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	$tmp0,-3		# 0x0ffffffc0ffffffc
226 my ($h0,$h1,$h2,$r0,$r1,$rs1,$d0,$d1,$d2) =
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
375 	 daddu	$d2,$h2,$padbit
381 	 daddu	$d2,$tmp1
382 	 daddu	$d2,$tmp0
399 	dmultu	($rs1,$d2)		# h2*5*r1
402 	mflo	($tmp2,$rs1,$d2)
404 	dmultu	($r0,$d2)		# h2*r0
407 	mflo	($tmp3,$r0,$d2)
430 $code.=<<___ if ($flavour =~ /nubi/i);	# optimize non-nubi epilogue
461 	li	$in0,-4			# final reduction
537 .asciiz	"Poly1305 for MIPS64, CRYPTOGAMS by \@dot-asm"
543 # 32-bit code path
735 my ($d0,$d1,$d2,$d3) =
746 	.mask	$SAVED_REGS_MASK,-4
758 $code.=<<___ if ($flavour =~ /nubi/i);	# optimize non-nubi prologue
800 	lw	$d2,8($inp)
811 	srlv	$at,$d2,$t1
812 	sllv	$d2,$d2,$shr
816 	or	$d2,$d2,$at
824 	sllv	$at,$d2,$t1
825 	srlv	$d2,$d2,$shr
829 	or	$d2,$d2,$at
837 	lwl	$d2,8+MSB($inp)
841 	lwr	$d2,8+LSB($inp)
848 	wsbh	$d2,$d2
852 	rotr	$d2,$d2,16
871 	srl	$at,$d2,24
873 	srl	$t0,$d2,8
875 	andi	$t1,$d2,0xFF00
876 	sll	$d2,$d2,24
879 	or	$d2,$at
883 	or	$d2,$t0
893 	srl	$t0,$h4,2		# modulo-scheduled reduction
909 	addu	$d2,$d2,$h2
911 	sltu	$h2,$d2,$h2
912 	addu	$d2,$d2,$h1
913 	sltu	$h1,$d2,$h1
925 	maddu	$rs2,$d2		# d2*s2
934 	maddu	$rs3,$d2		# d2*s3
943 	maddu	$r0,$d2			# d2*r0
954 	maddu	$r1,$d2			# d2*r1
980 	multu	($rs2,$d2)		# d2*s2
981 	mflo	($a3,$rs2,$d2)
982 	mfhi	($t1,$rs2,$d2)
1010 	multu	($rs3,$d2)		# d2*s3
1013 	mflo	($a3,$rs3,$d2)
1014 	mfhi	($t1,$rs3,$d2)
1048 	multu	($r0,$d2)		# d2*r0
1051 	mflo	($at,$r0,$d2)
1052 	mfhi	($t0,$r0,$d2)
1093 	multu	($r1,$d2)		# d2*r1
1097 	mflo	($at,$r1,$d2)
1098 	mfhi	($t0,$r1,$d2)
1140 $code.=<<___ if ($flavour =~ /nubi/i);	# optimize non-nubi prologue
1169 	li	$in0,-4			# final reduction
1265 .asciiz	"Poly1305 for MIPS32, CRYPTOGAMS by \@dot-asm"