV       EAX, 6
  RET
@Found7:
  MOV       EAX, 7
  RET
@Found8:
  MOV       EAX, 8
  RET
@@NotSmall:
  MOV       AH, AL
  ADD       EDX, ECX
  MOVD      MM0, EAX
  PSHUFW    MM0, MM0, 0
  PUSH      ECX
  NEG       ECX
@@First8:
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare Next 8 Bytes}
  PMOVMSKB  EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
  JGE       @@Last8
@@Align:
  LEA       EAX, [EDX+ECX]
  AND       EAX, 7
  SUB       ECX, EAX
@@Loop:                      {Loop Unrolled 2X}
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare Next 8 Bytes}
  PMOVMSKB  EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
{$IFNDEF NoUnroll}
  JGE       @@Last8
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare Next 8 Bytes}
  PMOVMSKB  EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
{$ENDIF}
  JL        @@loop
@@Last8:
  PCMPEQB   MM0, [EDX-8]
  POP       ECX              {Original Length}
  PMOVMSKB  EAX, MM0
  TEST      EAX, EAX
  JNZ       @@Matched2
  EMMS
  RET                        {Finished}
@@Matched:                   {Set Result from 1st Match in EcX}
  POP       EDX              {Original Length}
  ADD       ECX, EDX
@@Matched2:
  EMMS
  BSF       EDX, EAX
  LEA       EAX, [EDX+ECX-7]
end;

//Author:            John O'Harrow
//Date:              N/A
//Optimized for:     P4
//Instructionset(s): N/A
//Original Name:     CharPosJOH_SSE2

function CharPosFastcodeP4(Ch : Char; const Str : AnsiString) : Integer;
asm
  test      edx, edx
  jz        @@NullString
  mov       ecx, [edx-4]
  push      ebx
  mov       ebx, eax
  cmp       ecx, 16
  jl        @@Small
@@NotSmall:
  mov       ah, al           {Fill each Byte of XMM1 with AL}
  movd      xmm1, eax
  pshuflw   xmm1, xmm1, 0
  pshufd    xmm1, xmm1, 0
@@First16:
  movups    xmm0, [edx]      {Unaligned}
  pcmpeqb   xmm0, xmm1       {Compare First 16 Characters}
  pmovmskb  eax, xmm0
  test      eax, eax
  jnz       @@FoundStart     {Exit on any Match}
  cmp       ecx, 32
  jl        @@Medium         {If Length(Str) < 32, Check Remainder}
@@Align:
  sub       ecx, 16          {Align Block Reads}
  push      ecx
  mov       eax, edx
  neg       eax
  and       eax, 15
  add       edx, ecx
  neg       ecx
  add       ecx, eax
@@Loop:
  movaps    xmm0, [edx+ecx]  {Aligned}
  pcmpeqb   xmm0, xmm1       {Compare Next 16 Characters}
  pmovmskb  eax, xmm0
  test      eax, eax
  jnz       @@Found          {Exit on any Match}
  add       ecx, 16
  jle       @@Loop
  pop       eax              {Check Remaining Characters}
  add       edx, 16
  add       eax, ecx         {Count from Last Loop End Position}
  jmp       dword ptr [@@JumpTable2-ecx*4]
  nop
  nop
@@NullString:
  xor       eax, eax         {Result = 0}
  ret
  nop
@@FoundStart:
  bsf       eax, eax         {Get Set Bit}
  pop       ebx
  inc       eax              {Set Result}
  ret
  nop
  nop
@@Found:
  pop       edx
  bsf       eax, eax         {Get Set Bit}
  add       edx, ecx
  pop       ebx
  lea       eax, [eax+edx+1] {Set Result}
  ret
@@Medium:
  add       edx, ecx         {End of String}
  mov       eax, 16          {Count from 16}
  jmp       dword ptr [@@JumpTable1-64-ecx*4]
  nop
  nop
@@Small:
  add       edx, ecx         {End of String}
  xor       eax, eax         {Count from 0}
  jmp       dword ptr [@@JumpTable1-ecx*4]
  nop
@@JumpTable1:
  dd        @@NotFound, @@01, @@02, @@03, @@04, @@05, @@06, @@07
  dd        @@08, @@09, @@10, @@11, @@12, @@13, @@14, @@15, @@16
@@JumpTable2:
  dd        @@16, @@15, @@14, @@13, @@12, @@11, @@10, @@09, @@08
  dd        @@07, @@06, @@05, @@04, @@03, @@02, @@01, @@NotFound
@@16:
  add       eax, 1
  cmp       bl, [edx-16]
  je        @@Done
@@15:
  add       eax, 1
  cmp       bl, [edx-15]
  je        @@Done
@@14:
  add       eax, 1
  cmp       bl, [edx-14]
  je        @@Done
@@13:
  add       eax, 1
  cmp       bl, [edx-13]
  je        @@Done
@@12:
  add       eax, 1
  cmp       bl, [edx-12]
  je        @@Done
@@11:
  add       eax, 1
  cmp       bl, [edx-11]
  je        @@Done
@@10:
  add       eax, 1
  cmp       bl, [edx-10]
  je        @@Done
@@09:
  add       eax, 1
  cmp       bl, [edx-9]
  je        @@Done
@@08:
  add       eax, 1
  cmp       bl, [edx-8]
  je        @@Done
@@07:
  add       eax, 1
  cmp       bl, [edx-7]
  je        @@Done
@@06:
  add       eax, 1
  cmp       bl, [edx-6]
  je        @@Done
@@05:
  add       eax, 1
  cmp       bl, [edx-5]
  je        @@Done
@@04:
  add       eax, 1
  cmp       bl, [edx-4]
  je        @@Done
@@03:
  add       eax, 1
  cmp       bl, [edx-3]
  je        @@Done
@@02:
  add       eax, 1
  cmp       bl, [edx-2]
  je        @@Done
@@01:
  add       eax, 1
  cmp       bl, [edx-1]
  je        @@Done
@@NotFound:
  xor       eax, eax
  pop       ebx
  ret
@@Done:
  pop       ebx
end;

