sha1-armv4-large.pl revision 290207
1#!/usr/bin/env perl 2 3# ==================================================================== 4# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL 5# project. The module is, however, dual licensed under OpenSSL and 6# CRYPTOGAMS licenses depending on where you obtain it. For further 7# details see http://www.openssl.org/~appro/cryptogams/. 8# ==================================================================== 9 10# sha1_block procedure for ARMv4. 11# 12# January 2007. 13 14# Size/performance trade-off 15# ==================================================================== 16# impl size in bytes comp cycles[*] measured performance 17# ==================================================================== 18# thumb 304 3212 4420 19# armv4-small 392/+29% 1958/+64% 2250/+96% 20# armv4-compact 740/+89% 1552/+26% 1840/+22% 21# armv4-large 1420/+92% 1307/+19% 1370/+34%[***] 22# full unroll ~5100/+260% ~1260/+4% ~1300/+5% 23# ==================================================================== 24# thumb = same as 'small' but in Thumb instructions[**] and 25# with recurring code in two private functions; 26# small = detached Xload/update, loops are folded; 27# compact = detached Xload/update, 5x unroll; 28# large = interleaved Xload/update, 5x unroll; 29# full unroll = interleaved Xload/update, full unroll, estimated[!]; 30# 31# [*] Manually counted instructions in "grand" loop body. Measured 32# performance is affected by prologue and epilogue overhead, 33# i-cache availability, branch penalties, etc. 34# [**] While each Thumb instruction is twice smaller, they are not as 35# diverse as ARM ones: e.g., there are only two arithmetic 36# instructions with 3 arguments, no [fixed] rotate, addressing 37# modes are limited. As result it takes more instructions to do 38# the same job in Thumb, therefore the code is never twice as 39# small and always slower. 40# [***] which is also ~35% better than compiler generated code. Dual- 41# issue Cortex A8 core was measured to process input block in 42# ~990 cycles. 43 44# August 2010. 45# 46# Rescheduling for dual-issue pipeline resulted in 13% improvement on 47# Cortex A8 core and in absolute terms ~870 cycles per input block 48# [or 13.6 cycles per byte]. 49 50# February 2011. 51# 52# Profiler-assisted and platform-specific optimization resulted in 10% 53# improvement on Cortex A8 core and 12.2 cycles per byte. 54 55# September 2013. 56# 57# Add NEON implementation (see sha1-586.pl for background info). On 58# Cortex A8 it was measured to process one byte in 6.7 cycles or >80% 59# faster than integer-only code. Because [fully unrolled] NEON code 60# is ~2.5x larger and there are some redundant instructions executed 61# when processing last block, improvement is not as big for smallest 62# blocks, only ~30%. Snapdragon S4 is a tad faster, 6.4 cycles per 63# byte, which is also >80% faster than integer-only code. 64 65# May 2014. 