crypto/bn/asm/x86nt32.asm

   1         TITLE   bn_mulw.c
   2         .386P
   3 .model FLAT
   4 PUBLIC  _bn_mul_add_word
   5 _TEXT   SEGMENT
   6 ; File bn_mulw.c
   7 _bn_mul_add_word PROC NEAR
   8         push    ebp
   9         push    ebx
  10         push    esi
  11         push    edi
  12         mov     edi,DWORD PTR 20[esp]   ; r
  13         mov     ebx,DWORD PTR 24[esp]   ; a
  14         mov     ecx,DWORD PTR 32[esp]   ; w
  15         xor     esi,esi                 ; c=0
  16
  17         mov     ebp,DWORD PTR 28[esp]   ; num
  18         shr     ebp,2                   ; num/4
  19         jz      $L666
  20
  21 $L546:
  22         ; Round one
  23         mov     eax,DWORD PTR [ebx]     ; edx:eax = *a * w
  24         mul     ecx
  25         add     eax,DWORD PTR [edi]     ; *r+=ax
  26         adc     edx,0
  27         add     eax,esi                 ; edx:eax += c
  28         adc     edx,0
  29         mov     DWORD PTR [edi],eax     ; *r+=ax
  30         mov     esi,edx                 ; c = overflow
  31
  32         ; Round two
  33         mov     eax,DWORD PTR 4[ebx]    ; edx:eax = *a * w
  34         mul     ecx
  35         add     eax,DWORD PTR 4[edi]    ; *r+=ax
  36         adc     edx,0
  37         add     eax,esi                 ; edx:eax += c
  38         adc     edx,0
  39         mov     DWORD PTR 4[edi],eax    ; *r+=ax
  40         mov     esi,edx                 ; c = overflow
  41
  42         ; Round three
  43         mov     eax,DWORD PTR 8[ebx]    ; edx:eax = *a * w
  44         mul     ecx
  45         add     eax,DWORD PTR 8[edi]    ; *r+=ax
  46         adc     edx,0
  47         add     eax,esi                 ; edx:eax += c
  48         adc     edx,0
  49         mov     DWORD PTR 8[edi],eax    ; *r+=ax
  50         mov     esi,edx                 ; c = overflow
  51
  52         ; Round four
  53         mov     eax,DWORD PTR 12[ebx]   ; edx:eax = *a * w
  54         mul     ecx
  55         add     eax,DWORD PTR 12[edi]   ; *r+=ax
  56         adc     edx,0
  57         add     eax,esi                 ; edx:eax += c
  58         adc     edx,0
  59         mov     DWORD PTR 12[edi],eax   ; *r+=ax
  60         mov     esi,edx                 ; c = overflow
  61
  62         add     ebx,16
  63         add     edi,16
  64
  65         dec     ebp
  66         jz      $L666
  67         jmp     $L546
  68 $L666:
  69         mov     ebp,DWORD PTR 28[esp]   ; num
  70         and     ebp,3                   ; num%4
  71         jz      $L547
  72
  73         ; Round one
  74         mov     eax,DWORD PTR [ebx]     ; edx:eax = *a * w
  75         mul     ecx
  76         add     eax,DWORD PTR [edi]     ; *r+=ax
  77         adc     edx,0
  78         add     eax,esi                 ; edx:eax += c
  79         adc     edx,0
  80         mov     DWORD PTR [edi],eax     ; *r+=ax
  81         mov     esi,edx                 ; c = overflow
  82         dec     ebp
  83         jz      $L547
  84         ; Round two
  85         mov     eax,DWORD PTR 4[ebx]    ; edx:eax = *a * w
  86         mul     ecx
  87         add     eax,DWORD PTR 4[edi]    ; *r+=ax
  88         adc     edx,0
  89         add     eax,esi                 ; edx:eax += c
  90         adc     edx,0
  91         mov     DWORD PTR 4[edi],eax    ; *r+=ax
