]>
Commit | Line | Data |
---|---|---|
6aa36e8e RS |
1 | #! /usr/bin/env perl |
2 | # Copyright 2005-2016 The OpenSSL Project Authors. All Rights Reserved. | |
3 | # | |
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 | |
8 | ||
bcb43bb3 AP |
9 | |
10 | # ==================================================================== | |
e3713c36 | 11 | # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL |
2e21922e AP |
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/. | |
bcb43bb3 AP |
15 | # ==================================================================== |
16 | ||
aa2be094 AP |
17 | # October 2005 |
18 | # | |
bcb43bb3 AP |
19 | # "Teaser" Montgomery multiplication module for UltraSPARC. Why FPU? |
20 | # Because unlike integer multiplier, which simply stalls whole CPU, | |
21 | # FPU is fully pipelined and can effectively emit 48 bit partial | |
22 | # product every cycle. Why not blended SPARC v9? One can argue that | |
23 | # making this module dependent on UltraSPARC VIS extension limits its | |
a4d729f3 AP |
24 | # binary compatibility. Well yes, it does exclude SPARC64 prior-V(!) |
25 | # implementations from compatibility matrix. But the rest, whole Sun | |
26 | # UltraSPARC family and brand new Fujitsu's SPARC64 V, all support | |
27 | # VIS extension instructions used in this module. This is considered | |
73b979e6 AP |
28 | # good enough to not care about HAL SPARC64 users [if any] who have |
29 | # integer-only pure SPARCv9 module to "fall down" to. | |
bcb43bb3 AP |
30 | |
31 | # USI&II cores currently exhibit uniform 2x improvement [over pre- | |
32 | # bn_mul_mont codebase] for all key lengths and benchmarks. On USIII | |
33 | # performance improves few percents for shorter keys and worsens few | |
aa2be094 | 34 | # percents for longer keys. This is because USIII integer multiplier |
bcb43bb3 AP |
35 | # is >3x faster than USI&II one, which is harder to match [but see |
36 | # TODO list below]. It should also be noted that SPARC64 V features | |
37 | # out-of-order execution, which *might* mean that integer multiplier | |
a4d729f3 AP |
38 | # is pipelined, which in turn *might* be impossible to match... On |
39 | # additional note, SPARC64 V implements FP Multiply-Add instruction, | |
40 | # which is perfectly usable in this context... In other words, as far | |
73b979e6 | 41 | # as Fujitsu SPARC64 V goes, talk to the author:-) |
aa2be094 | 42 | |
a00e414f AP |
43 | # The implementation implies following "non-natural" limitations on |
44 | # input arguments: | |
aa2be094 AP |
45 | # - num may not be less than 4; |
46 | # - num has to be even; | |
aa2be094 AP |
47 | # Failure to meet either condition has no fatal effects, simply |
48 | # doesn't give any performance gain. | |
49 | ||
bcb43bb3 | 50 | # TODO: |
bcb43bb3 AP |
51 | # - modulo-schedule inner loop for better performance (on in-order |
52 | # execution core such as UltraSPARC this shall result in further | |
53 | # noticeable(!) improvement); | |
54 | # - dedicated squaring procedure[?]; | |
55 | ||
2e21922e AP |
56 | ###################################################################### |
57 | # November 2006 | |
58 | # | |
59 | # Modulo-scheduled inner loops allow to interleave floating point and | |
60 | # integer instructions and minimize Read-After-Write penalties. This | |
60250017 | 61 | # results in *further* 20-50% performance improvement [depending on |
2e21922e AP |
62 | # key length, more for longer keys] on USI&II cores and 30-80% - on |
63 | # USIII&IV. | |
64 | ||
6bd7a4d9 RL |
65 | $output = pop; |
66 | open STDOUT,">$output"; | |
67 | ||
a00e414f | 68 | $fname="bn_mul_mont_fpu"; |
eb77e888 AP |
69 | |
70 | $frame="STACK_FRAME"; | |
71 | $bias="STACK_BIAS"; | |
bcb43bb3 AP |
72 | $locals=64; |
73 | ||
74 | # In order to provide for 32-/64-bit ABI duality, I keep integers wider | |
75 | # than 32 bit in %g1-%g4 and %o0-%o5. %l0-%l7 and %i0-%i5 are used | |
76 | # exclusively for pointers, indexes and other small values... | |
77 | # int bn_mul_mont( | |
78 | $rp="%i0"; # BN_ULONG *rp, | |
