1#! /usr/bin/env perl
2# Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the Apache License 2.0 (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16
17# April 2007.
18#
19# Performance improvement over vanilla C code varies from 85% to 45%
20# depending on key length and benchmark. Unfortunately in this context
21# these are not very impressive results [for code that utilizes "wide"
22# 64x64=128-bit multiplication, which is not commonly available to C
23# programmers], at least hand-coded bn_asm.c replacement is known to
24# provide 30-40% better results for longest keys. Well, on a second
25# thought it's not very surprising, because z-CPUs are single-issue
26# and _strictly_ in-order execution, while bn_mul_mont is more or less
27# dependent on CPU ability to pipe-line instructions and have several
28# of them "in-flight" at the same time. I mean while other methods,
29# for example Karatsuba, aim to minimize amount of multiplications at
30# the cost of other operations increase, bn_mul_mont aim to neatly
31# "overlap" multiplications and the other operations [and on most
32# platforms even minimize the amount of the other operations, in
33# particular references to memory]. But it's possible to improve this
34# module performance by implementing dedicated squaring code-path and
35# possibly by unrolling loops...
36
37# January 2009.
38#
39# Reschedule to minimize/avoid Address Generation Interlock hazard,
40# make inner loops counter-based.
41
42# November 2010.
43#
44# Adapt for -m31 build. If kernel supports what's called "highgprs"
45# feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit
46# instructions and achieve "64-bit" performance even in 31-bit legacy
47# application context. The feature is not specific to any particular
48# processor, as long as it's "z-CPU". Latter implies that the code
49# remains z/Architecture specific. Compatibility with 32-bit BN_ULONG
50# is achieved by swapping words after 64-bit loads, follow _dswap-s.
51# On z990 it was measured to perform 2.6-2.2 times better than
52# compiler-generated code, less for longer keys...
53
54# $output is the last argument if it looks like a file (it has an extension)
55# $flavour is the first argument if it doesn't look like a file
56$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
57$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
58
59if ($flavour =~ /3[12]/) {
60	$SIZE_T=4;
61	$g="";
62} else {
63	$SIZE_T=8;
64	$g="g";
65}
66
67$output and open STDOUT,">$output";
68
69$stdframe=16*$SIZE_T+4*8;
70
71$mn0="%r0";
72$num="%r1";
73
74# int bn_mul_mont(
75$rp="%r2";		# BN_ULONG *rp,
76$ap="%r3";		# const BN_ULONG *ap,
77$bp="%r4";		# const BN_ULONG *bp,
78$np="%r5";		# const BN_ULONG *np,
79$n0="%r6";		# const BN_ULONG *n0,
80#$num="160(%r15)"	# int num);
81
82$bi="%r2";	# zaps rp
83$j="%r7";
84
85$ahi="%r8";
86$alo="%r9";
87$nhi="%r10";
88$nlo="%r11";
89$AHI="%r12";
90$NHI="%r13";
91$count="%r14";
92$sp="%r15";
93
94$code.=<<___;
95.text
96.globl	bn_mul_mont
97.type	bn_mul_mont,\@function
98bn_mul_mont:
99	lgf	$num,`$stdframe+$SIZE_T-4`($sp)	# pull $num
100	sla	$num,`log($SIZE_T)/log(2)`	# $num to enumerate bytes
101	la	$bp,0($num,$bp)
102
103	st${g}	%r2,2*$SIZE_T($sp)
104
105	cghi	$num,16		#
106	lghi	%r2,0		#
107	blr	%r14		# if($num<16) return 0;
108___
109$code.=<<___ if ($flavour =~ /3[12]/);
110	tmll	$num,4
111	bnzr	%r14		# if ($num&1) return 0;
112___
113$code.=<<___ if ($flavour !~ /3[12]/);
114	cghi	$num,96		#
115	bhr	%r14		# if($num>96) return 0;
116___
117$code.=<<___;
118	stm${g}	%r3,%r15,3*$SIZE_T($sp)
119
120	lghi	$rp,-$stdframe-8	# leave room for carry bit
121	lcgr	$j,$num		# -$num
122	lgr	%r0,$sp
