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

1 /* SPDX-License-Identifier: GPL-2.0-or-later */
5 # Copyright 2023- IBM Corp. All rights reserved
17 # 4.  c += d; b ^= c; b <<<= 7
24 # 4 blocks (a b c d)
43 #include <asm/asm-offsets.h>
44 #include <asm/asm-compat.h>
81 	stdu 1,-752(1)
100 	SAVE_GPR 31, 248, 1
114 	SAVE_VRS 31, 176, 9
133 	SAVE_VSX 31, 464, 9
149 	RESTORE_VRS 31, 176, 9
168 	RESTORE_VSX 31, 464, 9
187 	RESTORE_GPR 31, 248, 1
198 	vadduwm 0, 0, 4
214 	  vpermxor 31, 31, 19, 25
223 	  vadduwm 27, 27, 31
224 	vxor 4, 4, 8
235 	vrlw 4, 4, 25  #
244 	vadduwm 0, 0, 4
262 	  vpermxor 31, 31, 19, 25
271 	  vadduwm 27, 27, 31
274 	vxor 4, 4, 8
282 	vrlw 4, 4, 28  #
298 	vadduwm 3, 3, 4
308 	  vpermxor 31, 31, 16, 25
318 	  vadduwm 26, 26, 31
325 	vxor 4, 4, 9
336 	vrlw 4, 4, 25
346 	vadduwm 3, 3, 4
358 	  vpermxor 31, 31, 16, 25
368 	  vadduwm 26, 26, 31
378 	vxor 4, 4, 9
386 	vrlw 4, 4, 28
396 	vadduwm 0, 0, 4
408 	vxor 4, 4, 8
412 	vrlw 4, 4, 21
416 	vadduwm 0, 0, 4
428 	vxor 4, 4, 8
432 	vrlw 4, 4, 23
441 	vadduwm 3, 3, 4
453 	vxor 4, 4, 9
457 	vrlw 4, 4, 21
461 	vadduwm 3, 3, 4
473 	vxor 4, 4, 9
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
520 	add 16, 14, 4
525 	lxvw4x 4, 20, 9
536 	lxvw4x 15, 31, 9
542 	xxlxor \S+33, \S+33, 4
573 	stxvw4x \S+47, 31, 16
587 	# r17 - r31 mainly for Write_256 macro.
602 	li	31, 240
617 	vmrghw	4, 0, 1
619 	vsldoi	30, 4, 5, 8		# vr30 counter, 4 (0, 1, 2, 3)
639 	vspltisw 25, 4
645 	vadduwm	31, 30, 25		# counter = (0, 1, 2, 3) + (4, 4, 4, 4)
647 	xxlor	31, 32+31, 32+31
663 	xxspltw  32+4, 17, 0
692 	vadduwm	28, 28, 31	# increase counter
694 	xxspltw  32+31, 19, 3
707 	TP_4x 4, 5, 6, 7
726 	addi	15, 15, -256	# len -=256
728 	xxlor	5, 32+31, 32+31
729 	xxlor	32+31, 31, 31
730 	vadduwm	28, 28, 31
731 	xxlor	32+31, 5, 5
733 	TP_4x 16+4, 16+5, 16+6, 16+7
744 	addi	15, 15, -256	# len +=256
750 	vadduwm	31, 30, 24
752 	xxlor	31, 32+31, 32+31
785 	vspltw  4, 17, 0
807 	TP_4x 4, 5, 6, 7
814 	addi	15, 15, -256	# len += 256
817 	vspltisw 25, 4