crypto/bn/asm/x86-lnx.s

   1         .file   "bn_mulw.c"
   2         .version        "01.01"
   3 gcc2_compiled.:
   4 .text
   5         .align 16
   6 .globl bn_mul_add_word
   7         .type    bn_mul_add_word,@function
   8 bn_mul_add_word:
   9         pushl %ebp
  10         pushl %edi
  11         pushl %esi
  12         pushl %ebx
  13
  14         # ax            L(t)
  15         # dx            H(t)
  16         # bx            a
  17         # cx            w
  18         # di            r
  19         # si            c
  20         # bp            num
  21         xorl %esi,%esi          # c=0
  22         movl 20(%esp),%edi      # r => edi
  23         movl 24(%esp),%ebx      # a => exb
  24         movl 32(%esp),%ecx      # w => ecx
  25         movl 28(%esp),%ebp      # num => ebp
  26
  27         shrl $2,%ebp            # num/4
  28         je .L910
  29
  30         .align 4
  31 .L110:
  32         # Round 1
  33         movl %ecx,%eax          # w => eax
  34         mull (%ebx)             # w * *a
  35         addl (%edi),%eax        # *r+=L(t)
  36         adcl $0,%edx            # H(t)+= carry
  37         addl %esi,%eax          # L(t)+=c
  38         adcl $0,%edx            # H(t)+=carry
  39         movl %eax,(%edi)        # *r+=L(t)
  40         movl %edx,%esi          # c=H(t)
  41
  42         # Round 2
  43         movl %ecx,%eax          # w => eax
  44         mull 4(%ebx)            # w * *a
  45         addl 4(%edi),%eax       # *r+=L(t)
  46         adcl $0,%edx            # H(t)+= carry
  47         addl %esi,%eax          # L(t)+=c
  48         adcl $0,%edx            # H(t)+=carry
  49         movl %eax,4(%edi)       # *r+=L(t)
  50         movl %edx,%esi          # c=H(t)
  51
  52         # Round 3
  53         movl %ecx,%eax          # w => eax
  54         mull 8(%ebx)            # w * *a
  55         addl 8(%edi),%eax       # *r+=L(t)
  56         adcl $0,%edx            # H(t)+=carry
  57         addl %esi,%eax          # L(t)+=c
  58         adcl $0,%edx            # H(t)+=carry
  59         movl %eax,8(%edi)       # *r+=L(t)
  60         movl %edx,%esi          # c=H(t)
  61
  62         # Round 4
  63         movl %ecx,%eax          # w => eax
  64         mull 12(%ebx)           # w * *a
  65         addl 12(%edi),%eax      # *r+=L(t)
  66         adcl $0,%edx            # H(t)+=carry
  67         addl %esi,%eax          # L(t)+=c
  68         adcl $0,%edx            # H(t)+=carry
  69         movl %eax,12(%edi)      # *r+=L(t)
  70         movl %edx,%esi          # c=H(t)
  71
  72         addl $16,%ebx           # a+=4 (4 words)
  73         addl $16,%edi           # r+=4 (4 words)
  74
  75         decl %ebp               # --num
  76         je .L910
  77         jmp .L110
  78         .align 4
  79 .L910:
  80         movl 28(%esp),%ebp      # num => ebp
  81         andl $3,%ebp
  82         je .L111
  83
  84         # Round 1
  85         movl %ecx,%eax          # w => eax
  86         mull (%ebx)             # w * *a
  87         addl (%edi),%eax        # *r+=L(t)
  88         adcl $0,%edx            # H(t)+=carry
  89         addl %esi,%eax          # L(t)+=c
  90         adcl $0,%edx            # H(t)+=carry
  91         movl %eax,(%edi)        # *r+=L(t)
  92         movl %edx,%esi          # c=H(t)
  93         decl %ebp               # --num
  94         je .L111
  95
  96         # Round 2
  97         movl %ecx,%eax          # w => eax
  98         mull 4(%ebx)            # w * *a
  99         addl 4(%edi),%eax       # *r+=L(t)
 100         adcl $0,%edx            # H(t)+=carry
 101         addl %esi,%eax          # L(t)+=c
 102         adcl $0,%edx            # H(t)+=carry
 103         movl %eax,4(%edi)       # *r+=L(t)
 104         movl %edx,%esi          # c=H(t)
 105         decl %ebp               # --num
 106         je .L111
 107
 108         # Round 3
 109         movl %ecx,%eax          # w => eax
 110         mull 8(%ebx)            # w * *a
 111         addl 8(%edi),%eax       # *r+=L(t)
 112         adcl $0,%edx            # H(t)+=carry
 113         addl %esi,%eax          # L(t)+=c
 114         adcl $0,%edx            # H(t)+=carry
 115         movl %eax,8(%edi)       # *r+=L(t)
 116         movl %edx,%esi          # c=H(t)
 117
 118         .align 4
 119 .L111:
 120         movl %esi,%eax          # return(c)
 121         popl %ebx
 122         popl %esi
 123         popl %edi
 124         popl %ebp
 125         ret
 126 .Lfe1:
 127         .size    bn_mul_add_word,.Lfe1-bn_mul_add_word
 128         .align 16
 129 .globl bn_mul_word
 130         .type    bn_mul_word,@function
 131 bn_mul_word:
 132         pushl %ebp
 133         pushl %edi
 134         pushl %esi
 135         pushl %ebx
 136
 137         # ax            L(t)
 138         # dx            H(t)
 139         # bx            a
 140         # cx            w
 141         # di            r
 142         # num           bp