  92         mov     esi,edx                 ; c = overflow
  93         dec     ebp
  94         jz      $L547
  95         ; Round three
  96         mov     eax,DWORD PTR 8[ebx]    ; edx:eax = *a * w
  97         mul     ecx
  98         add     eax,DWORD PTR 8[edi]    ; *r+=ax
  99         adc     edx,0
 100         add     eax,esi                 ; edx:eax += c
 101         adc     edx,0
 102         mov     DWORD PTR 8[edi],eax    ; *r+=ax
 103         mov     esi,edx                 ; c = overflow
 104
 105 $L547:
 106         mov     eax,esi
 107         pop     edi
 108         pop     esi
 109         pop     ebx
 110         pop     ebp
 111         ret
 112 _bn_mul_add_word ENDP
 113 _TEXT   ENDS
 114 PUBLIC  _bn_mul_word
 115 _TEXT   SEGMENT
 116 _bn_mul_word PROC NEAR
 117         push    ebp
 118         push    ebx
 119         push    esi
 120         push    edi
 121
 122         mov     edi,DWORD PTR 20[esp]   ; r
 123         mov     ebx,DWORD PTR 24[esp]   ; a
 124         mov     ebp,DWORD PTR 28[esp]   ; num
 125         mov     ecx,DWORD PTR 32[esp]   ; w
 126         xor     esi,esi                 ; c=0
 127
 128         shr     ebp,2                   ; num/4
 129         jz      $L266
 130
 131 $L593:
 132         ; Round one
 133         mov     eax,DWORD PTR [ebx]     ; edx:eax= w * *a
 134         mul     ecx
 135         add     eax,esi                 ; edx:eax+=c
 136         adc     edx,0
 137         mov     DWORD PTR [edi],eax     ; *r=eax
 138         mov     esi,edx                 ; c=edx
 139         ; Round two
 140         mov     eax,DWORD PTR 4[ebx]    ; edx:eax= w * *a
 141         mul     ecx
 142         add     eax,esi                 ; edx:eax+=c
 143         adc     edx,0
 144         mov     DWORD PTR 4[edi],eax    ; *r=eax
 145         mov     esi,edx                 ; c=edx
 146         ; Round three
 147         mov     eax,DWORD PTR 8[ebx]    ; edx:eax= w * *a
 148         mul     ecx
 149         add     eax,esi                 ; edx:eax+=c
 150         adc     edx,0
 151         mov     DWORD PTR 8[edi],eax    ; *r=eax
 152         mov     esi,edx                 ; c=edx
 153         ; Round four
 154         mov     eax,DWORD PTR 12[ebx]   ; edx:eax= w * *a
 155         mul     ecx
 156         add     eax,esi                 ; edx:eax+=c
 157         adc     edx,0
 158         mov     DWORD PTR 12[edi],eax   ; *r=eax
 159         mov     esi,edx                 ; c=edx
 160
 161         add     ebx,16
 162         add     edi,16
 163
 164         dec     ebp
 165         jz      $L266
 166         jmp     $L593
 167 $L266:
 168         mov     ebp,DWORD PTR 28[esp]   ; num
 169         and     ebp,3
 170         jz      $L601
 171
 172         ; Round one
 173         mov     eax,DWORD PTR [ebx]     ; edx:eax= w * *a
 174         mul     ecx
 175         add     eax,esi                 ; edx:eax+=c
 176         adc     edx,0
 177         mov     DWORD PTR [edi],eax     ; *r=eax
 178         mov     esi,edx                 ; c=edx
 179         dec     ebp
 180         jz      $L601
 181         ; Round two
 182         mov     eax,DWORD PTR 4[ebx]    ; edx:eax= w * *a
 183         mul     ecx
 184         add     eax,esi                 ; edx:eax+=c
 185         adc     edx,0
