freepascal.compiler/rtl/x86_64/x86_64.inc

{
    This file is part of the Free Pascal run time library.
    Copyright (c) 2002 by Florian Klaempfl and Sergei Gorelkin
    Members of the Free Pascal development team

    Processor dependent implementation for the system unit for
    the x86-64 architecture

    See the file COPYING.FPC, included in this distribution,
    for details about the copyright.

    This program is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

 **********************************************************************}

{$asmmode GAS}

{****************************************************************************
                               Primitives
****************************************************************************}

{$ifndef win64}
  {$define use_fast_repmovstos} { REP STOS uses nonvolatile RDI and would require a stack frame on Win64 to be SEH-compliant. }
{$endif}

{$ifdef use_fast_repmovstos}
var
  fast_large_repmovstosb : boolean; { Enhanced REP MOVSB and STOSB (ERMSB) feature @ CPUID(7).ebx[9]. }
{$endif}
var
  has_sse41_support,fpc_cpuinit_performed : boolean;

{$define FPC_SYSTEM_HAS_SPTR}
Function Sptr : Pointer;assembler;nostackframe;
asm
        movq    %rsp,%rax
end;

{$IFNDEF INTERNAL_BACKTRACE}
{$define FPC_SYSTEM_HAS_GET_FRAME}
function get_frame:pointer;assembler;nostackframe;
asm
        movq    %rbp,%rax
end;
{$ENDIF not INTERNAL_BACKTRACE}

{$define FPC_SYSTEM_HAS_GET_PC_ADDR}
function get_pc_addr:pointer;assembler;nostackframe;
asm
        movq    (%rsp),%rax
end;

{$define FPC_SYSTEM_HAS_GET_CALLER_ADDR}
function get_caller_addr(framebp:pointer;addr:pointer=nil):pointer;
begin
  get_caller_addr:=framebp;
  if assigned(framebp) then
    get_caller_addr:=PPointer(framebp)[1];
end;


{$define FPC_SYSTEM_HAS_GET_CALLER_FRAME}
function get_caller_frame(framebp:pointer;addr:pointer=nil):pointer;
begin
  get_caller_frame:=framebp;
  if assigned(framebp) then
    get_caller_frame:=PPointer(framebp)^;
end;

// The following assembler procedures are disabled for FreeBSD due to
// multiple issues with its old GNU assembler (Mantis #19188).
// Even after fixing them, it can be enabled only for the trunk version,
// otherwise bootstrapping won't be possible.
// Modified to use oldbinutils as in cpu.pp source, to allow easier use for other targets.
{$ifdef freebsd}
  {$ifndef overridebinutils}
    {$define oldbinutils}
  {$endif}
{$endif freebsd}

{$ifndef oldbinutils}


{$ifndef FPC_SYSTEM_HAS_MOVE}
{$define FPC_SYSTEM_HAS_MOVE}
procedure Move(const source;var dest;count:SizeInt);[public, alias: 'FPC_MOVE'];assembler;nostackframe;
{ Linux: rdi source, rsi dest, rdx count
  win64: rcx source, rdx dest, r8 count }
const
  NtThreshold = 256 * 1024; { this limit must be processor-specific (1/2 L2 cache size) }
  PrefetchDistance = 512;
asm
{$ifndef win64}
    mov    %rdx, %r8
    mov    %rsi, %rdx
    mov    %rdi, %rcx
{$endif win64}

    cmp    $3, %r8
    jle    .L3OrLess
    cmp    $8, %r8
    jle    .L4to8
    cmp    $16, %r8
    jle    .L9to16
    movups (%rcx), %xmm4         { First and last 16 bytes, used both in .L33OrMore and 17–32 branch. }
    movups -16(%rcx,%r8), %xmm5
    cmp    $32, %r8
    jg     .L33OrMore
    movups %xmm4, (%rdx)         { 17–32 bytes }
    movups %xmm5, -16(%rdx,%r8)
    ret

    .balign 16
.L3OrLess:
    cmp    $1, %r8
    jl     .LZero
    movzbl (%rcx), %eax
    je     .LOne
    movzwl -2(%rcx,%r8), %r9d
    mov    %r9w, -2(%rdx,%r8)
.LOne:
    mov    %al, (%rdx)
.LZero:
    ret

.L4to8:
    mov    (%rcx), %eax
    mov    -4(%rcx,%r8), %r9d
    mov    %eax, (%rdx)
    mov    %r9d, -4(%rdx,%r8)
    ret

.L9to16:
    mov    (%rcx), %rax
    mov    -8(%rcx,%r8), %r9
    mov    %rax, (%rdx)
    mov    %r9, -8(%rdx,%r8)
.Lquit:
    ret
    .byte  102,102,102,102,102,102,102,102,102,102,102,144 { Turns .balign 16 before .Lloop32f into a no-op. }

.L33OrMore:
    movups -32(%rcx,%r8), %xmm3  { Second vector from the end. Wasted read if .Lback branch is taken (it uses second vector from the start instead), }
                                 { but -32(%rcx,%r8) is about to become not accessible so easily, .Lback is rare, and small .Lback is even rarer / matters even less. }

    sub    %rdx, %rcx            { rcx = src - dest }
    jz     .Lquit                { exit if src=dest }

    mov    %rcx, %rax
    neg    %rax
    cmp    %rax, %r8
    ja     .Lback                { count (r8) > unsigned(dest - src) (rax) if regions overlap }

    mov    %rdx, %r9             { remember original dest to write first 16 bytes }
    add    %rdx, %r8             { Move dest to the next 16-byte boundary. +16 if already aligned, as first 16 bytes will be writen separately anyway. }
    add    $16, %rdx
    and    $-16, %rdx
    sub    %rdx, %r8

.LRestAfterNTf:
    sub    $32, %r8              { During the N× loop, r8 is N bytes less than actually remained to allow sub N+jae .LLoop instead of sub N+cmp N+jae .LLoop. }
    jbe    .LPost32f
    cmp    $NtThreshold-32, %r8
    jae    .Lntf                 { might jump back right away after more checks, but the branch is taken only on huge moves so it's better to take these checks out of here... }

    .balign 16                   { no-op }
.Lloop32f:
    movups (%rcx,%rdx), %xmm0
    movaps %xmm0, (%rdx)
    movups 16(%rcx,%rdx), %xmm0
    movaps %xmm0, 16(%rdx)
    add    $32, %rdx
    sub    $32, %r8
    ja     .Lloop32f

.LPost32f:                       { +32 fixup not applied after 32× loop, r8 = remaining - 32 here. }
    movups %xmm3, (%rdx, %r8)
    movups %xmm5, 16(%rdx,%r8)   { Write first and last 16 bytes after everything else. }
    movups %xmm4, (%r9)          { Important for <16-byte step between src and dest. }
    ret

    .balign 16
.Lntf:
    cmp    $NtThreshold, %rcx    { Maybe change mind: don't bother bypassing cache if src and dest are close to each other }
    jb     .Lloop32f             { (this check is performed here to not stand in the way of smaller counts) }
    sub    $PrefetchDistance+32, %r8 { r8 = remaining - prefetch distance - bytes per loop (64), but 32 was subtracted already. }

    .balign 16                   { no-op }
.Lntloop64f:
    prefetchnta 0+PrefetchDistance(%rcx,%rdx,1)
    movups (%rcx,%rdx,1), %xmm0
    movntps %xmm0, (%rdx)
    movups 16(%rcx,%rdx,1), %xmm0
    movntps %xmm0, 16(%rdx)
    movups 32(%rcx,%rdx,1), %xmm0
    movntps %xmm0, 32(%rdx)
    movups 48(%rcx,%rdx,1), %xmm0
    movntps %xmm0, 48(%rdx)
    add    $64, %rdx
    sub    $64, %r8
    jae    .Lntloop64f

    sfence
    add    $PrefetchDistance+64, %r8
    jmpq   .LRestAfterNTf        { go handle remaining bytes }
    .byte  102,102,102,102,102,102,144 { Turns .balign 16 before .Lloop32b into a no-op. }

{ backwards move }
.Lback:
    movups 16(%rcx,%rdx), %xmm3  { Second vector from the start. }
    lea    (%rdx,%r8), %r9       { points to the end of dest; remember to write last 16 bytes }
    lea    -1(%r9), %r8          { move dest to the previous 16-byte boundary... }
    and    $-16, %r8
    sub    %rdx, %r8
    add    %r8, %rdx