//Author:            John O'Harrow
//Date:              N/A
//Optimized for:     Prescott
//Instructionset(s): N/A
//Original Name:     CharPosJOH_SSE2

function CharPosFastcodePrescott(Ch : Char; const Str : AnsiString) : Integer;
asm
  test      edx, edx
  jz        @@NullString
  mov       ecx, [edx-4]
  push      ebx
  mov       ebx, eax
  cmp       ecx, 16
  jl        @@Small
@@NotSmall:
  mov       ah, al           {Fill each Byte of XMM1 with AL}
  movd      xmm1, eax
  pshuflw   xmm1, xmm1, 0
  pshufd    xmm1, xmm1, 0
@@First16:
  movups    xmm0, [edx]      {Unaligned}
  pcmpeqb   xmm0, xmm1       {Compare First 16 Characters}
  pmovmskb  eax, xmm0
  test      eax, eax
  jnz       @@FoundStart     {Exit on any Match}
  cmp       ecx, 32
  jl        @@Medium         {If Length(Str) < 32, Check Remainder}
@@Align:
  sub       ecx, 16          {Align Block Reads}
  push      ecx
  mov       eax, edx
  neg       eax
  and       eax, 15
  add       edx, ecx
  neg       ecx
  add       ecx, eax
@@Loop:
  movaps    xmm0, [edx+ecx]  {Aligned}
  pcmpeqb   xmm0, xmm1       {Compare Next 16 Characters}
  pmovmskb  eax, xmm0
  test      eax, eax
  jnz       @@Found          {Exit on any Match}
  add       ecx, 16
  jle       @@Loop
  pop       eax              {Check Remaining Characters}
  add       edx, 16
  add       eax, ecx         {Count from Last Loop End Position}
  jmp       dword ptr [@@JumpTable2-ecx*4]
  nop
  nop
@@NullString:
  xor       eax, eax         {Result = 0}
  ret
  nop
@@FoundStart:
  bsf       eax, eax         {Get Set Bit}
  pop       ebx
  inc       eax              {Set Result}
  ret
  nop
  nop
@@Found:
  pop       edx
  bsf       eax, eax         {Get Set Bit}
  add       edx, ecx
  pop       ebx
  lea       eax, [eax+edx+1] {Set Result}
  ret
@@Medium:
  add       edx, ecx         {End of String}
  mov       eax, 16          {Count from 16}
  jmp       dword ptr [@@JumpTable1-64-ecx*4]
  nop
  nop
@@Small:
  add       edx, ecx         {End of String}
  xor       eax, eax         {Count from 0}
  jmp       dword ptr [@@JumpTable1-ecx*4]
  nop
@@JumpTable1:
  dd        @@NotFound, @@01, @@02, @@03, @@04, @@05, @@06, @@07
  dd        @@08, @@09, @@10, @@11, @@12, @@13, @@14, @@15, @@16
@@JumpTable2:
  dd        @@16, @@15, @@14, @@13, @@12, @@11, @@10, @@09, @@08
  dd        @@07, @@06, @@05, @@04, @@03, @@02, @@01, @@NotFound
@@16:
  add       eax, 1
  cmp       bl, [edx-16]
  je        @@Done
@@15:
  add       eax, 1
  cmp       bl, [edx-15]
  je        @@Done
@@14:
  add       eax, 1
  cmp       bl, [edx-14]
  je        @@Done
@@13:
  add       eax, 1
  cmp       bl, [edx-13]
  je        @@Done
@@12:
  add       eax, 1
  cmp       bl, [edx-12]
  je        @@Done
@@11:
  add       eax, 1
  cmp       bl, [edx-11]
  je        @@Done
@@10:
  add       eax, 1
  cmp       bl, [edx-10]
  je        @@Done
@@09:
  add       eax, 1
  cmp       bl, [edx-9]
  je        @@Done
@@08:
  add       eax, 1
  cmp       bl, [edx-8]
  je        @@Done
@@07:
  add       eax, 1
  cmp       bl, [edx-7]
