Home | History | Annotate | Download | only in fipsmodule
      1 #if defined(__x86_64__) && !defined(OPENSSL_NO_ASM)
      2 .text
      3 
      4 
      5 
      6 .globl	_bn_mul_mont
      7 .private_extern _bn_mul_mont
      8 
      9 .p2align	4
     10 _bn_mul_mont:
     11 
     12 	movl	%r9d,%r9d
     13 	movq	%rsp,%rax
     14 
     15 	testl	$3,%r9d
     16 	jnz	L$mul_enter
     17 	cmpl	$8,%r9d
     18 	jb	L$mul_enter
     19 	cmpq	%rsi,%rdx
     20 	jne	L$mul4x_enter
     21 	testl	$7,%r9d
     22 	jz	L$sqr8x_enter
     23 	jmp	L$mul4x_enter
     24 
     25 .p2align	4
     26 L$mul_enter:
     27 	pushq	%rbx
     28 
     29 	pushq	%rbp
     30 
     31 	pushq	%r12
     32 
     33 	pushq	%r13
     34 
     35 	pushq	%r14
     36 
     37 	pushq	%r15
     38 
     39 
     40 	negq	%r9
     41 	movq	%rsp,%r11
     42 	leaq	-16(%rsp,%r9,8),%r10
     43 	negq	%r9
     44 	andq	$-1024,%r10
     45 
     46 
     47 
     48 
     49 
     50 
     51 
     52 
     53 
     54 	subq	%r10,%r11
     55 	andq	$-4096,%r11
     56 	leaq	(%r10,%r11,1),%rsp
     57 	movq	(%rsp),%r11
     58 	cmpq	%r10,%rsp
     59 	ja	L$mul_page_walk
     60 	jmp	L$mul_page_walk_done
     61 
     62 .p2align	4
     63 L$mul_page_walk:
     64 	leaq	-4096(%rsp),%rsp
     65 	movq	(%rsp),%r11
     66 	cmpq	%r10,%rsp
     67 	ja	L$mul_page_walk
     68 L$mul_page_walk_done:
     69 
     70 	movq	%rax,8(%rsp,%r9,8)
     71 
     72 L$mul_body:
     73 	movq	%rdx,%r12
     74 	movq	(%r8),%r8
     75 	movq	(%r12),%rbx
     76 	movq	(%rsi),%rax
     77 
     78 	xorq	%r14,%r14
     79 	xorq	%r15,%r15
     80 
     81 	movq	%r8,%rbp
     82 	mulq	%rbx
     83 	movq	%rax,%r10
     84 	movq	(%rcx),%rax
     85 
     86 	imulq	%r10,%rbp
     87 	movq	%rdx,%r11
     88 
     89 	mulq	%rbp
     90 	addq	%rax,%r10
     91 	movq	8(%rsi),%rax
     92 	adcq	$0,%rdx
     93 	movq	%rdx,%r13
     94 
     95 	leaq	1(%r15),%r15
     96 	jmp	L$1st_enter
     97 
     98 .p2align	4
     99 L$1st:
    100 	addq	%rax,%r13
    101 	movq	(%rsi,%r15,8),%rax
    102 	adcq	$0,%rdx
    103 	addq	%r11,%r13
    104 	movq	%r10,%r11
    105 	adcq	$0,%rdx
    106 	movq	%r13,-16(%rsp,%r15,8)
    107 	movq	%rdx,%r13
    108 
    109 L$1st_enter:
    110 	mulq	%rbx
    111 	addq	%rax,%r11
    112 	movq	(%rcx,%r15,8),%rax
    113 	adcq	$0,%rdx
    114 	leaq	1(%r15),%r15
    115 	movq	%rdx,%r10
    116 
    117 	mulq	%rbp
    118 	cmpq	%r9,%r15
