1#!/usr/bin/env perl 2# 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# ==================================================================== 9# 10# July 2004 11# 12# 2.22x RC4 tune-up:-) It should be noted though that my hand [as in 13# "hand-coded assembler"] doesn't stand for the whole improvement 14# coefficient. It turned out that eliminating RC4_CHAR from config 15# line results in ~40% improvement (yes, even for C implementation). 16# Presumably it has everything to do with AMD cache architecture and 17# RAW or whatever penalties. Once again! The module *requires* config 18# line *without* RC4_CHAR! As for coding "secret," I bet on partial 19# register arithmetics. For example instead of 'inc %r8; and $255,%r8' 20# I simply 'inc %r8b'. Even though optimization manual discourages 21# to operate on partial registers, it turned out to be the best bet. 22# At least for AMD... How IA32E would perform remains to be seen... 23 24# November 2004 25# 26# As was shown by Marc Bevand reordering of couple of load operations 27# results in even higher performance gain of 3.3x:-) At least on 28# Opteron... For reference, 1x in this case is RC4_CHAR C-code 29# compiled with gcc 3.3.2, which performs at ~54MBps per 1GHz clock. 30# Latter means that if you want to *estimate* what to expect from 31# *your* Opteron, then multiply 54 by 3.3 and clock frequency in GHz. 32 33# November 2004 34# 35# Intel P4 EM64T core was found to run the AMD64 code really slow... 36# The only way to achieve comparable performance on P4 was to keep 37# RC4_CHAR. Kind of ironic, huh? As it's apparently impossible to 38# compose blended code, which would perform even within 30% marginal 39# on either AMD and Intel platforms, I implement both cases. See 40# rc4_skey.c for further details... 41 42# April 2005 43# 44# P4 EM64T core appears to be "allergic" to 64-bit inc/dec. Replacing 45# those with add/sub results in 50% performance improvement of folded 46# loop... 47 48# May 2005 49# 50# As was shown by Zou Nanhai loop unrolling can improve Intel EM64T 51# performance by >30% [unlike P4 32-bit case that is]. But this is 52# provided that loads are reordered even more aggressively! Both code 53# pathes, AMD64 and EM64T, reorder loads in essentially same manner 54# as my IA-64 implementation. On Opteron this resulted in modest 5% 55# improvement [I had to test it], while final Intel P4 performance 56# achieves respectful 432MBps on 2.8GHz processor now. For reference. 57# If executed on Xeon, current RC4_CHAR code-path is 2.7x faster than 58# RC4_INT code-path. While if executed on Opteron, it's only 25% 59# slower than the RC4_INT one [meaning that if CPU �-arch detection 60# is not implemented, then this final RC4_CHAR code-path should be 61# preferred, as it provides better *all-round* performance]. 