2 Copyright (C) 2010 Free Software Foundation, Inc.
3 Contributed by Intel Corporation.
4 This file is part of the GNU C Library.
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, write to the Free
18 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
24 #include "asm-syntax.h"
26 #define CFI_PUSH(REG) \
27 cfi_adjust_cfa_offset (4); \
28 cfi_rel_offset (REG, 0)
30 #define CFI_POP(REG) \
31 cfi_adjust_cfa_offset (-4); \
34 #define PUSH(REG) pushl REG; CFI_PUSH (REG)
35 #define POP(REG) popl REG; CFI_POP (REG)
45 # define SETRTNVAL movl DEST(%esp), %eax
49 # define ENTRANCE PUSH (%ebx);
50 # define RETURN_END POP (%ebx); ret
51 # define RETURN RETURN_END; CFI_PUSH (%ebx)
52 # define PARMS 8 /* Preserve EBX. */
53 # define JMPTBL(I, B) I - B
55 /* Load an entry in a jump table into EBX and branch to it. TABLE is a
56 jump table with relative offsets. */
57 # define BRANCH_TO_JMPTBL_ENTRY(TABLE) \
58 /* We first load PC into EBX. */ \
59 call __i686.get_pc_thunk.bx; \
60 /* Get the address of the jump table. */ \
61 add $(TABLE - .), %ebx; \
62 /* Get the entry and convert the relative offset to the \
63 absolute address. */ \
64 add (%ebx,%ecx,4), %ebx; \
66 /* We loaded the jump table and adjuested EDX. Go. */ \
69 .section .gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
70 .globl __i686.get_pc_thunk.bx
71 .hidden __i686.get_pc_thunk.bx
73 .type __i686.get_pc_thunk.bx,@function
74 __i686.get_pc_thunk.bx:
79 # define RETURN_END ret
80 # define RETURN RETURN_END
82 # define JMPTBL(I, B) I
84 /* Branch to an entry in a jump table. TABLE is a jump table with
86 # define BRANCH_TO_JMPTBL_ENTRY(TABLE) \
91 .section .text.sse2,"ax",@progbits
92 #if defined SHARED && !defined NOT_IN_libc && !defined USE_AS_BZERO
93 ENTRY (__memset_chk_sse2)
96 jb HIDDEN_JUMPTARGET (__chk_fail)
97 END (__memset_chk_sse2)
106 movzbl CHR(%esp), %eax
108 /* Fill the whole EAX with pattern. */
113 movl DEST(%esp), %edx
117 L(write_less32bytes):
118 BRANCH_TO_JMPTBL_ENTRY (L(table_less_32bytes))
121 .pushsection .rodata.sse2,"a",@progbits
123 L(table_less_32bytes):
124 .int JMPTBL (L(write_0bytes), L(table_less_32bytes))
125 .int JMPTBL (L(write_1bytes), L(table_less_32bytes))
126 .int JMPTBL (L(write_2bytes), L(table_less_32bytes))
127 .int JMPTBL (L(write_3bytes), L(table_less_32bytes))
128 .int JMPTBL (L(write_4bytes), L(table_less_32bytes))
129 .int JMPTBL (L(write_5bytes), L(table_less_32bytes))
130 .int JMPTBL (L(write_6bytes), L(table_less_32bytes))
131 .int JMPTBL (L(write_7bytes), L(table_less_32bytes))
132 .int JMPTBL (L(write_8bytes), L(table_less_32bytes))
133 .int JMPTBL (L(write_9bytes), L(table_less_32bytes))
