OpenPKG Specs / changeset

     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/cryptlib/bn-sol.s	Thu Jan 08 11:10:35 2009 +0100
     1.3 @@ -0,0 +1,2149 @@
     1.4 +	/ Don't even think of reading this code 
     1.5 +	/ It was automatically generated by x86.pl 
     1.6 +	/ Which is a perl program used to generate the x86 assember for 
     1.7 +	/ any of elf, a.out, BSDI, Win32, gaswin (for GNU as on Win32) or Solaris 
     1.8 +	/ eric <eay@cryptsoft.com> 
     1.9 +
    1.10 +	.file	"x86.s"
    1.11 +	.version	"01.01"
    1.12 +gcc2_compiled.:
    1.13 +.text
    1.14 +	.align 16
    1.15 +.globl bn_mul_add_words
    1.16 +	.type	bn_mul_add_words,@function
    1.17 +bn_mul_add_words:
    1.18 +	pushl	%ebp
    1.19 +	pushl	%ebx
    1.20 +	pushl	%esi
    1.21 +	pushl	%edi
    1.22 +
    1.23 +
    1.24 +	xorl	%esi,		%esi
    1.25 +	movl	20(%esp),	%edi
    1.26 +	movl	28(%esp),	%ecx
    1.27 +	movl	24(%esp),	%ebx
    1.28 +	andl	$4294967288,	%ecx
    1.29 +	movl	32(%esp),	%ebp
    1.30 +	pushl	%ecx
    1.31 +	jz	.L000maw_finish
    1.32 +.L001maw_loop:
    1.33 +	movl	%ecx,		(%esp)
    1.34 +	/ Round 0 
    1.35 +	movl	(%ebx),		%eax
    1.36 +	mull	%ebp
    1.37 +	addl	%esi,		%eax
    1.38 +	movl	(%edi),		%esi
    1.39 +	adcl	$0,		%edx
    1.40 +	addl	%esi,		%eax
    1.41 +	adcl	$0,		%edx
    1.42 +	movl	%eax,		(%edi)
    1.43 +	movl	%edx,		%esi
    1.44 +	/ Round 4 
    1.45 +	movl	4(%ebx),	%eax
    1.46 +	mull	%ebp
    1.47 +	addl	%esi,		%eax
    1.48 +	movl	4(%edi),	%esi
    1.49 +	adcl	$0,		%edx
    1.50 +	addl	%esi,		%eax
    1.51 +	adcl	$0,		%edx
    1.52 +	movl	%eax,		4(%edi)
    1.53 +	movl	%edx,		%esi
    1.54 +	/ Round 8 
    1.55 +	movl	8(%ebx),	%eax
    1.56 +	mull	%ebp
    1.57 +	addl	%esi,		%eax
    1.58 +	movl	8(%edi),	%esi
    1.59 +	adcl	$0,		%edx
    1.60 +	addl	%esi,		%eax
    1.61 +	adcl	$0,		%edx
    1.62 +	movl	%eax,		8(%edi)
    1.63 +	movl	%edx,		%esi
    1.64 +	/ Round 12 
    1.65 +	movl	12(%ebx),	%eax
    1.66 +	mull	%ebp
    1.67 +	addl	%esi,		%eax
    1.68 +	movl	12(%edi),	%esi
    1.69 +	adcl	$0,		%edx
    1.70 +	addl	%esi,		%eax
    1.71 +	adcl	$0,		%edx
    1.72 +	movl	%eax,		12(%edi)
    1.73 +	movl	%edx,		%esi
    1.74 +	/ Round 16 
    1.75 +	movl	16(%ebx),	%eax
    1.76 +	mull	%ebp
    1.77 +	addl	%esi,		%eax
    1.78 +	movl	16(%edi),	%esi
    1.79 +	adcl	$0,		%edx
    1.80 +	addl	%esi,		%eax
    1.81 +	adcl	$0,		%edx
    1.82 +	movl	%eax,		16(%edi)
    1.83 +	movl	%edx,		%esi
    1.84 +	/ Round 20 
    1.85 +	movl	20(%ebx),	%eax
    1.86 +	mull	%ebp
    1.87 +	addl	%esi,		%eax
    1.88 +	movl	20(%edi),	%esi
    1.89 +	adcl	$0,		%edx
    1.90 +	addl	%esi,		%eax
    1.91 +	adcl	$0,		%edx
    1.92 +	movl	%eax,		20(%edi)
    1.93 +	movl	%edx,		%esi
    1.94 +	/ Round 24 
    1.95 +	movl	24(%ebx),	%eax
    1.96 +	mull	%ebp
    1.97 +	addl	%esi,		%eax
    1.98 +	movl	24(%edi),	%esi
    1.99 +	adcl	$0,		%edx
   1.100 +	addl	%esi,		%eax
   1.101 +	adcl	$0,		%edx
   1.102 +	movl	%eax,		24(%edi)
   1.103 +	movl	%edx,		%esi
   1.104 +	/ Round 28 
   1.105 +	movl	28(%ebx),	%eax
   1.106 +	mull	%ebp
   1.107 +	addl	%esi,		%eax
   1.108 +	movl	28(%edi),	%esi
   1.109 +	adcl	$0,		%edx
   1.110 +	addl	%esi,		%eax
   1.111 +	adcl	$0,		%edx
   1.112 +	movl	%eax,		28(%edi)
   1.113 +	movl	%edx,		%esi
   1.114 +
   1.115 +	movl	(%esp),		%ecx
   1.116 +	addl	$32,		%ebx
   1.117 +	addl	$32,		%edi
   1.118 +	subl	$8,		%ecx
   1.119 +	jnz	.L001maw_loop
   1.120 +.L000maw_finish:
   1.121 +	movl	32(%esp),	%ecx
   1.122 +	andl	$7,		%ecx
   1.123 +	jnz	.L002maw_finish2
   1.124 +	jmp	.L003maw_end
   1.125 +.align 16
   1.126 +.L002maw_finish2:
   1.127 +	/ Tail Round 0 
   1.128 +	movl	(%ebx),		%eax
   1.129 +	mull	%ebp
   1.130 +	addl	%esi,		%eax
   1.131 +	movl	(%edi),		%esi
   1.132 +	adcl	$0,		%edx
   1.133 +	addl	%esi,		%eax
   1.134 +	adcl	$0,		%edx
   1.135 +	decl	%ecx
   1.136 +	movl	%eax,		(%edi)
   1.137 +	movl	%edx,		%esi
   1.138 +	jz	.L003maw_end
   1.139 +	/ Tail Round 1 
   1.140 +	movl	4(%ebx),	%eax
   1.141 +	mull	%ebp
   1.142 +	addl	%esi,		%eax
   1.143 +	movl	4(%edi),	%esi
   1.144 +	adcl	$0,		%edx
   1.145 +	addl	%esi,		%eax
   1.146 +	adcl	$0,		%edx
   1.147 +	decl	%ecx
   1.148 +	movl	%eax,		4(%edi)
   1.149 +	movl	%edx,		%esi
   1.150 +	jz	.L003maw_end
   1.151 +	/ Tail Round 2 
   1.152 +	movl	8(%ebx),	%eax
   1.153 +	mull	%ebp
   1.154 +	addl	%esi,		%eax
   1.155 +	movl	8(%edi),	%esi
   1.156 +	adcl	$0,		%edx
   1.157 +	addl	%esi,		%eax
   1.158 +	adcl	$0,		%edx
   1.159 +	decl	%ecx
   1.160 +	movl	%eax,		8(%edi)
   1.161 +	movl	%edx,		%esi
   1.162 +	jz	.L003maw_end
   1.163 +	/ Tail Round 3 
   1.164 +	movl	12(%ebx),	%eax
   1.165 +	mull	%ebp
   1.166 +	addl	%esi,		%eax
   1.167 +	movl	12(%edi),	%esi
   1.168 +	adcl	$0,		%edx
   1.169 +	addl	%esi,		%eax
   1.170 +	adcl	$0,		%edx
   1.171 +	decl	%ecx
   1.172 +	movl	%eax,		12(%edi)
   1.173 +	movl	%edx,		%esi
   1.174 +	jz	.L003maw_end
   1.175 +	/ Tail Round 4 
   1.176 +	movl	16(%ebx),	%eax
   1.177 +	mull	%ebp
   1.178 +	addl	%esi,		%eax
   1.179 +	movl	16(%edi),	%esi
   1.180 +	adcl	$0,		%edx
   1.181 +	addl	%esi,		%eax
   1.182 +	adcl	$0,		%edx
   1.183 +	decl	%ecx
   1.184 +	movl	%eax,		16(%edi)
   1.185 +	movl	%edx,		%esi
   1.186 +	jz	.L003maw_end
   1.187 +	/ Tail Round 5 
   1.188 +	movl	20(%ebx),	%eax
   1.189 +	mull	%ebp
   1.190 +	addl	%esi,		%eax
   1.191 +	movl	20(%edi),	%esi
   1.192 +	adcl	$0,		%edx
   1.193 +	addl	%esi,		%eax
   1.194 +	adcl	$0,		%edx
   1.195 +	decl	%ecx
   1.196 +	movl	%eax,		20(%edi)
   1.197 +	movl	%edx,		%esi
   1.198 +	jz	.L003maw_end
   1.199 +	/ Tail Round 6 
   1.200 +	movl	24(%ebx),	%eax
   1.201 +	mull	%ebp
   1.202 +	addl	%esi,		%eax
   1.203 +	movl	24(%edi),	%esi
   1.204 +	adcl	$0,		%edx
   1.205 +	addl	%esi,		%eax
   1.206 +	adcl	$0,		%edx
   1.207 +	movl	%eax,		24(%edi)
   1.208 +	movl	%edx,		%esi
   1.209 +.L003maw_end:
   1.210 +	movl	%esi,		%eax
   1.211 +	popl	%ecx
   1.212 +	popl	%edi
   1.213 +	popl	%esi
   1.214 +	popl	%ebx
   1.215 +	popl	%ebp
   1.216 +	ret
   1.217 +.bn_mul_add_words_end:
   1.218 +	.size	bn_mul_add_words,.bn_mul_add_words_end-bn_mul_add_words
   1.219 +.ident	"bn_mul_add_words"
   1.220 +.text
   1.221 +	.align 16
   1.222 +.globl bn_mul_words
   1.223 +	.type	bn_mul_words,@function
   1.224 +bn_mul_words:
   1.225 +	pushl	%ebp
   1.226 +	pushl	%ebx
   1.227 +	pushl	%esi
   1.228 +	pushl	%edi
   1.229 +
   1.230 +
   1.231 +	xorl	%esi,		%esi
   1.232 +	movl	20(%esp),	%edi
   1.233 +	movl	24(%esp),	%ebx
   1.234 +	movl	28(%esp),	%ebp
   1.235 +	movl	32(%esp),	%ecx
   1.236 +	andl	$4294967288,	%ebp
   1.237 +	jz	.L004mw_finish
   1.238 +.L005mw_loop:
   1.239 +	/ Round 0 
   1.240 +	movl	(%ebx),		%eax
   1.241 +	mull	%ecx
   1.242 +	addl	%esi,		%eax
   1.243 +	adcl	$0,		%edx
   1.244 +	movl	%eax,		(%edi)
   1.245 +	movl	%edx,		%esi
   1.246 +	/ Round 4 
   1.247 +	movl	4(%ebx),	%eax
   1.248 +	mull	%ecx
   1.249 +	addl	%esi,		%eax
   1.250 +	adcl	$0,		%edx
   1.251 +	movl	%eax,		4(%edi)
   1.252 +	movl	%edx,		%esi
   1.253 +	/ Round 8 
   1.254 +	movl	8(%ebx),	%eax
   1.255 +	mull	%ecx
   1.256 +	addl	%esi,		%eax
   1.257 +	adcl	$0,		%edx
   1.258 +	movl	%eax,		8(%edi)
   1.259 +	movl	%edx,		%esi
   1.260 +	/ Round 12 
   1.261 +	movl	12(%ebx),	%eax
   1.262 +	mull	%ecx
   1.263 +	addl	%esi,		%eax
   1.264 +	adcl	$0,		%edx
   1.265 +	movl	%eax,		12(%edi)
   1.266 +	movl	%edx,		%esi
   1.267 +	/ Round 16 
   1.268 +	movl	16(%ebx),	%eax
   1.269 +	mull	%ecx
   1.270 +	addl	%esi,		%eax
   1.271 +	adcl	$0,		%edx
   1.272 +	movl	%eax,		16(%edi)
   1.273 +	movl	%edx,		%esi
   1.274 +	/ Round 20 
   1.275 +	movl	20(%ebx),	%eax
   1.276 +	mull	%ecx
   1.277 +	addl	%esi,		%eax
   1.278 +	adcl	$0,		%edx