79 | $ap="%i1"; # const BN_ULONG *ap, | |
80 | $bp="%i2"; # const BN_ULONG *bp, | |
81 | $np="%i3"; # const BN_ULONG *np, | |
4d524040 | 82 | $n0="%i4"; # const BN_ULONG *n0, |
bcb43bb3 AP |
83 | $num="%i5"; # int num); |
84 | ||
aa2be094 | 85 | $tp="%l0"; # t[num] |
bcb43bb3 AP |
86 | $ap_l="%l1"; # a[num],n[num] are smashed to 32-bit words and saved |
87 | $ap_h="%l2"; # to these four vectors as double-precision FP values. | |
88 | $np_l="%l3"; # This way a bunch of fxtods are eliminated in second | |
89 | $np_h="%l4"; # loop and L1-cache aliasing is minimized... | |
90 | $i="%l5"; | |
91 | $j="%l6"; | |
92 | $mask="%l7"; # 16-bit mask, 0xffff | |
93 | ||
aa2be094 AP |
94 | $n0="%g4"; # reassigned(!) to "64-bit" register |
95 | $carry="%i4"; # %i4 reused(!) for a carry bit | |
bcb43bb3 AP |
96 | |
97 | # FP register naming chart | |
98 | # | |
99 | # ..HILO | |
100 | # dcba | |
101 | # -------- | |
102 | # LOa | |
103 | # LOb | |
104 | # LOc | |
105 | # LOd | |
106 | # HIa | |
107 | # HIb | |
108 | # HIc | |
109 | # HId | |
110 | # ..a | |
111 | # ..b | |
112 | $ba="%f0"; $bb="%f2"; $bc="%f4"; $bd="%f6"; | |
113 | $na="%f8"; $nb="%f10"; $nc="%f12"; $nd="%f14"; | |
114 | $alo="%f16"; $alo_="%f17"; $ahi="%f18"; $ahi_="%f19"; | |
115 | $nlo="%f20"; $nlo_="%f21"; $nhi="%f22"; $nhi_="%f23"; | |
116 | ||
117 | $dota="%f24"; $dotb="%f26"; | |
118 | ||
119 | $aloa="%f32"; $alob="%f34"; $aloc="%f36"; $alod="%f38"; | |
120 | $ahia="%f40"; $ahib="%f42"; $ahic="%f44"; $ahid="%f46"; | |
121 | $nloa="%f48"; $nlob="%f50"; $nloc="%f52"; $nlod="%f54"; | |
122 | $nhia="%f56"; $nhib="%f58"; $nhic="%f60"; $nhid="%f62"; | |
123 | ||
124 | $ASI_FL16_P=0xD2; # magic ASI value to engage 16-bit FP load | |
125 | ||
126 | $code=<<___; | |
eb77e888 AP |
127 | #include "sparc_arch.h" |
128 | ||
bcb43bb3 AP |
129 | .section ".text",#alloc,#execinstr |
130 | ||
131 | .global $fname | |
132 | .align 32 | |
133 | $fname: | |
aa2be094 | 134 | save %sp,-$frame-$locals,%sp |
6df8c74d | 135 | |
aa2be094 AP |
136 | cmp $num,4 |
137 | bl,a,pn %icc,.Lret | |
138 | clr %i0 | |
139 | andcc $num,1,%g0 ! $num has to be even... | |
140 | bnz,a,pn %icc,.Lret | |
141 | clr %i0 ! signal "unsupported input value" | |
760e3535 | 142 | |
aa2be094 | 143 | srl $num,1,$num |
760e3535 | 144 | sethi %hi(0xffff),$mask |
aa2be094 | 145 | ld [%i4+0],$n0 ! $n0 reassigned, remember? |
760e3535 | 146 | or $mask,%lo(0xffff),$mask |
aa2be094 AP |
147 | ld [%i4+4],%o0 |
148 | sllx %o0,32,%o0 | |
149 | or %o0,$n0,$n0 ! $n0=n0[1].n0[0] | |
6df8c74d | 150 | |
aa2be094 | 151 | sll $num,3,$num ! num*=8 |
bcb43bb3 AP |
152 | |
153 | add %sp,$bias,%o0 ! real top of stack | |
154 | sll $num,2,%o1 | |
155 | add %o1,$num,%o1 ! %o1=num*5 | |
156 | sub %o0,%o1,%o0 | |
bcb43bb3 | 157 | and %o0,-2048,%o0 ! optimize TLB utilization |
aa2be094 | 158 | sub %o0,$bias,%sp ! alloca(5*num*8) |
bcb43bb3 | 159 | |
aa2be094 | 160 | rd %asi,%o7 ! save %asi |
bcb43bb3 AP |
161 | add %sp,$bias+$frame+$locals,$tp |
162 | add $tp,$num,$ap_l | |
aa2be094 | 163 | add $ap_l,$num,$ap_l ! [an]p_[lh] point at the vectors' ends ! |
bcb43bb3 AP |
164 | add $ap_l,$num,$ap_h |
165 | add $ap_h,$num,$np_l | |
166 | add $np_l,$num,$np_h | |
167 | ||
168 | wr %g0,$ASI_FL16_P,%asi ! setup %asi for 16-bit FP loads | |
169 | ||
170 | add $rp,$num,$rp ! readjust input pointers to point | |
171 | add $ap,$num,$ap ! at the ends too... | |
172 | add $bp,$num,$bp | |
173 | add $np,$num,$np | |
174 | ||
aa2be094 | 175 | stx %o7,[%sp+$bias+$frame+48] ! save %asi |
bcb43bb3 | 176 | \f |
6df8c74d AP |
177 | sub %g0,$num,$i ! i=-num |
178 | sub %g0,$num,$j ! j=-num | |
bcb43bb3 AP |
179 | |
180 | add $ap,$j,%o3 | |
181 | add $bp,$i,%o4 | |
6df8c74d | 182 | |
87d3af64 AP |
183 | ld [%o3+4],%g1 ! bp[0] |
184 | ld [%o3+0],%o0 | |
185 | ld [%o4+4],%g5 ! ap[0] | |
186 | sllx %g1,32,%g1 | |
187 | ld [%o4+0],%o1 | |
188 | sllx %g5,32,%g5 | |
6df8c74d AP |
189 | or %g1,%o0,%o0 |
190 | or %g5,%o1,%o1 | |
191 | ||
aa2be094 | 192 | add $np,$j,%o5 |
bcb43bb3 AP |
193 | |
194 | mulx %o1,%o0,%o0 ! ap[0]*bp[0] | |