  je        @@Done
@@06:
  add       eax, 1
  cmp       bl, [edx-6]
  je        @@Done
@@05:
  add       eax, 1
  cmp       bl, [edx-5]
  je        @@Done
@@04:
  add       eax, 1
  cmp       bl, [edx-4]
  je        @@Done
@@03:
  add       eax, 1
  cmp       bl, [edx-3]
  je        @@Done
@@02:
  add       eax, 1
  cmp       bl, [edx-2]
  je        @@Done
@@01:
  add       eax, 1
  cmp       bl, [edx-1]
  je        @@Done
@@NotFound:
  xor       eax, eax
  pop       ebx
  ret
@@Done:
  pop       ebx
end;

//Author:            John O'Harrow
//Date:              N/A
//Optimized for:     XP
//Instructionset(s): N/A
//Original Name:     CharPosJOH_SSE

function CharPosFastcodeXP(Ch : Char; const Str : AnsiString) : Integer;
asm
  TEST      EDX, EDX         {Str = NIL?}
  JZ        @@NotFound       {Yes - Jump}
  MOV       ECX, [EDX-4]     {ECX = Length(Str)}
  CMP       ECX, 8
  JG        @@NotSmall
  TEST      ECX, ECX
  JZ        @@NotFound       {Exit if Length = 0}
@@Small:
  CMP       AL, [EDX]
  JZ        @Found1
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+1]
  JZ        @Found2
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+2]
  JZ        @Found3
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+3]
  JZ        @Found4
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+4]
  JZ        @Found5
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+5]
  JZ        @Found6
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+6]
  JZ        @Found7
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+7]
  JZ        @Found8
@@NotFound:
  XOR       EAX, EAX
  RET
@Found1:
  MOV       EAX, 1
  RET
@Found2:
  MOV       EAX, 2
  RET
@Found3:
  MOV       EAX, 3
  RET
@Found4:
  MOV       EAX, 4
  RET
@Found5:
  MOV       EAX, 5
  RET
@Found6:
  MOV       EAX, 6
  RET
@Found7:
  MOV       EAX, 7
  RET
@Found8:
  MOV       EAX, 8
  RET
@@NotSmall:
  MOV       AH, AL
  ADD       EDX, ECX
  MOVD      MM0, EAX
  PSHUFW    MM0, MM0, 0
  PUSH      ECX
  NEG       ECX
@@First8:
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare Next 8 Bytes}
  PMOVMSKB  EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
  JGE       @@Last8
@@Align:
  LEA       EAX, [EDX+ECX]
  AND       EAX, 7
  SUB       ECX, EAX
@@Loop:                      {Loop Unrolled 2X}
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare Next 8 Bytes}
  PMOVMSKB  EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
{$IFNDEF NoUnroll}
  JGE       @@Last8
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare Next 8 Bytes}
  PMOVMSKB  EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
{$ENDIF}
  JL        @@loop
@@Last8:
  PCMPEQB   MM0, [EDX-8]
  POP       ECX              {Original Length}
  PMOVMSKB  EAX, MM0
  TEST      EAX, EAX
  JNZ       @@Matched2
  EMMS
  RET                        {Finished}
@@Matched:                   {Set Result from 1st Match in EcX}
  POP       EDX              {Original Length}
  ADD       ECX, EDX
@@Matched2:
  EMMS
  BSF       EDX, EAX
  LEA       EAX, [EDX+ECX-7]
end;

