1 .explicit 2 .text 3 .ident "ia64.S, Version 2.1" 4 .ident "IA-64 ISA artwork by Andy Polyakov <appro@openssl.org>" 5 6 // Copyright 2001-2018 The OpenSSL Project Authors. All Rights Reserved. 7 // 8 // Licensed under the Apache License 2.0 (the "License"). You may not use 9 // this file except in compliance with the License. You can obtain a copy 10 // in the file LICENSE in the source distribution or at 11 // https://www.openssl.org/source/license.html 12 13 // 14 // ==================================================================== 15 // Written by Andy Polyakov <appro@openssl.org> for the OpenSSL 16 // project. 17 // 18 // Rights for redistribution and usage in source and binary forms are 19 // granted according to the License. Warranty of any kind is disclaimed. 20 // ==================================================================== 21 // 22 // Version 2.x is Itanium2 re-tune. Few words about how Itanium2 is 23 // different from Itanium to this module viewpoint. Most notably, is it 24 // "wider" than Itanium? Can you experience loop scalability as 25 // discussed in commentary sections? Not really:-( Itanium2 has 6 26 // integer ALU ports, i.e. it's 2 ports wider, but it's not enough to 27 // spin twice as fast, as I need 8 IALU ports. Amount of floating point 28 // ports is the same, i.e. 2, while I need 4. In other words, to this 29 // module Itanium2 remains effectively as "wide" as Itanium. Yet it's 30 // essentially different in respect to this module, and a re-tune was 31 // required. Well, because some instruction latencies has changed. Most 32 // noticeably those intensively used: 33 // 34 // Itanium Itanium2 35 // ldf8 9 6 L2 hit 36 // ld8 2 1 L1 hit 37 // getf 2 5 38 // xma[->getf] 7[+1] 4[+0] 39 // add[->st8] 1[+1] 1[+0] 40 // 41 // What does it mean? You might ratiocinate that the original code 42 // should run just faster... Because sum of latencies is smaller... 43 // Wrong! Note that getf latency increased. This means that if a loop is 44 // scheduled for lower latency (as they were), then it will suffer from 45 // stall condition and the code will therefore turn anti-scalable, e.g. 46 // original bn_mul_words spun at 5*n or 2.5 times slower than expected 47 // on Itanium2! What to do? Reschedule loops for Itanium2? But then 48 // Itanium would exhibit anti-scalability. So I've chosen to reschedule 49 // for worst latency for every instruction aiming for best *all-round* 50 // performance. 51 52 // Q. How much faster does it get? 53 // A. Here is the output from 'openssl speed rsa dsa' for vanilla 54 // 0.9.6a compiled with gcc version 2.96 20000731 (Red Hat 55 // Linux 7.1 2.96-81): 56 // 57 // sign verify sign/s verify/s 58 // rsa 512 bits 0.0036s 0.0003s 275.3 2999.2 59 // rsa 1024 bits 0.0203s 0.0011s 49.3 894.1 60 // rsa 2048 bits 0.1331s 0.0040s 7.5 250.9 61 // rsa 4096 bits 0.9270s 0.0147s 1.1 68.1 62 // sign verify sign/s verify/s 63 // dsa 512 bits 0.0035s 0.0043s 288.3 234.8 64 // dsa 1024 bits 0.0111s 0.0135s 90.0 74.2 65 // 66 // And here is similar output but for this assembler 67 // implementation:-) 68 // 69 // sign verify sign/s verify/s 70 // rsa 512 bits 0.0021s 0.0001s 549.4 9638.5 71 // rsa 1024 bits 0.0055s 0.0002s 183.8 4481.1 72 // rsa 2048 bits 0.0244s 0.0006s 41.4 1726.3 73 // rsa 4096 bits 0.1295s 0.0018s 7.7 561.5 74 // sign verify sign/s verify/s 75 // dsa 512 bits 0.0012s 0.0013s 891.9 756.6 76 // dsa 1024 bits 0.0023s 0.0028s 440.4 376.2 77 // 78 // Yes, you may argue that it's not fair comparison as it's 79 // possible to craft the C implementation with BN_UMULT_HIGH 80 // inline assembler macro. But of course! Here is the output 81 // with the macro: 82 // 83 // sign verify sign/s verify/s 84 // rsa 512 bits 0.0020s 0.0002s 495.0 6561.0 85 // rsa 1024 bits 0.0086s 0.0004s 116.2 2235.7 86 // rsa 2048 bits 0.0519s 0.0015s 19.3 667.3 87 // rsa 4096 bits 0.3464s 0.0053s 2.9 187.7 88 // sign verify sign/s verify/s 89 // dsa 512 bits 0.0016s 0.0020s 613.1 510.5 90 // dsa 1024 bits 0.0045s 0.0054s 221.0 183.9 91 // 92 // My code is still way faster, huh:-) And I believe that even 93 // higher performance can be achieved. Note that as keys get 94 // longer, performance gain is larger. Why? According to the 95 // profiler there is another player in the field, namely 96 // BN_from_montgomery consuming larger and larger portion of CPU 97 // time as keysize decreases. I therefore consider putting effort 98 // to assembler implementation of the following routine: 99 // 100 // void bn_mul_add_mont (BN_ULONG *rp,BN_ULONG *np,int nl,BN_ULONG n0) 101 // { 102 // int i,j; 103 // BN_ULONG v; 104 // 105 // for (i=0; i<nl; i++) 106 // { 107 // v=bn_mul_add_words(rp,np,nl,(rp[0]*n0)&BN_MASK2); 108 // nrp++; 109 // rp++; 110 // if (((nrp[-1]+=v)&BN_MASK2) < v) 111 // for (j=0; ((++nrp[j])&BN_MASK2) == 0; j++) ; 112 // } 113 // } 114 // 115 // It might as well be beneficial to implement even combaX 116 // variants, as it appears as it can literally unleash the 117 // performance (see comment section to bn_mul_comba8 below). 118 // 119 // And finally for your reference the output for 0.9.6a compiled 120 // with SGIcc version 0.01.0-12 (keep in mind that for the moment 121 // of this writing it's not possible to convince SGIcc to use 122 // BN_UMULT_HIGH inline assembler macro, yet the code is fast, 123 // i.e. for a compiler generated one:-): 124 // 125 // sign verify sign/s verify/s 126 // rsa 512 bits 0.0022s 0.0002s 452.7 5894.3 127 // rsa 1024 bits 0.0097s 0.0005s 102.7 2002.9 128 // rsa 2048 bits 0.0578s 0.0017s 17.3 600.2 129 // rsa 4096 bits 0.3838s 0.0061s 2.6 164.5 130 // sign verify sign/s verify/s 131 // dsa 512 bits 0.0018s 0.0022s 547.3 459.6 132 // dsa 1024 bits 0.0051s 0.0062s 196.6 161.3 133 // 134 // Oh! Benchmarks were performed on 733MHz Lion-class Itanium 135 // system running Redhat Linux 7.1 (very special thanks to Ray 136 // McCaffity of Williams Communications for providing an account). 137 // 138 // Q. What's the heck with 'rum 1<<5' at the end of every function? 