#! /usr/bin/env perl

# Copyright 2010-2020 The OpenSSL Project Authors. All Rights Reserved.

#

# Licensed under the OpenSSL license (the "License").  You may not use

# this file except in compliance with the License.  You can obtain a copy

# in the file LICENSE in the source distribution or at

# https://www.openssl.org/source/license.html

#

# ====================================================================

# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL

# project.

#

# Rights for redistribution and usage in source and binary forms are

# granted according to the OpenSSL license. Warranty of any kind is

# disclaimed.

# ====================================================================

# July 1999

#

# This is drop-in MIPS III/IV ISA replacement for crypto/bn/bn_asm.c.

#

# The module is designed to work with either of the "new" MIPS ABI(5),

# namely N32 or N64, offered by IRIX 6.x. It's not meant to work under

# IRIX 5.x not only because it doesn't support new ABIs but also

# because 5.x kernels put R4x00 CPU into 32-bit mode and all those

# 64-bit instructions (daddu, dmultu, etc.) found below gonna only

# cause illegal instruction exception:-(

#

# In addition the code depends on preprocessor flags set up by MIPSpro

# compiler driver (either as or cc) and therefore (probably?) can't be

# compiled by the GNU assembler. GNU C driver manages fine though...

# I mean as long as -mmips-as is specified or is the default option,

# because then it simply invokes /usr/bin/as which in turn takes

# perfect care of the preprocessor definitions. Another neat feature

# offered by the MIPSpro assembler is an optimization pass. This gave

# me the opportunity to have the code looking more regular as all those

# architecture dependent instruction rescheduling details were left to

# the assembler. Cool, huh?

#

# Performance improvement is astonishing! 'apps/openssl speed rsa dsa'

# goes way over 3 times faster!

#

#					<appro@openssl.org>

# October 2010

#

# Adapt the module even for 32-bit ABIs and other OSes. The former was

# achieved by mechanical replacement of 64-bit arithmetic instructions

# such as dmultu, daddu, etc. with their 32-bit counterparts and

# adjusting offsets denoting multiples of BN_ULONG. Above mentioned

# >3x performance improvement naturally does not apply to 32-bit code

# [because there is no instruction 32-bit compiler can't use], one

# has to content with 40-85% improvement depending on benchmark and

# key length, more for longer keys.

$flavour = shift || "o32";

while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {}

open STDOUT,">$output";

if ($flavour =~ /64|n32/i) {

	$LD="ld";

	$ST="sd";

	$MULTU="dmultu";

	$DIVU="ddivu";

	$ADDU="daddu";

	$SUBU="dsubu";

	$SRL="dsrl";

	$SLL="dsll";

	$BNSZ=8;

	$PTR_ADD="daddu";

	$PTR_SUB="dsubu";

	$SZREG=8;

	$REG_S="sd";

	$REG_L="ld";

} else {

	$LD="lw";

	$ST="sw";

	$MULTU="multu";

	$DIVU="divu";

	$ADDU="addu";

	$SUBU="subu";

	$SRL="srl";

	$SLL="sll";

	$BNSZ=4;

	$PTR_ADD="addu";

	$PTR_SUB="subu";

	$SZREG=4;

	$REG_S="sw";

	$REG_L="lw";

	$code="#if !(defined (__mips_isa_rev) && (__mips_isa_rev >= 6))\n.set     mips2\n#endif\n";

}

# Below is N32/64 register layout used in the original module.

#

($zero,$at,$v0,$v1)=map("\$$_",(0..3));

($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));

($t0,$t1,$t2,$t3,$t8,$t9)=map("\$$_",(12..15,24,25));

($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7)=map("\$$_",(16..23));

($gp,$sp,$fp,$ra)=map("\$$_",(28..31));

($ta0,$ta1,$ta2,$ta3)=($a4,$a5,$a6,$a7);

#

# No special adaptation is required for O32. NUBI on the other hand

# is treated by saving/restoring ($v1,$t0..$t3).

$gp=$v1 if ($flavour =~ /nubi/i);

$minus4=$v1;

$code.=<<___;