//Author:            Les Pawelzcyk
//Date:              N/A
//Optimized for:     Opteron
//Instructionset(s): N/A
//Original Name:     CharPosLPSSE2_3

function CharPosFastcodeOpteron(lnpChr: Char; const lcpStr: AnsiString): Integer;
register;
asm
 test       edx, edx
 jz         @strempty
 push       ebx
 mov        ecx, [edx-4]
 mov        ebx, eax
 cmp        ecx, 16
 jl         @smlcnt
 mov        ah, al
 movd       xmm1, eax
 pshuflw    xmm1, xmm1, 0
 pshufd     xmm1, xmm1, 0
 movups     xmm0, [edx]
 pcmpeqb    xmm0, xmm1
 pmovmskb   eax, xmm0
 test       eax, eax
 jnz        @found2
 cmp        ecx, 32
 jl         @medcnt
 mov        eax, edx
 not        eax
 sub        ecx, 16
 and        eax, 15
 mov        [esp-4], ecx
 add        eax, 1
 add        edx, ecx
 sub        ecx, eax
 neg        ecx
@loop:
 movaps     xmm0, [edx+ecx]
 pcmpeqb    xmm0, xmm1
 pmovmskb   eax, xmm0
 test       eax, eax
 jnz        @found
 add        ecx, 16
 js         @loop
 mov        eax, ecx
 lea        ecx, [ecx*8+offset @zerocnt-128]
 add        eax, [esp-4]
 add        edx, 16
 jmp        ecx
 ud2
@strempty:
 mov        eax, edx
 ret
@found:
 bsf        eax, eax
 mov        edx, [esp-4]
 lea        edx, [edx+ecx+1]
 pop        ebx
 add        eax, edx
 ret
@found2:
 bsf        eax, eax
 pop        ebx
 add        eax, 1
 ret
@medcnt:
 add        edx, ecx
 neg        ecx
 lea        ecx, [ecx*8+offset @zerocnt+128]
 mov        eax, 16
 jmp        ecx
 ud2
@smlcnt:
 add        edx, ecx
 neg        ecx
 lea        ecx, [ecx*8+offset @zerocnt]
 xor        eax, eax
 jmp        ecx
 ud2
 add        eax, 1
 cmp        bl, [edx-16]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-15]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-14]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-13]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-12]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-11]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-10]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-9]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-8]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-7]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-6]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-5]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-4]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-3]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-2]
 je         @smlfound
 add        eax, 1
 cmp        bl, [edx-1]
 je         @smlfound
@zerocnt:
 xor        eax, eax
@smlfound:
 pop        ebx
end;