195 | mulx $n0,%o0,%o0 ! ap[0]*bp[0]*n0 | |
aa2be094 | 196 | stx %o0,[%sp+$bias+$frame+0] |
bcb43bb3 | 197 | |
6df8c74d | 198 | ld [%o3+0],$alo_ ! load a[j] as pair of 32-bit words |
aa2be094 | 199 | fzeros $alo |
6df8c74d | 200 | ld [%o3+4],$ahi_ |
aa2be094 | 201 | fzeros $ahi |
6df8c74d | 202 | ld [%o5+0],$nlo_ ! load n[j] as pair of 32-bit words |
aa2be094 | 203 | fzeros $nlo |
6df8c74d | 204 | ld [%o5+4],$nhi_ |
aa2be094 | 205 | fzeros $nhi |
bcb43bb3 AP |
206 | |
207 | ! transfer b[i] to FPU as 4x16-bit values | |
6df8c74d | 208 | ldda [%o4+2]%asi,$ba |
bcb43bb3 | 209 | fxtod $alo,$alo |
6df8c74d | 210 | ldda [%o4+0]%asi,$bb |
bcb43bb3 | 211 | fxtod $ahi,$ahi |
6df8c74d | 212 | ldda [%o4+6]%asi,$bc |
bcb43bb3 | 213 | fxtod $nlo,$nlo |
6df8c74d | 214 | ldda [%o4+4]%asi,$bd |
bcb43bb3 AP |
215 | fxtod $nhi,$nhi |
216 | ||
217 | ! transfer ap[0]*b[0]*n0 to FPU as 4x16-bit values | |
aa2be094 | 218 | ldda [%sp+$bias+$frame+6]%asi,$na |
bcb43bb3 | 219 | fxtod $ba,$ba |
aa2be094 | 220 | ldda [%sp+$bias+$frame+4]%asi,$nb |
bcb43bb3 | 221 | fxtod $bb,$bb |
aa2be094 | 222 | ldda [%sp+$bias+$frame+2]%asi,$nc |
bcb43bb3 | 223 | fxtod $bc,$bc |
aa2be094 | 224 | ldda [%sp+$bias+$frame+0]%asi,$nd |
bcb43bb3 AP |
225 | fxtod $bd,$bd |
226 | ||
227 | std $alo,[$ap_l+$j] ! save smashed ap[j] in double format | |
228 | fxtod $na,$na | |
229 | std $ahi,[$ap_h+$j] | |
230 | fxtod $nb,$nb | |
231 | std $nlo,[$np_l+$j] ! save smashed np[j] in double format | |
232 | fxtod $nc,$nc | |
233 | std $nhi,[$np_h+$j] | |
234 | fxtod $nd,$nd | |
235 | ||
aa2be094 AP |
236 | fmuld $alo,$ba,$aloa |
237 | fmuld $nlo,$na,$nloa | |
238 | fmuld $alo,$bb,$alob | |
239 | fmuld $nlo,$nb,$nlob | |
240 | fmuld $alo,$bc,$aloc | |
aa2be094 | 241 | faddd $aloa,$nloa,$nloa |
6df8c74d | 242 | fmuld $nlo,$nc,$nloc |
aa2be094 | 243 | fmuld $alo,$bd,$alod |
aa2be094 | 244 | faddd $alob,$nlob,$nlob |
6df8c74d | 245 | fmuld $nlo,$nd,$nlod |
aa2be094 | 246 | fmuld $ahi,$ba,$ahia |
aa2be094 | 247 | faddd $aloc,$nloc,$nloc |
6df8c74d | 248 | fmuld $nhi,$na,$nhia |
aa2be094 | 249 | fmuld $ahi,$bb,$ahib |
aa2be094 | 250 | faddd $alod,$nlod,$nlod |
6df8c74d | 251 | fmuld $nhi,$nb,$nhib |
aa2be094 | 252 | fmuld $ahi,$bc,$ahic |
aa2be094 | 253 | faddd $ahia,$nhia,$nhia |
6df8c74d | 254 | fmuld $nhi,$nc,$nhic |
aa2be094 | 255 | fmuld $ahi,$bd,$ahid |
6df8c74d | 256 | faddd $ahib,$nhib,$nhib |
aa2be094 | 257 | fmuld $nhi,$nd,$nhid |
bcb43bb3 | 258 | |
bcb43bb3 AP |
259 | faddd $ahic,$nhic,$dota ! $nhic |
260 | faddd $ahid,$nhid,$dotb ! $nhid | |
261 | ||
262 | faddd $nloc,$nhia,$nloc | |
263 | faddd $nlod,$nhib,$nlod | |
264 | ||
265 | fdtox $nloa,$nloa | |
266 | fdtox $nlob,$nlob | |
267 | fdtox $nloc,$nloc | |
268 | fdtox $nlod,$nlod | |
269 | ||
270 | std $nloa,[%sp+$bias+$frame+0] | |
2e21922e | 271 | add $j,8,$j |
bcb43bb3 | 272 | std $nlob,[%sp+$bias+$frame+8] |
2e21922e | 273 | add $ap,$j,%o4 |
bcb43bb3 | 274 | std $nloc,[%sp+$bias+$frame+16] |
2e21922e | 275 | add $np,$j,%o5 |
bcb43bb3 | 276 | std $nlod,[%sp+$bias+$frame+24] |
bcb43bb3 | 277 | \f |
1c3d2b94 | 278 | ld [%o4+0],$alo_ ! load a[j] as pair of 32-bit words |
aa2be094 | 279 | fzeros $alo |
1c3d2b94 | 280 | ld [%o4+4],$ahi_ |
aa2be094 | 281 | fzeros $ahi |
1c3d2b94 | 282 | ld [%o5+0],$nlo_ ! load n[j] as pair of 32-bit words |
aa2be094 | 283 | fzeros $nlo |
1c3d2b94 | 284 | ld [%o5+4],$nhi_ |
aa2be094 | 285 | fzeros $nhi |
bcb43bb3 AP |
286 | |
287 | fxtod $alo,$alo | |
288 | fxtod $ahi,$ahi | |
289 | fxtod $nlo,$nlo | |
290 | fxtod $nhi,$nhi | |
291 | ||
2e21922e | 292 | ldx [%sp+$bias+$frame+0],%o0 |
aa2be094 | 293 | fmuld $alo,$ba,$aloa |
2e21922e | 294 | ldx [%sp+$bias+$frame+8],%o1 |
aa2be094 | 295 | fmuld $nlo,$na,$nloa |
2e21922e | 296 | ldx [%sp+$bias+$frame+16],%o2 |
aa2be094 | 297 | fmuld $alo,$bb,$alob |
2e21922e | 298 | ldx [%sp+$bias+$frame+24],%o3 |
aa2be094 | 299 | fmuld $nlo,$nb,$nlob |
2e21922e AP |
300 | |
301 | srlx %o0,16,%o7 | |
302 | std $alo,[$ap_l+$j] ! save smashed ap[j] in double format | |
aa2be094 | 303 | fmuld $alo,$bc,$aloc |
2e21922e AP |