#include "mips_arch.h"

#if defined(_MIPS_ARCH_MIPS64R6)

# define ddivu(rs,rt)

# define mfqt(rd,rs,rt)	ddivu	rd,rs,rt

# define mfrm(rd,rs,rt)	dmodu	rd,rs,rt

#elif defined(_MIPS_ARCH_MIPS32R6)

# define divu(rs,rt)

# define mfqt(rd,rs,rt)	divu	rd,rs,rt

# define mfrm(rd,rs,rt)	modu	rd,rs,rt

#else

# define $DIVU(rs,rt)	$DIVU	$zero,rs,rt

# define mfqt(rd,rs,rt)	mflo	rd

# define mfrm(rd,rs,rt)	mfhi	rd

#endif

.rdata

.asciiz	"mips3.s, Version 1.2"

.asciiz	"MIPS II/III/IV ISA artwork by Andy Polyakov <appro\@fy.chalmers.se>"

.text

.set	noat

.align	5

.globl	bn_mul_add_words

.ent	bn_mul_add_words

bn_mul_add_words:

	.set	noreorder

	bgtz	$a2,bn_mul_add_words_internal

	move	$v0,$zero

	jr	$ra

	move	$a0,$v0

.end	bn_mul_add_words

.align	5

.ent	bn_mul_add_words_internal

bn_mul_add_words_internal:

___

$code.=<<___ if ($flavour =~ /nubi/i);

	.frame	$sp,6*$SZREG,$ra

	.mask	0x8000f008,-$SZREG

	.set	noreorder

	$PTR_SUB $sp,6*$SZREG

	$REG_S	$ra,5*$SZREG($sp)

	$REG_S	$t3,4*$SZREG($sp)

	$REG_S	$t2,3*$SZREG($sp)

	$REG_S	$t1,2*$SZREG($sp)

	$REG_S	$t0,1*$SZREG($sp)

	$REG_S	$gp,0*$SZREG($sp)

___

$code.=<<___;

	.set	reorder

	li	$minus4,-4

	and	$ta0,$a2,$minus4

	beqz	$ta0,.L_bn_mul_add_words_tail

.L_bn_mul_add_words_loop:

	$LD	$t0,0($a1)

	$MULTU	($t0,$a3)

	$LD	$t1,0($a0)

	$LD	$t2,$BNSZ($a1)

	$LD	$t3,$BNSZ($a0)

	$LD	$ta0,2*$BNSZ($a1)

	$LD	$ta1,2*$BNSZ($a0)

	$ADDU	$t1,$v0

	sltu	$v0,$t1,$v0	# All manuals say it "compares 32-bit

				# values", but it seems to work fine

				# even on 64-bit registers.

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$t1,$at

	$ADDU	$v0,$t0

	 $MULTU	($t2,$a3)

	sltu	$at,$t1,$at

	$ST	$t1,0($a0)

	$ADDU	$v0,$at

	$LD	$ta2,3*$BNSZ($a1)

	$LD	$ta3,3*$BNSZ($a0)

	$ADDU	$t3,$v0

	sltu	$v0,$t3,$v0

	mflo	($at,$t2,$a3)

	mfhi	($t2,$t2,$a3)

	$ADDU	$t3,$at

	$ADDU	$v0,$t2

	 $MULTU	($ta0,$a3)

	sltu	$at,$t3,$at

	$ST	$t3,$BNSZ($a0)

	$ADDU	$v0,$at

	subu	$a2,4

	$PTR_ADD $a0,4*$BNSZ

	$PTR_ADD $a1,4*$BNSZ

	$ADDU	$ta1,$v0

	sltu	$v0,$ta1,$v0

	mflo	($at,$ta0,$a3)

	mfhi	($ta0,$ta0,$a3)

	$ADDU	$ta1,$at

	$ADDU	$v0,$ta0

	 $MULTU	($ta2,$a3)

	sltu	$at,$ta1,$at

	$ST	$ta1,-2*$BNSZ($a0)