//Author:            Chris Grant
//Date:              N/A
//Optimized for:     Pascal
//Instructionset(s): N/A
//Original Name:     CharPosCJGPas3_b

function CharPosFastcodePascal(Ch : Char; const Str : AnsiString) : Integer;
var
   Len, L2 : Integer;
   P1, P2, P3 : PChar;
label Finish, Small;
label label1, label2, label3, label4, label5, label6, label7, label8,
label9, label10, label11, label12, label13, label14, label15, label16;
label Found0, Found1, Found2, Found3, Found4, Found5, Found6, Found7,
Found8, Found9, Found10, Found11, Found12, Found13, Found14, Found15,
Found16;
label F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14,
F15, F16;
begin
  If Str='' then
    begin
      Result := 0;
      exit;
    end
  else
     begin
       Len := PInteger(Integer(Str)-4)^;
       P1 := Pointer(Str);   {Point to start of string}
       P2 := P1 + Len;       {Point to end of string}
       P3 := P2 - 16;        {Point to one sweep before end of string
                             this is only used as a reference}
       if (P1 < P3) then
         begin
           while 1 = 1 do
             begin
               if Ch = P1^    then goto Found0;
               if Ch = P1[1]  then goto Found1;
               if Ch = P1[2]  then goto Found2;
               if Ch = P1[3]  then goto Found3;
               if Ch = P1[4]  then goto Found4;
               if Ch = P1[5]  then goto Found5;
               if Ch = P1[6]  then goto Found6;
               if Ch = P1[7]  then goto Found7;
               if Ch = P1[8]  then goto Found8;
               if Ch = P1[9]  then goto Found9;
               if Ch = P1[10] then goto Found10;
               if Ch = P1[11] then goto Found11;
               if Ch = P1[12] then goto Found12;
               if Ch = P1[13] then goto Found13;
               if Ch = P1[14] then goto Found14;
               if Ch = P1[15] then goto Found15;

               Inc(P1, 16);
               if P1 > P3 then goto Small;
             end;
         end
         else goto Small;

       Found15: Result := 16 + Len + P1 - P2; exit;
       Found14: Result := 15 + Len + P1 - P2; exit;
       Found13: Result := 14 + Len + P1 - P2; exit;
       Found12: Result := 13 + Len + P1 - P2; exit;
       Found11: Result := 12 + Len + P1 - P2; exit;
       Found10: Result := 11 + Len + P1 - P2; exit;
       Found9:  Result := 10 + Len + P1 - P2; exit;
       Found8:  Result := 9  + Len + P1 - P2; exit;
       Found7:  Result := 8  + Len + P1 - P2; exit;
       Found6:  Result := 7  + Len + P1 - P2; exit;
       Found5:  Result := 6  + Len + P1 - P2; exit;
       Found4:  Result := 5  + Len + P1 - P2; exit;
       Found3:  Result := 4  + Len + P1 - P2; exit;
       Found2:  Result := 3  + Len + P1 - P2; exit;
       Found1:  Result := 2  + Len + P1 - P2; exit;
       Found0:  Result := 1  + Len + P1 - P2; exit;

       Small:
         L2 := P2 - P1;

         case L2 of
           0:  goto Finish;
           1:  goto label1;
           2:  goto label2;
           3:  goto label3;
           4:  goto label4;
           5:  goto label5;
           6:  goto label6;
           7:  goto label7;
           8:  goto label8;
           9:  goto label9;
           10: goto label10;
           11: goto label11;
           12: goto label12;
           13: goto label13;
           14: goto label14;
           15: goto label15;
         end;

       label16: if Ch = P1^       then goto F15;
       label15: if Ch = P1[L2-15] then goto F14;
       label14: if Ch = P1[L2-14] then goto F13;
       label13: if Ch = P1[L2-13] then goto F12;
       label12: if Ch = P1[L2-12] then goto F11;
       label11: if Ch = P1[L2-11] then goto F10;
       label10: if Ch = P1[L2-10] then goto F9;
       label9:  if Ch = P1[L2-9]  then goto F8;
       label8:  if Ch = P1[L2-8]  then goto F7;
       label7:  if Ch = P1[L2-7]  then goto F6;
       label6:  if Ch = P1[L2-6]  then goto F5;
       label5:  if Ch = P1[L2-5]  then goto F4;
       label4:  if Ch = P1[L2-4]  then goto F3;
       label3:  if Ch = P1[L2-3]  then goto F2;
       label2:  if Ch = P1[L2-2]  then goto F1;
       label1:  if Ch = P1[L2-1]  then goto F0;

       Finish:  Result := 0;        exit;