66# 67# Add ARMv8 code path performing at 2.35 cpb on Apple A7. 68 69while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {} 70open STDOUT,">$output"; 71 72$ctx="r0"; 73$inp="r1"; 74$len="r2"; 75$a="r3"; 76$b="r4"; 77$c="r5"; 78$d="r6"; 79$e="r7"; 80$K="r8"; 81$t0="r9"; 82$t1="r10"; 83$t2="r11"; 84$t3="r12"; 85$Xi="r14"; 86@V=($a,$b,$c,$d,$e); 87 88sub Xupdate { 89my ($a,$b,$c,$d,$e,$opt1,$opt2)=@_; 90$code.=<<___; 91 ldr $t0,[$Xi,#15*4] 92 ldr $t1,[$Xi,#13*4] 93 ldr $t2,[$Xi,#7*4] 94 add $e,$K,$e,ror#2 @ E+=K_xx_xx 95 ldr $t3,[$Xi,#2*4] 96 eor $t0,$t0,$t1 97 eor $t2,$t2,$t3 @ 1 cycle stall 98 eor $t1,$c,$d @ F_xx_xx 99 mov $t0,$t0,ror#31 100 add $e,$e,$a,ror#27 @ E+=ROR(A,27) 101 eor $t0,$t0,$t2,ror#31 102 str $t0,[$Xi,#-4]! 103 $opt1 @ F_xx_xx 104 $opt2 @ F_xx_xx 105 add $e,$e,$t0 @ E+=X[i] 106___ 107} 108 109sub BODY_00_15 { 110my ($a,$b,$c,$d,$e)=@_; 111$code.=<<___; 112#if __ARM_ARCH__<7 113 ldrb $t1,[$inp,#2] 114 ldrb $t0,[$inp,#3] 115 ldrb $t2,[$inp,#1] 116 add $e,$K,$e,ror#2 @ E+=K_00_19 117 ldrb $t3,[$inp],#4 118 orr $t0,$t0,$t1,lsl#8 119 eor $t1,$c,$d @ F_xx_xx 120 orr $t0,$t0,$t2,lsl#16 121 add $e,$e,$a,ror#27 @ E+=ROR(A,27) 122 orr $t0,$t0,$t3,lsl#24 123#else 124 ldr $t0,[$inp],#4 @ handles unaligned 125 add $e,$K,$e,ror#2 @ E+=K_00_19 126 eor $t1,$c,$d @ F_xx_xx 127 add $e,$e,$a,ror#27 @ E+=ROR(A,27) 128#ifdef __ARMEL__ 129 rev $t0,$t0 @ byte swap 130#endif 131#endif 132 and $t1,$b,$t1,ror#2 133 add $e,$e,$t0 @ E+=X[i] 134 eor $t1,$t1,$d,ror#2 @ F_00_19(B,C,D) 135 str $t0,[$Xi,#-4]! 136 add $e,$e,$t1 @ E+=F_00_19(B,C,D) 137___ 138} 139 140sub BODY_16_19 { 141my ($a,$b,$c,$d,$e)=@_; 142 &Xupdate(@_,"and $t1,$b,$t1,ror#2"); 143$code.=<<___; 144 eor $t1,$t1,$d,ror#2 @ F_00_19(B,C,D) 145 add $e,$e,$t1 @ E+=F_00_19(B,C,D) 146___ 147} 148 149sub BODY_20_39 { 150my ($a,$b,$c,$d,$e)=@_; 151 &Xupdate(@_,"eor $t1,$b,$t1,ror#2"); 152$code.=<<___; 153 add $e,$e,$t1 @ E+=F_20_39(B,C,D) 154___ 155} 156 157sub BODY_40_59 { 158my ($a,$b,$c,$d,$e)=@_; 159 &Xupdate(@_,"and $t1,$b,$t1,ror#2","and $t2,$c,$d"); 160$code.=<<___; 161 add $e,$e,$t1 @ E+=F_40_59(B,C,D) 162 add $e,$e,$t2,ror#2 163___ 164} 165 166$code=<<___; 167#include "arm_arch.h" 168 169.text 170.code 32 171 172.global sha1_block_data_order 173.type sha1_block_data_order,%function 174 175.align 5 176sha1_block_data_order: 177#if __ARM_MAX_ARCH__>=7 178 sub r3,pc,#8 @ sha1_block_data_order 179 ldr r12,.LOPENSSL_armcap 180 ldr r12,[r3,r12] @ OPENSSL_armcap_P 181 tst r12,#ARMV8_SHA1 182 bne .LARMv8 183 tst r12,#ARMV7_NEON 184 bne .LNEON 185#endif 186 stmdb sp!,{r4-r12,lr} 187 add $len,$inp,$len,lsl#6 @ $len to point at the end of $inp 188 ldmia $ctx,{$a,$b,$c,$d,$e} 189.Lloop: 190 ldr $K,.LK_00_19 191 mov $Xi,sp 192 sub sp,sp,#15*4 193 mov $c,$c,ror#30 194 mov $d,$d,ror#30 195 mov $e,$e,ror#30 @ [6] 196.L_00_15: 197___ 198for($i=0;$i<5;$i++) { 199 &BODY_00_15(@V); unshift(@V,pop(@V)); 200} 201$code.