| blob | 234034b0a0805dcc71521788516e317da955be7a |
1 #if defined(__i386__)
2 .file "src/crypto/bn/asm/x86-mont.S"
3 .text
4 .globl _bn_mul_mont
5 .private_extern _bn_mul_mont
6 .align 4
7 _bn_mul_mont:
8 L_bn_mul_mont_begin:
9 pushl %ebp
10 pushl %ebx
11 pushl %esi
12 pushl %edi
13 xorl %eax,%eax
14 movl 40(%esp),%edi
15 cmpl $4,%edi
16 jl L000just_leave
17 leal 20(%esp),%esi
18 leal 24(%esp),%edx
19 movl %esp,%ebp
20 addl $2,%edi
21 negl %edi
22 leal -32(%esp,%edi,4),%esp
23 negl %edi
24 movl %esp,%eax
25 subl %edx,%eax
26 andl $2047,%eax
27 subl %eax,%esp
28 xorl %esp,%edx
29 andl $2048,%edx
30 xorl $2048,%edx
31 subl %edx,%esp
32 andl $-64,%esp
33 movl (%esi),%eax
34 movl 4(%esi),%ebx
35 movl 8(%esi),%ecx
36 movl 12(%esi),%edx
37 movl 16(%esi),%esi
38 movl (%esi),%esi
39 movl %eax,4(%esp)
40 movl %ebx,8(%esp)
41 movl %ecx,12(%esp)
42 movl %edx,16(%esp)
43 movl %esi,20(%esp)
44 leal -3(%edi),%ebx
45 movl %ebp,24(%esp)
46 call L001PIC_me_up
47 L001PIC_me_up:
48 popl %eax
49 movl L_OPENSSL_ia32cap_P$non_lazy_ptr-L001PIC_me_up(%eax),%eax
50 btl $26,(%eax)
51 jnc L002non_sse2
52 movl $-1,%eax
53 movd %eax,%mm7
54 movl 8(%esp),%esi
55 movl 12(%esp),%edi
56 movl 16(%esp),%ebp
57 xorl %edx,%edx
58 xorl %ecx,%ecx
59 movd (%edi),%mm4
60 movd (%esi),%mm5
61 movd (%ebp),%mm3
62 pmuludq %mm4,%mm5
63 movq %mm5,%mm2
64 movq %mm5,%mm0
65 pand %mm7,%mm0
66 pmuludq 20(%esp),%mm5
67 pmuludq %mm5,%mm3
68 paddq %mm0,%mm3
69 movd 4(%ebp),%mm1
70 movd 4(%esi),%mm0
71 psrlq $32,%mm2
72 psrlq $32,%mm3
73 incl %ecx
74 .align 4,0x90
75 L0031st:
76 pmuludq %mm4,%mm0
77 pmuludq %mm5,%mm1
78 paddq %mm0,%mm2
79 paddq %mm1,%mm3
80 movq %mm2,%mm0
81 pand %mm7,%mm0
82 movd 4(%ebp,%ecx,4),%mm1
83 paddq %mm0,%mm3
84 movd 4(%esi,%ecx,4),%mm0
85 psrlq $32,%mm2
86 movd %mm3,28(%esp,%ecx,4)
87 psrlq $32,%mm3
88 leal 1(%ecx),%ecx
89 cmpl %ebx,%ecx
90 jl L0031st
91 pmuludq %mm4,%mm0
92 pmuludq %mm5,%mm1
93 paddq %mm0,%mm2
94 paddq %mm1,%mm3
95 movq %mm2,%mm0
96 pand %mm7,%mm0
97 paddq %mm0,%mm3
98 movd %mm3,28(%esp,%ecx,4)
99 psrlq $32,%mm2
100 psrlq $32,%mm3
101 paddq %mm2,%mm3
102 movq %mm3,32(%esp,%ebx,4)
103 incl %edx
104 L004outer:
105 xorl %ecx,%ecx
106 movd (%edi,%edx,4),%mm4
107 movd (%esi),%mm5
108 movd 32(%esp),%mm6
109 movd (%ebp),%mm3
110 pmuludq %mm4,%mm5
111 paddq %mm6,%mm5
112 movq %mm5,%mm0
113 movq %mm5,%mm2
114 pand %mm7,%mm0
115 pmuludq 20(%esp),%mm5
116 pmuludq %mm5,%mm3
117 paddq %mm0,%mm3
118 movd 36(%esp),%mm6
119 movd 4(%ebp),%mm1