   1.279 +	movl	%eax,		20(%edi)
   1.280 +	movl	%edx,		%esi
   1.281 +	/ Round 24 
   1.282 +	movl	24(%ebx),	%eax
   1.283 +	mull	%ecx
   1.284 +	addl	%esi,		%eax
   1.285 +	adcl	$0,		%edx
   1.286 +	movl	%eax,		24(%edi)
   1.287 +	movl	%edx,		%esi
   1.288 +	/ Round 28 
   1.289 +	movl	28(%ebx),	%eax
   1.290 +	mull	%ecx
   1.291 +	addl	%esi,		%eax
   1.292 +	adcl	$0,		%edx
   1.293 +	movl	%eax,		28(%edi)
   1.294 +	movl	%edx,		%esi
   1.295 +
   1.296 +	addl	$32,		%ebx
   1.297 +	addl	$32,		%edi
   1.298 +	subl	$8,		%ebp
   1.299 +	jz	.L004mw_finish
   1.300 +	jmp	.L005mw_loop
   1.301 +.L004mw_finish:
   1.302 +	movl	28(%esp),	%ebp
   1.303 +	andl	$7,		%ebp
   1.304 +	jnz	.L006mw_finish2
   1.305 +	jmp	.L007mw_end
   1.306 +.align 16
   1.307 +.L006mw_finish2:
   1.308 +	/ Tail Round 0 
   1.309 +	movl	(%ebx),		%eax
   1.310 +	mull	%ecx
   1.311 +	addl	%esi,		%eax
   1.312 +	adcl	$0,		%edx
   1.313 +	movl	%eax,		(%edi)
   1.314 +	movl	%edx,		%esi
   1.315 +	decl	%ebp
   1.316 +	jz	.L007mw_end
   1.317 +	/ Tail Round 1 
   1.318 +	movl	4(%ebx),	%eax
   1.319 +	mull	%ecx
   1.320 +	addl	%esi,		%eax
   1.321 +	adcl	$0,		%edx
   1.322 +	movl	%eax,		4(%edi)
   1.323 +	movl	%edx,		%esi
   1.324 +	decl	%ebp
   1.325 +	jz	.L007mw_end
   1.326 +	/ Tail Round 2 
   1.327 +	movl	8(%ebx),	%eax
   1.328 +	mull	%ecx
   1.329 +	addl	%esi,		%eax
   1.330 +	adcl	$0,		%edx
   1.331 +	movl	%eax,		8(%edi)
   1.332 +	movl	%edx,		%esi
   1.333 +	decl	%ebp
   1.334 +	jz	.L007mw_end
   1.335 +	/ Tail Round 3 
   1.336 +	movl	12(%ebx),	%eax
   1.337 +	mull	%ecx
   1.338 +	addl	%esi,		%eax
   1.339 +	adcl	$0,		%edx
   1.340 +	movl	%eax,		12(%edi)
   1.341 +	movl	%edx,		%esi
   1.342 +	decl	%ebp
   1.343 +	jz	.L007mw_end
   1.344 +	/ Tail Round 4 
   1.345 +	movl	16(%ebx),	%eax
   1.346 +	mull	%ecx
   1.347 +	addl	%esi,		%eax
   1.348 +	adcl	$0,		%edx
   1.349 +	movl	%eax,		16(%edi)
   1.350 +	movl	%edx,		%esi
   1.351 +	decl	%ebp
   1.352 +	jz	.L007mw_end
   1.353 +	/ Tail Round 5 
   1.354 +	movl	20(%ebx),	%eax
   1.355 +	mull	%ecx
   1.356 +	addl	%esi,		%eax
   1.357 +	adcl	$0,		%edx
   1.358 +	movl	%eax,		20(%edi)
   1.359 +	movl	%edx,		%esi
   1.360 +	decl	%ebp
   1.361 +	jz	.L007mw_end
   1.362 +	/ Tail Round 6 
   1.363 +	movl	24(%ebx),	%eax
   1.364 +	mull	%ecx
   1.365 +	addl	%esi,		%eax
   1.366 +	adcl	$0,		%edx
   1.367 +	movl	%eax,		24(%edi)
   1.368 +	movl	%edx,		%esi
   1.369 +.L007mw_end:
   1.370 +	movl	%esi,		%eax
   1.371 +	popl	%edi
   1.372 +	popl	%esi
   1.373 +	popl	%ebx
   1.374 +	popl	%ebp
   1.375 +	ret
   1.376 +.bn_mul_words_end:
   1.377 +	.size	bn_mul_words,.bn_mul_words_end-bn_mul_words
   1.378 +.ident	"bn_mul_words"
   1.379 +.text
   1.380 +	.align 16
   1.381 +.globl bn_sqr_words
   1.382 +	.type	bn_sqr_words,@function
   1.383 +bn_sqr_words:
   1.384 +	pushl	%ebp
   1.385 +	pushl	%ebx
   1.386 +	pushl	%esi
   1.387 +	pushl	%edi
   1.388 +
   1.389 +
   1.390 +	movl	20(%esp),	%esi
   1.391 +	movl	24(%esp),	%edi
   1.392 +	movl	28(%esp),	%ebx
   1.393 +	andl	$4294967288,	%ebx
   1.394 +	jz	.L008sw_finish
   1.395 +.L009sw_loop:
   1.396 +	/ Round 0 
   1.397 +	movl	(%edi),		%eax
   1.398 +	mull	%eax
   1.399 +	movl	%eax,		(%esi)
   1.400 +	movl	%edx,		4(%esi)
   1.401 +	/ Round 4 
   1.402 +	movl	4(%edi),	%eax
   1.403 +	mull	%eax
   1.404 +	movl	%eax,		8(%esi)
   1.405 +	movl	%edx,		12(%esi)
   1.406 +	/ Round 8 
   1.407 +	movl	8(%edi),	%eax
   1.408 +	mull	%eax
   1.409 +	movl	%eax,		16(%esi)
   1.410 +	movl	%edx,		20(%esi)
   1.411 +	/ Round 12 
   1.412 +	movl	12(%edi),	%eax
   1.413 +	mull	%eax
   1.414 +	movl	%eax,		24(%esi)
   1.415 +	movl	%edx,		28(%esi)
   1.416 +	/ Round 16 
   1.417 +	movl	16(%edi),	%eax
   1.418 +	mull	%eax
   1.419 +	movl	%eax,		32(%esi)
   1.420 +	movl	%edx,		36(%esi)
   1.421 +	/ Round 20 
   1.422 +	movl	20(%edi),	%eax
   1.423 +	mull	%eax
   1.424 +	movl	%eax,		40(%esi)
   1.425 +	movl	%edx,		44(%esi)
   1.426 +	/ Round 24 
   1.427 +	movl	24(%edi),	%eax
   1.428 +	mull	%eax
   1.429 +	movl	%eax,		48(%esi)
   1.430 +	movl	%edx,		52(%esi)
   1.431 +	/ Round 28 
   1.432 +	movl	28(%edi),	%eax
   1.433 +	mull	%eax
   1.434 +	movl	%eax,		56(%esi)
   1.435 +	movl	%edx,		60(%esi)
   1.436 +
   1.437 +	addl	$32,		%edi
   1.438 +	addl	$64,		%esi
   1.439 +	subl	$8,		%ebx
   1.440 +	jnz	.L009sw_loop
   1.441 +.L008sw_finish:
   1.442 +	movl	28(%esp),	%ebx
   1.443 +	andl	$7,		%ebx
   1.444 +	jz	.L010sw_end
   1.445 +	/ Tail Round 0 
   1.446 +	movl	(%edi),		%eax
   1.447 +	mull	%eax
   1.448 +	movl	%eax,		(%esi)
   1.449 +	decl	%ebx
   1.450 +	movl	%edx,		4(%esi)
   1.451 +	jz	.L010sw_end
   1.452 +	/ Tail Round 1 
   1.453 +	movl	4(%edi),	%eax
   1.454 +	mull	%eax
   1.455 +	movl	%eax,		8(%esi)
   1.456 +	decl	%ebx
   1.457 +	movl	%edx,		12(%esi)
   1.458 +	jz	.L010sw_end
   1.459 +	/ Tail Round 2 
   1.460 +	movl	8(%edi),	%eax
   1.461 +	mull	%eax
   1.462 +	movl	%eax,		16(%esi)
   1.463 +	decl	%ebx
   1.464 +	movl	%edx,		20(%esi)
   1.465 +	jz	.L010sw_end
   1.466 +	/ Tail Round 3 
   1.467 +	movl	12(%edi),	%eax
   1.468 +	mull	%eax
   1.469 +	movl	%eax,		24(%esi)
   1.470 +	decl	%ebx
   1.471 +	movl	%edx,		28(%esi)
   1.472 +	jz	.L010sw_end
   1.473 +	/ Tail Round 4 
   1.474 +	movl	16(%edi),	%eax
   1.475 +	mull	%eax
   1.476 +	movl	%eax,		32(%esi)
   1.477 +	decl	%ebx
   1.478 +	movl	%edx,		36(%esi)
   1.479 +	jz	.L010sw_end
   1.480 +	/ Tail Round 5 
   1.481 +	movl	20(%edi),	%eax
   1.482 +	mull	%eax
   1.483 +	movl	%eax,		40(%esi)
   1.484 +	decl	%ebx
   1.485 +	movl	%edx,		44(%esi)
   1.486 +	jz	.L010sw_end
   1.487 +	/ Tail Round 6 
   1.488 +	movl	24(%edi),	%eax
   1.489 +	mull	%eax
   1.490 +	movl	%eax,		48(%esi)
   1.491 +	movl	%edx,		52(%esi)
   1.492 +.L010sw_end:
   1.493 +	popl	%edi
   1.494 +	popl	%esi
   1.495 +	popl	%ebx
   1.496 +	popl	%ebp
   1.497 +	ret
   1.498 +.bn_sqr_words_end:
   1.499 +	.size	bn_sqr_words,.bn_sqr_words_end-bn_sqr_words
   1.500 +.ident	"bn_sqr_words"
   1.501 +.text
   1.502 +	.align 16
   1.503 +.globl bn_div_words
   1.504 +	.type	bn_div_words,@function
   1.505 +bn_div_words:
   1.506 +	pushl	%ebp
   1.507 +	pushl	%ebx
   1.508 +	pushl	%esi
   1.509 +	pushl	%edi
   1.510 +
   1.511 +	movl	20(%esp),	%edx
   1.512 +	movl	24(%esp),	%eax
   1.513 +	movl	28(%esp),	%ebx
   1.514 +	divl	%ebx
   1.515 +	popl	%edi
   1.516 +	popl	%esi
   1.517 +	popl	%ebx
   1.518 +	popl	%ebp
   1.519 +	ret
   1.520 +.bn_div_words_end:
   1.521 +	.size	bn_div_words,.bn_div_words_end-bn_div_words
   1.522 +.ident	"bn_div_words"
   1.523 +.text
   1.524 +	.align 16
   1.525 +.globl bn_add_words
   1.526 +	.type	bn_add_words,@function
   1.527 +bn_add_words:
   1.528 +	pushl	%ebp
   1.529 +	pushl	%ebx
   1.530 +	pushl	%esi
   1.531 +	pushl	%edi
   1.532 +
   1.533 +
   1.534 +	movl	20(%esp),	%ebx
   1.535 +	movl	24(%esp),	%esi
   1.536 +	movl	28(%esp),	%edi
   1.537 +	movl	32(%esp),	%ebp
   1.538 +	xorl	%eax,		%eax
   1.539 +	andl	$4294967288,	%ebp
   1.540 +	jz	.L011aw_finish
   1.541 +.L012aw_loop:
   1.542 +	/ Round 0 
   1.543 +	movl	(%esi),		%ecx
   1.544 +	movl	(%edi),		%edx
   1.545 +	addl	%eax,		%ecx
   1.546 +	movl	$0,		%eax
   1.547 +	adcl	%eax,		%eax
   1.548 +	addl	%edx,		%ecx
   1.549 +	adcl	$0,		%eax
   1.550 +	movl	%ecx,		(%ebx)
   1.551 +	/ Round 1 
   1.552 +	movl	4(%esi),	%ecx
   1.553 +	movl	4(%edi),	%edx
   1.554 +	addl	%eax,		%ecx
   1.555 +	movl	$0,		%eax
   1.556 +	adcl	%eax,		%eax
   1.557 +	addl	%edx,		%ecx
   1.558 +	adcl	$0,		%eax
   1.559 +	movl	%ecx,		4(%ebx)
   1.560 +	/ Round 2 
   1.561 +	movl	8(%esi),	%ecx
   1.562 +	movl	8(%edi),	%edx
   1.563 +	addl	%eax,		%ecx
   1.564 +	movl	$0,		%eax
   1.565 +	adcl	%eax,		%eax
   1.566 +	addl	%edx,		%ecx
   1.567 +	adcl	$0,		%eax
