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 IA-64. There are
13 # several possibilities for improvement:
15 # - modulo-scheduling outer loop would eliminate quite a number of
16 # stalls after ldf8, xma and getf.sig outside inner loop and
17 # improve shorter key performance;
18 # - shorter vector support [with input vectors being fetched only
19 # once] should be added;
20 # - 2x unroll with help of n0[1] would make the code scalable on
21 # "wider" IA-64, "wider" than Itanium 2 that is, which is not of
22 # acute interest, because upcoming Tukwila's individual cores are
23 # reportedly based on Itanium 2 design;
24 # - dedicated squaring procedure(?);
28 # Shorter vector support is implemented by zero-padding ap and np
29 # vectors up to 8 elements, or 512 bits. This means that 256-bit
30 # inputs will be processed only 2 times faster than 512-bit inputs,
31 # not 4 [as one would expect, because algorithm complexity is n^2].
32 # The reason for padding is that inputs shorter than 512 bits won't
33 # be processed faster anyway, because minimal critical path of the
34 # core loop happens to match 512-bit timing. Either way, it resulted
35 # in >100% improvement of 512-bit RSA sign benchmark and 50% - of
36 # 1024-bit one [in comparison to original version of *this* module].
38 # So far 'openssl speed rsa dsa' output on 900MHz Itanium 2 *with*
40 # sign verify sign/s verify/s
41 # rsa 512 bits 0.000290s 0.000024s 3452.8 42031.4
42 # rsa 1024 bits 0.000793s 0.000058s 1261.7 17172.0
43 # rsa 2048 bits 0.005908s 0.000148s 169.3 6754.0
44 # rsa 4096 bits 0.033456s 0.000469s 29.9 2133.6
45 # dsa 512 bits 0.000253s 0.000198s 3949.9 5057.0
46 # dsa 1024 bits 0.000585s 0.000607s 1708.4 1647.4
47 # dsa 2048 bits 0.001453s 0.001703s 688.1 587.4
49 # ... and *without* (but still with ia64.S):
51 # rsa 512 bits 0.000670s 0.000041s 1491.8 24145.5
52 # rsa 1024 bits 0.001988s 0.000080s 502.9 12499.3
53 # rsa 2048 bits 0.008702s 0.000189s 114.9 5293.9
54 # rsa 4096 bits 0.043860s 0.000533s 22.8 1875.9
55 # dsa 512 bits 0.000441s 0.000427s 2265.3 2340.6
56 # dsa 1024 bits 0.000823s 0.000867s 1215.6 1153.2
57 # dsa 2048 bits 0.001894s 0.002179s 528.1 458.9
59 # As it can be seen, RSA sign performance improves by 130-30%,
60 # hereafter less for longer keys, while verify - by 74-13%.
61 # DSA performance improves by 115-30%.
67 for (@ARGV) { $ADDP="add" if (/[\+DD|\-mlp]64/); }
68 } else { $ADDP="add"; }
74 // int bn_mul_mont (BN_ULONG *rp,const BN_ULONG *ap,
75 // const BN_ULONG *bp,const BN_ULONG *np,
76 // const BN_ULONG *n0p,int num);
83 { .mmi; cmp4.le p6,p7=2,r37;;
84 (p6) cmp4.lt.unc p8,p9=8,r37
87 (p9) br.cond.dptk.many bn_mul_mont_8
88 (p8) br.cond.dpnt.many bn_mul_mont_general
89 (p7) br.ret.spnt.many b0 };;
92 prevfs=r2; prevpr=r3; prevlc=r10; prevsp=r11;
94 rptr=r8; aptr=r9; bptr=r14; nptr=r15;
97 num=r18; len=r19; lc=r20;
98 topbit=r21; // carry bit from tmp[num]
105 .local bn_mul_mont_general#
106 .proc bn_mul_mont_general#
109 { .mmi; .save ar.pfs,prevfs
110 alloc prevfs=ar.pfs,6,2,0,8
114 { .mmi; .vframe prevsp