	$ADDU	$v0,$at

	and	$ta0,$a2,$minus4

	$ADDU	$ta3,$v0

	sltu	$v0,$ta3,$v0

	mflo	($at,$ta2,$a3)

	mfhi	($ta2,$ta2,$a3)

	$ADDU	$ta3,$at

	$ADDU	$v0,$ta2

	sltu	$at,$ta3,$at

	$ST	$ta3,-$BNSZ($a0)

	.set	noreorder

	bgtz	$ta0,.L_bn_mul_add_words_loop

	$ADDU	$v0,$at

	beqz	$a2,.L_bn_mul_add_words_return

	nop

.L_bn_mul_add_words_tail:

	.set	reorder

	$LD	$t0,0($a1)

	$MULTU	($t0,$a3)

	$LD	$t1,0($a0)

	subu	$a2,1

	$ADDU	$t1,$v0

	sltu	$v0,$t1,$v0

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$t1,$at

	$ADDU	$v0,$t0

	sltu	$at,$t1,$at

	$ST	$t1,0($a0)

	$ADDU	$v0,$at

	beqz	$a2,.L_bn_mul_add_words_return

	$LD	$t0,$BNSZ($a1)

	$MULTU	($t0,$a3)

	$LD	$t1,$BNSZ($a0)

	subu	$a2,1

	$ADDU	$t1,$v0

	sltu	$v0,$t1,$v0

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$t1,$at

	$ADDU	$v0,$t0

	sltu	$at,$t1,$at

	$ST	$t1,$BNSZ($a0)

	$ADDU	$v0,$at

	beqz	$a2,.L_bn_mul_add_words_return

	$LD	$t0,2*$BNSZ($a1)

	$MULTU	($t0,$a3)

	$LD	$t1,2*$BNSZ($a0)

	$ADDU	$t1,$v0

	sltu	$v0,$t1,$v0

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$t1,$at

	$ADDU	$v0,$t0

	sltu	$at,$t1,$at

	$ST	$t1,2*$BNSZ($a0)

	$ADDU	$v0,$at

.L_bn_mul_add_words_return:

	.set	noreorder

___

$code.=<<___ if ($flavour =~ /nubi/i);

	$REG_L	$t3,4*$SZREG($sp)

	$REG_L	$t2,3*$SZREG($sp)

	$REG_L	$t1,2*$SZREG($sp)

	$REG_L	$t0,1*$SZREG($sp)

	$REG_L	$gp,0*$SZREG($sp)

	$PTR_ADD $sp,6*$SZREG

___

$code.=<<___;

	jr	$ra

	move	$a0,$v0

.end	bn_mul_add_words_internal

.align	5

.globl	bn_mul_words

.ent	bn_mul_words

bn_mul_words:

	.set	noreorder

	bgtz	$a2,bn_mul_words_internal

	move	$v0,$zero

	jr	$ra

	move	$a0,$v0

.end	bn_mul_words

.align	5

.ent	bn_mul_words_internal

bn_mul_words_internal:

___

$code.=<<___ if ($flavour =~ /nubi/i);

	.frame	$sp,6*$SZREG,$ra

	.mask	0x8000f008,-$SZREG

	.set	noreorder

	$PTR_SUB $sp,6*$SZREG

	$REG_S	$ra,5*$SZREG($sp)

	$REG_S	$t3,4*$SZREG($sp)

	$REG_S	$t2,3*$SZREG($sp)

	$REG_S	$t1,2*$SZREG($sp)

	$REG_S	$t0,1*$SZREG($sp)

	$REG_S	$gp,0*$SZREG($sp)

___

$code.=<<___;

	.set	reorder

	li	$minus4,-4

	and	$ta0,$a2,$minus4

	beqz	$ta0,.L_bn_mul_words_tail

.L_bn_mul_words_loop:

	$LD	$t0,0($a1)

	$MULTU	($t0,$a3)

	$LD	$t2,$BNSZ($a1)

	$LD	$ta0,2*$BNSZ($a1)

	$LD	$ta2,3*$BNSZ($a1)

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$v0,$at

	sltu	$t1,$v0,$at

	 $MULTU	($t2,$a3)

