unit FastcodeStrLowerCaseUnit;

//Version : 0.1 Preliminary version
//Only direct calling supported

interface

function LowerCaseFastcodeP3(const S: String): String;
function LowerCaseFastcodeP4(const S: String): String;
function LowerCaseFastcodePrescott(const S: String): String;
function LowerCaseFastcodeXP(const S: String): String;
function LowerCaseFastcodeOpteron(const S: String): String;
function LowerCaseFastcodeRTL(const S: String): String;
function LowerCaseFastcodePascal(const S: String): String;
function LowerCaseFastcodeBlended(const S: String): String;

implementation

uses
 SysUtils;

var
 LookUpTable : array of Char;

procedure InitializeLookUpTable;
var
 I : Byte;
 S1, S2 : AnsiString;

begin
 SetLength(LookUpTable, 256);
 for I := 0 to 255 do
  begin
   S1 := Char(I);
   S2 := LowerCase(S1);
   LookUpTable[I] := S2[1];
  end;
end;

//Author:            Dennis Kjaer Christensen
//Date:              9/1 2004
//Optimized for:     P4
//Instructionset(s): IA32
//Original name:     LowerCaseDKC_IA32_1

function LowerCaseFastcodeP4(const S: String): String;
asm
 push  ebx
 push  esi
 push  edi
 mov   ebx,eax
 test  eax,eax
 jz    @LStrLenExit
 mov   eax,[eax-$04]
@LStrLenExit :
 mov   esi,eax
 mov   edi,edx
 //SetLength(Result, Max);
 mov   eax,edx
 mov   edx,esi
 call  System.@LStrSetLength
 //if Max > 0 then
 test  esi,esi
 jbe   @Exit1
 //PResult := Pointer(Result);
 mov   edi,[edi]
 sub   ebx,1
 sub   edi,1
 mov   ecx,[LookUpTable]
@RepeatBegin :
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 //until(CharNo >= Max);
 jnbe  @RepeatBegin
@Exit1 :
 pop   edi
 pop   esi
 pop   ebx
end;

//Author:            Aleksandr Sharahov
//Date:              N/A
//Optimized for:     Prescott
//Instructionset(s): IA32
//Original name:     LowerCaseShaPas2_d

{$UNDEF SaveQ} {$IFOPT Q+} {$Q-} {$DEFINE SaveQ} {$ENDIF}
{$UNDEF SaveR} {$IFOPT R+} {$R-} {$DEFINE SaveR} {$ENDIF}
function LowerCaseFastcodePrescott(const s: string): string;
var
  ch1, ch2, ch3, dist, term: integer;
  p: pchar;
label
  loop, last;
begin
  if s='' then begin;
    Result:=''; exit;
    end;

  p:=pointer(s);

  //If need pure Pascal change the next line to term:=Length(s);
  term:=pinteger(@p[-4])^;

  SetLength(Result,term);

  if term<>0 then begin;
    dist:=integer(Result);
    term:=integer(p+term);
    dist:=dist-integer(p)-4;

loop:
    ch1:=pinteger(p)^;
    ch3:=$7F7F7F7F;

    ch2:=ch1;
    ch3:=ch3 and ch1;

    ch2:=ch2 xor (-1);
    ch3:=ch3 + $25252525;

    ch2:=ch2 and $80808080;
    ch3:=ch3 and $7F7F7F7F;

    ch3:=ch3 + $1A1A1A1A;
    inc(p,4);

    ch3:=ch3 and ch2;
    if cardinal(p)>=cardinal(term) then goto last;

    ch3:=ch3 shr 2;
    ch1:=ch1 + ch3;
    pinteger(p+dist)^:=ch1;
    goto loop;

last:
    ch3:=ch3 shr 2;
    term:=term-integer(p);
    p:=p+dist;
    ch1:=ch1 + ch3;
    if term<-1
      then pword(p)^:=word(ch1)
      else pinteger(p)^:=ch1;
    end;
  end;
{$IFDEF SaveQ} {$Q+} {$UNDEF SaveQ} {$ENDIF}
{$IFDEF SaveR} {$R+} {$UNDEF SaveR} {$ENDIF}

