2 # Copyright 2010-2016 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the OpenSSL license (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
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 # ====================================================================
19 # The module implements "4-bit" GCM GHASH function and underlying
20 # single multiplication operation in GF(2^128). "4-bit" means that it
21 # uses 256 bytes per-key table [+128 bytes shared table]. Performance
22 # was measured to be ~18 cycles per processed byte on z10, which is
23 # almost 40% better than gcc-generated code. It should be noted that
24 # 18 cycles is worse result than expected: loop is scheduled for 12
25 # and the result should be close to 12. In the lack of instruction-
26 # level profiling data it's impossible to tell why...
30 # Adapt for -m31 build. If kernel supports what's called "highgprs"
31 # feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit
32 # instructions and achieve "64-bit" performance even in 31-bit legacy
33 # application context. The feature is not specific to any particular
34 # processor, as long as it's "z-CPU". Latter implies that the code
35 # remains z/Architecture specific. On z990 it was measured to perform
36 # 2.8x better than 32-bit code generated by gcc 4.3.
40 # Support for hardware KIMD-GHASH is verified to produce correct
41 # result and therefore is engaged. On z196 it was measured to process
42 # 8KB buffer ~7 faster than software implementation. It's not as
43 # impressive for smaller buffer sizes and for smallest 16-bytes buffer
44 # it's actually almost 2 times slower. Which is the reason why
45 # KIMD-GHASH is not used in gcm_gmult_4bit.
49 if ($flavour =~ /3[12]/) {
57 while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {}
58 open STDOUT,">$output";
65 $Xi="%r2"; # argument block
70 $rem0="%r6"; # variables
89 $code.=<<___ if(!$softonly && 0); # hardware is slow for single block...
90 larl %r1,OPENSSL_s390xcap_P
92 lg %r1,24(%r1) # load second word of kimd capabilities vector
93 tmhh %r1,0x4000 # check for function 65
95 stg %r0,16($sp) # arrange 16 bytes of zero input
97 lghi %r0,65 # function 65
98 la %r1,0($Xi) # H lies right after Xi in gcm128_context
101 .long 0xb93e0004 # kimd %r0,$inp
102 brc 1,.-4 # pay attention to "partial completion"
108 stm${g} %r6,%r14,6*$SIZE_T($sp)
113 larl $rem_4bit,rem_4bit
115 lg $Zlo,8+1($Xi) # Xi
117 .type gcm_gmult_4bit,\@function
118 .size gcm_gmult_4bit,(.-gcm_gmult_4bit)
120 .globl gcm_ghash_4bit
124 $code.=<<___ if(!$softonly);
125 larl %r1,OPENSSL_s390xcap_P
126 lg %r0,24(%r1) # load second word of kimd capabilities vector
127 tmhh %r0,0x4000 # check for function 65
129 lghi %r0,65 # function 65
130 la %r1,0($Xi) # H lies right after Xi in gcm128_context
131 .long 0xb93e0004 # kimd %r0,$inp
132 brc 1,.-4 # pay attention to "partial completion"
137 $code.=<<___ if ($flavour =~ /3[12]/);
141 stm${g} %r6,%r14,6*$SIZE_T($sp)
146 larl $rem_4bit,rem_4bit
148 lg $Zlo,8+1($Xi) # Xi
152 xg $Zhi,0($inp) # Xi ^= inp
161 srlg $xi,$Zlo,8 # extract second byte
167 lg $Zlo,8($nlo,$Htbl)
168 lg $Zhi,0($nlo,$Htbl)
179 xg $Zlo,8($nhi,$Htbl)
180 xg $Zhi,0($nhi,$Htbl)
191 xg $Zlo,8($nlo,$Htbl)
193 xg $Zhi,0($nlo,$Htbl)
195 xg $Zhi,0($rem0,$rem_4bit)
205 xg $Zlo,8($nhi,$Htbl)
206 xg $Zhi,0($nhi,$Htbl)
208 xg $Zhi,0($rem1,$rem_4bit)
213 brct $cnt,.Lghash_inner
217 xg $Zlo,8($nlo,$Htbl)
218 xg $Zhi,0($nlo,$Htbl)
220 xg $Zhi,0($rem0,$rem_4bit)
227 xg $Zlo,8($nhi,$Htbl)
228 xg $Zhi,0($nhi,$Htbl)
230 xg $Zhi,0($rem1,$rem_4bit)
232 lg $tmp,0($xi,$rem_4bit)
234 sllg $tmp,$tmp,4 # correct last rem_4bit[rem]
240 lm${g} %r6,%r14,6*$SIZE_T($sp)
242 .type gcm_ghash_4bit,\@function
243 .size gcm_ghash_4bit,(.-gcm_ghash_4bit)
247 .long `0x0000<<12`,0,`0x1C20<<12`,0,`0x3840<<12`,0,`0x2460<<12`,0
248 .long `0x7080<<12`,0,`0x6CA0<<12`,0,`0x48C0<<12`,0,`0x54E0<<12`,0
249 .long `0xE100<<12`,0,`0xFD20<<12`,0,`0xD940<<12`,0,`0xC560<<12`,0
250 .long `0x9180<<12`,0,`0x8DA0<<12`,0,`0xA9C0<<12`,0,`0xB5E0<<12`,0
251 .type rem_4bit,\@object
252 .size rem_4bit,(.-rem_4bit)
253 .string "GHASH for s390x, CRYPTOGAMS by <appro\@openssl.org>"
256 $code =~ s/\`([^\`]*)\`/eval $1/gem;