121 .rotf alo[6],nlo[4],ahi[8],nhi[6]
124 { .mmi; ldf8 bi=[bptr],8 // (*bp++)
125 ldf8 alo[4]=[aptr],16 // ap[0]
127 { .mmi; ldf8 alo[3]=[r30],16 // ap[1]
128 ldf8 alo[2]=[aptr],16 // ap[2]
130 { .mmi; ldf8 alo[1]=[r30] // ap[3]
133 { .mmi; $ADDP nptr=0,in3
136 { .mmi; ldf8 nlo[2]=[nptr],8 // np[0]
138 shladd r31=num,3,r31 };;
139 { .mmi; ldf8 nlo[1]=[nptr],8 // np[1]
142 { .mfb; and sp=-16,r31 // alloca
143 xmpy.hu ahi[2]=alo[4],bi // ap[0]*bp[0]
146 xmpy.lu alo[4]=alo[4],bi
147 brp.loop.imp .L1st_ctop,.L1st_cend-16
150 xma.hu ahi[1]=alo[3],bi,ahi[2] // ap[1]*bp[0]
153 xma.lu alo[3]=alo[3],bi,ahi[2]
154 mov pr.rot=0x20001f<<16
155 // ------^----- (p40) at first (p23)
156 // ----------^^ p[16:20]=1
159 xmpy.lu m0=alo[4],n0 // (ap[0]*bp[0])*n0
162 fcvt.fxu.s1 nhi[1]=f0
167 .pred.rel "mutex",p40,p42
168 { .mfi; (p16) ldf8 alo[0]=[aptr],8 // *(aptr++)
169 (p18) xma.hu ahi[0]=alo[2],bi,ahi[1]
170 (p40) add n[2]=n[2],a[2] } // (p23) }
171 { .mfi; (p18) ldf8 nlo[0]=[nptr],8 // *(nptr++)(p16)
172 (p18) xma.lu alo[2]=alo[2],bi,ahi[1]
173 (p42) add n[2]=n[2],a[2],1 };; // (p23)
174 { .mfi; (p21) getf.sig a[0]=alo[5]
175 (p20) xma.hu nhi[0]=nlo[2],m0,nhi[1]
176 (p42) cmp.leu p41,p39=n[2],a[2] } // (p23)
177 { .mfi; (p23) st8 [tp_1]=n[2],8
178 (p20) xma.lu nlo[2]=nlo[2],m0,nhi[1]
179 (p40) cmp.ltu p41,p39=n[2],a[2] } // (p23)
180 { .mmb; (p21) getf.sig n[0]=nlo[3]
182 br.ctop.sptk .L1st_ctop };;
185 { .mmi; getf.sig a[0]=ahi[6] // (p24)
187 add num=-1,num };; // num--
188 { .mmi; .pred.rel "mutex",p40,p42
189 (p40) add n[0]=n[0],a[0]
190 (p42) add n[0]=n[0],a[0],1
191 sub aptr=aptr,len };; // rewind
192 { .mmi; .pred.rel "mutex",p40,p42
193 (p40) cmp.ltu p41,p39=n[0],a[0]
194 (p42) cmp.leu p41,p39=n[0],a[0]
195 sub nptr=nptr,len };;
196 { .mmi; .pred.rel "mutex",p39,p41
197 (p39) add topbit=r0,r0
198 (p41) add topbit=r0,r0,1
200 { .mmi; st8 [tp_1]=n[0]
205 { .mmi; ldf8 bi=[bptr],8 // (*bp++)
206 ldf8 ahi[3]=[tptr] // tp[0]
208 { .mmi; ldf8 alo[4]=[aptr],16 // ap[0]
209 ldf8 alo[3]=[r30],16 // ap[1]
211 { .mfb; ldf8 alo[2]=[aptr],16 // ap[2]
212 xma.hu ahi[2]=alo[4],bi,ahi[3] // ap[0]*bp[i]+tp[0]
213 brp.loop.imp .Linner_ctop,.Linner_cend-16
215 { .mfb; ldf8 alo[1]=[r30] // ap[3]
216 xma.lu alo[4]=alo[4],bi,ahi[3]
218 { .mfi; ldf8 nlo[2]=[nptr],16 // np[0]
219 xma.hu ahi[1]=alo[3],bi,ahi[2] // ap[1]*bp[i]
221 { .mfi; ldf8 nlo[1]=[r31] // np[1]
222 xma.lu alo[3]=alo[3],bi,ahi[2]
223 mov pr.rot=0x20101f<<16
224 // ------^----- (p40) at first (p23)
225 // --------^--- (p30) at first (p22)
226 // ----------^^ p[16:20]=1
228 { .mfi; st8 [tptr]=r0 // tp[0] is already accounted
229 xmpy.lu m0=alo[4],n0 // (ap[0]*bp[i]+tp[0])*n0
232 fcvt.fxu.s1 nhi[1]=f0
235 // This loop spins in 4*(n+7) ticks on Itanium 2 and should spin in
236 // 7*(n+7) ticks on Itanium (the one codenamed Merced). Factor of 7
237 // in latter case accounts for two-tick pipeline stall, which means
238 // that its performance would be ~20% lower than optimal one. No
239 // attempt was made to address this, because original Itanium is
240 // hardly represented out in the wild...