134 .int JMPTBL (L(write_10bytes), L(table_less_32bytes))
135 .int JMPTBL (L(write_11bytes), L(table_less_32bytes))
136 .int JMPTBL (L(write_12bytes), L(table_less_32bytes))
137 .int JMPTBL (L(write_13bytes), L(table_less_32bytes))
138 .int JMPTBL (L(write_14bytes), L(table_less_32bytes))
139 .int JMPTBL (L(write_15bytes), L(table_less_32bytes))
140 .int JMPTBL (L(write_16bytes), L(table_less_32bytes))
141 .int JMPTBL (L(write_17bytes), L(table_less_32bytes))
142 .int JMPTBL (L(write_18bytes), L(table_less_32bytes))
143 .int JMPTBL (L(write_19bytes), L(table_less_32bytes))
144 .int JMPTBL (L(write_20bytes), L(table_less_32bytes))
145 .int JMPTBL (L(write_21bytes), L(table_less_32bytes))
146 .int JMPTBL (L(write_22bytes), L(table_less_32bytes))
147 .int JMPTBL (L(write_23bytes), L(table_less_32bytes))
148 .int JMPTBL (L(write_24bytes), L(table_less_32bytes))
149 .int JMPTBL (L(write_25bytes), L(table_less_32bytes))
150 .int JMPTBL (L(write_26bytes), L(table_less_32bytes))
151 .int JMPTBL (L(write_27bytes), L(table_less_32bytes))
152 .int JMPTBL (L(write_28bytes), L(table_less_32bytes))
153 .int JMPTBL (L(write_29bytes), L(table_less_32bytes))
154 .int JMPTBL (L(write_30bytes), L(table_less_32bytes))
155 .int JMPTBL (L(write_31bytes), L(table_less_32bytes))
239 /* ECX > 32 and EDX is 4 byte aligned. */
241 /* Fill xmm0 with the pattern. */
246 punpcklbw %xmm0, %xmm0
247 pshufd $0, %xmm0, %xmm0
251 /* ECX > 32 and EDX is not 16 byte aligned. */
264 jge L(128bytesormore)
266 L(aligned_16_less128bytes):
267 BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
271 #ifdef SHARED_CACHE_SIZE
273 mov $SHARED_CACHE_SIZE, %ebx
276 call __i686.get_pc_thunk.bx
277 add $_GLOBAL_OFFSET_TABLE_, %ebx
278 mov __x86_shared_cache_size@GOTOFF(%ebx), %ebx
281 mov __x86_shared_cache_size, %ebx
285 jae L(128bytesormore_nt_start)
288 #ifdef DATA_CACHE_SIZE
290 cmp $DATA_CACHE_SIZE, %ecx
293 call __i686.get_pc_thunk.bx
294 add $_GLOBAL_OFFSET_TABLE_, %ebx
295 cmp __x86_data_cache_size@GOTOFF(%ebx), %ecx
298 cmp __x86_data_cache_size, %ecx
302 jae L(128bytes_L2_normal)
304 L(128bytesormore_normal):
307 movdqa %xmm0, 0x10(%edx)
308 movdqa %xmm0, 0x20(%edx)
309 movdqa %xmm0, 0x30(%edx)
310 movdqa %xmm0, 0x40(%edx)
311 movdqa %xmm0, 0x50(%edx)
312 movdqa %xmm0, 0x60(%edx)
313 movdqa %xmm0, 0x70(%edx)
315 jl L(128bytesless_normal)
320 movdqa %xmm0, 0x10(%edx)
321 movdqa %xmm0, 0x20(%edx)
322 movdqa %xmm0, 0x30(%edx)
323 movdqa %xmm0, 0x40(%edx)
324 movdqa %xmm0, 0x50(%edx)
325 movdqa %xmm0, 0x60(%edx)
326 movdqa %xmm0, 0x70(%edx)
328 jge L(128bytesormore_normal)
330 L(128bytesless_normal):
332 BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
335 L(128bytes_L2_normal):