   1.568 +	movl	%ecx,		8(%ebx)
   1.569 +	/ Round 3 
   1.570 +	movl	12(%esi),	%ecx
   1.571 +	movl	12(%edi),	%edx
   1.572 +	addl	%eax,		%ecx
   1.573 +	movl	$0,		%eax
   1.574 +	adcl	%eax,		%eax
   1.575 +	addl	%edx,		%ecx
   1.576 +	adcl	$0,		%eax
   1.577 +	movl	%ecx,		12(%ebx)
   1.578 +	/ Round 4 
   1.579 +	movl	16(%esi),	%ecx
   1.580 +	movl	16(%edi),	%edx
   1.581 +	addl	%eax,		%ecx
   1.582 +	movl	$0,		%eax
   1.583 +	adcl	%eax,		%eax
   1.584 +	addl	%edx,		%ecx
   1.585 +	adcl	$0,		%eax
   1.586 +	movl	%ecx,		16(%ebx)
   1.587 +	/ Round 5 
   1.588 +	movl	20(%esi),	%ecx
   1.589 +	movl	20(%edi),	%edx
   1.590 +	addl	%eax,		%ecx
   1.591 +	movl	$0,		%eax
   1.592 +	adcl	%eax,		%eax
   1.593 +	addl	%edx,		%ecx
   1.594 +	adcl	$0,		%eax
   1.595 +	movl	%ecx,		20(%ebx)
   1.596 +	/ Round 6 
   1.597 +	movl	24(%esi),	%ecx
   1.598 +	movl	24(%edi),	%edx
   1.599 +	addl	%eax,		%ecx
   1.600 +	movl	$0,		%eax
   1.601 +	adcl	%eax,		%eax
   1.602 +	addl	%edx,		%ecx
   1.603 +	adcl	$0,		%eax
   1.604 +	movl	%ecx,		24(%ebx)
   1.605 +	/ Round 7 
   1.606 +	movl	28(%esi),	%ecx
   1.607 +	movl	28(%edi),	%edx
   1.608 +	addl	%eax,		%ecx
   1.609 +	movl	$0,		%eax
   1.610 +	adcl	%eax,		%eax
   1.611 +	addl	%edx,		%ecx
   1.612 +	adcl	$0,		%eax
   1.613 +	movl	%ecx,		28(%ebx)
   1.614 +
   1.615 +	addl	$32,		%esi
   1.616 +	addl	$32,		%edi
   1.617 +	addl	$32,		%ebx
   1.618 +	subl	$8,		%ebp
   1.619 +	jnz	.L012aw_loop
   1.620 +.L011aw_finish:
   1.621 +	movl	32(%esp),	%ebp
   1.622 +	andl	$7,		%ebp
   1.623 +	jz	.L013aw_end
   1.624 +	/ Tail Round 0 
   1.625 +	movl	(%esi),		%ecx
   1.626 +	movl	(%edi),		%edx
   1.627 +	addl	%eax,		%ecx
   1.628 +	movl	$0,		%eax
   1.629 +	adcl	%eax,		%eax
   1.630 +	addl	%edx,		%ecx
   1.631 +	adcl	$0,		%eax
   1.632 +	decl	%ebp
   1.633 +	movl	%ecx,		(%ebx)
   1.634 +	jz	.L013aw_end
   1.635 +	/ Tail Round 1 
   1.636 +	movl	4(%esi),	%ecx
   1.637 +	movl	4(%edi),	%edx
   1.638 +	addl	%eax,		%ecx
   1.639 +	movl	$0,		%eax
   1.640 +	adcl	%eax,		%eax
   1.641 +	addl	%edx,		%ecx
   1.642 +	adcl	$0,		%eax
   1.643 +	decl	%ebp
   1.644 +	movl	%ecx,		4(%ebx)
   1.645 +	jz	.L013aw_end
   1.646 +	/ Tail Round 2 
   1.647 +	movl	8(%esi),	%ecx
   1.648 +	movl	8(%edi),	%edx
   1.649 +	addl	%eax,		%ecx
   1.650 +	movl	$0,		%eax
   1.651 +	adcl	%eax,		%eax
   1.652 +	addl	%edx,		%ecx
   1.653 +	adcl	$0,		%eax
   1.654 +	decl	%ebp
   1.655 +	movl	%ecx,		8(%ebx)
   1.656 +	jz	.L013aw_end
   1.657 +	/ Tail Round 3 
   1.658 +	movl	12(%esi),	%ecx
   1.659 +	movl	12(%edi),	%edx
   1.660 +	addl	%eax,		%ecx
   1.661 +	movl	$0,		%eax
   1.662 +	adcl	%eax,		%eax
   1.663 +	addl	%edx,		%ecx
   1.664 +	adcl	$0,		%eax
   1.665 +	decl	%ebp
   1.666 +	movl	%ecx,		12(%ebx)
   1.667 +	jz	.L013aw_end
   1.668 +	/ Tail Round 4 
   1.669 +	movl	16(%esi),	%ecx
   1.670 +	movl	16(%edi),	%edx
   1.671 +	addl	%eax,		%ecx
   1.672 +	movl	$0,		%eax
   1.673 +	adcl	%eax,		%eax
   1.674 +	addl	%edx,		%ecx
   1.675 +	adcl	$0,		%eax
   1.676 +	decl	%ebp
   1.677 +	movl	%ecx,		16(%ebx)
   1.678 +	jz	.L013aw_end
   1.679 +	/ Tail Round 5 
   1.680 +	movl	20(%esi),	%ecx
   1.681 +	movl	20(%edi),	%edx
   1.682 +	addl	%eax,		%ecx
   1.683 +	movl	$0,		%eax
   1.684 +	adcl	%eax,		%eax
   1.685 +	addl	%edx,		%ecx
   1.686 +	adcl	$0,		%eax
   1.687 +	decl	%ebp
   1.688 +	movl	%ecx,		20(%ebx)
   1.689 +	jz	.L013aw_end
   1.690 +	/ Tail Round 6 
   1.691 +	movl	24(%esi),	%ecx
   1.692 +	movl	24(%edi),	%edx
   1.693 +	addl	%eax,		%ecx
   1.694 +	movl	$0,		%eax
   1.695 +	adcl	%eax,		%eax
   1.696 +	addl	%edx,		%ecx
   1.697 +	adcl	$0,		%eax
   1.698 +	movl	%ecx,		24(%ebx)
   1.699 +.L013aw_end:
   1.700 +	popl	%edi
   1.701 +	popl	%esi
   1.702 +	popl	%ebx
   1.703 +	popl	%ebp
   1.704 +	ret
   1.705 +.bn_add_words_end:
   1.706 +	.size	bn_add_words,.bn_add_words_end-bn_add_words
   1.707 +.ident	"bn_add_words"
   1.708 +.text
   1.709 +	.align 16
   1.710 +.globl bn_sub_words
   1.711 +	.type	bn_sub_words,@function
   1.712 +bn_sub_words:
   1.713 +	pushl	%ebp
   1.714 +	pushl	%ebx
   1.715 +	pushl	%esi
   1.716 +	pushl	%edi
   1.717 +
   1.718 +
   1.719 +	movl	20(%esp),	%ebx
   1.720 +	movl	24(%esp),	%esi
   1.721 +	movl	28(%esp),	%edi
   1.722 +	movl	32(%esp),	%ebp
   1.723 +	xorl	%eax,		%eax
   1.724 +	andl	$4294967288,	%ebp
   1.725 +	jz	.L014aw_finish
   1.726 +.L015aw_loop:
   1.727 +	/ Round 0 
   1.728 +	movl	(%esi),		%ecx
   1.729 +	movl	(%edi),		%edx
   1.730 +	subl	%eax,		%ecx
   1.731 +	movl	$0,		%eax
   1.732 +	adcl	%eax,		%eax
   1.733 +	subl	%edx,		%ecx
   1.734 +	adcl	$0,		%eax
   1.735 +	movl	%ecx,		(%ebx)
   1.736 +	/ Round 1 
   1.737 +	movl	4(%esi),	%ecx
   1.738 +	movl	4(%edi),	%edx
   1.739 +	subl	%eax,		%ecx
   1.740 +	movl	$0,		%eax
   1.741 +	adcl	%eax,		%eax
   1.742 +	subl	%edx,		%ecx
   1.743 +	adcl	$0,		%eax
   1.744 +	movl	%ecx,		4(%ebx)
   1.745 +	/ Round 2 
   1.746 +	movl	8(%esi),	%ecx
   1.747 +	movl	8(%edi),	%edx
   1.748 +	subl	%eax,		%ecx
   1.749 +	movl	$0,		%eax
   1.750 +	adcl	%eax,		%eax
   1.751 +	subl	%edx,		%ecx
   1.752 +	adcl	$0,		%eax
   1.753 +	movl	%ecx,		8(%ebx)
   1.754 +	/ Round 3 
   1.755 +	movl	12(%esi),	%ecx
   1.756 +	movl	12(%edi),	%edx
   1.757 +	subl	%eax,		%ecx
   1.758 +	movl	$0,		%eax
   1.759 +	adcl	%eax,		%eax
   1.760 +	subl	%edx,		%ecx
   1.761 +	adcl	$0,		%eax
   1.762 +	movl	%ecx,		12(%ebx)
   1.763 +	/ Round 4 
   1.764 +	movl	16(%esi),	%ecx
   1.765 +	movl	16(%edi),	%edx
   1.766 +	subl	%eax,		%ecx
   1.767 +	movl	$0,		%eax
   1.768 +	adcl	%eax,		%eax
   1.769 +	subl	%edx,		%ecx
   1.770 +	adcl	$0,		%eax
   1.771 +	movl	%ecx,		16(%ebx)
   1.772 +	/ Round 5 
   1.773 +	movl	20(%esi),	%ecx
   1.774 +	movl	20(%edi),	%edx
   1.775 +	subl	%eax,		%ecx
   1.776 +	movl	$0,		%eax
   1.777 +	adcl	%eax,		%eax
   1.778 +	subl	%edx,		%ecx
   1.779 +	adcl	$0,		%eax
   1.780 +	movl	%ecx,		20(%ebx)
   1.781 +	/ Round 6 
   1.782 +	movl	24(%esi),	%ecx
   1.783 +	movl	24(%edi),	%edx
   1.784 +	subl	%eax,		%ecx
   1.785 +	movl	$0,		%eax
   1.786 +	adcl	%eax,		%eax
   1.787 +	subl	%edx,		%ecx
   1.788 +	adcl	$0,		%eax
   1.789 +	movl	%ecx,		24(%ebx)
   1.790 +	/ Round 7 
   1.791 +	movl	28(%esi),	%ecx
   1.792 +	movl	28(%edi),	%edx
   1.793 +	subl	%eax,		%ecx
   1.794 +	movl	$0,		%eax
   1.795 +	adcl	%eax,		%eax
   1.796 +	subl	%edx,		%ecx
   1.797 +	adcl	$0,		%eax
   1.798 +	movl	%ecx,		28(%ebx)
   1.799 +
   1.800 +	addl	$32,		%esi
   1.801 +	addl	$32,		%edi
   1.802 +	addl	$32,		%ebx
   1.803 +	subl	$8,		%ebp
   1.804 +	jnz	.L015aw_loop
   1.805 +.L014aw_finish:
   1.806 +	movl	32(%esp),	%ebp
   1.807 +	andl	$7,		%ebp
   1.808 +	jz	.L016aw_end
   1.809 +	/ Tail Round 0 
   1.810 +	movl	(%esi),		%ecx
   1.811 +	movl	(%edi),		%edx
   1.812 +	subl	%eax,		%ecx
   1.813 +	movl	$0,		%eax
   1.814 +	adcl	%eax,		%eax
   1.815 +	subl	%edx,		%ecx
   1.816 +	adcl	$0,		%eax
   1.817 +	decl	%ebp
   1.818 +	movl	%ecx,		(%ebx)
   1.819 +	jz	.L016aw_end
   1.820 +	/ Tail Round 1 
   1.821 +	movl	4(%esi),	%ecx
   1.822 +	movl	4(%edi),	%edx
   1.823 +	subl	%eax,		%ecx
   1.824 +	movl	$0,		%eax
   1.825 +	adcl	%eax,		%eax
   1.826 +	subl	%edx,		%ecx
   1.827 +	adcl	$0,		%eax
   1.828 +	decl	%ebp
   1.829 +	movl	%ecx,		4(%ebx)
   1.830 +	jz	.L016aw_end
   1.831 +	/ Tail Round 2 
   1.832 +	movl	8(%esi),	%ecx
   1.833 +	movl	8(%edi),	%edx
   1.834 +	subl	%eax,		%ecx
   1.835 +	movl	$0,		%eax
   1.836 +	adcl	%eax,		%eax
   1.837 +	subl	%edx,		%ecx
   1.838 +	adcl	$0,		%eax
   1.839 +	decl	%ebp
   1.840 +	movl	%ecx,		8(%ebx)
   1.841 +	jz	.L016aw_end
   1.842 +	/ Tail Round 3 
   1.843 +	movl	12(%esi),	%ecx
   1.844 +	movl	12(%edi),	%edx
   1.845 +	subl	%eax,		%ecx
   1.846 +	movl	$0,		%eax
   1.847 +	adcl	%eax,		%eax