=<<___; 202 teq $Xi,sp 203 bne .L_00_15 @ [((11+4)*5+2)*3] 204 sub sp,sp,#25*4 205___ 206 &BODY_00_15(@V); unshift(@V,pop(@V)); 207 &BODY_16_19(@V); unshift(@V,pop(@V)); 208 &BODY_16_19(@V); unshift(@V,pop(@V)); 209 &BODY_16_19(@V); unshift(@V,pop(@V)); 210 &BODY_16_19(@V); unshift(@V,pop(@V)); 211$code.=<<___; 212 213 ldr $K,.LK_20_39 @ [+15+16*4] 214 cmn sp,#0 @ [+3], clear carry to denote 20_39 215.L_20_39_or_60_79: 216___ 217for($i=0;$i<5;$i++) { 218 &BODY_20_39(@V); unshift(@V,pop(@V)); 219} 220$code.=<<___; 221 teq $Xi,sp @ preserve carry 222 bne .L_20_39_or_60_79 @ [+((12+3)*5+2)*4] 223 bcs .L_done @ [+((12+3)*5+2)*4], spare 300 bytes 224 225 ldr $K,.LK_40_59 226 sub sp,sp,#20*4 @ [+2] 227.L_40_59: 228___ 229for($i=0;$i<5;$i++) { 230 &BODY_40_59(@V); unshift(@V,pop(@V)); 231} 232$code.=<<___; 233 teq $Xi,sp 234 bne .L_40_59 @ [+((12+5)*5+2)*4] 235 236 ldr $K,.LK_60_79 237 sub sp,sp,#20*4 238 cmp sp,#0 @ set carry to denote 60_79 239 b .L_20_39_or_60_79 @ [+4], spare 300 bytes 240.L_done: 241 add sp,sp,#80*4 @ "deallocate" stack frame 242 ldmia $ctx,{$K,$t0,$t1,$t2,$t3} 243 add $a,$K,$a 244 add $b,$t0,$b 245 add $c,$t1,$c,ror#2 246 add $d,$t2,$d,ror#2 247 add $e,$t3,$e,ror#2 248 stmia $ctx,{$a,$b,$c,$d,$e} 249 teq $inp,$len 250 bne .Lloop @ [+18], total 1307 251 252#if __ARM_ARCH__>=5 253 ldmia sp!,{r4-r12,pc} 254#else 255 ldmia sp!,{r4-r12,lr} 256 tst lr,#1 257 moveq pc,lr @ be binary compatible with V4, yet 258 bx lr @ interoperable with Thumb ISA:-) 259#endif 260.size sha1_block_data_order,.-sha1_block_data_order 261 262.align 5 263.LK_00_19: .word 0x5a827999 264.LK_20_39: .word 0x6ed9eba1 265.LK_40_59: .word 0x8f1bbcdc 266.LK_60_79: .word 0xca62c1d6 267#if __ARM_MAX_ARCH__>=7 268.LOPENSSL_armcap: 269.word OPENSSL_armcap_P-sha1_block_data_order 270#endif 271.asciz "SHA1 block transform for ARMv4/NEON/ARMv8, CRYPTOGAMS by <appro\@openssl.org>" 272.align 5 273___ 274##################################################################### 275# NEON stuff 276# 277{{{ 278my @V=($a,$b,$c,$d,$e); 279my ($K_XX_XX,$Ki,$t0,$t1,$Xfer,$saved_sp)=map("r$_",(8..12,14)); 280my $Xi=4; 281my @X=map("q$_",(8..11,0..3)); 282my @Tx=("q12","q13"); 283my ($K,$zero)=("q14","q15"); 284my $j=0; 285 286sub AUTOLOAD() # thunk [simplified] x86-style perlasm 287{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./; 288 my $arg = pop; 289 $arg = "#$arg" if ($arg*1 eq $arg); 290 $code .= "\t$opcode\t".join(',',@_,$arg)."