mips/crypto/poly1305-mips.pl

2 # SPDX-License-Identifier: GPL-1.0+ OR BSD-3-Clause
5 # Written by Andy Polyakov, @dot-asm, originally for the OpenSSL
16 # R1x000	~5.5/+130%	(big-endian)
17 # Octeon II	2.50/+70%	(little-endian)
21 # Add 32-bit code path.
25 # Modulo-scheduling reduction allows to omit dependency chain at the
30 # R1x000	~9.8/?		(big-endian)
31 # Octeon II	3.65/+140%	(little-endian)
32 # MT7621/1004K	4.75/?		(little-endian)
48 # - never ever touch $tp, "thread pointer", former $gp [o32 can be
50 # - copy return value to $t0, former $v0 [or to $a0 if you're adapting
52 # - on O32 populate $a4-$a7 with 'lw $aN,4*N($sp)' if necessary;
72 # 64-bit code path
204 	daddiu	$tmp0,-63		# 0x00000000ffffffc1
206 	daddiu	$tmp0,-1		# 0x0ffffffc0fffffff
209 	daddiu	$tmp0,-3		# 0x0ffffffc0ffffffc
249 	.mask	$SAVED_REGS_MASK|0x000c0000,-8
255 	.mask	$SAVED_REGS_MASK,-8
261 $code.=<<___ if ($flavour =~ /nubi/i);	# optimize non-nubi prologue
360 	dsrl	$tmp1,$h2,2		# modulo-scheduled reduction
430 $code.=<<___ if ($flavour =~ /nubi/i);	# optimize non-nubi epilogue
461 	li	$in0,-4			# final reduction
537 .asciiz	"Poly1305 for MIPS64, CRYPTOGAMS by \@dot-asm"
543 # 32-bit code path
547 my ($in0,$in1,$in2,$in3,$tmp0,$tmp1,$tmp2,$tmp3) =
610 	lw	$in2,8($inp)
621 	srlv	$tmp3,$in2,$tmp1
622 	sllv	$in2,$in2,$tmp0
626 	or	$in2,$in2,$tmp3
634 	sllv	$tmp3,$in2,$tmp1
635 	srlv	$in2,$in2,$tmp0
639 	or	$in2,$in2,$tmp3
647 	lwl	$in2,8+MSB($inp)
651 	lwr	$in2,8+LSB($inp)
658 	wsbh	$in2,$in2
662 	rotr	$in2,$in2,16
681 	srl	$tmp0,$in2,24
683 	srl	$tmp1,$in2,8
685 	andi	$tmp2,$in2,0xFF00
686 	sll	$in2,$in2,24
689 	or	$in2,$tmp0
693 	or	$in2,$tmp1
708 	and	$in2,$in2,$tmp0
713 	sw	$in2,28($ctx)
717 	srl	$tmp2,$in2,2
720 	addu	$in2,$in2,$tmp2
723 	sw	$in2,40($ctx)
746 	.mask	$SAVED_REGS_MASK,-4
758 $code.=<<___ if ($flavour =~ /nubi/i);	# optimize non-nubi prologue
893 	srl	$t0,$h4,2		# modulo-scheduled reduction
1140 $code.=<<___ if ($flavour =~ /nubi/i);	# optimize non-nubi prologue
1169 	li	$in0,-4			# final reduction
1183 	 sltu	$in2,$in1,$tmp1
1185 	 addu	$in2,$in2,$tmp2
1187 	 sltu	$in3,$in2,$tmp2
1199 	xor	$in2,$tmp2
1203 	and	$in2,$ctx
1207 	xor	$in2,$tmp2
1224 	addu	$in2,$tmp2
1225 	sltu	$tmp2,$in2,$tmp2
1226 	addu	$in2,$ctx
1227 	sltu	$ctx,$in2,$ctx
1245 	srl	$tmp0,$in2,8
1247 	srl	$tmp1,$in2,16
1249 	srl	$tmp2,$in2,24
1250 	sb	$in2, 8($mac)
1265 .asciiz	"Poly1305 for MIPS32, CRYPTOGAMS by \@dot-asm"