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 # ====================================================================
12 # "Teaser" Montgomery multiplication module for PowerPC. It's possible
13 # to gain a bit more by modulo-scheduling outer loop, then dedicated
14 # squaring procedure should give further 20% and code can be adapted
15 # for 32-bit application running on 64-bit CPU. As for the latter.
16 # It won't be able to achieve "native" 64-bit performance, because in
17 # 32-bit application context every addc instruction will have to be
18 # expanded as addc, twice right shift by 32 and finally adde, etc.
19 # So far RSA *sign* performance improvement over pre-bn_mul_mont asm
20 # for 64-bit application running on PPC970/G5 is:
29 if ($flavour =~ /32/) {
37 $LDU= "lwzu"; # load and update
38 $LDX= "lwzx"; # load indexed
40 $STU= "stwu"; # store and update
41 $STX= "stwx"; # store indexed
42 $STUX= "stwux"; # store indexed and update
43 $UMULL= "mullw"; # unsigned multiply low
44 $UMULH= "mulhwu"; # unsigned multiply high
45 $UCMP= "cmplw"; # unsigned compare
46 $SHRI= "srwi"; # unsigned shift right by immediate
49 } elsif ($flavour =~ /64/) {
56 # same as above, but 64-bit mnemonics...
58 $LDU= "ldu"; # load and update
59 $LDX= "ldx"; # load indexed
61 $STU= "stdu"; # store and update
62 $STX= "stdx"; # store indexed
63 $STUX= "stdux"; # store indexed and update
64 $UMULL= "mulld"; # unsigned multiply low
65 $UMULH= "mulhdu"; # unsigned multiply high
66 $UCMP= "cmpld"; # unsigned compare
67 $SHRI= "srdi"; # unsigned shift right by immediate
70 } else { die "nonsense $flavour"; }
72 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
73 ( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
74 ( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or
75 die "can't locate ppc-xlate.pl";
77 open STDOUT,"| $^X $xlate $flavour ".shift || die "can't call $xlate: $!";
87 $rp="r9"; # $rp is reassigned
91 # non-volatile registers
115 mr $rp,r3 ; $rp is reassigned
119 slwi $num,$num,`log($BNSZ)/log(2)`
121 addi $ovf,$num,`$FRAME+$RZONE`
122 subf $ovf,$ovf,$sp ; $sp-$ovf
123 and $ovf,$ovf,$tj ; minimize TLB usage
124 subf $ovf,$sp,$ovf ; $ovf-$sp
125 srwi $num,$num,`log($BNSZ)/log(2)`
128 $PUSH r14,`4*$SIZE_T`($sp)
129 $PUSH r15,`5*$SIZE_T`($sp)
130 $PUSH r16,`6*$SIZE_T`($sp)
131 $PUSH r17,`7*$SIZE_T`($sp)
132 $PUSH r18,`8*$SIZE_T`($sp)
133 $PUSH r19,`9*$SIZE_T`($sp)
134 $PUSH r20,`10*$SIZE_T`($sp)
135 $PUSH r21,`11*$SIZE_T`($sp)
136 $PUSH r22,`12*$SIZE_T`($sp)
137 $PUSH r23,`13*$SIZE_T`($sp)
138 $PUSH r24,`14*$SIZE_T`($sp)
139 $PUSH r25,`15*$SIZE_T`($sp)
