3
# ====================================================================
4
# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5
# project. The module is, however, dual licensed under OpenSSL and
6
# CRYPTOGAMS licenses depending on where you obtain it. For further
7
# details see http://www.openssl.org/~appro/cryptogams/.
8
# ====================================================================
10
# I let hardware handle unaligned input(*), except on page boundaries
11
# (see below for details). Otherwise straightforward implementation
12
# with X vector in register bank. The module is big-endian [which is
13
# not big deal as there're no little-endian targets left around].
15
# (*) this means that this module is inappropriate for PPC403? Does
16
# anybody know if pre-POWER3 can sustain unaligned load?
19
# ----------------------------------
20
# PPC970,gcc-4.0.0 +76% +59%
21
# Power6,xlc-7 +68% +33%
25
if ($flavour =~ /64/) {
31
} elsif ($flavour =~ /32/) {
37
} else { die "nonsense $flavour"; }
39
$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
40
( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
41
( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or
42
die "can't locate ppc-xlate.pl";
44
open STDOUT,"| $^X $xlate $flavour ".shift || die "can't call $xlate: $!";
64
@V=($A,$B,$C,$D,$E,$T);
65
@X=("r16","r17","r18","r19","r20","r21","r22","r23",
66
"r24","r25","r26","r27","r28","r29","r30","r31");
69
my ($i,$a,$b,$c,$d,$e,$f)=@_;
71
$code.=<<___ if ($i==0);
72
lwz @X[$i],`$i*4`($inp)
74
$code.=<<___ if ($i<15);
75
lwz @X[$j],`$j*4`($inp)
86
$code.=<<___ if ($i>=15);
89
xor @X[$j%16],@X[$j%16],@X[($j+2)%16]
92
xor @X[$j%16],@X[$j%16],@X[($j+8)%16]
97
xor @X[$j%16],@X[$j%16],@X[($j+13)%16]
99
rotlwi @X[$j%16],@X[$j%16],1
104
my ($i,$a,$b,$c,$d,$e,$f)=@_;
106
$code.=<<___ if ($i<79);
109
xor @X[$j%16],@X[$j%16],@X[($j+2)%16]
112
xor @X[$j%16],@X[$j%16],@X[($j+8)%16]
116
xor @X[$j%16],@X[$j%16],@X[($j+13)%16]
118
rotlwi @X[$j%16],@X[$j%16],1
120
$code.=<<___ if ($i==79);
138
my ($i,$a,$b,$c,$d,$e,$f)=@_;
143
xor @X[$j%16],@X[$j%16],@X[($j+2)%16]
146
xor @X[$j%16],@X[$j%16],@X[($j+8)%16]
150
xor @X[$j%16],@X[$j%16],@X[($j+13)%16]
153
rotlwi @X[$j%16],@X[$j%16],1
162
.globl .sha1_block_data_order
164
.sha1_block_data_order:
166
$STU $sp,`-($FRAME+64)`($sp)
167
$PUSH r0,`$FRAME-$SIZE_T*18`($sp)
168
$PUSH r15,`$FRAME-$SIZE_T*17`($sp)
169
$PUSH r16,`$FRAME-$SIZE_T*16`($sp)
170
$PUSH r17,`$FRAME-$SIZE_T*15`($sp)
171
$PUSH r18,`$FRAME-$SIZE_T*14`($sp)
172
$PUSH r19,`$FRAME-$SIZE_T*13`($sp)
173
$PUSH r20,`$FRAME-$SIZE_T*12`($sp)
174
$PUSH r21,`$FRAME-$SIZE_T*11`($sp)
175
$PUSH r22,`$FRAME-$SIZE_T*10`($sp)
176
$PUSH r23,`$FRAME-$SIZE_T*9`($sp)
177
$PUSH r24,`$FRAME-$SIZE_T*8`($sp)
178
$PUSH r25,`$FRAME-$SIZE_T*7`($sp)
179
$PUSH r26,`$FRAME-$SIZE_T*6`($sp)
180
$PUSH r27,`$FRAME-$SIZE_T*5`($sp)
181
$PUSH r28,`$FRAME-$SIZE_T*4`($sp)
182
$PUSH r29,`$FRAME-$SIZE_T*3`($sp)
183
$PUSH r30,`$FRAME-$SIZE_T*2`($sp)
184
$PUSH r31,`$FRAME-$SIZE_T*1`($sp)
194
bl Lsha1_block_private
196
$POP r0,`$FRAME-$SIZE_T*18`($sp)
197
$POP r15,`$FRAME-$SIZE_T*17`($sp)
198
$POP r16,`$FRAME-$SIZE_T*16`($sp)
199
$POP r17,`$FRAME-$SIZE_T*15`($sp)
200
$POP r18,`$FRAME-$SIZE_T*14`($sp)
201
$POP r19,`$FRAME-$SIZE_T*13`($sp)
202
$POP r20,`$FRAME-$SIZE_T*12`($sp)
203
$POP r21,`$FRAME-$SIZE_T*11`($sp)
204
$POP r22,`$FRAME-$SIZE_T*10`($sp)
205
$POP r23,`$FRAME-$SIZE_T*9`($sp)
206
$POP r24,`$FRAME-$SIZE_T*8`($sp)
207
$POP r25,`$FRAME-$SIZE_T*7`($sp)
208
$POP r26,`$FRAME-$SIZE_T*6`($sp)
209
$POP r27,`$FRAME-$SIZE_T*5`($sp)
210
$POP r28,`$FRAME-$SIZE_T*4`($sp)
211
$POP r29,`$FRAME-$SIZE_T*3`($sp)
212
$POP r30,`$FRAME-$SIZE_T*2`($sp)
213
$POP r31,`$FRAME-$SIZE_T*1`($sp)
215
addi $sp,$sp,`$FRAME+64`
219
# PowerPC specification allows an implementation to be ill-behaved
220
# upon unaligned access which crosses page boundary. "Better safe
221
# than sorry" principle makes me treat it specially. But I don't
222
# look for particular offending word, but rather for 64-byte input
223
# block which crosses the boundary. Once found that block is aligned
224
# and hashed separately...
229
andi. $t1,$t1,4095 ; distance to closest page boundary
230
srwi. $t1,$t1,6 ; t1/=64
233
ble- Laligned ; didn't cross the page boundary
236
bl Lsha1_block_private
240
addi r20,$sp,$FRAME ; spot below the frame
254
$PUSH $inp,`$FRAME-$SIZE_T*19`($sp)
258
bl Lsha1_block_private
259
$POP $inp,`$FRAME-$SIZE_T*19`($sp)
265
# This is private block function, which uses tailored calling
266
# interface, namely upon entry SHA_CTX is pre-loaded to given
267
# registers and counter register contains amount of chunks to
273
$code.=<<___; # load K_00_19
277
for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
278
$code.=<<___; # load K_20_39
282
for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
283
$code.=<<___; # load K_40_59
287
for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
288
$code.=<<___; # load K_60_79
292
for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
309
addi $inp,$inp,`16*4`
310
bdnz- Lsha1_block_private
314
.asciz "SHA1 block transform for PPC, CRYPTOGAMS by <appro\@fy.chalmers.se>"
317
$code =~ s/\`([^\`]*)\`/eval $1/gem;