 143         # si            c
 144         xorl %esi,%esi          # c=0
 145         movl 20(%esp),%edi      # r => edi
 146         movl 24(%esp),%ebx      # a => exb
 147         movl 28(%esp),%ebp      # num => bp
 148         movl 32(%esp),%ecx      # w => ecx
 149
 150         .align 4
 151 .L210:
 152         movl %ecx,%eax          # w => eax
 153         mull (%ebx)             # w * *a
 154         addl %esi,%eax          # L(t)+=c
 155         adcl $0,%edx            # H(t)+=carry
 156         movl %eax,(%edi)        # *r=L(t)
 157         movl %edx,%esi          # c=H(t)
 158         decl %ebp               # --num
 159         je .L211
 160
 161         movl %ecx,%eax          # w => eax
 162         mull 4(%ebx)            # w * *a
 163         addl %esi,%eax          # L(t)+=c
 164         adcl $0,%edx            # H(t)+=carry
 165         movl %eax,4(%edi)       # *r=L(t)
 166         movl %edx,%esi          # c=H(t)
 167         decl %ebp               # --num
 168         je .L211
 169
 170         movl %ecx,%eax          # w => eax
 171         mull 8(%ebx)            # w * *a
 172         addl %esi,%eax          # L(t)+=c
 173         adcl $0,%edx            # H(t)+=carry
 174         movl %eax,8(%edi)       # *r=L(t)
 175         movl %edx,%esi          # c=H(t)
 176         decl %ebp               # --num
 177         je .L211
 178
 179         movl %ecx,%eax          # w => eax
 180         mull 12(%ebx)           # w * *a
 181         addl %esi,%eax          # L(t)+=c
 182         adcl $0,%edx            # H(t)+=carry
 183         movl %eax,12(%edi)      # *r=L(t)
 184         movl %edx,%esi          # c=H(t)
 185         decl %ebp               # --num
 186         je .L211
 187
 188         addl $16,%ebx           # a+=4 (4 words)
 189         addl $16,%edi           # r+=4 (4 words)
 190
 191         jmp .L210
 192         .align 16
 193 .L211:
 194         movl %esi,%eax          # return(c)
 195         popl %ebx
 196         popl %esi
 197         popl %edi
 198         popl %ebp
 199         ret
 200 .Lfe2:
 201         .size    bn_mul_word,.Lfe2-bn_mul_word
 202
 203         .align 16
 204 .globl bn_sqr_words
 205         .type    bn_sqr_words,@function
 206 bn_sqr_words:
 207         pushl %edi
 208         pushl %esi
 209         pushl %ebx
 210         movl 16(%esp),%esi      # r
 211         movl 20(%esp),%edi      # a
 212         movl 24(%esp),%ebx      # n
 213         .align 4
 214         shrl $2,%ebx
 215         jz .L99
 216 .L28:
 217         movl (%edi),%eax        # get a
 218         mull %eax               # a*a
 219         movl %eax,(%esi)        # put low into return addr
 220         movl %edx,4(%esi)       # put high into return addr
 221
 222         movl 4(%edi),%eax       # get a
 223         mull %eax               # a*a
 224         movl %eax,8(%esi)       # put low into return addr
 225         movl %edx,12(%esi)      # put high into return addr
 226
 227         movl 8(%edi),%eax       # get a
 228         mull %eax               # a*a
 229         movl %eax,16(%esi)      # put low into return addr
 230         movl %edx,20(%esi)      # put high into return addr
 231
 232         movl 12(%edi),%eax      # get a
 233         mull %eax               # a*a
 234         movl %eax,24(%esi)      # put low into return addr
 235         movl %edx,28(%esi)      # put high into return addr
 236
 237         addl $16,%edi
 238         addl $32,%esi
 239         decl %ebx               # n-=4;
 240         jz .L99
 241         jmp .L28
 242         .align 16
 243 .L99:
 244         movl 24(%esp),%ebx      # n
 245         andl $3,%ebx
 246         jz .L29
 247         movl (%edi),%eax        # get a
 248         mull %eax               # a*a
 249         movl %eax,(%esi)        # put low into return addr
 250         movl %edx,4(%esi)       # put high into return addr
 251         decl %ebx               # n--;
 252         jz .L29
 253         movl 4(%edi),%eax       # get a
 254         mull %eax               # a*a
 255         movl %eax,8(%esi)       # put low into return addr
 256         movl %edx,12(%esi)      # put high into return addr
 257         decl %ebx               # n--;
 258         jz .L29
 259         movl 8(%edi),%eax       # get a
 260         mull %eax               # a*a
 261         movl %eax,16(%esi)      # put low into return addr
 262         movl %edx,20(%esi)      # put high into return addr
 263
 264 .L29:
 265         popl %ebx
 266         popl %esi
 267         popl %edi
 268         ret
 269 .Lfe3:
 270         .size    bn_sqr_words,.Lfe3-bn_sqr_words
 271
 272         .align 16
 273 .globl bn_div64
 274         .type    bn_div64,@function
 275 bn_div64:
 276         movl 4(%esp),%edx       # a
 277         movl 8(%esp),%eax       # b
 278         divl 12(%esp)           # ab/c
 279         ret
 280 .Lfe4:
 281         .size    bn_div64,.Lfe4-bn_div64
 282         .ident  "GCC: (GNU) 2.6.3"