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)
87 	SAVE_GPR 18, 144, 1
92 	SAVE_GPR 23, 184, 1
106 	SAVE_VRS 23, 48, 9
120 	SAVE_VSX 18, 256, 9
125 	SAVE_VSX 23, 336, 9
141 	RESTORE_VRS 23, 48, 9
155 	RESTORE_VSX 18, 256, 9
160 	RESTORE_VSX 23, 336, 9
174 	RESTORE_GPR 18, 144, 1
179 	RESTORE_GPR 23, 184, 1
204 	  vadduwm 18, 18, 22
205 	  vadduwm 19, 19, 23
213 	  vpermxor 30, 30, 18, 25
231 	  vxor 23, 23, 27
242 	  vrlw 23, 23, 25
250 	  vadduwm 18, 18, 22
251 	  vadduwm 19, 19, 23
261 	  vpermxor 30, 30, 18, 25
273 	xxlor	32+28, 23, 23
281 	  vxor 23, 23, 27
289 	  vrlw 23, 23, 28
301 	  vadduwm 18, 18, 23
310 	  vpermxor 29, 29, 18, 25
328 	  vxor 23, 23, 24
339 	  vrlw 23, 23, 25
349 	  vadduwm 18, 18, 23
360 	  vpermxor 29, 29, 18, 25
374 	xxlor	32+28, 23, 23
381 	  vxor 23, 23, 24
389 	  vrlw 23, 23, 28
432 	vrlw 4, 4, 23
433 	vrlw 5, 5, 23
434 	vrlw 6, 6, 23
435 	vrlw 7, 7, 23
474 	vrlw 5, 5, 23
475 	vrlw 6, 6, 23
476 	vrlw 7, 7, 23
477 	vrlw 4, 4, 23
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
523 	lxvw4x 2, 18, 9
528 	lxvw4x 7, 23, 9
557 	stxvw4x \S+40, 18, 16
563 	stxvw4x \S+45, 23, 16
587 	# r17 - r31 mainly for Write_256 macro.
589 	li	18, 32
594 	li	23, 112
609 	lxvw4x	50, 18, 3		#  vr18, key 2
622 	vspltisw 23, 7
636 	xxlor	18, 50, 50
652 	xxlor	23, 32+23, 32+23
667 	xxspltw  32+8, 18, 0
668 	xxspltw  32+9, 18, 1
669 	xxspltw  32+10, 18, 2
670 	xxspltw  32+11, 18, 3
679 	xxspltw  32+18, 16, 2
685 	xxspltw  32+23, 17, 3
686 	xxspltw  32+24, 18, 0
687 	xxspltw  32+25, 18, 1
688 	xxspltw  32+26, 18, 2
689 	xxspltw  32+27, 18, 3
717 	xxlor	50, 18, 18
726 	addi	15, 15, -256	# len -=256
739 	xxlor	34, 18, 18
744 	addi	15, 15, -256	# len +=256
766 	lxvw4x	50, 18, 3		#  vr18, key 2
770 	vspltisw 23, 7
789 	vspltw  8, 18, 0
790 	vspltw  9, 18, 1
791 	vspltw  10, 18, 2
792 	vspltw  11, 18, 3
814 	addi	15, 15, -256	# len += 256