   1.848 +	subl	%edx,		%ecx
   1.849 +	adcl	$0,		%eax
   1.850 +	decl	%ebp
   1.851 +	movl	%ecx,		12(%ebx)
   1.852 +	jz	.L016aw_end
   1.853 +	/ Tail Round 4 
   1.854 +	movl	16(%esi),	%ecx
   1.855 +	movl	16(%edi),	%edx
   1.856 +	subl	%eax,		%ecx
   1.857 +	movl	$0,		%eax
   1.858 +	adcl	%eax,		%eax
   1.859 +	subl	%edx,		%ecx
   1.860 +	adcl	$0,		%eax
   1.861 +	decl	%ebp
   1.862 +	movl	%ecx,		16(%ebx)
   1.863 +	jz	.L016aw_end
   1.864 +	/ Tail Round 5 
   1.865 +	movl	20(%esi),	%ecx
   1.866 +	movl	20(%edi),	%edx
   1.867 +	subl	%eax,		%ecx
   1.868 +	movl	$0,		%eax
   1.869 +	adcl	%eax,		%eax
   1.870 +	subl	%edx,		%ecx
   1.871 +	adcl	$0,		%eax
   1.872 +	decl	%ebp
   1.873 +	movl	%ecx,		20(%ebx)
   1.874 +	jz	.L016aw_end
   1.875 +	/ Tail Round 6 
   1.876 +	movl	24(%esi),	%ecx
   1.877 +	movl	24(%edi),	%edx
   1.878 +	subl	%eax,		%ecx
   1.879 +	movl	$0,		%eax
   1.880 +	adcl	%eax,		%eax
   1.881 +	subl	%edx,		%ecx
   1.882 +	adcl	$0,		%eax
   1.883 +	movl	%ecx,		24(%ebx)
   1.884 +.L016aw_end:
   1.885 +	popl	%edi
   1.886 +	popl	%esi
   1.887 +	popl	%ebx
   1.888 +	popl	%ebp
   1.889 +	ret
   1.890 +.bn_sub_words_end:
   1.891 +	.size	bn_sub_words,.bn_sub_words_end-bn_sub_words
   1.892 +.ident	"bn_sub_words"
   1.893 +.text
   1.894 +	.align 16
   1.895 +.globl bn_mul_comba8
   1.896 +	.type	bn_mul_comba8,@function
   1.897 +bn_mul_comba8:
   1.898 +	pushl	%esi
   1.899 +	movl	12(%esp),	%esi
   1.900 +	pushl	%edi
   1.901 +	movl	20(%esp),	%edi
   1.902 +	pushl	%ebp
   1.903 +	pushl	%ebx
   1.904 +	xorl	%ebx,		%ebx
   1.905 +	movl	(%esi),		%eax
   1.906 +	xorl	%ecx,		%ecx
   1.907 +	movl	(%edi),		%edx
   1.908 +	/ ################## Calculate word 0 
   1.909 +	xorl	%ebp,		%ebp
   1.910 +	/ mul a[0]*b[0] 
   1.911 +	mull	%edx
   1.912 +	addl	%eax,		%ebx
   1.913 +	movl	20(%esp),	%eax
   1.914 +	adcl	%edx,		%ecx
   1.915 +	movl	(%edi),		%edx
   1.916 +	adcl	$0,		%ebp
   1.917 +	movl	%ebx,		(%eax)
   1.918 +	movl	4(%esi),	%eax
   1.919 +	/ saved r[0] 
   1.920 +	/ ################## Calculate word 1 
   1.921 +	xorl	%ebx,		%ebx
   1.922 +	/ mul a[1]*b[0] 
   1.923 +	mull	%edx
   1.924 +	addl	%eax,		%ecx
   1.925 +	movl	(%esi),		%eax
   1.926 +	adcl	%edx,		%ebp
   1.927 +	movl	4(%edi),	%edx
   1.928 +	adcl	$0,		%ebx
   1.929 +	/ mul a[0]*b[1] 
   1.930 +	mull	%edx
   1.931 +	addl	%eax,		%ecx
   1.932 +	movl	20(%esp),	%eax
   1.933 +	adcl	%edx,		%ebp
   1.934 +	movl	(%edi),		%edx
   1.935 +	adcl	$0,		%ebx
   1.936 +	movl	%ecx,		4(%eax)
   1.937 +	movl	8(%esi),	%eax
   1.938 +	/ saved r[1] 
   1.939 +	/ ################## Calculate word 2 
   1.940 +	xorl	%ecx,		%ecx
   1.941 +	/ mul a[2]*b[0] 
   1.942 +	mull	%edx
   1.943 +	addl	%eax,		%ebp
   1.944 +	movl	4(%esi),	%eax
   1.945 +	adcl	%edx,		%ebx
   1.946 +	movl	4(%edi),	%edx
   1.947 +	adcl	$0,		%ecx
   1.948 +	/ mul a[1]*b[1] 
   1.949 +	mull	%edx
   1.950 +	addl	%eax,		%ebp
   1.951 +	movl	(%esi),		%eax
   1.952 +	adcl	%edx,		%ebx
   1.953 +	movl	8(%edi),	%edx
   1.954 +	adcl	$0,		%ecx
   1.955 +	/ mul a[0]*b[2] 
   1.956 +	mull	%edx
   1.957 +	addl	%eax,		%ebp
   1.958 +	movl	20(%esp),	%eax
   1.959 +	adcl	%edx,		%ebx
   1.960 +	movl	(%edi),		%edx
   1.961 +	adcl	$0,		%ecx
   1.962 +	movl	%ebp,		8(%eax)
   1.963 +	movl	12(%esi),	%eax
   1.964 +	/ saved r[2] 
   1.965 +	/ ################## Calculate word 3 
   1.966 +	xorl	%ebp,		%ebp
   1.967 +	/ mul a[3]*b[0] 
   1.968 +	mull	%edx
   1.969 +	addl	%eax,		%ebx
   1.970 +	movl	8(%esi),	%eax
   1.971 +	adcl	%edx,		%ecx
   1.972 +	movl	4(%edi),	%edx
   1.973 +	adcl	$0,		%ebp
   1.974 +	/ mul a[2]*b[1] 
   1.975 +	mull	%edx
   1.976 +	addl	%eax,		%ebx
   1.977 +	movl	4(%esi),	%eax
   1.978 +	adcl	%edx,		%ecx
   1.979 +	movl	8(%edi),	%edx
   1.980 +	adcl	$0,		%ebp
   1.981 +	/ mul a[1]*b[2] 
   1.982 +	mull	%edx
   1.983 +	addl	%eax,		%ebx
   1.984 +	movl	(%esi),		%eax
   1.985 +	adcl	%edx,		%ecx
   1.986 +	movl	12(%edi),	%edx
   1.987 +	adcl	$0,		%ebp
   1.988 +	/ mul a[0]*b[3] 
   1.989 +	mull	%edx
   1.990 +	addl	%eax,		%ebx
   1.991 +	movl	20(%esp),	%eax
   1.992 +	adcl	%edx,		%ecx
   1.993 +	movl	(%edi),		%edx
   1.994 +	adcl	$0,		%ebp
   1.995 +	movl	%ebx,		12(%eax)
   1.996 +	movl	16(%esi),	%eax
   1.997 +	/ saved r[3] 
   1.998 +	/ ################## Calculate word 4 
   1.999 +	xorl	%ebx,		%ebx
  1.1000 +	/ mul a[4]*b[0] 
  1.1001 +	mull	%edx
  1.1002 +	addl	%eax,		%ecx
  1.1003 +	movl	12(%esi),	%eax
  1.1004 +	adcl	%edx,		%ebp
  1.1005 +	movl	4(%edi),	%edx
  1.1006 +	adcl	$0,		%ebx
  1.1007 +	/ mul a[3]*b[1] 
  1.1008 +	mull	%edx
  1.1009 +	addl	%eax,		%ecx
  1.1010 +	movl	8(%esi),	%eax
  1.1011 +	adcl	%edx,		%ebp
  1.1012 +	movl	8(%edi),	%edx
  1.1013 +	adcl	$0,		%ebx
  1.1014 +	/ mul a[2]*b[2] 
  1.1015 +	mull	%edx
  1.1016 +	addl	%eax,		%ecx
  1.1017 +	movl	4(%esi),	%eax
  1.1018 +	adcl	%edx,		%ebp
  1.1019 +	movl	12(%edi),	%edx
  1.1020 +	adcl	$0,		%ebx
  1.1021 +	/ mul a[1]*b[3] 
  1.1022 +	mull	%edx
  1.1023 +	addl	%eax,		%ecx
  1.1024 +	movl	(%esi),		%eax
  1.1025 +	adcl	%edx,		%ebp
  1.1026 +	movl	16(%edi),	%edx
  1.1027 +	adcl	$0,		%ebx
  1.1028 +	/ mul a[0]*b[4] 
  1.1029 +	mull	%edx
  1.1030 +	addl	%eax,		%ecx
  1.1031 +	movl	20(%esp),	%eax
  1.1032 +	adcl	%edx,		%ebp
  1.1033 +	movl	(%edi),		%edx
  1.1034 +	adcl	$0,		%ebx
  1.1035 +	movl	%ecx,		16(%eax)
  1.1036 +	movl	20(%esi),	%eax
  1.1037 +	/ saved r[4] 
  1.1038 +	/ ################## Calculate word 5 
  1.1039 +	xorl	%ecx,		%ecx
  1.1040 +	/ mul a[5]*b[0] 
  1.1041 +	mull	%edx
  1.1042 +	addl	%eax,		%ebp
  1.1043 +	movl	16(%esi),	%eax
  1.1044 +	adcl	%edx,		%ebx
  1.1045 +	movl	4(%edi),	%edx
  1.1046 +	adcl	$0,		%ecx
  1.1047 +	/ mul a[4]*b[1] 
  1.1048 +	mull	%edx
  1.1049 +	addl	%eax,		%ebp
  1.1050 +	movl	12(%esi),	%eax
  1.1051 +	adcl	%edx,		%ebx
  1.1052 +	movl	8(%edi),	%edx
  1.1053 +	adcl	$0,		%ecx
  1.1054 +	/ mul a[3]*b[2] 
  1.1055 +	mull	%edx
  1.1056 +	addl	%eax,		%ebp
  1.1057 +	movl	8(%esi),	%eax
  1.1058 +	adcl	%edx,		%ebx
  1.1059 +	movl	12(%edi),	%edx
  1.1060 +	adcl	$0,		%ecx
  1.1061 +	/ mul a[2]*b[3] 
  1.1062 +	mull	%edx
  1.1063 +	addl	%eax,		%ebp
  1.1064 +	movl	4(%esi),	%eax
  1.1065 +	adcl	%edx,		%ebx
  1.1066 +	movl	16(%edi),	%edx
  1.1067 +	adcl	$0,		%ecx
  1.1068 +	/ mul a[1]*b[4] 
  1.1069 +	mull	%edx
  1.1070 +	addl	%eax,		%ebp
  1.1071 +	movl	(%esi),		%eax
  1.1072 +	adcl	%edx,		%ebx
  1.1073 +	movl	20(%edi),	%edx
  1.1074 +	adcl	$0,		%ecx
  1.1075 +	/ mul a[0]*b[5] 
  1.1076 +	mull	%edx
  1.1077 +	addl	%eax,		%ebp
  1.1078 +	movl	20(%esp),	%eax
  1.1079 +	adcl	%edx,		%ebx
  1.1080 +	movl	(%edi),		%edx
  1.1081 +	adcl	$0,		%ecx
  1.1082 +	movl	%ebp,		20(%eax)
  1.1083 +	movl	24(%esi),	%eax
  1.1084 +	/ saved r[5] 
  1.1085 +	/ ################## Calculate word 6 
  1.1086 +	xorl	%ebp,		%ebp
  1.1087 +	/ mul a[6]*b[0] 
  1.1088 +	mull	%edx
  1.1089 +	addl	%eax,		%ebx
  1.1090 +	movl	20(%esi),	%eax
  1.1091 +	adcl	%edx,		%ecx
  1.1092 +	movl	4(%edi),	%edx
  1.1093 +	adcl	$0,		%ebp
  1.1094 +	/ mul a[5]*b[1] 
  1.1095 +	mull	%edx
  1.1096 +	addl	%eax,		%ebx
  1.1097 +	movl	16(%esi),	%eax
  1.1098 +	adcl	%edx,		%ecx
  1.1099 +	movl	8(%edi),	%edx
  1.1100 +	adcl	$0,		%ebp
  1.1101 +	/ mul a[4]*b[2] 
  1.1102 +	mull	%edx
  1.1103 +	addl	%eax,		%ebx
  1.1104 +	movl	12(%esi),	%eax
  1.1105 +	adcl	%edx,		%ecx
  1.1106 +	movl	12(%edi),	%edx
  1.1107 +	adcl	$0,		%ebp
  1.1108 +	/ mul a[3]*b[3] 
  1.1109 +	mull	%edx
  1.1110 +	addl	%eax,		%ebx
  1.1111 +	movl	8(%esi),	%eax
  1.1112 +	adcl	%edx,		%ecx
  1.1113 +	movl	16(%edi),	%edx
  1.1114 +	adcl	$0,		%ebp
  1.1115 +	/ mul a[2]*b[4] 
  1.1116 +	mull	%edx
  1.1117 +	addl	%eax,		%ebx
  1.1118 +	movl	4(%esi),	%eax
  1.1119 +	adcl	%edx,		%ecx
  1.1120 +	movl	20(%edi),	%edx
  1.1121 +	adcl	$0,		%ebp
  1.1122 +	/ mul a[1]*b[5] 