304 | add %o7,%o1,%o1 |
305 | std $ahi,[$ap_h+$j] | |
306 | faddd $aloa,$nloa,$nloa | |
6df8c74d | 307 | fmuld $nlo,$nc,$nloc |
2e21922e AP |
308 | srlx %o1,16,%o7 |
309 | std $nlo,[$np_l+$j] ! save smashed np[j] in double format | |
aa2be094 | 310 | fmuld $alo,$bd,$alod |
2e21922e AP |
311 | add %o7,%o2,%o2 |
312 | std $nhi,[$np_h+$j] | |
313 | faddd $alob,$nlob,$nlob | |
6df8c74d | 314 | fmuld $nlo,$nd,$nlod |
2e21922e | 315 | srlx %o2,16,%o7 |
aa2be094 | 316 | fmuld $ahi,$ba,$ahia |
2e21922e AP |
317 | add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15] |
318 | faddd $aloc,$nloc,$nloc | |
6df8c74d | 319 | fmuld $nhi,$na,$nhia |
2e21922e AP |
320 | !and %o0,$mask,%o0 |
321 | !and %o1,$mask,%o1 | |
322 | !and %o2,$mask,%o2 | |
323 | !sllx %o1,16,%o1 | |
324 | !sllx %o2,32,%o2 | |
325 | !sllx %o3,48,%o7 | |
326 | !or %o1,%o0,%o0 | |
327 | !or %o2,%o0,%o0 | |
328 | !or %o7,%o0,%o0 ! 64-bit result | |
329 | srlx %o3,16,%g1 ! 34-bit carry | |
aa2be094 | 330 | fmuld $ahi,$bb,$ahib |
2e21922e | 331 | |
aa2be094 | 332 | faddd $alod,$nlod,$nlod |
6df8c74d | 333 | fmuld $nhi,$nb,$nhib |
aa2be094 | 334 | fmuld $ahi,$bc,$ahic |
aa2be094 | 335 | faddd $ahia,$nhia,$nhia |
6df8c74d | 336 | fmuld $nhi,$nc,$nhic |
aa2be094 | 337 | fmuld $ahi,$bd,$ahid |
aa2be094 | 338 | faddd $ahib,$nhib,$nhib |
6df8c74d | 339 | fmuld $nhi,$nd,$nhid |
bcb43bb3 AP |
340 | |
341 | faddd $dota,$nloa,$nloa | |
342 | faddd $dotb,$nlob,$nlob | |
343 | faddd $ahic,$nhic,$dota ! $nhic | |
344 | faddd $ahid,$nhid,$dotb ! $nhid | |
345 | ||
346 | faddd $nloc,$nhia,$nloc | |
347 | faddd $nlod,$nhib,$nlod | |
348 | ||
349 | fdtox $nloa,$nloa | |
350 | fdtox $nlob,$nlob | |
351 | fdtox $nloc,$nloc | |
352 | fdtox $nlod,$nlod | |
353 | ||
354 | std $nloa,[%sp+$bias+$frame+0] | |
355 | std $nlob,[%sp+$bias+$frame+8] | |
2e21922e | 356 | addcc $j,8,$j |
bcb43bb3 | 357 | std $nloc,[%sp+$bias+$frame+16] |
2e21922e | 358 | bz,pn %icc,.L1stskip |
bcb43bb3 | 359 | std $nlod,[%sp+$bias+$frame+24] |
1c3d2b94 | 360 | \f |
23296942 | 361 | .align 32 ! incidentally already aligned ! |
1c3d2b94 | 362 | .L1st: |
1c3d2b94 AP |
363 | add $ap,$j,%o4 |
364 | add $np,$j,%o5 | |
365 | ld [%o4+0],$alo_ ! load a[j] as pair of 32-bit words | |
366 | fzeros $alo | |
367 | ld [%o4+4],$ahi_ | |
368 | fzeros $ahi | |
369 | ld [%o5+0],$nlo_ ! load n[j] as pair of 32-bit words | |
370 | fzeros $nlo | |
371 | ld [%o5+4],$nhi_ | |
372 | fzeros $nhi | |
373 | ||
374 | fxtod $alo,$alo | |
375 | fxtod $ahi,$ahi | |
376 | fxtod $nlo,$nlo | |
377 | fxtod $nhi,$nhi | |
378 | ||
2e21922e | 379 | ldx [%sp+$bias+$frame+0],%o0 |
1c3d2b94 | 380 | fmuld $alo,$ba,$aloa |
2e21922e | 381 | ldx [%sp+$bias+$frame+8],%o1 |
1c3d2b94 | 382 | fmuld $nlo,$na,$nloa |
2e21922e | 383 | ldx [%sp+$bias+$frame+16],%o2 |
1c3d2b94 | 384 | fmuld $alo,$bb,$alob |
2e21922e | 385 | ldx [%sp+$bias+$frame+24],%o3 |
1c3d2b94 | 386 | fmuld $nlo,$nb,$nlob |
2e21922e AP |
387 | |
388 | srlx %o0,16,%o7 | |
389 | std $alo,[$ap_l+$j] ! save smashed ap[j] in double format | |
1c3d2b94 | 390 | fmuld $alo,$bc,$aloc |
2e21922e AP |
391 | add %o7,%o1,%o1 |
392 | std $ahi,[$ap_h+$j] | |
393 | faddd $aloa,$nloa,$nloa | |
1c3d2b94 | 394 | fmuld $nlo,$nc,$nloc |
2e21922e AP |
395 | srlx %o1,16,%o7 |
396 | std $nlo,[$np_l+$j] ! save smashed np[j] in double format | |
1c3d2b94 | 397 | fmuld $alo,$bd,$alod |
2e21922e AP |
398 | add %o7,%o2,%o2 |
399 | std $nhi,[$np_h+$j] | |
400 | faddd $alob,$nlob,$nlob | |
1c3d2b94 | 401 | fmuld $nlo,$nd,$nlod |
2e21922e | 402 | srlx %o2,16,%o7 |
1c3d2b94 | 403 | fmuld $ahi,$ba,$ahia |
2e21922e AP |
404 | add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15] |
405 | and %o0,$mask,%o0 | |
406 | faddd $aloc,$nloc,$nloc | |
1c3d2b94 | 407 | fmuld $nhi,$na,$nhia |
2e21922e AP |
408 | and %o1,$mask,%o1 |
409 | and %o2,$mask,%o2 | |
1c3d2b94 | 410 | fmuld $ahi,$bb,$ahib |
2e21922e AP |
411 | sllx %o1,16,%o1 |
412 | faddd $alod,$nlod,$nlod | |
1c3d2b94 | 413 | fmuld $nhi,$nb,$nhib |
2e21922e | 414 | sllx %o2,32,%o2 |
1c3d2b94 | 415 | fmuld $ahi,$bc,$ahic |
2e21922e AP |