       F0:      Result := Len;      exit;
       F1:      Result := Len - 1;  exit;
       F2:      Result := Len - 2;  exit;
       F3:      Result := Len - 3;  exit;
       F4:      Result := Len - 4;  exit;
       F5:      Result := Len - 5;  exit;
       F6:      Result := Len - 6;  exit;
       F7:      Result := Len - 7;  exit;
       F8:      Result := Len - 8;  exit;
       F9:      Result := Len - 9;  exit;
       F10:     Result := Len - 10; exit;
       F11:     Result := Len - 11; exit;
       F12:     Result := Len - 12; exit;
       F13:     Result := Len - 13; exit;
       F14:     Result := Len - 14; exit;
       F15:     Result := Len - 15; exit;
     end;
end;

//Author:            Chris Grant
//Date:              N/A
//Optimized for:     RTL
//Instructionset(s): N/A
//Original Name:     CharPosCJGPas3_b

function CharPosFastcodeRTL(Ch : Char; const Str : AnsiString) : Integer;
var
   Len, L2 : Integer;
   P1, P2, P3 : PChar;
label Finish, Small;
label label1, label2, label3, label4, label5, label6, label7, label8,
label9, label10, label11, label12, label13, label14, label15, label16;
label Found0, Found1, Found2, Found3, Found4, Found5, Found6, Found7,
Found8, Found9, Found10, Found11, Found12, Found13, Found14, Found15,
Found16;
label F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14,
F15, F16;
begin
  If Str='' then
    begin
      Result := 0;
      exit;
    end
  else
     begin
       Len := PInteger(Integer(Str)-4)^;
       P1 := Pointer(Str);   {Point to start of string}
       P2 := P1 + Len;       {Point to end of string}
       P3 := P2 - 16;        {Point to one sweep before end of string
                             this is only used as a reference}
       if (P1 < P3) then
         begin
           while 1 = 1 do
             begin
               if Ch = P1^    then goto Found0;
               if Ch = P1[1]  then goto Found1;
               if Ch = P1[2]  then goto Found2;
               if Ch = P1[3]  then goto Found3;
               if Ch = P1[4]  then goto Found4;
               if Ch = P1[5]  then goto Found5;
               if Ch = P1[6]  then goto Found6;
               if Ch = P1[7]  then goto Found7;
               if Ch = P1[8]  then goto Found8;
               if Ch = P1[9]  then goto Found9;
               if Ch = P1[10] then goto Found10;
               if Ch = P1[11] then goto Found11;
               if Ch = P1[12] then goto Found12;
               if Ch = P1[13] then goto Found13;
               if Ch = P1[14] then goto Found14;
               if Ch = P1[15] then goto Found15;

               Inc(P1, 16);
               if P1 > P3 then goto Small;
             end;
         end
         else goto Small;

       Found15: Result := 16 + Len + P1 - P2; exit;
       Found14: Result := 15 + Len + P1 - P2; exit;
       Found13: Result := 14 + Len + P1 - P2; exit;
       Found12: Result := 13 + Len + P1 - P2; exit;
       Found11: Result := 12 + Len + P1 - P2; exit;
       Found10: Result := 11 + Len + P1 - P2; exit;
       Found9:  Result := 10 + Len + P1 - P2; exit;
       Found8:  Result := 9  + Len + P1 - P2; exit;
       Found7:  Result := 8  + Len + P1 - P2; exit;
       Found6:  Result := 7  + Len + P1 - P2; exit;
       Found5:  Result := 6  + Len + P1 - P2; exit;
       Found4:  Result := 5  + Len + P1 - P2; exit;
       Found3:  Result := 4  + Len + P1 - P2; exit;
       Found2:  Result := 3  + Len + P1 - P2; exit;
       Found1:  Result := 2  + Len + P1 - P2; exit;
       Found0:  Result := 1  + Len + P1 - P2; exit;