120 movd 4(%esi),%mm0
121 psrlq $32,%mm2
122 psrlq $32,%mm3
123 paddq %mm6,%mm2
124 incl %ecx
125 decl %ebx
126 L005inner:
127 pmuludq %mm4,%mm0
128 pmuludq %mm5,%mm1
129 paddq %mm0,%mm2
130 paddq %mm1,%mm3
131 movq %mm2,%mm0
132 movd 36(%esp,%ecx,4),%mm6
133 pand %mm7,%mm0
134 movd 4(%ebp,%ecx,4),%mm1
135 paddq %mm0,%mm3
136 movd 4(%esi,%ecx,4),%mm0
137 psrlq $32,%mm2
138 movd %mm3,28(%esp,%ecx,4)
139 psrlq $32,%mm3
140 paddq %mm6,%mm2
141 decl %ebx
142 leal 1(%ecx),%ecx
143 jnz L005inner
144 movl %ecx,%ebx
145 pmuludq %mm4,%mm0
146 pmuludq %mm5,%mm1
147 paddq %mm0,%mm2
148 paddq %mm1,%mm3
149 movq %mm2,%mm0
150 pand %mm7,%mm0
151 paddq %mm0,%mm3
152 movd %mm3,28(%esp,%ecx,4)
153 psrlq $32,%mm2
154 psrlq $32,%mm3
155 movd 36(%esp,%ebx,4),%mm6
156 paddq %mm2,%mm3
157 paddq %mm6,%mm3
158 movq %mm3,32(%esp,%ebx,4)
159 leal 1(%edx),%edx
160 cmpl %ebx,%edx
161 jle L004outer
162 emms
163 jmp L006common_tail
164 .align 4,0x90
165 L002non_sse2:
166 movl 8(%esp),%esi
167 leal 1(%ebx),%ebp
168 movl 12(%esp),%edi
169 xorl %ecx,%ecx
170 movl %esi,%edx
171 andl $1,%ebp
172 subl %edi,%edx
173 leal 4(%edi,%ebx,4),%eax
174 orl %edx,%ebp
175 movl (%edi),%edi
176 jz L007bn_sqr_mont
177 movl %eax,28(%esp)
178 movl (%esi),%eax
179 xorl %edx,%edx
180 .align 4,0x90
181 L008mull:
182 movl %edx,%ebp
183 mull %edi
184 addl %eax,%ebp
185 leal 1(%ecx),%ecx
186 adcl $0,%edx
187 movl (%esi,%ecx,4),%eax
188 cmpl %ebx,%ecx
189 movl %ebp,28(%esp,%ecx,4)
190 jl L008mull
191 movl %edx,%ebp
192 mull %edi
193 movl 20(%esp),%edi
194 addl %ebp,%eax
195 movl 16(%esp),%esi
196 adcl $0,%edx
197 imull 32(%esp),%edi
198 movl %eax,32(%esp,%ebx,4)
199 xorl %ecx,%ecx
200 movl %edx,36(%esp,%ebx,4)
201 movl %ecx,40(%esp,%ebx,4)
202 movl (%esi),%eax
203 mull %edi
204 addl 32(%esp),%eax
205 movl 4(%esi),%eax
206 adcl $0,%edx
207 incl %ecx
208 jmp L0092ndmadd
209 .align 4,0x90
210 L0101stmadd:
211 movl %edx,%ebp
212 mull %edi
213 addl 32(%esp,%ecx,4),%ebp
214 leal 1(%ecx),%ecx
215 adcl $0,%edx
216 addl %eax,%ebp
217 movl (%esi,%ecx,4),%eax
218 adcl $0,%edx
219 cmpl %ebx,%ecx
220 movl %ebp,28(%esp,%ecx,4)
221 jl L0101stmadd
222 movl %edx,%ebp
223 mull %edi
224 addl 32(%esp,%ebx,4),%eax
225 movl 20(%esp),%edi
226 adcl $0,%edx
227 movl 16(%esp),%esi
228 addl %eax,%ebp
229 adcl $0,%edx
230 imull 32(%esp),%edi
231 xorl %ecx,%ecx
232 addl 36(%esp,%ebx,4),%edx
233 movl %ebp,32(%esp,%ebx,4)
234 adcl $0,%ecx
235 movl (%esi),%eax
236 movl %edx,36(%esp,%ebx,4)
237 movl %ecx,40(%esp,%ebx,4)