336 prefetcht0 0x380(%edx)
337 prefetcht0 0x3c0(%edx)
340 movaps %xmm0, 0x10(%edx)
341 movaps %xmm0, 0x20(%edx)
342 movaps %xmm0, 0x30(%edx)
343 movaps %xmm0, 0x40(%edx)
344 movaps %xmm0, 0x50(%edx)
345 movaps %xmm0, 0x60(%edx)
346 movaps %xmm0, 0x70(%edx)
349 jge L(128bytes_L2_normal)
351 L(128bytesless_L2_normal):
352 BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
354 L(128bytesormore_nt_start):
357 L(128bytesormore_shared_cache_loop):
358 prefetcht0 0x3c0(%edx)
359 prefetcht0 0x380(%edx)
362 movdqa %xmm0, 0x10(%edx)
363 movdqa %xmm0, 0x20(%edx)
364 movdqa %xmm0, 0x30(%edx)
365 movdqa %xmm0, 0x40(%edx)
366 movdqa %xmm0, 0x50(%edx)
367 movdqa %xmm0, 0x60(%edx)
368 movdqa %xmm0, 0x70(%edx)
371 jge L(128bytesormore_shared_cache_loop)
373 jb L(shared_cache_loop_end)
375 L(128bytesormore_nt):
377 movntdq %xmm0, (%edx)
378 movntdq %xmm0, 0x10(%edx)
379 movntdq %xmm0, 0x20(%edx)
380 movntdq %xmm0, 0x30(%edx)
381 movntdq %xmm0, 0x40(%edx)
382 movntdq %xmm0, 0x50(%edx)
383 movntdq %xmm0, 0x60(%edx)
384 movntdq %xmm0, 0x70(%edx)
387 jge L(128bytesormore_nt)
389 L(shared_cache_loop_end):
390 #if defined DATA_CACHE_SIZE || !defined SHARED
393 BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
396 .pushsection .rodata.sse2,"a",@progbits
398 L(table_16_128bytes):
399 .int JMPTBL (L(aligned_16_0bytes), L(table_16_128bytes))
400 .int JMPTBL (L(aligned_16_1bytes), L(table_16_128bytes))
401 .int JMPTBL (L(aligned_16_2bytes), L(table_16_128bytes))
402 .int JMPTBL (L(aligned_16_3bytes), L(table_16_128bytes))
403 .int JMPTBL (L(aligned_16_4bytes), L(table_16_128bytes))
404 .int JMPTBL (L(aligned_16_5bytes), L(table_16_128bytes))
405 .int JMPTBL (L(aligned_16_6bytes), L(table_16_128bytes))
406 .int JMPTBL (L(aligned_16_7bytes), L(table_16_128bytes))
407 .int JMPTBL (L(aligned_16_8bytes), L(table_16_128bytes))
408 .int JMPTBL (L(aligned_16_9bytes), L(table_16_128bytes))
409 .int JMPTBL (L(aligned_16_10bytes), L(table_16_128bytes))
410 .int JMPTBL (L(aligned_16_11bytes), L(table_16_128bytes))
411 .int JMPTBL (L(aligned_16_12bytes), L(table_16_128bytes))
412 .int JMPTBL (L(aligned_16_13bytes), L(table_16_128bytes))
413 .int JMPTBL (L(aligned_16_14bytes), L(table_16_128bytes))
414 .int JMPTBL (L(aligned_16_15bytes), L(table_16_128bytes))
415 .int JMPTBL (L(aligned_16_16bytes), L(table_16_128bytes))
416 .int JMPTBL (L(aligned_16_17bytes), L(table_16_128bytes))
417 .int JMPTBL (L(aligned_16_18bytes), L(table_16_128bytes))
418 .int JMPTBL (L(aligned_16_19bytes), L(table_16_128bytes))
419 .int JMPTBL (L(aligned_16_20bytes), L(table_16_128bytes))
420 .int JMPTBL (L(aligned_16_21bytes), L(table_16_128bytes))