123	la	$rp,0($rp,$sp)
124	la	$sp,0($j,$rp)	# alloca
125	st${g}	%r0,0($sp)	# back chain
126
127	sra	$num,3		# restore $num
128	la	$bp,0($j,$bp)	# restore $bp
129	ahi	$num,-1		# adjust $num for inner loop
130	lg	$n0,0($n0)	# pull n0
131	_dswap	$n0
132
133	lg	$bi,0($bp)
134	_dswap	$bi
135	lg	$alo,0($ap)
136	_dswap	$alo
137	mlgr	$ahi,$bi	# ap[0]*bp[0]
138	lgr	$AHI,$ahi
139
140	lgr	$mn0,$alo	# "tp[0]"*n0
141	msgr	$mn0,$n0
142
143	lg	$nlo,0($np)	#
144	_dswap	$nlo
145	mlgr	$nhi,$mn0	# np[0]*m1
146	algr	$nlo,$alo	# +="tp[0]"
147	lghi	$NHI,0
148	alcgr	$NHI,$nhi
149
150	la	$j,8		# j=1
151	lr	$count,$num
152
153.align	16
154.L1st:
155	lg	$alo,0($j,$ap)
156	_dswap	$alo
157	mlgr	$ahi,$bi	# ap[j]*bp[0]
158	algr	$alo,$AHI
159	lghi	$AHI,0
160	alcgr	$AHI,$ahi
161
162	lg	$nlo,0($j,$np)
163	_dswap	$nlo
164	mlgr	$nhi,$mn0	# np[j]*m1
165	algr	$nlo,$NHI
166	lghi	$NHI,0
167	alcgr	$nhi,$NHI	# +="tp[j]"
168	algr	$nlo,$alo
169	alcgr	$NHI,$nhi
170
171	stg	$nlo,$stdframe-8($j,$sp)	# tp[j-1]=
172	la	$j,8($j)	# j++
173	brct	$count,.L1st
174
175	algr	$NHI,$AHI
176	lghi	$AHI,0
177	alcgr	$AHI,$AHI	# upmost overflow bit
178	stg	$NHI,$stdframe-8($j,$sp)
179	stg	$AHI,$stdframe($j,$sp)
180	la	$bp,8($bp)	# bp++
181
182.Louter:
183	lg	$bi,0($bp)	# bp[i]
184	_dswap	$bi
185	lg	$alo,0($ap)
186	_dswap	$alo
187	mlgr	$ahi,$bi	# ap[0]*bp[i]
188	alg	$alo,$stdframe($sp)	# +=tp[0]
189	lghi	$AHI,0
190	alcgr	$AHI,$ahi
191
192	lgr	$mn0,$alo
193	msgr	$mn0,$n0	# tp[0]*n0
194
195	lg	$nlo,0($np)	# np[0]
196	_dswap	$nlo
197	mlgr	$nhi,$mn0	# np[0]*m1
198	algr	$nlo,$alo	# +="tp[0]"
199	lghi	$NHI,0
200	alcgr	$NHI,$nhi
201
202	la	$j,8		# j=1
203	lr	$count,$num
204
205.align	16
206.Linner:
207	lg	$alo,0($j,$ap)
208	_dswap	$alo
209	mlgr	$ahi,$bi	# ap[j]*bp[i]
210	algr	$alo,$AHI
211	lghi	$AHI,0
212	alcgr	$ahi,$AHI
213	alg	$alo,$stdframe($j,$sp)# +=tp[j]
214	alcgr	$AHI,$ahi
215
216	lg	$nlo,0($j,$np)
217	_dswap	$nlo
218	mlgr	$nhi,$mn0	# np[j]*m1
219	algr	$nlo,$NHI
220	lghi	$NHI,0
221	alcgr	$nhi,$NHI
222	algr	$nlo,$alo	# +="tp[j]"
223	alcgr	$NHI,$nhi
224
225	stg	$nlo,$stdframe-8($j,$sp)	# tp[j-1]=
226	la	$j,8($j)	# j++
227	brct	$count,.Linner
228
229	algr	$NHI,$AHI
230	lghi	$AHI,0
231	alcgr	$AHI,$AHI
232	alg	$NHI,$stdframe($j,$sp)# accumulate previous upmost overflow bit
233	lghi	$ahi,0
234	alcgr	$AHI,$ahi	# new upmost overflow bit
235	stg	$NHI,$stdframe-8($j,$sp)
236	stg	$AHI,$stdframe($j,$sp)
237
238	la	$bp,8($bp)	# bp++
239	cl${g}	$bp,`$stdframe+8+4*$SIZE_T`($j,$sp)	# compare to &bp[num]
240	jne	.Louter
241
242	l${g}	$rp,`$stdframe+8+2*$SIZE_T`($j,$sp)	# reincarnate rp
243	la	$ap,$stdframe($sp)
244	ahi	$num,1		# restore $num, incidentally clears "borrow"
245
246	la	$j,0
247	lr	$count,$num
248.Lsub:	lg	$alo,0($j,$ap)
249	lg	$nlo,0($j,$np)
250	_dswap	$nlo
251	slbgr	$alo,$nlo
252	stg	$alo,0($j,$rp)
253	la	$j,8($j)
254	brct	$count,.Lsub
255	lghi	$ahi,0
256	slbgr	$AHI,$ahi	# handle upmost carry
257	lghi	$NHI,-1
258	xgr	$NHI,$AHI
259
260	la	$j,0
261	lgr	$count,$num
262.Lcopy:	lg	$ahi,$stdframe($j,$sp)	# conditional copy
263	lg	$alo,0($j,$rp)
264	ngr	$ahi,$AHI
265	ngr	$alo,$NHI
266	ogr	$alo,$ahi
267	_dswap	$alo
268	stg	$j,$stdframe($j,$sp)	# zap tp
269	stg	$alo,0($j,$rp)
270	la	$j,8($j)
271	brct	$count,.Lcopy
272
273	la	%r1,`$stdframe+8+6*$SIZE_T`($j,$sp)
274	lm${g}	%r6,%r15,0(%r1)
275	lghi	%r2,1		# signal "processed"
276	br	%r14
277.size	bn_mul_mont,.-bn_mul_mont
278.string	"Montgomery Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>"
279___
280
281foreach (split("\n",$code)) {
282	s/\`([^\`]*)\`/eval $1/ge;
283	s/_dswap\s+(%r[0-9]+)/sprintf("rllg\t%s,%s,32",$1,$1) if($SIZE_T==4)/e;
284	print $_,"\n";
285}
286close STDOUT or die "error closing STDOUT: $!";
287