| blob | 1569b2cff1738b661f73f9aa5808299406c18caa |
1 #if defined(__i386__)
2 .file "src/crypto/bn/asm/x86-mont.S"
3 .text
4 .globl bn_mul_mont
5 .hidden bn_mul_mont
6 .type bn_mul_mont,@function
7 .align 16
8 bn_mul_mont:
9 .L_bn_mul_mont_begin:
10 pushl %ebp
11 pushl %ebx
12 pushl %esi
13 pushl %edi
14 xorl %eax,%eax
15 movl 40(%esp),%edi
16 cmpl $4,%edi
17 jl .L000just_leave
18 leal 20(%esp),%esi
19 leal 24(%esp),%edx
20 movl %esp,%ebp
21 addl $2,%edi
22 negl %edi
23 leal -32(%esp,%edi,4),%esp
24 negl %edi
25 movl %esp,%eax
26 subl %edx,%eax
27 andl $2047,%eax
28 subl %eax,%esp
29 xorl %esp,%edx
30 andl $2048,%edx
31 xorl $2048,%edx
32 subl %edx,%esp
33 andl $-64,%esp
34 movl (%esi),%eax
35 movl 4(%esi),%ebx
36 movl 8(%esi),%ecx
37 movl 12(%esi),%edx
38 movl 16(%esi),%esi
39 movl (%esi),%esi
40 movl %eax,4(%esp)
41 movl %ebx,8(%esp)
42 movl %ecx,12(%esp)
43 movl %edx,16(%esp)
44 movl %esi,20(%esp)
45 leal -3(%edi),%ebx
46 movl %ebp,24(%esp)
47 call .L001PIC_me_up
48 .L001PIC_me_up:
49 popl %eax
50 leal OPENSSL_ia32cap_P-.L001PIC_me_up(%eax),%eax
51 btl $26,(%eax)
52 jnc .L002non_sse2
53 movl $-1,%eax
54 movd %eax,%mm7
55 movl 8(%esp),%esi
56 movl 12(%esp),%edi
57 movl 16(%esp),%ebp
58 xorl %edx,%edx
59 xorl %ecx,%ecx
60 movd (%edi),%mm4
61 movd (%esi),%mm5
62 movd (%ebp),%mm3
63 pmuludq %mm4,%mm5
64 movq %mm5,%mm2
65 movq %mm5,%mm0
66 pand %mm7,%mm0
67 pmuludq 20(%esp),%mm5
68 pmuludq %mm5,%mm3
69 paddq %mm0,%mm3
70 movd 4(%ebp),%mm1
71 movd 4(%esi),%mm0
72 psrlq $32,%mm2
73 psrlq $32,%mm3
74 incl %ecx
75 .align 16
76 .L0031st:
77 pmuludq %mm4,%mm0
78 pmuludq %mm5,%mm1
79 paddq %mm0,%mm2
80 paddq %mm1,%mm3
81 movq %mm2,%mm0
82 pand %mm7,%mm0
83 movd 4(%ebp,%ecx,4),%mm1
84 paddq %mm0,%mm3
85 movd 4(%esi,%ecx,4),%mm0
86 psrlq $32,%mm2
87 movd %mm3,28(%esp,%ecx,4)
88 psrlq $32,%mm3
89 leal 1(%ecx),%ecx
90 cmpl %ebx,%ecx
91 jl .L0031st
92 pmuludq %mm4,%mm0
93 pmuludq %mm5,%mm1
94 paddq %mm0,%mm2
95 paddq %mm1,%mm3
96 movq %mm2,%mm0
97 pand %mm7,%mm0
98 paddq %mm0,%mm3
99 movd %mm3,28(%esp,%ecx,4)
100 psrlq $32,%mm2
101 psrlq $32,%mm3
102 paddq %mm2,%mm3
103 movq %mm3,32(%esp,%ebx,4)
104 incl %edx
105 .L004outer:
106 xorl %ecx,%ecx
107 movd (%edi,%edx,4),%mm4
108 movd (%esi),%mm5
109 movd 32(%esp),%mm6
110 movd (%ebp),%mm3
111 pmuludq %mm4,%mm5
112 paddq %mm6,%mm5
113 movq %mm5,%mm0
114 movq %mm5,%mm2
115 pand %mm7,%mm0
116 pmuludq 20(%esp),%mm5
117 pmuludq %mm5,%mm3
118 paddq %mm0,%mm3
119 movd 36(%esp),%mm6