416 | sllx %o3,48,%o7 |
417 | or %o1,%o0,%o0 | |
418 | faddd $ahia,$nhia,$nhia | |
1c3d2b94 | 419 | fmuld $nhi,$nc,$nhic |
2e21922e | 420 | or %o2,%o0,%o0 |
1c3d2b94 | 421 | fmuld $ahi,$bd,$ahid |
2e21922e AP |
422 | or %o7,%o0,%o0 ! 64-bit result |
423 | faddd $ahib,$nhib,$nhib | |
1c3d2b94 | 424 | fmuld $nhi,$nd,$nhid |
2e21922e AP |
425 | addcc %g1,%o0,%o0 |
426 | faddd $dota,$nloa,$nloa | |
427 | srlx %o3,16,%g1 ! 34-bit carry | |
428 | faddd $dotb,$nlob,$nlob | |
429 | bcs,a %xcc,.+8 | |
430 | add %g1,1,%g1 | |
431 | ||
432 | stx %o0,[$tp] ! tp[j-1]= | |
1c3d2b94 | 433 | |
1c3d2b94 AP |
434 | faddd $ahic,$nhic,$dota ! $nhic |
435 | faddd $ahid,$nhid,$dotb ! $nhid | |
436 | ||
437 | faddd $nloc,$nhia,$nloc | |
438 | faddd $nlod,$nhib,$nlod | |
439 | ||
440 | fdtox $nloa,$nloa | |
441 | fdtox $nlob,$nlob | |
442 | fdtox $nloc,$nloc | |
443 | fdtox $nlod,$nlod | |
444 | ||
445 | std $nloa,[%sp+$bias+$frame+0] | |
446 | std $nlob,[%sp+$bias+$frame+8] | |
447 | std $nloc,[%sp+$bias+$frame+16] | |
448 | std $nlod,[%sp+$bias+$frame+24] | |
449 | ||
aa2be094 AP |
450 | addcc $j,8,$j |
451 | bnz,pt %icc,.L1st | |
bcb43bb3 | 452 | add $tp,8,$tp |
1c3d2b94 AP |
453 | \f |
454 | .L1stskip: | |
ebae8092 AP |
455 | fdtox $dota,$dota |
456 | fdtox $dotb,$dotb | |
457 | ||
1c3d2b94 AP |
458 | ldx [%sp+$bias+$frame+0],%o0 |
459 | ldx [%sp+$bias+$frame+8],%o1 | |
460 | ldx [%sp+$bias+$frame+16],%o2 | |
461 | ldx [%sp+$bias+$frame+24],%o3 | |
462 | ||
463 | srlx %o0,16,%o7 | |
ebae8092 | 464 | std $dota,[%sp+$bias+$frame+32] |
1c3d2b94 | 465 | add %o7,%o1,%o1 |
ebae8092 | 466 | std $dotb,[%sp+$bias+$frame+40] |
1c3d2b94 AP |
467 | srlx %o1,16,%o7 |
468 | add %o7,%o2,%o2 | |
469 | srlx %o2,16,%o7 | |
470 | add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15] | |
471 | and %o0,$mask,%o0 | |
472 | and %o1,$mask,%o1 | |
473 | and %o2,$mask,%o2 | |
474 | sllx %o1,16,%o1 | |
475 | sllx %o2,32,%o2 | |
476 | sllx %o3,48,%o7 | |
477 | or %o1,%o0,%o0 | |
478 | or %o2,%o0,%o0 | |
479 | or %o7,%o0,%o0 ! 64-bit result | |
ebae8092 | 480 | ldx [%sp+$bias+$frame+32],%o4 |
1c3d2b94 | 481 | addcc %g1,%o0,%o0 |
ebae8092 | 482 | ldx [%sp+$bias+$frame+40],%o5 |
1c3d2b94 AP |
483 | srlx %o3,16,%g1 ! 34-bit carry |
484 | bcs,a %xcc,.+8 | |
485 | add %g1,1,%g1 | |
486 | ||
487 | stx %o0,[$tp] ! tp[j-1]= | |
488 | add $tp,8,$tp | |
bcb43bb3 | 489 | |
ebae8092 AP |
490 | srlx %o4,16,%o7 |
491 | add %o7,%o5,%o5 | |
492 | and %o4,$mask,%o4 | |
493 | sllx %o5,16,%o7 | |
494 | or %o7,%o4,%o4 | |
495 | addcc %g1,%o4,%o4 | |
496 | srlx %o5,48,%g1 | |
bcb43bb3 AP |
497 | bcs,a %xcc,.+8 |
498 | add %g1,1,%g1 | |
499 | ||
500 | mov %g1,$carry | |
ebae8092 | 501 | stx %o4,[$tp] ! tp[num-1]= |
bcb43bb3 AP |
502 | \f |
503 | ba .Louter | |
504 | add $i,8,$i | |
505 | .align 32 | |
506 | .Louter: | |
6df8c74d | 507 | sub %g0,$num,$j ! j=-num |
bcb43bb3 AP |
508 | add %sp,$bias+$frame+$locals,$tp |
509 | ||
87d3af64 | 510 | add $ap,$j,%o3 |
bcb43bb3 | 511 | add $bp,$i,%o4 |
6df8c74d | 512 | |
87d3af64 AP |
513 | ld [%o3+4],%g1 ! bp[i] |
514 | ld [%o3+0],%o0 | |
515 | ld [%o4+4],%g5 ! ap[0] | |
516 | sllx %g1,32,%g1 | |
517 | ld [%o4+0],%o1 | |
518 | sllx %g5,32,%g5 | |
6df8c74d AP |
519 | or %g1,%o0,%o0 |
520 | or %g5,%o1,%o1 | |
521 | ||
bcb43bb3 AP |
522 | ldx [$tp],%o2 ! tp[0] |
523 | mulx %o1,%o0,%o0 | |
524 | addcc %o2,%o0,%o0 | |
525 | mulx $n0,%o0,%o0 ! (ap[0]*bp[i]+t[0])*n0 | |
aa2be094 | 526 | stx %o0,[%sp+$bias+$frame+0] |
bcb43bb3 | 527 | |
bcb43bb3 | 528 | ! transfer b[i] to FPU as 4x16-bit values |
6df8c74d AP |
529 | ldda [%o4+2]%asi,$ba |
530 | ldda [%o4+0]%asi,$bb | |
531 | ldda [%o4+6]%asi,$bc | |
532 | ldda [%o4+4]%asi,$bd | |
bcb43bb3 AP |
533 | |
534 | ! transfer (ap[0]*b[i]+t[0])*n0 to FPU as 4x16-bit values | |
aa2be094 | 535 | ldda [%sp+$bias+$frame+6]%asi,$na |
bcb43bb3 | 536 | fxtod $ba,$ba |
aa2be094 | 537 | ldda [%sp+$bias+$frame+4]%asi,$nb |
bcb43bb3 | 538 | fxtod $bb,$bb |
aa2be094 | 539 | ldda [%sp+$bias+$frame+2]%asi,$nc |
bcb43bb3 | 540 | fxtod $bc,$bc |
aa2be094 | 541 | ldda [%sp+$bias+$frame+0]%asi,$nd |