139 // A. Well, by clearing the "upper FP registers written" bit of the 140 // User Mask I want to excuse the kernel from preserving upper 141 // (f32-f128) FP register bank over process context switch, thus 142 // minimizing bus bandwidth consumption during the switch (i.e. 143 // after PKI operation completes and the program is off doing 144 // something else like bulk symmetric encryption). Having said 145 // this, I also want to point out that it might be good idea 146 // to compile the whole toolkit (as well as majority of the 147 // programs for that matter) with -mfixed-range=f32-f127 command 148 // line option. No, it doesn't prevent the compiler from writing 149 // to upper bank, but at least discourages to do so. If you don't 150 // like the idea you have the option to compile the module with 151 // -Drum=nop.m in command line. 152 // 153 154 #if defined(_HPUX_SOURCE) && !defined(_LP64) 155 #define ADDP addp4 156 #else 157 #define ADDP add 158 #endif 159 #ifdef __VMS 160 .alias abort, "decc$abort" 161 #endif 162 163 #if 1 164 // 165 // bn_[add|sub]_words routines. 166 // 167 // Loops are spinning in 2*(n+5) ticks on Itanium (provided that the 168 // data reside in L1 cache, i.e. 2 ticks away). It's possible to 169 // compress the epilogue and get down to 2*n+6, but at the cost of 170 // scalability (the neat feature of this implementation is that it 171 // shall automagically spin in n+5 on "wider" IA-64 implementations:-) 172 // I consider that the epilogue is short enough as it is to trade tiny 173 // performance loss on Itanium for scalability. 174 // 175 // BN_ULONG bn_add_words(BN_ULONG *rp, BN_ULONG *ap, BN_ULONG *bp,int num) 176 // 177 .global bn_add_words# 178 .proc bn_add_words# 179 .align 64 180 .skip 32 // makes the loop body aligned at 64-byte boundary 181 bn_add_words: 182 .prologue 183 .save ar.pfs,r2 184 { .mii; alloc r2=ar.pfs,4,12,0,16 185 cmp4.le p6,p0=r35,r0 };; 186 { .mfb; mov r8=r0 // return value 187 (p6) br.ret.spnt.many b0 };; 188 189 { .mib; sub r10=r35,r0,1 190 .save ar.lc,r3 191 mov r3=ar.lc 192 brp.loop.imp .L_bn_add_words_ctop,.L_bn_add_words_cend-16 193 } 194 { .mib; ADDP r14=0,r32 // rp 195 .save pr,r9 196 mov r9=pr };; 197 .body 198 { .mii; ADDP r15=0,r33 // ap 199 mov ar.lc=r10 200 mov ar.ec=6 } 201 { .mib; ADDP r16=0,r34 // bp 202 mov pr.rot=1<<16 };; 203 204 .L_bn_add_words_ctop: 205 { .mii; (p16) ld8 r32=[r16],8 // b=*(bp++) 206 (p18) add r39=r37,r34 207 (p19) cmp.ltu.unc p56,p0=r40,r38 } 208 { .mfb; (p0) nop.m 0x0 209 (p0) nop.f 0x0 210 (p0) nop.b 0x0 } 211 { .mii; (p16) ld8 r35=[r15],8 // a=*(ap++) 212 (p58) cmp.eq.or p57,p0=-1,r41 // (p20) 213 (p58) add r41=1,r41 } // (p20) 214 { .mfb; (p21) st8 [r14]=r42,8 // *(rp++)=r 215 (p0) nop.f 0x0 216 br.ctop.sptk .L_bn_add_words_ctop };; 217 .L_bn_add_words_cend: 218 219 { .mii; 220 (p59) add r8=1,r8 // return value 221 mov pr=r9,0x1ffff 222 mov ar.lc=r3 } 223 { .mbb; nop.b 0x0 224 br.ret.sptk.many b0 };; 225 .endp bn_add_words# 226 227 // 228 // BN_ULONG bn_sub_words(BN_ULONG *rp, BN_ULONG *ap, BN_ULONG *bp,int num) 229 // 230 .global bn_sub_words# 231 .proc bn_sub_words# 232 .align 64 233 .skip 32 // makes the loop body aligned at 64-byte boundary 234 bn_sub_words: 235 .prologue 236 .save ar.pfs,r2 237 { .mii; alloc r2=ar.pfs,4,12,0,16 238 cmp4.le p6,p0=r35,r0 };; 239 { .mfb; mov r8=r0 // return value 240 (p6) br.ret.spnt.many b0 };; 241 242 { .mib; sub r10=r35,r0,1 243 .save ar.lc,r3 244 mov r3=ar.lc 245 brp.loop.imp .L_bn_sub_words_ctop,.L_bn_sub_words_cend-16 246 } 247 { .mib; ADDP r14=0,r32 // rp 248 .save pr,r9 249 mov r9=pr };; 250 .body 251 { .mii; ADDP r15=0,r33 // ap 252 mov ar.lc=r10 253 mov ar.ec=6 } 254 { .mib; ADDP r16=0,r34 // bp 255 mov pr.rot=1<<16 };; 256 257 .L_bn_sub_words_ctop: 258 { .mii; (p16) ld8 r32=[r16],8 // b=*(bp++) 259 (p18) sub r39=r37,r34 260 (p19) cmp.gtu.unc p56,p0=r40,r38 } 261 { .mfb; (p0) nop.m 0x0 262 (p0) nop.f 0x0 263 (p0) nop.b 0x0 } 264 { .mii; (p16) ld8 r35=[r15],8 // a=*(ap++) 265 (p58) cmp.eq.or p57,p0=0,r41 // (p20) 266 (p58) add r41=-1,r41 } // (p20) 267 { .mbb; (p21) st8 [r14]=r42,8 // *(rp++)=r 268 (p0) nop.b 0x0 269 br.ctop.sptk .L_bn_sub_words_ctop };; 270 .L_bn_sub_words_cend: 271 272 { .mii; 273 (p59) add r8=1,r8 // return value 274 mov pr=r9,0x1ffff 275 mov ar.lc=r3 } 276 { .mbb; nop.b 0x0 277 br.ret.sptk.many b0 };; 278 .endp bn_sub_words# 279 #endif 280 281 #if 0 282 #define XMA_TEMPTATION 283 #endif 284 285 #if 1 286 // 287 // BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w) 288 // 289 .global bn_mul_words# 290 .proc bn_mul_words# 291 .align 64 292 .skip 32 // makes the loop body aligned at 64-byte boundary 293 bn_mul_words: 294 .prologue 295 .save ar.pfs,r2 296 #ifdef XMA_TEMPTATION 297 { .mfi; alloc r2=ar.pfs,4,0,0,0 };; 298 #else 299 { .mfi; alloc r2=ar.pfs,4,12,0,16 };; 300 #endif 301 { .mib; mov r8=r0 // return value 302 cmp4.le p6,p0=r34,r0 303 (p6) br.ret.spnt.many b0 };; 304 305 { .mii; sub r10=r34,r0,1 306 .save ar.lc,r3 307 mov r3=ar.lc 308 .save pr,r9 309 mov r9=pr };; 310 311 .body 312 { .mib; setf.sig f8=r35 // w 313 mov pr.rot=0x800001<<16 314 // ------^----- serves as (p50) at first (p27) 315 brp.loop.imp .L_bn_mul_words_ctop,.L_bn_mul_words_cend-16 316 } 317 318 #ifndef XMA_TEMPTATION 319 320 { .mmi; ADDP r14=0,r32 // rp 321 ADDP r15=0,r33 // ap 322 mov ar.lc=r10 } 323 { .mmi; mov r40=0 // serves as r35 at first (p27) 324 mov ar.ec=13 };; 325 326 // This loop spins in 2*(n+12) ticks. It's scheduled for data in Itanium 327 // L2 cache (i.e. 9 ticks away) as floating point load/store instructions 328 // bypass L1 cache and L2 latency is actually best-case scenario for 329 // ldf8. The loop is not scalable and shall run in 2*(n+12) even on 330 // "wider" IA-64 implementations. It's a trade-off here. n+24 loop 331 // would give us ~5% in *overall* performance improvement on "wider" 332 // IA-64, but would hurt Itanium for about same because of longer 333 // epilogue. As it's a matter of few percents in either case I've 334 // chosen to trade the scalability for development time (you can see 335 // this very instruction sequence in bn_mul_add_words loop which in 336 // turn is scalable). 