crypto/bn/asm/x86-bsdi.s

   1         .file   "bn_mulw.c"
   2         .version        "01.01"
   3 gcc2_compiled.:
   4 .text
   5         .align 4
   6 .globl _bn_mul_add_word
   7 _bn_mul_add_word:
   8         pushl %ebp
   9         pushl %edi
  10         pushl %esi
  11         pushl %ebx
  12
  13         # ax            L(t)
  14         # dx            H(t)
  15         # bx            a
  16         # cx            w
  17         # di            r
  18         # si            c
  19         # bp            num
  20         xorl %esi,%esi          # c=0
  21         movl 20(%esp),%edi      # r => edi
  22         movl 24(%esp),%ebx      # a => exb
  23         movl 32(%esp),%ecx      # w => ecx
  24         movl 28(%esp),%ebp      # num => ebp
  25
  26         shrl $2,%ebp            # num/4
  27         je .L910
  28
  29 #       .align 4
  30 .L110:
  31         # Round 1
  32         movl %ecx,%eax          # w => eax
  33         mull (%ebx)             # w * *a
  34         addl (%edi),%eax        # *r+=L(t)
  35         adcl $0,%edx            # H(t)+= carry
  36         addl %esi,%eax          # L(t)+=c
  37         adcl $0,%edx            # H(t)+=carry
  38         movl %eax,(%edi)        # *r+=L(t)
  39         movl %edx,%esi          # c=H(t)
  40
  41         # Round 2
  42         movl %ecx,%eax          # w => eax
  43         mull 4(%ebx)            # w * *a
  44         addl 4(%edi),%eax       # *r+=L(t)
  45         adcl $0,%edx            # H(t)+= carry
  46         addl %esi,%eax          # L(t)+=c
  47         adcl $0,%edx            # H(t)+=carry
  48         movl %eax,4(%edi)       # *r+=L(t)
  49         movl %edx,%esi          # c=H(t)
  50
  51         # Round 3
  52         movl %ecx,%eax          # w => eax
  53         mull 8(%ebx)            # w * *a
  54         addl 8(%edi),%eax       # *r+=L(t)
  55         adcl $0,%edx            # H(t)+=carry
  56         addl %esi,%eax          # L(t)+=c
  57         adcl $0,%edx            # H(t)+=carry
  58         movl %eax,8(%edi)       # *r+=L(t)
  59         movl %edx,%esi          # c=H(t)
  60
  61         # Round 4
  62         movl %ecx,%eax          # w => eax
  63         mull 12(%ebx)           # w * *a
  64         addl 12(%edi),%eax      # *r+=L(t)
  65         adcl $0,%edx            # H(t)+=carry
  66         addl %esi,%eax          # L(t)+=c
  67         adcl $0,%edx            # H(t)+=carry
  68         movl %eax,12(%edi)      # *r+=L(t)
  69         movl %edx,%esi          # c=H(t)
  70
  71         addl $16,%ebx           # a+=4 (4 words)
  72         addl $16,%edi           # r+=4 (4 words)
  73
  74         decl %ebp               # --num
  75         je .L910
  76         jmp .L110
  77 #       .align 4
  78 .L910:
  79         movl 28(%esp),%ebp      # num => ebp
  80         andl $3,%ebp
  81         je .L111
  82
  83         # Round 1
  84         movl %ecx,%eax          # w => eax
  85         mull (%ebx)             # w * *a
  86         addl (%edi),%eax        # *r+=L(t)
  87         adcl $0,%edx            # H(t)+=carry
  88         addl %esi,%eax          # L(t)+=c
  89         adcl $0,%edx            # H(t)+=carry
  90         movl %eax,(%edi)        # *r+=L(t)
  91         movl %edx,%esi          # c=H(t)
  92         decl %ebp               # --num
  93         je .L111
  94
  95         # Round 2
  96         movl %ecx,%eax          # w => eax
  97         mull 4(%ebx)            # w * *a
  98         addl 4(%edi),%eax       # *r+=L(t)
  99         adcl $0,%edx            # H(t)+=carry
 100         addl %esi,%eax          # L(t)+=c
 101         adcl $0,%edx            # H(t)+=carry
 102         movl %eax,4(%edi)       # *r+=L(t)
 103         movl %edx,%esi          # c=H(t)
 104         decl %ebp               # --num
 105         je .L111
 106
 107         # Round 3
 108         movl %ecx,%eax          # w => eax
 109         mull 8(%ebx)            # w * *a
 110         addl 8(%edi),%eax       # *r+=L(t)
 111         adcl $0,%edx            # H(t)+=carry
 112         addl %esi,%eax          # L(t)+=c
 113         adcl $0,%edx            # H(t)+=carry
 114         movl %eax,8(%edi)       # *r+=L(t)
 115         movl %edx,%esi          # c=H(t)
 116
 117 #       .align 4
 118 .L111:
 119         movl %esi,%eax          # return(c)
 120         popl %ebx
 121         popl %esi
 122         popl %edi
 123         popl %ebp
 124         ret
 125 .Lfe1:
 126         .align 4
 127 .globl _bn_mul_word
 128 _bn_mul_word:
 129         pushl %ebp
 130         pushl %edi
 131         pushl %esi
 132         pushl %ebx
 133
 134         # ax            L(t)
 135         # dx            H(t)
 136         # bx            a
 137         # cx            w
 138         # di            r