bcb43bb3 AP |
542 | fxtod $bd,$bd |
543 | ldd [$ap_l+$j],$alo ! load a[j] in double format | |
544 | fxtod $na,$na | |
545 | ldd [$ap_h+$j],$ahi | |
546 | fxtod $nb,$nb | |
547 | ldd [$np_l+$j],$nlo ! load n[j] in double format | |
548 | fxtod $nc,$nc | |
549 | ldd [$np_h+$j],$nhi | |
550 | fxtod $nd,$nd | |
551 | ||
aa2be094 AP |
552 | fmuld $alo,$ba,$aloa |
553 | fmuld $nlo,$na,$nloa | |
554 | fmuld $alo,$bb,$alob | |
555 | fmuld $nlo,$nb,$nlob | |
556 | fmuld $alo,$bc,$aloc | |
aa2be094 | 557 | faddd $aloa,$nloa,$nloa |
6df8c74d | 558 | fmuld $nlo,$nc,$nloc |
aa2be094 | 559 | fmuld $alo,$bd,$alod |
aa2be094 | 560 | faddd $alob,$nlob,$nlob |
6df8c74d | 561 | fmuld $nlo,$nd,$nlod |
aa2be094 | 562 | fmuld $ahi,$ba,$ahia |
aa2be094 | 563 | faddd $aloc,$nloc,$nloc |
6df8c74d | 564 | fmuld $nhi,$na,$nhia |
aa2be094 | 565 | fmuld $ahi,$bb,$ahib |
aa2be094 | 566 | faddd $alod,$nlod,$nlod |
6df8c74d | 567 | fmuld $nhi,$nb,$nhib |
aa2be094 | 568 | fmuld $ahi,$bc,$ahic |
aa2be094 | 569 | faddd $ahia,$nhia,$nhia |
6df8c74d | 570 | fmuld $nhi,$nc,$nhic |
aa2be094 | 571 | fmuld $ahi,$bd,$ahid |
6df8c74d | 572 | faddd $ahib,$nhib,$nhib |
aa2be094 | 573 | fmuld $nhi,$nd,$nhid |
bcb43bb3 | 574 | |
bcb43bb3 AP |
575 | faddd $ahic,$nhic,$dota ! $nhic |
576 | faddd $ahid,$nhid,$dotb ! $nhid | |
577 | ||
578 | faddd $nloc,$nhia,$nloc | |
579 | faddd $nlod,$nhib,$nlod | |
580 | ||
581 | fdtox $nloa,$nloa | |
582 | fdtox $nlob,$nlob | |
583 | fdtox $nloc,$nloc | |
584 | fdtox $nlod,$nlod | |
585 | ||
586 | std $nloa,[%sp+$bias+$frame+0] | |
587 | std $nlob,[%sp+$bias+$frame+8] | |
588 | std $nloc,[%sp+$bias+$frame+16] | |
2e21922e | 589 | add $j,8,$j |
bcb43bb3 | 590 | std $nlod,[%sp+$bias+$frame+24] |
2e21922e AP |
591 | \f |
592 | ldd [$ap_l+$j],$alo ! load a[j] in double format | |
593 | ldd [$ap_h+$j],$ahi | |
594 | ldd [$np_l+$j],$nlo ! load n[j] in double format | |
595 | ldd [$np_h+$j],$nhi | |
596 | ||
597 | fmuld $alo,$ba,$aloa | |
598 | fmuld $nlo,$na,$nloa | |
599 | fmuld $alo,$bb,$alob | |
600 | fmuld $nlo,$nb,$nlob | |
601 | fmuld $alo,$bc,$aloc | |
bcb43bb3 | 602 | ldx [%sp+$bias+$frame+0],%o0 |
2e21922e AP |
603 | faddd $aloa,$nloa,$nloa |
604 | fmuld $nlo,$nc,$nloc | |
bcb43bb3 | 605 | ldx [%sp+$bias+$frame+8],%o1 |
2e21922e | 606 | fmuld $alo,$bd,$alod |
bcb43bb3 | 607 | ldx [%sp+$bias+$frame+16],%o2 |
2e21922e AP |
608 | faddd $alob,$nlob,$nlob |
609 | fmuld $nlo,$nd,$nlod | |
bcb43bb3 | 610 | ldx [%sp+$bias+$frame+24],%o3 |
2e21922e | 611 | fmuld $ahi,$ba,$ahia |
bcb43bb3 AP |
612 | |
613 | srlx %o0,16,%o7 | |
2e21922e AP |
614 | faddd $aloc,$nloc,$nloc |
615 | fmuld $nhi,$na,$nhia | |
bcb43bb3 | 616 | add %o7,%o1,%o1 |
2e21922e | 617 | fmuld $ahi,$bb,$ahib |
bcb43bb3 | 618 | srlx %o1,16,%o7 |
2e21922e AP |
619 | faddd $alod,$nlod,$nlod |
620 | fmuld $nhi,$nb,$nhib | |
bcb43bb3 | 621 | add %o7,%o2,%o2 |
2e21922e | 622 | fmuld $ahi,$bc,$ahic |
bcb43bb3 | 623 | srlx %o2,16,%o7 |
2e21922e AP |
624 | faddd $ahia,$nhia,$nhia |
625 | fmuld $nhi,$nc,$nhic | |
bcb43bb3 AP |
626 | add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15] |
627 | ! why? | |
628 | and %o0,$mask,%o0 | |
2e21922e | 629 | fmuld $ahi,$bd,$ahid |
bcb43bb3 AP |
630 | and %o1,$mask,%o1 |
631 | and %o2,$mask,%o2 | |
2e21922e AP |
632 | faddd $ahib,$nhib,$nhib |
633 | fmuld $nhi,$nd,$nhid | |
bcb43bb3 | 634 | sllx %o1,16,%o1 |
2e21922e | 635 | faddd $dota,$nloa,$nloa |
bcb43bb3 | 636 | sllx %o2,32,%o2 |
2e21922e | 637 | faddd $dotb,$nlob,$nlob |
bcb43bb3 AP |
638 | sllx %o3,48,%o7 |
639 | or %o1,%o0,%o0 | |
2e21922e | 640 | faddd $ahic,$nhic,$dota ! $nhic |
bcb43bb3 | 641 | or %o2,%o0,%o0 |
2e21922e | 642 | faddd $ahid,$nhid,$dotb ! $nhid |
bcb43bb3 AP |
643 | or %o7,%o0,%o0 ! 64-bit result |
644 | ldx [$tp],%o7 | |
2e21922e | 645 | faddd $nloc,$nhia,$nloc |
bcb43bb3 AP |
646 | addcc %o7,%o0,%o0 |
647 | ! end-of-why? | |
2e21922e | 648 | faddd $nlod,$nhib,$nlod |
bcb43bb3 | 649 | srlx %o3,16,%g1 ! 34-bit carry |
2e21922e | 650 | fdtox $nloa,$nloa |
bcb43bb3 AP |
651 | bcs,a %xcc,.+8 |
652 | add %g1,1,%g1 | |
bcb43bb3 | 653 | |
