Lines Matching +full:ror +full:- +full:default
1 // This file is generated from a similarly-named Perl script in the BoringSSL
14 // Copyright 2014-2016 The OpenSSL Project Authors. All Rights Reserved.
33 // over code generated with "default" compiler:
35 // SHA256-hw SHA256(*) SHA512
36 // Apple A7 1.97 10.5 (+33%) 6.73 (-1%(**))
37 // Cortex-A53 2.38 15.5 (+115%) 10.0 (+150%(***))
38 // Cortex-A57 2.31 11.6 (+86%) 7.51 (+260%(***))
40 // X-Gene 20.0 (+100%) 12.8 (+300%(***))
45 // (**) The result is a trade-off: it's possible to improve it by
47 // on Cortex-A53 (or by 4 cycles per round).
48 // (***) Super-impressive coefficients over gcc-generated code are
50 // generated with -mgeneral-regs-only is significanty faster
51 // and the gap is only 40-90%.
54 # include <ring-core/arm_arch.h>
67 stp x29,x30,[sp,#-128]!
94 ror x16,x24,#14
96 eor x6,x24,x24,ror#23
102 eor x16,x16,x6,ror#18 // Sigma1(e)
103 ror x6,x20,#28
105 eor x17,x20,x20,ror#5
110 eor x17,x6,x17,ror#34 // Sigma0(a)
119 ror x16,x23,#14
121 eor x7,x23,x23,ror#23
127 eor x16,x16,x7,ror#18 // Sigma1(e)
128 ror x7,x27,#28
130 eor x17,x27,x27,ror#5
135 eor x17,x7,x17,ror#34 // Sigma0(a)
143 ror x16,x22,#14
145 eor x8,x22,x22,ror#23
151 eor x16,x16,x8,ror#18 // Sigma1(e)
152 ror x8,x26,#28
154 eor x17,x26,x26,ror#5
159 eor x17,x8,x17,ror#34 // Sigma0(a)
168 ror x16,x21,#14
170 eor x9,x21,x21,ror#23
176 eor x16,x16,x9,ror#18 // Sigma1(e)
177 ror x9,x25,#28
179 eor x17,x25,x25,ror#5
184 eor x17,x9,x17,ror#34 // Sigma0(a)
192 ror x16,x20,#14
194 eor x10,x20,x20,ror#23
200 eor x16,x16,x10,ror#18 // Sigma1(e)
201 ror x10,x24,#28
203 eor x17,x24,x24,ror#5
208 eor x17,x10,x17,ror#34 // Sigma0(a)
217 ror x16,x27,#14
219 eor x11,x27,x27,ror#23
225 eor x16,x16,x11,ror#18 // Sigma1(e)
226 ror x11,x23,#28
228 eor x17,x23,x23,ror#5
233 eor x17,x11,x17,ror#34 // Sigma0(a)
241 ror x16,x26,#14
243 eor x12,x26,x26,ror#23
249 eor x16,x16,x12,ror#18 // Sigma1(e)
250 ror x12,x22,#28
252 eor x17,x22,x22,ror#5
257 eor x17,x12,x17,ror#34 // Sigma0(a)
266 ror x16,x25,#14
268 eor x13,x25,x25,ror#23
274 eor x16,x16,x13,ror#18 // Sigma1(e)
275 ror x13,x21,#28
277 eor x17,x21,x21,ror#5
282 eor x17,x13,x17,ror#34 // Sigma0(a)
290 ror x16,x24,#14
292 eor x14,x24,x24,ror#23
298 eor x16,x16,x14,ror#18 // Sigma1(e)
299 ror x14,x20,#28
301 eor x17,x20,x20,ror#5
306 eor x17,x14,x17,ror#34 // Sigma0(a)
315 ror x16,x23,#14
317 eor x15,x23,x23,ror#23
323 eor x16,x16,x15,ror#18 // Sigma1(e)
324 ror x15,x27,#28
326 eor x17,x27,x27,ror#5
331 eor x17,x15,x17,ror#34 // Sigma0(a)
339 ror x16,x22,#14
341 eor x0,x22,x22,ror#23
347 eor x16,x16,x0,ror#18 // Sigma1(e)
