1 | default rel
|
---|
2 | %define XMMWORD
|
---|
3 | %define YMMWORD
|
---|
4 | %define ZMMWORD
|
---|
5 | section .text code align=64
|
---|
6 |
|
---|
7 |
|
---|
8 | EXTERN OPENSSL_ia32cap_P
|
---|
9 |
|
---|
10 | global bn_mul_mont
|
---|
11 |
|
---|
12 | ALIGN 16
|
---|
13 | bn_mul_mont:
|
---|
14 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
---|
15 | mov QWORD[16+rsp],rsi
|
---|
16 | mov rax,rsp
|
---|
17 | $L$SEH_begin_bn_mul_mont:
|
---|
18 | mov rdi,rcx
|
---|
19 | mov rsi,rdx
|
---|
20 | mov rdx,r8
|
---|
21 | mov rcx,r9
|
---|
22 | mov r8,QWORD[40+rsp]
|
---|
23 | mov r9,QWORD[48+rsp]
|
---|
24 |
|
---|
25 |
|
---|
26 |
|
---|
27 | mov r9d,r9d
|
---|
28 | mov rax,rsp
|
---|
29 |
|
---|
30 | test r9d,3
|
---|
31 | jnz NEAR $L$mul_enter
|
---|
32 | cmp r9d,8
|
---|
33 | jb NEAR $L$mul_enter
|
---|
34 | mov r11d,DWORD[((OPENSSL_ia32cap_P+8))]
|
---|
35 | cmp rdx,rsi
|
---|
36 | jne NEAR $L$mul4x_enter
|
---|
37 | test r9d,7
|
---|
38 | jz NEAR $L$sqr8x_enter
|
---|
39 | jmp NEAR $L$mul4x_enter
|
---|
40 |
|
---|
41 | ALIGN 16
|
---|
42 | $L$mul_enter:
|
---|
43 | push rbx
|
---|
44 |
|
---|
45 | push rbp
|
---|
46 |
|
---|
47 | push r12
|
---|
48 |
|
---|
49 | push r13
|
---|
50 |
|
---|
51 | push r14
|
---|
52 |
|
---|
53 | push r15
|
---|
54 |
|
---|
55 |
|
---|
56 | neg r9
|
---|
57 | mov r11,rsp
|
---|
58 | lea r10,[((-16))+r9*8+rsp]
|
---|
59 | neg r9
|
---|
60 | and r10,-1024
|
---|
61 |
|
---|
62 |
|
---|
63 |
|
---|
64 |
|
---|
65 |
|
---|
66 |
|
---|
67 |
|
---|
68 |
|
---|
69 |
|
---|
70 | sub r11,r10
|
---|
71 | and r11,-4096
|
---|
72 | lea rsp,[r11*1+r10]
|
---|
73 | mov r11,QWORD[rsp]
|
---|
74 | cmp rsp,r10
|
---|
75 | ja NEAR $L$mul_page_walk
|
---|
76 | jmp NEAR $L$mul_page_walk_done
|
---|
77 |
|
---|
78 | ALIGN 16
|
---|
79 | $L$mul_page_walk:
|
---|
80 | lea rsp,[((-4096))+rsp]
|
---|
81 | mov r11,QWORD[rsp]
|
---|
82 | cmp rsp,r10
|
---|
83 | ja NEAR $L$mul_page_walk
|
---|
84 | $L$mul_page_walk_done:
|
---|
85 |
|
---|
86 | mov QWORD[8+r9*8+rsp],rax
|
---|
87 |
|
---|
88 | $L$mul_body:
|
---|
89 | mov r12,rdx
|
---|
90 | mov r8,QWORD[r8]
|
---|
91 | mov rbx,QWORD[r12]
|
---|
92 | mov rax,QWORD[rsi]
|
---|
93 |
|
---|
94 | xor r14,r14
|
---|
95 | xor r15,r15
|
---|
96 |
|
---|
97 | mov rbp,r8
|
---|
98 | mul rbx
|
---|
99 | mov r10,rax
|
---|
100 | mov rax,QWORD[rcx]
|
---|
101 |
|
---|
102 | imul rbp,r10
|
---|
103 | mov r11,rdx
|
---|
104 |
|
---|
105 | mul rbp
|
---|
106 | add r10,rax
|
---|
107 | mov rax,QWORD[8+rsi]
|
---|
108 | adc rdx,0
|
---|
109 | mov r13,rdx
|
---|
110 |
|
---|
111 | lea r15,[1+r15]
|
---|
112 | jmp NEAR $L$1st_enter
|
---|
113 |
|
---|
114 | ALIGN 16
|
---|
115 | $L$1st:
|
---|
116 | add r13,rax
|
---|
117 | mov rax,QWORD[r15*8+rsi]
|
---|
118 | adc rdx,0
|
---|
119 | add r13,r11
|
---|
120 | mov r11,r10
|
---|
121 | adc rdx,0
|
---|
122 | mov QWORD[((-16))+r15*8+rsp],r13
|
---|
123 | mov r13,rdx
|
---|
124 |
|
---|
125 | $L$1st_enter:
|
---|
126 | mul rbx
|
---|
127 | add r11,rax
|
---|
128 | mov rax,QWORD[r15*8+rcx]
|
---|
129 | adc rdx,0
|
---|
130 | lea r15,[1+r15]
|
---|
131 | mov r10,rdx
|
---|
132 |
|
---|
133 | mul rbp
|
---|
134 | cmp r15,r9
|
---|
135 | jne NEAR $L$1st
|
---|
136 |
|
---|
137 | add r13,rax
|
---|
138 | mov rax,QWORD[rsi]
|
---|
139 | adc rdx,0
|
---|
140 | add r13,r11
|
---|
141 | adc rdx,0
|
---|
142 | mov QWORD[((-16))+r15*8+rsp],r13
|
---|
143 | mov r13,rdx
|
---|
144 | mov r11,r10
|
---|
145 |
|
---|
146 | xor rdx,rdx
|
---|
147 | add r13,r11
|
---|
148 | adc rdx,0
|
---|
149 | mov QWORD[((-8))+r9*8+rsp],r13
|
---|
150 | mov QWORD[r9*8+rsp],rdx
|
---|
151 |
|
---|
152 | lea r14,[1+r14]
|
---|
153 | jmp NEAR $L$outer
|
---|
154 | ALIGN 16
|
---|
155 | $L$outer:
|
---|
156 | mov rbx,QWORD[r14*8+r12]
|
---|
157 | xor r15,r15
|
---|
158 | mov rbp,r8
|
---|
159 | mov r10,QWORD[rsp]
|
---|
160 | mul rbx
|
---|
161 | add r10,rax
|
---|
162 | mov rax,QWORD[rcx]
|
---|
163 | adc rdx,0
|
---|
164 |
|
---|
165 | imul rbp,r10
|
---|
166 | mov r11,rdx
|
---|
167 |
|
---|
168 | mul rbp
|
---|
169 | add r10,rax
|
---|
170 | mov rax,QWORD[8+rsi]
|
---|
171 | adc rdx,0
|
---|
172 | mov r10,QWORD[8+rsp]
|
---|
173 | mov r13,rdx
|
---|
174 |
|
---|
175 | lea r15,[1+r15]
|
---|
176 | jmp NEAR $L$inner_enter
|
---|
177 |
|
---|
178 | ALIGN 16
|
---|
179 | $L$inner:
|
---|
180 | add r13,rax
|
---|
181 | mov rax,QWORD[r15*8+rsi]
|
---|
182 | adc rdx,0
|
---|
183 | add r13,r10
|
---|
184 | mov r10,QWORD[r15*8+rsp]
|
---|
185 | adc rdx,0
|
---|
186 | mov QWORD[((-16))+r15*8+rsp],r13
|
---|
187 | mov r13,rdx
|
---|
188 |
|
---|
189 | $L$inner_enter:
|
---|
190 | mul rbx
|
---|
191 | add r11,rax
|
---|
192 | mov rax,QWORD[r15*8+rcx]
|
---|
193 | adc rdx,0
|
---|
194 | add r10,r11
|
---|
195 | mov r11,rdx
|
---|
196 | adc r11,0
|
---|
197 | lea r15,[1+r15]
|
---|
198 |
|
---|
199 | mul rbp
|
---|
200 | cmp r15,r9
|
---|
201 | jne NEAR $L$inner
|
---|
202 |
|
---|
203 | add r13,rax
|
---|
204 | mov rax,QWORD[rsi]
|
---|
205 | adc rdx,0
|
---|
206 | add r13,r10
|
---|
207 | mov r10,QWORD[r15*8+rsp]
|
---|
208 | adc rdx,0
|
---|
209 | mov QWORD[((-16))+r15*8+rsp],r13
|
---|
210 | mov r13,rdx