421 .int JMPTBL (L(aligned_16_22bytes), L(table_16_128bytes))
422 .int JMPTBL (L(aligned_16_23bytes), L(table_16_128bytes))
423 .int JMPTBL (L(aligned_16_24bytes), L(table_16_128bytes))
424 .int JMPTBL (L(aligned_16_25bytes), L(table_16_128bytes))
425 .int JMPTBL (L(aligned_16_26bytes), L(table_16_128bytes))
426 .int JMPTBL (L(aligned_16_27bytes), L(table_16_128bytes))
427 .int JMPTBL (L(aligned_16_28bytes), L(table_16_128bytes))
428 .int JMPTBL (L(aligned_16_29bytes), L(table_16_128bytes))
429 .int JMPTBL (L(aligned_16_30bytes), L(table_16_128bytes))
430 .int JMPTBL (L(aligned_16_31bytes), L(table_16_128bytes))
431 .int JMPTBL (L(aligned_16_32bytes), L(table_16_128bytes))
432 .int JMPTBL (L(aligned_16_33bytes), L(table_16_128bytes))
433 .int JMPTBL (L(aligned_16_34bytes), L(table_16_128bytes))
434 .int JMPTBL (L(aligned_16_35bytes), L(table_16_128bytes))
435 .int JMPTBL (L(aligned_16_36bytes), L(table_16_128bytes))
436 .int JMPTBL (L(aligned_16_37bytes), L(table_16_128bytes))
437 .int JMPTBL (L(aligned_16_38bytes), L(table_16_128bytes))
438 .int JMPTBL (L(aligned_16_39bytes), L(table_16_128bytes))
439 .int JMPTBL (L(aligned_16_40bytes), L(table_16_128bytes))
440 .int JMPTBL (L(aligned_16_41bytes), L(table_16_128bytes))
441 .int JMPTBL (L(aligned_16_42bytes), L(table_16_128bytes))
442 .int JMPTBL (L(aligned_16_43bytes), L(table_16_128bytes))
443 .int JMPTBL (L(aligned_16_44bytes), L(table_16_128bytes))
444 .int JMPTBL (L(aligned_16_45bytes), L(table_16_128bytes))
445 .int JMPTBL (L(aligned_16_46bytes), L(table_16_128bytes))
446 .int JMPTBL (L(aligned_16_47bytes), L(table_16_128bytes))
447 .int JMPTBL (L(aligned_16_48bytes), L(table_16_128bytes))
448 .int JMPTBL (L(aligned_16_49bytes), L(table_16_128bytes))
449 .int JMPTBL (L(aligned_16_50bytes), L(table_16_128bytes))
450 .int JMPTBL (L(aligned_16_51bytes), L(table_16_128bytes))
451 .int JMPTBL (L(aligned_16_52bytes), L(table_16_128bytes))
452 .int JMPTBL (L(aligned_16_53bytes), L(table_16_128bytes))
453 .int JMPTBL (L(aligned_16_54bytes), L(table_16_128bytes))
454 .int JMPTBL (L(aligned_16_55bytes), L(table_16_128bytes))
455 .int JMPTBL (L(aligned_16_56bytes), L(table_16_128bytes))
456 .int JMPTBL (L(aligned_16_57bytes), L(table_16_128bytes))
457 .int JMPTBL (L(aligned_16_58bytes), L(table_16_128bytes))
458 .int JMPTBL (L(aligned_16_59bytes), L(table_16_128bytes))
459 .int JMPTBL (L(aligned_16_60bytes), L(table_16_128bytes))
460 .int JMPTBL (L(aligned_16_61bytes), L(table_16_128bytes))
461 .int JMPTBL (L(aligned_16_62bytes), L(table_16_128bytes))
462 .int JMPTBL (L(aligned_16_63bytes), L(table_16_128bytes))
463 .int JMPTBL (L(aligned_16_64bytes), L(table_16_128bytes))
464 .int JMPTBL (L(aligned_16_65bytes), L(table_16_128bytes))