	$ST	$v0,0($a0)

	$ADDU	$v0,$t1,$t0

	subu	$a2,4

	$PTR_ADD $a0,4*$BNSZ

	$PTR_ADD $a1,4*$BNSZ

	mflo	($at,$t2,$a3)

	mfhi	($t2,$t2,$a3)

	$ADDU	$v0,$at

	sltu	$t3,$v0,$at

	 $MULTU	($ta0,$a3)

	$ST	$v0,-3*$BNSZ($a0)

	$ADDU	$v0,$t3,$t2

	mflo	($at,$ta0,$a3)

	mfhi	($ta0,$ta0,$a3)

	$ADDU	$v0,$at

	sltu	$ta1,$v0,$at

	 $MULTU	($ta2,$a3)

	$ST	$v0,-2*$BNSZ($a0)

	$ADDU	$v0,$ta1,$ta0

	and	$ta0,$a2,$minus4

	mflo	($at,$ta2,$a3)

	mfhi	($ta2,$ta2,$a3)

	$ADDU	$v0,$at

	sltu	$ta3,$v0,$at

	$ST	$v0,-$BNSZ($a0)

	.set	noreorder

	bgtz	$ta0,.L_bn_mul_words_loop

	$ADDU	$v0,$ta3,$ta2

	beqz	$a2,.L_bn_mul_words_return

	nop

.L_bn_mul_words_tail:

	.set	reorder

	$LD	$t0,0($a1)

	$MULTU	($t0,$a3)

	subu	$a2,1

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$v0,$at

	sltu	$t1,$v0,$at

	$ST	$v0,0($a0)

	$ADDU	$v0,$t1,$t0

	beqz	$a2,.L_bn_mul_words_return

	$LD	$t0,$BNSZ($a1)

	$MULTU	($t0,$a3)

	subu	$a2,1

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$v0,$at

	sltu	$t1,$v0,$at

	$ST	$v0,$BNSZ($a0)

	$ADDU	$v0,$t1,$t0

	beqz	$a2,.L_bn_mul_words_return

	$LD	$t0,2*$BNSZ($a1)

	$MULTU	($t0,$a3)

	mflo	($at,$t0,$a3)

	mfhi	($t0,$t0,$a3)

	$ADDU	$v0,$at

	sltu	$t1,$v0,$at

	$ST	$v0,2*$BNSZ($a0)

	$ADDU	$v0,$t1,$t0

.L_bn_mul_words_return:

	.set	noreorder

___

$code.=<<___ if ($flavour =~ /nubi/i);

	$REG_L	$t3,4*$SZREG($sp)

	$REG_L	$t2,3*$SZREG($sp)

	$REG_L	$t1,2*$SZREG($sp)

	$REG_L	$t0,1*$SZREG($sp)

	$REG_L	$gp,0*$SZREG($sp)

	$PTR_ADD $sp,6*$SZREG

___

$code.=<<___;

	jr	$ra

	move	$a0,$v0

.end	bn_mul_words_internal

.align	5

.globl	bn_sqr_words

.ent	bn_sqr_words

bn_sqr_words:

	.set	noreorder

	bgtz	$a2,bn_sqr_words_internal

	move	$v0,$zero

	jr	$ra

	move	$a0,$v0

.end	bn_sqr_words

.align	5

.ent	bn_sqr_words_internal

bn_sqr_words_internal:

___

$code.=<<___ if ($flavour =~ /nubi/i);

	.frame	$sp,6*$SZREG,$ra

	.mask	0x8000f008,-$SZREG

	.set	noreorder

	$PTR_SUB $sp,6*$SZREG

	$REG_S	$ra,5*$SZREG($sp)

	$REG_S	$t3,4*$SZREG($sp)

	$REG_S	$t2,3*$SZREG($sp)

	$REG_S	$t1,2*$SZREG($sp)

	$REG_S	$t0,1*$SZREG($sp)

	$REG_S	$gp,0*$SZREG($sp)

___

$code.=<<___;

	.set	reorder

	li	$minus4,-4

	and	$ta0,$a2,$minus4

	beqz	$ta0,.L_bn_sqr_words_tail

.L_bn_sqr_words_loop:

	$LD	$t0,0($a1)

	$MULTU	($t0,$t0)

	$LD	$t2,$BNSZ($a1)

	$LD	$ta0,2*$BNSZ($a1)

	$LD	$ta2,3*$BNSZ($a1)

	mflo	($t1,$t0,$t0)

	mfhi	($t0,$t0,$t0)

	$ST	$t1,0($a0)

	$ST	$t0,$BNSZ($a0)

	$MULTU	($t2,$t2)

	subu	$a2,4

	$PTR_ADD $a0,8*$BNSZ

	$PTR_ADD $a1,4*$BNSZ

	mflo	($t3,$t2,$t2)

	mfhi	($t2,$t2,$t2)

	$ST	$t3,-6*$BNSZ($a0)

	$ST	$t2,-5*$BNSZ($a0)

	$MULTU	($ta0,$ta0)

	mflo	($ta1,$ta0,$ta0)

	mfhi	($ta0,$ta0,$ta0)

	$ST	$ta1,-4*$BNSZ($a0)

	$ST	$ta0,-3*$BNSZ($a0)

	$MULTU	($ta2,$ta2)

	and	$ta0,$a2,$minus4

	mflo	($ta3,$ta2,$ta2)

	mfhi	($ta2,$ta2,$ta2)

	$ST	$ta3,-2*$BNSZ($a0)

	.set	noreorder

	bgtz	$ta0,.L_bn_sqr_words_loop

	$ST	$ta2,-$BNSZ($a0)

	beqz	$a2,.L_bn_sqr_words_return

	nop

.L_bn_sqr_words_tail:

	.set	reorder

	$LD	$t0,0($a1)

	$MULTU	($t0,$t0)

	subu	$a2,1

	mflo	($t1,$t0,$t0)

	mfhi	($t0,$t0,$t0)

	$ST	$t1,0($a0)

	$ST	$t0,$BNSZ($a0)

	beqz	$a2,.L_bn_sqr_words_return

	$LD	$t0,$BNSZ($a1)

	$MULTU	($t0,$t0)

	subu	$a2,1

	mflo	($t1,$t0,$t0)

	mfhi	($t0,$t0,$t0)

	$ST	$t1,2*$BNSZ($a0)

	$ST	$t0,3*$BNSZ($a0)

	beqz	$a2,.L_bn_sqr_words_return

	$LD	$t0,2*$BNSZ($a1)

	$MULTU	($t0,$t0)

	mflo	($t1,$t0,$t0)

	mfhi	($t0,$t0,$t0)

	$ST	$t1,4*$BNSZ($a0)

	$ST	$t0,5*$BNSZ($a0)

.L_bn_sqr_words_return:

	.set	noreorder

___

$code.=<<___ if ($flavour =~ /nubi/i);

	$REG_L	$t3,4*$SZREG($sp)

	$REG_L	$t2,3*$SZREG($sp)

	$REG_L	$t1,2*$SZREG($sp)

	$REG_L	$t0,1*$SZREG($sp)

	$REG_L	$gp,0*$SZREG($sp)

	$PTR_ADD $sp,6*$SZREG

___

$code.=<<___;

	jr	$ra

	move	$a0,$v0

.end	bn_sqr_words_internal

.align	5

.globl	bn_add_words

.ent	bn_add_words

bn_add_words:

	.set	noreorder

	bgtz	$a3,bn_add_words_internal

	move	$v0,$zero

	jr	$ra

	move	$a0,$v0

.end	bn_add_words

.align	5

.ent	bn_add_words_internal

bn_add_words_internal:

___

$code.=<<___ if ($flavour =~ /nubi/i);

	.frame	$sp,6*$SZREG,$ra

	.mask	0x8000f008,-$SZREG

	.set	noreorder

	$PTR_SUB $sp,6*$SZREG

	$REG_S	$ra,5*$SZREG($sp)

	$REG_S	$t3,4*$SZREG($sp)

	$REG_S	$t2,3*$SZREG($sp)