    119 	jne	L$1st
    120 
    121 	addq	%rax,%r13
    122 	movq	(%rsi),%rax
    123 	adcq	$0,%rdx
    124 	addq	%r11,%r13
    125 	adcq	$0,%rdx
    126 	movq	%r13,-16(%rsp,%r15,8)
    127 	movq	%rdx,%r13
    128 	movq	%r10,%r11
    129 
    130 	xorq	%rdx,%rdx
    131 	addq	%r11,%r13
    132 	adcq	$0,%rdx
    133 	movq	%r13,-8(%rsp,%r9,8)
    134 	movq	%rdx,(%rsp,%r9,8)
    135 
    136 	leaq	1(%r14),%r14
    137 	jmp	L$outer
    138 .p2align	4
    139 L$outer:
    140 	movq	(%r12,%r14,8),%rbx
    141 	xorq	%r15,%r15
    142 	movq	%r8,%rbp
    143 	movq	(%rsp),%r10
    144 	mulq	%rbx
    145 	addq	%rax,%r10
    146 	movq	(%rcx),%rax
    147 	adcq	$0,%rdx
    148 
    149 	imulq	%r10,%rbp
    150 	movq	%rdx,%r11
    151 
    152 	mulq	%rbp
    153 	addq	%rax,%r10
    154 	movq	8(%rsi),%rax
    155 	adcq	$0,%rdx
    156 	movq	8(%rsp),%r10
    157 	movq	%rdx,%r13
    158 
    159 	leaq	1(%r15),%r15
    160 	jmp	L$inner_enter
    161 
    162 .p2align	4
    163 L$inner:
    164 	addq	%rax,%r13
    165 	movq	(%rsi,%r15,8),%rax
    166 	adcq	$0,%rdx
    167 	addq	%r10,%r13
    168 	movq	(%rsp,%r15,8),%r10
    169 	adcq	$0,%rdx
    170 	movq	%r13,-16(%rsp,%r15,8)
    171 	movq	%rdx,%r13
    172 
    173 L$inner_enter:
    174 	mulq	%rbx
    175 	addq	%rax,%r11
    176 	movq	(%rcx,%r15,8),%rax
    177 	adcq	$0,%rdx
    178 	addq	%r11,%r10
    179 	movq	%rdx,%r11
    180 	adcq	$0,%r11
    181 	leaq	1(%r15),%r15
    182 
    183 	mulq	%rbp
    184 	cmpq	%r9,%r15
    185 	jne	L$inner
    186 
    187 	addq	%rax,%r13
    188 	movq	(%rsi),%rax
    189 	adcq	$0,%rdx
    190 	addq	%r10,%r13
    191 	movq	(%rsp,%r15,8),%r10
    192 	adcq	$0,%rdx
    193 	movq	%r13,-16(%rsp,%r15,8)
    194 	movq	%rdx,%r13
    195 
    196 	xorq	%rdx,%rdx
    197 	addq	%r11,%r13
    198 	adcq	$0,%rdx
    199 	addq	%r10,%r13
    200 	adcq	$0,%rdx
    201 	movq	%r13,-8(%rsp,%r9,8)
    202 	movq	%rdx,(%rsp,%r9,8)
    203 
    204 	leaq	1(%r14),%r14
    205 	cmpq	%r9,%r14
    206 	jb	L$outer
    207 
    208 	xorq	%r14,%r14
    209 	movq	(%rsp),%rax
    210 	leaq	(%rsp),%rsi
    211 	movq	%r9,%r15
    212 	jmp	L$sub
    213 .p2align	4
    214 L$sub:
    215 	sbbq	(%rcx,%r14,8),%rax
    216 	movq	%rax,(%rdi,%r14,8)
    217 	movq	8(%rsi,%r14,8),%rax
    218 	leaq	1(%r14),%r14
    219 	decq	%r15
    220 	jnz	L$sub
    221 
    222 	sbbq	$0,%rax
    223 	xorq	%r14,%r14
    224 	andq	%rax,%rsi
    225 	notq	%rax
    226 	movq	%rdi,%rcx