\n"; 291} 292 293sub body_00_19 () { 294 ( 295 '($a,$b,$c,$d,$e)=@V;'. # '$code.="@ $j\n";'. 296 '&bic ($t0,$d,$b)', 297 '&add ($e,$e,$Ki)', # e+=X[i]+K 298 '&and ($t1,$c,$b)', 299 '&ldr ($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))', 300 '&add ($e,$e,$a,"ror#27")', # e+=ROR(A,27) 301 '&eor ($t1,$t1,$t0)', # F_00_19 302 '&mov ($b,$b,"ror#2")', # b=ROR(b,2) 303 '&add ($e,$e,$t1);'. # e+=F_00_19 304 '$j++; unshift(@V,pop(@V));' 305 ) 306} 307sub body_20_39 () { 308 ( 309 '($a,$b,$c,$d,$e)=@V;'. # '$code.="@ $j\n";'. 310 '&eor ($t0,$b,$d)', 311 '&add ($e,$e,$Ki)', # e+=X[i]+K 312 '&ldr ($Ki,sprintf "[sp,#%d]",4*(($j+1)&15)) if ($j<79)', 313 '&eor ($t1,$t0,$c)', # F_20_39 314 '&add ($e,$e,$a,"ror#27")', # e+=ROR(A,27) 315 '&mov ($b,$b,"ror#2")', # b=ROR(b,2) 316 '&add ($e,$e,$t1);'. # e+=F_20_39 317 '$j++; unshift(@V,pop(@V));' 318 ) 319} 320sub body_40_59 () { 321 ( 322 '($a,$b,$c,$d,$e)=@V;'. # '$code.="@ $j\n";'. 323 '&add ($e,$e,$Ki)', # e+=X[i]+K 324 '&and ($t0,$c,$d)', 325 '&ldr ($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))', 326 '&add ($e,$e,$a,"ror#27")', # e+=ROR(A,27) 327 '&eor ($t1,$c,$d)', 328 '&add ($e,$e,$t0)', 329 '&and ($t1,$t1,$b)', 330 '&mov ($b,$b,"ror#2")', # b=ROR(b,2) 331 '&add ($e,$e,$t1);'. # e+=F_40_59 332 '$j++; unshift(@V,pop(@V));' 333 ) 334} 335 336sub Xupdate_16_31 () 337{ use integer; 338 my $body = shift; 339 my @insns = (&$body,&$body,&$body,&$body); 340 my ($a,$b,$c,$d,$e); 341 342 &vext_8 (@X[0],@X[-4&7],@X[-3&7],8); # compose "X[-14]" in "X[0]" 343 eval(shift(@insns)); 344 eval(shift(@insns)); 345 eval(shift(@insns)); 346 &vadd_i32 (@Tx[1],@X[-1&7],$K); 347 eval(shift(@insns)); 348 &vld1_32 ("{$K\[]}","[$K_XX_XX,:32]!") if ($Xi%5==0); 349 eval(shift(@insns)); 350 &vext_8 (@Tx[0],@X[-1&7],$zero,4); # "X[-3]", 3 words 351 eval(shift(@insns)); 352 eval(shift(@insns)); 353 eval(shift(@insns)); 354 &veor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 355 eval(shift(@insns)); 356 eval(shift(@insns)); 357 &veor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" 358 eval(shift(@insns)); 359 eval(shift(@insns)); 360 &veor (@Tx[0],@Tx[0],@X[0]); # "X[0]"^="X[-3]"^"X[-8] 361 eval(shift(@insns)); 362 eval(shift(@insns)); 363 &vst1_32 ("{@Tx[1]}","[$Xfer,:128]!"); # X[]+K xfer 364 &sub ($Xfer,$Xfer,64) if ($Xi%4==0); 365 eval(shift(@insns)); 366 eval(shift(@insns)); 367 &vext_8 (@Tx[1],$zero,@Tx[0],4); # "X[0]"<<96, extract one dword 368 eval(shift(@insns)); 369 eval(shift(@insns)); 370 &vadd_i32 (@X[0],@Tx[0],@Tx[0]); 371 eval(shift(@insns)); 372 eval(shift(@insns)); 373 &vsri_32 (@X[0],@Tx[0],31); # "X[0]"<<<=1 374 eval(shift(@insns)); 375 eval(shift(@insns)); 376 eval(shift(@insns)); 377 &vshr_u32 (@Tx[0],@Tx[1],30); 378 eval(shift(@insns)); 379 eval(shift(@insns)); 380 &vshl_u32 (@Tx[1],@Tx[1],2); 381 eval(shift(@insns)); 382 eval(shift(@insns)); 383 &veor (@X[0],@X[0],@Tx[0]); 384 eval(shift(@insns)); 385 eval(shift(@insns)); 386 &veor (@X[0],@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2 387 388 foreach (@insns) { eval; } # remaining instructions [if any] 389 390 $Xi++; push(@X,shift(@X)); # "rotate" X[] 391} 392 393sub Xupdate_32_79 () 394{ use integer; 395 my $body = shift; 396 my @insns = (&$body,&$body,&$body,&$body); 397 my ($a,$b,$c,$d,$e); 398 399 &vext_8 (@Tx[0],@X[-2&7],@X[-1&7],8); # compose "X[-6]" 400 eval(shift(@insns)); 401 eval(shift(@insns)); 402 eval(shift(@insns)); 403 &veor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 404 eval(shift(@insns)); 405 eval(shift(@insns)); 406 &veor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 407 eval(shift(@insns)); 408 eval(shift(@insns)); 409 &vadd_i32 (@Tx[1],@X[-1&7],$K); 410 eval(shift(@insns)); 411 &vld1_32 ("{$K\[]}","[$K_XX_XX,:32]!") if ($Xi%5==0); 412 eval(shift(@insns)); 413 &veor (@Tx[0],@Tx[0],@X[0]); # "X[-6]"^="X[0]" 414 eval(shift(@insns)); 415 eval(shift(@insns)); 416 &vshr_u32 (@X[0],@Tx[0],30); 417 eval(shift(@insns)); 418 eval(shift(@insns)); 419 &vst1_32 ("{@Tx[1]}","[$Xfer,:128]!"); # X[]+K xfer 420 &sub ($Xfer,$Xfer,64) if ($Xi%4==0); 421 eval(shift(@insns)); 422 eval(shift(@insns)); 423 &vsli_32 (@X[0],@Tx[0],2); # "X[0]"="X[-6]"<<<2 424 425 foreach (@insns) { eval; } # remaining instructions [if any] 426 427 $Xi++; push(@X,shift(@X)); # "rotate" X[] 428} 429 430sub Xuplast_80 () 431{ use integer; 432 my $body = shift; 433 my @insns = (&$body,&$body,&$body,&$body); 434 my ($a,$b,$c,$d,$e); 435 436 &vadd_i32 (@Tx[1],@X[-1&7],$K); 437 eval(shift(@insns)); 438 eval(shift(@insns)); 439 &vst1_32 ("{@Tx[1]}","[$Xfer,:128]!"); 440 &sub ($Xfer,$Xfer,64); 441 442 &teq ($inp,$len); 443 &sub ($K_XX_XX,$K_XX_XX,16); # rewind $K_XX_XX 444 &subeq ($inp,$inp,64); # reload last block to avoid SEGV 445 &vld1_8 ("{@X[-4&7]-@X[-3&7]}","[$inp]!"); 446 eval(shift(@insns)); 447 eval(shift(@insns)); 448 &vld1_8 ("{@X[-2&7]-@X[-1&7]}","[$inp]!"); 449 eval(shift(@insns)); 450 eval(shift(@insns)); 451 &vld1_32 ("{$K\[]}","[$K_XX_XX,:32]!"); # load K_00_19 452 eval(shift(@insns)); 453 eval(shift(@insns)); 454 &vrev32_8 (@X[-4&7],@X[-4&7]); 455 456 foreach (@insns) { eval; } # remaining instructions 457 458 $Xi=0; 459} 460 461sub Xloop() 462{ use integer; 463 my $body = shift; 464 my @insns = (&$body,&$body,&$body,&$body); 465 my ($a,$b,$c,$d,$e); 466 467 &vrev32_8 (@X[($Xi-3)&7],@X[($Xi-3)&7]); 468 eval(shift(@insns)); 469 eval(shift(@insns)); 470 &vadd_i32 (@X[$Xi&7],@X[($Xi-4)&7],$K); 471 eval(shift(@insns)); 472 eval(shift(@insns)); 473 &vst1_32 ("{@X[$Xi&7]}","[$Xfer,:128]!");# X[]+K xfer to IALU 474 475 foreach (@insns) { eval; } 476 477 $Xi++; 478} 479 480$code.