.LRestAfterNTb:
    sub    $32, %r8
    jbe    .LPost32b
    cmp    $NtThreshold-32, %r8
    jae    .Lntb

    .balign 16                   { no-op }
.Lloop32b:
    sub    $32, %rdx
    movups 16(%rcx,%rdx), %xmm0
    movaps %xmm0, 16(%rdx)
    movups (%rcx,%rdx), %xmm0
    movaps %xmm0, (%rdx)
    sub    $32, %r8
    ja     .Lloop32b

.LPost32b:
    sub    %r8, %rdx
    movups %xmm3, -16(%rdx)
    movups %xmm4, -32(%rdx)
    movups %xmm5, -16(%r9)
    ret

    .balign 16
.Lntb:
    cmp    $-NtThreshold,%rcx
    jnb    .Lloop32b
    sub    $PrefetchDistance+32, %r8

    .balign 16                   { no-op }
.Lntloop64b:
    prefetchnta -PrefetchDistance(%rcx,%rdx,1)
    sub    $64, %rdx
    movups 48(%rcx,%rdx,1), %xmm0
    movntps %xmm0, 48(%rdx)
    movups 32(%rcx,%rdx,1), %xmm0
    movntps %xmm0, 32(%rdx)
    movups 16(%rcx,%rdx,1), %xmm0
    movntps %xmm0, 16(%rdx)
    movups (%rcx,%rdx,1), %xmm0
    movntps %xmm0, (%rdx)
    sub    $64, %r8
    jae    .Lntloop64b

    sfence
    add    $PrefetchDistance+64, %r8
    jmpq   .LRestAfterNTb
end;
{$endif FPC_SYSTEM_HAS_MOVE}

{$if not defined(FPC_SYSTEM_HAS_FILLCHAR)
  or not defined(FPC_SYSTEM_HAS_FILLWORD)
  or not defined(FPC_SYSTEM_HAS_FILLDWORD)
  or not defined(FPC_SYSTEM_HAS_FILLQWORD)}
procedure FillXxxx_MoreThanTwoXmms; assembler; nostackframe;
{ Input:
  rcx = 'x'
  rdx = byte count
  xmm0 = pattern for ALIGNED writes
  First and last 16 bytes are written. }
const
{$ifdef use_fast_repmovstos}
  ErmsThreshold = 1536;
{$endif}
  NtThreshold = 4 * 1024 * 1024;
asm
    { x can start and end misaligned on the vector boundary:

      x = ~~][H1][H2][...][T2][T1]~
          [UH]                 [UT]

      UH (“unaligned head”) potentially overlaps with H1 and is already written with 'movdqu' by the caller.
      At least 1 of its bytes is exclusive to it, i.e. if x is already aligned, H1 starts at byte 16.

      H1 and so on are called “aligned heads” or just “heads”.
      T1 and so on are called “aligned tails” or just “tails”.

      UT (“unaligned tail”) is written by the caller as well.
      At least 1 of its bytes is exclusive to it as well, that’s why 65 is subtracted below instead of 64. }

    lea    -65(%rcx,%rdx), %rax
    and    $-16, %rax { rax = “T4” (possibly fictive). }
    mov    %rax, %rdx { Remember T4 to rdx. }
    and    $-16, %rcx { rcx = H1 − 16. }
    sub    %rcx, %rax { rax = aligned byte count − 48. }
    movdqa %xmm0, 16(%rcx) { Write H1. }
    cmp    $32-48, %rax
    jle    .LOneAlignedTailWrite
    movdqa %xmm0, 32(%rcx) { Write H2. }
    cmp    $64-48, %rax
    jle    .LTwoAlignedTailWrites
    sub    $48, %rax { rax = aligned byte count − 96 (32 bytes already written + 64 bytes written after loop). }
    jle    .LFourAlignedTailWrites

    add    $48, %rcx { rcx = H3. }
{$ifdef use_fast_repmovstos}
    cmp    $ErmsThreshold-64, %rax { Need to write aligned byte count − 32 bytes already written. rax = aligned byte count − 96, so compare rax + 64 to ErmsThreshold, or rax to ErmsThreshold − 64. }
    jae    .LRepStos
{$else}
    cmp    $NtThreshold, %rax
    jae    .L64xNT_Body
{$endif}

.balign 16
.L64x_Body:
    movdqa %xmm0, (%rcx)
    movdqa %xmm0, 16(%rcx)
    movdqa %xmm0, 32(%rcx)
    movdqa %xmm0, 48(%rcx)
    add    $64, %rcx
    sub    $64, %rax
    ja     .L64x_Body

.LFourAlignedTailWrites:
    movdqa %xmm0, (%rdx) { T4 }
    movdqa %xmm0, 16(%rdx) { T3 }
.LTwoAlignedTailWrites:
    movdqa %xmm0, 32(%rdx) { T2 }
.LOneAlignedTailWrite:
    movdqa %xmm0, 48(%rdx) { T1 }
    ret

{$ifdef use_fast_repmovstos}
.LRepStos:
{$ifdef FPC_PIC}
    movq   fast_large_repmovstosb@GOTPCREL(%rip), %r8
    cmpb   $1, (%r8)
{$else FPC_PIC}
    cmpb   $1, fast_large_repmovstosb(%rip)
{$endif FPC_PIC}
    jne    .LRepStosIsNotBetter
{$ifdef win64}
    push   %rdi { For tests on Windows; however this is SEH incompliant so the entire use_fast_repmovstos branch is disabled by default! }
{$endif}
    mov    %rcx, %rdi { rdi = REP STOS destination. }
    lea    64(%rax), %rcx
    shr    $3, %rcx { rcx = count of REP STOSQ blocks up to T1 (might be 1 more than strictly required if T1 and UT overlap is 8 or more, don’t care). }
    movq   %xmm0, %rax { recover pattern for aligned writes back to GPR :) }
    rep stosq
{$ifdef win64}
    pop    %rdi
{$endif}
    ret
{$endif}

.LRepStosIsNotBetter:
    cmp    $NtThreshold-64, %rax
    jb     .L64x_Body

.balign 16
.L64xNT_Body:
    movntdq %xmm0, (%rcx)
    movntdq %xmm0, 16(%rcx)
    movntdq %xmm0, 32(%rcx)
    movntdq %xmm0, 48(%rcx)
    add    $64, %rcx
    sub    $64, %rax
    ja     .L64xNT_Body
    sfence
    jmp    .LFourAlignedTailWrites
end;
{$endif FPC_SYSTEM_HAS_FILLxxxx}

{$ifndef FPC_SYSTEM_HAS_FILLCHAR}
{$define FPC_SYSTEM_HAS_FILLCHAR}
Procedure FillChar(var x;count:SizeInt;value:byte);assembler;nostackframe;
  asm
{ win64: rcx dest, rdx count, r8b value
  linux: rdi dest, rsi count, rdx value }
    movzbl {$ifdef win64} %r8b {$else} %dl {$endif}, %eax
    imul   $0x01010101, %eax
{$ifndef win64}
    mov    %rsi, %rdx
    mov    %rdi, %rcx
{$endif win64}

    cmp    $3, %rdx
    jle    .L3OrLess
    cmp    $16, %rdx
    jl     .L4to15

    movd   %eax, %xmm0
    pshufd $0, %xmm0, %xmm0
    movdqu %xmm0, (%rcx)
    movdqu %xmm0, -16(%rcx,%rdx)
    cmp    $32, %rdx
    jg     FillXxxx_MoreThanTwoXmms
    ret

.L4to15:
    mov    %eax, (%rcx)
    cmp    $8, %edx
    jle    .LLast4
    mov    %eax, 4(%rcx)
    mov    %eax, -8(%rcx,%rdx)
.LLast4:
    mov    %eax, -4(%rcx,%rdx)
    ret

.L3OrLess:
    test   %rdx, %rdx
    jle    .LQuit
    mov    %al, (%rcx)
    mov    %al, -1(%rcx,%rdx)
    shr    $1, %edx
    mov    %al, (%rcx,%rdx)
.LQuit:
  end;
{$endif FPC_SYSTEM_HAS_FILLCHAR}