    227 	andq	%rax,%rcx
    228 	movq	%r9,%r15
    229 	orq	%rcx,%rsi
    230 .p2align	4
    231 L$copy:
    232 	movq	(%rsi,%r14,8),%rax
    233 	movq	%r14,(%rsp,%r14,8)
    234 	movq	%rax,(%rdi,%r14,8)
    235 	leaq	1(%r14),%r14
    236 	subq	$1,%r15
    237 	jnz	L$copy
    238 
    239 	movq	8(%rsp,%r9,8),%rsi
    240 
    241 	movq	$1,%rax
    242 	movq	-48(%rsi),%r15
    243 
    244 	movq	-40(%rsi),%r14
    245 
    246 	movq	-32(%rsi),%r13
    247 
    248 	movq	-24(%rsi),%r12
    249 
    250 	movq	-16(%rsi),%rbp
    251 
    252 	movq	-8(%rsi),%rbx
    253 
    254 	leaq	(%rsi),%rsp
    255 
    256 L$mul_epilogue:
    257 	.byte	0xf3,0xc3
    258 
    259 
    260 
    261 .p2align	4
    262 bn_mul4x_mont:
    263 
    264 	movl	%r9d,%r9d
    265 	movq	%rsp,%rax
    266 
    267 L$mul4x_enter:
    268 	pushq	%rbx
    269 
    270 	pushq	%rbp
    271 
    272 	pushq	%r12
    273 
    274 	pushq	%r13
    275 
    276 	pushq	%r14
    277 
    278 	pushq	%r15
    279 
    280 
    281 	negq	%r9
    282 	movq	%rsp,%r11
    283 	leaq	-32(%rsp,%r9,8),%r10
    284 	negq	%r9
    285 	andq	$-1024,%r10
    286 
    287 	subq	%r10,%r11
    288 	andq	$-4096,%r11
    289 	leaq	(%r10,%r11,1),%rsp
    290 	movq	(%rsp),%r11
    291 	cmpq	%r10,%rsp
    292 	ja	L$mul4x_page_walk
    293 	jmp	L$mul4x_page_walk_done
    294 
    295 L$mul4x_page_walk:
    296 	leaq	-4096(%rsp),%rsp
    297 	movq	(%rsp),%r11
    298 	cmpq	%r10,%rsp
    299 	ja	L$mul4x_page_walk
    300 L$mul4x_page_walk_done:
    301 
    302 	movq	%rax,8(%rsp,%r9,8)
    303 
    304 L$mul4x_body:
    305 	movq	%rdi,16(%rsp,%r9,8)
    306 	movq	%rdx,%r12
    307 	movq	(%r8),%r8
    308 	movq	(%r12),%rbx
    309 	movq	(%rsi),%rax
    310 
    311 	xorq	%r14,%r14
    312 	xorq	%r15,%r15
    313 
    314 	movq	%r8,%rbp
    315 	mulq	%rbx
    316 	movq	%rax,%r10
    317 	movq	(%rcx),%rax
    318 
    319 	imulq	%r10,%rbp
    320 	movq	%rdx,%r11
    321 
    322 	mulq	%rbp
    323 	addq	%rax,%r10
    324 	movq	8(%rsi),%rax
    325 	adcq	$0,%rdx
    326 	movq	%rdx,%rdi
    327 
    328 	mulq	%rbx
    329 	addq	%rax,%r11
    330 	movq	8(%rcx),%rax
    331 	adcq	$0,%rdx
    332 	movq	%rdx,%r10
    333 
    334 	mulq	%rbp
    335 	addq	%rax,%rdi
    336 	movq	16(%rsi),%rax
    337 	adcq	$0,%rdx
    338 	addq	%r11,%rdi
    339 	leaq	4(%r15),%r15
    340 	adcq	$0,%rdx
    341 	movq	%rdi,(%rsp)
    342 	movq	%rdx,%r13
    343 	jmp	L$1st4x
    344 .p2align	4