//Author:            Dennis Kjaer Christensen
//Date:              9/1 2004
//Optimized for:     P3
//Instructionset(s): IA32
//Original name:     LowerCaseDKC_IA32_1

function LowerCaseFastcodeP3(const S: String): String;
asm
 push  ebx
 push  esi
 push  edi
 mov   ebx,eax
 test  eax,eax
 jz    @LStrLenExit
 mov   eax,[eax-$04]
@LStrLenExit :
 mov   esi,eax
 mov   edi,edx
 //SetLength(Result, Max);
 mov   eax,edx
 mov   edx,esi
 call  System.@LStrSetLength
 //if Max > 0 then
 test  esi,esi
 jbe   @Exit1
 //PResult := Pointer(Result);
 mov   edi,[edi]
 sub   ebx,1
 sub   edi,1
 mov   ecx,[LookUpTable]
@RepeatBegin :
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 jbe   @Exit1
 movzx edx,[ebx+esi]
 movzx edx,[ecx+edx]
 mov   [edi+esi],dl
 sub   esi,1
 //until(CharNo >= Max);
 jnbe  @RepeatBegin
@Exit1 :
 pop   edi
 pop   esi
 pop   ebx
end;

//Author:            Aleksandr Sharahov
//Date:              N/A
//Optimized for:     XP
//Instructionset(s): IA32
//Original name:     LowerCaseShaPas2_d

{$UNDEF SaveQ} {$IFOPT Q+} {$Q-} {$DEFINE SaveQ} {$ENDIF}
{$UNDEF SaveR} {$IFOPT R+} {$R-} {$DEFINE SaveR} {$ENDIF}
function LowerCaseFastcodeXP(const s: string): string;
var
  ch1, ch2, ch3, dist, term: integer;
  p: pchar;
label
  loop, last;
begin
  if s='' then begin;
    Result:=''; exit;
    end;

  p:=pointer(s);

  //If need pure Pascal change the next line to term:=Length(s);
  term:=pinteger(@p[-4])^;

  SetLength(Result,term);

  if term<>0 then begin;
    dist:=integer(Result);
    term:=integer(p+term);
    dist:=dist-integer(p)-4;

loop:
    ch1:=pinteger(p)^;
    ch3:=$7F7F7F7F;

    ch2:=ch1;
    ch3:=ch3 and ch1;

    ch2:=ch2 xor (-1);
    ch3:=ch3 + $25252525;

    ch2:=ch2 and $80808080;
    ch3:=ch3 and $7F7F7F7F;

    ch3:=ch3 + $1A1A1A1A;
    inc(p,4);

    ch3:=ch3 and ch2;
    if cardinal(p)>=cardinal(term) then goto last;

    ch3:=ch3 shr 2;
    ch1:=ch1 + ch3;
    pinteger(p+dist)^:=ch1;
    goto loop;

last:
    ch3:=ch3 shr 2;
    term:=term-integer(p);
    p:=p+dist;
    ch1:=ch1 + ch3;
    if term<-1
      then pword(p)^:=word(ch1)
      else pinteger(p)^:=ch1;
    end;
  end;
{$IFDEF SaveQ} {$Q+} {$UNDEF SaveQ} {$ENDIF}
{$IFDEF SaveR} {$R+} {$UNDEF SaveR} {$ENDIF}

//Author:            Aleksandr Sharahov
//Date:              N/A
//Optimized for:     Opteron
//Instructionset(s): IA32
//Original name:     LowerCaseShaPas2_d

{$UNDEF SaveQ} {$IFOPT Q+} {$Q-} {$DEFINE SaveQ} {$ENDIF}
{$UNDEF SaveR} {$IFOPT R+} {$R-} {$DEFINE SaveR} {$ENDIF}
function LowerCaseFastcodeOpteron(const s: string): string;
var
  ch1, ch2, ch3, dist, term: integer;
  p: pchar;
label
  loop, last;
begin
  if s='' then begin;
    Result:=''; exit;
    end;

  p:=pointer(s);