337 .L_bn_mul_words_ctop: 338 { .mfi; (p25) getf.sig r36=f52 // low 339 (p21) xmpy.lu f48=f37,f8 340 (p28) cmp.ltu p54,p50=r41,r39 } 341 { .mfi; (p16) ldf8 f32=[r15],8 342 (p21) xmpy.hu f40=f37,f8 343 (p0) nop.i 0x0 };; 344 { .mii; (p25) getf.sig r32=f44 // high 345 .pred.rel "mutex",p50,p54 346 (p50) add r40=r38,r35 // (p27) 347 (p54) add r40=r38,r35,1 } // (p27) 348 { .mfb; (p28) st8 [r14]=r41,8 349 (p0) nop.f 0x0 350 br.ctop.sptk .L_bn_mul_words_ctop };; 351 .L_bn_mul_words_cend: 352 353 { .mii; nop.m 0x0 354 .pred.rel "mutex",p51,p55 355 (p51) add r8=r36,r0 356 (p55) add r8=r36,r0,1 } 357 { .mfb; nop.m 0x0 358 nop.f 0x0 359 nop.b 0x0 } 360 361 #else // XMA_TEMPTATION 362 363 setf.sig f37=r0 // serves as carry at (p18) tick 364 mov ar.lc=r10 365 mov ar.ec=5;; 366 367 // Most of you examining this code very likely wonder why in the name 368 // of Intel the following loop is commented out? Indeed, it looks so 369 // neat that you find it hard to believe that it's something wrong 370 // with it, right? The catch is that every iteration depends on the 371 // result from previous one and the latter isn't available instantly. 372 // The loop therefore spins at the latency of xma minus 1, or in other 373 // words at 6*(n+4) ticks:-( Compare to the "production" loop above 374 // that runs in 2*(n+11) where the low latency problem is worked around 375 // by moving the dependency to one-tick latent integer ALU. Note that 376 // "distance" between ldf8 and xma is not latency of ldf8, but the 377 // *difference* between xma and ldf8 latencies. 378 .L_bn_mul_words_ctop: 379 { .mfi; (p16) ldf8 f32=[r33],8 380 (p18) xma.hu f38=f34,f8,f39 } 381 { .mfb; (p20) stf8 [r32]=f37,8 382 (p18) xma.lu f35=f34,f8,f39 383 br.ctop.sptk .L_bn_mul_words_ctop };; 384 .L_bn_mul_words_cend: 385 386 getf.sig r8=f41 // the return value 387 388 #endif // XMA_TEMPTATION 389 390 { .mii; nop.m 0x0 391 mov pr=r9,0x1ffff 392 mov ar.lc=r3 } 393 { .mfb; rum 1<<5 // clear um.mfh 394 nop.f 0x0 395 br.ret.sptk.many b0 };; 396 .endp bn_mul_words# 397 #endif 398 399 #if 1 400 // 401 // BN_ULONG bn_mul_add_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w) 402 // 403 .global bn_mul_add_words# 404 .proc bn_mul_add_words# 405 .align 64 406 .skip 48 // makes the loop body aligned at 64-byte boundary 407 bn_mul_add_words: 408 .prologue 409 .save ar.pfs,r2 410 { .mmi; alloc r2=ar.pfs,4,4,0,8 411 cmp4.le p6,p0=r34,r0 412 .save ar.lc,r3 413 mov r3=ar.lc };; 414 { .mib; mov r8=r0 // return value 415 sub r10=r34,r0,1 416 (p6) br.ret.spnt.many b0 };; 417 418 { .mib; setf.sig f8=r35 // w 419 .save pr,r9 420 mov r9=pr 421 brp.loop.imp .L_bn_mul_add_words_ctop,.L_bn_mul_add_words_cend-16 422 } 423 .body 424 { .mmi; ADDP r14=0,r32 // rp 425 ADDP r15=0,r33 // ap 426 mov ar.lc=r10 } 427 { .mii; ADDP r16=0,r32 // rp copy 428 mov pr.rot=0x2001<<16 429 // ------^----- serves as (p40) at first (p27) 430 mov ar.ec=11 };; 431 432 // This loop spins in 3*(n+10) ticks on Itanium and in 2*(n+10) on 433 // Itanium 2. Yes, unlike previous versions it scales:-) Previous 434 // version was performing *all* additions in IALU and was starving 435 // for those even on Itanium 2. In this version one addition is 436 // moved to FPU and is folded with multiplication. This is at cost 437 // of propagating the result from previous call to this subroutine 438 // to L2 cache... In other words negligible even for shorter keys. 439 // *Overall* performance improvement [over previous version] varies 440 // from 11 to 22 percent depending on key length. 441 .L_bn_mul_add_words_ctop: 442 .pred.rel "mutex",p40,p42 443 { .mfi; (p23) getf.sig r36=f45 // low 444 (p20) xma.lu f42=f36,f8,f50 // low 445 (p40) add r39=r39,r35 } // (p27) 446 { .mfi; (p16) ldf8 f32=[r15],8 // *(ap++) 447 (p20) xma.hu f36=f36,f8,f50 // high 448 (p42) add r39=r39,r35,1 };; // (p27) 449 { .mmi; (p24) getf.sig r32=f40 // high 450 (p16) ldf8 f46=[r16],8 // *(rp1++) 451 (p40) cmp.ltu p41,p39=r39,r35 } // (p27) 452 { .mib; (p26) st8 [r14]=r39,8 // *(rp2++) 453 (p42) cmp.leu p41,p39=r39,r35 // (p27) 454 br.ctop.sptk .L_bn_mul_add_words_ctop};; 455 .L_bn_mul_add_words_cend: 456 457 { .mmi; .pred.rel "mutex",p40,p42 458 (p40) add r8=r35,r0 459 (p42) add r8=r35,r0,1 460 mov pr=r9,0x1ffff } 461 { .mib; rum 1<<5 // clear um.mfh 462 mov ar.lc=r3 463 br.ret.sptk.many b0 };; 464 .endp bn_mul_add_words# 465 #endif 466 467 #if 1 468 // 469 // void bn_sqr_words(BN_ULONG *rp, BN_ULONG *ap, int num) 470 // 471 .global bn_sqr_words# 472 .proc bn_sqr_words# 473 .align 64 474 .skip 32 // makes the loop body aligned at 64-byte boundary 475 bn_sqr_words: 476 .prologue 477 .save ar.pfs,r2 478 { .mii; alloc r2=ar.pfs,3,0,0,0 479 sxt4 r34=r34 };; 480 { .mii; cmp.le p6,p0=r34,r0 481 mov r8=r0 } // return value 482 { .mfb; ADDP r32=0,r32 483 nop.f 0x0 484 (p6) br.ret.spnt.many b0 };; 485 486 { .mii; sub r10=r34,r0,1 487 .save ar.lc,r3 488 mov r3=ar.lc 489 .save pr,r9 490 mov r9=pr };; 491 492 .body 493 { .mib; ADDP r33=0,r33 494 mov pr.rot=1<<16 495 brp.loop.imp .L_bn_sqr_words_ctop,.L_bn_sqr_words_cend-16 496 } 497 { .mii; add r34=8,r32 498 mov ar.lc=r10 499 mov ar.ec=18 };; 500 501 // 2*(n+17) on Itanium, (n+17) on "wider" IA-64 implementations. It's 502 // possible to compress the epilogue (I'm getting tired to write this 503 // comment over and over) and get down to 2*n+16 at the cost of 504 // scalability. The decision will very likely be reconsidered after the 505 // benchmark program is profiled. I.e. if performance gain on Itanium 506 // will appear larger than loss on "wider" IA-64, then the loop should 507 // be explicitly split and the epilogue compressed. 508 .L_bn_sqr_words_ctop: 509 { .mfi; (p16) ldf8 f32=[r33],8 510 (p25) xmpy.lu f42=f41,f41 511 (p0) nop.i 0x0 } 512 { .mib; (p33) stf8 [r32]=f50,16 513 (p0) nop.i 0x0 514 (p0) nop.b 0x0 } 515 { .mfi; (p0) nop.m 0x0 516 (p25) xmpy.hu f52=f41,f41 517 (p0) nop.i 0x0 } 518 { .mib; (p33) stf8 [r34]=f60,16 519 (p0) nop.i 0x0 520 br.ctop.sptk .L_bn_sqr_words_ctop };; 521 .L_bn_sqr_words_cend: 522 523 { .mii; nop.m 0x0 524 mov pr=r9,0x1ffff 525 mov ar.lc=r3 } 526 { .mfb; rum 1<<5 // clear um.mfh 527 nop.f 0x0 528 br.ret.sptk.many b0 };; 529 .endp bn_sqr_words# 530 #endif 531 532 #if 1 533 // Apparently we win nothing by implementing special bn_sqr_comba8. 