62 63# March 2007 64# 65# Intel Core2 was observed to perform poorly on both code paths:-( It 66# apparently suffers from some kind of partial register stall, which 67# occurs in 64-bit mode only [as virtually identical 32-bit loop was 68# observed to outperform 64-bit one by almost 50%]. Adding two movzb to 69# cloop1 boosts its performance by 80%! This loop appears to be optimal 70# fit for Core2 and therefore the code was modified to skip cloop8 on 71# this CPU. 72 73# May 2010 74# 75# Intel Westmere was observed to perform suboptimally. Adding yet 76# another movzb to cloop1 improved performance by almost 50%! Core2 77# performance is improved too, but nominally... 78 79# May 2011 80# 81# The only code path that was not modified is P4-specific one. Non-P4 82# Intel code path optimization is heavily based on submission by Maxim 83# Perminov, Maxim Locktyukhin and Jim Guilford of Intel. I've used 84# some of the ideas even in attempt to optmize the original RC4_INT 85# code path... Current performance in cycles per processed byte (less 86# is better) and improvement coefficients relative to previous 87# version of this module are: 88# 89# Opteron 5.3/+0%(*) 90# P4 6.5 91# Core2 6.2/+15%(**) 92# Westmere 4.2/+60% 93# Sandy Bridge 4.2/+120% 94# Atom 9.3/+80% 95# 96# (*) But corresponding loop has less instructions, which should have 97# positive effect on upcoming Bulldozer, which has one less ALU. 98# For reference, Intel code runs at 6.8 cpb rate on Opteron. 99# (**) Note that Core2 result is ~15% lower than corresponding result 100# for 32-bit code, meaning that it's possible to improve it, 101# but more than likely at the cost of the others (see rc4-586.pl 102# to get the idea)... 103 104$flavour = shift; 105$output = shift; 106if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } 107 108$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); 109 110$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 111( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or 112( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or 113die "can't locate x86_64-xlate.pl"; 114 115open STDOUT,"| $^X $xlate $flavour $output"; 116 117$dat="%rdi"; # arg1 118$len="%rsi"; # arg2 119$inp="%rdx"; # arg3 120$out="%rcx"; # arg4 121 122{ 123$code=<<___; 124.text 125.extern OPENSSL_ia32cap_P 126 127.globl RC4 128.type RC4,\@function,4 129.align 16 130RC4: or $len,$len 131 jne .Lentry 132 ret 133.Lentry: 134 push %rbx 135 push %r12 136 push %r13 137.Lprologue: 138 mov $len,%r11 139 mov $inp,%r12 140 mov $out,%r13 141___ 142my $len="%r11"; # reassign input arguments 143my $inp="%r12"; 144my $out="%r13"; 145 146my @XX=("%r10","%rsi"); 147my @TX=("%rax","%rbx"); 148my $YY="%rcx"; 149my $TY="%rdx"; 150 151$code.