  //If need pure Pascal change the next line to term:=Length(s);
  term:=pinteger(@p[-4])^;

  SetLength(Result,term);

  if term<>0 then begin;
    dist:=integer(Result);
    term:=integer(p+term);
    dist:=dist-integer(p)-4;

loop:
    ch1:=pinteger(p)^;
    ch3:=$7F7F7F7F;

    ch2:=ch1;
    ch3:=ch3 and ch1;

    ch2:=ch2 xor (-1);
    ch3:=ch3 + $25252525;

    ch2:=ch2 and $80808080;
    ch3:=ch3 and $7F7F7F7F;

    ch3:=ch3 + $1A1A1A1A;
    inc(p,4);

    ch3:=ch3 and ch2;
    if cardinal(p)>=cardinal(term) then goto last;

    ch3:=ch3 shr 2;
    ch1:=ch1 + ch3;
    pinteger(p+dist)^:=ch1;
    goto loop;

last:
    ch3:=ch3 shr 2;
    term:=term-integer(p);
    p:=p+dist;
    ch1:=ch1 + ch3;
    if term<-1
      then pword(p)^:=word(ch1)
      else pinteger(p)^:=ch1;
    end;
  end;
{$IFDEF SaveQ} {$Q+} {$UNDEF SaveQ} {$ENDIF}
{$IFDEF SaveR} {$R+} {$UNDEF SaveR} {$ENDIF}

//Author:            Aleksandr Sharahov
//Date:              N/A
//Optimized for:     RTL
//Instructionset(s): IA32
//Original name:     LowerCaseShaPas2_d

{$UNDEF SaveQ} {$IFOPT Q+} {$Q-} {$DEFINE SaveQ} {$ENDIF}
{$UNDEF SaveR} {$IFOPT R+} {$R-} {$DEFINE SaveR} {$ENDIF}
function LowerCaseFastcodeRTL(const s: string): string;
var
  ch1, ch2, ch3, dist, term: integer;
  p: pchar;
label
  loop, last;
begin
  if s='' then begin;
    Result:=''; exit;
    end;

  p:=pointer(s);

  //If need pure Pascal change the next line to term:=Length(s);
  term:=pinteger(@p[-4])^;

  SetLength(Result,term);

  if term<>0 then begin;
    dist:=integer(Result);
    term:=integer(p+term);
    dist:=dist-integer(p)-4;

loop:
    ch1:=pinteger(p)^;
    ch3:=$7F7F7F7F;

    ch2:=ch1;
    ch3:=ch3 and ch1;

    ch2:=ch2 xor (-1);
    ch3:=ch3 + $25252525;

    ch2:=ch2 and $80808080;
    ch3:=ch3 and $7F7F7F7F;

    ch3:=ch3 + $1A1A1A1A;
    inc(p,4);

    ch3:=ch3 and ch2;
    if cardinal(p)>=cardinal(term) then goto last;

    ch3:=ch3 shr 2;
    ch1:=ch1 + ch3;
    pinteger(p+dist)^:=ch1;
    goto loop;

last:
    ch3:=ch3 shr 2;
    term:=term-integer(p);
    p:=p+dist;
    ch1:=ch1 + ch3;
    if term<-1
      then pword(p)^:=word(ch1)
      else pinteger(p)^:=ch1;
    end;
  end;
{$IFDEF SaveQ} {$Q+} {$UNDEF SaveQ} {$ENDIF}
{$IFDEF SaveR} {$R+} {$UNDEF SaveR} {$ENDIF}

//Author:            Aleksandr Sharahov
//Date:              N/A
//Optimized for:     Pascal
//Instructionset(s): IA32
//Original name:     LowerCaseShaPas2_d