120 movd 4(%ebp),%mm1
121 movd 4(%esi),%mm0
122 psrlq $32,%mm2
123 psrlq $32,%mm3
124 paddq %mm6,%mm2
125 incl %ecx
126 decl %ebx
127 .L005inner:
128 pmuludq %mm4,%mm0
129 pmuludq %mm5,%mm1
130 paddq %mm0,%mm2
131 paddq %mm1,%mm3
132 movq %mm2,%mm0
133 movd 36(%esp,%ecx,4),%mm6
134 pand %mm7,%mm0
135 movd 4(%ebp,%ecx,4),%mm1
136 paddq %mm0,%mm3
137 movd 4(%esi,%ecx,4),%mm0
138 psrlq $32,%mm2
139 movd %mm3,28(%esp,%ecx,4)
140 psrlq $32,%mm3
141 paddq %mm6,%mm2
142 decl %ebx
143 leal 1(%ecx),%ecx
144 jnz .L005inner
145 movl %ecx,%ebx
146 pmuludq %mm4,%mm0
147 pmuludq %mm5,%mm1
148 paddq %mm0,%mm2
149 paddq %mm1,%mm3
150 movq %mm2,%mm0
151 pand %mm7,%mm0
152 paddq %mm0,%mm3
153 movd %mm3,28(%esp,%ecx,4)
154 psrlq $32,%mm2
155 psrlq $32,%mm3
156 movd 36(%esp,%ebx,4),%mm6
157 paddq %mm2,%mm3
158 paddq %mm6,%mm3
159 movq %mm3,32(%esp,%ebx,4)
160 leal 1(%edx),%edx
161 cmpl %ebx,%edx
162 jle .L004outer
163 emms
164 jmp .L006common_tail
165 .align 16
166 .L002non_sse2:
167 movl 8(%esp),%esi
168 leal 1(%ebx),%ebp
169 movl 12(%esp),%edi
170 xorl %ecx,%ecx
171 movl %esi,%edx
172 andl $1,%ebp
173 subl %edi,%edx
174 leal 4(%edi,%ebx,4),%eax
175 orl %edx,%ebp
176 movl (%edi),%edi
177 jz .L007bn_sqr_mont
178 movl %eax,28(%esp)
179 movl (%esi),%eax
180 xorl %edx,%edx
181 .align 16
182 .L008mull:
183 movl %edx,%ebp
184 mull %edi
185 addl %eax,%ebp
186 leal 1(%ecx),%ecx
187 adcl $0,%edx
188 movl (%esi,%ecx,4),%eax
189 cmpl %ebx,%ecx
190 movl %ebp,28(%esp,%ecx,4)
191 jl .L008mull
192 movl %edx,%ebp
193 mull %edi
194 movl 20(%esp),%edi
195 addl %ebp,%eax
196 movl 16(%esp),%esi
197 adcl $0,%edx
198 imull 32(%esp),%edi
199 movl %eax,32(%esp,%ebx,4)
200 xorl %ecx,%ecx
201 movl %edx,36(%esp,%ebx,4)
202 movl %ecx,40(%esp,%ebx,4)
203 movl (%esi),%eax
204 mull %edi
205 addl 32(%esp),%eax
206 movl 4(%esi),%eax
207 adcl $0,%edx
208 incl %ecx
209 jmp .L0092ndmadd
210 .align 16
211 .L0101stmadd:
212 movl %edx,%ebp
213 mull %edi
214 addl 32(%esp,%ecx,4),%ebp
215 leal 1(%ecx),%ecx
216 adcl $0,%edx
217 addl %eax,%ebp
218 movl (%esi,%ecx,4),%eax
219 adcl $0,%edx
220 cmpl %ebx,%ecx
221 movl %ebp,28(%esp,%ecx,4)
222 jl .L0101stmadd
223 movl %edx,%ebp
224 mull %edi
225 addl 32(%esp,%ebx,4),%eax
226 movl 20(%esp),%edi
227 adcl $0,%edx
228 movl 16(%esp),%esi
229 addl %eax,%ebp
230 adcl $0,%edx
231 imull 32(%esp),%edi
232 xorl %ecx,%ecx
233 addl 36(%esp,%ebx,4),%edx
234 movl %ebp,32(%esp,%ebx,4)
235 adcl $0,%ecx
236 movl (%esi),%eax