243 .pred.rel "mutex",p40,p42
244 .pred.rel "mutex",p30,p32
245 { .mfi; (p16) ldf8 alo[0]=[aptr],8 // *(aptr++)
246 (p18) xma.hu ahi[0]=alo[2],bi,ahi[1]
247 (p40) add n[2]=n[2],a[2] } // (p23)
248 { .mfi; (p16) nop.m 0
249 (p18) xma.lu alo[2]=alo[2],bi,ahi[1]
250 (p42) add n[2]=n[2],a[2],1 };; // (p23)
251 { .mfi; (p21) getf.sig a[0]=alo[5]
253 (p40) cmp.ltu p41,p39=n[2],a[2] } // (p23)
254 { .mfi; (p21) ld8 t[0]=[tptr],8
256 (p42) cmp.leu p41,p39=n[2],a[2] };; // (p23)
257 { .mfi; (p18) ldf8 nlo[0]=[nptr],8 // *(nptr++)
258 (p20) xma.hu nhi[0]=nlo[2],m0,nhi[1]
259 (p30) add a[1]=a[1],t[1] } // (p22)
260 { .mfi; (p16) nop.m 0
261 (p20) xma.lu nlo[2]=nlo[2],m0,nhi[1]
262 (p32) add a[1]=a[1],t[1],1 };; // (p22)
263 { .mmi; (p21) getf.sig n[0]=nlo[3]
265 (p30) cmp.ltu p31,p29=a[1],t[1] } // (p22)
266 { .mmb; (p23) st8 [tp_1]=n[2],8
267 (p32) cmp.leu p31,p29=a[1],t[1] // (p22)
268 br.ctop.sptk .Linner_ctop };;
271 { .mmi; getf.sig a[0]=ahi[6] // (p24)
275 { .mmi; .pred.rel "mutex",p31,p33
276 (p31) add a[0]=a[0],topbit
277 (p33) add a[0]=a[0],topbit,1
279 { .mfi; .pred.rel "mutex",p31,p33
280 (p31) cmp.ltu p32,p30=a[0],topbit
281 (p33) cmp.leu p32,p30=a[0],topbit
283 { .mfi; .pred.rel "mutex",p40,p42
284 (p40) add n[0]=n[0],a[0]
285 (p42) add n[0]=n[0],a[0],1
287 { .mmi; .pred.rel "mutex",p44,p46
288 (p40) cmp.ltu p41,p39=n[0],a[0]
289 (p42) cmp.leu p41,p39=n[0],a[0]
290 (p32) add topbit=r0,r0,1 }
292 { .mmi; st8 [tp_1]=n[0],8
294 sub aptr=aptr,len };; // rewind
295 { .mmi; sub nptr=nptr,len
296 (p41) add topbit=r0,r0,1
298 { .mmb; add tp_1=8,sp
299 add num=-1,num // num--
300 (p6) br.cond.sptk.many .Louter };;
303 brp.loop.imp .Lsub_ctop,.Lsub_cend-16
306 mov pr.rot=0x10001<<16
307 // ------^---- (p33) at first (p17)
314 .pred.rel "mutex",p33,p35
315 { .mfi; (p16) ld8 t[0]=[tptr],8 // t=*(tp++)
317 (p33) sub n[1]=t[1],n[1] } // (p17)
318 { .mfi; (p16) ld8 n[0]=[nptr],8 // n=*(np++)
320 (p35) sub n[1]=t[1],n[1],1 };; // (p17)
321 { .mib; (p18) st8 [rptr]=n[2],8 // *(rp++)=r
322 (p33) cmp.gtu p34,p32=n[1],t[1] // (p17)
324 { .mib; (p18) nop.m 0
325 (p35) cmp.geu p34,p32=n[1],t[1] // (p17)
326 br.ctop.sptk .Lsub_ctop };;
329 { .mmb; .pred.rel "mutex",p34,p36
330 (p34) sub topbit=topbit,r0 // (p19)
331 (p36) sub topbit=topbit,r0,1
332 brp.loop.imp .Lcopy_ctop,.Lcopy_cend-16
334 { .mmb; sub rptr=rptr,len // rewind
337 { .mmi; and aptr=tptr,topbit
338 andcm bptr=rptr,topbit
340 { .mii; or nptr=aptr,bptr
345 { .mmb; (p16) ld8 n[0]=[nptr],8
346 (p18) st8 [tptr]=r0,8
348 { .mmb; (p16) nop.m 0
349 (p18) st8 [rptr]=n[2],8
350 br.ctop.sptk .Lcopy_ctop };;
353 { .mmi; mov ret0=1 // signal "handled"
354 rum 1<<5 // clear um.mfh
358 mov pr=prevpr,0x1ffff
359 br.ret.sptk.many b0 };;
360 .endp bn_mul_mont_general#