141 $LD $n0,0($n0) ; pull n0[0] value
142 addi $num,$num,-2 ; adjust $num for counter register
144 $LD $m0,0($bp) ; m0=bp[0]
145 $LD $aj,0($ap) ; ap[0]
147 $UMULL $lo0,$aj,$m0 ; ap[0]*bp[0]
150 $LD $aj,$BNSZ($ap) ; ap[1]
151 $LD $nj,0($np) ; np[0]
153 $UMULL $m1,$lo0,$n0 ; "tp[0]"*n0
155 $UMULL $alo,$aj,$m0 ; ap[1]*bp[0]
158 $UMULL $lo1,$nj,$m1 ; np[0]*m1
160 $LD $nj,$BNSZ($np) ; np[1]
164 $UMULL $nlo,$nj,$m1 ; np[1]*m1
171 $LDX $aj,$ap,$j ; ap[j]
173 $LDX $nj,$np,$j ; np[j]
175 $UMULL $alo,$aj,$m0 ; ap[j]*bp[0]
179 $UMULL $nlo,$nj,$m1 ; np[j]*m1
180 addc $lo1,$lo1,$lo0 ; np[j]*m1+ap[j]*bp[0]
183 $ST $lo1,0($tp) ; tp[j-1]
185 addi $j,$j,$BNSZ ; j++
186 addi $tp,$tp,$BNSZ ; tp++
194 addc $lo1,$lo1,$lo0 ; np[j]*m1+ap[j]*bp[0]
196 $ST $lo1,0($tp) ; tp[j-1]
200 addze $ovf,$ovf ; upmost overflow bit
206 $LDX $m0,$bp,$i ; m0=bp[i]
207 $LD $aj,0($ap) ; ap[0]
209 $LD $tj,$FRAME($sp) ; tp[0]
210 $UMULL $lo0,$aj,$m0 ; ap[0]*bp[i]
212 $LD $aj,$BNSZ($ap) ; ap[1]
213 $LD $nj,0($np) ; np[0]
214 addc $lo0,$lo0,$tj ; ap[0]*bp[i]+tp[0]
215 $UMULL $alo,$aj,$m0 ; ap[j]*bp[i]
217 $UMULL $m1,$lo0,$n0 ; tp[0]*n0
219 $UMULL $lo1,$nj,$m1 ; np[0]*m1
221 $LD $nj,$BNSZ($np) ; np[1]
223 $UMULL $nlo,$nj,$m1 ; np[1]*m1
231 $LDX $aj,$ap,$j ; ap[j]
233 $LD $tj,$BNSZ($tp) ; tp[j]
235 $LDX $nj,$np,$j ; np[j]
237 $UMULL $alo,$aj,$m0 ; ap[j]*bp[i]
240 addc $lo0,$lo0,$tj ; ap[j]*bp[i]+tp[j]
241 $UMULL $nlo,$nj,$m1 ; np[j]*m1
244 addc $lo1,$lo1,$lo0 ; np[j]*m1+ap[j]*bp[i]+tp[j]
245 addi $j,$j,$BNSZ ; j++
247 $ST $lo1,0($tp) ; tp[j-1]
248 addi $tp,$tp,$BNSZ ; tp++
251 $LD $tj,$BNSZ($tp) ; tp[j]
254 addc $lo0,$lo0,$tj ; ap[j]*bp[i]+tp[j]
259 addc $lo1,$lo1,$lo0 ; np[j]*m1+ap[j]*bp[i]+tp[j]
261 $ST $lo1,0($tp) ; tp[j-1]
263 addic $ovf,$ovf,-1 ; move upmost overflow to XER[CA]
269 slwi $tj,$num,`log($BNSZ)/log(2)`
274 addi $num,$num,2 ; restore $num
275 subfc $j,$j,$j ; j=0 and "clear" XER[CA]
280 Lsub: $LDX $tj,$tp,$j
282 subfe $aj,$nj,$tj ; tp[j]-np[j]
289 subfe $ovf,$j,$ovf ; handle upmost overflow bit
292 or $ap,$ap,$np ; ap=borrow?tp:rp
295 Lcopy: ; copy or in-place refresh
298 $STX $j,$tp,$j ; zap at once
302 $POP r14,`4*$SIZE_T`($sp)
303 $POP r15,`5*$SIZE_T`($sp)
304 $POP r16,`6*$SIZE_T`($sp)
305 $POP r17,`7*$SIZE_T`($sp)
306 $POP r18,`8*$SIZE_T`($sp)
307 $POP r19,`9*$SIZE_T`($sp)
308 $POP r20,`10*$SIZE_T`($sp)
309 $POP r21,`11*$SIZE_T`($sp)
310 $POP r22,`12*$SIZE_T`($sp)
311 $POP r23,`13*$SIZE_T`($sp)
312 $POP r24,`14*$SIZE_T`($sp)
313 $POP r25,`15*$SIZE_T`($sp)
318 .asciz "Montgomery Multiplication for PPC, CRYPTOGAMS by <appro\@fy.chalmers.se>"
321 $code =~ s/\`([^\`]*)\`/eval $1/gem;