{$ifndef FPC_SYSTEM_HAS_FILLWORD}
{$define FPC_SYSTEM_HAS_FILLWORD}
procedure FillWord(var x;count:SizeInt;value:word);assembler;nostackframe;
  asm
{$ifdef win64}
    movzwl %r8w, %eax
    shl    $16, %r8d
    or     %r8d, %eax
{$else}
    movzwl %dx, %eax
    shl    $16, %edx
    or     %edx, %eax
    mov    %rsi, %rdx
    mov    %rdi, %rcx
{$endif}

    cmp    $3, %rdx
    jle    .L3OrLess
    cmp    $8, %rdx
    jle    .L4to8

    movd   %eax, %xmm0
    pshufd $0, %xmm0, %xmm0 { xmm0 = pattern for unaligned writes }
    movdqu %xmm0, (%rcx)
    movdqu %xmm0, -16(%rcx,%rdx,2)
    cmp    $16, %rdx
    jg     .LMoreThanTwoXMMs
    ret

.LMoreThanTwoXMMs:
    shl    $1, %rdx { rdx = byte count }
    mov    %rcx, %r8
    shl    $3, %ecx
    rol    %cl, %eax { misalign the pattern by the misalignment of x }
    mov    %r8, %rcx
    movd   %eax, %xmm0
    pshufd $0, %xmm0, %xmm0 { xmm0 = pattern for aligned writes }
    jmp    FillXxxx_MoreThanTwoXmms

.L4to8:
    mov    %eax, %r8d
    shl    $32, %r8
    or     %r8, %rax
    mov    %rax, (%rcx)
    mov    %rax, -8(%rcx,%rdx,2)
    ret

.L3OrLess:
    test   %rdx, %rdx
    jle    .LQuit
    mov    %ax, (%rcx)
    mov    %ax, -2(%rcx,%rdx,2)
    shr    $1, %edx
    mov    %ax, (%rcx,%rdx,2)
.LQuit:
  end;
{$endif FPC_SYSTEM_HAS_FILLWORD}

{$ifndef FPC_SYSTEM_HAS_FILLDWORD}
{$define FPC_SYSTEM_HAS_FILLDWORD}
procedure FillDWord(var x;count:SizeInt;value:DWord);assembler;nostackframe;
  asm
{$ifdef win64}
    mov    %r8d, %eax
{$else}
    mov    %edx, %eax
    mov    %rsi, %rdx
    mov    %rdi, %rcx
{$endif win64}

    cmp    $3, %rdx
    jle    .L3OrLess
    cmp    $8, %rdx
    jle    .L4to8

    movd   %eax, %xmm0
    pshufd $0, %xmm0, %xmm0 { xmm0 = pattern for unaligned writes }
    movdqu %xmm0, (%rcx)
    movdqu %xmm0, -16(%rcx,%rdx,4)

    shl    $2, %rdx { rdx = byte count }
    mov    %rcx, %r8
    shl    $3, %ecx
    rol    %cl, %eax { misalign the pattern by the misalignment of x }
    mov    %r8, %rcx
    movd   %eax, %xmm0
    pshufd $0, %xmm0, %xmm0 { xmm0 = pattern for aligned writes }
    jmp    FillXxxx_MoreThanTwoXmms

.L4to8:
{$ifndef win64} { on win64, eax = r8d already. }
    mov    %eax, %r8d
{$endif}
    shl    $32, %r8
    or     %r8, %rax
    mov    %rax, (%rcx)
    mov    %rax, 8(%rcx)
    mov    %rax, -16(%rcx,%rdx,4)
    mov    %rax, -8(%rcx,%rdx,4)
    ret

.L3OrLess:
    test   %rdx, %rdx
    jle    .LQuit
    mov    %eax, (%rcx)
    mov    %eax, -4(%rcx,%rdx,4)
    shr    $1, %edx
    mov    %eax, (%rcx,%rdx,4)
.LQuit:
  end;
{$endif FPC_SYSTEM_HAS_FILLDWORD}

{$ifndef FPC_SYSTEM_HAS_FILLQWORD}
{$define FPC_SYSTEM_HAS_FILLQWORD}
procedure FillQWord(var x;count:SizeInt;value:QWord);assembler;nostackframe;
  asm
{$ifdef win64}
    mov    %r8, %rax
{$else}
    mov    %rdx, %rax
    mov    %rsi, %rdx
    mov    %rdi, %rcx
{$endif win64}

    cmp    $2, %rdx
    jle    .L2OrLess
    cmp    $6, %rdx
    jle    .L3to6

    movq   %rax, %xmm0
    punpcklqdq %xmm0, %xmm0 { xmm0 = pattern for unaligned writes }
    movdqu %xmm0, (%rcx)
    movdqu %xmm0, -16(%rcx,%rdx,8)

    shl    $3, %rdx { rdx = byte count }
    mov    %rcx, %r8
    shl    $3, %ecx
    rol    %cl, %rax { misalign the pattern by the misalignment of x }
    mov    %r8, %rcx
    movq   %rax, %xmm0
    punpcklqdq %xmm0, %xmm0 { xmm0 = pattern for aligned writes }
    jmp    FillXxxx_MoreThanTwoXmms

.L3to6:
    mov    %rax, (%rcx)
    mov    %rax, 8(%rcx)
    mov    %rax, 16(%rcx)
    mov    %rax, -24(%rcx,%rdx,8)
    mov    %rax, -16(%rcx,%rdx,8)
    mov    %rax, -8(%rcx,%rdx,8)
    ret

.L2OrLess:
    test   %rdx, %rdx
    jle    .LQuit
    mov    %rax, (%rcx)
    mov    %rax, -8(%rcx,%rdx,8)
.LQuit:
  end;
{$endif FPC_SYSTEM_HAS_FILLQWORD}

{$ifndef FPC_SYSTEM_HAS_INDEXBYTE}
{$define FPC_SYSTEM_HAS_INDEXBYTE}
function IndexByte(Const buf;len:SizeInt;b:byte):SizeInt; assembler; nostackframe;
{ win64: rcx buf, rdx len, r8b word
  linux: rdi buf, rsi len, rdx word }
asm
    test   len, len
    jz     .Lnotfound                  { exit if len=0 }

    movd   {$ifdef win64} %r8d {$else} %edx {$endif}, %xmm1
    mov    {$ifdef win64} %ecx {$else} %edi {$endif}, %eax
    punpcklbw  %xmm1, %xmm1
    punpcklbw  %xmm1, %xmm1
    and    $4095, %eax
    pshufd $0, %xmm1, %xmm1

    cmp    $4080, %eax
    ja     .LCrossPage

    movdqu    ({$ifdef win64} %rcx {$else} %rdi {$endif}), %xmm0 { Analyze first 16 bytes, unaligned. }
    pcmpeqb   %xmm1, %xmm0
    pmovmskb  %xmm0, %eax
    test      %eax, %eax
    jz        .LContinueAligned

    bsf    %eax, %eax
    cmp    len, %rax
    jae    .Lnotfound
    ret

    .byte  {$ifndef win64}102,102,102,102,{$endif}102,102,102,102,102,102,102,102,102,144 { Make .balign 16 before .Lloop a no-op. }
.LContinueAligned:
    cmp    $16, len                    { Length might be explicitly set to 16 or less; if so, skip a bit of work. }
    jbe    .Lnotfound                  { (Or rather, this check is *required* unless jumping to .Lcontinue instead of going directly to .Lloop) }

{$ifdef win64}
    mov    %rcx, %r8                   { r8 = original ptr, rcx = buf + 16 for aligning & shifts. }
    add    $16, %rcx
{$else}
    lea    16(%rdi), %rcx              { rdi = original ptr, rcx = buf + 16 for aligning & shifts. }
{$endif}
    and    $-0x10, %rcx                { first aligned address after buf }
    sub    {$ifdef win64} %r8 {$else} %rdi {$endif}, %rcx { rcx=number of valid bytes, r8/rdi=original ptr }

    .balign 16
.Lloop:
    movdqa ({$ifdef win64} %r8 {$else} %rdi {$endif},%rcx), %xmm0 { r8/rdi and rcx may have any values, }
    add    $16, %rcx                   { but their sum is evenly divisible by 16. }
    pcmpeqb %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    test   %eax, %eax
    jnz    .Lmatch
.Lcontinue:
    cmp    %rcx, len
    ja     .Lloop
.Lnotfound:
    or     $-1, %rax
    ret