|
---|
211 |
|
---|
212 | xor rdx,rdx
|
---|
213 | add r13,r11
|
---|
214 | adc rdx,0
|
---|
215 | add r13,r10
|
---|
216 | adc rdx,0
|
---|
217 | mov QWORD[((-8))+r9*8+rsp],r13
|
---|
218 | mov QWORD[r9*8+rsp],rdx
|
---|
219 |
|
---|
220 | lea r14,[1+r14]
|
---|
221 | cmp r14,r9
|
---|
222 | jb NEAR $L$outer
|
---|
223 |
|
---|
224 | xor r14,r14
|
---|
225 | mov rax,QWORD[rsp]
|
---|
226 | mov r15,r9
|
---|
227 |
|
---|
228 | ALIGN 16
|
---|
229 | $L$sub: sbb rax,QWORD[r14*8+rcx]
|
---|
230 | mov QWORD[r14*8+rdi],rax
|
---|
231 | mov rax,QWORD[8+r14*8+rsp]
|
---|
232 | lea r14,[1+r14]
|
---|
233 | dec r15
|
---|
234 | jnz NEAR $L$sub
|
---|
235 |
|
---|
236 | sbb rax,0
|
---|
237 | mov rbx,-1
|
---|
238 | xor rbx,rax
|
---|
239 | xor r14,r14
|
---|
240 | mov r15,r9
|
---|
241 |
|
---|
242 | $L$copy:
|
---|
243 | mov rcx,QWORD[r14*8+rdi]
|
---|
244 | mov rdx,QWORD[r14*8+rsp]
|
---|
245 | and rcx,rbx
|
---|
246 | and rdx,rax
|
---|
247 | mov QWORD[r14*8+rsp],r9
|
---|
248 | or rdx,rcx
|
---|
249 | mov QWORD[r14*8+rdi],rdx
|
---|
250 | lea r14,[1+r14]
|
---|
251 | sub r15,1
|
---|
252 | jnz NEAR $L$copy
|
---|
253 |
|
---|
254 | mov rsi,QWORD[8+r9*8+rsp]
|
---|
255 |
|
---|
256 | mov rax,1
|
---|
257 | mov r15,QWORD[((-48))+rsi]
|
---|
258 |
|
---|
259 | mov r14,QWORD[((-40))+rsi]
|
---|
260 |
|
---|
261 | mov r13,QWORD[((-32))+rsi]
|
---|
262 |
|
---|
263 | mov r12,QWORD[((-24))+rsi]
|
---|
264 |
|
---|
265 | mov rbp,QWORD[((-16))+rsi]
|
---|
266 |
|
---|
267 | mov rbx,QWORD[((-8))+rsi]
|
---|
268 |
|
---|
269 | lea rsp,[rsi]
|
---|
270 |
|
---|
271 | $L$mul_epilogue:
|
---|
272 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
---|
273 | mov rsi,QWORD[16+rsp]
|
---|
274 | DB 0F3h,0C3h ;repret
|
---|
275 |
|
---|
276 | $L$SEH_end_bn_mul_mont:
|
---|
277 |
|
---|
278 | ALIGN 16
|
---|
279 | bn_mul4x_mont:
|
---|
280 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
---|
281 | mov QWORD[16+rsp],rsi
|
---|
282 | mov rax,rsp
|
---|
283 | $L$SEH_begin_bn_mul4x_mont:
|
---|
284 | mov rdi,rcx
|
---|
285 | mov rsi,rdx
|
---|
286 | mov rdx,r8
|
---|
287 | mov rcx,r9
|
---|
288 | mov r8,QWORD[40+rsp]
|
---|
289 | mov r9,QWORD[48+rsp]
|
---|
290 |
|
---|
291 |
|
---|
292 |
|
---|
293 | mov r9d,r9d
|
---|
294 | mov rax,rsp
|
---|
295 |
|
---|
296 | $L$mul4x_enter:
|
---|
297 | and r11d,0x80100
|
---|
298 | cmp r11d,0x80100
|
---|
299 | je NEAR $L$mulx4x_enter
|
---|
300 | push rbx
|
---|
301 |
|
---|
302 | push rbp
|
---|
303 |
|
---|
304 | push r12
|
---|
305 |
|
---|
306 | push r13
|
---|
307 |
|
---|
308 | push r14
|
---|
309 |
|
---|
310 | push r15
|
---|
311 |
|
---|
312 |
|
---|
313 | neg r9
|
---|
314 | mov r11,rsp
|
---|
315 | lea r10,[((-32))+r9*8+rsp]
|
---|
316 | neg r9
|
---|
317 | and r10,-1024
|
---|
318 |
|
---|
319 | sub r11,r10
|
---|
320 | and r11,-4096
|
---|
321 | lea rsp,[r11*1+r10]
|
---|
322 | mov r11,QWORD[rsp]
|
---|
323 | cmp rsp,r10
|
---|
324 | ja NEAR $L$mul4x_page_walk
|
---|
325 | jmp NEAR $L$mul4x_page_walk_done
|
---|
326 |
|
---|
327 | $L$mul4x_page_walk:
|
---|
328 | lea rsp,[((-4096))+rsp]
|
---|
329 | mov r11,QWORD[rsp]
|
---|
330 | cmp rsp,r10
|
---|
331 | ja NEAR $L$mul4x_page_walk
|
---|
332 | $L$mul4x_page_walk_done:
|
---|
333 |
|
---|
334 | mov QWORD[8+r9*8+rsp],rax
|
---|
335 |
|
---|
336 | $L$mul4x_body:
|
---|
337 | mov QWORD[16+r9*8+rsp],rdi
|
---|
338 | mov r12,rdx
|
---|
339 | mov r8,QWORD[r8]
|
---|
340 | mov rbx,QWORD[r12]
|
---|
341 | mov rax,QWORD[rsi]
|
---|
342 |
|
---|
343 | xor r14,r14
|
---|
344 | xor r15,r15
|
---|
345 |
|
---|
346 | mov rbp,r8
|
---|
347 | mul rbx
|
---|
348 | mov r10,rax
|
---|
349 | mov rax,QWORD[rcx]
|
---|
350 |
|
---|
351 | imul rbp,r10
|
---|
352 | mov r11,rdx
|
---|
353 |
|
---|
354 | mul rbp
|
---|
355 | add r10,rax
|
---|
356 | mov rax,QWORD[8+rsi]
|
---|
357 | adc rdx,0
|
---|
358 | mov rdi,rdx
|
---|
359 |
|
---|
360 | mul rbx
|
---|
361 | add r11,rax
|
---|
362 | mov rax,QWORD[8+rcx]
|
---|
363 | adc rdx,0
|
---|
364 | mov r10,rdx
|
---|
365 |
|
---|
366 | mul rbp
|
---|
367 | add rdi,rax
|
---|
368 | mov rax,QWORD[16+rsi]
|
---|
369 | adc rdx,0
|
---|
370 | add rdi,r11
|
---|
371 | lea r15,[4+r15]
|
---|
372 | adc rdx,0
|
---|
373 | mov QWORD[rsp],rdi
|
---|
374 | mov r13,rdx
|
---|
375 | jmp NEAR $L$1st4x
|
---|
376 | ALIGN 16
|
---|
377 | $L$1st4x:
|
---|
378 | mul rbx
|
---|
379 | add r10,rax
|
---|
380 | mov rax,QWORD[((-16))+r15*8+rcx]
|
---|
381 | adc rdx,0
|
---|
382 | mov r11,rdx
|
---|
383 |
|
---|
384 | mul rbp
|
---|
385 | add r13,rax
|
---|
386 | mov rax,QWORD[((-8))+r15*8+rsi]
|
---|
387 | adc rdx,0
|
---|
388 | add r13,r10
|
---|
389 | adc rdx,0
|
---|
390 | mov QWORD[((-24))+r15*8+rsp],r13
|
---|
391 | mov rdi,rdx
|
---|
392 |
|
---|
393 | mul rbx
|
---|
394 | add r11,rax
|
---|
395 | mov rax,QWORD[((-8))+r15*8+rcx]
|
---|
396 | adc rdx,0
|
---|
397 | mov r10,rdx
|
---|
398 |
|
---|
399 | mul rbp
|
---|
400 | add rdi,rax
|
---|
401 | mov rax,QWORD[r15*8+rsi]
|
---|
402 | adc rdx,0
|
---|
403 | add rdi,r11
|
---|
404 | adc rdx,0
|
---|
405 | mov QWORD[((-16))+r15*8+rsp],rdi
|
---|
406 | mov r13,rdx
|
---|
407 |
|
---|
408 | mul rbx
|
---|
409 | add r10,rax
|