    345 L$1st4x:
    346 	mulq	%rbx
    347 	addq	%rax,%r10
    348 	movq	-16(%rcx,%r15,8),%rax
    349 	adcq	$0,%rdx
    350 	movq	%rdx,%r11
    351 
    352 	mulq	%rbp
    353 	addq	%rax,%r13
    354 	movq	-8(%rsi,%r15,8),%rax
    355 	adcq	$0,%rdx
    356 	addq	%r10,%r13
    357 	adcq	$0,%rdx
    358 	movq	%r13,-24(%rsp,%r15,8)
    359 	movq	%rdx,%rdi
    360 
    361 	mulq	%rbx
    362 	addq	%rax,%r11
    363 	movq	-8(%rcx,%r15,8),%rax
    364 	adcq	$0,%rdx
    365 	movq	%rdx,%r10
    366 
    367 	mulq	%rbp
    368 	addq	%rax,%rdi
    369 	movq	(%rsi,%r15,8),%rax
    370 	adcq	$0,%rdx
    371 	addq	%r11,%rdi
    372 	adcq	$0,%rdx
    373 	movq	%rdi,-16(%rsp,%r15,8)
    374 	movq	%rdx,%r13
    375 
    376 	mulq	%rbx
    377 	addq	%rax,%r10
    378 	movq	(%rcx,%r15,8),%rax
    379 	adcq	$0,%rdx
    380 	movq	%rdx,%r11
    381 
    382 	mulq	%rbp
    383 	addq	%rax,%r13
    384 	movq	8(%rsi,%r15,8),%rax
    385 	adcq	$0,%rdx
    386 	addq	%r10,%r13
    387 	adcq	$0,%rdx
    388 	movq	%r13,-8(%rsp,%r15,8)
    389 	movq	%rdx,%rdi
    390 
    391 	mulq	%rbx
    392 	addq	%rax,%r11
    393 	movq	8(%rcx,%r15,8),%rax
    394 	adcq	$0,%rdx
    395 	leaq	4(%r15),%r15
    396 	movq	%rdx,%r10
    397 
    398 	mulq	%rbp
    399 	addq	%rax,%rdi
    400 	movq	-16(%rsi,%r15,8),%rax
    401 	adcq	$0,%rdx
    402 	addq	%r11,%rdi
    403 	adcq	$0,%rdx
    404 	movq	%rdi,-32(%rsp,%r15,8)
    405 	movq	%rdx,%r13
    406 	cmpq	%r9,%r15
    407 	jb	L$1st4x
    408 
    409 	mulq	%rbx
    410 	addq	%rax,%r10
    411 	movq	-16(%rcx,%r15,8),%rax
    412 	adcq	$0,%rdx
    413 	movq	%rdx,%r11
    414 
    415 	mulq	%rbp
    416 	addq	%rax,%r13
    417 	movq	-8(%rsi,%r15,8),%rax
    418 	adcq	$0,%rdx
    419 	addq	%r10,%r13
    420 	adcq	$0,%rdx
    421 	movq	%r13,-24(%rsp,%r15,8)
    422 	movq	%rdx,%rdi
    423 
    424 	mulq	%rbx
    425 	addq	%rax,%r11
    426 	movq	-8(%rcx,%r15,8),%rax
    427 	adcq	$0,%rdx
    428 	movq	%rdx,%r10
    429 
    430 	mulq	%rbp
    431 	addq	%rax,%rdi
    432 	movq	(%rsi),%rax
    433 	adcq	$0,%rdx
    434 	addq	%r11,%rdi
    435 	adcq	$0,%rdx
    436 	movq	%rdi,-16(%rsp,%r15,8)
    437 	movq	%rdx,%r13
    438 
    439 	xorq	%rdi,%rdi
    440 	addq	%r10,%r13
    441 	adcq	$0,%rdi
    442 	movq	%r13,-8(%rsp,%r15,8)
    443 	movq	%rdi,(%rsp,%r15,8)
    444 
    445 	leaq	1(%r14),%r14
    446 .p2align	2
    447 L$outer4x:
    448 	movq	(%r12,%r14,8),%rbx
    449 	xorq	%r15,%r15