238 mull %edi
239 addl 32(%esp),%eax
240 movl 4(%esi),%eax
241 adcl $0,%edx
242 movl $1,%ecx
243 .align 4,0x90
244 L0092ndmadd:
245 movl %edx,%ebp
246 mull %edi
247 addl 32(%esp,%ecx,4),%ebp
248 leal 1(%ecx),%ecx
249 adcl $0,%edx
250 addl %eax,%ebp
251 movl (%esi,%ecx,4),%eax
252 adcl $0,%edx
253 cmpl %ebx,%ecx
254 movl %ebp,24(%esp,%ecx,4)
255 jl L0092ndmadd
256 movl %edx,%ebp
257 mull %edi
258 addl 32(%esp,%ebx,4),%ebp
259 adcl $0,%edx
260 addl %eax,%ebp
261 adcl $0,%edx
262 movl %ebp,28(%esp,%ebx,4)
263 xorl %eax,%eax
264 movl 12(%esp),%ecx
265 addl 36(%esp,%ebx,4),%edx
266 adcl 40(%esp,%ebx,4),%eax
267 leal 4(%ecx),%ecx
268 movl %edx,32(%esp,%ebx,4)
269 cmpl 28(%esp),%ecx
270 movl %eax,36(%esp,%ebx,4)
271 je L006common_tail
272 movl (%ecx),%edi
273 movl 8(%esp),%esi
274 movl %ecx,12(%esp)
275 xorl %ecx,%ecx
276 xorl %edx,%edx
277 movl (%esi),%eax
278 jmp L0101stmadd
279 .align 4,0x90
280 L007bn_sqr_mont:
281 movl %ebx,(%esp)
282 movl %ecx,12(%esp)
283 movl %edi,%eax
284 mull %edi
285 movl %eax,32(%esp)
286 movl %edx,%ebx
287 shrl $1,%edx
288 andl $1,%ebx
289 incl %ecx
290 .align 4,0x90
291 L011sqr:
292 movl (%esi,%ecx,4),%eax
293 movl %edx,%ebp
294 mull %edi
295 addl %ebp,%eax
296 leal 1(%ecx),%ecx
297 adcl $0,%edx
298 leal (%ebx,%eax,2),%ebp
299 shrl $31,%eax
300 cmpl (%esp),%ecx
301 movl %eax,%ebx
302 movl %ebp,28(%esp,%ecx,4)
303 jl L011sqr
304 movl (%esi,%ecx,4),%eax
305 movl %edx,%ebp
306 mull %edi
307 addl %ebp,%eax
308 movl 20(%esp),%edi
309 adcl $0,%edx
310 movl 16(%esp),%esi
311 leal (%ebx,%eax,2),%ebp
312 imull 32(%esp),%edi
313 shrl $31,%eax
314 movl %ebp,32(%esp,%ecx,4)
315 leal (%eax,%edx,2),%ebp
316 movl (%esi),%eax
317 shrl $31,%edx
318 movl %ebp,36(%esp,%ecx,4)
319 movl %edx,40(%esp,%ecx,4)
320 mull %edi
321 addl 32(%esp),%eax
322 movl %ecx,%ebx
323 adcl $0,%edx
324 movl 4(%esi),%eax
325 movl $1,%ecx
326 .align 4,0x90
327 L0123rdmadd:
328 movl %edx,%ebp
329 mull %edi
330 addl 32(%esp,%ecx,4),%ebp
331 adcl $0,%edx
332 addl %eax,%ebp
333 movl 4(%esi,%ecx,4),%eax
334 adcl $0,%edx
335 movl %ebp,28(%esp,%ecx,4)
336 movl %edx,%ebp
337 mull %edi
338 addl 36(%esp,%ecx,4),%ebp
339 leal 2(%ecx),%ecx
340 adcl $0,%edx
341 addl %eax,%ebp
342 movl (%esi,%ecx,4),%eax
343 adcl $0,%edx
344 cmpl %ebx,%ecx
345 movl %ebp,24(%esp,%ecx,4)
346 jl L0123rdmadd
347 movl %edx,%ebp
348 mull %edi
349 addl 32(%esp,%ebx,4),%ebp
350 adcl $0,%edx
351 addl %eax,%ebp
352 adcl $0,%edx
353 movl %ebp,28(%esp,%ebx,4)
354 movl 12(%esp),%ecx