.LCrossPage:
{$ifdef win64}
    mov    %rcx, %r8                   { r8 = original ptr, rcx = buf + 16 for aligning & shifts. }
    add    $16, %rcx
{$else}
    lea    16(%rdi), %rcx              { rdi = original ptr, rcx = buf + 16 for aligning & shifts. }
{$endif}
    and    $-0x10, %rcx                { first aligned address after buf }
    movdqa -16(%rcx), %xmm0            { Fetch first 16 bytes (up to 15 bytes before target) }
    sub    {$ifdef win64} %r8 {$else} %rdi {$endif}, %rcx { rcx=number of valid bytes, r8/rdi=original ptr }

    pcmpeqb %xmm1, %xmm0               { compare with pattern and get bitmask }
    pmovmskb %xmm0, %eax

    shl    %cl, %eax                   { shift valid bits into high word }
    and    $0xffff0000, %eax           { clear low word containing invalid bits }
    shr    %cl, %eax                   { shift back }
    jz     .Lcontinue
.Lmatch:
    bsf    %eax, %eax
    lea    -16(%rcx,%rax), %rax
    cmp    %rax, len                   { check against the buffer length }
    jbe    .Lnotfound
end;
{$endif FPC_SYSTEM_HAS_INDEXBYTE}

{$ifndef FPC_SYSTEM_HAS_INDEXWORD}
{$define FPC_SYSTEM_HAS_INDEXWORD}
function IndexWord(Const buf;len:SizeInt;b:word):SizeInt; assembler; nostackframe;
{ win64: rcx buf, rdx len, r8b word
  linux: rdi buf, rsi len, rdx word }
asm
    test   len, len
    jz     .Lnotfound                  { exit if len=0 }
    movd   {$ifdef win64} %r8d {$else} %edx {$endif}, %xmm1
{$ifdef win64}
    mov    %rcx, %r8                   { r8 = original ptr, rcx = buf + 16 for aligning & shifts. }
    add    $16, %rcx
{$else}
    lea    16(%rdi), %rcx              { rdi = original ptr, rcx = buf + 16 for aligning & shifts. }
{$endif}
    punpcklwd  %xmm1, %xmm1
    and    $-0x10, %rcx
    pshufd $0, %xmm1, %xmm1
    movdqa -16(%rcx), %xmm0            { Fetch first 16 bytes (up to 14 bytes before target) }
    sub    {$ifdef win64} %r8 {$else} %rdi {$endif}, %rcx { rcx=number of valid bytes }

    test   $1, {$ifdef win64} %r8b {$else} %dil {$endif} { if buffer isn't aligned to word boundary, }
    jnz    .Lunaligned                 { use a different algorithm }

    pcmpeqw  %xmm1, %xmm0
    pmovmskb %xmm0, %eax

    shl    %cl, %eax
    and    $0xffff0000, %eax
    shr    %cl, %eax
    shr    $1, %ecx                    { bytes->words }
    test   %eax, %eax
    jz     .Lcontinue
.Lmatch:
    bsf    %eax, %eax
    shr    $1, %eax                    { in words }
    lea    -8(%rcx,%rax), %rax
    cmp    %rax, len
    jbe    .Lnotfound                  { if match is after the specified length, ignore it }
    retq

    .balign 16
.Lloop:
    movdqa ({$ifdef win64} %r8 {$else} %rdi {$endif},%rcx,2), %xmm0
    add    $8, %rcx
    pcmpeqw  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    test   %eax, %eax
    jnz    .Lmatch
.Lcontinue:
    cmp    %rcx, len
    ja     .Lloop

.Lnotfound:
    or    $-1, %rax
    retq

.Lunaligned:
    movdqa  %xmm1, %xmm2               { (mis)align the pattern (in this particular case: }
    psllw   $8, %xmm1                  {   swap bytes of each word of pattern) }
    psrlw   $8, %xmm2
    por     %xmm2, %xmm1

    pcmpeqb  %xmm1, %xmm0
    pmovmskb %xmm0, %eax

    shl    %cl, %eax
    and    $0xffff0000, %eax
    shr    %cl, %eax

    add    len, len                    { length words -> bytes }
    xor    %r10d, %r10d                { nothing to merge yet }
    jmp    .Lcontinue_u

    .balign 16
.Lloop_u:
    movdqa ({$ifdef win64} %r8 {$else} %rdi {$endif},%rcx), %xmm0
    add    $16, %rcx
    pcmpeqb %xmm1, %xmm0               { compare by bytes }
    shr    $16, %r10d                  { bit 16 shifts into 0 }
    pmovmskb %xmm0, %eax
.Lcontinue_u:
    shl    $1, %eax                    { 15:0 -> 16:1 }
    or     %r10d, %eax                 { merge bit 0 from previous round }
    mov    %eax, %r10d
    shr    $1, %eax                    { now AND together adjacent pairs of bits }
    and    %r10d, %eax
    and    $0x5555, %eax               { also reset odd bits }
    jnz    .Lmatch_u
    cmpq   %rcx, len
    ja     .Lloop_u

.Lnotfound_u:
    or     $-1, %rax
    retq
.Lmatch_u:
    bsf    %eax, %eax
    lea    -16(%rcx,%rax), %rax
    cmp    %rax, len
    jbe    .Lnotfound_u                { if match is after the specified length, ignore it }
    sar    $1, %rax                    { in words }
end;
{$endif FPC_SYSTEM_HAS_INDEXWORD}

{$ifndef FPC_SYSTEM_HAS_INDEXDWORD}
{$define FPC_SYSTEM_HAS_INDEXDWORD}
function IndexDWord(Const buf;len:SizeInt;b:dword):SizeInt; assembler; nostackframe;
asm
{$ifdef win64}
    mov      %rcx, %rax
{$else}
    mov      %rdx, %r8
    mov      %rsi, %rdx
    mov      %rdi, %rax
{$endif}
    cmp      $4, %rdx
    jle      .LDwordwise_Prepare
    sub      $4, %rdx
    movd     %r8d, %xmm1
    pshufd   $0, %xmm1, %xmm1
.balign 16
.L4x_Body:
    movdqu   (%rax), %xmm0
    pcmpeqd  %xmm1, %xmm0
    pmovmskb %xmm0, %r8d
    test     %r8d, %r8d
    jnz      .LFoundAtMask
    add      $16, %rax
    sub      $4, %rdx
    jg       .L4x_Body

    lea      (%rax,%rdx,4), %rax
    movdqu   (%rax), %xmm0
    pcmpeqd  %xmm1, %xmm0
    pmovmskb %xmm0, %r8d
    test     %r8d, %r8d
    jnz      .LFoundAtMask
    or       $-1, %rax
    ret

.balign 16 { no-op }
.LDwordwise_Body:
    cmp      (%rax), %r8d
    je       .LFoundAtRax
    add      $4, %rax
.LDwordwise_Prepare:
    sub      $1, %rdx
    jae      .LDwordwise_Body
    or       $-1, %rax
    ret

.LFoundAtMask:
    bsf      %r8d, %r8d
    add      %r8, %rax
.LFoundAtRax:
    sub      {$ifdef win64} %rcx {$else} %rdi {$endif}, %rax
    shr      $2, %rax
end;
{$endif FPC_SYSTEM_HAS_INDEXDWORD}

{$ifndef FPC_SYSTEM_HAS_INDEXQWORD}
{$define FPC_SYSTEM_HAS_INDEXQWORD}
function IndexQWord_Plain(Const buf;len:SizeInt;b:QWord):SizeInt; assembler; nostackframe;
{ win64: rcx=buf, rdx=len, r8=b
  else:  rdi=buf, rsi=len, rdx=b }
asm
    mov      buf, %rax
    sub      $8, %rax
.balign 16
.LQwordwise_Next:
    add      $8, %rax
    sub      $1, len
    jb       .LNothing
    cmpq     b, (%rax)
    jne      .LQwordwise_Next
    sub      buf, %rax
    shr      $3, %rax
    ret