---|
410 | mov rax,QWORD[r15*8+rcx]
|
---|
411 | adc rdx,0
|
---|
412 | mov r11,rdx
|
---|
413 |
|
---|
414 | mul rbp
|
---|
415 | add r13,rax
|
---|
416 | mov rax,QWORD[8+r15*8+rsi]
|
---|
417 | adc rdx,0
|
---|
418 | add r13,r10
|
---|
419 | adc rdx,0
|
---|
420 | mov QWORD[((-8))+r15*8+rsp],r13
|
---|
421 | mov rdi,rdx
|
---|
422 |
|
---|
423 | mul rbx
|
---|
424 | add r11,rax
|
---|
425 | mov rax,QWORD[8+r15*8+rcx]
|
---|
426 | adc rdx,0
|
---|
427 | lea r15,[4+r15]
|
---|
428 | mov r10,rdx
|
---|
429 |
|
---|
430 | mul rbp
|
---|
431 | add rdi,rax
|
---|
432 | mov rax,QWORD[((-16))+r15*8+rsi]
|
---|
433 | adc rdx,0
|
---|
434 | add rdi,r11
|
---|
435 | adc rdx,0
|
---|
436 | mov QWORD[((-32))+r15*8+rsp],rdi
|
---|
437 | mov r13,rdx
|
---|
438 | cmp r15,r9
|
---|
439 | jb NEAR $L$1st4x
|
---|
440 |
|
---|
441 | mul rbx
|
---|
442 | add r10,rax
|
---|
443 | mov rax,QWORD[((-16))+r15*8+rcx]
|
---|
444 | adc rdx,0
|
---|
445 | mov r11,rdx
|
---|
446 |
|
---|
447 | mul rbp
|
---|
448 | add r13,rax
|
---|
449 | mov rax,QWORD[((-8))+r15*8+rsi]
|
---|
450 | adc rdx,0
|
---|
451 | add r13,r10
|
---|
452 | adc rdx,0
|
---|
453 | mov QWORD[((-24))+r15*8+rsp],r13
|
---|
454 | mov rdi,rdx
|
---|
455 |
|
---|
456 | mul rbx
|
---|
457 | add r11,rax
|
---|
458 | mov rax,QWORD[((-8))+r15*8+rcx]
|
---|
459 | adc rdx,0
|
---|
460 | mov r10,rdx
|
---|
461 |
|
---|
462 | mul rbp
|
---|
463 | add rdi,rax
|
---|
464 | mov rax,QWORD[rsi]
|
---|
465 | adc rdx,0
|
---|
466 | add rdi,r11
|
---|
467 | adc rdx,0
|
---|
468 | mov QWORD[((-16))+r15*8+rsp],rdi
|
---|
469 | mov r13,rdx
|
---|
470 |
|
---|
471 | xor rdi,rdi
|
---|
472 | add r13,r10
|
---|
473 | adc rdi,0
|
---|
474 | mov QWORD[((-8))+r15*8+rsp],r13
|
---|
475 | mov QWORD[r15*8+rsp],rdi
|
---|
476 |
|
---|
477 | lea r14,[1+r14]
|
---|
478 | ALIGN 4
|
---|
479 | $L$outer4x:
|
---|
480 | mov rbx,QWORD[r14*8+r12]
|
---|
481 | xor r15,r15
|
---|
482 | mov r10,QWORD[rsp]
|
---|
483 | mov rbp,r8
|
---|
484 | mul rbx
|
---|
485 | add r10,rax
|
---|
486 | mov rax,QWORD[rcx]
|
---|
487 | adc rdx,0
|
---|
488 |
|
---|
489 | imul rbp,r10
|
---|
490 | mov r11,rdx
|
---|
491 |
|
---|
492 | mul rbp
|
---|
493 | add r10,rax
|
---|
494 | mov rax,QWORD[8+rsi]
|
---|
495 | adc rdx,0
|
---|
496 | mov rdi,rdx
|
---|
497 |
|
---|
498 | mul rbx
|
---|
499 | add r11,rax
|
---|
500 | mov rax,QWORD[8+rcx]
|
---|
501 | adc rdx,0
|
---|
502 | add r11,QWORD[8+rsp]
|
---|
503 | adc rdx,0
|
---|
504 | mov r10,rdx
|
---|
505 |
|
---|
506 | mul rbp
|
---|
507 | add rdi,rax
|
---|
508 | mov rax,QWORD[16+rsi]
|
---|
509 | adc rdx,0
|
---|
510 | add rdi,r11
|
---|
511 | lea r15,[4+r15]
|
---|
512 | adc rdx,0
|
---|
513 | mov QWORD[rsp],rdi
|
---|
514 | mov r13,rdx
|
---|
515 | jmp NEAR $L$inner4x
|
---|
516 | ALIGN 16
|
---|
517 | $L$inner4x:
|
---|
518 | mul rbx
|
---|
519 | add r10,rax
|
---|
520 | mov rax,QWORD[((-16))+r15*8+rcx]
|
---|
521 | adc rdx,0
|
---|
522 | add r10,QWORD[((-16))+r15*8+rsp]
|
---|
523 | adc rdx,0
|
---|
524 | mov r11,rdx
|
---|
525 |
|
---|
526 | mul rbp
|
---|
527 | add r13,rax
|
---|
528 | mov rax,QWORD[((-8))+r15*8+rsi]
|
---|
529 | adc rdx,0
|
---|
530 | add r13,r10
|
---|
531 | adc rdx,0
|
---|
532 | mov QWORD[((-24))+r15*8+rsp],r13
|
---|
533 | mov rdi,rdx
|
---|
534 |
|
---|
535 | mul rbx
|
---|
536 | add r11,rax
|
---|
537 | mov rax,QWORD[((-8))+r15*8+rcx]
|
---|
538 | adc rdx,0
|
---|
539 | add r11,QWORD[((-8))+r15*8+rsp]
|
---|
540 | adc rdx,0
|
---|
541 | mov r10,rdx
|
---|
542 |
|
---|
543 | mul rbp
|
---|
544 | add rdi,rax
|
---|
545 | mov rax,QWORD[r15*8+rsi]
|
---|
546 | adc rdx,0
|
---|
547 | add rdi,r11
|
---|
548 | adc rdx,0
|
---|
549 | mov QWORD[((-16))+r15*8+rsp],rdi
|
---|
550 | mov r13,rdx
|
---|
551 |
|
---|
552 | mul rbx
|
---|
553 | add r10,rax
|
---|
554 | mov rax,QWORD[r15*8+rcx]
|
---|
555 | adc rdx,0
|
---|
556 | add r10,QWORD[r15*8+rsp]
|
---|
557 | adc rdx,0
|
---|
558 | mov r11,rdx
|
---|
559 |
|
---|
560 | mul rbp
|
---|
561 | add r13,rax
|
---|
562 | mov rax,QWORD[8+r15*8+rsi]
|
---|
563 | adc rdx,0
|
---|
564 | add r13,r10
|
---|
565 | adc rdx,0
|
---|
566 | mov QWORD[((-8))+r15*8+rsp],r13
|
---|
567 | mov rdi,rdx
|
---|
568 |
|
---|
569 | mul rbx
|
---|
570 | add r11,rax
|
---|
571 | mov rax,QWORD[8+r15*8+rcx]
|
---|
572 | adc rdx,0
|
---|
573 | add r11,QWORD[8+r15*8+rsp]
|
---|
574 | adc rdx,0
|
---|
575 | lea r15,[4+r15]
|
---|
576 | mov r10,rdx
|
---|
577 |
|
---|
578 | mul rbp
|
---|
579 | add rdi,rax
|
---|
580 | mov rax,QWORD[((-16))+r15*8+rsi]
|
---|
581 | adc rdx,0
|
---|
582 | add rdi,r11
|
---|
583 | adc rdx,0
|
---|
584 | mov QWORD[((-32))+r15*8+rsp],rdi
|
---|
585 | mov r13,rdx
|
---|
586 | cmp r15,r9
|
---|
587 | jb NEAR $L$inner4x
|
---|
588 |
|
---|
589 | mul rbx
|
---|
590 | add r10,rax
|
---|
591 | mov rax,QWORD[((-16))+r15*8+rcx]
|
---|
592 | adc rdx,0
|
---|
593 | add r10,QWORD[((-16))+r15*8+rsp]
|
---|
594 | adc rdx,0
|
---|
595 | mov r11,rdx
|
---|
596 |
|
---|
597 | mul rbp
|
---|
598 | add r13,rax
|
---|
599 | mov rax,QWORD[((-8))+r15*8+rsi]
|
---|
600 | adc rdx,0
|
---|
601 | add r13,r10
|
---|
602 | adc rdx,0
|
---|
603 | mov QWORD[((-24))+r15*8+rsp],r13
|
---|