       Small:
         L2 := P2 - P1;

         case L2 of
           0:  goto Finish;
           1:  goto label1;
           2:  goto label2;
           3:  goto label3;
           4:  goto label4;
           5:  goto label5;
           6:  goto label6;
           7:  goto label7;
           8:  goto label8;
           9:  goto label9;
           10: goto label10;
           11: goto label11;
           12: goto label12;
           13: goto label13;
           14: goto label14;
           15: goto label15;
         end;

       label16: if Ch = P1^       then goto F15;
       label15: if Ch = P1[L2-15] then goto F14;
       label14: if Ch = P1[L2-14] then goto F13;
       label13: if Ch = P1[L2-13] then goto F12;
       label12: if Ch = P1[L2-12] then goto F11;
       label11: if Ch = P1[L2-11] then goto F10;
       label10: if Ch = P1[L2-10] then goto F9;
       label9:  if Ch = P1[L2-9]  then goto F8;
       label8:  if Ch = P1[L2-8]  then goto F7;
       label7:  if Ch = P1[L2-7]  then goto F6;
       label6:  if Ch = P1[L2-6]  then goto F5;
       label5:  if Ch = P1[L2-5]  then goto F4;
       label4:  if Ch = P1[L2-4]  then goto F3;
       label3:  if Ch = P1[L2-3]  then goto F2;
       label2:  if Ch = P1[L2-2]  then goto F1;
       label1:  if Ch = P1[L2-1]  then goto F0;

       Finish:  Result := 0;        exit;

       F0:      Result := Len;      exit;
       F1:      Result := Len - 1;  exit;
       F2:      Result := Len - 2;  exit;
       F3:      Result := Len - 3;  exit;
       F4:      Result := Len - 4;  exit;
       F5:      Result := Len - 5;  exit;
       F6:      Result := Len - 6;  exit;
       F7:      Result := Len - 7;  exit;
       F8:      Result := Len - 8;  exit;
       F9:      Result := Len - 9;  exit;
       F10:     Result := Len - 10; exit;
       F11:     Result := Len - 11; exit;
       F12:     Result := Len - 12; exit;
       F13:     Result := Len - 13; exit;
       F14:     Result := Len - 14; exit;
       F15:     Result := Len - 15; exit;
     end;
end;

//Author:            John O'Harrow
//Date:              N/A
//Optimized for:     Blended
//Instructionset(s): N/A
//Original Name:     CharPosJOH_MMX

function CharPosFastcodeBlended(Ch : Char; const Str : AnsiString) : Integer;
asm
  TEST      EDX, EDX         {Str = NIL?}
  JZ        @@NotFound       {Yes - Jump}
  MOV       ECX, [EDX-4]     {ECX = Length(Str)}
  CMP       ECX, 8
  JG        @@NotSmall
  TEST      ECX, ECX
  JZ        @@NotFound       {Exit if Length = 0}
@@Small:
  CMP       AL, [EDX]
  JZ        @Found1
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+1]
  JZ        @Found2
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+2]
  JZ        @Found3
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+3]
  JZ        @Found4
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+4]
  JZ        @Found5
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+5]
  JZ        @Found6
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+6]
  JZ        @Found7
  DEC       ECX
  JZ        @@NotFound
  CMP       AL, [EDX+7]
  JZ        @Found8
@@NotFound:
  XOR       EAX, EAX
  RET
@Found1:
  MOV       EAX, 1
  RET
@Found2:
  MOV       EAX, 2
  RET
@Found3:
  MOV       EAX, 3
  RET
@Found4:
  MOV       EAX, 4
  RET
@Found5:
  MOV       EAX, 5
  RET
@Found6:
  MOV       EAX, 6
  RET
@Found7:
  MOV       EAX, 7
  RET
@Found8:
  MOV       EAX, 8
  RET