.LNothing:
    mov      $-1, %rax
end;

function {$ifdef CPUX86_HAS_SSE4_1} IndexQWord {$else} IndexQWord_SSE41 {$endif}(Const buf;len:SizeInt;b:QWord):SizeInt; assembler; nostackframe;
{ win64: rcx=buf, rdx=len, r8=b
  else:  rdi=buf, rsi=len, rdx=b }
asm
    cmp      $6, len
    jle      IndexQWord_Plain
    mov      buf, %rax
    movq     {$ifdef win64} %r8 {$else} %rdx {$endif}, %xmm0
    punpcklqdq %xmm0, %xmm0 { xmm0 = pattern of 'b's. }
    sub      $6, len
.balign 16
.L6x_Loop:
    movdqu   (%rax), %xmm1
    pcmpeqq  %xmm0, %xmm1 { xmm1 = cmpeq(vec 0, pattern) }
    movdqu   16(%rax), %xmm2
    pcmpeqq  %xmm0, %xmm2
    por      %xmm1, %xmm2 { xmm2 = cmpeq(vec 0, pattern) or cmpeq(vec 1, pattern) }
    movdqu   32(%rax), %xmm3
    pcmpeqq  %xmm0, %xmm3
    por      %xmm2, %xmm3 { xmm3 = cmpeq(vec 0, pattern) or cmpeq(vec 1, pattern) or cmpeq(vec 2, pattern) }
    ptest    %xmm3, %xmm3
    jnz      .LFound
    add      $48, %rax
    sub      $6, len
    jge      .L6x_Loop
    lea      (%rax,{$ifdef win64} %rdx {$else} %rsi {$endif},8), %rax { Point to last 3 vectors. }
    cmp      $-5, len
    jge      .L6x_Loop { Reuse .L6x_Loop to compare last 3 vectors, if not compared already. }
    mov      $-1, %rax
    ret

.LFound:
    sub      buf, %rax
    ptest    %xmm1, %xmm1
    jnz      .LFoundAtXmm1
    ptest    %xmm2, %xmm2
    jnz      .LFoundAtXmm2
    add      $16, %rax
    movdqa   %xmm3, %xmm2
.LFoundAtXmm2:
    add      $16, %rax
    movdqa   %xmm2, %xmm1
.LFoundAtXmm1:
    pmovmskb %xmm1, %ecx
    bsf      %ecx, %ecx
    add      %rcx, %rax
    shr      $3, %rax
end;

{$ifndef CPUX86_HAS_SSE4_1}
function IndexQWord_Dispatch(const buf;len:SizeInt;b:QWord):SizeInt; forward;

var
  IndexQWord_Impl: function(const buf;len:SizeInt;b:QWord):SizeInt = @IndexQWord_Dispatch;

function IndexQWord_Dispatch(const buf;len:SizeInt;b:QWord):SizeInt;
begin
  if not fpc_cpuinit_performed then
    exit(IndexQWord_Plain(buf,len,b));
  if has_sse41_support then
    IndexQWord_Impl:=@IndexQWord_SSE41
  else
    IndexQWord_Impl:=@IndexQWord_Plain;
  result:=IndexQWord_Impl(buf,len,b);
end;

function IndexQWord(const buf;len:SizeInt;b:QWord):SizeInt;
begin
  result:=IndexQWord_Impl(buf,len,b);
end;
{$endif ndef CPUX86_HAS_SSE4_1}
{$endif FPC_SYSTEM_HAS_INDEXQWORD}

{$endif freebsd}

{$ifndef FPC_SYSTEM_HAS_COMPAREBYTE}
{$define FPC_SYSTEM_HAS_COMPAREBYTE}
function CompareByte(Const buf1,buf2;len:SizeInt):SizeInt; assembler; nostackframe;
{ win64: rcx buf, rdx buf, r8 len
  linux: rdi buf, rsi buf, rdx len }
asm
{$ifndef win64}
    mov      %rdx, %r8
    mov      %rsi, %rdx
    mov      %rdi, %rcx
{$endif win64}
    { rcx = buf1, rdx = buf2, r8 = len }
    cmp      $1, %r8
    jle      .L1OrLess

    cmp      $16, %r8
    jae      .LVecOrMore

    { 2 to 15 bytes: check for page cross. Pessimistic variant that has false positives, but is faster. }
    mov      %ecx, %eax
    or       %edx, %eax
    and      $4095, %eax
    cmp      $4080, %eax
    ja       .LCantOverReadBoth

    { Over-read both as XMMs. }
    movdqu   (%rcx), %xmm0
    movdqu   (%rdx), %xmm1
    pcmpeqb  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jz       .LNothing
    bsf      %eax, %eax
    cmp      %r8d, %eax { Ignore garbage beyond 'len'. }
    jae      .LNothing
    movzbl   (%rdx,%rax), %edx
    movzbl   (%rcx,%rax), %eax
    sub      %rdx, %rax
    ret

.balign 16
.LNothing:
    xor      %eax, %eax
    ret

.LAligned32xLoop_TwoVectorsDiffer:
    add      %rcx, %rdx { restore rdx = buf2 }
    pmovmskb %xmm0, %r8d { Is there a difference in the first vector? }
    inc      %r8w
    jz       .LVec1Differs { No difference in the first vector, xmm0 is all ones, eax = pmovmskb(pcmpeqb(buf1 + 16, buf2 + 16)) from the loop body. }
    mov      %r8d, %eax
.LVec0Differs:
    bsf      %eax, %eax
    movzbl   (%rdx,%rax), %edx
    movzbl   (%rcx,%rax), %eax
    sub      %rdx, %rax
    ret
    .byte    102,102,102,102,102,102,102,144 { Turn .balign 16 before .LAligned32xLoop_Body into a no-op. }

.LVecOrMore:
    { Compare first vectors. }
    movdqu   (%rcx), %xmm0
    movdqu   (%rdx), %xmm1
    pcmpeqb  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVec0Differs

    sub      $32, %r8
    jbe      .LLastVec

    { Compare second vectors. }
    movdqu   16(%rcx), %xmm0
    movdqu   16(%rdx), %xmm1
    pcmpeqb  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVec1Differs

    cmp      $32, %r8
    jbe      .LLastTwoVectors

    { More than four vectors: aligned loop. }
    sub      %rcx, %rdx { rdx = buf2 - buf1 }
    jz       .LNothing { Exit if buf1 = buf2. }
    lea      -32(%rcx,%r8), %r8 { buffer end - last two vectors handled separately - first two vectors already analyzed (by the fact ecx was still len - 32). }
    and      $-16, %rcx { Align buf1. First two vectors already analyzed are skipped by +32 on the first loop iteration. }
    sub      %rcx, %r8 { r8 = count to be handled with loop }
.balign 16 { no-op }
.LAligned32xLoop_Body:
    add      $32, %rcx
    { Compare two XMMs, reduce the result with 'and'. }
    movdqu   (%rdx,%rcx), %xmm0
    pcmpeqb  (%rcx), %xmm0 { xmm0 = pcmpeqb(buf1, buf2) }
    movdqu   16(%rdx,%rcx), %xmm1
    pcmpeqb  16(%rcx), %xmm1
    pand     %xmm0, %xmm1 { xmm1 = xmm0 and pcmpeqb(buf1 + 16, buf2 + 16) }
    pmovmskb %xmm1, %eax
    inc      %ax
    jnz      .LAligned32xLoop_TwoVectorsDiffer
    sub      $32, %r8
    ja       .LAligned32xLoop_Body
    add      %rcx, %rdx { restore rdx = buf2 }
    add      $32, %r8
.LLastTwoVectors:
    movdqu   (%rcx,%r8), %xmm0
    movdqu   (%rdx,%r8), %xmm1
    pcmpeqb  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVecEm2Differs
.LLastVec:
    movdqu   16(%rcx,%r8), %xmm0
    movdqu   16(%rdx,%r8), %xmm1
    pcmpeqb  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVecEm1Differs
    xor      %eax, %eax
    ret

.LVec1Differs:
    xor      %r8d, %r8d
.LVecEm1Differs:
    add      $16, %r8
.LVecEm2Differs:
    bsf      %eax, %eax
    add      %r8, %rax
    movzbl   (%rdx,%rax), %edx
    movzbl   (%rcx,%rax), %eax
    sub      %rdx, %rax
    ret