604 | mov rdi,rdx
|
---|
605 |
|
---|
606 | mul rbx
|
---|
607 | add r11,rax
|
---|
608 | mov rax,QWORD[((-8))+r15*8+rcx]
|
---|
609 | adc rdx,0
|
---|
610 | add r11,QWORD[((-8))+r15*8+rsp]
|
---|
611 | adc rdx,0
|
---|
612 | lea r14,[1+r14]
|
---|
613 | mov r10,rdx
|
---|
614 |
|
---|
615 | mul rbp
|
---|
616 | add rdi,rax
|
---|
617 | mov rax,QWORD[rsi]
|
---|
618 | adc rdx,0
|
---|
619 | add rdi,r11
|
---|
620 | adc rdx,0
|
---|
621 | mov QWORD[((-16))+r15*8+rsp],rdi
|
---|
622 | mov r13,rdx
|
---|
623 |
|
---|
624 | xor rdi,rdi
|
---|
625 | add r13,r10
|
---|
626 | adc rdi,0
|
---|
627 | add r13,QWORD[r9*8+rsp]
|
---|
628 | adc rdi,0
|
---|
629 | mov QWORD[((-8))+r15*8+rsp],r13
|
---|
630 | mov QWORD[r15*8+rsp],rdi
|
---|
631 |
|
---|
632 | cmp r14,r9
|
---|
633 | jb NEAR $L$outer4x
|
---|
634 | mov rdi,QWORD[16+r9*8+rsp]
|
---|
635 | lea r15,[((-4))+r9]
|
---|
636 | mov rax,QWORD[rsp]
|
---|
637 | mov rdx,QWORD[8+rsp]
|
---|
638 | shr r15,2
|
---|
639 | lea rsi,[rsp]
|
---|
640 | xor r14,r14
|
---|
641 |
|
---|
642 | sub rax,QWORD[rcx]
|
---|
643 | mov rbx,QWORD[16+rsi]
|
---|
644 | mov rbp,QWORD[24+rsi]
|
---|
645 | sbb rdx,QWORD[8+rcx]
|
---|
646 |
|
---|
647 | $L$sub4x:
|
---|
648 | mov QWORD[r14*8+rdi],rax
|
---|
649 | mov QWORD[8+r14*8+rdi],rdx
|
---|
650 | sbb rbx,QWORD[16+r14*8+rcx]
|
---|
651 | mov rax,QWORD[32+r14*8+rsi]
|
---|
652 | mov rdx,QWORD[40+r14*8+rsi]
|
---|
653 | sbb rbp,QWORD[24+r14*8+rcx]
|
---|
654 | mov QWORD[16+r14*8+rdi],rbx
|
---|
655 | mov QWORD[24+r14*8+rdi],rbp
|
---|
656 | sbb rax,QWORD[32+r14*8+rcx]
|
---|
657 | mov rbx,QWORD[48+r14*8+rsi]
|
---|
658 | mov rbp,QWORD[56+r14*8+rsi]
|
---|
659 | sbb rdx,QWORD[40+r14*8+rcx]
|
---|
660 | lea r14,[4+r14]
|
---|
661 | dec r15
|
---|
662 | jnz NEAR $L$sub4x
|
---|
663 |
|
---|
664 | mov QWORD[r14*8+rdi],rax
|
---|
665 | mov rax,QWORD[32+r14*8+rsi]
|
---|
666 | sbb rbx,QWORD[16+r14*8+rcx]
|
---|
667 | mov QWORD[8+r14*8+rdi],rdx
|
---|
668 | sbb rbp,QWORD[24+r14*8+rcx]
|
---|
669 | mov QWORD[16+r14*8+rdi],rbx
|
---|
670 |
|
---|
671 | sbb rax,0
|
---|
672 | mov QWORD[24+r14*8+rdi],rbp
|
---|
673 | pxor xmm0,xmm0
|
---|
674 | DB 102,72,15,110,224
|
---|
675 | pcmpeqd xmm5,xmm5
|
---|
676 | pshufd xmm4,xmm4,0
|
---|
677 | mov r15,r9
|
---|
678 | pxor xmm5,xmm4
|
---|
679 | shr r15,2
|
---|
680 | xor eax,eax
|
---|
681 |
|
---|
682 | jmp NEAR $L$copy4x
|
---|
683 | ALIGN 16
|
---|
684 | $L$copy4x:
|
---|
685 | movdqa xmm1,XMMWORD[rax*1+rsp]
|
---|
686 | movdqu xmm2,XMMWORD[rax*1+rdi]
|
---|
687 | pand xmm1,xmm4
|
---|
688 | pand xmm2,xmm5
|
---|
689 | movdqa xmm3,XMMWORD[16+rax*1+rsp]
|
---|
690 | movdqa XMMWORD[rax*1+rsp],xmm0
|
---|
691 | por xmm1,xmm2
|
---|
692 | movdqu xmm2,XMMWORD[16+rax*1+rdi]
|
---|
693 | movdqu XMMWORD[rax*1+rdi],xmm1
|
---|
694 | pand xmm3,xmm4
|
---|
695 | pand xmm2,xmm5
|
---|
696 | movdqa XMMWORD[16+rax*1+rsp],xmm0
|
---|
697 | por xmm3,xmm2
|
---|
698 | movdqu XMMWORD[16+rax*1+rdi],xmm3
|
---|
699 | lea rax,[32+rax]
|
---|
700 | dec r15
|
---|
701 | jnz NEAR $L$copy4x
|
---|
702 | mov rsi,QWORD[8+r9*8+rsp]
|
---|
703 |
|
---|
704 | mov rax,1
|
---|
705 | mov r15,QWORD[((-48))+rsi]
|
---|
706 |
|
---|
707 | mov r14,QWORD[((-40))+rsi]
|
---|
708 |
|
---|
709 | mov r13,QWORD[((-32))+rsi]
|
---|
710 |
|
---|
711 | mov r12,QWORD[((-24))+rsi]
|
---|
712 |
|
---|
713 | mov rbp,QWORD[((-16))+rsi]
|
---|
714 |
|
---|
715 | mov rbx,QWORD[((-8))+rsi]
|
---|
716 |
|
---|
717 | lea rsp,[rsi]
|
---|
718 |
|
---|
719 | $L$mul4x_epilogue:
|
---|
720 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
---|
721 | mov rsi,QWORD[16+rsp]
|
---|
722 | DB 0F3h,0C3h ;repret
|
---|
723 |
|
---|
724 | $L$SEH_end_bn_mul4x_mont:
|
---|
725 | EXTERN bn_sqrx8x_internal
|
---|
726 | EXTERN bn_sqr8x_internal
|
---|
727 |
|
---|
728 |
|
---|
729 | ALIGN 32
|
---|
730 | bn_sqr8x_mont:
|
---|
731 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
---|
732 | mov QWORD[16+rsp],rsi
|
---|
733 | mov rax,rsp
|
---|
734 | $L$SEH_begin_bn_sqr8x_mont:
|
---|
735 | mov rdi,rcx
|
---|
736 | mov rsi,rdx
|
---|
737 | mov rdx,r8
|
---|
738 | mov rcx,r9
|
---|
739 | mov r8,QWORD[40+rsp]
|
---|
740 | mov r9,QWORD[48+rsp]
|
---|
741 |
|
---|
742 |
|
---|
743 |
|
---|
744 | mov rax,rsp
|
---|
745 |
|
---|
746 | $L$sqr8x_enter:
|
---|
747 | push rbx
|
---|
748 |
|
---|
749 | push rbp
|
---|
750 |
|
---|
751 | push r12
|
---|
752 |
|
---|
753 | push r13
|
---|
754 |
|
---|
755 | push r14
|
---|
756 |
|
---|
757 | push r15
|
---|
758 |
|
---|
759 | $L$sqr8x_prologue:
|
---|
760 |
|
---|
761 | mov r10d,r9d
|
---|
762 | shl r9d,3
|
---|
763 | shl r10,3+2
|
---|
764 | neg r9
|
---|
765 |
|
---|
766 |
|
---|
767 |
|
---|
768 |
|
---|
769 |
|
---|
770 |
|
---|
771 | lea r11,[((-64))+r9*2+rsp]
|
---|
772 | mov rbp,rsp
|
---|
773 | mov r8,QWORD[r8]
|
---|
774 | sub r11,rsi
|
---|
775 | and r11,4095
|
---|
776 | cmp r10,r11
|
---|
777 | jb NEAR $L$sqr8x_sp_alt
|
---|
778 | sub rbp,r11
|
---|
779 | lea rbp,[((-64))+r9*2+rbp]
|
---|
780 | jmp NEAR $L$sqr8x_sp_done
|
---|
781 |
|
---|
782 | ALIGN 32
|
---|
783 | $L$sqr8x_sp_alt:
|
---|
784 | lea r10,[((4096-64))+r9*2]
|
---|
785 | lea rbp,[((-64))+r9*2+rbp]
|
---|