=<<___; 152 xor $XX[0],$XX[0] 153 xor $YY,$YY 154 155 lea 8($dat),$dat 156 mov -8($dat),$XX[0]#b 157 mov -4($dat),$YY#b 158 cmpl \$-1,256($dat) 159 je .LRC4_CHAR 160 mov OPENSSL_ia32cap_P(%rip),%r8d 161 xor $TX[1],$TX[1] 162 inc $XX[0]#b 163 sub $XX[0],$TX[1] 164 sub $inp,$out 165 movl ($dat,$XX[0],4),$TX[0]#d 166 test \$-16,$len 167 jz .Lloop1 168 bt \$30,%r8d # Intel CPU? 169 jc .Lintel 170 and \$7,$TX[1] 171 lea 1($XX[0]),$XX[1] 172 jz .Loop8 173 sub $TX[1],$len 174.Loop8_warmup: 175 add $TX[0]#b,$YY#b 176 movl ($dat,$YY,4),$TY#d 177 movl $TX[0]#d,($dat,$YY,4) 178 movl $TY#d,($dat,$XX[0],4) 179 add $TY#b,$TX[0]#b 180 inc $XX[0]#b 181 movl ($dat,$TX[0],4),$TY#d 182 movl ($dat,$XX[0],4),$TX[0]#d 183 xorb ($inp),$TY#b 184 movb $TY#b,($out,$inp) 185 lea 1($inp),$inp 186 dec $TX[1] 187 jnz .Loop8_warmup 188 189 lea 1($XX[0]),$XX[1] 190 jmp .Loop8 191.align 16 192.Loop8: 193___ 194for ($i=0;$i<8;$i++) { 195$code.=<<___ if ($i==7); 196 add \$8,$XX[1]#b 197___ 198$code.=<<___; 199 add $TX[0]#b,$YY#b 200 movl ($dat,$YY,4),$TY#d 201 movl $TX[0]#d,($dat,$YY,4) 202 movl `4*($i==7?-1:$i)`($dat,$XX[1],4),$TX[1]#d 203 ror \$8,%r8 # ror is redundant when $i=0 204 movl $TY#d,4*$i($dat,$XX[0],4) 205 add $TX[0]#b,$TY#b 206 movb ($dat,$TY,4),%r8b 207___ 208push(@TX,shift(@TX)); #push(@XX,shift(@XX)); # "rotate" registers 209} 210$code.=<<___; 211 add \$8,$XX[0]#b 212 ror \$8,%r8 213 sub \$8,$len 214 215 xor ($inp),%r8 216 mov %r8,($out,$inp) 217 lea 8($inp),$inp 218 219 test \$-8,$len 220 jnz .Loop8 221 cmp \$0,$len 222 jne .Lloop1 223 jmp .Lexit 224 225.align 16 226.Lintel: 227 test \$-32,$len 228 jz .Lloop1 229 and \$15,$TX[1] 230 jz .Loop16_is_hot 231 sub $TX[1],$len 232.Loop16_warmup: 233 add $TX[0]#b,$YY#b 234 movl ($dat,$YY,4),$TY#d 235 movl $TX[0]#d,($dat,$YY,4) 236 movl $TY#d,($dat,$XX[0],4) 237 add $TY#b,$TX[0]#b 238 inc $XX[0]#b 239 movl ($dat,$TX[0],4),$TY#d 240 movl ($dat,$XX[0],4),$TX[0]#d 241 xorb ($inp),$TY#b 242 movb $TY#b,($out,$inp) 243 lea 1($inp),$inp 244 dec $TX[1] 245 jnz .Loop16_warmup 246 247 mov $YY,$TX[1] 248 xor $YY,$YY 249 mov $TX[1]#b,$YY#b 250 251.Loop16_is_hot: 252 lea ($dat,$XX[0],4),$XX[1] 253___ 254sub RC4_loop { 255 my $i=shift; 256 my $j=$i<0?0:$i; 257 my $xmm="%xmm".($j&1); 258 259 $code.=" add \$16,$XX[0]#b\n" if ($i==15); 260 $code.=" movdqu ($inp),%xmm2\n" if ($i==15); 261 $code.=" add $TX[0]#b,$YY#b\n" if ($i<=0); 262 $code.=" movl ($dat,$YY,4),$TY#d\n"; 263 $code.=" pxor %xmm0,%xmm2\n" if ($i==0); 264 $code.=" psllq \$8,%xmm1\n" if ($i==0); 265 $code.=" pxor $xmm,$xmm\n" if ($i<=1); 266 $code.=" movl $TX[0]#d,($dat,$YY,4)\n"; 267 $code.=" add $TY#b,$TX[0]#b\n"; 268 $code.=" movl `4*($j+1)`($XX[1]),$TX[1]#d\n" if ($i<15); 269 $code.=" movz $TX[0]#b,$TX[0]#d\n"; 270 $code.=" movl $TY#d,4*$j($XX[1])\n"; 271 $code.=" pxor %xmm1,%xmm2\n" if ($i==0); 272 $code.=" lea ($dat,$XX[0],4),$XX[1]\n" if ($i==15); 273 $code.=" add $TX[1]#b,$YY#b\n" if ($i<15); 274 $code.=" pinsrw \$`($j>>1)&7`,($dat,$TX[0],4),$xmm\n"; 275 $code.=" movdqu %xmm2,($out,$inp)\n" if ($i==0); 276 $code.=" lea 16($inp),$inp\n" if ($i==0); 277 $code.