362 a1=r16; a2=r17; a3=r18; a4=r19; a5=r20; a6=r21; a7=r22; a8=r23;
363 n1=r24; n2=r25; n3=r26; n4=r27; n5=r28; n6=r29; n7=r30; n8=r31;
366 ai0=f8; ai1=f9; ai2=f10; ai3=f11; ai4=f12; ai5=f13; ai6=f14; ai7=f15;
367 ni0=f16; ni1=f17; ni2=f18; ni3=f19; ni4=f20; ni5=f21; ni6=f22; ni7=f23;
370 .skip 48 // aligns loop body
371 .local bn_mul_mont_8#
375 { .mmi; .save ar.pfs,prevfs
376 alloc prevfs=ar.pfs,6,2,0,8
381 { .mmi; add r17=-6*16,sp
386 { .mmi; .save.gf 0,0x10
387 stf.spill [sp]=f16,-16
389 stf.spill [r17]=f17,32
390 add r16=-5*16,prevsp};;
391 { .mmi; .save.gf 0,0x40
392 stf.spill [r16]=f18,32
394 stf.spill [r17]=f19,32
396 { .mmi; .save.gf 0,0x100
397 stf.spill [r16]=f20,32
399 stf.spill [r17]=f21,32
401 { .mmi; .save.gf 0,0x400
408 .rotf bj[8],mj[2],tf[2],alo[10],ahi[10],nlo[10],nhi[10]
411 // load input vectors padding them to 8 elements
412 { .mmi; ldf8 ai0=[aptr],16 // ap[0]
413 ldf8 ai1=[r29],16 // ap[1]
415 { .mmi; $ADDP r30=8,in2
418 { .mmi; ldf8 bj[7]=[bptr],16 // bp[0]
419 ldf8 bj[6]=[r30],16 // bp[1]
420 cmp4.le p4,p5=3,in5 }
421 { .mmi; ldf8 ni0=[nptr],16 // np[0]
422 ldf8 ni1=[r31],16 // np[1]
423 cmp4.le p6,p7=4,in5 };;
425 { .mfi; (p4)ldf8 ai2=[aptr],16 // ap[2]
427 cmp4.le p8,p9=5,in5 }
428 { .mfi; (p6)ldf8 ai3=[r29],16 // ap[3]
430 cmp4.le p10,p11=6,in5 }
431 { .mfi; (p4)ldf8 bj[5]=[bptr],16 // bp[2]
432 (p5)fcvt.fxu bj[5]=f0
433 cmp4.le p12,p13=7,in5 }
434 { .mfi; (p6)ldf8 bj[4]=[r30],16 // bp[3]
435 (p7)fcvt.fxu bj[4]=f0
436 cmp4.le p14,p15=8,in5 }
437 { .mfi; (p4)ldf8 ni2=[nptr],16 // np[2]
440 { .mfi; (p6)ldf8 ni3=[r31],16 // np[3]
444 { .mfi; ldf8 n0=[in4]
448 { .mfi; (p8)ldf8 ai4=[aptr],16 // ap[4]
451 { .mfi; (p10)ldf8 ai5=[r29],16 // ap[5]
454 { .mfi; (p8)ldf8 bj[3]=[bptr],16 // bp[4]
455 (p9)fcvt.fxu bj[3]=f0
457 { .mfi; (p10)ldf8 bj[2]=[r30],16 // bp[5]
458 (p11)fcvt.fxu bj[2]=f0
460 { .mfi; (p8)ldf8 ni4=[nptr],16 // np[4]
463 { .mfi; (p10)ldf8 ni5=[r31],16 // np[5]
467 { .mfi; (p12)ldf8 ai6=[aptr],16 // ap[6]
470 { .mfi; (p14)ldf8 ai7=[r29],16 // ap[7]
473 { .mfi; (p12)ldf8 bj[1]=[bptr],16 // bp[6]
474 (p13)fcvt.fxu bj[1]=f0
476 { .mfi; (p14)ldf8 bj[0]=[r30],16 // bp[7]
477 (p15)fcvt.fxu bj[0]=f0
479 { .mfi; (p12)ldf8 ni6=[nptr],16 // np[6]
482 { .mfb; (p14)ldf8 ni7=[r31],16 // np[7]
484 brp.loop.imp .Louter_8_ctop,.Louter_8_cend-16
487 // The loop is scheduled for 32*n ticks on Itanium 2. Actual attempt
488 // to measure with help of Interval Time Counter indicated that the
489 // factor is a tad higher: 33 or 34, if not 35. Exact measurement and
490 // addressing the issue is problematic, because I don't have access
491 // to platform-specific instruction-level profiler. On Itanium it
492 // should run in 56*n ticks, because of higher xma latency...