  1.1123 +	mull	%edx
  1.1124 +	addl	%eax,		%ebx
  1.1125 +	movl	(%esi),		%eax
  1.1126 +	adcl	%edx,		%ecx
  1.1127 +	movl	24(%edi),	%edx
  1.1128 +	adcl	$0,		%ebp
  1.1129 +	/ mul a[0]*b[6] 
  1.1130 +	mull	%edx
  1.1131 +	addl	%eax,		%ebx
  1.1132 +	movl	20(%esp),	%eax
  1.1133 +	adcl	%edx,		%ecx
  1.1134 +	movl	(%edi),		%edx
  1.1135 +	adcl	$0,		%ebp
  1.1136 +	movl	%ebx,		24(%eax)
  1.1137 +	movl	28(%esi),	%eax
  1.1138 +	/ saved r[6] 
  1.1139 +	/ ################## Calculate word 7 
  1.1140 +	xorl	%ebx,		%ebx
  1.1141 +	/ mul a[7]*b[0] 
  1.1142 +	mull	%edx
  1.1143 +	addl	%eax,		%ecx
  1.1144 +	movl	24(%esi),	%eax
  1.1145 +	adcl	%edx,		%ebp
  1.1146 +	movl	4(%edi),	%edx
  1.1147 +	adcl	$0,		%ebx
  1.1148 +	/ mul a[6]*b[1] 
  1.1149 +	mull	%edx
  1.1150 +	addl	%eax,		%ecx
  1.1151 +	movl	20(%esi),	%eax
  1.1152 +	adcl	%edx,		%ebp
  1.1153 +	movl	8(%edi),	%edx
  1.1154 +	adcl	$0,		%ebx
  1.1155 +	/ mul a[5]*b[2] 
  1.1156 +	mull	%edx
  1.1157 +	addl	%eax,		%ecx
  1.1158 +	movl	16(%esi),	%eax
  1.1159 +	adcl	%edx,		%ebp
  1.1160 +	movl	12(%edi),	%edx
  1.1161 +	adcl	$0,		%ebx
  1.1162 +	/ mul a[4]*b[3] 
  1.1163 +	mull	%edx
  1.1164 +	addl	%eax,		%ecx
  1.1165 +	movl	12(%esi),	%eax
  1.1166 +	adcl	%edx,		%ebp
  1.1167 +	movl	16(%edi),	%edx
  1.1168 +	adcl	$0,		%ebx
  1.1169 +	/ mul a[3]*b[4] 
  1.1170 +	mull	%edx
  1.1171 +	addl	%eax,		%ecx
  1.1172 +	movl	8(%esi),	%eax
  1.1173 +	adcl	%edx,		%ebp
  1.1174 +	movl	20(%edi),	%edx
  1.1175 +	adcl	$0,		%ebx
  1.1176 +	/ mul a[2]*b[5] 
  1.1177 +	mull	%edx
  1.1178 +	addl	%eax,		%ecx
  1.1179 +	movl	4(%esi),	%eax
  1.1180 +	adcl	%edx,		%ebp
  1.1181 +	movl	24(%edi),	%edx
  1.1182 +	adcl	$0,		%ebx
  1.1183 +	/ mul a[1]*b[6] 
  1.1184 +	mull	%edx
  1.1185 +	addl	%eax,		%ecx
  1.1186 +	movl	(%esi),		%eax
  1.1187 +	adcl	%edx,		%ebp
  1.1188 +	movl	28(%edi),	%edx
  1.1189 +	adcl	$0,		%ebx
  1.1190 +	/ mul a[0]*b[7] 
  1.1191 +	mull	%edx
  1.1192 +	addl	%eax,		%ecx
  1.1193 +	movl	20(%esp),	%eax
  1.1194 +	adcl	%edx,		%ebp
  1.1195 +	movl	4(%edi),	%edx
  1.1196 +	adcl	$0,		%ebx
  1.1197 +	movl	%ecx,		28(%eax)
  1.1198 +	movl	28(%esi),	%eax
  1.1199 +	/ saved r[7] 
  1.1200 +	/ ################## Calculate word 8 
  1.1201 +	xorl	%ecx,		%ecx
  1.1202 +	/ mul a[7]*b[1] 
  1.1203 +	mull	%edx
  1.1204 +	addl	%eax,		%ebp
  1.1205 +	movl	24(%esi),	%eax
  1.1206 +	adcl	%edx,		%ebx
  1.1207 +	movl	8(%edi),	%edx
  1.1208 +	adcl	$0,		%ecx
  1.1209 +	/ mul a[6]*b[2] 
  1.1210 +	mull	%edx
  1.1211 +	addl	%eax,		%ebp
  1.1212 +	movl	20(%esi),	%eax
  1.1213 +	adcl	%edx,		%ebx
  1.1214 +	movl	12(%edi),	%edx
  1.1215 +	adcl	$0,		%ecx
  1.1216 +	/ mul a[5]*b[3] 
  1.1217 +	mull	%edx
  1.1218 +	addl	%eax,		%ebp
  1.1219 +	movl	16(%esi),	%eax
  1.1220 +	adcl	%edx,		%ebx
  1.1221 +	movl	16(%edi),	%edx
  1.1222 +	adcl	$0,		%ecx
  1.1223 +	/ mul a[4]*b[4] 
  1.1224 +	mull	%edx
  1.1225 +	addl	%eax,		%ebp
  1.1226 +	movl	12(%esi),	%eax
  1.1227 +	adcl	%edx,		%ebx
  1.1228 +	movl	20(%edi),	%edx
  1.1229 +	adcl	$0,		%ecx
  1.1230 +	/ mul a[3]*b[5] 
  1.1231 +	mull	%edx
  1.1232 +	addl	%eax,		%ebp
  1.1233 +	movl	8(%esi),	%eax
  1.1234 +	adcl	%edx,		%ebx
  1.1235 +	movl	24(%edi),	%edx
  1.1236 +	adcl	$0,		%ecx
  1.1237 +	/ mul a[2]*b[6] 
  1.1238 +	mull	%edx
  1.1239 +	addl	%eax,		%ebp
  1.1240 +	movl	4(%esi),	%eax
  1.1241 +	adcl	%edx,		%ebx
  1.1242 +	movl	28(%edi),	%edx
  1.1243 +	adcl	$0,		%ecx
  1.1244 +	/ mul a[1]*b[7] 
  1.1245 +	mull	%edx
  1.1246 +	addl	%eax,		%ebp
  1.1247 +	movl	20(%esp),	%eax
  1.1248 +	adcl	%edx,		%ebx
  1.1249 +	movl	8(%edi),	%edx
  1.1250 +	adcl	$0,		%ecx
  1.1251 +	movl	%ebp,		32(%eax)
  1.1252 +	movl	28(%esi),	%eax
  1.1253 +	/ saved r[8] 
  1.1254 +	/ ################## Calculate word 9 
  1.1255 +	xorl	%ebp,		%ebp
  1.1256 +	/ mul a[7]*b[2] 
  1.1257 +	mull	%edx
  1.1258 +	addl	%eax,		%ebx
  1.1259 +	movl	24(%esi),	%eax
  1.1260 +	adcl	%edx,		%ecx
  1.1261 +	movl	12(%edi),	%edx
  1.1262 +	adcl	$0,		%ebp
  1.1263 +	/ mul a[6]*b[3] 
  1.1264 +	mull	%edx
  1.1265 +	addl	%eax,		%ebx
  1.1266 +	movl	20(%esi),	%eax
  1.1267 +	adcl	%edx,		%ecx
  1.1268 +	movl	16(%edi),	%edx
  1.1269 +	adcl	$0,		%ebp
  1.1270 +	/ mul a[5]*b[4] 
  1.1271 +	mull	%edx
  1.1272 +	addl	%eax,		%ebx
  1.1273 +	movl	16(%esi),	%eax
  1.1274 +	adcl	%edx,		%ecx
  1.1275 +	movl	20(%edi),	%edx
  1.1276 +	adcl	$0,		%ebp
  1.1277 +	/ mul a[4]*b[5] 
  1.1278 +	mull	%edx
  1.1279 +	addl	%eax,		%ebx
  1.1280 +	movl	12(%esi),	%eax
  1.1281 +	adcl	%edx,		%ecx
  1.1282 +	movl	24(%edi),	%edx
  1.1283 +	adcl	$0,		%ebp
  1.1284 +	/ mul a[3]*b[6] 
  1.1285 +	mull	%edx
  1.1286 +	addl	%eax,		%ebx
  1.1287 +	movl	8(%esi),	%eax
  1.1288 +	adcl	%edx,		%ecx
  1.1289 +	movl	28(%edi),	%edx
  1.1290 +	adcl	$0,		%ebp
  1.1291 +	/ mul a[2]*b[7] 
  1.1292 +	mull	%edx
  1.1293 +	addl	%eax,		%ebx
  1.1294 +	movl	20(%esp),	%eax
  1.1295 +	adcl	%edx,		%ecx
  1.1296 +	movl	12(%edi),	%edx
  1.1297 +	adcl	$0,		%ebp
  1.1298 +	movl	%ebx,		36(%eax)
  1.1299 +	movl	28(%esi),	%eax
  1.1300 +	/ saved r[9] 
  1.1301 +	/ ################## Calculate word 10 
  1.1302 +	xorl	%ebx,		%ebx
  1.1303 +	/ mul a[7]*b[3] 
  1.1304 +	mull	%edx
  1.1305 +	addl	%eax,		%ecx
  1.1306 +	movl	24(%esi),	%eax
  1.1307 +	adcl	%edx,		%ebp
  1.1308 +	movl	16(%edi),	%edx
  1.1309 +	adcl	$0,		%ebx
  1.1310 +	/ mul a[6]*b[4] 
  1.1311 +	mull	%edx
  1.1312 +	addl	%eax,		%ecx
  1.1313 +	movl	20(%esi),	%eax
  1.1314 +	adcl	%edx,		%ebp
  1.1315 +	movl	20(%edi),	%edx
  1.1316 +	adcl	$0,		%ebx
  1.1317 +	/ mul a[5]*b[5] 
  1.1318 +	mull	%edx
  1.1319 +	addl	%eax,		%ecx
  1.1320 +	movl	16(%esi),	%eax
  1.1321 +	adcl	%edx,		%ebp
  1.1322 +	movl	24(%edi),	%edx
  1.1323 +	adcl	$0,		%ebx
  1.1324 +	/ mul a[4]*b[6] 
  1.1325 +	mull	%edx
  1.1326 +	addl	%eax,		%ecx
  1.1327 +	movl	12(%esi),	%eax
  1.1328 +	adcl	%edx,		%ebp
  1.1329 +	movl	28(%edi),	%edx
  1.1330 +	adcl	$0,		%ebx
  1.1331 +	/ mul a[3]*b[7] 
  1.1332 +	mull	%edx
  1.1333 +	addl	%eax,		%ecx
  1.1334 +	movl	20(%esp),	%eax
  1.1335 +	adcl	%edx,		%ebp
  1.1336 +	movl	16(%edi),	%edx
  1.1337 +	adcl	$0,		%ebx
  1.1338 +	movl	%ecx,		40(%eax)
  1.1339 +	movl	28(%esi),	%eax
  1.1340 +	/ saved r[10] 
  1.1341 +	/ ################## Calculate word 11 
  1.1342 +	xorl	%ecx,		%ecx
  1.1343 +	/ mul a[7]*b[4] 
  1.1344 +	mull	%edx
  1.1345 +	addl	%eax,		%ebp
  1.1346 +	movl	24(%esi),	%eax
  1.1347 +	adcl	%edx,		%ebx
  1.1348 +	movl	20(%edi),	%edx
  1.1349 +	adcl	$0,		%ecx
  1.1350 +	/ mul a[6]*b[5] 
  1.1351 +	mull	%edx
  1.1352 +	addl	%eax,		%ebp
  1.1353 +	movl	20(%esi),	%eax
  1.1354 +	adcl	%edx,		%ebx
  1.1355 +	movl	24(%edi),	%edx
  1.1356 +	adcl	$0,		%ecx
  1.1357 +	/ mul a[5]*b[6] 
  1.1358 +	mull	%edx
  1.1359 +	addl	%eax,		%ebp
  1.1360 +	movl	16(%esi),	%eax
  1.1361 +	adcl	%edx,		%ebx
  1.1362 +	movl	28(%edi),	%edx
  1.1363 +	adcl	$0,		%ecx
  1.1364 +	/ mul a[4]*b[7] 
  1.1365 +	mull	%edx
  1.1366 +	addl	%eax,		%ebp
  1.1367 +	movl	20(%esp),	%eax
  1.1368 +	adcl	%edx,		%ebx
  1.1369 +	movl	20(%edi),	%edx
  1.1370 +	adcl	$0,		%ecx
  1.1371 +	movl	%ebp,		44(%eax)
  1.1372 +	movl	28(%esi),	%eax
  1.1373 +	/ saved r[11] 
  1.1374 +	/ ################## Calculate word 12 
  1.1375 +	xorl	%ebp,		%ebp
  1.1376 +	/ mul a[7]*b[5] 
  1.1377 +	mull	%edx
  1.1378 +	addl	%eax,		%ebx
  1.1379 +	movl	24(%esi),	%eax
  1.1380 +	adcl	%edx,		%ecx
  1.1381 +	movl	24(%edi),	%edx
  1.1382 +	adcl	$0,		%ebp
  1.1383 +	/ mul a[6]*b[6] 
  1.1384 +	mull	%edx
  1.1385 +	addl	%eax,		%ebx
  1.1386 +	movl	20(%esi),	%eax
  1.1387 +	adcl	%edx,		%ecx
  1.1388 +	movl	28(%edi),	%edx
  1.1389 +	adcl	$0,		%ebp
  1.1390 +	/ mul a[5]*b[7] 
  1.1391 +	mull	%edx
  1.1392 +	addl	%eax,		%ebx
  1.1393 +	movl	20(%esp),	%eax
  1.1394 +	adcl	%edx,		%ecx
  1.1395 +	movl	24(%edi),	%edx