786 | sub r11,r10
|
---|
787 | mov r10,0
|
---|
788 | cmovc r11,r10
|
---|
789 | sub rbp,r11
|
---|
790 | $L$sqr8x_sp_done:
|
---|
791 | and rbp,-64
|
---|
792 | mov r11,rsp
|
---|
793 | sub r11,rbp
|
---|
794 | and r11,-4096
|
---|
795 | lea rsp,[rbp*1+r11]
|
---|
796 | mov r10,QWORD[rsp]
|
---|
797 | cmp rsp,rbp
|
---|
798 | ja NEAR $L$sqr8x_page_walk
|
---|
799 | jmp NEAR $L$sqr8x_page_walk_done
|
---|
800 |
|
---|
801 | ALIGN 16
|
---|
802 | $L$sqr8x_page_walk:
|
---|
803 | lea rsp,[((-4096))+rsp]
|
---|
804 | mov r10,QWORD[rsp]
|
---|
805 | cmp rsp,rbp
|
---|
806 | ja NEAR $L$sqr8x_page_walk
|
---|
807 | $L$sqr8x_page_walk_done:
|
---|
808 |
|
---|
809 | mov r10,r9
|
---|
810 | neg r9
|
---|
811 |
|
---|
812 | mov QWORD[32+rsp],r8
|
---|
813 | mov QWORD[40+rsp],rax
|
---|
814 |
|
---|
815 | $L$sqr8x_body:
|
---|
816 |
|
---|
817 | DB 102,72,15,110,209
|
---|
818 | pxor xmm0,xmm0
|
---|
819 | DB 102,72,15,110,207
|
---|
820 | DB 102,73,15,110,218
|
---|
821 | mov eax,DWORD[((OPENSSL_ia32cap_P+8))]
|
---|
822 | and eax,0x80100
|
---|
823 | cmp eax,0x80100
|
---|
824 | jne NEAR $L$sqr8x_nox
|
---|
825 |
|
---|
826 | call bn_sqrx8x_internal
|
---|
827 |
|
---|
828 |
|
---|
829 |
|
---|
830 |
|
---|
831 | lea rbx,[rcx*1+r8]
|
---|
832 | mov r9,rcx
|
---|
833 | mov rdx,rcx
|
---|
834 | DB 102,72,15,126,207
|
---|
835 | sar rcx,3+2
|
---|
836 | jmp NEAR $L$sqr8x_sub
|
---|
837 |
|
---|
838 | ALIGN 32
|
---|
839 | $L$sqr8x_nox:
|
---|
840 | call bn_sqr8x_internal
|
---|
841 |
|
---|
842 |
|
---|
843 |
|
---|
844 |
|
---|
845 | lea rbx,[r9*1+rdi]
|
---|
846 | mov rcx,r9
|
---|
847 | mov rdx,r9
|
---|
848 | DB 102,72,15,126,207
|
---|
849 | sar rcx,3+2
|
---|
850 | jmp NEAR $L$sqr8x_sub
|
---|
851 |
|
---|
852 | ALIGN 32
|
---|
853 | $L$sqr8x_sub:
|
---|
854 | mov r12,QWORD[rbx]
|
---|
855 | mov r13,QWORD[8+rbx]
|
---|
856 | mov r14,QWORD[16+rbx]
|
---|
857 | mov r15,QWORD[24+rbx]
|
---|
858 | lea rbx,[32+rbx]
|
---|
859 | sbb r12,QWORD[rbp]
|
---|
860 | sbb r13,QWORD[8+rbp]
|
---|
861 | sbb r14,QWORD[16+rbp]
|
---|
862 | sbb r15,QWORD[24+rbp]
|
---|
863 | lea rbp,[32+rbp]
|
---|
864 | mov QWORD[rdi],r12
|
---|
865 | mov QWORD[8+rdi],r13
|
---|
866 | mov QWORD[16+rdi],r14
|
---|
867 | mov QWORD[24+rdi],r15
|
---|
868 | lea rdi,[32+rdi]
|
---|
869 | inc rcx
|
---|
870 | jnz NEAR $L$sqr8x_sub
|
---|
871 |
|
---|
872 | sbb rax,0
|
---|
873 | lea rbx,[r9*1+rbx]
|
---|
874 | lea rdi,[r9*1+rdi]
|
---|
875 |
|
---|
876 | DB 102,72,15,110,200
|
---|
877 | pxor xmm0,xmm0
|
---|
878 | pshufd xmm1,xmm1,0
|
---|
879 | mov rsi,QWORD[40+rsp]
|
---|
880 |
|
---|
881 | jmp NEAR $L$sqr8x_cond_copy
|
---|
882 |
|
---|
883 | ALIGN 32
|
---|
884 | $L$sqr8x_cond_copy:
|
---|
885 | movdqa xmm2,XMMWORD[rbx]
|
---|
886 | movdqa xmm3,XMMWORD[16+rbx]
|
---|
887 | lea rbx,[32+rbx]
|
---|
888 | movdqu xmm4,XMMWORD[rdi]
|
---|
889 | movdqu xmm5,XMMWORD[16+rdi]
|
---|
890 | lea rdi,[32+rdi]
|
---|
891 | movdqa XMMWORD[(-32)+rbx],xmm0
|
---|
892 | movdqa XMMWORD[(-16)+rbx],xmm0
|
---|
893 | movdqa XMMWORD[(-32)+rdx*1+rbx],xmm0
|
---|
894 | movdqa XMMWORD[(-16)+rdx*1+rbx],xmm0
|
---|
895 | pcmpeqd xmm0,xmm1
|
---|
896 | pand xmm2,xmm1
|
---|
897 | pand xmm3,xmm1
|
---|
898 | pand xmm4,xmm0
|
---|
899 | pand xmm5,xmm0
|
---|
900 | pxor xmm0,xmm0
|
---|
901 | por xmm4,xmm2
|
---|
902 | por xmm5,xmm3
|
---|
903 | movdqu XMMWORD[(-32)+rdi],xmm4
|
---|
904 | movdqu XMMWORD[(-16)+rdi],xmm5
|
---|
905 | add r9,32
|
---|
906 | jnz NEAR $L$sqr8x_cond_copy
|
---|
907 |
|
---|
908 | mov rax,1
|
---|
909 | mov r15,QWORD[((-48))+rsi]
|
---|
910 |
|
---|
911 | mov r14,QWORD[((-40))+rsi]
|
---|
912 |
|
---|
913 | mov r13,QWORD[((-32))+rsi]
|
---|
914 |
|
---|
915 | mov r12,QWORD[((-24))+rsi]
|
---|
916 |
|
---|
917 | mov rbp,QWORD[((-16))+rsi]
|
---|
918 |
|
---|
919 | mov rbx,QWORD[((-8))+rsi]
|
---|
920 |
|
---|
921 | lea rsp,[rsi]
|
---|
922 |
|
---|
923 | $L$sqr8x_epilogue:
|
---|
924 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
---|
925 | mov rsi,QWORD[16+rsp]
|
---|
926 | DB 0F3h,0C3h ;repret
|
---|
927 |
|
---|
928 | $L$SEH_end_bn_sqr8x_mont:
|
---|
929 |
|
---|
930 | ALIGN 32
|
---|
931 | bn_mulx4x_mont:
|
---|
932 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
---|
933 | mov QWORD[16+rsp],rsi
|
---|
934 | mov rax,rsp
|
---|
935 | $L$SEH_begin_bn_mulx4x_mont:
|
---|
936 | mov rdi,rcx
|
---|
937 | mov rsi,rdx
|
---|
938 | mov rdx,r8
|
---|
939 | mov rcx,r9
|
---|
940 | mov r8,QWORD[40+rsp]
|
---|
941 | mov r9,QWORD[48+rsp]
|
---|
942 |
|
---|
943 |
|
---|
944 |
|
---|
945 | mov rax,rsp
|
---|
946 |
|
---|
947 | $L$mulx4x_enter:
|
---|
948 | push rbx
|
---|
949 |
|
---|
950 | push rbp
|
---|
951 |
|
---|
952 | push r12
|
---|
953 |
|
---|
954 | push r13
|
---|
955 |
|
---|
956 | push r14
|
---|
957 |
|
---|
958 | push r15
|
---|
959 |
|
---|
960 | $L$mulx4x_prologue:
|
---|
961 |
|
---|
962 | shl r9d,3
|
---|
963 | xor r10,r10
|
---|
964 | sub r10,r9
|
---|
965 | mov r8,QWORD[r8]
|
---|
966 | lea rbp,[((-72))+r10*1+rsp]
|
---|
967 | and rbp,-128
|
---|
968 | mov r11,rsp
|
---|
969 | sub r11,rbp
|
---|
970 | and r11,-4096
|
---|
971 | lea rsp,[rbp*1+r11]
|
---|
972 | mov r10,QWORD[rsp]
|
---|