494 .pred.rel "mutex",p40,p42
495 .pred.rel "mutex",p48,p50
496 { .mfi; (p16) nop.m 0 // 0:
497 (p16) xma.hu ahi[0]=ai0,bj[7],tf[1] // ap[0]*b[i]+t[0]
498 (p40) add a3=a3,n3 } // (p17) a3+=n3
499 { .mfi; (p42) add a3=a3,n3,1
500 (p16) xma.lu alo[0]=ai0,bj[7],tf[1]
502 { .mii; (p17) getf.sig a7=alo[8] // 1:
503 (p48) add t[6]=t[6],a3 // (p17) t[6]+=a3
504 (p50) add t[6]=t[6],a3,1 };;
505 { .mfi; (p17) getf.sig a8=ahi[8] // 2:
506 (p17) xma.hu nhi[7]=ni6,mj[1],nhi[6] // np[6]*m0
507 (p40) cmp.ltu p43,p41=a3,n3 }
508 { .mfi; (p42) cmp.leu p43,p41=a3,n3
509 (p17) xma.lu nlo[7]=ni6,mj[1],nhi[6]
511 { .mii; (p17) getf.sig n5=nlo[6] // 3:
512 (p48) cmp.ltu p51,p49=t[6],a3
513 (p50) cmp.leu p51,p49=t[6],a3 };;
514 .pred.rel "mutex",p41,p43
515 .pred.rel "mutex",p49,p51
516 { .mfi; (p16) nop.m 0 // 4:
517 (p16) xma.hu ahi[1]=ai1,bj[7],ahi[0] // ap[1]*b[i]
518 (p41) add a4=a4,n4 } // (p17) a4+=n4
519 { .mfi; (p43) add a4=a4,n4,1
520 (p16) xma.lu alo[1]=ai1,bj[7],ahi[0]
522 { .mfi; (p49) add t[5]=t[5],a4 // 5: (p17) t[5]+=a4
523 (p16) xmpy.lu mj[0]=alo[0],n0 // (ap[0]*b[i]+t[0])*n0
524 (p51) add t[5]=t[5],a4,1 };;
525 { .mfi; (p16) nop.m 0 // 6:
526 (p17) xma.hu nhi[8]=ni7,mj[1],nhi[7] // np[7]*m0
527 (p41) cmp.ltu p42,p40=a4,n4 }
528 { .mfi; (p43) cmp.leu p42,p40=a4,n4
529 (p17) xma.lu nlo[8]=ni7,mj[1],nhi[7]
531 { .mii; (p17) getf.sig n6=nlo[7] // 7:
532 (p49) cmp.ltu p50,p48=t[5],a4
533 (p51) cmp.leu p50,p48=t[5],a4 };;
534 .pred.rel "mutex",p40,p42
535 .pred.rel "mutex",p48,p50
536 { .mfi; (p16) nop.m 0 // 8:
537 (p16) xma.hu ahi[2]=ai2,bj[7],ahi[1] // ap[2]*b[i]
538 (p40) add a5=a5,n5 } // (p17) a5+=n5
539 { .mfi; (p42) add a5=a5,n5,1
540 (p16) xma.lu alo[2]=ai2,bj[7],ahi[1]
542 { .mii; (p16) getf.sig a1=alo[1] // 9:
543 (p48) add t[4]=t[4],a5 // p(17) t[4]+=a5
544 (p50) add t[4]=t[4],a5,1 };;
545 { .mfi; (p16) nop.m 0 // 10:
546 (p16) xma.hu nhi[0]=ni0,mj[0],alo[0] // np[0]*m0
547 (p40) cmp.ltu p43,p41=a5,n5 }
548 { .mfi; (p42) cmp.leu p43,p41=a5,n5
549 (p16) xma.lu nlo[0]=ni0,mj[0],alo[0]
551 { .mii; (p17) getf.sig n7=nlo[8] // 11:
552 (p48) cmp.ltu p51,p49=t[4],a5
553 (p50) cmp.leu p51,p49=t[4],a5 };;
554 .pred.rel "mutex",p41,p43
555 .pred.rel "mutex",p49,p51
556 { .mfi; (p17) getf.sig n8=nhi[8] // 12:
557 (p16) xma.hu ahi[3]=ai3,bj[7],ahi[2] // ap[3]*b[i]
558 (p41) add a6=a6,n6 } // (p17) a6+=n6
559 { .mfi; (p43) add a6=a6,n6,1
560 (p16) xma.lu alo[3]=ai3,bj[7],ahi[2]
562 { .mii; (p16) getf.sig a2=alo[2] // 13:
563 (p49) add t[3]=t[3],a6 // (p17) t[3]+=a6
564 (p51) add t[3]=t[3],a6,1 };;
565 { .mfi; (p16) nop.m 0 // 14:
566 (p16) xma.hu nhi[1]=ni1,mj[0],nhi[0] // np[1]*m0
567 (p41) cmp.ltu p42,p40=a6,n6 }
568 { .mfi; (p43) cmp.leu p42,p40=a6,n6
569 (p16) xma.lu nlo[1]=ni1,mj[0],nhi[0]
571 { .mii; (p16) nop.m 0 // 15:
572 (p49) cmp.ltu p50,p48=t[3],a6
573 (p51) cmp.leu p50,p48=t[3],a6 };;
574 .pred.rel "mutex",p40,p42
575 .pred.rel "mutex",p48,p50
576 { .mfi; (p16) nop.m 0 // 16:
577 (p16) xma.hu ahi[4]=ai4,bj[7],ahi[3] // ap[4]*b[i]
578 (p40) add a7=a7,n7 } // (p17) a7+=n7
579 { .mfi; (p42) add a7=a7,n7,1
580 (p16) xma.lu alo[4]=ai4,bj[7],ahi[3]
582 { .mii; (p16) getf.sig a3=alo[3] // 17:
583 (p48) add t[2]=t[2],a7 // (p17) t[2]+=a7
584 (p50) add t[2]=t[2],a7,1 };;
585 { .mfi; (p16) nop.m 0 // 18:
586 (p16) xma.hu nhi[2]=ni2,mj[0],nhi[1] // np[2]*m0
587 (p40) cmp.ltu p43,p41=a7,n7 }
588 { .mfi; (p42) cmp.leu p43,p41=a7,n7
589 (p16) xma.lu nlo[2]=ni2,mj[0],nhi[1]
591 { .mii; (p16) getf.sig n1=nlo[1] // 19:
592 (p48) cmp.ltu p51,p49=t[2],a7
593 (p50) cmp.leu p51,p49=t[2],a7 };;
594 .pred.rel "mutex",p41,p43
595 .pred.rel "mutex",p49,p51
596 { .mfi; (p16) nop.m 0 // 20:
597 (p16) xma.hu ahi[5]=ai5,bj[7],ahi[4] // ap[5]*b[i]
598 (p41) add a8=a8,n8 } // (p17) a8+=n8
599 { .mfi; (p43) add a8=a8,n8,1
600 (p16) xma.lu alo[5]=ai5,bj[7],ahi[4]
602 { .mii; (p16) getf.sig a4=alo[4] // 21:
603 (p49) add t[1]=t[1],a8 // (p17) t[1]+=a8
604 (p51) add t[1]=t[1],a8,1 };;
605 { .mfi; (p16) nop.m 0 // 22:
606 (p16) xma.hu nhi[3]=ni3,mj[0],nhi[2] // np[3]*m0
607 (p41) cmp.ltu p42,p40=a8,n8 }
608 { .mfi; (p43) cmp.leu p42,p40=a8,n8
609 (p16) xma.lu nlo[3]=ni3,mj[0],nhi[2]
611 { .mii; (p16) getf.sig n2=nlo[2] // 23:
612 (p49) cmp.ltu p50,p48=t[1],a8
613 (p51) cmp.leu p50,p48=t[1],a8 };;
614 { .mfi; (p16) nop.m 0 // 24:
615 (p16) xma.hu ahi[6]=ai6,bj[7],ahi[5] // ap[6]*b[i]
616 (p16) add a1=a1,n1 } // (p16) a1+=n1
617 { .mfi; (p16) nop.m 0
618 (p16) xma.lu alo[6]=ai6,bj[7],ahi[5]
619 (p17) mov t[0]=r0 };;
620 { .mii; (p16) getf.sig a5=alo[5] // 25:
621 (p16) add t0=t[7],a1 // (p16) t[7]+=a1
622 (p42) add t[0]=t[0],r0,1 };;
623 { .mfi; (p16) setf.sig tf[0]=t0 // 26:
624 (p16) xma.hu nhi[4]=ni4,mj[0],nhi[3] // np[4]*m0
625 (p50) add t[0]=t[0],r0,1 }
626 { .mfi; (p16) cmp.ltu.unc p42,p40=a1,n1
627 (p16) xma.lu nlo[4]=ni4,mj[0],nhi[3]
629 { .mii; (p16) getf.sig n3=nlo[3] // 27:
630 (p16) cmp.ltu.unc p50,p48=t0,a1
632 .pred.rel "mutex",p40,p42
633 .pred.rel "mutex",p48,p50
634 { .mfi; (p16) nop.m 0 // 28:
635 (p16) xma.hu ahi[7]=ai7,bj[7],ahi[6] // ap[7]*b[i]
636 (p40) add a2=a2,n2 } // (p16) a2+=n2
637 { .mfi; (p42) add a2=a2,n2,1
638 (p16) xma.lu alo[7]=ai7,bj[7],ahi[6]
640 { .mii; (p16) getf.sig a6=alo[6] // 29:
641 (p48) add t[6]=t[6],a2 // (p16) t[6]+=a2
642 (p50) add t[6]=t[6],a2,1 };;
643 { .mfi; (p16) nop.m 0 // 30:
644 (p16) xma.hu nhi[5]=ni5,mj[0],nhi[4] // np[5]*m0
645 (p40) cmp.ltu p41,p39=a2,n2 }
646 { .mfi; (p42) cmp.leu p41,p39=a2,n2
647 (p16) xma.lu nlo[5]=ni5,mj[0],nhi[4]
649 { .mfi; (p16) getf.sig n4=nlo[4] // 31:
651 (p48) cmp.ltu p49,p47=t[6],a2 }
652 { .mfb; (p50) cmp.leu p49,p47=t[6],a2
654 br.ctop.sptk.many .Louter_8_ctop };;
657 // above loop has to execute one more time, without (p16), which is
658 // replaced with merged move of np[8] to GPR bank
659 .pred.rel "mutex",p40,p42
660 .pred.rel "mutex",p48,p50
661 { .mmi; (p0) getf.sig n1=ni0 // 0:
662 (p40) add a3=a3,n3 // (p17) a3+=n3
663 (p42) add a3=a3,n3,1 };;
664 { .mii; (p17) getf.sig a7=alo[8] // 1:
665 (p48) add t[6]=t[6],a3 // (p17) t[6]+=a3
666 (p50) add t[6]=t[6],a3,1 };;
667 { .mfi; (p17) getf.sig a8=ahi[8] // 2:
668 (p17) xma.hu nhi[7]=ni6,mj[1],nhi[6] // np[6]*m0
669 (p40) cmp.ltu p43,p41=a3,n3 }
670 { .mfi; (p42) cmp.leu p43,p41=a3,n3
671 (p17) xma.lu nlo[7]=ni6,mj[1],nhi[6]
673 { .mii; (p17) getf.sig n5=nlo[6] // 3:
674 (p48) cmp.ltu p51,p49=t[6],a3
675 (p50) cmp.leu p51,p49=t[6],a3 };;
676 .pred.rel "mutex",p41,p43
677 .pred.rel "mutex",p49,p51
678 { .mmi; (p0) getf.sig n2=ni1 // 4:
679 (p41) add a4=a4,n4 // (p17) a4+=n4
680 (p43) add a4=a4,n4,1 };;
681 { .mfi; (p49) add t[5]=t[5],a4 // 5: (p17) t[5]+=a4
683 (p51) add t[5]=t[5],a4,1 };;
684 { .mfi; (p0) getf.sig n3=ni2 // 6:
685 (p17) xma.hu nhi[8]=ni7,mj[1],nhi[7] // np[7]*m0
686 (p41) cmp.ltu p42,p40=a4,n4 }
687 { .mfi; (p43) cmp.leu p42,p40=a4,n4
688 (p17) xma.lu nlo[8]=ni7,mj[1],nhi[7]
690 { .mii; (p17) getf.sig n6=nlo[7] // 7:
691 (p49) cmp.ltu p50,p48=t[5],a4
692 (p51) cmp.leu p50,p48=t[5],a4 };;
693 .pred.rel "mutex",p40,p42
694 .pred.rel "mutex",p48,p50
695 { .mii; (p0) getf.sig n4=ni3 // 8:
696 (p40) add a5=a5,n5 // (p17) a5+=n5
697 (p42) add a5=a5,n5,1 };;
698 { .mii; (p0) nop.m 0 // 9:
699 (p48) add t[4]=t[4],a5 // p(17) t[4]+=a5
700 (p50) add t[4]=t[4],a5,1 };;
701 { .mii; (p0) nop.m 0 // 10:
702 (p40) cmp.ltu p43,p41=a5,n5
703 (p42) cmp.leu p43,p41=a5,n5 };;
704 { .mii; (p17) getf.sig n7=nlo[8] // 11:
705 (p48) cmp.ltu p51,p49=t[4],a5
706 (p50) cmp.leu p51,p49=t[4],a5 };;
707 .pred.rel "mutex",p41,p43
708 .pred.rel "mutex",p49,p51
709 { .mii; (p17) getf.sig n8=nhi[8] // 12:
710 (p41) add a6=a6,n6 // (p17) a6+=n6
711 (p43) add a6=a6,n6,1 };;
712 { .mii; (p0) getf.sig n5=ni4 // 13:
713 (p49) add t[3]=t[3],a6 // (p17) t[3]+=a6
714 (p51) add t[3]=t[3],a6,1 };;
715 { .mii; (p0) nop.m 0 // 14:
716 (p41) cmp.ltu p42,p40=a6,n6
717 (p43) cmp.leu p42,p40=a6,n6 };;
718 { .mii; (p0) getf.sig n6=ni5 // 15:
719 (p49) cmp.ltu p50,p48=t[3],a6
720 (p51) cmp.leu p50,p48=t[3],a6 };;
721 .pred.rel "mutex",p40,p42
722 .pred.rel "mutex",p48,p50
723 { .mii; (p0) nop.m 0 // 16:
724 (p40) add a7=a7,n7 // (p17) a7+=n7
725 (p42) add a7=a7,n7,1 };;
726 { .mii; (p0) nop.m 0 // 17:
727 (p48) add t[2]=t[2],a7 // (p17) t[2]+=a7
728 (p50) add t[2]=t[2],a7,1 };;
729 { .mii; (p0) nop.m 0 // 18:
730 (p40) cmp.ltu p43,p41=a7,n7
731 (p42) cmp.leu p43,p41=a7,n7 };;
732 { .mii; (p0) getf.sig n7=ni6 // 19:
733 (p48) cmp.ltu p51,p49=t[2],a7
734 (p50) cmp.leu p51,p49=t[2],a7 };;
735 .pred.rel "mutex",p41,p43
736 .pred.rel "mutex",p49,p51
737 { .mii; (p0) nop.m 0 // 20:
738 (p41) add a8=a8,n8 // (p17) a8+=n8
739 (p43) add a8=a8,n8,1 };;
740 { .mmi; (p0) nop.m 0 // 21:
741 (p49) add t[1]=t[1],a8 // (p17) t[1]+=a8
742 (p51) add t[1]=t[1],a8,1 }
743 { .mmi; (p17) mov t[0]=r0
744 (p41) cmp.ltu p42,p40=a8,n8
745 (p43) cmp.leu p42,p40=a8,n8 };;