=<<___; 481#if __ARM_MAX_ARCH__>=7 482.arch armv7-a 483.fpu neon 484 485.type sha1_block_data_order_neon,%function 486.align 4 487sha1_block_data_order_neon: 488.LNEON: 489 stmdb sp!,{r4-r12,lr} 490 add $len,$inp,$len,lsl#6 @ $len to point at the end of $inp 491 @ dmb @ errata #451034 on early Cortex A8 492 @ vstmdb sp!,{d8-d15} @ ABI specification says so 493 mov $saved_sp,sp 494 sub sp,sp,#64 @ alloca 495 adr $K_XX_XX,.LK_00_19 496 bic sp,sp,#15 @ align for 128-bit stores 497 498 ldmia $ctx,{$a,$b,$c,$d,$e} @ load context 499 mov $Xfer,sp 500 501 vld1.8 {@X[-4&7]-@X[-3&7]},[$inp]! @ handles unaligned 502 veor $zero,$zero,$zero 503 vld1.8 {@X[-2&7]-@X[-1&7]},[$inp]! 504 vld1.32 {${K}\[]},[$K_XX_XX,:32]! @ load K_00_19 505 vrev32.8 @X[-4&7],@X[-4&7] @ yes, even on 506 vrev32.8 @X[-3&7],@X[-3&7] @ big-endian... 507 vrev32.8 @X[-2&7],@X[-2&7] 508 vadd.i32 @X[0],@X[-4&7],$K 509 vrev32.8 @X[-1&7],@X[-1&7] 510 vadd.i32 @X[1],@X[-3&7],$K 511 vst1.32 {@X[0]},[$Xfer,:128]! 512 vadd.i32 @X[2],@X[-2&7],$K 513 vst1.32 {@X[1]},[$Xfer,:128]! 514 vst1.32 {@X[2]},[$Xfer,:128]! 515 ldr $Ki,[sp] @ big RAW stall 516 517.Loop_neon: 518___ 519 &Xupdate_16_31(\&body_00_19); 520 &Xupdate_16_31(\&body_00_19); 521 &Xupdate_16_31(\&body_00_19); 522 &Xupdate_16_31(\&body_00_19); 523 &Xupdate_32_79(\&body_00_19); 524 &Xupdate_32_79(\&body_20_39); 525 &Xupdate_32_79(\&body_20_39); 526 &Xupdate_32_79(\&body_20_39); 527 &Xupdate_32_79(\&body_20_39); 528 &Xupdate_32_79(\&body_20_39); 529 &Xupdate_32_79(\&body_40_59); 530 &Xupdate_32_79(\&body_40_59); 531 &Xupdate_32_79(\&body_40_59); 532 &Xupdate_32_79(\&body_40_59); 533 &Xupdate_32_79(\&body_40_59); 534 &Xupdate_32_79(\&body_20_39); 535 &Xuplast_80(\&body_20_39); 536 &Xloop(\&body_20_39); 537 &Xloop(\&body_20_39); 538 &Xloop(\&body_20_39); 539$code.=<<___; 540 ldmia $ctx,{$Ki,$t0,$t1,$Xfer} @ accumulate context 541 add $a,$a,$Ki 542 ldr $Ki,[$ctx,#16] 543 add $b,$b,$t0 544 add $c,$c,$t1 545 add $d,$d,$Xfer 546 moveq sp,$saved_sp 547 add $e,$e,$Ki 548 ldrne $Ki,[sp] 549 stmia $ctx,{$a,$b,$c,$d,$e} 550 addne $Xfer,sp,#3*16 551 bne .Loop_neon 552 553 @ vldmia sp!,{d8-d15} 554 ldmia sp!,{r4-r12,pc} 555.size sha1_block_data_order_neon,.-sha1_block_data_order_neon 556#endif 557___ 558}}} 559##################################################################### 560# ARMv8 stuff 561# 562{{{ 563my ($ABCD,$E,$E0,$E1)=map("q$_",(0..3)); 564my @MSG=map("q$_",(4..7)); 565my @Kxx=map("q$_",(8..11)); 566my ($W0,$W1,$ABCD_SAVE)=map("q$_",(12..14)); 567 568$code.=<<___; 569#if __ARM_MAX_ARCH__>=7 570.type sha1_block_data_order_armv8,%function 571.align 5 572sha1_block_data_order_armv8: 573.LARMv8: 574 vstmdb sp!,{d8-d15} @ ABI specification says so 575 576 veor $E,$E,$E 577 adr r3,.LK_00_19 578 vld1.32 {$ABCD},[$ctx]! 