bcb43bb3 AP |
654 | fdtox $nlob,$nlob |
655 | fdtox $nloc,$nloc | |
656 | fdtox $nlod,$nlod | |
657 | ||
658 | std $nloa,[%sp+$bias+$frame+0] | |
659 | std $nlob,[%sp+$bias+$frame+8] | |
2e21922e | 660 | addcc $j,8,$j |
bcb43bb3 | 661 | std $nloc,[%sp+$bias+$frame+16] |
2e21922e | 662 | bz,pn %icc,.Linnerskip |
bcb43bb3 | 663 | std $nlod,[%sp+$bias+$frame+24] |
1c3d2b94 | 664 | \f |
ebae8092 AP |
665 | ba .Linner |
666 | nop | |
667 | .align 32 | |
1c3d2b94 | 668 | .Linner: |
2e21922e AP |
669 | ldd [$ap_l+$j],$alo ! load a[j] in double format |
670 | ldd [$ap_h+$j],$ahi | |
671 | ldd [$np_l+$j],$nlo ! load n[j] in double format | |
672 | ldd [$np_h+$j],$nhi | |
673 | ||
674 | fmuld $alo,$ba,$aloa | |
675 | fmuld $nlo,$na,$nloa | |
676 | fmuld $alo,$bb,$alob | |
677 | fmuld $nlo,$nb,$nlob | |
678 | fmuld $alo,$bc,$aloc | |
bcb43bb3 | 679 | ldx [%sp+$bias+$frame+0],%o0 |
2e21922e AP |
680 | faddd $aloa,$nloa,$nloa |
681 | fmuld $nlo,$nc,$nloc | |
bcb43bb3 | 682 | ldx [%sp+$bias+$frame+8],%o1 |
2e21922e | 683 | fmuld $alo,$bd,$alod |
bcb43bb3 | 684 | ldx [%sp+$bias+$frame+16],%o2 |
2e21922e AP |
685 | faddd $alob,$nlob,$nlob |
686 | fmuld $nlo,$nd,$nlod | |
bcb43bb3 | 687 | ldx [%sp+$bias+$frame+24],%o3 |
2e21922e | 688 | fmuld $ahi,$ba,$ahia |
bcb43bb3 AP |
689 | |
690 | srlx %o0,16,%o7 | |
2e21922e AP |
691 | faddd $aloc,$nloc,$nloc |
692 | fmuld $nhi,$na,$nhia | |
bcb43bb3 | 693 | add %o7,%o1,%o1 |
2e21922e | 694 | fmuld $ahi,$bb,$ahib |
bcb43bb3 | 695 | srlx %o1,16,%o7 |
2e21922e AP |
696 | faddd $alod,$nlod,$nlod |
697 | fmuld $nhi,$nb,$nhib | |
bcb43bb3 | 698 | add %o7,%o2,%o2 |
2e21922e | 699 | fmuld $ahi,$bc,$ahic |
bcb43bb3 | 700 | srlx %o2,16,%o7 |
2e21922e AP |
701 | faddd $ahia,$nhia,$nhia |
702 | fmuld $nhi,$nc,$nhic | |
bcb43bb3 AP |
703 | add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15] |
704 | and %o0,$mask,%o0 | |
2e21922e | 705 | fmuld $ahi,$bd,$ahid |
bcb43bb3 AP |
706 | and %o1,$mask,%o1 |
707 | and %o2,$mask,%o2 | |
2e21922e AP |
708 | faddd $ahib,$nhib,$nhib |
709 | fmuld $nhi,$nd,$nhid | |
bcb43bb3 | 710 | sllx %o1,16,%o1 |
2e21922e | 711 | faddd $dota,$nloa,$nloa |
bcb43bb3 | 712 | sllx %o2,32,%o2 |
2e21922e | 713 | faddd $dotb,$nlob,$nlob |
bcb43bb3 AP |
714 | sllx %o3,48,%o7 |
715 | or %o1,%o0,%o0 | |
2e21922e | 716 | faddd $ahic,$nhic,$dota ! $nhic |
bcb43bb3 | 717 | or %o2,%o0,%o0 |
2e21922e | 718 | faddd $ahid,$nhid,$dotb ! $nhid |
bcb43bb3 | 719 | or %o7,%o0,%o0 ! 64-bit result |
2e21922e | 720 | faddd $nloc,$nhia,$nloc |
bcb43bb3 | 721 | addcc %g1,%o0,%o0 |
ebae8092 | 722 | ldx [$tp+8],%o7 ! tp[j] |
2e21922e | 723 | faddd $nlod,$nhib,$nlod |
bcb43bb3 | 724 | srlx %o3,16,%g1 ! 34-bit carry |
2e21922e | 725 | fdtox $nloa,$nloa |
bcb43bb3 AP |
726 | bcs,a %xcc,.+8 |
727 | add %g1,1,%g1 | |
2e21922e | 728 | fdtox $nlob,$nlob |
bcb43bb3 | 729 | addcc %o7,%o0,%o0 |
2e21922e | 730 | fdtox $nloc,$nloc |
bcb43bb3 AP |
731 | bcs,a %xcc,.+8 |
732 | add %g1,1,%g1 | |
733 | ||
734 | stx %o0,[$tp] ! tp[j-1] | |
2e21922e | 735 | fdtox $nlod,$nlod |
1c3d2b94 AP |
736 | |
737 | std $nloa,[%sp+$bias+$frame+0] | |
738 | std $nlob,[%sp+$bias+$frame+8] | |
739 | std $nloc,[%sp+$bias+$frame+16] | |
aa2be094 | 740 | addcc $j,8,$j |
2e21922e | 741 | std $nlod,[%sp+$bias+$frame+24] |
aa2be094 | 742 | bnz,pt %icc,.Linner |
bcb43bb3 | 743 | add $tp,8,$tp |
1c3d2b94 AP |
744 | \f |
745 | .Linnerskip: | |
2e21922e AP |
746 | fdtox $dota,$dota |
747 | fdtox $dotb,$dotb | |
748 | ||
1c3d2b94 AP |
749 | ldx [%sp+$bias+$frame+0],%o0 |
750 | ldx [%sp+$bias+$frame+8],%o1 | |
751 | ldx [%sp+$bias+$frame+16],%o2 | |
752 | ldx [%sp+$bias+$frame+24],%o3 | |
753 | ||
754 | srlx %o0,16,%o7 | |
2e21922e | 755 | std $dota,[%sp+$bias+$frame+32] |
1c3d2b94 | 756 | add %o7,%o1,%o1 |
2e21922e | 757 | std $dotb,[%sp+$bias+$frame+40] |
1c3d2b94 AP |
758 | srlx %o1,16,%o7 |
759 | add %o7,%o2,%o2 | |
760 | srlx %o2,16,%o7 | |
761 | add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15] | |
762 | and %o0,$mask,%o0 | |
763 | and %o1,$mask,%o1 | |
764 | and %o2,$mask,%o2 | |