355 xorl %eax,%eax
356 movl 8(%esp),%esi
357 addl 36(%esp,%ebx,4),%edx
358 adcl 40(%esp,%ebx,4),%eax
359 movl %edx,32(%esp,%ebx,4)
360 cmpl %ebx,%ecx
361 movl %eax,36(%esp,%ebx,4)
362 je L006common_tail
363 movl 4(%esi,%ecx,4),%edi
364 leal 1(%ecx),%ecx
365 movl %edi,%eax
366 movl %ecx,12(%esp)
367 mull %edi
368 addl 32(%esp,%ecx,4),%eax
369 adcl $0,%edx
370 movl %eax,32(%esp,%ecx,4)
371 xorl %ebp,%ebp
372 cmpl %ebx,%ecx
373 leal 1(%ecx),%ecx
374 je L013sqrlast
375 movl %edx,%ebx
376 shrl $1,%edx
377 andl $1,%ebx
378 .align 4,0x90
379 L014sqradd:
380 movl (%esi,%ecx,4),%eax
381 movl %edx,%ebp
382 mull %edi
383 addl %ebp,%eax
384 leal (%eax,%eax,1),%ebp
385 adcl $0,%edx
386 shrl $31,%eax
387 addl 32(%esp,%ecx,4),%ebp
388 leal 1(%ecx),%ecx
389 adcl $0,%eax
390 addl %ebx,%ebp
391 adcl $0,%eax
392 cmpl (%esp),%ecx
393 movl %ebp,28(%esp,%ecx,4)
394 movl %eax,%ebx
395 jle L014sqradd
396 movl %edx,%ebp
397 addl %edx,%edx
398 shrl $31,%ebp
399 addl %ebx,%edx
400 adcl $0,%ebp
401 L013sqrlast:
402 movl 20(%esp),%edi
403 movl 16(%esp),%esi
404 imull 32(%esp),%edi
405 addl 32(%esp,%ecx,4),%edx
406 movl (%esi),%eax
407 adcl $0,%ebp
408 movl %edx,32(%esp,%ecx,4)
409 movl %ebp,36(%esp,%ecx,4)
410 mull %edi
411 addl 32(%esp),%eax
412 leal -1(%ecx),%ebx
413 adcl $0,%edx
414 movl $1,%ecx
415 movl 4(%esi),%eax
416 jmp L0123rdmadd
417 .align 4,0x90
418 L006common_tail:
419 movl 16(%esp),%ebp
420 movl 4(%esp),%edi
421 leal 32(%esp),%esi
422 movl (%esi),%eax
423 movl %ebx,%ecx
424 xorl %edx,%edx
425 .align 4,0x90
426 L015sub:
427 sbbl (%ebp,%edx,4),%eax
428 movl %eax,(%edi,%edx,4)
429 decl %ecx
430 movl 4(%esi,%edx,4),%eax
431 leal 1(%edx),%edx
432 jge L015sub
433 sbbl $0,%eax
434 .align 4,0x90
435 L016copy:
436 movl (%esi,%ebx,4),%edx
437 movl (%edi,%ebx,4),%ebp
438 xorl %ebp,%edx
439 andl %eax,%edx
440 xorl %ebp,%edx
441 movl %ecx,(%esi,%ebx,4)
442 movl %edx,(%edi,%ebx,4)
443 decl %ebx
444 jge L016copy
445 movl 24(%esp),%esp
446 movl $1,%eax
447 L000just_leave:
448 popl %edi
449 popl %esi
450 popl %ebx
451 popl %ebp
452 ret
453 .byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
454 .byte 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
455 .byte 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
456 .byte 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
457 .byte 111,114,103,62,0
458 .section __IMPORT,__pointers,non_lazy_symbol_pointers
459 L_OPENSSL_ia32cap_P$non_lazy_ptr:
460 .indirect_symbol _OPENSSL_ia32cap_P
461 .long 0
462 #endif