579 vld1.32 {$E\[0]},[$ctx] 580 sub $ctx,$ctx,#16 581 vld1.32 {@Kxx[0]\[]},[r3,:32]! 582 vld1.32 {@Kxx[1]\[]},[r3,:32]! 583 vld1.32 {@Kxx[2]\[]},[r3,:32]! 584 vld1.32 {@Kxx[3]\[]},[r3,:32] 585 586.Loop_v8: 587 vld1.8 {@MSG[0]-@MSG[1]},[$inp]! 588 vld1.8 {@MSG[2]-@MSG[3]},[$inp]! 589 vrev32.8 @MSG[0],@MSG[0] 590 vrev32.8 @MSG[1],@MSG[1] 591 592 vadd.i32 $W0,@Kxx[0],@MSG[0] 593 vrev32.8 @MSG[2],@MSG[2] 594 vmov $ABCD_SAVE,$ABCD @ offload 595 subs $len,$len,#1 596 597 vadd.i32 $W1,@Kxx[0],@MSG[1] 598 vrev32.8 @MSG[3],@MSG[3] 599 sha1h $E1,$ABCD @ 0 600 sha1c $ABCD,$E,$W0 601 vadd.i32 $W0,@Kxx[$j],@MSG[2] 602 sha1su0 @MSG[0],@MSG[1],@MSG[2] 603___ 604for ($j=0,$i=1;$i<20-3;$i++) { 605my $f=("c","p","m","p")[$i/5]; 606$code.=<<___; 607 sha1h $E0,$ABCD @ $i 608 sha1$f $ABCD,$E1,$W1 609 vadd.i32 $W1,@Kxx[$j],@MSG[3] 610 sha1su1 @MSG[0],@MSG[3] 611___ 612$code.=<<___ if ($i<20-4); 613 sha1su0 @MSG[1],@MSG[2],@MSG[3] 614___ 615 ($E0,$E1)=($E1,$E0); ($W0,$W1)=($W1,$W0); 616 push(@MSG,shift(@MSG)); $j++ if ((($i+3)%5)==0); 617} 618$code.=<<___; 619 sha1h $E0,$ABCD @ $i 620 sha1p $ABCD,$E1,$W1 621 vadd.i32 $W1,@Kxx[$j],@MSG[3] 622 623 sha1h $E1,$ABCD @ 18 624 sha1p $ABCD,$E0,$W0 625 626 sha1h $E0,$ABCD @ 19 627 sha1p $ABCD,$E1,$W1 628 629 vadd.i32 $E,$E,$E0 630 vadd.i32 $ABCD,$ABCD,$ABCD_SAVE 631 bne .Loop_v8 632 633 vst1.32 {$ABCD},[$ctx]! 634 vst1.32 {$E\[0]},[$ctx] 635 636 vldmia sp!,{d8-d15} 637 ret @ bx lr 638.size sha1_block_data_order_armv8,.-sha1_block_data_order_armv8 639#endif 640___ 641}}} 642$code.=<<___; 643#if __ARM_MAX_ARCH__>=7 644.comm OPENSSL_armcap_P,4,4 645#endif 646___ 647 648{ my %opcode = ( 649 "sha1c" => 0xf2000c40, "sha1p" => 0xf2100c40, 650 "sha1m" => 0xf2200c40, "sha1su0" => 0xf2300c40, 651 "sha1h" => 0xf3b902c0, "sha1su1" => 0xf3ba0380 ); 652 653 sub unsha1 { 654 my ($mnemonic,$arg)=@_; 655 656 if ($arg =~ m/q([0-9]+)(?:,\s*q([0-9]+))?,\s*q([0-9]+)/o) { 657 my $word = $opcode{$mnemonic}|(($1&7)<<13)|(($1&8)<<19) 658 |(($2&7)<<17)|(($2&8)<<4) 659 |(($3&7)<<1) |(($3&8)<<2); 660 # since ARMv7 instructions are always encoded little-endian. 661 # correct solution is to use .inst directive, but older 662 # assemblers don't implement it:-( 663 sprintf ".byte\t0x%02x,0x%02x,0x%02x,0x%02x\t@ %s %s", 664 $word&0xff,($word>>8)&0xff, 665 ($word>>16)&0xff,($word>>24)&0xff, 666 $mnemonic,$arg; 667 } 668 } 669} 670 671foreach (split($/,$code)) { 672 s/{q([0-9]+)\[\]}/sprintf "{d%d[],d%d[]}",2*$1,2*$1+1/eo or 673 s/{q([0-9]+)\[0\]}/sprintf "{d%d[0]}",2*$1/eo; 674 675 s/\b(sha1\w+)\s+(q.*)/unsha1($1,$2)/geo; 676 677 s/\bret\b/bx lr/o or 678 s/\bbx\s+lr\b/.word\t0xe12fff1e/o; # make it possible to compile with -march=armv4 679 680 print $_,$/; 681} 682 683close STDOUT; # enforce flush 684