  1.1396 +	adcl	$0,		%ebp
  1.1397 +	movl	%ebx,		48(%eax)
  1.1398 +	movl	28(%esi),	%eax
  1.1399 +	/ saved r[12] 
  1.1400 +	/ ################## Calculate word 13 
  1.1401 +	xorl	%ebx,		%ebx
  1.1402 +	/ mul a[7]*b[6] 
  1.1403 +	mull	%edx
  1.1404 +	addl	%eax,		%ecx
  1.1405 +	movl	24(%esi),	%eax
  1.1406 +	adcl	%edx,		%ebp
  1.1407 +	movl	28(%edi),	%edx
  1.1408 +	adcl	$0,		%ebx
  1.1409 +	/ mul a[6]*b[7] 
  1.1410 +	mull	%edx
  1.1411 +	addl	%eax,		%ecx
  1.1412 +	movl	20(%esp),	%eax
  1.1413 +	adcl	%edx,		%ebp
  1.1414 +	movl	28(%edi),	%edx
  1.1415 +	adcl	$0,		%ebx
  1.1416 +	movl	%ecx,		52(%eax)
  1.1417 +	movl	28(%esi),	%eax
  1.1418 +	/ saved r[13] 
  1.1419 +	/ ################## Calculate word 14 
  1.1420 +	xorl	%ecx,		%ecx
  1.1421 +	/ mul a[7]*b[7] 
  1.1422 +	mull	%edx
  1.1423 +	addl	%eax,		%ebp
  1.1424 +	movl	20(%esp),	%eax
  1.1425 +	adcl	%edx,		%ebx
  1.1426 +	adcl	$0,		%ecx
  1.1427 +	movl	%ebp,		56(%eax)
  1.1428 +	/ saved r[14] 
  1.1429 +	/ save r[15] 
  1.1430 +	movl	%ebx,		60(%eax)
  1.1431 +	popl	%ebx
  1.1432 +	popl	%ebp
  1.1433 +	popl	%edi
  1.1434 +	popl	%esi
  1.1435 +	ret
  1.1436 +.L_bn_mul_comba8_end:
  1.1437 +	.size	bn_mul_comba8,.L_bn_mul_comba8_end-bn_mul_comba8
  1.1438 +.ident	"desasm.pl"
  1.1439 +.text
  1.1440 +	.align 16
  1.1441 +.globl bn_mul_comba4
  1.1442 +	.type	bn_mul_comba4,@function
  1.1443 +bn_mul_comba4:
  1.1444 +	pushl	%esi
  1.1445 +	movl	12(%esp),	%esi
  1.1446 +	pushl	%edi
  1.1447 +	movl	20(%esp),	%edi
  1.1448 +	pushl	%ebp
  1.1449 +	pushl	%ebx
  1.1450 +	xorl	%ebx,		%ebx
  1.1451 +	movl	(%esi),		%eax
  1.1452 +	xorl	%ecx,		%ecx
  1.1453 +	movl	(%edi),		%edx
  1.1454 +	/ ################## Calculate word 0 
  1.1455 +	xorl	%ebp,		%ebp
  1.1456 +	/ mul a[0]*b[0] 
  1.1457 +	mull	%edx
  1.1458 +	addl	%eax,		%ebx
  1.1459 +	movl	20(%esp),	%eax
  1.1460 +	adcl	%edx,		%ecx
  1.1461 +	movl	(%edi),		%edx
  1.1462 +	adcl	$0,		%ebp
  1.1463 +	movl	%ebx,		(%eax)
  1.1464 +	movl	4(%esi),	%eax
  1.1465 +	/ saved r[0] 
  1.1466 +	/ ################## Calculate word 1 
  1.1467 +	xorl	%ebx,		%ebx
  1.1468 +	/ mul a[1]*b[0] 
  1.1469 +	mull	%edx
  1.1470 +	addl	%eax,		%ecx
  1.1471 +	movl	(%esi),		%eax
  1.1472 +	adcl	%edx,		%ebp
  1.1473 +	movl	4(%edi),	%edx
  1.1474 +	adcl	$0,		%ebx
  1.1475 +	/ mul a[0]*b[1] 
  1.1476 +	mull	%edx
  1.1477 +	addl	%eax,		%ecx
  1.1478 +	movl	20(%esp),	%eax
  1.1479 +	adcl	%edx,		%ebp
  1.1480 +	movl	(%edi),		%edx
  1.1481 +	adcl	$0,		%ebx
  1.1482 +	movl	%ecx,		4(%eax)
  1.1483 +	movl	8(%esi),	%eax
  1.1484 +	/ saved r[1] 
  1.1485 +	/ ################## Calculate word 2 
  1.1486 +	xorl	%ecx,		%ecx
  1.1487 +	/ mul a[2]*b[0] 
  1.1488 +	mull	%edx
  1.1489 +	addl	%eax,		%ebp
  1.1490 +	movl	4(%esi),	%eax
  1.1491 +	adcl	%edx,		%ebx
  1.1492 +	movl	4(%edi),	%edx
  1.1493 +	adcl	$0,		%ecx
  1.1494 +	/ mul a[1]*b[1] 
  1.1495 +	mull	%edx
  1.1496 +	addl	%eax,		%ebp
  1.1497 +	movl	(%esi),		%eax
  1.1498 +	adcl	%edx,		%ebx
  1.1499 +	movl	8(%edi),	%edx
  1.1500 +	adcl	$0,		%ecx
  1.1501 +	/ mul a[0]*b[2] 
  1.1502 +	mull	%edx
  1.1503 +	addl	%eax,		%ebp
  1.1504 +	movl	20(%esp),	%eax
  1.1505 +	adcl	%edx,		%ebx
  1.1506 +	movl	(%edi),		%edx
  1.1507 +	adcl	$0,		%ecx
  1.1508 +	movl	%ebp,		8(%eax)
  1.1509 +	movl	12(%esi),	%eax
  1.1510 +	/ saved r[2] 
  1.1511 +	/ ################## Calculate word 3 
  1.1512 +	xorl	%ebp,		%ebp
  1.1513 +	/ mul a[3]*b[0] 
  1.1514 +	mull	%edx
  1.1515 +	addl	%eax,		%ebx
  1.1516 +	movl	8(%esi),	%eax
  1.1517 +	adcl	%edx,		%ecx
  1.1518 +	movl	4(%edi),	%edx
  1.1519 +	adcl	$0,		%ebp
  1.1520 +	/ mul a[2]*b[1] 
  1.1521 +	mull	%edx
  1.1522 +	addl	%eax,		%ebx
  1.1523 +	movl	4(%esi),	%eax
  1.1524 +	adcl	%edx,		%ecx
  1.1525 +	movl	8(%edi),	%edx
  1.1526 +	adcl	$0,		%ebp
  1.1527 +	/ mul a[1]*b[2] 
  1.1528 +	mull	%edx
  1.1529 +	addl	%eax,		%ebx
  1.1530 +	movl	(%esi),		%eax
  1.1531 +	adcl	%edx,		%ecx
  1.1532 +	movl	12(%edi),	%edx
  1.1533 +	adcl	$0,		%ebp
  1.1534 +	/ mul a[0]*b[3] 
  1.1535 +	mull	%edx
  1.1536 +	addl	%eax,		%ebx
  1.1537 +	movl	20(%esp),	%eax
  1.1538 +	adcl	%edx,		%ecx
  1.1539 +	movl	4(%edi),	%edx
  1.1540 +	adcl	$0,		%ebp
  1.1541 +	movl	%ebx,		12(%eax)
  1.1542 +	movl	12(%esi),	%eax
  1.1543 +	/ saved r[3] 
  1.1544 +	/ ################## Calculate word 4 
  1.1545 +	xorl	%ebx,		%ebx
  1.1546 +	/ mul a[3]*b[1] 
  1.1547 +	mull	%edx
  1.1548 +	addl	%eax,		%ecx
  1.1549 +	movl	8(%esi),	%eax
  1.1550 +	adcl	%edx,		%ebp
  1.1551 +	movl	8(%edi),	%edx
  1.1552 +	adcl	$0,		%ebx
  1.1553 +	/ mul a[2]*b[2] 
  1.1554 +	mull	%edx
  1.1555 +	addl	%eax,		%ecx
  1.1556 +	movl	4(%esi),	%eax
  1.1557 +	adcl	%edx,		%ebp
  1.1558 +	movl	12(%edi),	%edx
  1.1559 +	adcl	$0,		%ebx
  1.1560 +	/ mul a[1]*b[3] 
  1.1561 +	mull	%edx
  1.1562 +	addl	%eax,		%ecx
  1.1563 +	movl	20(%esp),	%eax
  1.1564 +	adcl	%edx,		%ebp
  1.1565 +	movl	8(%edi),	%edx
  1.1566 +	adcl	$0,		%ebx
  1.1567 +	movl	%ecx,		16(%eax)
  1.1568 +	movl	12(%esi),	%eax
  1.1569 +	/ saved r[4] 
  1.1570 +	/ ################## Calculate word 5 
  1.1571 +	xorl	%ecx,		%ecx
  1.1572 +	/ mul a[3]*b[2] 
  1.1573 +	mull	%edx
  1.1574 +	addl	%eax,		%ebp
  1.1575 +	movl	8(%esi),	%eax
  1.1576 +	adcl	%edx,		%ebx
  1.1577 +	movl	12(%edi),	%edx
  1.1578 +	adcl	$0,		%ecx
  1.1579 +	/ mul a[2]*b[3] 
  1.1580 +	mull	%edx
  1.1581 +	addl	%eax,		%ebp
  1.1582 +	movl	20(%esp),	%eax
  1.1583 +	adcl	%edx,		%ebx
  1.1584 +	movl	12(%edi),	%edx
  1.1585 +	adcl	$0,		%ecx
  1.1586 +	movl	%ebp,		20(%eax)
  1.1587 +	movl	12(%esi),	%eax
  1.1588 +	/ saved r[5] 
  1.1589 +	/ ################## Calculate word 6 
  1.1590 +	xorl	%ebp,		%ebp
  1.1591 +	/ mul a[3]*b[3] 
  1.1592 +	mull	%edx
  1.1593 +	addl	%eax,		%ebx
  1.1594 +	movl	20(%esp),	%eax
  1.1595 +	adcl	%edx,		%ecx
  1.1596 +	adcl	$0,		%ebp
  1.1597 +	movl	%ebx,		24(%eax)
  1.1598 +	/ saved r[6] 
  1.1599 +	/ save r[7] 
  1.1600 +	movl	%ecx,		28(%eax)
  1.1601 +	popl	%ebx
  1.1602 +	popl	%ebp
  1.1603 +	popl	%edi
  1.1604 +	popl	%esi
  1.1605 +	ret
  1.1606 +.L_bn_mul_comba4_end:
  1.1607 +	.size	bn_mul_comba4,.L_bn_mul_comba4_end-bn_mul_comba4
  1.1608 +.ident	"desasm.pl"
  1.1609 +.text
  1.1610 +	.align 16
  1.1611 +.globl bn_sqr_comba8
  1.1612 +	.type	bn_sqr_comba8,@function
  1.1613 +bn_sqr_comba8:
  1.1614 +	pushl	%esi
  1.1615 +	pushl	%edi
  1.1616 +	pushl	%ebp
  1.1617 +	pushl	%ebx
  1.1618 +	movl	20(%esp),	%edi
  1.1619 +	movl	24(%esp),	%esi
  1.1620 +	xorl	%ebx,		%ebx
  1.1621 +	xorl	%ecx,		%ecx
  1.1622 +	movl	(%esi),		%eax
  1.1623 +	/ ############### Calculate word 0 
  1.1624 +	xorl	%ebp,		%ebp
  1.1625 +	/ sqr a[0]*a[0] 
  1.1626 +	mull	%eax
  1.1627 +	addl	%eax,		%ebx
  1.1628 +	adcl	%edx,		%ecx
  1.1629 +	movl	(%esi),		%edx
  1.1630 +	adcl	$0,		%ebp
  1.1631 +	movl	%ebx,		(%edi)
  1.1632 +	movl	4(%esi),	%eax
  1.1633 +	/ saved r[0] 
  1.1634 +	/ ############### Calculate word 1 
  1.1635 +	xorl	%ebx,		%ebx
  1.1636 +	/ sqr a[1]*a[0] 
  1.1637 +	mull	%edx
  1.1638 +	addl	%eax,		%eax
  1.1639 +	adcl	%edx,		%edx
  1.1640 +	adcl	$0,		%ebx
  1.1641 +	addl	%eax,		%ecx
  1.1642 +	adcl	%edx,		%ebp
  1.1643 +	movl	8(%esi),	%eax
  1.1644 +	adcl	$0,		%ebx
  1.1645 +	movl	%ecx,		4(%edi)
  1.1646 +	movl	(%esi),		%edx
  1.1647 +	/ saved r[1] 
  1.1648 +	/ ############### Calculate word 2 
  1.1649 +	xorl	%ecx,		%ecx
  1.1650 +	/ sqr a[2]*a[0] 
  1.1651 +	mull	%edx
  1.1652 +	addl	%eax,		%eax
  1.1653 +	adcl	%edx,		%edx
  1.1654 +	adcl	$0,		%ecx
  1.1655 +	addl	%eax,		%ebp
  1.1656 +	adcl	%edx,		%ebx
  1.1657 +	movl	4(%esi),	%eax
  1.1658 +	adcl	$0,		%ecx
  1.1659 +	/ sqr a[1]*a[1] 
  1.1660 +	mull	%eax
  1.1661 +	addl	%eax,		%ebp
  1.1662 +	adcl	%edx,		%ebx
  1.1663 +	movl	(%esi),		%edx
  1.1664 +	adcl	$0,		%ecx
  1.1665 +	movl	%ebp,		8(%edi)
