3 # ====================================================================
4 # Written by Andy Polyakov <appro@openssl.org> 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 # SPARCv9 VIS3 Montgomery multiplicaion procedure suitable for T3 and
13 # onward. There are three new instructions used here: umulxhi,
14 # addxc[cc] and initializing store. On T3 RSA private key operations
15 # are 1.54/1.87/2.11/2.26 times faster for 512/1024/2048/4096-bit key
16 # lengths. This is without dedicated squaring procedure. On T4
17 # corresponding coefficients are 1.47/2.10/2.80/2.90x, which is mostly
18 # for reference purposes, because T4 has dedicated Montgomery
19 # multiplication and squaring *instructions* that deliver even more.
22 open STDOUT,">$output";
24 $frame = "STACK_FRAME";
28 #include "sparc_arch.h"
31 .register %g2,#scratch
32 .register %g3,#scratch
35 .section ".text",#alloc,#execinstr
38 ($n0,$m0,$m1,$lo0,$hi0, $lo1,$hi1,$aj,$alo,$nj,$nlo,$tj)=
39 (map("%g$_",(1..5)),map("%o$_",(0..5,7)));
42 $rp="%o0"; # BN_ULONG *rp,
43 $ap="%o1"; # const BN_ULONG *ap,
44 $bp="%o2"; # const BN_ULONG *bp,
45 $np="%o3"; # const BN_ULONG *np,
46 $n0p="%o4"; # const BN_ULONG *n0,
47 $num="%o5"; # int num); # caller ensures that num is even
50 .globl bn_mul_mont_vis3
53 add %sp, $bias, %g4 ! real top of stack
54 sll $num, 2, $num ! size in bytes
56 andn %g5, 63, %g5 ! buffer size rounded up to 64 bytes
58 add %g5, %g1, %g1 ! 3*buffer size
60 andn %g1, 63, %g1 ! align at 64 byte
61 sub %g1, $frame, %g1 ! new top of stack
67 # +-------------------------------+<----- %sp
69 # +-------------------------------+<----- aligned at 64 bytes
71 # +-------------------------------+
74 # +-------------------------------+<----- aligned at 64 bytes
75 # | __int64 ap[1..0] | converted ap[]
76 # +-------------------------------+
77 # | __int64 np[1..0] | converted np[]
78 # +-------------------------------+
79 # | __int64 ap[3..2] |
82 # +-------------------------------+
83 ($rp,$ap,$bp,$np,$n0p,$num)=map("%i$_",(0..5));
84 ($t0,$t1,$t2,$t3,$cnt,$tp,$bufsz,$anp)=map("%l$_",(0..7));
87 ld [$n0p+0], $t0 ! pull n0[0..1] value
88 add %sp, $bias+$frame, $tp
91 ld [$bp+0], $t2 ! m0=bp[0]
97 ld [$ap+0], $t0 ! ap[0]
102 ld [$ap+8], $t2 ! ap[1]
107 stx $aj, [$anp] ! converted ap[0]
109 mulx $aj, $m0, $lo0 ! ap[0]*bp[0]
110 umulxhi $aj, $m0, $hi0
112 ld [$np+0], $t0 ! np[0]
117 ld [$np+8], $t2 ! np[1]
122 stx $nj, [$anp+8] ! converted np[0]
124 mulx $lo0, $n0, $m1 ! "tp[0]"*n0
125 stx $aj, [$anp+16] ! converted ap[1]
127 mulx $aj, $m0, $alo ! ap[1]*bp[0]
128 umulxhi $aj, $m0, $aj ! ahi=aj
130 mulx $nj, $m1, $lo1 ! np[0]*m1
131 umulxhi $nj, $m1, $hi1
135 stx $nj, [$anp+24] ! converted np[1]
138 addcc $lo0, $lo1, $lo1
139 addxc %g0, $hi1, $hi1
141 mulx $nj, $m1, $nlo ! np[1]*m1
142 umulxhi $nj, $m1, $nj ! nhi=nj
145 sub $num, 24, $cnt ! cnt=num-3
149 ld [$ap+0], $t0 ! ap[j]
150 addcc $alo, $hi0, $lo0
157 stx $aj, [$anp] ! converted ap[j]
159 ld [$np+0], $t2 ! np[j]
160 addcc $nlo, $hi1, $lo1
162 addxc $nj, %g0, $hi1 ! nhi=nj
166 mulx $aj, $m0, $alo ! ap[j]*bp[0]
168 umulxhi $aj, $m0, $aj ! ahi=aj
169 stx $nj, [$anp+8] ! converted np[j]
170 add $anp, 16, $anp ! anp++
172 mulx $nj, $m1, $nlo ! np[j]*m1
173 addcc $lo0, $lo1, $lo1 ! np[j]*m1+ap[j]*bp[0]
174 umulxhi $nj, $m1, $nj ! nhi=nj
175 addxc %g0, $hi1, $hi1
176 stx $lo1, [$tp] ! tp[j-1]
177 add $tp, 8, $tp ! tp++
180 sub $cnt, 8, $cnt ! j--