534 // Yes, it is possible to reduce the number of multiplications by 535 // almost factor of two, but then the amount of additions would 536 // increase by factor of two (as we would have to perform those 537 // otherwise performed by xma ourselves). Normally we would trade 538 // anyway as multiplications are way more expensive, but not this 539 // time... Multiplication kernel is fully pipelined and as we drain 540 // one 128-bit multiplication result per clock cycle multiplications 541 // are effectively as inexpensive as additions. Special implementation 542 // might become of interest for "wider" IA-64 implementation as you'll 543 // be able to get through the multiplication phase faster (there won't 544 // be any stall issues as discussed in the commentary section below and 545 // you therefore will be able to employ all 4 FP units)... But these 546 // Itanium days it's simply too hard to justify the effort so I just 547 // drop down to bn_mul_comba8 code:-) 548 // 549 // void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a) 550 // 551 .global bn_sqr_comba8# 552 .proc bn_sqr_comba8# 553 .align 64 554 bn_sqr_comba8: 555 .prologue 556 .save ar.pfs,r2 557 #if defined(_HPUX_SOURCE) && !defined(_LP64) 558 { .mii; alloc r2=ar.pfs,2,1,0,0 559 addp4 r33=0,r33 560 addp4 r32=0,r32 };; 561 { .mii; 562 #else 563 { .mii; alloc r2=ar.pfs,2,1,0,0 564 #endif 565 mov r34=r33 566 add r14=8,r33 };; 567 .body 568 { .mii; add r17=8,r34 569 add r15=16,r33 570 add r18=16,r34 } 571 { .mfb; add r16=24,r33 572 br .L_cheat_entry_point8 };; 573 .endp bn_sqr_comba8# 574 #endif 575 576 #if 1 577 // I've estimated this routine to run in ~120 ticks, but in reality 578 // (i.e. according to ar.itc) it takes ~160 ticks. Are those extra 579 // cycles consumed for instructions fetch? Or did I misinterpret some 580 // clause in Itanium µ-architecture manual? Comments are welcomed and 581 // highly appreciated. 582 // 583 // On Itanium 2 it takes ~190 ticks. This is because of stalls on 584 // result from getf.sig. I do nothing about it at this point for 585 // reasons depicted below. 586 // 587 // However! It should be noted that even 160 ticks is darn good result 588 // as it's over 10 (yes, ten, spelled as t-e-n) times faster than the 589 // C version (compiled with gcc with inline assembler). I really 590 // kicked compiler's butt here, didn't I? Yeah! This brings us to the 591 // following statement. It's damn shame that this routine isn't called 592 // very often nowadays! According to the profiler most CPU time is 593 // consumed by bn_mul_add_words called from BN_from_montgomery. In 594 // order to estimate what we're missing, I've compared the performance 595 // of this routine against "traditional" implementation, i.e. against 596 // following routine: 597 // 598 // void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) 599 // { r[ 8]=bn_mul_words( &(r[0]),a,8,b[0]); 600 // r[ 9]=bn_mul_add_words(&(r[1]),a,8,b[1]); 601 // r[10]=bn_mul_add_words(&(r[2]),a,8,b[2]); 602 // r[11]=bn_mul_add_words(&(r[3]),a,8,b[3]); 603 // r[12]=bn_mul_add_words(&(r[4]),a,8,b[4]); 604 // r[13]=bn_mul_add_words(&(r[5]),a,8,b[5]); 605 // r[14]=bn_mul_add_words(&(r[6]),a,8,b[6]); 606 // r[15]=bn_mul_add_words(&(r[7]),a,8,b[7]); 607 // } 608 // 609 // The one below is over 8 times faster than the one above:-( Even 610 // more reasons to "combafy" bn_mul_add_mont... 611 // 612 // And yes, this routine really made me wish there were an optimizing 613 // assembler! It also feels like it deserves a dedication. 614 // 615 // To my wife for being there and to my kids... 616 // 617 // void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) 618 // 619 #define carry1 r14 620 #define carry2 r15 621 #define carry3 r34 622 .global bn_mul_comba8# 623 .proc bn_mul_comba8# 624 .align 64 625 bn_mul_comba8: 626 .prologue 627 .save ar.pfs,r2 628 #if defined(_HPUX_SOURCE) && !defined(_LP64) 629 { .mii; alloc r2=ar.pfs,3,0,0,0 630 addp4 r33=0,r33 631 addp4 r34=0,r34 };; 632 { .mii; addp4 r32=0,r32 633 #else 634 { .mii; alloc r2=ar.pfs,3,0,0,0 635 #endif 636 add r14=8,r33 637 add r17=8,r34 } 638 .body 639 { .mii; add r15=16,r33 640 add r18=16,r34 641 add r16=24,r33 } 642 .L_cheat_entry_point8: 643 { .mmi; add r19=24,r34 644 645 ldf8 f32=[r33],32 };; 646 647 { .mmi; ldf8 f120=[r34],32 648 ldf8 f121=[r17],32 } 649 { .mmi; ldf8 f122=[r18],32 650 ldf8 f123=[r19],32 };; 651 { .mmi; ldf8 f124=[r34] 652 ldf8 f125=[r17] } 653 { .mmi; ldf8 f126=[r18] 654 ldf8 f127=[r19] } 655 656 { .mmi; ldf8 f33=[r14],32 657 ldf8 f34=[r15],32 } 658 { .mmi; ldf8 f35=[r16],32;; 659 ldf8 f36=[r33] } 660 { .mmi; ldf8 f37=[r14] 661 ldf8 f38=[r15] } 662 { .mfi; ldf8 f39=[r16] 663 // -------\ Entering multiplier's heaven /------- 664 // ------------\ /------------ 665 // -----------------\ /----------------- 666 // ----------------------\/---------------------- 667 xma.hu f41=f32,f120,f0 } 668 { .mfi; xma.lu f40=f32,f120,f0 };; // (*) 669 { .mfi; xma.hu f51=f32,f121,f0 } 670 { .mfi; xma.lu f50=f32,f121,f0 };; 671 { .mfi; xma.hu f61=f32,f122,f0 } 672 { .mfi; xma.lu f60=f32,f122,f0 };; 673 { .mfi; xma.hu f71=f32,f123,f0 } 674 { .mfi; xma.lu f70=f32,f123,f0 };; 675 { .mfi; xma.hu f81=f32,f124,f0 } 676 { .mfi; xma.lu f80=f32,f124,f0 };; 677 { .mfi; xma.hu f91=f32,f125,f0 } 678 { .mfi; xma.lu f90=f32,f125,f0 };; 679 { .mfi; xma.hu f101=f32,f126,f0 } 680 { .mfi; xma.lu f100=f32,f126,f0 };; 681 { .mfi; xma.hu f111=f32,f127,f0 } 682 { .mfi; xma.lu f110=f32,f127,f0 };;// 683 // (*) You can argue that splitting at every second bundle would 684 // prevent "wider" IA-64 implementations from achieving the peak 685 // performance. Well, not really... The catch is that if you 686 // intend to keep 4 FP units busy by splitting at every fourth 687 // bundle and thus perform these 16 multiplications in 4 ticks, 688 // the first bundle *below* would stall because the result from 689 // the first xma bundle *above* won't be available for another 3 690 // ticks (if not more, being an optimist, I assume that "wider" 691 // implementation will have same latency:-). This stall will hold 692 // you back and the performance would be as if every second bundle 693 // were split *anyway*... 