    450 	movq	(%rsp),%r10
    451 	movq	%r8,%rbp
    452 	mulq	%rbx
    453 	addq	%rax,%r10
    454 	movq	(%rcx),%rax
    455 	adcq	$0,%rdx
    456 
    457 	imulq	%r10,%rbp
    458 	movq	%rdx,%r11
    459 
    460 	mulq	%rbp
    461 	addq	%rax,%r10
    462 	movq	8(%rsi),%rax
    463 	adcq	$0,%rdx
    464 	movq	%rdx,%rdi
    465 
    466 	mulq	%rbx
    467 	addq	%rax,%r11
    468 	movq	8(%rcx),%rax
    469 	adcq	$0,%rdx
    470 	addq	8(%rsp),%r11
    471 	adcq	$0,%rdx
    472 	movq	%rdx,%r10
    473 
    474 	mulq	%rbp
    475 	addq	%rax,%rdi
    476 	movq	16(%rsi),%rax
    477 	adcq	$0,%rdx
    478 	addq	%r11,%rdi
    479 	leaq	4(%r15),%r15
    480 	adcq	$0,%rdx
    481 	movq	%rdi,(%rsp)
    482 	movq	%rdx,%r13
    483 	jmp	L$inner4x
    484 .p2align	4
    485 L$inner4x:
    486 	mulq	%rbx
    487 	addq	%rax,%r10
    488 	movq	-16(%rcx,%r15,8),%rax
    489 	adcq	$0,%rdx
    490 	addq	-16(%rsp,%r15,8),%r10
    491 	adcq	$0,%rdx
    492 	movq	%rdx,%r11
    493 
    494 	mulq	%rbp
    495 	addq	%rax,%r13
    496 	movq	-8(%rsi,%r15,8),%rax
    497 	adcq	$0,%rdx
    498 	addq	%r10,%r13
    499 	adcq	$0,%rdx
    500 	movq	%r13,-24(%rsp,%r15,8)
    501 	movq	%rdx,%rdi
    502 
    503 	mulq	%rbx
    504 	addq	%rax,%r11
    505 	movq	-8(%rcx,%r15,8),%rax
    506 	adcq	$0,%rdx
    507 	addq	-8(%rsp,%r15,8),%r11
    508 	adcq	$0,%rdx
    509 	movq	%rdx,%r10
    510 
    511 	mulq	%rbp
    512 	addq	%rax,%rdi
    513 	movq	(%rsi,%r15,8),%rax
    514 	adcq	$0,%rdx
    515 	addq	%r11,%rdi
    516 	adcq	$0,%rdx
    517 	movq	%rdi,-16(%rsp,%r15,8)
    518 	movq	%rdx,%r13
    519 
    520 	mulq	%rbx
    521 	addq	%rax,%r10
    522 	movq	(%rcx,%r15,8),%rax
    523 	adcq	$0,%rdx
    524 	addq	(%rsp,%r15,8),%r10
    525 	adcq	$0,%rdx
    526 	movq	%rdx,%r11
    527 
    528 	mulq	%rbp
    529 	addq	%rax,%r13
    530 	movq	8(%rsi,%r15,8),%rax
    531 	adcq	$0,%rdx
    532 	addq	%r10,%r13
    533 	adcq	$0,%rdx
    534 	movq	%r13,-8(%rsp,%r15,8)
    535 	movq	%rdx,%rdi
    536 
    537 	mulq	%rbx
    538 	addq	%rax,%r11
    539 	movq	8(%rcx,%r15,8),%rax
    540 	adcq	$0,%rdx
    541 	addq	8(%rsp,%r15,8),%r11
    542 	adcq	$0,%rdx
    543 	leaq	4(%r15),%r15
    544 	movq	%rdx,%r10
    545 
    546 	mulq	%rbp
    547 	addq	%rax,%rdi
    548 	movq	-16(%rsi,%r15,8),%rax
    549 	adcq	$0,%rdx
    550 	addq	%r11,%rdi
    551 	adcq	$0,%rdx
    552 	movq	%rdi,-32(%rsp,%r15,8)
    553 	movq	%rdx,%r13
    554 	cmpq	%r9,%r15