.LCantOverReadBoth:
    cmp      $8, %r8d
    ja       .L9to15
    cmp      $3, %r8d
    jle      .L2to3
    mov      (%rcx), %eax
    mov      (%rdx), %r9d
    cmp      %r9d, %eax
    jne      .L4xOr8xDiffer
    mov      -4(%rcx,%r8), %eax
    mov      -4(%rdx,%r8), %r9d
    cmp      %r9d, %eax
    jne      .L4xOr8xDiffer
    xor      %eax, %eax
    ret

.L9to15:
    mov      (%rcx), %rax
    mov      (%rdx), %r9
    cmp      %r9, %rax
    jne      .L4xOr8xDiffer
    mov      -8(%rcx,%r8), %rax
    mov      -8(%rdx,%r8), %r9
    cmp      %r9, %rax
    jne      .L4xOr8xDiffer
    xor      %eax, %eax
    ret

.L4xOr8xDiffer:
    bswap    %r9
    bswap    %rax
    cmp      %r9, %rax
    sbb      %rax, %rax
    or       $1, %rax
    ret

.L2to3:
    movzwl   (%rcx), %eax
    bswap    %eax
    shr      $1, %eax
    mov      -1(%rcx,%r8), %al
    movzwl   (%rdx), %ecx
    bswap    %ecx
    shr      $1, %ecx
    mov      -1(%rdx,%r8), %cl
    sub      %rcx, %rax
    ret

.L1OrLess:
    jl       .LUnbounded_Prepare
    movzbl   (%rcx), %eax
    movzbl   (%rdx), %edx
    sub      %rdx, %rax
    ret

.LUnbounded_Prepare:
    sub      %rcx, %rdx { rdx = buf2 - buf1 }
    test     %r8, %r8
    jnz      .LUnbounded_Body
    xor      %eax, %eax
    ret

.balign 16
.LUnbounded_Next:
    add      $1, %rcx
.LUnbounded_Body:
    movzbl   (%rdx,%rcx), %eax
    cmp      %al, (%rcx)
    je       .LUnbounded_Next
    sbb      %rax, %rax
    or       $1, %rax
end;
{$endif FPC_SYSTEM_HAS_COMPAREBYTE}


{$ifndef FPC_SYSTEM_HAS_COMPAREWORD}
{$define FPC_SYSTEM_HAS_COMPAREWORD}
function CompareWord(Const buf1,buf2;len:SizeInt):SizeInt; assembler; nostackframe;
asm
{$ifndef win64}
    mov      %rdx, %r8
    mov      %rsi, %rdx
    mov      %rdi, %rcx
{$endif win64}
    sub      %rcx, %rdx { rdx = buf2 - buf1 }
    cmp      $1, %r8
    jle      .LWordwise_Prepare
    mov      %r8, %rax
    shr      $62, %rax
    jnz      .LWordwise_Prepare
    cmp      $8, %r8
    jge      .LVecOrMore

    lea      (%rdx,%rcx), %eax
    or       %ecx, %eax
    and      $4095, %eax
    cmp      $4080, %eax
    ja       .LWordwise_Prepare
    movdqu   (%rdx,%rcx), %xmm0
    movdqu   (%rcx), %xmm1
    pcmpeqw  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    shl      $1, %r8 { convert to bytes }
    inc      %ax
    jz       .LNothing
    bsf      %eax, %eax
    cmp      %r8d, %eax
    jb       .LSubtractWords
.LNothing:
    xor      %eax, %eax
    ret

.balign 16
.LWordwise_Body:
    movzwl  (%rdx,%rcx), %eax
    cmp     %ax, (%rcx)
    jne     .LDoSbb
    add     $2, %rcx
.LWordwise_Prepare:
    sub     $1, %r8
    jae     .LWordwise_Body
    xor     %eax, %eax
    ret

.LDoSbb:
    sbb      %rax, %rax
    or       $1, %rax
    ret

.LVec0Differs:
    bsf      %eax, %eax
.LSubtractWords:
    add      %rcx, %rdx { recover rdx = buf2 }
    movzwl   (%rdx,%rax), %edx
    movzwl   (%rcx,%rax), %eax
    sub      %rdx, %rax
    ret

    .byte    102,102,102,102,102,102,102,102,102,102,102,102,102,102,144 { Turn .balign 16 before .LAligned8xLoop_Body into a no-op. }
.LVecOrMore:
    movdqu   (%rdx,%rcx), %xmm0 { Compare first vectors. }
    movdqu   (%rcx), %xmm1
    pcmpeqw  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVec0Differs

    shl      $1, %r8 { convert to bytes }
    sub      $32, %r8 { first 16 bytes already analyzed + last 16 bytes analyzed separately }
    jle      .LLastVec

    test     %rdx, %rdx
    jz       .LNothing { Exit if buf1 = buf2. }
    mov      %rcx, %r9 { save original buf1 to recover word position if byte mismatch found (aligned loop works in bytes to support misaligned buf1). }
    add      %rcx, %r8
    and      $-16, %rcx { align buf1; +16 is performed by the loop. }
    sub      %rcx, %r8

.balign 16 { no-op }
.LAligned8xLoop_Body:
    add      $16, %rcx
    movdqu   (%rdx,%rcx), %xmm0
    pcmpeqb  (%rcx), %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LAligned8xLoop_VecDiffers
    sub      $16, %r8
    ja       .LAligned8xLoop_Body
.LLastVec:
    lea      16(%rcx,%r8), %rcx { point to the last 16 bytes }
    movdqu   (%rdx,%rcx), %xmm0
    movdqu   (%rcx), %xmm1
    pcmpeqw  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVec0Differs
    xor      %eax, %eax
    ret

.LAligned8xLoop_VecDiffers:
    bsf      %eax, %eax
    add      %rax, %rcx
    sub      %r9, %rcx
    and      $-2, %rcx
    add      %r9, %rcx
    movzwl   (%rdx,%rcx), %edx
    movzwl   (%rcx), %eax
    sub      %rdx, %rax
end;
{$endif FPC_SYSTEM_HAS_COMPAREWORD}


{$ifndef FPC_SYSTEM_HAS_COMPAREDWORD}
{$define FPC_SYSTEM_HAS_COMPAREDWORD}
function CompareDWord(Const buf1,buf2;len:SizeInt):SizeInt; assembler; nostackframe;
asm
{$ifndef win64}
    mov      %rdx, %r8
    mov      %rsi, %rdx
    mov      %rdi, %rcx
{$endif win64}
    sub      %rcx, %rdx { rdx = buf2 - buf1 }
    cmp      $4, %r8
    jle      .LDwordwise_Prepare
    mov      %r8, %rax
    shr      $61, %rax
    jnz      .LDwordwise_Prepare

    movdqu   (%rdx,%rcx), %xmm0 { Compare first vectors. }
    movdqu   (%rcx), %xmm1
    pcmpeqd  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVec0Differs

    shl      $2, %r8 { convert to bytes }
    sub      $32, %r8 { first 16 bytes already analyzed + last 16 bytes analyzed separately }
    jle      .LLastVec

    test     %rdx, %rdx
    jz       .LNothing { Exit if buf1 = buf2. }
    mov      %rcx, %r9 { save original buf1 to recover word position if byte mismatch found (aligned loop works in bytes to support misaligned buf1). }
    add      %rcx, %r8
    and      $-16, %rcx { align buf1; +16 is performed by the loop. }
    sub      %rcx, %r8

.balign 16
.LAligned4xLoop_Body:
    add      $16, %rcx
    movdqu   (%rdx,%rcx), %xmm0
    pcmpeqb  (%rcx), %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LAligned4xLoop_VecDiffers
    sub      $16, %r8
    ja       .LAligned4xLoop_Body
.LLastVec:
    lea      16(%rcx,%r8), %rcx { point to the last 16 bytes }
    movdqu   (%rdx,%rcx), %xmm0
    movdqu   (%rcx), %xmm1
    pcmpeqd  %xmm1, %xmm0
    pmovmskb %xmm0, %eax
    inc      %ax
    jnz      .LVec0Differs
.LNothing:
    xor      %eax, %eax
    ret

.LVec0Differs:
    bsf      %eax, %eax
    add      %rcx, %rdx { recover rdx = buf2 }
    mov      (%rdx,%rax), %edx
    cmp      %edx, (%rcx,%rax)
    sbb      %rax, %rax
    or       $1, %rax
    ret