  1.1666 +	movl	12(%esi),	%eax
  1.1667 +	/ saved r[2] 
  1.1668 +	/ ############### Calculate word 3 
  1.1669 +	xorl	%ebp,		%ebp
  1.1670 +	/ sqr a[3]*a[0] 
  1.1671 +	mull	%edx
  1.1672 +	addl	%eax,		%eax
  1.1673 +	adcl	%edx,		%edx
  1.1674 +	adcl	$0,		%ebp
  1.1675 +	addl	%eax,		%ebx
  1.1676 +	adcl	%edx,		%ecx
  1.1677 +	movl	8(%esi),	%eax
  1.1678 +	adcl	$0,		%ebp
  1.1679 +	movl	4(%esi),	%edx
  1.1680 +	/ sqr a[2]*a[1] 
  1.1681 +	mull	%edx
  1.1682 +	addl	%eax,		%eax
  1.1683 +	adcl	%edx,		%edx
  1.1684 +	adcl	$0,		%ebp
  1.1685 +	addl	%eax,		%ebx
  1.1686 +	adcl	%edx,		%ecx
  1.1687 +	movl	16(%esi),	%eax
  1.1688 +	adcl	$0,		%ebp
  1.1689 +	movl	%ebx,		12(%edi)
  1.1690 +	movl	(%esi),		%edx
  1.1691 +	/ saved r[3] 
  1.1692 +	/ ############### Calculate word 4 
  1.1693 +	xorl	%ebx,		%ebx
  1.1694 +	/ sqr a[4]*a[0] 
  1.1695 +	mull	%edx
  1.1696 +	addl	%eax,		%eax
  1.1697 +	adcl	%edx,		%edx
  1.1698 +	adcl	$0,		%ebx
  1.1699 +	addl	%eax,		%ecx
  1.1700 +	adcl	%edx,		%ebp
  1.1701 +	movl	12(%esi),	%eax
  1.1702 +	adcl	$0,		%ebx
  1.1703 +	movl	4(%esi),	%edx
  1.1704 +	/ sqr a[3]*a[1] 
  1.1705 +	mull	%edx
  1.1706 +	addl	%eax,		%eax
  1.1707 +	adcl	%edx,		%edx
  1.1708 +	adcl	$0,		%ebx
  1.1709 +	addl	%eax,		%ecx
  1.1710 +	adcl	%edx,		%ebp
  1.1711 +	movl	8(%esi),	%eax
  1.1712 +	adcl	$0,		%ebx
  1.1713 +	/ sqr a[2]*a[2] 
  1.1714 +	mull	%eax
  1.1715 +	addl	%eax,		%ecx
  1.1716 +	adcl	%edx,		%ebp
  1.1717 +	movl	(%esi),		%edx
  1.1718 +	adcl	$0,		%ebx
  1.1719 +	movl	%ecx,		16(%edi)
  1.1720 +	movl	20(%esi),	%eax
  1.1721 +	/ saved r[4] 
  1.1722 +	/ ############### Calculate word 5 
  1.1723 +	xorl	%ecx,		%ecx
  1.1724 +	/ sqr a[5]*a[0] 
  1.1725 +	mull	%edx
  1.1726 +	addl	%eax,		%eax
  1.1727 +	adcl	%edx,		%edx
  1.1728 +	adcl	$0,		%ecx
  1.1729 +	addl	%eax,		%ebp
  1.1730 +	adcl	%edx,		%ebx
  1.1731 +	movl	16(%esi),	%eax
  1.1732 +	adcl	$0,		%ecx
  1.1733 +	movl	4(%esi),	%edx
  1.1734 +	/ sqr a[4]*a[1] 
  1.1735 +	mull	%edx
  1.1736 +	addl	%eax,		%eax
  1.1737 +	adcl	%edx,		%edx
  1.1738 +	adcl	$0,		%ecx
  1.1739 +	addl	%eax,		%ebp
  1.1740 +	adcl	%edx,		%ebx
  1.1741 +	movl	12(%esi),	%eax
  1.1742 +	adcl	$0,		%ecx
  1.1743 +	movl	8(%esi),	%edx
  1.1744 +	/ sqr a[3]*a[2] 
  1.1745 +	mull	%edx
  1.1746 +	addl	%eax,		%eax
  1.1747 +	adcl	%edx,		%edx
  1.1748 +	adcl	$0,		%ecx
  1.1749 +	addl	%eax,		%ebp
  1.1750 +	adcl	%edx,		%ebx
  1.1751 +	movl	24(%esi),	%eax
  1.1752 +	adcl	$0,		%ecx
  1.1753 +	movl	%ebp,		20(%edi)
  1.1754 +	movl	(%esi),		%edx
  1.1755 +	/ saved r[5] 
  1.1756 +	/ ############### Calculate word 6 
  1.1757 +	xorl	%ebp,		%ebp
  1.1758 +	/ sqr a[6]*a[0] 
  1.1759 +	mull	%edx
  1.1760 +	addl	%eax,		%eax
  1.1761 +	adcl	%edx,		%edx
  1.1762 +	adcl	$0,		%ebp
  1.1763 +	addl	%eax,		%ebx
  1.1764 +	adcl	%edx,		%ecx
  1.1765 +	movl	20(%esi),	%eax
  1.1766 +	adcl	$0,		%ebp
  1.1767 +	movl	4(%esi),	%edx
  1.1768 +	/ sqr a[5]*a[1] 
  1.1769 +	mull	%edx
  1.1770 +	addl	%eax,		%eax
  1.1771 +	adcl	%edx,		%edx
  1.1772 +	adcl	$0,		%ebp
  1.1773 +	addl	%eax,		%ebx
  1.1774 +	adcl	%edx,		%ecx
  1.1775 +	movl	16(%esi),	%eax
  1.1776 +	adcl	$0,		%ebp
  1.1777 +	movl	8(%esi),	%edx
  1.1778 +	/ sqr a[4]*a[2] 
  1.1779 +	mull	%edx
  1.1780 +	addl	%eax,		%eax
  1.1781 +	adcl	%edx,		%edx
  1.1782 +	adcl	$0,		%ebp
  1.1783 +	addl	%eax,		%ebx
  1.1784 +	adcl	%edx,		%ecx
  1.1785 +	movl	12(%esi),	%eax
  1.1786 +	adcl	$0,		%ebp
  1.1787 +	/ sqr a[3]*a[3] 
  1.1788 +	mull	%eax
  1.1789 +	addl	%eax,		%ebx
  1.1790 +	adcl	%edx,		%ecx
  1.1791 +	movl	(%esi),		%edx
  1.1792 +	adcl	$0,		%ebp
  1.1793 +	movl	%ebx,		24(%edi)
  1.1794 +	movl	28(%esi),	%eax
  1.1795 +	/ saved r[6] 
  1.1796 +	/ ############### Calculate word 7 
  1.1797 +	xorl	%ebx,		%ebx
  1.1798 +	/ sqr a[7]*a[0] 
  1.1799 +	mull	%edx
  1.1800 +	addl	%eax,		%eax
  1.1801 +	adcl	%edx,		%edx
  1.1802 +	adcl	$0,		%ebx
  1.1803 +	addl	%eax,		%ecx
  1.1804 +	adcl	%edx,		%ebp
  1.1805 +	movl	24(%esi),	%eax
  1.1806 +	adcl	$0,		%ebx
  1.1807 +	movl	4(%esi),	%edx
  1.1808 +	/ sqr a[6]*a[1] 
  1.1809 +	mull	%edx
  1.1810 +	addl	%eax,		%eax
  1.1811 +	adcl	%edx,		%edx
  1.1812 +	adcl	$0,		%ebx
  1.1813 +	addl	%eax,		%ecx
  1.1814 +	adcl	%edx,		%ebp
  1.1815 +	movl	20(%esi),	%eax
  1.1816 +	adcl	$0,		%ebx
  1.1817 +	movl	8(%esi),	%edx
  1.1818 +	/ sqr a[5]*a[2] 
  1.1819 +	mull	%edx
  1.1820 +	addl	%eax,		%eax
  1.1821 +	adcl	%edx,		%edx
  1.1822 +	adcl	$0,		%ebx
  1.1823 +	addl	%eax,		%ecx
  1.1824 +	adcl	%edx,		%ebp
  1.1825 +	movl	16(%esi),	%eax
  1.1826 +	adcl	$0,		%ebx
  1.1827 +	movl	12(%esi),	%edx
  1.1828 +	/ sqr a[4]*a[3] 
  1.1829 +	mull	%edx
  1.1830 +	addl	%eax,		%eax
  1.1831 +	adcl	%edx,		%edx
  1.1832 +	adcl	$0,		%ebx
  1.1833 +	addl	%eax,		%ecx
  1.1834 +	adcl	%edx,		%ebp
  1.1835 +	movl	28(%esi),	%eax
  1.1836 +	adcl	$0,		%ebx
  1.1837 +	movl	%ecx,		28(%edi)
  1.1838 +	movl	4(%esi),	%edx
  1.1839 +	/ saved r[7] 
  1.1840 +	/ ############### Calculate word 8 
  1.1841 +	xorl	%ecx,		%ecx
  1.1842 +	/ sqr a[7]*a[1] 
  1.1843 +	mull	%edx
  1.1844 +	addl	%eax,		%eax
  1.1845 +	adcl	%edx,		%edx
  1.1846 +	adcl	$0,		%ecx
  1.1847 +	addl	%eax,		%ebp
  1.1848 +	adcl	%edx,		%ebx
  1.1849 +	movl	24(%esi),	%eax
  1.1850 +	adcl	$0,		%ecx
  1.1851 +	movl	8(%esi),	%edx
  1.1852 +	/ sqr a[6]*a[2] 
  1.1853 +	mull	%edx
  1.1854 +	addl	%eax,		%eax
  1.1855 +	adcl	%edx,		%edx
  1.1856 +	adcl	$0,		%ecx
  1.1857 +	addl	%eax,		%ebp
  1.1858 +	adcl	%edx,		%ebx
  1.1859 +	movl	20(%esi),	%eax
  1.1860 +	adcl	$0,		%ecx
  1.1861 +	movl	12(%esi),	%edx
  1.1862 +	/ sqr a[5]*a[3] 
  1.1863 +	mull	%edx
  1.1864 +	addl	%eax,		%eax
  1.1865 +	adcl	%edx,		%edx
  1.1866 +	adcl	$0,		%ecx
  1.1867 +	addl	%eax,		%ebp
  1.1868 +	adcl	%edx,		%ebx
  1.1869 +	movl	16(%esi),	%eax
  1.1870 +	adcl	$0,		%ecx
  1.1871 +	/ sqr a[4]*a[4] 
  1.1872 +	mull	%eax
  1.1873 +	addl	%eax,		%ebp
  1.1874 +	adcl	%edx,		%ebx
  1.1875 +	movl	8(%esi),	%edx
  1.1876 +	adcl	$0,		%ecx
  1.1877 +	movl	%ebp,		32(%edi)
  1.1878 +	movl	28(%esi),	%eax
  1.1879 +	/ saved r[8] 
  1.1880 +	/ ############### Calculate word 9 
  1.1881 +	xorl	%ebp,		%ebp
  1.1882 +	/ sqr a[7]*a[2] 
  1.1883 +	mull	%edx
  1.1884 +	addl	%eax,		%eax
  1.1885 +	adcl	%edx,		%edx
  1.1886 +	adcl	$0,		%ebp
  1.1887 +	addl	%eax,		%ebx
  1.1888 +	adcl	%edx,		%ecx
  1.1889 +	movl	24(%esi),	%eax
  1.1890 +	adcl	$0,		%ebp
  1.1891 +	movl	12(%esi),	%edx
  1.1892 +	/ sqr a[6]*a[3] 
  1.1893 +	mull	%edx
  1.1894 +	addl	%eax,		%eax
  1.1895 +	adcl	%edx,		%edx
  1.1896 +	adcl	$0,		%ebp
  1.1897 +	addl	%eax,		%ebx
  1.1898 +	adcl	%edx,		%ecx
  1.1899 +	movl	20(%esi),	%eax
  1.1900 +	adcl	$0,		%ebp
  1.1901 +	movl	16(%esi),	%edx
  1.1902 +	/ sqr a[5]*a[4] 
  1.1903 +	mull	%edx
  1.1904 +	addl	%eax,		%eax
  1.1905 +	adcl	%edx,		%edx
  1.1906 +	adcl	$0,		%ebp
  1.1907 +	addl	%eax,		%ebx
  1.1908 +	adcl	%edx,		%ecx
  1.1909 +	movl	28(%esi),	%eax
  1.1910 +	adcl	$0,		%ebp
  1.1911 +	movl	%ebx,		36(%edi)
  1.1912 +	movl	12(%esi),	%edx
  1.1913 +	/ saved r[9] 
  1.1914 +	/ ############### Calculate word 10 
  1.1915 +	xorl	%ebx,		%ebx
  1.1916 +	/ sqr a[7]*a[3] 
  1.1917 +	mull	%edx
  1.1918 +	addl	%eax,		%eax
  1.1919 +	adcl	%edx,		%edx
  1.1920 +	adcl	$0,		%ebx
  1.1921 +	addl	%eax,		%ecx
  1.1922 +	adcl	%edx,		%ebp
  1.1923 +	movl	24(%esi),	%eax
  1.1924 +	adcl	$0,		%ebx
  1.1925 +	movl	16(%esi),	%edx
  1.1926 +	/ sqr a[6]*a[4] 
  1.1927 +	mull	%edx
  1.1928 +	addl	%eax,		%eax
  1.1929 +	adcl	%edx,		%edx
  1.1930 +	adcl	$0,		%ebx
  1.1931 +	addl	%eax,		%ecx
  1.1932 +	adcl	%edx,		%ebp
  1.1933 +	movl	20(%esi),	%eax
  1.1934 +	adcl	$0,		%ebx
  1.1935 +	/ sqr a[5]*a[5] 
  1.1936 +	mull	%eax
  1.1937 +	addl	%eax,		%ecx
  1.1938 +	adcl	%edx,		%ebp
  1.1939 +	movl	16(%esi),	%edx
  1.1940 +	adcl	$0,		%ebx