237 movl %edx,36(%esp,%ebx,4)
238 movl %ecx,40(%esp,%ebx,4)
239 mull %edi
240 addl 32(%esp),%eax
241 movl 4(%esi),%eax
242 adcl $0,%edx
243 movl $1,%ecx
244 .align 16
245 .L0092ndmadd:
246 movl %edx,%ebp
247 mull %edi
248 addl 32(%esp,%ecx,4),%ebp
249 leal 1(%ecx),%ecx
250 adcl $0,%edx
251 addl %eax,%ebp
252 movl (%esi,%ecx,4),%eax
253 adcl $0,%edx
254 cmpl %ebx,%ecx
255 movl %ebp,24(%esp,%ecx,4)
256 jl .L0092ndmadd
257 movl %edx,%ebp
258 mull %edi
259 addl 32(%esp,%ebx,4),%ebp
260 adcl $0,%edx
261 addl %eax,%ebp
262 adcl $0,%edx
263 movl %ebp,28(%esp,%ebx,4)
264 xorl %eax,%eax
265 movl 12(%esp),%ecx
266 addl 36(%esp,%ebx,4),%edx
267 adcl 40(%esp,%ebx,4),%eax
268 leal 4(%ecx),%ecx
269 movl %edx,32(%esp,%ebx,4)
270 cmpl 28(%esp),%ecx
271 movl %eax,36(%esp,%ebx,4)
272 je .L006common_tail
273 movl (%ecx),%edi
274 movl 8(%esp),%esi
275 movl %ecx,12(%esp)
276 xorl %ecx,%ecx
277 xorl %edx,%edx
278 movl (%esi),%eax
279 jmp .L0101stmadd
280 .align 16
281 .L007bn_sqr_mont:
282 movl %ebx,(%esp)
283 movl %ecx,12(%esp)
284 movl %edi,%eax
285 mull %edi
286 movl %eax,32(%esp)
287 movl %edx,%ebx
288 shrl $1,%edx
289 andl $1,%ebx
290 incl %ecx
291 .align 16
292 .L011sqr:
293 movl (%esi,%ecx,4),%eax
294 movl %edx,%ebp
295 mull %edi
296 addl %ebp,%eax
297 leal 1(%ecx),%ecx
298 adcl $0,%edx
299 leal (%ebx,%eax,2),%ebp
300 shrl $31,%eax
301 cmpl (%esp),%ecx
302 movl %eax,%ebx
303 movl %ebp,28(%esp,%ecx,4)
304 jl .L011sqr
305 movl (%esi,%ecx,4),%eax
306 movl %edx,%ebp
307 mull %edi
308 addl %ebp,%eax
309 movl 20(%esp),%edi
310 adcl $0,%edx
311 movl 16(%esp),%esi
312 leal (%ebx,%eax,2),%ebp
313 imull 32(%esp),%edi
314 shrl $31,%eax
315 movl %ebp,32(%esp,%ecx,4)
316 leal (%eax,%edx,2),%ebp
317 movl (%esi),%eax
318 shrl $31,%edx
319 movl %ebp,36(%esp,%ecx,4)
320 movl %edx,40(%esp,%ecx,4)
321 mull %edi
322 addl 32(%esp),%eax
323 movl %ecx,%ebx
324 adcl $0,%edx
325 movl 4(%esi),%eax
326 movl $1,%ecx
327 .align 16
328 .L0123rdmadd:
329 movl %edx,%ebp
330 mull %edi
331 addl 32(%esp,%ecx,4),%ebp
332 adcl $0,%edx
333 addl %eax,%ebp
334 movl 4(%esi,%ecx,4),%eax
335 adcl $0,%edx
336 movl %ebp,28(%esp,%ecx,4)
337 movl %edx,%ebp
338 mull %edi
339 addl 36(%esp,%ecx,4),%ebp
340 leal 2(%ecx),%ecx
341 adcl $0,%edx
342 addl %eax,%ebp
343 movl (%esi,%ecx,4),%eax
344 adcl $0,%edx
345 cmpl %ebx,%ecx
346 movl %ebp,24(%esp,%ecx,4)
347 jl .L0123rdmadd
348 movl %edx,%ebp
349 mull %edi
350 addl 32(%esp,%ebx,4),%ebp
351 adcl $0,%edx
352 addl %eax,%ebp