348 ror x0,x26,#28
350 eor x17,x26,x26,ror#5
355 eor x17,x0,x17,ror#34 // Sigma0(a)
365 ror x16,x21,#14
367 eor x6,x21,x21,ror#23
373 eor x16,x16,x6,ror#18 // Sigma1(e)
374 ror x6,x25,#28
376 eor x17,x25,x25,ror#5
381 eor x17,x6,x17,ror#34 // Sigma0(a)
390 ror x16,x20,#14
392 eor x7,x20,x20,ror#23
398 eor x16,x16,x7,ror#18 // Sigma1(e)
399 ror x7,x24,#28
401 eor x17,x24,x24,ror#5
406 eor x17,x7,x17,ror#34 // Sigma0(a)
416 ror x16,x27,#14
418 eor x8,x27,x27,ror#23
424 eor x16,x16,x8,ror#18 // Sigma1(e)
425 ror x8,x23,#28
427 eor x17,x23,x23,ror#5
432 eor x17,x8,x17,ror#34 // Sigma0(a)
442 ror x16,x26,#14
444 eor x9,x26,x26,ror#23
450 eor x16,x16,x9,ror#18 // Sigma1(e)
451 ror x9,x22,#28
453 eor x17,x22,x22,ror#5
458 eor x17,x9,x17,ror#34 // Sigma0(a)
468 ror x16,x25,#14
470 ror x9,x4,#1
472 ror x8,x1,#19
474 ror x10,x21,#28
476 eor x16,x16,x25,ror#18
477 eor x9,x9,x4,ror#8
480 eor x16,x16,x25,ror#41 // Sigma1(e)
481 eor x10,x10,x21,ror#34
484 eor x8,x8,x1,ror#61
488 eor x17,x10,x21,ror#39 // Sigma0(a)
500 ror x16,x24,#14
502 ror x10,x5,#1
504 ror x9,x2,#19
506 ror x11,x20,#28
508 eor x16,x16,x24,ror#18
509 eor x10,x10,x5,ror#8
512 eor x16,x16,x24,ror#41 // Sigma1(e)
513 eor x11,x11,x20,ror#34
516 eor x9,x9,x2,ror#61
520 eor x17,x11,x20,ror#39 // Sigma0(a)
531 ror x16,x23,#14
533 ror x11,x6,#1
535 ror x10,x3,#19
537 ror x12,x27,#28
539 eor x16,x16,x23,ror#18
540 eor x11,x11,x6,ror#8
543 eor x16,x16,x23,ror#41 // Sigma1(e)
544 eor x12,x12,x27,ror#34
547 eor x10,x10,x3,ror#61
551 eor x17,x12,x27,ror#39 // Sigma0(a)
562 ror x16,x22,#14
564 ror x12,x7,#1
566 ror x11,x4,#19
568 ror x13,x26,#28
570 eor x16,x16,x22,ror#18
571 eor x12,x12,x7,ror#8
574 eor x16,x16,x22,ror#41 // Sigma1(e)
575 eor x13,x13,x26,ror#34
578 eor x11,x11,x4,ror#61
582 eor x17,x13,x26,ror#39 // Sigma0(a)
593 ror x16,x21,#14
595 ror x13,x8,#1
597 ror x12,x5,#19
599 ror x14,x25,#28
601 eor x16,x16,x21,ror#18
602 eor x13,x13,x8,ror#8
605 eor x16,x16,x21,ror#41 // Sigma1(e)
606 eor x14,x14,x25,ror#34
609 eor x12,x12,x5,ror#61
613 eor x17,x14,x25,ror#39 // Sigma0(a)
624 ror x16,x20,#14
626 ror x14,x9,#1
628 ror x13,x6,#19
630 ror x15,x24,#28
632 eor x16,x16,x20,ror#18
633 eor x14,x14,x9,ror#8
636 eor x16,x16,x20,ror#41 // Sigma1(e)
637 eor x15,x15,x24,ror#34
640 eor x13,x13,x6,ror#61
644 eor x17,x15,x24,ror#39 // Sigma0(a)
655 ror x16,x27,#14
657 ror x15,x10,#1
659 ror x14,x7,#19
661 ror x0,x23,#28
663 eor x16,x16,x27,ror#18
664 eor x15,x15,x10,ror#8
667 eor x16,x16,x27,ror#41 // Sigma1(e)
668 eor x0,x0,x23,ror#34
671 eor x14,x14,x7,ror#61
675 eor x17,x0,x23,ror#39 // Sigma0(a)
686 ror x16,x26,#14
688 ror x0,x11,#1
690 ror x15,x8,#19