 186         mov     DWORD PTR 4[edi],eax    ; *r=eax
 187         mov     esi,edx                 ; c=edx
 188         dec     ebp
 189         jz      $L601
 190         ; Round three
 191         mov     eax,DWORD PTR 8[ebx]    ; edx:eax= w * *a
 192         mul     ecx
 193         add     eax,esi                 ; edx:eax+=c
 194         adc     edx,0
 195         mov     DWORD PTR 8[edi],eax    ; *r=eax
 196         mov     esi,edx                 ; c=edx
 197
 198 $L601:
 199         mov     eax,esi
 200         pop     edi
 201         pop     esi
 202         pop     ebx
 203         pop     ebp
 204         ret
 205 _bn_mul_word ENDP
 206 _TEXT   ENDS
 207 PUBLIC  _bn_sqr_words
 208 _TEXT   SEGMENT
 209 _bn_sqr_words PROC NEAR
 210         push    ebx
 211         push    esi
 212         push    edi
 213         mov     esi,DWORD PTR 16[esp]   ; r
 214         mov     edi,DWORD PTR 20[esp]   ; a
 215         mov     ebx,DWORD PTR 24[esp]   ; num
 216
 217         shr     ebx,2                   ; num/4
 218         jz      $L111
 219 $L640:
 220         ; Round 1
 221         mov     eax, DWORD PTR [edi]
 222         mul     eax                     ; *a * *a
 223         mov     DWORD PTR [esi],eax
 224         mov     DWORD PTR 4[esi],edx
 225         ; Round 2
 226         mov     eax, DWORD PTR 4[edi]
 227         mul     eax                     ; *a * *a
 228         mov     DWORD PTR 8[esi],eax
 229         mov     DWORD PTR 12[esi],edx
 230         ; Round 3
 231         mov     eax, DWORD PTR 8[edi]
 232         mul     eax                     ; *a * *a
 233         mov     DWORD PTR 16[esi],eax
 234         mov     DWORD PTR 20[esi],edx
 235         ; Round 4
 236         mov     eax, DWORD PTR 12[edi]
 237         mul     eax                     ; *a * *a
 238         mov     DWORD PTR 24[esi],eax
 239         mov     DWORD PTR 28[esi],edx
 240
 241         add     edi,16
 242         add     esi,32
 243
 244         dec     ebx
 245         jz      $L111
 246         jmp     $L640
 247 $L111:
 248         mov     ebx,DWORD PTR 24[esp]   ; num
 249         and     ebx,3                   ; num%3
 250         jz      $L645
 251
 252         ; Round 1
 253         mov     eax, DWORD PTR [edi]
 254         mul     eax                     ; *a * *a
 255         mov     DWORD PTR [esi],eax
 256         mov     DWORD PTR 4[esi],edx
 257         dec     ebx
 258         jz      $L645
 259         ; Round 2
 260         mov     eax, DWORD PTR 4[edi]
 261         mul     eax                     ; *a * *a
 262         mov     DWORD PTR 8[esi],eax
 263         mov     DWORD PTR 12[esi],edx
 264         dec     ebx
 265         jz      $L645
 266         ; Round 3
 267         mov     eax, DWORD PTR 8[edi]
 268         mul     eax                     ; *a * *a
 269         mov     DWORD PTR 16[esi],eax
 270         mov     DWORD PTR 20[esi],edx
 271
 272 $L645:
 273         pop     edi
 274         pop     esi
 275         pop     ebx
 276         ret
 277 _bn_sqr_words ENDP
 278 _TEXT   ENDS
 279 PUBLIC  _bn_div64
 280 _TEXT   SEGMENT
 281 _bn_div64 PROC NEAR
 282         mov     edx, DWORD PTR 4[esp]
 283         mov     eax, DWORD PTR 8[esp]
 284         div     DWORD PTR 12[esp]
 285         ret
 286 _bn_div64 ENDP
 287 _TEXT   ENDS
 288 END