=" movl ($XX[1]),$TX[1]#d\n" if ($i==15); 278} 279 RC4_loop(-1); 280$code.=<<___; 281 jmp .Loop16_enter 282.align 16 283.Loop16: 284___ 285 286for ($i=0;$i<16;$i++) { 287 $code.=".Loop16_enter:\n" if ($i==1); 288 RC4_loop($i); 289 push(@TX,shift(@TX)); # "rotate" registers 290} 291$code.=<<___; 292 mov $YY,$TX[1] 293 xor $YY,$YY # keyword to partial register 294 sub \$16,$len 295 mov $TX[1]#b,$YY#b 296 test \$-16,$len 297 jnz .Loop16 298 299 psllq \$8,%xmm1 300 pxor %xmm0,%xmm2 301 pxor %xmm1,%xmm2 302 movdqu %xmm2,($out,$inp) 303 lea 16($inp),$inp 304 305 cmp \$0,$len 306 jne .Lloop1 307 jmp .Lexit 308 309.align 16 310.Lloop1: 311 add $TX[0]#b,$YY#b 312 movl ($dat,$YY,4),$TY#d 313 movl $TX[0]#d,($dat,$YY,4) 314 movl $TY#d,($dat,$XX[0],4) 315 add $TY#b,$TX[0]#b 316 inc $XX[0]#b 317 movl ($dat,$TX[0],4),$TY#d 318 movl ($dat,$XX[0],4),$TX[0]#d 319 xorb ($inp),$TY#b 320 movb $TY#b,($out,$inp) 321 lea 1($inp),$inp 322 dec $len 323 jnz .Lloop1 324 jmp .Lexit 325 326.align 16 327.LRC4_CHAR: 328 add \$1,$XX[0]#b 329 movzb ($dat,$XX[0]),$TX[0]#d 330 test \$-8,$len 331 jz .Lcloop1 332 jmp .Lcloop8 333.align 16 334.Lcloop8: 335 mov ($inp),%r8d 336 mov 4($inp),%r9d 337___ 338# unroll 2x4-wise, because 64-bit rotates kill Intel P4... 339for ($i=0;$i<4;$i++) { 340$code.=<<___; 341 add $TX[0]#b,$YY#b 342 lea 1($XX[0]),$XX[1] 343 movzb ($dat,$YY),$TY#d 344 movzb $XX[1]#b,$XX[1]#d 345 movzb ($dat,$XX[1]),$TX[1]#d 346 movb $TX[0]#b,($dat,$YY) 347 cmp $XX[1],$YY 348 movb $TY#b,($dat,$XX[0]) 349 jne .Lcmov$i # Intel cmov is sloooow... 350 mov $TX[0],$TX[1] 351.Lcmov$i: 352 add $TX[0]#b,$TY#b 353 xor ($dat,$TY),%r8b 354 ror \$8,%r8d 355___ 356push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers 357} 358for ($i=4;$i<8;$i++) { 359$code.=<<___; 360 add $TX[0]#b,$YY#b 361 lea 1($XX[0]),$XX[1] 362 movzb ($dat,$YY),$TY#d 363 movzb $XX[1]#b,$XX[1]#d 364 movzb ($dat,$XX[1]),$TX[1]#d 365 movb $TX[0]#b,($dat,$YY) 366 cmp $XX[1],$YY 367 movb $TY#b,($dat,$XX[0]) 368 jne .Lcmov$i # Intel cmov is sloooow... 369 mov $TX[0],$TX[1] 370.Lcmov$i: 371 add $TX[0]#b,$TY#b 372 xor ($dat,$TY),%r9b 373 ror \$8,%r9d 374___ 375push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers 376} 377$code.=<<___; 378 lea -8($len),$len 379 mov %r8d,($out) 380 lea 8($inp),$inp 381 mov %r9d,4($out) 382 lea 8($out),$out 383 384 test \$-8,$len 385 jnz .Lcloop8 386 cmp \$0,$len 387 jne .Lcloop1 388 jmp .Lexit 389___ 390$code.=<<___; 391.align 16 392.Lcloop1: 393 add $TX[0]#b,$YY#b 394 movzb $YY#b,$YY#d 395 movzb ($dat,$YY),$TY#d 396 movb $TX[0]#b,($dat,$YY) 397 movb $TY#b,($dat,$XX[0]) 398 add $TX[0]#b,$TY#b 399 add \$1,$XX[0]#b 400 movzb $TY#b,$TY#d 401 movzb $XX[0]#b,$XX[0]#d 402 movzb ($dat,$TY),$TY#d 403 movzb ($dat,$XX[0]),$TX[0]#d 404 xorb ($inp),$TY#b 405 lea 1($inp),$inp 406 movb $TY#b,($out) 407 lea 1($out),$out 408 sub \$1,$len 409 jnz .Lcloop1 410 jmp .Lexit 411 412.align 16 413.Lexit: 414 sub \$1,$XX[0]#b 415 movl $XX[0]#d,-8($dat) 416 movl $YY#d,-4($dat) 417 418 mov (%rsp),%r13 419 mov 8(%rsp),%r12 420 mov 16(%rsp),%rbx 421 add \$24,%rsp 422.Lepilogue: 423 ret 424.size RC4,.