692 ror x1,x22,#28
694 eor x16,x16,x26,ror#18
695 eor x0,x0,x11,ror#8
698 eor x16,x16,x26,ror#41 // Sigma1(e)
699 eor x1,x1,x22,ror#34
702 eor x15,x15,x8,ror#61
706 eor x17,x1,x22,ror#39 // Sigma0(a)
717 ror x16,x25,#14
719 ror x1,x12,#1
721 ror x0,x9,#19
723 ror x2,x21,#28
725 eor x16,x16,x25,ror#18
726 eor x1,x1,x12,ror#8
729 eor x16,x16,x25,ror#41 // Sigma1(e)
730 eor x2,x2,x21,ror#34
733 eor x0,x0,x9,ror#61
737 eor x17,x2,x21,ror#39 // Sigma0(a)
748 ror x16,x24,#14
750 ror x2,x13,#1
752 ror x1,x10,#19
754 ror x3,x20,#28
756 eor x16,x16,x24,ror#18
757 eor x2,x2,x13,ror#8
760 eor x16,x16,x24,ror#41 // Sigma1(e)
761 eor x3,x3,x20,ror#34
764 eor x1,x1,x10,ror#61
768 eor x17,x3,x20,ror#39 // Sigma0(a)
779 ror x16,x23,#14
781 ror x3,x14,#1
783 ror x2,x11,#19
785 ror x4,x27,#28
787 eor x16,x16,x23,ror#18
788 eor x3,x3,x14,ror#8
791 eor x16,x16,x23,ror#41 // Sigma1(e)
792 eor x4,x4,x27,ror#34
795 eor x2,x2,x11,ror#61
799 eor x17,x4,x27,ror#39 // Sigma0(a)
810 ror x16,x22,#14
812 ror x4,x15,#1
814 ror x3,x12,#19
816 ror x5,x26,#28
818 eor x16,x16,x22,ror#18
819 eor x4,x4,x15,ror#8
822 eor x16,x16,x22,ror#41 // Sigma1(e)
823 eor x5,x5,x26,ror#34
826 eor x3,x3,x12,ror#61
830 eor x17,x5,x26,ror#39 // Sigma0(a)
841 ror x16,x21,#14
843 ror x5,x0,#1
845 ror x4,x13,#19
847 ror x6,x25,#28
849 eor x16,x16,x21,ror#18
850 eor x5,x5,x0,ror#8
853 eor x16,x16,x21,ror#41 // Sigma1(e)
854 eor x6,x6,x25,ror#34
857 eor x4,x4,x13,ror#61
861 eor x17,x6,x25,ror#39 // Sigma0(a)
872 ror x16,x20,#14
874 ror x6,x1,#1
876 ror x5,x14,#19
878 ror x7,x24,#28
880 eor x16,x16,x20,ror#18
881 eor x6,x6,x1,ror#8
884 eor x16,x16,x20,ror#41 // Sigma1(e)
885 eor x7,x7,x24,ror#34
888 eor x5,x5,x14,ror#61
892 eor x17,x7,x24,ror#39 // Sigma0(a)
903 ror x16,x27,#14
905 ror x7,x2,#1
907 ror x6,x15,#19
909 ror x8,x23,#28
911 eor x16,x16,x27,ror#18
912 eor x7,x7,x2,ror#8
915 eor x16,x16,x27,ror#41 // Sigma1(e)
916 eor x8,x8,x23,ror#34
919 eor x6,x6,x15,ror#61
923 eor x17,x8,x23,ror#39 // Sigma0(a)
934 ror x16,x26,#14
936 ror x8,x3,#1
938 ror x7,x0,#19
940 ror x9,x22,#28
942 eor x16,x16,x26,ror#18
943 eor x8,x8,x3,ror#8
946 eor x16,x16,x26,ror#41 // Sigma1(e)
947 eor x9,x9,x22,ror#34
950 eor x7,x7,x0,ror#61
954 eor x17,x9,x22,ror#39 // Sigma0(a)
965 ror x16,x25,#14
967 ror x9,x4,#1
969 ror x8,x1,#19
971 ror x10,x21,#28
973 eor x16,x16,x25,ror#18
974 eor x9,x9,x4,ror#8
977 eor x16,x16,x25,ror#41 // Sigma1(e)
978 eor x10,x10,x21,ror#34
981 eor x8,x8,x1,ror#61
985 eor x17,x10,x21,ror#39 // Sigma0(a)
1029 .size sha512_block_data_order,.-sha512_block_data_order
1076 .size .LK512,.-.LK512
1082 .section .note.GNU-stack,"",%progbits