973 | cmp rsp,rbp
|
---|
974 | ja NEAR $L$mulx4x_page_walk
|
---|
975 | jmp NEAR $L$mulx4x_page_walk_done
|
---|
976 |
|
---|
977 | ALIGN 16
|
---|
978 | $L$mulx4x_page_walk:
|
---|
979 | lea rsp,[((-4096))+rsp]
|
---|
980 | mov r10,QWORD[rsp]
|
---|
981 | cmp rsp,rbp
|
---|
982 | ja NEAR $L$mulx4x_page_walk
|
---|
983 | $L$mulx4x_page_walk_done:
|
---|
984 |
|
---|
985 | lea r10,[r9*1+rdx]
|
---|
986 |
|
---|
987 |
|
---|
988 |
|
---|
989 |
|
---|
990 |
|
---|
991 |
|
---|
992 |
|
---|
993 |
|
---|
994 |
|
---|
995 |
|
---|
996 |
|
---|
997 |
|
---|
998 | mov QWORD[rsp],r9
|
---|
999 | shr r9,5
|
---|
1000 | mov QWORD[16+rsp],r10
|
---|
1001 | sub r9,1
|
---|
1002 | mov QWORD[24+rsp],r8
|
---|
1003 | mov QWORD[32+rsp],rdi
|
---|
1004 | mov QWORD[40+rsp],rax
|
---|
1005 |
|
---|
1006 | mov QWORD[48+rsp],r9
|
---|
1007 | jmp NEAR $L$mulx4x_body
|
---|
1008 |
|
---|
1009 | ALIGN 32
|
---|
1010 | $L$mulx4x_body:
|
---|
1011 | lea rdi,[8+rdx]
|
---|
1012 | mov rdx,QWORD[rdx]
|
---|
1013 | lea rbx,[((64+32))+rsp]
|
---|
1014 | mov r9,rdx
|
---|
1015 |
|
---|
1016 | mulx rax,r8,QWORD[rsi]
|
---|
1017 | mulx r14,r11,QWORD[8+rsi]
|
---|
1018 | add r11,rax
|
---|
1019 | mov QWORD[8+rsp],rdi
|
---|
1020 | mulx r13,r12,QWORD[16+rsi]
|
---|
1021 | adc r12,r14
|
---|
1022 | adc r13,0
|
---|
1023 |
|
---|
1024 | mov rdi,r8
|
---|
1025 | imul r8,QWORD[24+rsp]
|
---|
1026 | xor rbp,rbp
|
---|
1027 |
|
---|
1028 | mulx r14,rax,QWORD[24+rsi]
|
---|
1029 | mov rdx,r8
|
---|
1030 | lea rsi,[32+rsi]
|
---|
1031 | adcx r13,rax
|
---|
1032 | adcx r14,rbp
|
---|
1033 |
|
---|
1034 | mulx r10,rax,QWORD[rcx]
|
---|
1035 | adcx rdi,rax
|
---|
1036 | adox r10,r11
|
---|
1037 | mulx r11,rax,QWORD[8+rcx]
|
---|
1038 | adcx r10,rax
|
---|
1039 | adox r11,r12
|
---|
1040 | DB 0xc4,0x62,0xfb,0xf6,0xa1,0x10,0x00,0x00,0x00
|
---|
1041 | mov rdi,QWORD[48+rsp]
|
---|
1042 | mov QWORD[((-32))+rbx],r10
|
---|
1043 | adcx r11,rax
|
---|
1044 | adox r12,r13
|
---|
1045 | mulx r15,rax,QWORD[24+rcx]
|
---|
1046 | mov rdx,r9
|
---|
1047 | mov QWORD[((-24))+rbx],r11
|
---|
1048 | adcx r12,rax
|
---|
1049 | adox r15,rbp
|
---|
1050 | lea rcx,[32+rcx]
|
---|
1051 | mov QWORD[((-16))+rbx],r12
|
---|
1052 |
|
---|
1053 | jmp NEAR $L$mulx4x_1st
|
---|
1054 |
|
---|
1055 | ALIGN 32
|
---|
1056 | $L$mulx4x_1st:
|
---|
1057 | adcx r15,rbp
|
---|
1058 | mulx rax,r10,QWORD[rsi]
|
---|
1059 | adcx r10,r14
|
---|
1060 | mulx r14,r11,QWORD[8+rsi]
|
---|
1061 | adcx r11,rax
|
---|
1062 | mulx rax,r12,QWORD[16+rsi]
|
---|
1063 | adcx r12,r14
|
---|
1064 | mulx r14,r13,QWORD[24+rsi]
|
---|
1065 | DB 0x67,0x67
|
---|
1066 | mov rdx,r8
|
---|
1067 | adcx r13,rax
|
---|
1068 | adcx r14,rbp
|
---|
1069 | lea rsi,[32+rsi]
|
---|
1070 | lea rbx,[32+rbx]
|
---|
1071 |
|
---|
1072 | adox r10,r15
|
---|
1073 | mulx r15,rax,QWORD[rcx]
|
---|
1074 | adcx r10,rax
|
---|
1075 | adox r11,r15
|
---|
1076 | mulx r15,rax,QWORD[8+rcx]
|
---|
1077 | adcx r11,rax
|
---|
1078 | adox r12,r15
|
---|
1079 | mulx r15,rax,QWORD[16+rcx]
|
---|
1080 | mov QWORD[((-40))+rbx],r10
|
---|
1081 | adcx r12,rax
|
---|
1082 | mov QWORD[((-32))+rbx],r11
|
---|
1083 | adox r13,r15
|
---|
1084 | mulx r15,rax,QWORD[24+rcx]
|
---|
1085 | mov rdx,r9
|
---|
1086 | mov QWORD[((-24))+rbx],r12
|
---|
1087 | adcx r13,rax
|
---|
1088 | adox r15,rbp
|
---|
1089 | lea rcx,[32+rcx]
|
---|
1090 | mov QWORD[((-16))+rbx],r13
|
---|
1091 |
|
---|
1092 | dec rdi
|
---|
1093 | jnz NEAR $L$mulx4x_1st
|
---|
1094 |
|
---|
1095 | mov rax,QWORD[rsp]
|
---|
1096 | mov rdi,QWORD[8+rsp]
|
---|
1097 | adc r15,rbp
|
---|
1098 | add r14,r15
|
---|
1099 | sbb r15,r15
|
---|
1100 | mov QWORD[((-8))+rbx],r14
|
---|
1101 | jmp NEAR $L$mulx4x_outer
|
---|
1102 |
|
---|
1103 | ALIGN 32
|
---|
1104 | $L$mulx4x_outer:
|
---|
1105 | mov rdx,QWORD[rdi]
|
---|
1106 | lea rdi,[8+rdi]
|
---|
1107 | sub rsi,rax
|
---|
1108 | mov QWORD[rbx],r15
|
---|
1109 | lea rbx,[((64+32))+rsp]
|
---|
1110 | sub rcx,rax
|
---|
1111 |
|
---|
1112 | mulx r11,r8,QWORD[rsi]
|
---|
1113 | xor ebp,ebp
|
---|
1114 | mov r9,rdx
|
---|
1115 | mulx r12,r14,QWORD[8+rsi]
|
---|
1116 | adox r8,QWORD[((-32))+rbx]
|
---|
1117 | adcx r11,r14
|
---|
1118 | mulx r13,r15,QWORD[16+rsi]
|
---|
1119 | adox r11,QWORD[((-24))+rbx]
|
---|
1120 | adcx r12,r15
|
---|
1121 | adox r12,QWORD[((-16))+rbx]
|
---|
1122 | adcx r13,rbp
|
---|
1123 | adox r13,rbp
|
---|
1124 |
|
---|
1125 | mov QWORD[8+rsp],rdi
|
---|
1126 | mov r15,r8
|
---|
1127 | imul r8,QWORD[24+rsp]
|
---|
1128 | xor ebp,ebp
|
---|
1129 |
|
---|
1130 | mulx r14,rax,QWORD[24+rsi]
|
---|
1131 | mov rdx,r8
|
---|
1132 | adcx r13,rax
|
---|
1133 | adox r13,QWORD[((-8))+rbx]
|
---|
1134 | adcx r14,rbp
|
---|
1135 | lea rsi,[32+rsi]
|
---|
1136 | adox r14,rbp
|
---|
1137 |
|
---|
1138 | mulx r10,rax,QWORD[rcx]
|
---|
1139 | adcx r15,rax
|
---|
1140 | adox r10,r11
|
---|
1141 | mulx r11,rax,QWORD[8+rcx]
|
---|
1142 | adcx r10,rax
|
---|
1143 | adox r11,r12
|
---|
1144 | mulx r12,rax,QWORD[16+rcx]
|
---|
1145 | mov QWORD[((-32))+rbx],r10
|
---|
1146 | adcx r11,rax
|
---|
1147 | adox r12,r13
|
---|
1148 | mulx r15,rax,QWORD[24+rcx]
|
---|
1149 | mov rdx,r9
|
---|
1150 | mov QWORD[((-24))+rbx],r11