-RC4 425___ 426} 427 428$idx="%r8"; 429$ido="%r9"; 430 431$code.=<<___; 432.globl private_RC4_set_key 433.type private_RC4_set_key,\@function,3 434.align 16 435private_RC4_set_key: 436 lea 8($dat),$dat 437 lea ($inp,$len),$inp 438 neg $len 439 mov $len,%rcx 440 xor %eax,%eax 441 xor $ido,$ido 442 xor %r10,%r10 443 xor %r11,%r11 444 445 mov OPENSSL_ia32cap_P(%rip),$idx#d 446 bt \$20,$idx#d # RC4_CHAR? 447 jc .Lc1stloop 448 jmp .Lw1stloop 449 450.align 16 451.Lw1stloop: 452 mov %eax,($dat,%rax,4) 453 add \$1,%al 454 jnc .Lw1stloop 455 456 xor $ido,$ido 457 xor $idx,$idx 458.align 16 459.Lw2ndloop: 460 mov ($dat,$ido,4),%r10d 461 add ($inp,$len,1),$idx#b 462 add %r10b,$idx#b 463 add \$1,$len 464 mov ($dat,$idx,4),%r11d 465 cmovz %rcx,$len 466 mov %r10d,($dat,$idx,4) 467 mov %r11d,($dat,$ido,4) 468 add \$1,$ido#b 469 jnc .Lw2ndloop 470 jmp .Lexit_key 471 472.align 16 473.Lc1stloop: 474 mov %al,($dat,%rax) 475 add \$1,%al 476 jnc .Lc1stloop 477 478 xor $ido,$ido 479 xor $idx,$idx 480.align 16 481.Lc2ndloop: 482 mov ($dat,$ido),%r10b 483 add ($inp,$len),$idx#b 484 add %r10b,$idx#b 485 add \$1,$len 486 mov ($dat,$idx),%r11b 487 jnz .Lcnowrap 488 mov %rcx,$len 489.Lcnowrap: 490 mov %r10b,($dat,$idx) 491 mov %r11b,($dat,$ido) 492 add \$1,$ido#b 493 jnc .Lc2ndloop 494 movl \$-1,256($dat) 495 496.align 16 497.Lexit_key: 498 xor %eax,%eax 499 mov %eax,-8($dat) 500 mov %eax,-4($dat) 501 ret 502.size private_RC4_set_key,.-private_RC4_set_key 503 504.globl RC4_options 505.type RC4_options,\@abi-omnipotent 506.align 16 507RC4_options: 508 lea .Lopts(%rip),%rax 509 mov OPENSSL_ia32cap_P(%rip),%edx 510 bt \$20,%edx 511 jc .L8xchar 512 bt \$30,%edx 513 jnc .Ldone 514 add \$25,%rax 515 ret 516.L8xchar: 517 add \$12,%rax 518.Ldone: 519 ret 520.align 64 521.Lopts: 522.asciz "rc4(8x,int)" 523.asciz "rc4(8x,char)" 524.asciz "rc4(16x,int)" 525.asciz "RC4 for x86_64, CRYPTOGAMS by <appro\@openssl.org>" 526.align 64 527.size RC4_options,.-RC4_options 528___ 529 530# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, 531# CONTEXT *context,DISPATCHER_CONTEXT *disp) 532if ($win64) { 533$rec="%rcx"; 534$frame="%rdx"; 535$context="%r8"; 536$disp="%r9"; 537 538$code.=<<___; 539.extern __imp_RtlVirtualUnwind 540.type stream_se_handler,\@abi-omnipotent 541.align 16 542stream_se_handler: 543 push %rsi 544 push %rdi 545 push %rbx 546 push %rbp 547 push %r12 548 push %r13 549 push %r14 550 push %r15 551 pushfq 552 sub \$64,%rsp 553 554 mov 120($context),%rax # pull context->Rax 555 mov 248($context),%rbx # pull context->Rip 556 557 lea .Lprologue(%rip),%r10 558 cmp %r10,%rbx # context->Rip<prologue label 559 jb .Lin_prologue 560 561 mov 