    555 	jb	L$inner4x
    556 
    557 	mulq	%rbx
    558 	addq	%rax,%r10
    559 	movq	-16(%rcx,%r15,8),%rax
    560 	adcq	$0,%rdx
    561 	addq	-16(%rsp,%r15,8),%r10
    562 	adcq	$0,%rdx
    563 	movq	%rdx,%r11
    564 
    565 	mulq	%rbp
    566 	addq	%rax,%r13
    567 	movq	-8(%rsi,%r15,8),%rax
    568 	adcq	$0,%rdx
    569 	addq	%r10,%r13
    570 	adcq	$0,%rdx
    571 	movq	%r13,-24(%rsp,%r15,8)
    572 	movq	%rdx,%rdi
    573 
    574 	mulq	%rbx
    575 	addq	%rax,%r11
    576 	movq	-8(%rcx,%r15,8),%rax
    577 	adcq	$0,%rdx
    578 	addq	-8(%rsp,%r15,8),%r11
    579 	adcq	$0,%rdx
    580 	leaq	1(%r14),%r14
    581 	movq	%rdx,%r10
    582 
    583 	mulq	%rbp
    584 	addq	%rax,%rdi
    585 	movq	(%rsi),%rax
    586 	adcq	$0,%rdx
    587 	addq	%r11,%rdi
    588 	adcq	$0,%rdx
    589 	movq	%rdi,-16(%rsp,%r15,8)
    590 	movq	%rdx,%r13
    591 
    592 	xorq	%rdi,%rdi
    593 	addq	%r10,%r13
    594 	adcq	$0,%rdi
    595 	addq	(%rsp,%r9,8),%r13
    596 	adcq	$0,%rdi
    597 	movq	%r13,-8(%rsp,%r15,8)
    598 	movq	%rdi,(%rsp,%r15,8)
    599 
    600 	cmpq	%r9,%r14
    601 	jb	L$outer4x
    602 	movq	16(%rsp,%r9,8),%rdi
    603 	leaq	-4(%r9),%r15
    604 	movq	0(%rsp),%rax
    605 	pxor	%xmm0,%xmm0
    606 	movq	8(%rsp),%rdx
    607 	shrq	$2,%r15
    608 	leaq	(%rsp),%rsi
    609 	xorq	%r14,%r14
    610 
    611 	subq	0(%rcx),%rax
    612 	movq	16(%rsi),%rbx
    613 	movq	24(%rsi),%rbp
    614 	sbbq	8(%rcx),%rdx
    615 	jmp	L$sub4x
    616 .p2align	4
    617 L$sub4x:
    618 	movq	%rax,0(%rdi,%r14,8)
    619 	movq	%rdx,8(%rdi,%r14,8)
    620 	sbbq	16(%rcx,%r14,8),%rbx
    621 	movq	32(%rsi,%r14,8),%rax
    622 	movq	40(%rsi,%r14,8),%rdx
    623 	sbbq	24(%rcx,%r14,8),%rbp
    624 	movq	%rbx,16(%rdi,%r14,8)
    625 	movq	%rbp,24(%rdi,%r14,8)
    626 	sbbq	32(%rcx,%r14,8),%rax
    627 	movq	48(%rsi,%r14,8),%rbx
    628 	movq	56(%rsi,%r14,8),%rbp
    629 	sbbq	40(%rcx,%r14,8),%rdx
    630 	leaq	4(%r14),%r14
    631 	decq	%r15
    632 	jnz	L$sub4x
    633 
    634 	movq	%rax,0(%rdi,%r14,8)
    635 	movq	32(%rsi,%r14,8),%rax
    636 	sbbq	16(%rcx,%r14,8),%rbx
    637 	movq	%rdx,8(%rdi,%r14,8)
    638 	sbbq	24(%rcx,%r14,8),%rbp
    639 	movq	%rbx,16(%rdi,%r14,8)
    640 
    641 	sbbq	$0,%rax
    642 	movq	%rbp,24(%rdi,%r14,8)
    643 	xorq	%r14,%r14
    644 	andq	%rax,%rsi
    645 	notq	%rax
    646 	movq	%rdi,%rcx
    647 	andq	%rax,%rcx
    648 	leaq	-4(%r9),%r15
    649 	orq	%rcx,%rsi