.LAligned4xLoop_VecDiffers:
    bsf      %eax, %eax
    add      %rax, %rcx
    sub      %r9, %rcx
    and      $-4, %rcx
    add      %r9, %rcx
    mov      (%rdx,%rcx), %edx
    cmp      %edx, (%rcx)
.LDoSbb:
    sbb      %rax, %rax
    or       $1, %rax
    ret

.balign 16
.LDwordwise_Body:
    mov     (%rdx,%rcx), %eax
    cmp     %eax, (%rcx)
    jne     .LDoSbb
    add     $4, %rcx
.LDwordwise_Prepare:
    sub     $1, %r8
    jae     .LDwordwise_Body
    xor     %eax, %eax
end;
{$endif FPC_SYSTEM_HAS_COMPAREDWORD}


{$define FPC_SYSTEM_HAS_DECLOCKED_LONGINT}
{ does a thread save inc/dec }
function declocked(var l : longint) : boolean;assembler; nostackframe;
  asm
     { this check should be done because a lock takes a lot }
     { of time!                                             }
{$ifdef FPC_PIC}
     movq       IsMultithread@GOTPCREL(%rip),%rax
     cmpl       $0,(%rax)
{$else FPC_PIC}
     cmpl       $0,IsMultithread(%rip)
{$endif FPC_PIC}
     jz         .Ldeclockedskiplock
     .byte      0xF0 // LOCK prefix.
.Ldeclockedskiplock:
     decl       {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
     setzb      %al
  end;


{$define FPC_SYSTEM_HAS_DECLOCKED_INT64}
function declocked(var l : int64) : boolean;assembler; nostackframe;
  asm
     { this check should be done because a lock takes a lot }
     { of time!                                             }
{$ifdef FPC_PIC}
     movq       IsMultithread@GOTPCREL(%rip),%rax
     cmpl       $0,(%rax)
{$else FPC_PIC}
     cmpl       $0,IsMultithread(%rip)
{$endif FPC_PIC}
     jz         .Ldeclockedskiplock
     .byte      0xF0 // LOCK prefix.
.Ldeclockedskiplock:
     decq       {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
     setzb      %al
  end;


{$define FPC_SYSTEM_HAS_INCLOCKED_LONGINT}
procedure inclocked(var l : longint);assembler; nostackframe;

  asm
     { this check should be done because a lock takes a lot }
     { of time!                                             }
{$ifdef FPC_PIC}
     movq       IsMultithread@GOTPCREL(%rip),%rax
     cmpl       $0,(%rax)
{$else FPC_PIC}
     cmpl       $0,IsMultithread(%rip)
{$endif FPC_PIC}
     jz         .Linclockedskiplock
     .byte      0xF0 // LOCK prefix.
.Linclockedskiplock:
     incl       {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
  end;


{$define FPC_SYSTEM_HAS_INCLOCKED_INT64}
procedure inclocked(var l : int64);assembler; nostackframe;

  asm
     { this check should be done because a lock takes a lot }
     { of time!                                             }
{$ifdef FPC_PIC}
     movq       IsMultithread@GOTPCREL(%rip),%rax
     cmpl       $0,(%rax)
{$else FPC_PIC}
     cmpl       $0,IsMultithread(%rip)
{$endif FPC_PIC}
     jz         .Linclockedskiplock
     .byte      0xF0 // LOCK prefix.
.Linclockedskiplock:
     incq       {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
  end;


{$ifndef VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_CMP_XCHG_8}
function fpc_atomic_cmp_xchg_8(var Target: shortint; NewValue: shortint; Comparand: shortint): shortint; assembler; nostackframe;
asm
        movl            {$ifdef win64} %r8d {$else} %edx {$endif},%eax
        lock
        cmpxchgb        NewValue,({$ifdef win64} %rcx {$else} %rdi {$endif})
end;

{$define FPC_SYSTEM_HAS_ATOMIC_CMP_XCHG_16}
function fpc_atomic_cmp_xchg_16(var Target: smallint; NewValue: smallint; Comparand: smallint): smallint; assembler; nostackframe;
asm
        movl            {$ifdef win64} %r8d {$else} %edx {$endif},%eax
        lock
        cmpxchgw        NewValue,({$ifdef win64} %rcx {$else} %rdi {$endif})
end;

{$define FPC_SYSTEM_HAS_ATOMIC_SUB_32}
function fpc_atomic_sub_32(var Target: longint; Value: longint): longint; assembler; nostackframe;
asm
        negl    Value
        lock
        xaddl   Value,({$ifdef win64} %rcx {$else} %rdi {$endif})
        movl    Value,%eax
end;

{$define FPC_SYSTEM_HAS_ATOMIC_SUB_64}
function fpc_atomic_sub_64(var Target: int64; Value: int64): int64; assembler; nostackframe;
asm
        negq    Value
        lock
        xaddq   Value,({$ifdef win64} %rcx {$else} %rdi {$endif})
        movq    Value,%rax
end;
{$endif VER3_2}


{$ifdef VER3_2}
function InterLockedDecrement (var Target: longint) : longint; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_DEC_32}
function fpc_atomic_dec_32 (var Target: longint) : longint; assembler; nostackframe;
{$endif VER3_2}
asm
        movl    $-1,%eax
        lock
        xaddl   %eax, {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
        decl    %eax
end;


{$ifdef VER3_2}
function InterLockedIncrement (var Target: longint) : longint; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_INC_32}
function fpc_atomic_inc_32 (var Target: longint) : longint; assembler; nostackframe;
{$endif VER3_2}
asm
        movl    $1,%eax
        lock
        xaddl   %eax, {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
        incl    %eax
end;


{$ifdef VER3_2}
function InterLockedExchange (var Target: longint;Source : longint) : longint; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_XCHG_32}
function fpc_atomic_xchg_32 (var Target: longint;Source : longint) : longint; assembler; nostackframe;
{$endif VER3_2}
asm
        xchgl   ({$ifdef win64} %rcx {$else} %rdi {$endif}),Source
        movl    Source,%eax
end;


{$ifdef VER3_2}
function InterLockedExchangeAdd (var Target: longint;Source : longint) : longint; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_ADD_32}
function fpc_atomic_add_32 (var Target: longint;Value : longint) : longint; assembler; nostackframe;
{$endif VER3_2}
asm
        lock
        xaddl   {$ifdef VER3_2} Source {$else} Value {$endif},({$ifdef win64} %rcx {$else} %rdi {$endif})
        movl    {$ifdef VER3_2} Source {$else} Value {$endif},%eax
end;


{$ifdef VER3_2}
function InterLockedCompareExchange(var Target: longint; NewValue, Comperand : longint): longint; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_CMP_XCHG_32}
function fpc_atomic_cmp_xchg_32 (var Target: longint; NewValue, Comparand : longint) : longint; assembler; nostackframe;
{$endif VER3_2}
asm
        movl            {$ifdef VER3_2} Comperand {$else} Comparand {$endif},%eax
        lock
        cmpxchgl        NewValue,({$ifdef win64} %rcx {$else} %rdi {$endif})
end;


{$ifdef VER3_2}
function InterLockedDecrement64 (var Target: int64) : int64; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_DEC_64}
function fpc_atomic_dec_64 (var Target: int64) : int64; assembler; nostackframe;
{$endif VER3_2}
asm
        movq    $-1,%rax
        lock
        xaddq   %rax, {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
        decq    %rax
end;


{$ifdef VER3_2}
function InterLockedIncrement64 (var Target: int64) : int64; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_INC_64}
function fpc_atomic_inc_64 (var Target: int64) : int64; assembler; nostackframe;
{$endif VER3_2}
asm
        movq    $1,%rax
        lock
        xaddq   %rax, {$ifdef win64} (%rcx) {$else} (%rdi) {$endif}
        incq    %rax
end;


{$ifdef VER3_2}
function InterLockedExchange64 (var Target: int64;Source : int64) : int64; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_XCHG_64}
function fpc_atomic_xchg_64 (var Target: int64;Source: int64) : int64; assembler; nostackframe;
{$endif VER3_2}
asm
        xchgq   ({$ifdef win64} %rcx {$else} %rdi {$endif}),Source
        movq    Source,%rax
end;