  1.1941 +	movl	%ecx,		40(%edi)
  1.1942 +	movl	28(%esi),	%eax
  1.1943 +	/ saved r[10] 
  1.1944 +	/ ############### Calculate word 11 
  1.1945 +	xorl	%ecx,		%ecx
  1.1946 +	/ sqr a[7]*a[4] 
  1.1947 +	mull	%edx
  1.1948 +	addl	%eax,		%eax
  1.1949 +	adcl	%edx,		%edx
  1.1950 +	adcl	$0,		%ecx
  1.1951 +	addl	%eax,		%ebp
  1.1952 +	adcl	%edx,		%ebx
  1.1953 +	movl	24(%esi),	%eax
  1.1954 +	adcl	$0,		%ecx
  1.1955 +	movl	20(%esi),	%edx
  1.1956 +	/ sqr a[6]*a[5] 
  1.1957 +	mull	%edx
  1.1958 +	addl	%eax,		%eax
  1.1959 +	adcl	%edx,		%edx
  1.1960 +	adcl	$0,		%ecx
  1.1961 +	addl	%eax,		%ebp
  1.1962 +	adcl	%edx,		%ebx
  1.1963 +	movl	28(%esi),	%eax
  1.1964 +	adcl	$0,		%ecx
  1.1965 +	movl	%ebp,		44(%edi)
  1.1966 +	movl	20(%esi),	%edx
  1.1967 +	/ saved r[11] 
  1.1968 +	/ ############### Calculate word 12 
  1.1969 +	xorl	%ebp,		%ebp
  1.1970 +	/ sqr a[7]*a[5] 
  1.1971 +	mull	%edx
  1.1972 +	addl	%eax,		%eax
  1.1973 +	adcl	%edx,		%edx
  1.1974 +	adcl	$0,		%ebp
  1.1975 +	addl	%eax,		%ebx
  1.1976 +	adcl	%edx,		%ecx
  1.1977 +	movl	24(%esi),	%eax
  1.1978 +	adcl	$0,		%ebp
  1.1979 +	/ sqr a[6]*a[6] 
  1.1980 +	mull	%eax
  1.1981 +	addl	%eax,		%ebx
  1.1982 +	adcl	%edx,		%ecx
  1.1983 +	movl	24(%esi),	%edx
  1.1984 +	adcl	$0,		%ebp
  1.1985 +	movl	%ebx,		48(%edi)
  1.1986 +	movl	28(%esi),	%eax
  1.1987 +	/ saved r[12] 
  1.1988 +	/ ############### Calculate word 13 
  1.1989 +	xorl	%ebx,		%ebx
  1.1990 +	/ sqr a[7]*a[6] 
  1.1991 +	mull	%edx
  1.1992 +	addl	%eax,		%eax
  1.1993 +	adcl	%edx,		%edx
  1.1994 +	adcl	$0,		%ebx
  1.1995 +	addl	%eax,		%ecx
  1.1996 +	adcl	%edx,		%ebp
  1.1997 +	movl	28(%esi),	%eax
  1.1998 +	adcl	$0,		%ebx
  1.1999 +	movl	%ecx,		52(%edi)
  1.2000 +	/ saved r[13] 
  1.2001 +	/ ############### Calculate word 14 
  1.2002 +	xorl	%ecx,		%ecx
  1.2003 +	/ sqr a[7]*a[7] 
  1.2004 +	mull	%eax
  1.2005 +	addl	%eax,		%ebp
  1.2006 +	adcl	%edx,		%ebx
  1.2007 +	adcl	$0,		%ecx
  1.2008 +	movl	%ebp,		56(%edi)
  1.2009 +	/ saved r[14] 
  1.2010 +	movl	%ebx,		60(%edi)
  1.2011 +	popl	%ebx
  1.2012 +	popl	%ebp
  1.2013 +	popl	%edi
  1.2014 +	popl	%esi
  1.2015 +	ret
  1.2016 +.L_bn_sqr_comba8_end:
  1.2017 +	.size	bn_sqr_comba8,.L_bn_sqr_comba8_end-bn_sqr_comba8
  1.2018 +.ident	"desasm.pl"
  1.2019 +.text
  1.2020 +	.align 16
  1.2021 +.globl bn_sqr_comba4
  1.2022 +	.type	bn_sqr_comba4,@function
  1.2023 +bn_sqr_comba4:
  1.2024 +	pushl	%esi
  1.2025 +	pushl	%edi
  1.2026 +	pushl	%ebp
  1.2027 +	pushl	%ebx
  1.2028 +	movl	20(%esp),	%edi
  1.2029 +	movl	24(%esp),	%esi
  1.2030 +	xorl	%ebx,		%ebx
  1.2031 +	xorl	%ecx,		%ecx
  1.2032 +	movl	(%esi),		%eax
  1.2033 +	/ ############### Calculate word 0 
  1.2034 +	xorl	%ebp,		%ebp
  1.2035 +	/ sqr a[0]*a[0] 
  1.2036 +	mull	%eax
  1.2037 +	addl	%eax,		%ebx
  1.2038 +	adcl	%edx,		%ecx
  1.2039 +	movl	(%esi),		%edx
  1.2040 +	adcl	$0,		%ebp
  1.2041 +	movl	%ebx,		(%edi)
  1.2042 +	movl	4(%esi),	%eax
  1.2043 +	/ saved r[0] 
  1.2044 +	/ ############### Calculate word 1 
  1.2045 +	xorl	%ebx,		%ebx
  1.2046 +	/ sqr a[1]*a[0] 
  1.2047 +	mull	%edx
  1.2048 +	addl	%eax,		%eax
  1.2049 +	adcl	%edx,		%edx
  1.2050 +	adcl	$0,		%ebx
  1.2051 +	addl	%eax,		%ecx
  1.2052 +	adcl	%edx,		%ebp
  1.2053 +	movl	8(%esi),	%eax
  1.2054 +	adcl	$0,		%ebx
  1.2055 +	movl	%ecx,		4(%edi)
  1.2056 +	movl	(%esi),		%edx
  1.2057 +	/ saved r[1] 
  1.2058 +	/ ############### Calculate word 2 
  1.2059 +	xorl	%ecx,		%ecx
  1.2060 +	/ sqr a[2]*a[0] 
  1.2061 +	mull	%edx
  1.2062 +	addl	%eax,		%eax
  1.2063 +	adcl	%edx,		%edx
  1.2064 +	adcl	$0,		%ecx
  1.2065 +	addl	%eax,		%ebp
  1.2066 +	adcl	%edx,		%ebx
  1.2067 +	movl	4(%esi),	%eax
  1.2068 +	adcl	$0,		%ecx
  1.2069 +	/ sqr a[1]*a[1] 
  1.2070 +	mull	%eax
  1.2071 +	addl	%eax,		%ebp
  1.2072 +	adcl	%edx,		%ebx
  1.2073 +	movl	(%esi),		%edx
  1.2074 +	adcl	$0,		%ecx
  1.2075 +	movl	%ebp,		8(%edi)
  1.2076 +	movl	12(%esi),	%eax
  1.2077 +	/ saved r[2] 
  1.2078 +	/ ############### Calculate word 3 
  1.2079 +	xorl	%ebp,		%ebp
  1.2080 +	/ sqr a[3]*a[0] 
  1.2081 +	mull	%edx
  1.2082 +	addl	%eax,		%eax
  1.2083 +	adcl	%edx,		%edx
  1.2084 +	adcl	$0,		%ebp
  1.2085 +	addl	%eax,		%ebx
  1.2086 +	adcl	%edx,		%ecx
  1.2087 +	movl	8(%esi),	%eax
  1.2088 +	adcl	$0,		%ebp
  1.2089 +	movl	4(%esi),	%edx
  1.2090 +	/ sqr a[2]*a[1] 
  1.2091 +	mull	%edx
  1.2092 +	addl	%eax,		%eax
  1.2093 +	adcl	%edx,		%edx
  1.2094 +	adcl	$0,		%ebp
  1.2095 +	addl	%eax,		%ebx
  1.2096 +	adcl	%edx,		%ecx
  1.2097 +	movl	12(%esi),	%eax
  1.2098 +	adcl	$0,		%ebp
  1.2099 +	movl	%ebx,		12(%edi)
  1.2100 +	movl	4(%esi),	%edx
  1.2101 +	/ saved r[3] 
  1.2102 +	/ ############### Calculate word 4 
  1.2103 +	xorl	%ebx,		%ebx
  1.2104 +	/ sqr a[3]*a[1] 
  1.2105 +	mull	%edx
  1.2106 +	addl	%eax,		%eax
  1.2107 +	adcl	%edx,		%edx
  1.2108 +	adcl	$0,		%ebx
  1.2109 +	addl	%eax,		%ecx
  1.2110 +	adcl	%edx,		%ebp
  1.2111 +	movl	8(%esi),	%eax
  1.2112 +	adcl	$0,		%ebx
  1.2113 +	/ sqr a[2]*a[2] 
  1.2114 +	mull	%eax
  1.2115 +	addl	%eax,		%ecx
  1.2116 +	adcl	%edx,		%ebp
  1.2117 +	movl	8(%esi),	%edx
  1.2118 +	adcl	$0,		%ebx
  1.2119 +	movl	%ecx,		16(%edi)
  1.2120 +	movl	12(%esi),	%eax
  1.2121 +	/ saved r[4] 
  1.2122 +	/ ############### Calculate word 5 
  1.2123 +	xorl	%ecx,		%ecx
  1.2124 +	/ sqr a[3]*a[2] 
  1.2125 +	mull	%edx
  1.2126 +	addl	%eax,		%eax
  1.2127 +	adcl	%edx,		%edx
  1.2128 +	adcl	$0,		%ecx
  1.2129 +	addl	%eax,		%ebp
  1.2130 +	adcl	%edx,		%ebx
  1.2131 +	movl	12(%esi),	%eax
  1.2132 +	adcl	$0,		%ecx
  1.2133 +	movl	%ebp,		20(%edi)
  1.2134 +	/ saved r[5] 
  1.2135 +	/ ############### Calculate word 6 
  1.2136 +	xorl	%ebp,		%ebp
  1.2137 +	/ sqr a[3]*a[3] 
  1.2138 +	mull	%eax
  1.2139 +	addl	%eax,		%ebx
  1.2140 +	adcl	%edx,		%ecx
  1.2141 +	adcl	$0,		%ebp
  1.2142 +	movl	%ebx,		24(%edi)
  1.2143 +	/ saved r[6] 
  1.2144 +	movl	%ecx,		28(%edi)
  1.2145 +	popl	%ebx
  1.2146 +	popl	%ebp
  1.2147 +	popl	%edi
  1.2148 +	popl	%esi
  1.2149 +	ret
  1.2150 +.L_bn_sqr_comba4_end:
  1.2151 +	.size	bn_sqr_comba4,.L_bn_sqr_comba4_end-bn_sqr_comba4
  1.2152 +.ident	"desasm.pl"

     2.1 --- a/cryptlib/cryptlib.spec	Thu Jan 08 11:09:42 2009 +0100
     2.2 +++ b/cryptlib/cryptlib.spec	Thu Jan 08 11:10:35 2009 +0100
     2.3 @@ -37,17 +37,18 @@
     2.4  Group:        Cryptography
     2.5  License:      MIT-style
     2.6  Version:      %{V_full}
     2.7 -Release:      20080704
     2.8 +Release:      20090106
     2.9  
    2.10  #   list of sources
    2.11  Source0:      ftp://ftp.franken.de/pub/crypt/cryptlib/cl%{V_comp}.zip
    2.12  Source1:      ftp://ftp.franken.de/pub/crypt/cryptlib/manual.pdf
    2.13 +Source2:      bn-sol.s
    2.14  Patch0:       cryptlib.patch
    2.15  
    2.16  #   build information
    2.17  Prefix:       %{l_prefix}
    2.18  BuildRoot:    %{l_buildroot}
    2.19 -BuildPreReq:  OpenPKG, openpkg >= 20040130, infozip, make, gcc, perl
    2.20 +BuildPreReq:  OpenPKG, openpkg >= 20040130, infozip, make, gcc
    2.21  PreReq:       OpenPKG, openpkg >= 20040130
    2.22  AutoReq:      no
    2.23  AutoReqProv:  no
    2.24 @@ -79,11 +80,14 @@
    2.25  %prep
    2.26      %setup -q -c -T
    2.27      %{l_prefix}/bin/unzip -q -x %{SOURCE0}
    2.28 +    cp %{SOURCE bn-sol.s} bn/
    2.29      %patch -p0
    2.30 -    %{l_prefix}/bin/perl -p -i -e 's;\r$;;m' tools/*.pl tools/*.sh
    2.31 +    %{l_shtool} subst -e 's;\r$;;' tools/*.pl tools/*.sh
    2.32 +    %{l_shtool} subst -e 's;-xO3 *;;g' 's; *-O;;g' makefile
    2.33 +    %{l_shtool} subst -e 's; *-O[1234];;g' 's; *-O;;g' makefile
    2.34  
    2.35  %build
    2.36 -    %{l_make} %{l_mflags} \
    2.37 +    %{l_make} %{l_mflags -O} \
    2.38          CC="%{l_cc} %{l_cflags -O}"
    2.39  
    2.40  %install