|
---|
1151 | lea rcx,[32+rcx]
|
---|
1152 | adcx r12,rax
|
---|
1153 | adox r15,rbp
|
---|
1154 | mov rdi,QWORD[48+rsp]
|
---|
1155 | mov QWORD[((-16))+rbx],r12
|
---|
1156 |
|
---|
1157 | jmp NEAR $L$mulx4x_inner
|
---|
1158 |
|
---|
1159 | ALIGN 32
|
---|
1160 | $L$mulx4x_inner:
|
---|
1161 | mulx rax,r10,QWORD[rsi]
|
---|
1162 | adcx r15,rbp
|
---|
1163 | adox r10,r14
|
---|
1164 | mulx r14,r11,QWORD[8+rsi]
|
---|
1165 | adcx r10,QWORD[rbx]
|
---|
1166 | adox r11,rax
|
---|
1167 | mulx rax,r12,QWORD[16+rsi]
|
---|
1168 | adcx r11,QWORD[8+rbx]
|
---|
1169 | adox r12,r14
|
---|
1170 | mulx r14,r13,QWORD[24+rsi]
|
---|
1171 | mov rdx,r8
|
---|
1172 | adcx r12,QWORD[16+rbx]
|
---|
1173 | adox r13,rax
|
---|
1174 | adcx r13,QWORD[24+rbx]
|
---|
1175 | adox r14,rbp
|
---|
1176 | lea rsi,[32+rsi]
|
---|
1177 | lea rbx,[32+rbx]
|
---|
1178 | adcx r14,rbp
|
---|
1179 |
|
---|
1180 | adox r10,r15
|
---|
1181 | mulx r15,rax,QWORD[rcx]
|
---|
1182 | adcx r10,rax
|
---|
1183 | adox r11,r15
|
---|
1184 | mulx r15,rax,QWORD[8+rcx]
|
---|
1185 | adcx r11,rax
|
---|
1186 | adox r12,r15
|
---|
1187 | mulx r15,rax,QWORD[16+rcx]
|
---|
1188 | mov QWORD[((-40))+rbx],r10
|
---|
1189 | adcx r12,rax
|
---|
1190 | adox r13,r15
|
---|
1191 | mulx r15,rax,QWORD[24+rcx]
|
---|
1192 | mov rdx,r9
|
---|
1193 | mov QWORD[((-32))+rbx],r11
|
---|
1194 | mov QWORD[((-24))+rbx],r12
|
---|
1195 | adcx r13,rax
|
---|
1196 | adox r15,rbp
|
---|
1197 | lea rcx,[32+rcx]
|
---|
1198 | mov QWORD[((-16))+rbx],r13
|
---|
1199 |
|
---|
1200 | dec rdi
|
---|
1201 | jnz NEAR $L$mulx4x_inner
|
---|
1202 |
|
---|
1203 | mov rax,QWORD[rsp]
|
---|
1204 | mov rdi,QWORD[8+rsp]
|
---|
1205 | adc r15,rbp
|
---|
1206 | sub rbp,QWORD[rbx]
|
---|
1207 | adc r14,r15
|
---|
1208 | sbb r15,r15
|
---|
1209 | mov QWORD[((-8))+rbx],r14
|
---|
1210 |
|
---|
1211 | cmp rdi,QWORD[16+rsp]
|
---|
1212 | jne NEAR $L$mulx4x_outer
|
---|
1213 |
|
---|
1214 | lea rbx,[64+rsp]
|
---|
1215 | sub rcx,rax
|
---|
1216 | neg r15
|
---|
1217 | mov rdx,rax
|
---|
1218 | shr rax,3+2
|
---|
1219 | mov rdi,QWORD[32+rsp]
|
---|
1220 | jmp NEAR $L$mulx4x_sub
|
---|
1221 |
|
---|
1222 | ALIGN 32
|
---|
1223 | $L$mulx4x_sub:
|
---|
1224 | mov r11,QWORD[rbx]
|
---|
1225 | mov r12,QWORD[8+rbx]
|
---|
1226 | mov r13,QWORD[16+rbx]
|
---|
1227 | mov r14,QWORD[24+rbx]
|
---|
1228 | lea rbx,[32+rbx]
|
---|
1229 | sbb r11,QWORD[rcx]
|
---|
1230 | sbb r12,QWORD[8+rcx]
|
---|
1231 | sbb r13,QWORD[16+rcx]
|
---|
1232 | sbb r14,QWORD[24+rcx]
|
---|
1233 | lea rcx,[32+rcx]
|
---|
1234 | mov QWORD[rdi],r11
|
---|
1235 | mov QWORD[8+rdi],r12
|
---|
1236 | mov QWORD[16+rdi],r13
|
---|
1237 | mov QWORD[24+rdi],r14
|
---|
1238 | lea rdi,[32+rdi]
|
---|
1239 | dec rax
|
---|
1240 | jnz NEAR $L$mulx4x_sub
|
---|
1241 |
|
---|
1242 | sbb r15,0
|
---|
1243 | lea rbx,[64+rsp]
|
---|
1244 | sub rdi,rdx
|
---|
1245 |
|
---|
1246 | DB 102,73,15,110,207
|
---|
1247 | pxor xmm0,xmm0
|
---|
1248 | pshufd xmm1,xmm1,0
|
---|
1249 | mov rsi,QWORD[40+rsp]
|
---|
1250 |
|
---|
1251 | jmp NEAR $L$mulx4x_cond_copy
|
---|
1252 |
|
---|
1253 | ALIGN 32
|
---|
1254 | $L$mulx4x_cond_copy:
|
---|
1255 | movdqa xmm2,XMMWORD[rbx]
|
---|
1256 | movdqa xmm3,XMMWORD[16+rbx]
|
---|
1257 | lea rbx,[32+rbx]
|
---|
1258 | movdqu xmm4,XMMWORD[rdi]
|
---|
1259 | movdqu xmm5,XMMWORD[16+rdi]
|
---|
1260 | lea rdi,[32+rdi]
|
---|
1261 | movdqa XMMWORD[(-32)+rbx],xmm0
|
---|
1262 | movdqa XMMWORD[(-16)+rbx],xmm0
|
---|
1263 | pcmpeqd xmm0,xmm1
|
---|
1264 | pand xmm2,xmm1
|
---|
1265 | pand xmm3,xmm1
|
---|
1266 | pand xmm4,xmm0
|
---|
1267 | pand xmm5,xmm0
|
---|
1268 | pxor xmm0,xmm0
|
---|
1269 | por xmm4,xmm2
|
---|
1270 | por xmm5,xmm3
|
---|
1271 | movdqu XMMWORD[(-32)+rdi],xmm4
|
---|
1272 | movdqu XMMWORD[(-16)+rdi],xmm5
|
---|
1273 | sub rdx,32
|
---|
1274 | jnz NEAR $L$mulx4x_cond_copy
|
---|
1275 |
|
---|
1276 | mov QWORD[rbx],rdx
|
---|
1277 |
|
---|
1278 | mov rax,1
|
---|
1279 | mov r15,QWORD[((-48))+rsi]
|
---|
1280 |
|
---|
1281 | mov r14,QWORD[((-40))+rsi]
|
---|
1282 |
|
---|
1283 | mov r13,QWORD[((-32))+rsi]
|
---|
1284 |
|
---|
1285 | mov r12,QWORD[((-24))+rsi]
|
---|
1286 |
|
---|
1287 | mov rbp,QWORD[((-16))+rsi]
|
---|
1288 |
|
---|
1289 | mov rbx,QWORD[((-8))+rsi]
|
---|
1290 |
|
---|
1291 | lea rsp,[rsi]
|
---|
1292 |
|
---|
1293 | $L$mulx4x_epilogue:
|
---|
1294 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
---|
1295 | mov rsi,QWORD[16+rsp]
|
---|
1296 | DB 0F3h,0C3h ;repret
|
---|
1297 |
|
---|
1298 | $L$SEH_end_bn_mulx4x_mont:
|
---|
1299 | DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
|
---|
1300 | DB 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
|
---|
1301 | DB 54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83
|
---|
1302 | DB 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
|
---|
1303 | DB 115,108,46,111,114,103,62,0
|
---|
1304 | ALIGN 16
|
---|
1305 | EXTERN __imp_RtlVirtualUnwind
|
---|
1306 |
|
---|
1307 | ALIGN 16
|
---|
1308 | mul_handler:
|
---|
1309 | push rsi
|
---|
1310 | push rdi
|
---|
1311 | push rbx
|
---|
1312 | push rbp
|
---|
1313 | push r12
|
---|
1314 | push r13
|
---|
1315 | push r14
|
---|