    650 	shrq	$2,%r15
    651 
    652 	movdqu	(%rsi),%xmm1
    653 	movdqa	%xmm0,(%rsp)
    654 	movdqu	%xmm1,(%rdi)
    655 	jmp	L$copy4x
    656 .p2align	4
    657 L$copy4x:
    658 	movdqu	16(%rsi,%r14,1),%xmm2
    659 	movdqu	32(%rsi,%r14,1),%xmm1
    660 	movdqa	%xmm0,16(%rsp,%r14,1)
    661 	movdqu	%xmm2,16(%rdi,%r14,1)
    662 	movdqa	%xmm0,32(%rsp,%r14,1)
    663 	movdqu	%xmm1,32(%rdi,%r14,1)
    664 	leaq	32(%r14),%r14
    665 	decq	%r15
    666 	jnz	L$copy4x
    667 
    668 	movdqu	16(%rsi,%r14,1),%xmm2
    669 	movdqa	%xmm0,16(%rsp,%r14,1)
    670 	movdqu	%xmm2,16(%rdi,%r14,1)
    671 	movq	8(%rsp,%r9,8),%rsi
    672 
    673 	movq	$1,%rax
    674 	movq	-48(%rsi),%r15
    675 
    676 	movq	-40(%rsi),%r14
    677 
    678 	movq	-32(%rsi),%r13
    679 
    680 	movq	-24(%rsi),%r12
    681 
    682 	movq	-16(%rsi),%rbp
    683 
    684 	movq	-8(%rsi),%rbx
    685 
    686 	leaq	(%rsi),%rsp
    687 
    688 L$mul4x_epilogue:
    689 	.byte	0xf3,0xc3
    690 
    691 
    692 
    693 
    694 
    695 .p2align	5
    696 bn_sqr8x_mont:
    697 
    698 	movq	%rsp,%rax
    699 
    700 L$sqr8x_enter:
    701 	pushq	%rbx
    702 
    703 	pushq	%rbp
    704 
    705 	pushq	%r12
    706 
    707 	pushq	%r13
    708 
    709 	pushq	%r14
    710 
    711 	pushq	%r15
    712 
    713 L$sqr8x_prologue:
    714 
    715 	movl	%r9d,%r10d
    716 	shll	$3,%r9d
    717 	shlq	$3+2,%r10
    718 	negq	%r9
    719 
    720 
    721 
    722 
    723 
    724 
    725 	leaq	-64(%rsp,%r9,2),%r11
    726 	movq	%rsp,%rbp
    727 	movq	(%r8),%r8
    728 	subq	%rsi,%r11
    729 	andq	$4095,%r11
    730 	cmpq	%r11,%r10
    731 	jb	L$sqr8x_sp_alt
    732 	subq	%r11,%rbp
    733 	leaq	-64(%rbp,%r9,2),%rbp
    734 	jmp	L$sqr8x_sp_done
    735 
    736 .p2align	5
    737 L$sqr8x_sp_alt:
    738 	leaq	4096-64(,%r9,2),%r10
    739 	leaq	-64(%rbp,%r9,2),%rbp
    740 	subq	%r10,%r11
    741 	movq	$0,%r10
    742 	cmovcq	%r10,%r11
    743 	subq	%r11,%rbp
    744 L$sqr8x_sp_done:
    745 	andq	$-64,%rbp
    746 	movq	%rsp,%r11
    747 	subq	%rbp,%r11
    748 	andq	$-4096,%r11
    749 	leaq	(%r11,%rbp,1),%rsp
    750 	movq	(%rsp),%r10
    751 	cmpq	%rbp,%rsp
    752 	ja	L$sqr8x_page_walk
    753 	jmp	L$sqr8x_page_walk_done
    754 
    755 .p2align	4
    756 L$sqr8x_page_walk:
    757 	leaq	-4096(%rsp),%rsp
    758 	movq	(%rsp),%r10
    759 	cmpq	%rbp,%rsp
    760 	ja	L$sqr8x_page_walk
    761 L$sqr8x_page_walk_done:
    762 