 139         # num           bp
 140         # si            c
 141         xorl %esi,%esi          # c=0
 142         movl 20(%esp),%edi      # r => edi
 143         movl 24(%esp),%ebx      # a => exb
 144         movl 28(%esp),%ebp      # num => bp
 145         movl 32(%esp),%ecx      # w => ecx
 146
 147 #       .align 4
 148 .L210:
 149         movl %ecx,%eax          # w => eax
 150         mull (%ebx)             # w * *a
 151         addl %esi,%eax          # L(t)+=c
 152         adcl $0,%edx            # H(t)+=carry
 153         movl %eax,(%edi)        # *r=L(t)
 154         movl %edx,%esi          # c=H(t)
 155         decl %ebp               # --num
 156         je .L211
 157
 158         movl %ecx,%eax          # w => eax
 159         mull 4(%ebx)            # w * *a
 160         addl %esi,%eax          # L(t)+=c
 161         adcl $0,%edx            # H(t)+=carry
 162         movl %eax,4(%edi)       # *r=L(t)
 163         movl %edx,%esi          # c=H(t)
 164         decl %ebp               # --num
 165         je .L211
 166
 167         movl %ecx,%eax          # w => eax
 168         mull 8(%ebx)            # w * *a
 169         addl %esi,%eax          # L(t)+=c
 170         adcl $0,%edx            # H(t)+=carry
 171         movl %eax,8(%edi)       # *r=L(t)
 172         movl %edx,%esi          # c=H(t)
 173         decl %ebp               # --num
 174         je .L211
 175
 176         movl %ecx,%eax          # w => eax
 177         mull 12(%ebx)           # w * *a
 178         addl %esi,%eax          # L(t)+=c
 179         adcl $0,%edx            # H(t)+=carry
 180         movl %eax,12(%edi)      # *r=L(t)
 181         movl %edx,%esi          # c=H(t)
 182         decl %ebp               # --num
 183         je .L211
 184
 185         addl $16,%ebx           # a+=4 (4 words)
 186         addl $16,%edi           # r+=4 (4 words)
 187
 188         jmp .L210
 189 #       .align 4
 190 .L211:
 191         movl %esi,%eax          # return(c)
 192         popl %ebx
 193         popl %esi
 194         popl %edi
 195         popl %ebp
 196         ret
 197 .Lfe2:
 198         .align 4
 199 .globl _bn_sqr_words
 200 _bn_sqr_words:
 201         pushl %edi
 202         pushl %esi
 203         pushl %ebx
 204         movl 16(%esp),%esi      # r
 205         movl 20(%esp),%edi      # a
 206         movl 24(%esp),%ebx      # n
 207 #       .align 4
 208         shrl $2,%ebx
 209         jz .L99
 210 .L28:
 211         movl (%edi),%eax        # get a
 212         mull %eax               # a*a
 213         movl %eax,(%esi)        # put low into return addr
 214         movl %edx,4(%esi)       # put high into return addr
 215
 216         movl 4(%edi),%eax       # get a
 217         mull %eax               # a*a
 218         movl %eax,8(%esi)       # put low into return addr
 219         movl %edx,12(%esi)      # put high into return addr
 220
 221         movl 8(%edi),%eax       # get a
 222         mull %eax               # a*a
 223         movl %eax,16(%esi)      # put low into return addr
 224         movl %edx,20(%esi)      # put high into return addr
 225
 226         movl 12(%edi),%eax      # get a
 227         mull %eax               # a*a
 228         movl %eax,24(%esi)      # put low into return addr
 229         movl %edx,28(%esi)      # put high into return addr
 230
 231         addl $16,%edi
 232         addl $32,%esi
 233         decl %ebx               # n-=4;
 234         jz .L99
 235         jmp .L28
 236 #       .align 4
 237 .L99:
 238         movl 24(%esp),%ebx      # n
 239         andl $3,%ebx
 240         jz .L29
 241         movl (%edi),%eax        # get a
 242         mull %eax               # a*a
 243         movl %eax,(%esi)        # put low into return addr
 244         movl %edx,4(%esi)       # put high into return addr
 245         decl %ebx               # n--;
 246         jz .L29
 247         movl 4(%edi),%eax       # get a
 248         mull %eax               # a*a
 249         movl %eax,8(%esi)       # put low into return addr
 250         movl %edx,12(%esi)      # put high into return addr
 251         decl %ebx               # n--;
 252         jz .L29
 253         movl 8(%edi),%eax       # get a
 254         mull %eax               # a*a
 255         movl %eax,16(%esi)      # put low into return addr
 256         movl %edx,20(%esi)      # put high into return addr
 257
 258 .L29:
 259         popl %ebx
 260         popl %esi
 261         popl %edi
 262         ret
 263 .Lfe3:
 264         .align 4
 265 .globl _bn_div64
 266 _bn_div64:
 267         movl 4(%esp),%edx       # a
 268         movl 8(%esp),%eax       # b
 269         divl 12(%esp)           # ab/c
 270         ret
 271 .Lfe4:
 272         .ident  "GCC: (GNU) 2.6.3"