182 addcc $alo, $hi0, $lo0
183 addxc $aj, %g0, $hi0 ! ahi=aj
185 addcc $nlo, $hi1, $lo1
187 addcc $lo0, $lo1, $lo1 ! np[j]*m1+ap[j]*bp[0]
188 addxc %g0, $hi1, $hi1
189 stx $lo1, [$tp] ! tp[j-1]
192 addcc $hi0, $hi1, $hi1
193 addxc %g0, %g0, $ovf ! upmost overflow bit
198 sub $num, 16, $i ! i=num-2
202 ld [$bp+0], $t2 ! m0=bp[i]
205 sub $anp, $num, $anp ! rewind
211 ldx [$anp+0], $aj ! ap[0]
213 ldx [$anp+8], $nj ! np[0]
215 mulx $aj, $m0, $lo0 ! ap[0]*bp[i]
216 ldx [$tp], $tj ! tp[0]
217 umulxhi $aj, $m0, $hi0
218 ldx [$anp+16], $aj ! ap[1]
219 addcc $lo0, $tj, $lo0 ! ap[0]*bp[i]+tp[0]
220 mulx $aj, $m0, $alo ! ap[1]*bp[i]
221 addxc %g0, $hi0, $hi0
222 mulx $lo0, $n0, $m1 ! tp[0]*n0
223 umulxhi $aj, $m0, $aj ! ahi=aj
224 mulx $nj, $m1, $lo1 ! np[0]*m1
225 umulxhi $nj, $m1, $hi1
226 ldx [$anp+24], $nj ! np[1]
228 addcc $lo1, $lo0, $lo1
229 mulx $nj, $m1, $nlo ! np[1]*m1
230 addxc %g0, $hi1, $hi1
231 umulxhi $nj, $m1, $nj ! nhi=nj
234 sub $num, 24, $cnt ! cnt=num-3
237 addcc $alo, $hi0, $lo0
238 ldx [$tp+8], $tj ! tp[j]
239 addxc $aj, %g0, $hi0 ! ahi=aj
240 ldx [$anp+0], $aj ! ap[j]
241 addcc $nlo, $hi1, $lo1
242 mulx $aj, $m0, $alo ! ap[j]*bp[i]
243 addxc $nj, %g0, $hi1 ! nhi=nj
244 ldx [$anp+8], $nj ! np[j]
246 umulxhi $aj, $m0, $aj ! ahi=aj
247 addcc $lo0, $tj, $lo0 ! ap[j]*bp[i]+tp[j]
248 mulx $nj, $m1, $nlo ! np[j]*m1
249 addxc %g0, $hi0, $hi0
250 umulxhi $nj, $m1, $nj ! nhi=nj
251 addcc $lo1, $lo0, $lo1 ! np[j]*m1+ap[j]*bp[i]+tp[j]
252 addxc %g0, $hi1, $hi1
253 stx $lo1, [$tp] ! tp[j-1]
255 brnz,pt $cnt, .Linner
258 ldx [$tp+8], $tj ! tp[j]
259 addcc $alo, $hi0, $lo0
260 addxc $aj, %g0, $hi0 ! ahi=aj
261 addcc $lo0, $tj, $lo0 ! ap[j]*bp[i]+tp[j]
262 addxc %g0, $hi0, $hi0
264 addcc $nlo, $hi1, $lo1
265 addxc $nj, %g0, $hi1 ! nhi=nj
266 addcc $lo1, $lo0, $lo1 ! np[j]*m1+ap[j]*bp[i]+tp[j]
267 addxc %g0, $hi1, $hi1
268 stx $lo1, [$tp] ! tp[j-1]
270 subcc %g0, $ovf, %g0 ! move upmost overflow to CCR.xcc
271 addxccc $hi1, $hi0, $hi1
279 sub $anp, $num, $anp ! rewind
283 subcc $num, 8, $cnt ! cnt=num-1 and clear CCR.xcc
291 subccc $tj, $nj, $t2 ! tp[j]-np[j]
296 st $t2, [$rp-4] ! reverse order
301 sub $anp, $num, $anp ! rewind
306 subc $ovf, %g0, $ovf ! handle upmost overflow bit
309 or $np, $ap, $ap ! ap=borrow?tp:rp
314 .Lcopy: ! copy or in-place refresh
320 stx %g0, [$anp] ! zap
323 st $t3, [$rp+0] ! flip order
332 .type bn_mul_mont_vis3, #function
333 .size bn_mul_mont_vis3, .-bn_mul_mont_vis3
334 .asciz "Montgomery Multiplication for SPARCv9 VIS3, CRYPTOGAMS by <appro\@openssl.org>"
338 # Purpose of these subroutines is to explicitly encode VIS instructions,
339 # so that one can compile the module without having to specify VIS
340 # extensions on compiler command line, e.g. -xarch=v9 vs. -xarch=v9a.
341 # Idea is to reserve for option to produce "universal" binary and let
342 # programmer detect if current CPU is VIS capable at run-time.
344 my ($mnemonic,$rs1,$rs2,$rd)=@_;
345 my %bias = ( "g" => 0, "o" => 8, "l" => 16, "i" => 24 );
347 my %visopf = ( "addxc" => 0x011,
349 "umulxhi" => 0x016 );
351 $ref = "$mnemonic\t$rs1,$rs2,$rd";
353 if ($opf=$visopf{$mnemonic}) {
354 foreach ($rs1,$rs2,$rd) {
355 return $ref if (!/%([goli])([0-9])/);
359 return sprintf ".word\t0x%08x !%s",
360 0x81b00000|$rd<<25|$rs1<<14|$opf<<5|$rs2,
367 foreach (split("\n",$code)) {
368 s/\`([^\`]*)\`/eval $1/ge;
370 s/\b(umulxhi|addxc[c]{0,2})\s+(%[goli][0-7]),\s*(%[goli][0-7]),\s*(%[goli][0-7])/