694 { .mfi; getf.sig r16=f40 695 xma.hu f42=f33,f120,f41 696 add r33=8,r32 } 697 { .mfi; xma.lu f41=f33,f120,f41 };; 698 { .mfi; getf.sig r24=f50 699 xma.hu f52=f33,f121,f51 } 700 { .mfi; xma.lu f51=f33,f121,f51 };; 701 { .mfi; st8 [r32]=r16,16 702 xma.hu f62=f33,f122,f61 } 703 { .mfi; xma.lu f61=f33,f122,f61 };; 704 { .mfi; xma.hu f72=f33,f123,f71 } 705 { .mfi; xma.lu f71=f33,f123,f71 };; 706 { .mfi; xma.hu f82=f33,f124,f81 } 707 { .mfi; xma.lu f81=f33,f124,f81 };; 708 { .mfi; xma.hu f92=f33,f125,f91 } 709 { .mfi; xma.lu f91=f33,f125,f91 };; 710 { .mfi; xma.hu f102=f33,f126,f101 } 711 { .mfi; xma.lu f101=f33,f126,f101 };; 712 { .mfi; xma.hu f112=f33,f127,f111 } 713 { .mfi; xma.lu f111=f33,f127,f111 };;// 714 //-------------------------------------------------// 715 { .mfi; getf.sig r25=f41 716 xma.hu f43=f34,f120,f42 } 717 { .mfi; xma.lu f42=f34,f120,f42 };; 718 { .mfi; getf.sig r16=f60 719 xma.hu f53=f34,f121,f52 } 720 { .mfi; xma.lu f52=f34,f121,f52 };; 721 { .mfi; getf.sig r17=f51 722 xma.hu f63=f34,f122,f62 723 add r25=r25,r24 } 724 { .mfi; xma.lu f62=f34,f122,f62 725 mov carry1=0 };; 726 { .mfi; cmp.ltu p6,p0=r25,r24 727 xma.hu f73=f34,f123,f72 } 728 { .mfi; xma.lu f72=f34,f123,f72 };; 729 { .mfi; st8 [r33]=r25,16 730 xma.hu f83=f34,f124,f82 731 (p6) add carry1=1,carry1 } 732 { .mfi; xma.lu f82=f34,f124,f82 };; 733 { .mfi; xma.hu f93=f34,f125,f92 } 734 { .mfi; xma.lu f92=f34,f125,f92 };; 735 { .mfi; xma.hu f103=f34,f126,f102 } 736 { .mfi; xma.lu f102=f34,f126,f102 };; 737 { .mfi; xma.hu f113=f34,f127,f112 } 738 { .mfi; xma.lu f112=f34,f127,f112 };;// 739 //-------------------------------------------------// 740 { .mfi; getf.sig r18=f42 741 xma.hu f44=f35,f120,f43 742 add r17=r17,r16 } 743 { .mfi; xma.lu f43=f35,f120,f43 };; 744 { .mfi; getf.sig r24=f70 745 xma.hu f54=f35,f121,f53 } 746 { .mfi; mov carry2=0 747 xma.lu f53=f35,f121,f53 };; 748 { .mfi; getf.sig r25=f61 749 xma.hu f64=f35,f122,f63 750 cmp.ltu p7,p0=r17,r16 } 751 { .mfi; add r18=r18,r17 752 xma.lu f63=f35,f122,f63 };; 753 { .mfi; getf.sig r26=f52 754 xma.hu f74=f35,f123,f73 755 (p7) add carry2=1,carry2 } 756 { .mfi; cmp.ltu p7,p0=r18,r17 757 xma.lu f73=f35,f123,f73 758 add r18=r18,carry1 };; 759 { .mfi; 760 xma.hu f84=f35,f124,f83 761 (p7) add carry2=1,carry2 } 762 { .mfi; cmp.ltu p7,p0=r18,carry1 763 xma.lu f83=f35,f124,f83 };; 764 { .mfi; st8 [r32]=r18,16 765 xma.hu f94=f35,f125,f93 766 (p7) add carry2=1,carry2 } 767 { .mfi; xma.lu f93=f35,f125,f93 };; 768 { .mfi; xma.hu f104=f35,f126,f103 } 769 { .mfi; xma.lu f103=f35,f126,f103 };; 770 { .mfi; xma.hu f114=f35,f127,f113 } 771 { .mfi; mov carry1=0 772 xma.lu f113=f35,f127,f113 773 add r25=r25,r24 };;// 774 //-------------------------------------------------// 775 { .mfi; getf.sig r27=f43 776 xma.hu f45=f36,f120,f44 777 cmp.ltu p6,p0=r25,r24 } 778 { .mfi; xma.lu f44=f36,f120,f44 779 add r26=r26,r25 };; 780 { .mfi; getf.sig r16=f80 781 xma.hu f55=f36,f121,f54 782 (p6) add carry1=1,carry1 } 783 { .mfi; xma.lu f54=f36,f121,f54 };; 784 { .mfi; getf.sig r17=f71 785 xma.hu f65=f36,f122,f64 786 cmp.ltu p6,p0=r26,r25 } 787 { .mfi; xma.lu f64=f36,f122,f64 788 add r27=r27,r26 };; 789 { .mfi; getf.sig r18=f62 790 xma.hu f75=f36,f123,f74 791 (p6) add carry1=1,carry1 } 792 { .mfi; cmp.ltu p6,p0=r27,r26 793 xma.lu f74=f36,f123,f74 794 add r27=r27,carry2 };; 795 { .mfi; getf.sig r19=f53 796 xma.hu f85=f36,f124,f84 797 (p6) add carry1=1,carry1 } 798 { .mfi; xma.lu f84=f36,f124,f84 799 cmp.ltu p6,p0=r27,carry2 };; 800 { .mfi; st8 [r33]=r27,16 801 xma.hu f95=f36,f125,f94 802 (p6) add carry1=1,carry1 } 803 { .mfi; xma.lu f94=f36,f125,f94 };; 804 { .mfi; xma.hu f105=f36,f126,f104 } 805 { .mfi; mov carry2=0 806 xma.lu f104=f36,f126,f104 807 add r17=r17,r16 };; 808 { .mfi; xma.hu f115=f36,f127,f114 809 cmp.ltu p7,p0=r17,r16 } 810 { .mfi; xma.lu f114=f36,f127,f114 811 add r18=r18,r17 };;// 812 //-------------------------------------------------// 813 { .mfi; getf.sig r20=f44 814 xma.hu f46=f37,f120,f45 815 (p7) add carry2=1,carry2 } 816 { .mfi; cmp.ltu p7,p0=r18,r17 817 xma.lu f45=f37,f120,f45 818 add r19=r19,r18 };; 819 { .mfi; getf.sig r24=f90 820 xma.hu f56=f37,f121,f55 } 821 { .mfi; xma.lu f55=f37,f121,f55 };; 822 { .mfi; getf.sig r25=f81 823 xma.hu f66=f37,f122,f65 824 (p7) add carry2=1,carry2 } 825 { .mfi; cmp.ltu p7,p0=r19,r18 826 xma.lu f65=f37,f122,f65 827 add r20=r20,r19 };; 828 { .mfi; getf.sig r26=f72 829 xma.hu f76=f37,f123,f75 830 (p7) add carry2=1,carry2 } 831 { .mfi; cmp.ltu p7,p0=r20,r19 832 xma.lu f75=f37,f123,f75 833 add r20=r20,carry1 };; 834 { .mfi; getf.sig r27=f63 835 xma.hu f86=f37,f124,f85 836 (p7) add carry2=1,carry2 } 837 { .mfi; xma.lu f85=f37,f124,f85 838 cmp.ltu p7,p0=r20,carry1 };; 839 { .mfi; getf.sig r28=f54 840 xma.hu f96=f37,f125,f95 841 (p7) add carry2=1,carry2 } 842 { .mfi; st8 [r32]=r20,16 843 xma.lu f95=f37,f125,f95 };; 844 { .mfi; xma.hu f106=f37,f126,f105 } 845 { .mfi; mov carry1=0 846 xma.lu f105=f37,f126,f105 847 add r25=r25,r24 };; 848 { .mfi; xma.hu f116=f37,f127,f115 849 cmp.ltu p6,p0=r25,r24 } 850 { .mfi; xma.lu f115=f37,f127,f115 851 add r26=r26,r25 };;// 852 //-------------------------------------------------// 853 { .mfi; getf.sig r29=f45 854 xma.hu f47=f38,f120,f46 855 (p6) add carry1=1,carry1 } 856 { .mfi; cmp.ltu p6,p0=r26,r25 857 xma.lu f46=f38,f120,f46 858 add r27=r27,r26 };; 859 { .mfi; getf.sig r16=f100 860 xma.hu f57=f38,f121,f56 861 (p6) add carry1=1,carry1 } 862 { .mfi; cmp.ltu p6,p0=r27,r26 863 xma.lu f56=f38,f121,f56 864 add r28=r28,r27 };; 865 { .mfi; getf.sig r17=f91 866 xma.hu f67=f38,f122,f66 867 (p6) add carry1=1,carry1 } 868 { .mfi; cmp.ltu p6,p0=r28,r27 869 xma.lu f66=f38,f122,f66 870 add r29=r29,r28 };; 871 { .mfi; getf.sig r18=f82 872 xma.hu f77=f38,f123,f76 873 (p6) add carry1=1,carry1 } 874 { .mfi; cmp.ltu p6,p0=r29,r28 875 xma.lu f76=f38,f123,f76 876 add r29=r29,carry2 };; 877 { .mfi; getf.sig r19=f73 878 xma.hu f87=f38,f124,f86 879 (p6) add carry1=1,carry1 } 880 { .mfi; xma.lu f86=f38,f124,f86 881 cmp.ltu p6,p0=r29,carry2 };; 882 { .mfi; getf.sig r20=f64 883 xma.hu f97=f38,f125,f96 884 (p6) add carry1=1,carry1 } 885 { .mfi; st8 [r33]=r29,16 886 xma.lu f96=f38,f125,f96 };; 887 { .mfi; getf.sig r21=f55 888 xma.hu f107=f38,f126,f106 } 889 { .mfi; mov carry2=0 890 xma.lu f106=f38,f126,f106 891 add r17=r17,r16 };; 892 { .mfi; xma.hu f117=f38,f127,f116 893 cmp.ltu p7,p0=r17,r16 } 894 { .mfi; xma.lu f116=f38,f127,f116 895 add r18=r18,r17 };;// 896 //-------------------------------------------------// 897 { .mfi; getf.sig r22=f46 898 xma.hu f48=f39,f120,f47 899 (p7) add carry2=1,carry2 } 900 { .mfi; cmp.ltu p7,p0=r18,r17 901 xma.lu f47=f39,f120,f47 902 add r19=r19,r18 };; 903 { .mfi; getf.sig