    763 	movq	%r9,%r10
    764 	negq	%r9
    765 
    766 	movq	%r8,32(%rsp)
    767 	movq	%rax,40(%rsp)
    768 
    769 L$sqr8x_body:
    770 
    771 .byte	102,72,15,110,209
    772 	pxor	%xmm0,%xmm0
    773 .byte	102,72,15,110,207
    774 .byte	102,73,15,110,218
    775 	call	_bn_sqr8x_internal
    776 
    777 
    778 
    779 
    780 	leaq	(%rdi,%r9,1),%rbx
    781 	movq	%r9,%rcx
    782 	movq	%r9,%rdx
    783 .byte	102,72,15,126,207
    784 	sarq	$3+2,%rcx
    785 	jmp	L$sqr8x_sub
    786 
    787 .p2align	5
    788 L$sqr8x_sub:
    789 	movq	0(%rbx),%r12
    790 	movq	8(%rbx),%r13
    791 	movq	16(%rbx),%r14
    792 	movq	24(%rbx),%r15
    793 	leaq	32(%rbx),%rbx
    794 	sbbq	0(%rbp),%r12
    795 	sbbq	8(%rbp),%r13
    796 	sbbq	16(%rbp),%r14
    797 	sbbq	24(%rbp),%r15
    798 	leaq	32(%rbp),%rbp
    799 	movq	%r12,0(%rdi)
    800 	movq	%r13,8(%rdi)
    801 	movq	%r14,16(%rdi)
    802 	movq	%r15,24(%rdi)
    803 	leaq	32(%rdi),%rdi
    804 	incq	%rcx
    805 	jnz	L$sqr8x_sub
    806 
    807 	sbbq	$0,%rax
    808 	leaq	(%rbx,%r9,1),%rbx
    809 	leaq	(%rdi,%r9,1),%rdi
    810 
    811 .byte	102,72,15,110,200
    812 	pxor	%xmm0,%xmm0
    813 	pshufd	$0,%xmm1,%xmm1
    814 	movq	40(%rsp),%rsi
    815 
    816 	jmp	L$sqr8x_cond_copy
    817 
    818 .p2align	5
    819 L$sqr8x_cond_copy:
    820 	movdqa	0(%rbx),%xmm2
    821 	movdqa	16(%rbx),%xmm3
    822 	leaq	32(%rbx),%rbx
    823 	movdqu	0(%rdi),%xmm4
    824 	movdqu	16(%rdi),%xmm5
    825 	leaq	32(%rdi),%rdi
    826 	movdqa	%xmm0,-32(%rbx)
    827 	movdqa	%xmm0,-16(%rbx)
    828 	movdqa	%xmm0,-32(%rbx,%rdx,1)
    829 	movdqa	%xmm0,-16(%rbx,%rdx,1)
    830 	pcmpeqd	%xmm1,%xmm0
    831 	pand	%xmm1,%xmm2
    832 	pand	%xmm1,%xmm3
    833 	pand	%xmm0,%xmm4
    834 	pand	%xmm0,%xmm5
    835 	pxor	%xmm0,%xmm0
    836 	por	%xmm2,%xmm4
    837 	por	%xmm3,%xmm5
    838 	movdqu	%xmm4,-32(%rdi)
    839 	movdqu	%xmm5,-16(%rdi)
    840 	addq	$32,%r9
    841 	jnz	L$sqr8x_cond_copy
    842 
    843 	movq	$1,%rax
    844 	movq	-48(%rsi),%r15
    845 
    846 	movq	-40(%rsi),%r14
    847 
    848 	movq	-32(%rsi),%r13
    849 
    850 	movq	-24(%rsi),%r12
    851 
    852 	movq	-16(%rsi),%rbp
    853 
    854 	movq	-8(%rsi),%rbx
    855 
    856 	leaq	(%rsi),%rsp
    857 
    858 L$sqr8x_epilogue:
    859 	.byte	0xf3,0xc3
    860 
    861 
    862 .byte	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
    863 .p2align	4
    864 #endif
    865