{$UNDEF SaveQ} {$IFOPT Q+} {$Q-} {$DEFINE SaveQ} {$ENDIF}
{$UNDEF SaveR} {$IFOPT R+} {$R-} {$DEFINE SaveR} {$ENDIF}
function LowerCaseFastcodePascal(const s: string): string;
var
  ch1, ch2, ch3, dist, term: integer;
  p: pchar;
label
  loop, last;
begin
  if s='' then begin;
    Result:=''; exit;
    end;

  p:=pointer(s);

  //If need pure Pascal change the next line to term:=Length(s);
  term:=pinteger(@p[-4])^;

  SetLength(Result,term);

  if term<>0 then begin;
    dist:=integer(Result);
    term:=integer(p+term);
    dist:=dist-integer(p)-4;

loop:
    ch1:=pinteger(p)^;
    ch3:=$7F7F7F7F;

    ch2:=ch1;
    ch3:=ch3 and ch1;

    ch2:=ch2 xor (-1);
    ch3:=ch3 + $25252525;

    ch2:=ch2 and $80808080;
    ch3:=ch3 and $7F7F7F7F;

    ch3:=ch3 + $1A1A1A1A;
    inc(p,4);

    ch3:=ch3 and ch2;
    if cardinal(p)>=cardinal(term) then goto last;

    ch3:=ch3 shr 2;
    ch1:=ch1 + ch3;
    pinteger(p+dist)^:=ch1;
    goto loop;

last:
    ch3:=ch3 shr 2;
    term:=term-integer(p);
    p:=p+dist;
    ch1:=ch1 + ch3;
    if term<-1
      then pword(p)^:=word(ch1)
      else pinteger(p)^:=ch1;
    end;
  end;
{$IFDEF SaveQ} {$Q+} {$UNDEF SaveQ} {$ENDIF}
{$IFDEF SaveR} {$R+} {$UNDEF SaveR} {$ENDIF}

//Author:            Aleksandr Sharahov
//Date:              N/A
//Optimized for:     Blended
//Instructionset(s): IA32
//Original name:     LowerCaseShaPas2_d

{$UNDEF SaveQ} {$IFOPT Q+} {$Q-} {$DEFINE SaveQ} {$ENDIF}
{$UNDEF SaveR} {$IFOPT R+} {$R-} {$DEFINE SaveR} {$ENDIF}
function LowerCaseFastcodeBlended(const s: string): string;
var
  ch1, ch2, ch3, dist, term: integer;
  p: pchar;
label
  loop, last;
begin
  if s='' then begin;
    Result:=''; exit;
    end;

  p:=pointer(s);

  //If need pure Pascal change the next line to term:=Length(s);
  term:=pinteger(@p[-4])^;

  SetLength(Result,term);

  if term<>0 then begin;
    dist:=integer(Result);
    term:=integer(p+term);
    dist:=dist-integer(p)-4;

loop:
    ch1:=pinteger(p)^;
    ch3:=$7F7F7F7F;

    ch2:=ch1;
    ch3:=ch3 and ch1;

    ch2:=ch2 xor (-1);
    ch3:=ch3 + $25252525;

    ch2:=ch2 and $80808080;
    ch3:=ch3 and $7F7F7F7F;

    ch3:=ch3 + $1A1A1A1A;
    inc(p,4);

    ch3:=ch3 and ch2;
    if cardinal(p)>=cardinal(term) then goto last;

    ch3:=ch3 shr 2;
    ch1:=ch1 + ch3;
    pinteger(p+dist)^:=ch1;
    goto loop;

last:
    ch3:=ch3 shr 2;
    term:=term-integer(p);
    p:=p+dist;
    ch1:=ch1 + ch3;
    if term<-1
      then pword(p)^:=word(ch1)
      else pinteger(p)^:=ch1;
    end;
  end;
{$IFDEF SaveQ} {$Q+} {$UNDEF SaveQ} {$ENDIF}
{$IFDEF SaveR} {$R+} {$UNDEF SaveR} {$ENDIF}

initialization

 InitializeLookUpTable;

end.