1316 | push r15
|
---|
1317 | pushfq
|
---|
1318 | sub rsp,64
|
---|
1319 |
|
---|
1320 | mov rax,QWORD[120+r8]
|
---|
1321 | mov rbx,QWORD[248+r8]
|
---|
1322 |
|
---|
1323 | mov rsi,QWORD[8+r9]
|
---|
1324 | mov r11,QWORD[56+r9]
|
---|
1325 |
|
---|
1326 | mov r10d,DWORD[r11]
|
---|
1327 | lea r10,[r10*1+rsi]
|
---|
1328 | cmp rbx,r10
|
---|
1329 | jb NEAR $L$common_seh_tail
|
---|
1330 |
|
---|
1331 | mov rax,QWORD[152+r8]
|
---|
1332 |
|
---|
1333 | mov r10d,DWORD[4+r11]
|
---|
1334 | lea r10,[r10*1+rsi]
|
---|
1335 | cmp rbx,r10
|
---|
1336 | jae NEAR $L$common_seh_tail
|
---|
1337 |
|
---|
1338 | mov r10,QWORD[192+r8]
|
---|
1339 | mov rax,QWORD[8+r10*8+rax]
|
---|
1340 |
|
---|
1341 | jmp NEAR $L$common_pop_regs
|
---|
1342 |
|
---|
1343 |
|
---|
1344 |
|
---|
1345 | ALIGN 16
|
---|
1346 | sqr_handler:
|
---|
1347 | push rsi
|
---|
1348 | push rdi
|
---|
1349 | push rbx
|
---|
1350 | push rbp
|
---|
1351 | push r12
|
---|
1352 | push r13
|
---|
1353 | push r14
|
---|
1354 | push r15
|
---|
1355 | pushfq
|
---|
1356 | sub rsp,64
|
---|
1357 |
|
---|
1358 | mov rax,QWORD[120+r8]
|
---|
1359 | mov rbx,QWORD[248+r8]
|
---|
1360 |
|
---|
1361 | mov rsi,QWORD[8+r9]
|
---|
1362 | mov r11,QWORD[56+r9]
|
---|
1363 |
|
---|
1364 | mov r10d,DWORD[r11]
|
---|
1365 | lea r10,[r10*1+rsi]
|
---|
1366 | cmp rbx,r10
|
---|
1367 | jb NEAR $L$common_seh_tail
|
---|
1368 |
|
---|
1369 | mov r10d,DWORD[4+r11]
|
---|
1370 | lea r10,[r10*1+rsi]
|
---|
1371 | cmp rbx,r10
|
---|
1372 | jb NEAR $L$common_pop_regs
|
---|
1373 |
|
---|
1374 | mov rax,QWORD[152+r8]
|
---|
1375 |
|
---|
1376 | mov r10d,DWORD[8+r11]
|
---|
1377 | lea r10,[r10*1+rsi]
|
---|
1378 | cmp rbx,r10
|
---|
1379 | jae NEAR $L$common_seh_tail
|
---|
1380 |
|
---|
1381 | mov rax,QWORD[40+rax]
|
---|
1382 |
|
---|
1383 | $L$common_pop_regs:
|
---|
1384 | mov rbx,QWORD[((-8))+rax]
|
---|
1385 | mov rbp,QWORD[((-16))+rax]
|
---|
1386 | mov r12,QWORD[((-24))+rax]
|
---|
1387 | mov r13,QWORD[((-32))+rax]
|
---|
1388 | mov r14,QWORD[((-40))+rax]
|
---|
1389 | mov r15,QWORD[((-48))+rax]
|
---|
1390 | mov QWORD[144+r8],rbx
|
---|
1391 | mov QWORD[160+r8],rbp
|
---|
1392 | mov QWORD[216+r8],r12
|
---|
1393 | mov QWORD[224+r8],r13
|
---|
1394 | mov QWORD[232+r8],r14
|
---|
1395 | mov QWORD[240+r8],r15
|
---|
1396 |
|
---|
1397 | $L$common_seh_tail:
|
---|
1398 | mov rdi,QWORD[8+rax]
|
---|
1399 | mov rsi,QWORD[16+rax]
|
---|
1400 | mov QWORD[152+r8],rax
|
---|
1401 | mov QWORD[168+r8],rsi
|
---|
1402 | mov QWORD[176+r8],rdi
|
---|
1403 |
|
---|
1404 | mov rdi,QWORD[40+r9]
|
---|
1405 | mov rsi,r8
|
---|
1406 | mov ecx,154
|
---|
1407 | DD 0xa548f3fc
|
---|
1408 |
|
---|
1409 | mov rsi,r9
|
---|
1410 | xor rcx,rcx
|
---|
1411 | mov rdx,QWORD[8+rsi]
|
---|
1412 | mov r8,QWORD[rsi]
|
---|
1413 | mov r9,QWORD[16+rsi]
|
---|
1414 | mov r10,QWORD[40+rsi]
|
---|
1415 | lea r11,[56+rsi]
|
---|
1416 | lea r12,[24+rsi]
|
---|
1417 | mov QWORD[32+rsp],r10
|
---|
1418 | mov QWORD[40+rsp],r11
|
---|
1419 | mov QWORD[48+rsp],r12
|
---|
1420 | mov QWORD[56+rsp],rcx
|
---|
1421 | call QWORD[__imp_RtlVirtualUnwind]
|
---|
1422 |
|
---|
1423 | mov eax,1
|
---|
1424 | add rsp,64
|
---|
1425 | popfq
|
---|
1426 | pop r15
|
---|
1427 | pop r14
|
---|
1428 | pop r13
|
---|
1429 | pop r12
|
---|
1430 | pop rbp
|
---|
1431 | pop rbx
|
---|
1432 | pop rdi
|
---|
1433 | pop rsi
|
---|
1434 | DB 0F3h,0C3h ;repret
|
---|
1435 |
|
---|
1436 |
|
---|
1437 | section .pdata rdata align=4
|
---|
1438 | ALIGN 4
|
---|
1439 | DD $L$SEH_begin_bn_mul_mont wrt ..imagebase
|
---|
1440 | DD $L$SEH_end_bn_mul_mont wrt ..imagebase
|
---|
1441 | DD $L$SEH_info_bn_mul_mont wrt ..imagebase
|
---|
1442 |
|
---|
1443 | DD $L$SEH_begin_bn_mul4x_mont wrt ..imagebase
|
---|
1444 | DD $L$SEH_end_bn_mul4x_mont wrt ..imagebase
|
---|
1445 | DD $L$SEH_info_bn_mul4x_mont wrt ..imagebase
|
---|
1446 |
|
---|
1447 | DD $L$SEH_begin_bn_sqr8x_mont wrt ..imagebase
|
---|
1448 | DD $L$SEH_end_bn_sqr8x_mont wrt ..imagebase
|
---|
1449 | DD $L$SEH_info_bn_sqr8x_mont wrt ..imagebase
|
---|
1450 | DD $L$SEH_begin_bn_mulx4x_mont wrt ..imagebase
|
---|
1451 | DD $L$SEH_end_bn_mulx4x_mont wrt ..imagebase
|
---|
1452 | DD $L$SEH_info_bn_mulx4x_mont wrt ..imagebase
|
---|
1453 | section .xdata rdata align=8
|
---|
1454 | ALIGN 8
|
---|
1455 | $L$SEH_info_bn_mul_mont:
|
---|
1456 | DB 9,0,0,0
|
---|
1457 | DD mul_handler wrt ..imagebase
|
---|
1458 | DD $L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
|
---|
1459 | $L$SEH_info_bn_mul4x_mont:
|
---|
1460 | DB 9,0,0,0
|
---|
1461 | DD mul_handler wrt ..imagebase
|
---|
1462 | DD $L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase
|
---|
1463 | $L$SEH_info_bn_sqr8x_mont:
|
---|
1464 | DB 9,0,0,0
|
---|
1465 | DD sqr_handler wrt ..imagebase
|
---|
1466 | DD $L$sqr8x_prologue wrt ..imagebase,$L$sqr8x_body wrt ..imagebase,$L$sqr8x_epilogue wrt ..imagebase
|
---|
1467 | ALIGN 8
|
---|
1468 | $L$SEH_info_bn_mulx4x_mont:
|
---|
1469 | DB 9,0,0,0
|
---|
1470 | DD sqr_handler wrt ..imagebase
|
---|
1471 | DD $L$mulx4x_prologue wrt ..imagebase,$L$mulx4x_body wrt ..imagebase,$L$mulx4x_epilogue wrt ..imagebase
|
---|
1472 | ALIGN 8
|
---|