{$ifdef VER3_2}
function InterLockedExchangeAdd64 (var Target: int64;Source : int64) : int64; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_ADD_64}
function fpc_atomic_add_64 (var Target: int64;Value: int64) : int64; assembler; nostackframe;
{$endif VER3_2}
asm
        lock
        xaddq   {$ifdef VER3_2} Source {$else} Value {$endif},({$ifdef win64} %rcx {$else} %rdi {$endif})
        movq    {$ifdef VER3_2} Source {$else} Value {$endif},%rax
end;


{$ifdef VER3_2}
function InterLockedCompareExchange64(var Target: int64; NewValue, Comperand : int64): int64; assembler; nostackframe;
{$else VER3_2}
{$define FPC_SYSTEM_HAS_ATOMIC_CMP_XCHG_64}
function fpc_atomic_cmp_xchg_64 (var Target: int64; NewValue, Comparand : int64) : int64; [public, alias:'FPC_ATOMIC_CMP_XCHG_64']; assembler; nostackframe;
{$endif VER3_2}
asm
        movq            {$ifdef VER3_2} Comperand {$else} Comparand {$endif},%rax
        lock
        cmpxchgq        NewValue,({$ifdef win64} %rcx {$else} %rdi {$endif})
end;


{****************************************************************************
                                  FPU
****************************************************************************}

const
  { Internal constants for use in system unit }
  FPU_Invalid = 1;
  FPU_Denormal = 2;
  FPU_DivisionByZero = 4;
  FPU_Overflow = 8;
  FPU_Underflow = $10;
  FPU_StackUnderflow = $20;
  FPU_StackOverflow = $40;
  FPU_ExceptionMask = $ff;

  MM_Invalid = 1;
  MM_Denormal = 2;
  MM_DivisionByZero = 4;
  MM_Overflow = 8;
  MM_Underflow = $10;
  MM_Precicion = $20;
  MM_ExceptionMask = $3f;

  MM_MaskInvalidOp = %0000000010000000;
  MM_MaskDenorm    = %0000000100000000;
  MM_MaskDivZero   = %0000001000000000;
  MM_MaskOverflow  = %0000010000000000;
  MM_MaskUnderflow = %0000100000000000;
  MM_MaskPrecision = %0001000000000000;

{$define FPC_SYSTEM_HAS_FPC_CPUINIT}
procedure fpc_cpuinit;
  var
    _eax,cpuid7_ebx,cpuid1_ecx : dword;
  begin
    { don't let libraries influence the FPU cw set by the host program }
    if IsLibrary then
      begin
        Default8087CW:=Get8087CW;
        DefaultMXCSR:=GetMXCSR;
      end;
    SysResetFPU;
    asm
      xorl %eax,%eax
      cpuid
      movl %eax,_eax
      movl $1,%eax
      xorl %ecx,%ecx
      cpuid
      movl %ecx,cpuid1_ecx
    end ['eax', 'ebx', 'ecx', 'edx'];
    has_sse41_support:=boolean(cpuid1_ecx shr 19 and 1);
    if _eax>=7 then
      begin
        asm
          movl $7,%eax
          xorl %ecx,%ecx
          cpuid
          movl %ebx,cpuid7_ebx
        end ['eax', 'ebx', 'ecx', 'edx'];
{$ifdef use_fast_repmovstos}
        fast_large_repmovstosb:=cpuid7_ebx and (1 shl 9)<>0;
{$endif}
        { XGETBV support? }
        if (cpuid1_ecx and $8000000)<>0 then
          begin
            asm
              xorl %ecx,%ecx
              .byte   0x0f,0x01,0xd0 { xgetbv }
              movl %eax,_eax
            end ['eax', 'rcx', 'edx'];
            if (_eax and 6)=6 then
              begin
                has_avx_support:=(cpuid1_ecx and $10000000)<>0;
                has_avx2_support:=(cpuid7_ebx and $20)<>0;
              end;
          end;
      end;
    fpc_cpuinit_performed:=true;
  end;

{$define FPC_SYSTEM_HAS_SYSINITFPU}
Procedure SysInitFPU;
begin
end;


{$define FPC_SYSTEM_HAS_SYSRESETFPU}
Procedure SysResetFPU;assembler;nostackframe;
  asm
    fninit
    fwait
{$ifdef FPC_PIC}
    movq  Default8087CW@GOTPCREL(%rip),%rax
    fldcw (%rax)
    movq  DefaultMXCSR@GOTPCREL(%rip),%rax
    ldmxcsr (%rax)
{$else FPC_PIC}
    fldcw   Default8087CW(%rip)
    ldmxcsr DefaultMXCSR(%rip)
{$endif FPC_PIC}
  end;


{$ifndef FPC_SYSTEM_HAS_MEM_BARRIER}
{$define FPC_SYSTEM_HAS_MEM_BARRIER}

procedure ReadBarrier;assembler;nostackframe;
asm
  lfence
end;

procedure ReadDependencyBarrier;assembler;nostackframe;
asm
  { reads imply barrier on earlier reads depended on }
end;

procedure ReadWriteBarrier;assembler;nostackframe;
asm
  mfence
end;

procedure WriteBarrier;assembler;nostackframe;
asm
  sfence
end;

{$endif}

{****************************************************************************
                          Math Routines
****************************************************************************}

{$define FPC_SYSTEM_HAS_SWAPENDIAN}

{ SwapEndian(<16 Bit>) being inlined is faster than using assembler }
function SwapEndian(const AValue: SmallInt): SmallInt;{$ifdef SYSTEMINLINE}inline;{$endif}
  begin
    { the extra Word type cast is necessary because the "AValue shr 8" }
    { is turned into "longint(AValue) shr 8", so if AValue < 0 then    }
    { the sign bits from the upper 16 bits are shifted in rather than  }
    { zeroes.                                                          }
    Result := SmallInt(((Word(AValue) shr 8) or (Word(AValue) shl 8)) and $ffff);
  end;


function SwapEndian(const AValue: Word): Word;{$ifdef SYSTEMINLINE}inline;{$endif}
  begin
    Result := ((AValue shr 8) or (AValue shl 8)) and $ffff;
  end;


function SwapEndian(const AValue: LongInt): LongInt; assembler; nostackframe;
asm
{$ifdef win64}
  movl %ecx, %eax
{$else win64}
  movl %edi, %eax
{$endif win64}
  bswap %eax
end;


function SwapEndian(const AValue: DWord): DWord; assembler; nostackframe;
asm
{$ifdef win64}
  movl %ecx, %eax
{$else win64}
  movl %edi, %eax
{$endif win64}
  bswap %eax
end;


function SwapEndian(const AValue: Int64): Int64; assembler; nostackframe;
asm
{$ifdef win64}
  movq %rcx, %rax
{$else win64}
  movq %rdi, %rax
{$endif win64}
  bswap %rax
end;


function SwapEndian(const AValue: QWord): QWord; assembler; nostackframe;
asm
{$ifdef win64}
  movq %rcx, %rax
{$else win64}
  movq %rdi, %rax
{$endif win64}
  bswap %rax
end;


{$ifndef win64}
{$define FPC_SYSTEM_HAS_U128_DIV_U64_TO_U64}
function u128_div_u64_to_u64( const xh, xl: qword; const y: qword; out quotient, remainder: qword ): boolean;nostackframe;assembler;
{
  SysV:
  xh: RDI
  xl: RSI
  y: RDX
  quotient: RCX
  remainder: R8
}
label
  dodiv;
asm
  cmpq %rdi,%rdx
  ja   dodiv
  xorl %eax,%eax
  ret
dodiv:
  movq %rdx,%r9
  movq %rsi,%rax
  movq %rdi,%rdx
  divq %r9
  movq %rax,(%rcx)
  movq %rdx,(%r8)
  movl $1,%eax
end;
{$endif win64}

{$ifndef FPC_SYSTEM_HAS_UMUL64X64_128}
{$define FPC_SYSTEM_HAS_UMUL64X64_128}
function UMul64x64_128(a,b: uint64; out rHi: uint64): uint64; assembler; nostackframe;
{ Win64: rcx = a, rdx = b, r8 = rHi.
  SysV:  rdi = a, rsi = b, rdx = rHi. }
asm
{$ifndef win64}
  mov %rdx, %rcx { rcx = rHi, as rdx is used for mul. }
{$endif}
  mov a, %rax
  mul b
  mov %rdx, {$ifdef win64} (%r8) {$else} (%rcx) {$endif}
end;
{$endif FPC_SYSTEM_HAS_UMUL64X64_128}