powerpc/crypto/chacha-p10le-8x.S

1 /* SPDX-License-Identifier: GPL-2.0-or-later */
5 # Copyright 2023- IBM Corp. All rights reserved
43 #include <asm/asm-offsets.h>
44 #include <asm/asm-compat.h>
81 	stdu 1,-752(1)
200 	vadduwm 2, 2, 6
219 	vadduwm 11, 11, 15
226 	vxor 6, 6, 10
227 	vxor 7, 7, 11
237 	vrlw 6, 6, 25
246 	vadduwm 2, 2, 6
267 	vadduwm 11, 11, 15
276 	vxor 6, 6, 10
277 	vxor 7, 7, 11
284 	vrlw 6, 6, 28
296 	vadduwm 1, 1, 6
315 	vadduwm 11, 11, 12
323 	vxor 6, 6, 11
334 	vrlw 6, 6, 25
344 	vadduwm 1, 1, 6
365 	vadduwm 11, 11, 12
376 	vxor 6, 6, 11
384 	vrlw 6, 6, 28
398 	vadduwm 2, 2, 6
407 	vadduwm 11, 11, 15
410 	vxor 6, 6, 10
411 	vxor 7, 7, 11
414 	vrlw 6, 6, 21
418 	vadduwm 2, 2, 6
427 	vadduwm 11, 11, 15
430 	vxor 6, 6, 10
431 	vxor 7, 7, 11
434 	vrlw 6, 6, 23
439 	vadduwm 1, 1, 6
447 	vadduwm 11, 11, 12
451 	vxor 6, 6, 11
455 	vrlw 6, 6, 21
459 	vadduwm 1, 1, 6
467 	vadduwm 11, 11, 12
471 	vxor 6, 6, 11
475 	vrlw 6, 6, 23
483 	xxmrghw  11, 32+\a2, 32+\a3	# a2, a3, b2, b3
486 	xxpermdi	32+\a0, 10, 11, 0	# a0, a1, a2, a3
487 	xxpermdi	32+\a1, 10, 11, 3	# b0, b1, b2, b3
494 	vadduwm \S+0, \S+0, 16-\S
495 	vadduwm \S+4, \S+4, 17-\S
496 	vadduwm \S+8, \S+8, 18-\S
497 	vadduwm \S+12, \S+12, 19-\S
499 	vadduwm \S+1, \S+1, 16-\S
500 	vadduwm \S+5, \S+5, 17-\S
501 	vadduwm \S+9, \S+9, 18-\S
502 	vadduwm \S+13, \S+13, 19-\S
504 	vadduwm \S+2, \S+2, 16-\S
505 	vadduwm \S+6, \S+6, 17-\S
506 	vadduwm \S+10, \S+10, 18-\S
507 	vadduwm \S+14, \S+14, 19-\S
509 	vadduwm	\S+3, \S+3, 16-\S
510 	vadduwm	\S+7, \S+7, 17-\S
511 	vadduwm	\S+11, \S+11, 18-\S
512 	vadduwm	\S+15, \S+15, 19-\S
527 	lxvw4x 6, 22, 9
532 	lxvw4x 11, 27, 9
544 	xxlxor \S+41, \S+41, 6
549 	xxlxor \S+46, \S+46, 11
582 	cmpdi	6, 0
587 	# r17 - r31 mainly for Write_256 macro.
604 	mr 15, 6			# len
624 	addis	11, 2, permx@toc@ha
625 	addi	11, 11, permx@toc@l
626 	lxvw4x	32+20, 0, 11
627 	lxvw4x	32+22, 17, 11
654 	cmpdi	6, 512
665 	xxspltw  32+6, 17, 2
670 	xxspltw  32+11, 18, 3
707 	TP_4x 4, 5, 6, 7
708 	TP_4x 8, 9, 10, 11
726 	addi	15, 15, -256	# len -=256
733 	TP_4x 16+4, 16+5, 16+6, 16+7
734 	TP_4x 16+8, 16+9, 16+10, 16+11
744 	addi	15, 15, -256	# len +=256
771 	addis	11, 2, permx@toc@ha
772 	addi	11, 11, permx@toc@l
773 	lxvw4x	32+20, 0, 11
774 	lxvw4x	32+22, 17, 11
787 	vspltw  6, 17, 2
792 	vspltw  11, 18, 3
807 	TP_4x 4, 5, 6, 7
808 	TP_4x 8, 9, 10, 11
814 	addi	15, 15, -256	# len += 256