	$REG_S	$t1,2*$SZREG($sp)

	$REG_S	$t0,1*$SZREG($sp)

	$REG_S	$gp,0*$SZREG($sp)

___

$code.=<<___;

	.set	reorder

	li	$minus4,-4

	and	$at,$a3,$minus4

	beqz	$at,.L_bn_add_words_tail

.L_bn_add_words_loop:

	$LD	$t0,0($a1)

	$LD	$ta0,0($a2)

	subu	$a3,4

	$LD	$t1,$BNSZ($a1)

	and	$at,$a3,$minus4

	$LD	$t2,2*$BNSZ($a1)

	$PTR_ADD $a2,4*$BNSZ

	$LD	$t3,3*$BNSZ($a1)

	$PTR_ADD $a0,4*$BNSZ

	$LD	$ta1,-3*$BNSZ($a2)

	$PTR_ADD $a1,4*$BNSZ

	$LD	$ta2,-2*$BNSZ($a2)

	$LD	$ta3,-$BNSZ($a2)

	$ADDU	$ta0,$t0

	sltu	$t8,$ta0,$t0

	$ADDU	$t0,$ta0,$v0

	sltu	$v0,$t0,$ta0

	$ST	$t0,-4*$BNSZ($a0)

	$ADDU	$v0,$t8

	$ADDU	$ta1,$t1

	sltu	$t9,$ta1,$t1

	$ADDU	$t1,$ta1,$v0

	sltu	$v0,$t1,$ta1

	$ST	$t1,-3*$BNSZ($a0)

	$ADDU	$v0,$t9

	$ADDU	$ta2,$t2

	sltu	$t8,$ta2,$t2

	$ADDU	$t2,$ta2,$v0

	sltu	$v0,$t2,$ta2

	$ST	$t2,-2*$BNSZ($a0)

	$ADDU	$v0,$t8

	$ADDU	$ta3,$t3

	sltu	$t9,$ta3,$t3

	$ADDU	$t3,$ta3,$v0

	sltu	$v0,$t3,$ta3

	$ST	$t3,-$BNSZ($a0)

	.set	noreorder

	bgtz	$at,.L_bn_add_words_loop

	$ADDU	$v0,$t9

	beqz	$a3,.L_bn_add_words_return

	nop

.L_bn_add_words_tail:

	.set	reorder

	$LD	$t0,0($a1)

	$LD	$ta0,0($a2)

	$ADDU	$ta0,$t0

	subu	$a3,1

	sltu	$t8,$ta0,$t0

	$ADDU	$t0,$ta0,$v0

	sltu	$v0,$t0,$ta0

	$ST	$t0,0($a0)

	$ADDU	$v0,$t8

	beqz	$a3,.L_bn_add_words_return

	$LD	$t1,$BNSZ($a1)

	$LD	$ta1,$BNSZ($a2)

	$ADDU	$ta1,$t1

	subu	$a3,1

	sltu	$t9,$ta1,$t1

	$ADDU	$t1,$ta1,$v0

	sltu	$v0,$t1,$ta1

	$ST	$t1,$BNSZ($a0)

	$ADDU	$v0,$t9

	beqz	$a3,.L_bn_add_words_return

	$LD	$t2,2*$BNSZ($a1)

	$LD	$ta2,2*$BNSZ($a2)

	$ADDU	$ta2,$t2

	sltu	$t8,$ta2,$t2

	$ADDU	$t2,$ta2,$v0

	sltu	$v0,$t2,$ta2

	$ST	$t2,2*$BNSZ($a0)

	$ADDU	$v0,$t8

.L_bn_add_words_return:

	.set	noreorder

___

$code.=<<___ if ($flavour =~ /nubi/i);

	$REG_L	$t3,4*$SZREG($sp)

	$REG_L	$t2,3*$SZREG($sp)

	$REG_L	$t1,2*$SZREG($sp)

	$REG_L	$t0,1*$SZREG($sp)

	$REG_L	$gp,0*$SZREG($sp)