353 adcl $0,%edx
354 movl %ebp,28(%esp,%ebx,4)
355 movl 12(%esp),%ecx
356 xorl %eax,%eax
357 movl 8(%esp),%esi
358 addl 36(%esp,%ebx,4),%edx
359 adcl 40(%esp,%ebx,4),%eax
360 movl %edx,32(%esp,%ebx,4)
361 cmpl %ebx,%ecx
362 movl %eax,36(%esp,%ebx,4)
363 je .L006common_tail
364 movl 4(%esi,%ecx,4),%edi
365 leal 1(%ecx),%ecx
366 movl %edi,%eax
367 movl %ecx,12(%esp)
368 mull %edi
369 addl 32(%esp,%ecx,4),%eax
370 adcl $0,%edx
371 movl %eax,32(%esp,%ecx,4)
372 xorl %ebp,%ebp
373 cmpl %ebx,%ecx
374 leal 1(%ecx),%ecx
375 je .L013sqrlast
376 movl %edx,%ebx
377 shrl $1,%edx
378 andl $1,%ebx
379 .align 16
380 .L014sqradd:
381 movl (%esi,%ecx,4),%eax
382 movl %edx,%ebp
383 mull %edi
384 addl %ebp,%eax
385 leal (%eax,%eax,1),%ebp
386 adcl $0,%edx
387 shrl $31,%eax
388 addl 32(%esp,%ecx,4),%ebp
389 leal 1(%ecx),%ecx
390 adcl $0,%eax
391 addl %ebx,%ebp
392 adcl $0,%eax
393 cmpl (%esp),%ecx
394 movl %ebp,28(%esp,%ecx,4)
395 movl %eax,%ebx
396 jle .L014sqradd
397 movl %edx,%ebp
398 addl %edx,%edx
399 shrl $31,%ebp
400 addl %ebx,%edx
401 adcl $0,%ebp
402 .L013sqrlast:
403 movl 20(%esp),%edi
404 movl 16(%esp),%esi
405 imull 32(%esp),%edi
406 addl 32(%esp,%ecx,4),%edx
407 movl (%esi),%eax
408 adcl $0,%ebp
409 movl %edx,32(%esp,%ecx,4)
410 movl %ebp,36(%esp,%ecx,4)
411 mull %edi
412 addl 32(%esp),%eax
413 leal -1(%ecx),%ebx
414 adcl $0,%edx
415 movl $1,%ecx
416 movl 4(%esi),%eax
417 jmp .L0123rdmadd
418 .align 16
419 .L006common_tail:
420 movl 16(%esp),%ebp
421 movl 4(%esp),%edi
422 leal 32(%esp),%esi
423 movl (%esi),%eax
424 movl %ebx,%ecx
425 xorl %edx,%edx
426 .align 16
427 .L015sub:
428 sbbl (%ebp,%edx,4),%eax
429 movl %eax,(%edi,%edx,4)
430 decl %ecx
431 movl 4(%esi,%edx,4),%eax
432 leal 1(%edx),%edx
433 jge .L015sub
434 sbbl $0,%eax
435 .align 16
436 .L016copy:
437 movl (%esi,%ebx,4),%edx
438 movl (%edi,%ebx,4),%ebp
439 xorl %ebp,%edx
440 andl %eax,%edx
441 xorl %ebp,%edx
442 movl %ecx,(%esi,%ebx,4)
443 movl %edx,(%edi,%ebx,4)
444 decl %ebx
445 jge .L016copy
446 movl 24(%esp),%esp
447 movl $1,%eax
448 .L000just_leave:
449 popl %edi
450 popl %esi
451 popl %ebx
452 popl %ebp
453 ret
454 .size bn_mul_mont,.-.L_bn_mul_mont_begin
455 .byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
456 .byte 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
457 .byte 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
458 .byte 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
459 .byte 111,114,103,62,0
460 #endif