465 .int JMPTBL (L(aligned_16_66bytes), L(table_16_128bytes))
466 .int JMPTBL (L(aligned_16_67bytes), L(table_16_128bytes))
467 .int JMPTBL (L(aligned_16_68bytes), L(table_16_128bytes))
468 .int JMPTBL (L(aligned_16_69bytes), L(table_16_128bytes))
469 .int JMPTBL (L(aligned_16_70bytes), L(table_16_128bytes))
470 .int JMPTBL (L(aligned_16_71bytes), L(table_16_128bytes))
471 .int JMPTBL (L(aligned_16_72bytes), L(table_16_128bytes))
472 .int JMPTBL (L(aligned_16_73bytes), L(table_16_128bytes))
473 .int JMPTBL (L(aligned_16_74bytes), L(table_16_128bytes))
474 .int JMPTBL (L(aligned_16_75bytes), L(table_16_128bytes))
475 .int JMPTBL (L(aligned_16_76bytes), L(table_16_128bytes))
476 .int JMPTBL (L(aligned_16_77bytes), L(table_16_128bytes))
477 .int JMPTBL (L(aligned_16_78bytes), L(table_16_128bytes))
478 .int JMPTBL (L(aligned_16_79bytes), L(table_16_128bytes))
479 .int JMPTBL (L(aligned_16_80bytes), L(table_16_128bytes))
480 .int JMPTBL (L(aligned_16_81bytes), L(table_16_128bytes))
481 .int JMPTBL (L(aligned_16_82bytes), L(table_16_128bytes))
482 .int JMPTBL (L(aligned_16_83bytes), L(table_16_128bytes))
483 .int JMPTBL (L(aligned_16_84bytes), L(table_16_128bytes))
484 .int JMPTBL (L(aligned_16_85bytes), L(table_16_128bytes))
485 .int JMPTBL (L(aligned_16_86bytes), L(table_16_128bytes))
486 .int JMPTBL (L(aligned_16_87bytes), L(table_16_128bytes))
487 .int JMPTBL (L(aligned_16_88bytes), L(table_16_128bytes))
488 .int JMPTBL (L(aligned_16_89bytes), L(table_16_128bytes))
489 .int JMPTBL (L(aligned_16_90bytes), L(table_16_128bytes))
490 .int JMPTBL (L(aligned_16_91bytes), L(table_16_128bytes))
491 .int JMPTBL (L(aligned_16_92bytes), L(table_16_128bytes))
492 .int JMPTBL (L(aligned_16_93bytes), L(table_16_128bytes))
493 .int JMPTBL (L(aligned_16_94bytes), L(table_16_128bytes))
494 .int JMPTBL (L(aligned_16_95bytes), L(table_16_128bytes))
495 .int JMPTBL (L(aligned_16_96bytes), L(table_16_128bytes))
496 .int JMPTBL (L(aligned_16_97bytes), L(table_16_128bytes))
497 .int JMPTBL (L(aligned_16_98bytes), L(table_16_128bytes))
498 .int JMPTBL (L(aligned_16_99bytes), L(table_16_128bytes))
499 .int JMPTBL (L(aligned_16_100bytes), L(table_16_128bytes))
500 .int JMPTBL (L(aligned_16_101bytes), L(table_16_128bytes))
501 .int JMPTBL (L(aligned_16_102bytes), L(table_16_128bytes))
502 .int JMPTBL (L(aligned_16_103bytes), L(table_16_128bytes))
503 .int JMPTBL (L(aligned_16_104bytes), L(table_16_128bytes))
504 .int JMPTBL (L(aligned_16_105bytes), L(table_16_128bytes))
505 .int JMPTBL (L(aligned_16_106bytes), L(table_16_128bytes))
506 .int JMPTBL (L(aligned_16_107bytes), L(table_16_128bytes))
507 .int JMPTBL (L(aligned_16_108bytes), L(table_16_128bytes))