	$PTR_ADD $sp,6*$SZREG

___

$code.=<<___;

	jr	$ra

	move	$a0,$v0

.end	bn_add_words_internal

.align	5

.globl	bn_sub_words

.ent	bn_sub_words

bn_sub_words:

	.set	noreorder

	bgtz	$a3,bn_sub_words_internal

	move	$v0,$zero

	jr	$ra

	move	$a0,$zero

.end	bn_sub_words

.align	5

.ent	bn_sub_words_internal

bn_sub_words_internal:

___

$code.=<<___ if ($flavour =~ /nubi/i);

	.frame	$sp,6*$SZREG,$ra

	.mask	0x8000f008,-$SZREG

	.set	noreorder

	$PTR_SUB $sp,6*$SZREG

	$REG_S	$ra,5*$SZREG($sp)

	$REG_S	$t3,4*$SZREG($sp)

	$REG_S	$t2,3*$SZREG($sp)

	$REG_S	$t1,2*$SZREG($sp)

	$REG_S	$t0,1*$SZREG($sp)

	$REG_S	$gp,0*$SZREG($sp)

___

$code.=<<___;

	.set	reorder

	li	$minus4,-4

	and	$at,$a3,$minus4

	beqz	$at,.L_bn_sub_words_tail

.L_bn_sub_words_loop:

	$LD	$t0,0($a1)

	$LD	$ta0,0($a2)

	subu	$a3,4

	$LD	$t1,$BNSZ($a1)

	and	$at,$a3,$minus4

	$LD	$t2,2*$BNSZ($a1)

	$PTR_ADD $a2,4*$BNSZ

	$LD	$t3,3*$BNSZ($a1)

	$PTR_ADD $a0,4*$BNSZ

	$LD	$ta1,-3*$BNSZ($a2)

	$PTR_ADD $a1,4*$BNSZ

	$LD	$ta2,-2*$BNSZ($a2)

	$LD	$ta3,-$BNSZ($a2)

	sltu	$t8,$t0,$ta0

	$SUBU	$ta0,$t0,$ta0

	$SUBU	$t0,$ta0,$v0

	sgtu	$v0,$t0,$ta0

	$ST	$t0,-4*$BNSZ($a0)

	$ADDU	$v0,$t8

	sltu	$t9,$t1,$ta1

	$SUBU	$ta1,$t1,$ta1

	$SUBU	$t1,$ta1,$v0

	sgtu	$v0,$t1,$ta1

	$ST	$t1,-3*$BNSZ($a0)

	$ADDU	$v0,$t9

	sltu	$t8,$t2,$ta2

	$SUBU	$ta2,$t2,$ta2

	$SUBU	$t2,$ta2,$v0

	sgtu	$v0,$t2,$ta2

	$ST	$t2,-2*$BNSZ($a0)

	$ADDU	$v0,$t8

	sltu	$t9,$t3,$ta3

	$SUBU	$ta3,$t3,$ta3

	$SUBU	$t3,$ta3,$v0

	sgtu	$v0,$t3,$ta3

	$ST	$t3,-$BNSZ($a0)

	.set	noreorder

	bgtz	$at,.L_bn_sub_words_loop

	$ADDU	$v0,$t9

	beqz	$a3,.L_bn_sub_words_return

	nop

.L_bn_sub_words_tail:

	.set	reorder

	$LD	$t0,0($a1)

	$LD	$ta0,0($a2)

	subu	$a3,1

	sltu	$t8,$t0,$ta0

	$SUBU	$ta0,$t0,$ta0

	$SUBU	$t0,$ta0,$v0

	sgtu	$v0,$t0,$ta0

	$ST	$t0,0($a0)

	$ADDU	$v0,$t8

	beqz	$a3,.L_bn_sub_words_return

	$LD	$t1,$BNSZ($a1)

	subu	$a3,1

	$LD	$ta1,$BNSZ($a2)

	sltu	$t9,$t1,$ta1

	$SUBU	$ta1,$t1,$ta1

	$SUBU	$t1,$ta1,$v0

	sgtu	$v0,$t1,$ta1

	$ST	$t1,$BNSZ($a0)