746 { .mmi; (p0) getf.sig n8=ni7 // 22:
747 (p49) cmp.ltu p50,p48=t[1],a8
748 (p51) cmp.leu p50,p48=t[1],a8 }
749 { .mmi; (p42) add t[0]=t[0],r0,1
750 (p0) add r16=-7*16,prevsp
751 (p0) add r17=-6*16,prevsp };;
753 // subtract np[8] from carrybit|tmp[8]
754 // carrybit|tmp[8] layout upon exit from above loop is:
755 // t[0]|t[1]|t[2]|t[3]|t[4]|t[5]|t[6]|t[7]|t0 (least significant)
756 { .mmi; (p50)add t[0]=t[0],r0,1
759 { .mmi; cmp.gtu p34,p32=n1,t0;;
760 .pred.rel "mutex",p32,p34
762 (p34)sub n2=t[7],n2,1 };;
763 { .mii; (p32)cmp.gtu p35,p33=n2,t[7]
764 (p34)cmp.geu p35,p33=n2,t[7];;
765 .pred.rel "mutex",p33,p35
766 (p33)sub n3=t[6],n3 }
767 { .mmi; (p35)sub n3=t[6],n3,1;;
768 (p33)cmp.gtu p34,p32=n3,t[6]
769 (p35)cmp.geu p34,p32=n3,t[6] };;
770 .pred.rel "mutex",p32,p34
771 { .mii; (p32)sub n4=t[5],n4
772 (p34)sub n4=t[5],n4,1;;
773 (p32)cmp.gtu p35,p33=n4,t[5] }
774 { .mmi; (p34)cmp.geu p35,p33=n4,t[5];;
775 .pred.rel "mutex",p33,p35
777 (p35)sub n5=t[4],n5,1 };;
778 { .mii; (p33)cmp.gtu p34,p32=n5,t[4]
779 (p35)cmp.geu p34,p32=n5,t[4];;
780 .pred.rel "mutex",p32,p34
781 (p32)sub n6=t[3],n6 }
782 { .mmi; (p34)sub n6=t[3],n6,1;;
783 (p32)cmp.gtu p35,p33=n6,t[3]
784 (p34)cmp.geu p35,p33=n6,t[3] };;
785 .pred.rel "mutex",p33,p35
786 { .mii; (p33)sub n7=t[2],n7
787 (p35)sub n7=t[2],n7,1;;
788 (p33)cmp.gtu p34,p32=n7,t[2] }
789 { .mmi; (p35)cmp.geu p34,p32=n7,t[2];;
790 .pred.rel "mutex",p32,p34
792 (p34)sub n8=t[1],n8,1 };;
793 { .mii; (p32)cmp.gtu p35,p33=n8,t[1]
794 (p34)cmp.geu p35,p33=n8,t[1];;
795 .pred.rel "mutex",p33,p35
796 (p33)sub a8=t[0],r0 }
797 { .mmi; (p35)sub a8=t[0],r0,1;;
798 (p33)cmp.gtu p34,p32=a8,t[0]
799 (p35)cmp.geu p34,p32=a8,t[0] };;
801 // save the result, either tmp[num] or tmp[num]-np[num]
802 .pred.rel "mutex",p32,p34
803 { .mmi; (p32)st8 [rptr]=n1,8
805 add r19=-4*16,prevsp};;
806 { .mmb; (p32)st8 [rptr]=n2,8
807 (p34)st8 [rptr]=t[7],8
808 (p5)br.cond.dpnt.few .Ldone };;
809 { .mmb; (p32)st8 [rptr]=n3,8
810 (p34)st8 [rptr]=t[6],8
811 (p7)br.cond.dpnt.few .Ldone };;
812 { .mmb; (p32)st8 [rptr]=n4,8
813 (p34)st8 [rptr]=t[5],8
814 (p9)br.cond.dpnt.few .Ldone };;
815 { .mmb; (p32)st8 [rptr]=n5,8
816 (p34)st8 [rptr]=t[4],8
817 (p11)br.cond.dpnt.few .Ldone };;
818 { .mmb; (p32)st8 [rptr]=n6,8
819 (p34)st8 [rptr]=t[3],8
820 (p13)br.cond.dpnt.few .Ldone };;
821 { .mmb; (p32)st8 [rptr]=n7,8
822 (p34)st8 [rptr]=t[2],8
823 (p15)br.cond.dpnt.few .Ldone };;
824 { .mmb; (p32)st8 [rptr]=n8,8
825 (p34)st8 [rptr]=t[1],8
828 { .mmi; ldf.fill f16=[r16],64
829 ldf.fill f17=[r17],64
831 { .mmi; ldf.fill f18=[r18],64
832 ldf.fill f19=[r19],64
833 mov pr=prevpr,0x1ffff };;
834 { .mmi; ldf.fill f20=[r16]
837 { .mmi; ldf.fill f22=[r18]
839 mov ret0=1 } // signal "handled"
843 br.ret.sptk.many b0 };;
846 .type copyright#,\@object
848 stringz "Montgomery multiplication for IA-64, CRYPTOGAMS by <appro\@openssl.org>"
851 open STDOUT,">$output" if $output;