@@NotSmall:                  {Length(Str) > 8}
  MOV       AH, AL
  ADD       EDX, ECX
  MOVD      MM0, EAX
  PUNPCKLWD MM0, MM0
  PUNPCKLDQ MM0, MM0
  PUSH      ECX              {Save Length}
  NEG       ECX
@@First8:
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare All 8 Bytes}
  PACKSSWB  MM1, MM1         {Pack Result into 4 Bytes}
  MOVD      EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
  JGE       @@Last8
@@Align:                     {Align to Previous 8 Byte Boundary}
  LEA       EAX, [EDX+ECX]
  AND       EAX, 7           {EAX -> 0 or 4}
  SUB       ECX, EAX
@@Loop:
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare All 8 Bytes}
  PACKSSWB  MM1, MM1         {Pack Result into 4 Bytes}
  MOVD      EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
{$IFNDEF NoUnroll}
  JGE       @@Last8
  MOVQ      MM1, [EDX+ECX]
  ADD       ECX, 8
  PCMPEQB   MM1, MM0         {Compare All 8 Bytes}
  PACKSSWB  MM1, MM1         {Pack Result into 4 Bytes}
  MOVD      EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched        {Exit on Match at any Position}
  CMP       ECX, -8          {Check if Next Loop would pass String End}
{$ENDIF}
  JL        @@Loop
@@Last8:
  MOVQ      MM1, [EDX-8]     {Position for Last 8 Used Characters}
  POP       EDX              {Original Length}
  PCMPEQB   MM1, MM0         {Compare All 8 Bytes}
  PACKSSWB  MM1, MM1         {Pack Result into 4 Bytes}
  MOVD      EAX, MM1
  TEST      EAX, EAX
  JNZ       @@Matched2       {Exit on Match at any Position}
  EMMS
  RET                        {Finished - Not Found}
@@Matched:                   {Set Result from 1st Match in EDX}
  POP       EDX              {Original Length}
  ADD       EDX, ECX
@@Matched2:
  EMMS
  SUB       EDX, 8           {Adjust for Extra ADD ECX,8 in Loop}
  TEST      AL, AL
  JNZ       @@MatchDone      {Match at Position 1 or 2}
  TEST      AH, AH
  JNZ       @@Match1         {Match at Position 3 or 4}
  SHR       EAX, 16
  TEST      AL, AL
  JNZ       @@Match2         {Match at Position 5 or 6}
  SHR       EAX, 8
  ADD       EDX, 6
  JMP       @@MatchDone
@@Match2:
  ADD       EDX, 4
  JMP       @@MatchDone
@@Match1:
  SHR       EAX, 8           {AL <- AH}
  ADD       EDX, 2
@@MatchDone:
  XOR       EAX, 2
  AND       EAX, 3           {EAX <- 1 or 2}
  ADD       EAX, EDX
end;

////////////////////////////////////////////////////////////////////////////////

type
  TFunc = function (Ch : Char; const Str : AnsiString) : Integer;

const
  NewFuncs: array[TCPUType] of TFunc = (
    CharPosFastcodeP3,
    CharPosFastcodeP4,
    CharPosFastcodeXP,
    CharPosFastcodeOpteron,
    CharPosFastcodeBlended,//CharPosFastcodePrescott,
    CharPosFastcodeBlended
  );

function CharPosStub(Ch : Char; const Str : AnsiString) : Integer; 
begin
  Result := Pos(Ch, Str);
end;

var
  CodeHook: TCodeHook;
  OldFunc_: TFunc; // OldFunc_ & NewFunc_ have same type against overloads ambiguties
  OldFunc: Pointer absolute OldFunc_;
  OldFuncSave: Pointer;
  OldProtect: DWORD;
  NewFunc_: TFunc;
  NewFunc: Pointer absolute NewFunc_;

initialization
begin
///  CharPosStub(#0, '');

  OldFunc_ := CharPosStub;
  OldFuncSave := OldFunc;
  if not VirtualProtect(OldFunc, 3 + SizeOf(TJump), PAGE_READWRITE, OldProtect) then
  begin
///    RaiseLastOSError;
  end;
  Inc(Integer(OldFunc), 3);
  Assert(PBYTE(OldFunc)^=$E8);// asserts we have a call to _LStrCharPos
  Inc(Integer(OldFunc));
  OldFunc := Pointer(Integer(OldFunc) + SizeOf(Pointer) + PInteger(OldFunc)^); // calculate true address of _LStrCharPos
  if not VirtualProtect(OldFuncSave, 3 + SizeOf(TJump), OldProtect, nil) then
  begin
///    RaiseLastOSError;
  end;
  if not FlushInstructionCache(GetCurrentProcess, OldFuncSave, 3 + SizeOf(TJump)) then
  begin
    RaiseLastOSError;
  end;

  NewFunc_ := NewFuncs[GetCPUType];
  CodeHook := TCodeHook.Create(OldFunc, NewFunc);
///  CharPos('', '');
end;

end.