508 .int JMPTBL (L(aligned_16_109bytes), L(table_16_128bytes))
509 .int JMPTBL (L(aligned_16_110bytes), L(table_16_128bytes))
510 .int JMPTBL (L(aligned_16_111bytes), L(table_16_128bytes))
511 .int JMPTBL (L(aligned_16_112bytes), L(table_16_128bytes))
512 .int JMPTBL (L(aligned_16_113bytes), L(table_16_128bytes))
513 .int JMPTBL (L(aligned_16_114bytes), L(table_16_128bytes))
514 .int JMPTBL (L(aligned_16_115bytes), L(table_16_128bytes))
515 .int JMPTBL (L(aligned_16_116bytes), L(table_16_128bytes))
516 .int JMPTBL (L(aligned_16_117bytes), L(table_16_128bytes))
517 .int JMPTBL (L(aligned_16_118bytes), L(table_16_128bytes))
518 .int JMPTBL (L(aligned_16_119bytes), L(table_16_128bytes))
519 .int JMPTBL (L(aligned_16_120bytes), L(table_16_128bytes))
520 .int JMPTBL (L(aligned_16_121bytes), L(table_16_128bytes))
521 .int JMPTBL (L(aligned_16_122bytes), L(table_16_128bytes))
522 .int JMPTBL (L(aligned_16_123bytes), L(table_16_128bytes))
523 .int JMPTBL (L(aligned_16_124bytes), L(table_16_128bytes))
524 .int JMPTBL (L(aligned_16_125bytes), L(table_16_128bytes))
525 .int JMPTBL (L(aligned_16_126bytes), L(table_16_128bytes))
526 .int JMPTBL (L(aligned_16_127bytes), L(table_16_128bytes))
530 L(aligned_16_112bytes):
531 movdqa %xmm0, -112(%edx)
532 L(aligned_16_96bytes):
533 movdqa %xmm0, -96(%edx)
534 L(aligned_16_80bytes):
535 movdqa %xmm0, -80(%edx)
536 L(aligned_16_64bytes):
537 movdqa %xmm0, -64(%edx)
538 L(aligned_16_48bytes):
539 movdqa %xmm0, -48(%edx)
540 L(aligned_16_32bytes):
541 movdqa %xmm0, -32(%edx)
542 L(aligned_16_16bytes):
543 movdqa %xmm0, -16(%edx)
544 L(aligned_16_0bytes):
549 L(aligned_16_113bytes):
550 movdqa %xmm0, -113(%edx)
551 L(aligned_16_97bytes):
552 movdqa %xmm0, -97(%edx)
553 L(aligned_16_81bytes):
554 movdqa %xmm0, -81(%edx)
555 L(aligned_16_65bytes):
556 movdqa %xmm0, -65(%edx)
557 L(aligned_16_49bytes):
558 movdqa %xmm0, -49(%edx)
559 L(aligned_16_33bytes):
560 movdqa %xmm0, -33(%edx)
561 L(aligned_16_17bytes):
562 movdqa %xmm0, -17(%edx)
563 L(aligned_16_1bytes):
569 L(aligned_16_114bytes):
570 movdqa %xmm0, -114(%edx)
571 L(aligned_16_98bytes):
572 movdqa %xmm0, -98(%edx)
573 L(aligned_16_82bytes):
574 movdqa %xmm0, -82(%edx)
575 L(aligned_16_66bytes):
576 movdqa %xmm0, -66(%edx)
577 L(aligned_16_50bytes):
578 movdqa %xmm0, -50(%edx)
579 L(aligned_16_34bytes):
580 movdqa %xmm0, -34(%edx)
581 L(aligned_16_18bytes):
582 movdqa %xmm0, -18(%edx)
583 L(aligned_16_2bytes):
589 L(aligned_16_115bytes):
590 movdqa %xmm0, -115(%edx)
591 L(aligned_16_99bytes):
592 movdqa %xmm0, -99(%edx)
593 L(aligned_16_83bytes):
594 movdqa %xmm0, -83(%edx)
595 L(aligned_16_67bytes):
596 movdqa %xmm0, -67(%edx)