	$ADDU	$v0,$t9

	beqz	$a3,.L_bn_sub_words_return

	$LD	$t2,2*$BNSZ($a1)

	$LD	$ta2,2*$BNSZ($a2)

	sltu	$t8,$t2,$ta2

	$SUBU	$ta2,$t2,$ta2

	$SUBU	$t2,$ta2,$v0

	sgtu	$v0,$t2,$ta2

	$ST	$t2,2*$BNSZ($a0)

	$ADDU	$v0,$t8

.L_bn_sub_words_return:

	.set	noreorder

___

$code.=<<___ if ($flavour =~ /nubi/i);

	$REG_L	$t3,4*$SZREG($sp)

	$REG_L	$t2,3*$SZREG($sp)

	$REG_L	$t1,2*$SZREG($sp)

	$REG_L	$t0,1*$SZREG($sp)

	$REG_L	$gp,0*$SZREG($sp)

	$PTR_ADD $sp,6*$SZREG

___

$code.=<<___;

	jr	$ra

	move	$a0,$v0

.end	bn_sub_words_internal

#if 0

/*

 * The bn_div_3_words entry point is re-used for constant-time interface.

 * Implementation is retained as historical reference.

 */

.align 5

.globl	bn_div_3_words

.ent	bn_div_3_words

bn_div_3_words:

	.set	noreorder

	move	$a3,$a0		# we know that bn_div_words does not

				# touch $a3, $ta2, $ta3 and preserves $a2

				# so that we can save two arguments

				# and return address in registers

				# instead of stack:-)

	$LD	$a0,($a3)

	move	$ta2,$a1

	bne	$a0,$a2,bn_div_3_words_internal

	$LD	$a1,-$BNSZ($a3)

	li	$v0,-1

	jr	$ra

	move	$a0,$v0

.end	bn_div_3_words

.align	5

.ent	bn_div_3_words_internal

bn_div_3_words_internal:

___

$code.=<<___ if ($flavour =~ /nubi/i);

	.frame	$sp,6*$SZREG,$ra

	.mask	0x8000f008,-$SZREG

	.set	noreorder

	$PTR_SUB $sp,6*$SZREG

	$REG_S	$ra,5*$SZREG($sp)

	$REG_S	$t3,4*$SZREG($sp)

	$REG_S	$t2,3*$SZREG($sp)

	$REG_S	$t1,2*$SZREG($sp)

	$REG_S	$t0,1*$SZREG($sp)

	$REG_S	$gp,0*$SZREG($sp)

___

$code.=<<___;

	.set	reorder

	move	$ta3,$ra

	bal	bn_div_words_internal

	move	$ra,$ta3

	$MULTU	($ta2,$v0)

	$LD	$t2,-2*$BNSZ($a3)

	move	$ta0,$zero

	mfhi	($t1,$ta2,$v0)

	mflo	($t0,$ta2,$v0)

	sltu	$t8,$t1,$a1

.L_bn_div_3_words_inner_loop:

	bnez	$t8,.L_bn_div_3_words_inner_loop_done

	sgeu	$at,$t2,$t0

	seq	$t9,$t1,$a1

	and	$at,$t9

	sltu	$t3,$t0,$ta2

	$ADDU	$a1,$a2

	$SUBU	$t1,$t3

	$SUBU	$t0,$ta2

	sltu	$t8,$t1,$a1

	sltu	$ta0,$a1,$a2

	or	$t8,$ta0

	.set	noreorder

	beqz	$at,.L_bn_div_3_words_inner_loop

	$SUBU	$v0,1

	$ADDU	$v0,1

	.set	reorder

.L_bn_div_3_words_inner_loop_done:

	.set	noreorder

___

$code.=<<___ if ($flavour =~ /nubi/i);

	$REG_L	$t3,4*$SZREG($sp)

	$REG_L	$t2,3*$SZREG($sp)

	$REG_L	$t1,2*$SZREG($sp)

	$REG_L	$t0,1*$SZREG($sp)

	$REG_L	$gp,0*$SZREG($sp)

	$PTR_ADD $sp,6*$SZREG