765 | sllx %o1,16,%o1 | |
766 | sllx %o2,32,%o2 | |
767 | sllx %o3,48,%o7 | |
768 | or %o1,%o0,%o0 | |
769 | or %o2,%o0,%o0 | |
2e21922e | 770 | ldx [%sp+$bias+$frame+32],%o4 |
1c3d2b94 | 771 | or %o7,%o0,%o0 ! 64-bit result |
2e21922e | 772 | ldx [%sp+$bias+$frame+40],%o5 |
1c3d2b94 | 773 | addcc %g1,%o0,%o0 |
2e21922e | 774 | ldx [$tp+8],%o7 ! tp[j] |
1c3d2b94 AP |
775 | srlx %o3,16,%g1 ! 34-bit carry |
776 | bcs,a %xcc,.+8 | |
777 | add %g1,1,%g1 | |
778 | ||
1c3d2b94 AP |
779 | addcc %o7,%o0,%o0 |
780 | bcs,a %xcc,.+8 | |
781 | add %g1,1,%g1 | |
782 | ||
783 | stx %o0,[$tp] ! tp[j-1] | |
784 | add $tp,8,$tp | |
bcb43bb3 | 785 | |
2e21922e AP |
786 | srlx %o4,16,%o7 |
787 | add %o7,%o5,%o5 | |
788 | and %o4,$mask,%o4 | |
789 | sllx %o5,16,%o7 | |
790 | or %o7,%o4,%o4 | |
791 | addcc %g1,%o4,%o4 | |
792 | srlx %o5,48,%g1 | |
bcb43bb3 AP |
793 | bcs,a %xcc,.+8 |
794 | add %g1,1,%g1 | |
795 | ||
2e21922e AP |
796 | addcc $carry,%o4,%o4 |
797 | stx %o4,[$tp] ! tp[num-1] | |
bcb43bb3 AP |
798 | mov %g1,$carry |
799 | bcs,a %xcc,.+8 | |
800 | add $carry,1,$carry | |
801 | ||
aa2be094 AP |
802 | addcc $i,8,$i |
803 | bnz %icc,.Louter | |
bcb43bb3 AP |
804 | nop |
805 | \f | |
7d9cf7c0 | 806 | add $tp,8,$tp ! adjust tp to point at the end |
7d9cf7c0 | 807 | orn %g0,%g0,%g4 |
7d9cf7c0 | 808 | sub %g0,$num,%o7 ! n=-num |
23296942 | 809 | ba .Lsub |
673c55a2 | 810 | subcc %g0,%g0,%g0 ! clear %icc.c |
23296942 AP |
811 | |
812 | .align 32 | |
bcb43bb3 | 813 | .Lsub: |
87d3af64 AP |
814 | ldx [$tp+%o7],%o0 |
815 | add $np,%o7,%g1 | |
816 | ld [%g1+0],%o2 | |
817 | ld [%g1+4],%o3 | |
818 | srlx %o0,32,%o1 | |
819 | subccc %o0,%o2,%o2 | |
820 | add $rp,%o7,%g1 | |
821 | subccc %o1,%o3,%o3 | |
822 | st %o2,[%g1+0] | |
aa2be094 AP |
823 | add %o7,8,%o7 |
824 | brnz,pt %o7,.Lsub | |
87d3af64 | 825 | st %o3,[%g1+4] |
7d9cf7c0 | 826 | subc $carry,0,%g4 |
6df8c74d | 827 | sub %g0,$num,%o7 ! n=-num |
23296942 AP |
828 | ba .Lcopy |
829 | nop | |
bcb43bb3 | 830 | |
23296942 | 831 | .align 32 |
bcb43bb3 | 832 | .Lcopy: |
aa2be094 | 833 | ldx [$tp+%o7],%o0 |
87d3af64 | 834 | add $rp,%o7,%g1 |
7d9cf7c0 AP |
835 | ld [%g1+0],%o2 |
836 | ld [%g1+4],%o3 | |
837 | stx %g0,[$tp+%o7] | |
838 | and %o0,%g4,%o0 | |
839 | srlx %o0,32,%o1 | |
840 | andn %o2,%g4,%o2 | |
841 | andn %o3,%g4,%o3 | |
842 | or %o2,%o0,%o0 | |
843 | or %o3,%o1,%o1 | |
87d3af64 | 844 | st %o0,[%g1+0] |
aa2be094 AP |
845 | add %o7,8,%o7 |
846 | brnz,pt %o7,.Lcopy | |
87d3af64 | 847 | st %o1,[%g1+4] |
6df8c74d | 848 | sub %g0,$num,%o7 ! n=-num |
bcb43bb3 | 849 | |
bcb43bb3 | 850 | .Lzap: |
aa2be094 AP |
851 | stx %g0,[$ap_l+%o7] |
852 | stx %g0,[$ap_h+%o7] | |
853 | stx %g0,[$np_l+%o7] | |
854 | stx %g0,[$np_h+%o7] | |
855 | add %o7,8,%o7 | |
856 | brnz,pt %o7,.Lzap | |
bcb43bb3 AP |
857 | nop |
858 | ||
859 | ldx [%sp+$bias+$frame+48],%o7 | |
860 | wr %g0,%o7,%asi ! restore %asi | |
861 | ||
862 | mov 1,%i0 | |
aa2be094 | 863 | .Lret: |
bcb43bb3 AP |
864 | ret |
865 | restore | |
866 | .type $fname,#function | |
867 | .size $fname,(.-$fname) | |
46f4e1be | 868 | .asciz "Montgomery Multiplication for UltraSPARC, CRYPTOGAMS by <appro\@openssl.org>" |
23296942 | 869 | .align 32 |
bcb43bb3 AP |
870 | ___ |
871 | ||
872 | $code =~ s/\`([^\`]*)\`/eval($1)/gem; | |
3b4a0225 AP |
873 | |
874 | # Below substitution makes it possible to compile without demanding | |
478b50cf | 875 | # VIS extensions on command line, e.g. -xarch=v9 vs. -xarch=v9a. I |
3b4a0225 AP |
876 | # dare to do this, because VIS capability is detected at run-time now |
877 | # and this routine is not called on CPU not capable to execute it. Do | |
878 | # note that fzeros is not the only VIS dependency! Another dependency | |
879 | # is implicit and is just _a_ numerical value loaded to %asi register, | |
880 | # which assembler can't recognize as VIS specific... | |
881 | $code =~ s/fzeros\s+%f([0-9]+)/ | |
882 | sprintf(".word\t0x%x\t! fzeros %%f%d",0x81b00c20|($1<<25),$1) | |
883 | /gem; | |
884 | ||
bcb43bb3 | 885 | print $code; |
3b4a0225 | 886 | # flush |
bcb43bb3 | 887 | close STDOUT; |