597 L(aligned_16_51bytes):
598 movdqa %xmm0, -51(%edx)
599 L(aligned_16_35bytes):
600 movdqa %xmm0, -35(%edx)
601 L(aligned_16_19bytes):
602 movdqa %xmm0, -19(%edx)
603 L(aligned_16_3bytes):
610 L(aligned_16_116bytes):
611 movdqa %xmm0, -116(%edx)
612 L(aligned_16_100bytes):
613 movdqa %xmm0, -100(%edx)
614 L(aligned_16_84bytes):
615 movdqa %xmm0, -84(%edx)
616 L(aligned_16_68bytes):
617 movdqa %xmm0, -68(%edx)
618 L(aligned_16_52bytes):
619 movdqa %xmm0, -52(%edx)
620 L(aligned_16_36bytes):
621 movdqa %xmm0, -36(%edx)
622 L(aligned_16_20bytes):
623 movdqa %xmm0, -20(%edx)
624 L(aligned_16_4bytes):
630 L(aligned_16_117bytes):
631 movdqa %xmm0, -117(%edx)
632 L(aligned_16_101bytes):
633 movdqa %xmm0, -101(%edx)
634 L(aligned_16_85bytes):
635 movdqa %xmm0, -85(%edx)
636 L(aligned_16_69bytes):
637 movdqa %xmm0, -69(%edx)
638 L(aligned_16_53bytes):
639 movdqa %xmm0, -53(%edx)
640 L(aligned_16_37bytes):
641 movdqa %xmm0, -37(%edx)
642 L(aligned_16_21bytes):
643 movdqa %xmm0, -21(%edx)
644 L(aligned_16_5bytes):
651 L(aligned_16_118bytes):
652 movdqa %xmm0, -118(%edx)
653 L(aligned_16_102bytes):
654 movdqa %xmm0, -102(%edx)
655 L(aligned_16_86bytes):
656 movdqa %xmm0, -86(%edx)
657 L(aligned_16_70bytes):
658 movdqa %xmm0, -70(%edx)
659 L(aligned_16_54bytes):
660 movdqa %xmm0, -54(%edx)
661 L(aligned_16_38bytes):
662 movdqa %xmm0, -38(%edx)
663 L(aligned_16_22bytes):
664 movdqa %xmm0, -22(%edx)
665 L(aligned_16_6bytes):
672 L(aligned_16_119bytes):
673 movdqa %xmm0, -119(%edx)
674 L(aligned_16_103bytes):
675 movdqa %xmm0, -103(%edx)
676 L(aligned_16_87bytes):
677 movdqa %xmm0, -87(%edx)
678 L(aligned_16_71bytes):
679 movdqa %xmm0, -71(%edx)
680 L(aligned_16_55bytes):
681 movdqa %xmm0, -55(%edx)
682 L(aligned_16_39bytes):
683 movdqa %xmm0, -39(%edx)
684 L(aligned_16_23bytes):
685 movdqa %xmm0, -23(%edx)
686 L(aligned_16_7bytes):
694 L(aligned_16_120bytes):
695 movdqa %xmm0, -120(%edx)
696 L(aligned_16_104bytes):
697 movdqa %xmm0, -104(%edx)
698 L(aligned_16_88bytes):
699 movdqa %xmm0, -88(%edx)
700 L(aligned_16_72bytes):
701 movdqa %xmm0, -72(%edx)
702 L(aligned_16_56bytes):
703 movdqa %xmm0, -56(%edx)
704 L(aligned_16_40bytes):
705 movdqa %xmm0, -40(%edx)
706 L(aligned_16_24bytes):
707 movdqa %xmm0, -24(%edx)
708 L(aligned_16_8bytes):
714 L(aligned_16_121bytes):
715 movdqa %xmm0, -121(%edx)
716 L(aligned_16_105bytes):
717 movdqa %xmm0, -105(%edx)
718 L(aligned_16_89bytes):
719 movdqa %xmm0, -89(%edx)
720 L(aligned_16_73bytes):
721 movdqa %xmm0, -73(%edx)
722 L(aligned_16_57bytes):
723 movdqa %xmm0, -57(%edx)
724 L(aligned_16_41bytes):
725 movdqa %xmm0, -41(%edx)
726 L(aligned_16_25bytes):
727 movdqa %xmm0, -25(%edx)