r24=f110 904 xma.hu f58=f39,f121,f57 905 (p7) add carry2=1,carry2 } 906 { .mfi; cmp.ltu p7,p0=r19,r18 907 xma.lu f57=f39,f121,f57 908 add r20=r20,r19 };; 909 { .mfi; getf.sig r25=f101 910 xma.hu f68=f39,f122,f67 911 (p7) add carry2=1,carry2 } 912 { .mfi; cmp.ltu p7,p0=r20,r19 913 xma.lu f67=f39,f122,f67 914 add r21=r21,r20 };; 915 { .mfi; getf.sig r26=f92 916 xma.hu f78=f39,f123,f77 917 (p7) add carry2=1,carry2 } 918 { .mfi; cmp.ltu p7,p0=r21,r20 919 xma.lu f77=f39,f123,f77 920 add r22=r22,r21 };; 921 { .mfi; getf.sig r27=f83 922 xma.hu f88=f39,f124,f87 923 (p7) add carry2=1,carry2 } 924 { .mfi; cmp.ltu p7,p0=r22,r21 925 xma.lu f87=f39,f124,f87 926 add r22=r22,carry1 };; 927 { .mfi; getf.sig r28=f74 928 xma.hu f98=f39,f125,f97 929 (p7) add carry2=1,carry2 } 930 { .mfi; xma.lu f97=f39,f125,f97 931 cmp.ltu p7,p0=r22,carry1 };; 932 { .mfi; getf.sig r29=f65 933 xma.hu f108=f39,f126,f107 934 (p7) add carry2=1,carry2 } 935 { .mfi; st8 [r32]=r22,16 936 xma.lu f107=f39,f126,f107 };; 937 { .mfi; getf.sig r30=f56 938 xma.hu f118=f39,f127,f117 } 939 { .mfi; xma.lu f117=f39,f127,f117 };;// 940 //-------------------------------------------------// 941 // Leaving multiplier's heaven... Quite a ride, huh? 942 943 { .mii; getf.sig r31=f47 944 add r25=r25,r24 945 mov carry1=0 };; 946 { .mii; getf.sig r16=f111 947 cmp.ltu p6,p0=r25,r24 948 add r26=r26,r25 };; 949 { .mfb; getf.sig r17=f102 } 950 { .mii; 951 (p6) add carry1=1,carry1 952 cmp.ltu p6,p0=r26,r25 953 add r27=r27,r26 };; 954 { .mfb; nop.m 0x0 } 955 { .mii; 956 (p6) add carry1=1,carry1 957 cmp.ltu p6,p0=r27,r26 958 add r28=r28,r27 };; 959 { .mii; getf.sig r18=f93 960 add r17=r17,r16 961 mov carry3=0 } 962 { .mii; 963 (p6) add carry1=1,carry1 964 cmp.ltu p6,p0=r28,r27 965 add r29=r29,r28 };; 966 { .mii; getf.sig r19=f84 967 cmp.ltu p7,p0=r17,r16 } 968 { .mii; 969 (p6) add carry1=1,carry1 970 cmp.ltu p6,p0=r29,r28 971 add r30=r30,r29 };; 972 { .mii; getf.sig r20=f75 973 add r18=r18,r17 } 974 { .mii; 975 (p6) add carry1=1,carry1 976 cmp.ltu p6,p0=r30,r29 977 add r31=r31,r30 };; 978 { .mfb; getf.sig r21=f66 } 979 { .mii; (p7) add carry3=1,carry3 980 cmp.ltu p7,p0=r18,r17 981 add r19=r19,r18 } 982 { .mfb; nop.m 0x0 } 983 { .mii; 984 (p6) add carry1=1,carry1 985 cmp.ltu p6,p0=r31,r30 986 add r31=r31,carry2 };; 987 { .mfb; getf.sig r22=f57 } 988 { .mii; (p7) add carry3=1,carry3 989 cmp.ltu p7,p0=r19,r18 990 add r20=r20,r19 } 991 { .mfb; nop.m 0x0 } 992 { .mii; 993 (p6) add carry1=1,carry1 994 cmp.ltu p6,p0=r31,carry2 };; 995 { .mfb; getf.sig r23=f48 } 996 { .mii; (p7) add carry3=1,carry3 997 cmp.ltu p7,p0=r20,r19 998 add r21=r21,r20 } 999 { .mii; 1000 (p6) add carry1=1,carry1 } 1001 { .mfb; st8 [r33]=r31,16 };; 1002 1003 { .mfb; getf.sig r24=f112 } 1004 { .mii; (p7) add carry3=1,carry3 1005 cmp.ltu p7,p0=r21,r20 1006 add r22=r22,r21 };; 1007 { .mfb; getf.sig r25=f103 } 1008 { .mii; (p7) add carry3=1,carry3 1009 cmp.ltu p7,p0=r22,r21 1010 add r23=r23,r22 };; 1011 { .mfb; getf.sig r26=f94 } 1012 { .mii; (p7) add carry3=1,carry3 1013 cmp.ltu p7,p0=r23,r22 1014 add r23=r23,carry1 };; 1015 { .mfb; getf.sig r27=f85 } 1016 { .mii; (p7) add carry3=1,carry3 1017 cmp.ltu p7,p8=r23,carry1};; 1018 { .mii; getf.sig r28=f76 1019 add r25=r25,r24 1020 mov carry1=0 } 1021 { .mii; st8 [r32]=r23,16 1022 (p7) add carry2=1,carry3 1023 (p8) add carry2=0,carry3 };; 1024 1025 { .mfb; nop.m 0x0 } 1026 { .mii; getf.sig r29=f67 1027 cmp.ltu p6,p0=r25,r24 1028 add r26=r26,r25 };; 1029 { .mfb; getf.sig r30=f58 } 1030 { .mii; 1031 (p6) add carry1=1,carry1 1032 cmp.ltu p6,p0=r26,r25 1033 add r27=r27,r26 };; 1034 { .mfb; getf.sig r16=f113 } 1035 { .mii; 1036 (p6) add carry1=1,carry1 1037 cmp.ltu p6,p0=r27,r26 1038 add r28=r28,r27 };; 1039 { .mfb; getf.sig r17=f104 } 1040 { .mii; 1041 (p6) add carry1=1,carry1 1042 cmp.ltu p6,p0=r28,r27 1043 add r29=r29,r28 };; 1044 { .mfb; getf.sig r18=f95 } 1045 { .mii; 1046 (p6) add carry1=1,carry1 1047 cmp.ltu p6,p0=r29,r28 1048 add r30=r30,r29 };; 1049 { .mii; getf.sig r19=f86 1050 add r17=r17,r16 1051 mov carry3=0 } 1052 { .mii; 1053 (p6) add carry1=1,carry1 1054 cmp.ltu p6,p0=r30,r29 1055 add r30=r30,carry2 };; 1056 { .mii; getf.sig r20=f77 1057 cmp.ltu p7,p0=r17,r16 1058 add r18=r18,r17 } 1059 { .mii; 1060 (p6) add carry1=1,carry1 1061 cmp.ltu p6,p0=r30,carry2 };; 1062 { .mfb; getf.sig r21=f68 } 1063 { .mii; st8 [r33]=r30,16 1064 (p6) add carry1=1,carry1 };; 1065 1066 { .mfb; getf.sig r24=f114 } 1067 { .mii; (p7) add carry3=1,carry3 1068 cmp.ltu p7,p0=r18,r17 1069 add r19=r19,r18 };; 1070 { .mfb; getf.sig r25=f105 } 1071 { .mii; (p7) add carry3=1,carry3 1072 cmp.ltu p7,p0=r19,r18 1073 add r20=r20,r19 };; 1074 { .mfb; getf.sig r26=f96 } 1075 { .mii; (p7) add carry3=1,carry3 1076 cmp.ltu p7,p0=r20,r19 1077 add r21=r21,r20 };; 1078 { .mfb; getf.sig r27=f87 } 1079 { .mii; (p7) add carry3=1,carry3 1080 cmp.ltu p7,p0=r21,r20 1081 add r21=r21,carry1 };; 1082 { .mib; getf.sig r28=f78 1083 add r25=r25,r24 } 1084 { .mib; (p7) add carry3=1,carry3 1085 cmp.ltu p7,p8=r21,carry1};; 1086 { .mii; st8 [r32]=r21,16 1087 (p7) add carry2=1,carry3 1088 (p8) add carry2=0,carry3 } 1089 1090 { .mii; mov carry1=0 1091 cmp.ltu p6,p0=r25,r24 1092 add r26=r26,r25 };; 1093 { .mfb; getf.sig r16=f115 } 1094 { .mii; 1095 (p6) add carry1=1,carry1 1096 cmp.ltu p6,p0=r26,r25 1097 add r27=r27,r26 };; 1098 { .mfb; getf.sig r17=f106 } 1099 { .mii; 1100 (p6) add carry1=1,carry1 1101 cmp.ltu p6,p0=r27,r26 1102 add r28=r28,r27 };; 1103 { .mfb; getf.sig r18=f97 } 1104 { .mii; 1105 (p6) add carry1=1,carry1 1106 cmp.ltu p6,p0=r28,r27 1107 add r28=r28,carry2 };; 1108 { .mib; getf.sig r19=f88 1109 add r17=r17,r16 } 1110 { .mib; 1111 (p6) add carry1=1,carry1 1112 cmp.ltu p6,p0=r28,carry2 };; 1113 { .mii; st8 [r33]=r28,16 1114 (p6) add carry1=1,carry1 } 1115 1116 { .mii; mov carry2=0 1117 cmp.ltu p7,p0=r17,r16 1118 add r18=r18,r17 };; 1119 { .mfb; getf.sig r24=f116 } 1120 { .mii; (p7) add carry2=1,carry2 1121 cmp.ltu p7,p0=r18,r17 1122 add r19=r19,r18 };; 1123 { .mfb; getf.sig r25=f107 } 1124 { .mii; (p7) add carry2=1,carry2 1125 cmp.ltu p7,p0=r19,r18 1126 add r19=r19,carry1 };; 1127 { .mfb; getf.sig r26=f98 } 1128 { .mii; (p7) add carry2=1,carry2 1129 cmp.ltu p7,p0=r19,carry1};; 1130 { .mii; st8 [r32]=r19,16 1131 (p7) add carry2=1,carry2 } 1132 1133 { .mfb; add r25=r25,r24 };; 1134 1135 { .mfb; getf.sig r16=f117 } 1136 { .mii; mov carry1=0 1137 cmp.ltu p6,p0=r25,r24 1138 add r26=r26,r25 };; 1139 { .mfb; getf.sig r17=f108 } 1140 { .mii; 1141 (p6) add carry1=1,carry1 1142 cmp.ltu p6,p0=r26,r25 1143 add r26=r26,carry2 };; 1144 { .mfb; nop.m 0x0 } 1145 { .mii; 1146 (p6) add carry1=1,carry1 1147 cmp.ltu p6,p0=r26,carry2 };; 1148 { .mii; st8 [r33]=r26,16 1149 (p6) add carry1=1,carry1 } 1150 1151 { .mfb; add r17=r17,r16 };; 1152 { .mfb; getf.sig r24=f118 } 1153 { .mii; mov carry2=0 1154 cmp.ltu p7,p0=r17,r16 1155 add r17=r17,carry1 };; 1156 { .mii; (p7) add carry2=1,carry2 1157 cmp.ltu p7,p0=r17,carry1};; 1158 { .mii; st8 [r32]=r17 1159 (p7) add carry2=1,carry2 };; 1160 { .mfb; add r24=r24,carry2 };; 1161 { .mib; st8 [r33]=r24 } 1162 1163 { .mib; rum 1<<5 // clear um.mfh 1164 br.ret.sptk.many b0 };; 1165 .endp bn_mul_comba8# 1166 #undef carry3 1167 #undef carry2 1168 #undef carry1 1169 #endif 1170 1171 #if 1 1172 // It's possible to make it faster (see comment to bn_sqr_comba8), but 1173 // I reckon it doesn't worth the effort. Basically because the routine 1174 // (actually both of them) practically never called... So I just play 1175 // same trick as with bn_sqr_comba8. 1176 // 1177 // void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a) 1178 // 1179 .global bn_sqr_comba4# 1180 .proc bn_sqr_comba4# 1181 .align 64 1182 bn_sqr_comba4: 1183 .prologue 1184 .save ar.pfs,r2 1185 #if defined(_HPUX_SOURCE) && !defined(_LP64) 1186 { .mii; alloc r2=ar.pfs,2,1,0,0 1187 addp4 r32=0,r32 1188 addp4 r33=0,r33 };; 1189 { .mii; 1190 #else 1191 { .mii; alloc r2=ar.pfs,2,1,0,0 1192 #endif 1193 mov r34=r33 1194 add r14=8,r33 };; 1195 .body 1196 { .mii; add r17=8,r34 1197 add r15=16,r33 1198 add r18=16,r34 } 1199 { .mfb; add r16=24,r33 1200 br .L_cheat_entry_point4 };; 1201 .endp bn_sqr_comba4# 1202 #endif 1203 1204 #if 1 1205 // Runs in ~115 cycles and ~4.5 times faster than C. Well, whatever... 1206 // 1207 // void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) 1208 // 1209 #define carry1 r14 1210 #define carry2 r15 1211 .global bn_mul_comba4# 1212 .proc bn_mul_comba4# 1213 .align 64 1214 bn_mul_comba4: 1215 .prologue 1216 .save ar.pfs,r2 1217 #if defined(_HPUX_SOURCE) && !defined(_LP64) 1218 { .mii; alloc r2=ar.pfs,3,0,0,0 1219 addp4 r33=0,r33 1220 addp4 r34=0,r34 };; 1221 { .mii; addp4 r32=0,r32 1222 #else 1223 { .mii; alloc r2=ar.pfs,3,0,0,0 1224 #endif 1225 add r14=8,r33 1226 add r17=8,r34 } 1227 .body 1228 { .mii; add r15=16,r33 1229 add r18=16,r34 1230 add r16=24,r33 };; 1231 .L_cheat_entry_point4: 1232 { .mmi; add r19=24,r34 1233 1234 ldf8 f32=[r33] } 1235 1236 { .mmi; ldf8 f120=[r34] 1237 ldf8 f121=[r17] };; 1238 { .mmi; ldf8 f122=[r18] 1239 ldf8 f123=[r19] } 1240 1241 { .mmi; ldf8 f33=[r14] 1242 ldf8 f34=[r15] } 1243 { .mfi; ldf8 f35=[r16] 1244 1245 xma.hu f41=f32,f120,f0 } 1246 { .mfi; xma.lu f40=f32,f120,f0 };; 1247 { .mfi; xma.hu f51=f32,f121,f0 } 1248 { .mfi; xma.lu f50=f32,f121,f0 };; 1249 { .mfi; xma.hu f61=f32,f122,f0 } 1250 { .mfi; xma.lu f60=f32,f122,f0 };; 1251 { .mfi; xma.hu f71=f32,f123,f0 } 1252 { .mfi; xma.lu f70=f32,f123,f0 };;// 1253 // Major stall takes place here, and 3 more places below. Result from 1254 // first xma is not available for another 3 ticks. 1255 { .mfi; getf.sig r16=f40 1256 xma.hu f42=f33,f120,f41 1257 add r33=8,r32 } 1258 { .mfi; xma.lu f41=f33,f120,f41 };; 1259 { .mfi; getf.sig r24=f50 1260 xma.hu f52=f33,f121,f51 } 1261 { .mfi; xma.lu f51=f33,f121,f51 };; 1262 { .mfi; st8 [r32]=r16,16 1263 xma.hu f62=f33,f122,f61 } 1264 { .mfi; xma.lu f61=f33,f122,f61 };; 1265 { .mfi; xma.hu f72=f33,f123,f71 } 1266 { .mfi; xma.lu f71=f33,f123,f71 };;// 1267 //-------------------------------------------------// 1268 { .mfi; getf.sig r25=f41 1269 xma.hu f43=f34,f120,f42 } 1270 { .mfi; xma.lu f42=f34,f120,f42 };; 1271 { .mfi; getf.sig r16=f60 1272 xma.hu f53=f34,f121,f52 } 1273 { .mfi; xma.lu f52=f34,f121,f52 };; 1274 { .mfi; getf.sig r17=f51 1275 xma.hu f63=f34,f122,f62 1276 add r25=r25,r24 } 1277 { .mfi; mov carry1=0 1278 xma.lu f62=f34,f122,f62 };; 1279 { .mfi; st8 [r33]=r25,16 1280 xma.hu f73=f34,f123,f72 1281 cmp.ltu p6,p0=r25,r24 } 1282 { .mfi; xma.lu f72=f34,f123,f72 };;// 1283 //-------------------------------------------------// 1284 { .mfi; getf.sig r18=f42 1285 xma.hu f44=f35,f120,f43 1286 (p6) add carry1=1,carry1 } 1287 { .mfi; add r17=r17,r16 1288 xma.lu f43=f35,f120,f43 1289 mov carry2=0 };; 1290 { .mfi; getf.sig r24=f70 1291 xma.hu f54=f35,f121,f53 1292 cmp.ltu p7,p0=r17,r16 } 1293 { .mfi; xma.lu f53=f35,f121,f53 };; 1294 { .mfi; getf.sig r25=f61 1295 xma.hu f64=f35,f122,f63 1296 add r18=r18,r17 } 1297 { .mfi; xma.lu f63=f35,f122,f63 1298 (p7) add carry2=1,carry2 };; 1299 { .mfi; getf.sig r26=f52 1300 xma.hu f74=f35,f123,f73 1301 cmp.ltu p7,p0=r18,r17 } 1302 { .mfi; xma.lu f73=f35,f123,f73 1303 add r18=r18,carry1 };; 1304 //-------------------------------------------------// 1305 { .mii; st8 [r32]=r18,16 1306 (p7) add carry2=1,carry2 1307 cmp.ltu p7,p0=r18,carry1 };; 1308 1309 { .mfi; getf.sig r27=f43 // last major stall 1310 (p7) add carry2=1,carry2 };; 1311 { .mii; getf.sig r16=f71 1312 add r25=r25,r24 1313 mov carry1=0 };; 1314 { .mii; getf.sig r17=f62 1315 cmp.ltu p6,p0=r25,r24 1316 add r26=r26,r25 };; 1317 { .mii; 1318 (p6) add carry1=1,carry1 1319 cmp.ltu p6,p0=r26,r25 1320 add r27=r27,r26 };; 1321 { .mii; 1322 (p6) add carry1=1,carry1 1323 cmp.ltu p6,p0=r27,r26 1324 add r27=r27,carry2 };; 1325 { .mii; getf.sig r18=f53 1326 (p6) add carry1=1,carry1 1327 cmp.ltu p6,p0=r27,carry2 };; 1328 { .mfi; st8 [r33]=r27,16 1329 (p6) add carry1=1,carry1 } 1330 1331 { .mii; getf.sig r19=f44 1332 add r17=r17,r16 1333 mov carry2=0 };; 1334 { .mii; getf.sig r24=f72 1335 cmp.ltu p7,p0=r17,r16 1336 add r18=r18,r17 };; 1337 { .mii; (p7) add carry2=1,carry2 1338 cmp.ltu p7,p0=r18,r17 1339 add r19=r19,r18 };; 1340 { .mii; (p7) add carry2=1,carry2 1341 cmp.ltu p7,p0=r19,r18 1342 add r19=r19,carry1 };; 1343 { .mii; getf.sig r25=f63 1344 (p7) add carry2=1,carry2 1345 cmp.ltu p7,p0=r19,carry1};; 1346 { .mii; st8 [r32]=r19,16 1347 (p7) add carry2=1,carry2 } 1348 1349 { .mii; getf.sig r26=f54 1350 add r25=r25,r24 1351 mov carry1=0 };; 1352 { .mii; getf.sig r16=f73 1353 cmp.ltu p6,p0=r25,r24 1354 