test/sysperf/memzero.S

   1
   2                 /*
   3                  * memcpy.S
   4                  *
   5                  * $DragonFly: src/test/sysperf/memzero.S,v 1.1 2004/04/29 16:14:53 dillon Exp $
   6                  */
   7                 .text
   8                 .globl  dozero1
   9                 .globl  dozero2
  10                 .globl  dozero3
  11                 .globl  dozero4
  12                 .globl  dozero5
  13                 .globl  dozero6
  14                 .globl  dozero7
  15                 .globl  fpcleanup
  16
  17                 .p2align 4,0x90
  18 dozero1:
  19                 pushl   %esi
  20                 pushl   %edi
  21                 pushl   %ecx
  22                 pushl   %ebx
  23
  24                 movl    4+16(%esp),%edi
  25                 movl    8+16(%esp),%ecx
  26                 shrl    $2,%ecx
  27                 subl    %eax,%eax
  28                 cld
  29                 rep
  30                 stosl
  31                 popl    %ebx
  32                 popl    %ecx
  33                 popl    %edi
  34                 popl    %esi
  35                 ret
  36
  37                 .p2align 4,0x90
  38 dozero2:
  39                 pushl   %esi
  40                 pushl   %edi
  41                 pushl   %ecx
  42                 pushl   %ebx
  43
  44                 movl    4+16(%esp),%edi
  45                 movl    8+16(%esp),%ecx
  46                 addl    %ecx,%esi
  47                 addl    %ecx,%edi
  48                 shrl    $2,%ecx
  49                 subl    %eax,%eax
  50                 std
  51                 rep
  52                 stosl
  53                 popl    %ebx
  54                 popl    %ecx
  55                 popl    %edi
  56                 popl    %esi
  57                 ret
  58
  59                 .p2align 4,0x90
  60 dozero3:
  61                 pushl   %esi
  62                 pushl   %edi
  63                 pushl   %ecx
  64                 pushl   %ebx
  65
  66                 movl    4+16(%esp),%edi
  67                 movl    8+16(%esp),%ecx
  68                 movl    $8,%edx
  69                 subl    %eax,%eax
  70                 .p2align 4,0x90
  71 1:
  72                 subl    %edx,%ecx
  73                 movl    %eax,(%edi)
  74                 movl    %eax,4(%edi)
  75                 addl    %edx,%edi
  76                 testl   %ecx,%ecx
  77                 jnz     1b
  78
  79                 popl    %ebx
  80                 popl    %ecx
  81                 popl    %edi
  82                 popl    %esi
  83                 ret
  84
  85                 .p2align 4,0x90
  86 dozero4:
  87                 pushl   %esi
  88                 pushl   %edi
  89                 pushl   %ecx
  90                 pushl   %ebx
  91
  92                 movl    4+16(%esp),%edi
  93                 movl    8+16(%esp),%ecx
  94                 subl    %eax,%eax
  95                 .p2align 4,0x90
  96 1:
  97                 subl    $16,%ecx
  98                 movnti  %eax,0(%edi)
  99                 movnti  %eax,4(%edi)
 100                 movnti  %eax,8(%edi)
 101                 movnti  %eax,12(%edi)
 102                 addl    $16,%edi
 103                 testl   %ecx,%ecx
 104                 jnz     1b
 105
 106                 popl    %ebx
 107                 popl    %ecx
 108                 popl    %edi
 109                 popl    %esi
 110                 ret
 111
 112                 .p2align 4,0x90
 113 dozero5:
 114                 pushl   %esi
 115                 pushl   %edi
 116                 pushl   %ecx
 117                 pushl   %ebx
 118
 119                 movl    4+16(%esp),%edi
 120                 movl    8+16(%esp),%ecx
 121
 122                 subl    $108,%esp
 123                 fnsave  0(%esp)
 124                 fninit
 125                 fldz
 126                 .p2align 4,0x90
 127 1:
 128                 subl    $64,%ecx
 129                 movq    %mm0,(%edi)
 130                 movq    %mm1,8(%edi)
 131                 movq    %mm2,16(%edi)
 132                 movq    %mm3,24(%edi)
 133                 movq    %mm4,32(%edi)
 134                 movq    %mm5,40(%edi)
 135                 movq    %mm6,48(%edi)
 136                 movq    %mm7,56(%edi)
 137                 addl    $64,%edi
 138                 testl   %ecx,%ecx
 139                 jnz     1b
 140
 141                 frstor  0(%esp)
 142                 addl    $108,%esp
 143
 144                 popl    %ebx
 145                 popl    %ecx
 146                 popl    %edi
 147                 popl    %esi
 148                 ret
 149
 150                 .p2align 4,0x90
 151 dozero6:
 152                 pushl   %esi
 153                 pushl   %edi
 154                 pushl   %ecx
 155                 pushl   %ebx
 156
 157                 movl    4+16(%esp),%edi
 158                 movl    8+16(%esp),%ecx
 159                 movl    $16,%eax
 160                 .p2align 4,0x90
 161 1:
 162                 subl    %eax,%ecx
 163                 movq    %mm0,(%edi)
 164                 movq    %mm1,8(%edi)
 165                 addl    %eax,%edi
 166                 testl   %ecx,%ecx
 167                 jnz     1b
 168
 169                 popl    %ebx
 170                 popl    %ecx
 171                 popl    %edi
 172                 popl    %esi
 173                 ret
 174
 175                 .p2align 4,0x90
 176 dozero7:
 177                 pushl   %esi
 178                 pushl   %edi
 179                 pushl   %ecx
 180                 pushl   %ebx
 181
 182                 movl    4+16(%esp),%edi
 183                 movl    8+16(%esp),%ecx
 184                 movl    $32,%eax
 185                 .p2align 4,0x90
 186 1:
 187                 subl    %eax,%ecx
 188                 movntdq %xmm0,(%edi)
 189                 movntdq %xmm1,16(%edi)
 190                 addl    %eax,%edi
 191                 testl   %ecx,%ecx
 192                 jnz     1b
 193                 sfence
 194
 195                 popl    %ebx
 196                 popl    %ecx
 197                 popl    %edi
 198                 popl    %esi
 199                 ret
 200
 201                 .p2align 4,0x90
 202 fpcleanup:
 203                 fninit
 204                 ret
 205