728 L(aligned_16_9bytes):
735 L(aligned_16_122bytes):
736 movdqa %xmm0, -122(%edx)
737 L(aligned_16_106bytes):
738 movdqa %xmm0, -106(%edx)
739 L(aligned_16_90bytes):
740 movdqa %xmm0, -90(%edx)
741 L(aligned_16_74bytes):
742 movdqa %xmm0, -74(%edx)
743 L(aligned_16_58bytes):
744 movdqa %xmm0, -58(%edx)
745 L(aligned_16_42bytes):
746 movdqa %xmm0, -42(%edx)
747 L(aligned_16_26bytes):
748 movdqa %xmm0, -26(%edx)
749 L(aligned_16_10bytes):
750 movq %xmm0, -10(%edx)
756 L(aligned_16_123bytes):
757 movdqa %xmm0, -123(%edx)
758 L(aligned_16_107bytes):
759 movdqa %xmm0, -107(%edx)
760 L(aligned_16_91bytes):
761 movdqa %xmm0, -91(%edx)
762 L(aligned_16_75bytes):
763 movdqa %xmm0, -75(%edx)
764 L(aligned_16_59bytes):
765 movdqa %xmm0, -59(%edx)
766 L(aligned_16_43bytes):
767 movdqa %xmm0, -43(%edx)
768 L(aligned_16_27bytes):
769 movdqa %xmm0, -27(%edx)
770 L(aligned_16_11bytes):
771 movq %xmm0, -11(%edx)
778 L(aligned_16_124bytes):
779 movdqa %xmm0, -124(%edx)
780 L(aligned_16_108bytes):
781 movdqa %xmm0, -108(%edx)
782 L(aligned_16_92bytes):
783 movdqa %xmm0, -92(%edx)
784 L(aligned_16_76bytes):
785 movdqa %xmm0, -76(%edx)
786 L(aligned_16_60bytes):
787 movdqa %xmm0, -60(%edx)
788 L(aligned_16_44bytes):
789 movdqa %xmm0, -44(%edx)
790 L(aligned_16_28bytes):
791 movdqa %xmm0, -28(%edx)
792 L(aligned_16_12bytes):
793 movq %xmm0, -12(%edx)
799 L(aligned_16_125bytes):
800 movdqa %xmm0, -125(%edx)
801 L(aligned_16_109bytes):
802 movdqa %xmm0, -109(%edx)
803 L(aligned_16_93bytes):
804 movdqa %xmm0, -93(%edx)
805 L(aligned_16_77bytes):
806 movdqa %xmm0, -77(%edx)
807 L(aligned_16_61bytes):
808 movdqa %xmm0, -61(%edx)
809 L(aligned_16_45bytes):
810 movdqa %xmm0, -45(%edx)
811 L(aligned_16_29bytes):
812 movdqa %xmm0, -29(%edx)
813 L(aligned_16_13bytes):
814 movq %xmm0, -13(%edx)
821 L(aligned_16_126bytes):
822 movdqa %xmm0, -126(%edx)
823 L(aligned_16_110bytes):
824 movdqa %xmm0, -110(%edx)
825 L(aligned_16_94bytes):
826 movdqa %xmm0, -94(%edx)
827 L(aligned_16_78bytes):
828 movdqa %xmm0, -78(%edx)
829 L(aligned_16_62bytes):
830 movdqa %xmm0, -62(%edx)
831 L(aligned_16_46bytes):
832 movdqa %xmm0, -46(%edx)
833 L(aligned_16_30bytes):
834 movdqa %xmm0, -30(%edx)
835 L(aligned_16_14bytes):
836 movq %xmm0, -14(%edx)
843 L(aligned_16_127bytes):
844 movdqa %xmm0, -127(%edx)
845 L(aligned_16_111bytes):
846 movdqa %xmm0, -111(%edx)
847 L(aligned_16_95bytes):
848 movdqa %xmm0, -95(%edx)
849 L(aligned_16_79bytes):
850 movdqa %xmm0, -79(%edx)
851 L(aligned_16_63bytes):
852 movdqa %xmm0, -63(%edx)
853 L(aligned_16_47bytes):
854 movdqa %xmm0, -47(%edx)
855 L(aligned_16_31bytes):
856 movdqa %xmm0, -31(%edx)
857 L(aligned_16_15bytes):
858 movq %xmm0, -15(%edx)