add r26=r26,r25 };; 1355 { .mii; 1356 (p6) add carry1=1,carry1 1357 cmp.ltu p6,p0=r26,r25 1358 add r26=r26,carry2 };; 1359 { .mii; getf.sig r17=f64 1360 (p6) add carry1=1,carry1 1361 cmp.ltu p6,p0=r26,carry2 };; 1362 { .mii; st8 [r33]=r26,16 1363 (p6) add carry1=1,carry1 } 1364 1365 { .mii; getf.sig r24=f74 1366 add r17=r17,r16 1367 mov carry2=0 };; 1368 { .mii; cmp.ltu p7,p0=r17,r16 1369 add r17=r17,carry1 };; 1370 1371 { .mii; (p7) add carry2=1,carry2 1372 cmp.ltu p7,p0=r17,carry1};; 1373 { .mii; st8 [r32]=r17,16 1374 (p7) add carry2=1,carry2 };; 1375 1376 { .mii; add r24=r24,carry2 };; 1377 { .mii; st8 [r33]=r24 } 1378 1379 { .mib; rum 1<<5 // clear um.mfh 1380 br.ret.sptk.many b0 };; 1381 .endp bn_mul_comba4# 1382 #undef carry2 1383 #undef carry1 1384 #endif 1385 1386 #if 1 1387 // 1388 // BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d) 1389 // 1390 // In the nutshell it's a port of my MIPS III/IV implementation. 1391 // 1392 #define AT r14 1393 #define H r16 1394 #define HH r20 1395 #define L r17 1396 #define D r18 1397 #define DH r22 1398 #define I r21 1399 1400 #if 0 1401 // Some preprocessors (most notably HP-UX) appear to be allergic to 1402 // macros enclosed to parenthesis [as these three were]. 1403 #define cont p16 1404 #define break p0 // p20 1405 #define equ p24 1406 #else 1407 cont=p16 1408 break=p0 1409 equ=p24 1410 #endif 1411 1412 .global abort# 1413 .global bn_div_words# 1414 .proc bn_div_words# 1415 .align 64 1416 bn_div_words: 1417 .prologue 1418 .save ar.pfs,r2 1419 { .mii; alloc r2=ar.pfs,3,5,0,8 1420 .save b0,r3 1421 mov r3=b0 1422 .save pr,r10 1423 mov r10=pr };; 1424 { .mmb; cmp.eq p6,p0=r34,r0 1425 mov r8=-1 1426 (p6) br.ret.spnt.many b0 };; 1427 1428 .body 1429 { .mii; mov H=r32 // save h 1430 mov ar.ec=0 // don't rotate at exit 1431 mov pr.rot=0 } 1432 { .mii; mov L=r33 // save l 1433 mov r25=r0 // needed if abort is called on VMS 1434 mov r36=r0 };; 1435 1436 .L_divw_shift: // -vv- note signed comparison 1437 { .mfi; (p0) cmp.lt p16,p0=r0,r34 // d 1438 (p0) shladd r33=r34,1,r0 } 1439 { .mfb; (p0) add r35=1,r36 1440 (p0) nop.f 0x0 1441 (p16) br.wtop.dpnt .L_divw_shift };; 1442 1443 { .mii; mov D=r34 1444 shr.u DH=r34,32 1445 sub r35=64,r36 };; 1446 { .mii; setf.sig f7=DH 1447 shr.u AT=H,r35 1448 mov I=r36 };; 1449 { .mib; cmp.ne p6,p0=r0,AT 1450 shl H=H,r36 1451 (p6) br.call.spnt.clr b0=abort };; // overflow, die... 1452 1453 { .mfi; fcvt.xuf.s1 f7=f7 1454 shr.u AT=L,r35 };; 1455 { .mii; shl L=L,r36 1456 or H=H,AT };; 1457 1458 { .mii; nop.m 0x0 1459 cmp.leu p6,p0=D,H;; 1460 (p6) sub H=H,D } 1461 1462 { .mlx; setf.sig f14=D 1463 movl AT=0xffffffff };; 1464 /////////////////////////////////////////////////////////// 1465 { .mii; setf.sig f6=H 1466 shr.u HH=H,32;; 1467 cmp.eq p6,p7=HH,DH };; 1468 { .mfb; 1469 (p6) setf.sig f8=AT 1470 (p7) fcvt.xuf.s1 f6=f6 1471 (p7) br.call.sptk b6=.L_udiv64_32_b6 };; 1472 1473 { .mfi; getf.sig r33=f8 // q 1474 xmpy.lu f9=f8,f14 } 1475 { .mfi; xmpy.hu f10=f8,f14 1476 shrp H=H,L,32 };; 1477 1478 { .mmi; getf.sig r35=f9 // tl 1479 getf.sig r31=f10 };; // th 1480 1481 .L_divw_1st_iter: 1482 { .mii; (p0) add r32=-1,r33 1483 (p0) cmp.eq equ,cont=HH,r31 };; 1484 { .mii; (p0) cmp.ltu p8,p0=r35,D 1485 (p0) sub r34=r35,D 1486 (equ) cmp.leu break,cont=r35,H };; 1487 { .mib; (cont) cmp.leu cont,break=HH,r31 1488 (p8) add r31=-1,r31 1489 (cont) br.wtop.spnt .L_divw_1st_iter };; 1490 /////////////////////////////////////////////////////////// 1491 { .mii; sub H=H,r35 1492 shl r8=r33,32 1493 shl L=L,32 };; 1494 /////////////////////////////////////////////////////////// 1495 { .mii; setf.sig f6=H 1496 shr.u HH=H,32;; 1497 cmp.eq p6,p7=HH,DH };; 1498 { .mfb; 1499 (p6) setf.sig f8=AT 1500 (p7) fcvt.xuf.s1 f6=f6 1501 (p7) br.call.sptk b6=.L_udiv64_32_b6 };; 1502 1503 { .mfi; getf.sig r33=f8 // q 1504 xmpy.lu f9=f8,f14 } 1505 { .mfi; xmpy.hu f10=f8,f14 1506 shrp H=H,L,32 };; 1507 1508 { .mmi; getf.sig r35=f9 // tl 1509 getf.sig r31=f10 };; // th 1510 1511 .L_divw_2nd_iter: 1512 { .mii; (p0) add r32=-1,r33 1513 (p0) cmp.eq equ,cont=HH,r31 };; 1514 { .mii; (p0) cmp.ltu p8,p0=r35,D 1515 (p0) sub r34=r35,D 1516 (equ) cmp.leu break,cont=r35,H };; 1517 { .mib; (cont) cmp.leu cont,break=HH,r31 1518 (p8) add r31=-1,r31 1519 (cont) br.wtop.spnt .L_divw_2nd_iter };; 1520 /////////////////////////////////////////////////////////// 1521 { .mii; sub H=H,r35 1522 or r8=r8,r33 1523 mov ar.pfs=r2 };; 1524 { .mii; shr.u r9=H,I // remainder if anybody wants it 1525 mov pr=r10,0x1ffff } 1526 { .mfb; br.ret.sptk.many b0 };; 1527 1528 // Unsigned 64 by 32 (well, by 64 for the moment) bit integer division 1529 // procedure. 1530 // 1531 // inputs: f6 = (double)a, f7 = (double)b 1532 // output: f8 = (int)(a/b) 1533 // clobbered: f8,f9,f10,f11,pred 1534 pred=p15 1535 // This snippet is based on text found in the "Divide, Square 1536 // Root and Remainder" section at 1537 // http://www.intel.com/software/products/opensource/libraries/num.htm. 1538 // Yes, I admit that the referred code was used as template, 1539 // but after I realized that there hardly is any other instruction 1540 // sequence which would perform this operation. I mean I figure that 1541 // any independent attempt to implement high-performance division 1542 // will result in code virtually identical to the Intel code. It 1543 // should be noted though that below division kernel is 1 cycle 1544 // faster than Intel one (note commented splits:-), not to mention 1545 // original prologue (rather lack of one) and epilogue. 1546 .align 32 1547 .skip 16 1548 .L_udiv64_32_b6: 1549 frcpa.s1 f8,pred=f6,f7;; // [0] y0 = 1 / b 1550 1551 (pred) fnma.s1 f9=f7,f8,f1 // [5] e0 = 1 - b * y0 1552 (pred) fmpy.s1 f10=f6,f8;; // [5] q0 = a * y0 1553 (pred) fmpy.s1 f11=f9,f9 // [10] e1 = e0 * e0 1554 (pred) fma.s1 f10=f9,f10,f10;; // [10] q1 = q0 + e0 * q0 1555 (pred) fma.s1 f8=f9,f8,f8 //;; // [15] y1 = y0 + e0 * y0 1556 (pred) fma.s1 f9=f11,f10,f10;; // [15] q2 = q1 + e1 * q1 1557 (pred) fma.s1 f8=f11,f8,f8 //;; // [20] y2 = y1 + e1 * y1 1558 (pred) fnma.s1 f10=f7,f9,f6;; // [20] r2 = a - b * q2 1559 (pred) fma.s1 f8=f10,f8,f9;; // [25] q3 = q2 + r2 * y2 1560 1561 fcvt.fxu.trunc.s1 f8=f8 // [30] q = trunc(q3) 1562 br.ret.sptk.many b6;; 1563 .endp bn_div_words# 1564 #endif 1565