152($context),%rax # pull context->Rsp 562 563 lea .Lepilogue(%rip),%r10 564 cmp %r10,%rbx # context->Rip>=epilogue label 565 jae .Lin_prologue 566 567 lea 24(%rax),%rax 568 569 mov -8(%rax),%rbx 570 mov -16(%rax),%r12 571 mov -24(%rax),%r13 572 mov %rbx,144($context) # restore context->Rbx 573 mov %r12,216($context) # restore context->R12 574 mov %r13,224($context) # restore context->R13 575 576.Lin_prologue: 577 mov 8(%rax),%rdi 578 mov 16(%rax),%rsi 579 mov %rax,152($context) # restore context->Rsp 580 mov %rsi,168($context) # restore context->Rsi 581 mov %rdi,176($context) # restore context->Rdi 582 583 jmp .Lcommon_seh_exit 584.size stream_se_handler,.-stream_se_handler 585 586.type key_se_handler,\@abi-omnipotent 587.align 16 588key_se_handler: 589 push %rsi 590 push %rdi 591 push %rbx 592 push %rbp 593 push %r12 594 push %r13 595 push %r14 596 push %r15 597 pushfq 598 sub \$64,%rsp 599 600 mov 152($context),%rax # pull context->Rsp 601 mov 8(%rax),%rdi 602 mov 16(%rax),%rsi 603 mov %rsi,168($context) # restore context->Rsi 604 mov %rdi,176($context) # restore context->Rdi 605 606.Lcommon_seh_exit: 607 608 mov 40($disp),%rdi # disp->ContextRecord 609 mov $context,%rsi # context 610 mov \$154,%ecx # sizeof(CONTEXT) 611 .long 0xa548f3fc # cld; rep movsq 612 613 mov $disp,%rsi 614 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER 615 mov 8(%rsi),%rdx # arg2, disp->ImageBase 616 mov 0(%rsi),%r8 # arg3, disp->ControlPc 617 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry 618 mov 40(%rsi),%r10 # disp->ContextRecord 619 lea 56(%rsi),%r11 # &disp->HandlerData 620 lea 24(%rsi),%r12 # &disp->EstablisherFrame 621 mov %r10,32(%rsp) # arg5 622 mov %r11,40(%rsp) # arg6 623 mov %r12,48(%rsp) # arg7 624 mov %rcx,56(%rsp) # arg8, (NULL) 625 call *__imp_RtlVirtualUnwind(%rip) 626 627 mov \$1,%eax # ExceptionContinueSearch 628 add \$64,%rsp 629 popfq 630 pop %r15 631 pop %r14 632 pop %r13 633 pop %r12 634 pop %rbp 635 pop %rbx 636 pop %rdi 637 pop %rsi 638 ret 639.size key_se_handler,.-key_se_handler 640 641.section .pdata 642.align 4 643 .rva .LSEH_begin_RC4 644 .rva .LSEH_end_RC4 645 .rva .LSEH_info_RC4 646 647 .rva .LSEH_begin_private_RC4_set_key 648 .rva .LSEH_end_private_RC4_set_key 649 .rva .LSEH_info_private_RC4_set_key 650 651.section .xdata 652.align 8 653.LSEH_info_RC4: 654 .byte 9,0,0,0 655 .rva stream_se_handler 656.LSEH_info_private_RC4_set_key: 657 .byte 9,0,0,0 658 .rva key_se_handler 659___ 660} 661 662sub reg_part { 663my ($reg,$conv)=@_; 664 if ($reg =~ /%r[0-9]+/) { $reg .= $conv; } 665 elsif ($conv eq "b") { $reg =~ s/%[er]([^x]+)x?/%$1l/; } 666 elsif ($conv eq "w") { $reg =~ s/%[er](.+)/%$1/; } 667 elsif ($conv eq "d") { $reg =~ s/%[er](.+)/%e$1/; } 668 return $reg; 669} 670 671$code =~ s/(%[a-z0-9]+)#([bwd])/reg_part($1,$2)/gem; 672$code =~ s/\`([^\`]*)\`/eval $1/gem; 673 674print $code; 675 676close STDOUT; 677