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# 2.22x RC4 tune-up:-) It should be noted though that my hand [as in 11# "hand-coded assembler"] doesn't stand for the whole improvement 12# coefficient. It turned out that eliminating RC4_CHAR from config 13# line results in ~40% improvement (yes, even for C implementation). 14# Presumably it has everything to do with AMD cache architecture and 15# RAW or whatever penalties. Once again! The module *requires* config 16# line *without* RC4_CHAR! As for coding "secret," I bet on partial 17# register arithmetics. For example instead of 'inc %r8; and $255,%r8' 18# I simply 'inc %r8b'. Even though optimization manual discourages 19# to operate on partial registers, it turned out to be the best bet. 20# At least for AMD... How IA32E would perform remains to be seen... 21 22# As was shown by Marc Bevand reordering of couple of load operations 23# results in even higher performance gain of 3.3x:-) At least on 24# Opteron... For reference, 1x in this case is RC4_CHAR C-code 25# compiled with gcc 3.3.2, which performs at ~54MBps per 1GHz clock. 26# Latter means that if you want to *estimate* what to expect from 27# *your* Opteron, then multiply 54 by 3.3 and clock frequency in GHz. 28 29# Intel P4 EM64T core was found to run the AMD64 code really slow... 30# The only way to achieve comparable performance on P4 was to keep 31# RC4_CHAR. Kind of ironic, huh? As it's apparently impossible to 32# compose blended code, which would perform even within 30% marginal 33# on either AMD and Intel platforms, I implement both cases. See 34# rc4_skey.c for further details... 35 36# P4 EM64T core appears to be "allergic" to 64-bit inc/dec. Replacing 37# those with add/sub results in 50% performance improvement of folded 38# loop... 39 40# As was shown by Zou Nanhai loop unrolling can improve Intel EM64T 41# performance by >30% [unlike P4 32-bit case that is]. But this is 42# provided that loads are reordered even more aggressively! Both code 43# pathes, AMD64 and EM64T, reorder loads in essentially same manner 44# as my IA-64 implementation. On Opteron this resulted in modest 5% 45# improvement [I had to test it], while final Intel P4 performance 46# achieves respectful 432MBps on 2.8GHz processor now. For reference. 47# If executed on Xeon, current RC4_CHAR code-path is 2.7x faster than 48# RC4_INT code-path. While if executed on Opteron, it's only 25% 49# slower than the RC4_INT one [meaning that if CPU �-arch detection 50# is not implemented, then this final RC4_CHAR code-path should be 51# preferred, as it provides better *all-round* performance]. 52 53# Intel Core2 was observed to perform poorly on both code paths:-( It 54# apparently suffers from some kind of partial register stall, which 55# occurs in 64-bit mode only [as virtually identical 32-bit loop was 56# observed to outperform 64-bit one by almost 50%]. Adding two movzb to 57# cloop1 boosts its performance by 80%! This loop appears to be optimal 58# fit for Core2 and therefore the code was modified to skip cloop8 on 59# this CPU. 60 61$flavour = shift; 62$output = shift; 63if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } 64 65$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); 66 67$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 68( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or 69( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or 70die "can't locate x86_64-xlate.pl"; 71 72open STDOUT,"| $^X $xlate $flavour $output"; 73 74$dat="%rdi"; # arg1 75$len="%rsi"; # arg2 76$inp="%rdx"; # arg3 77$out="%rcx"; # arg4 78 79@XX=("%r8","%r10"); 80@TX=("%r9","%r11"); 81$YY="%r12"; 82$TY="%r13"; 83 84$code=<<___; 85.text 86 87.globl RC4 88.type RC4,\@function,4 89.align 16 90RC4: or $len,$len 91 jne .Lentry 92 ret 93.Lentry: 94 push %rbx 95 push %r12 96 push %r13 97.Lprologue: 98 99 add \$8,$dat 100 movl -8($dat),$XX[0]#d 101 movl -4($dat),$YY#d 102 cmpl \$-1,256($dat) 103 je .LRC4_CHAR 104 inc $XX[0]#b 105 movl ($dat,$XX[0],4),$TX[0]#d 106 test \$-8,$len 107 jz .Lloop1 108 jmp .Lloop8 109.align 16 110.Lloop8: 111___ 112for ($i=0;$i<8;$i++) { 113$code.=<<___; 114 add $TX[0]#b,$YY#b 115 mov $XX[0],$XX[1] 116 movl ($dat,$YY,4),$TY#d 117 ror \$8,%rax # ror is redundant when $i=0 118 inc $XX[1]#b 119 movl ($dat,$XX[1],4),$TX[1]#d 120 cmp $XX[1],$YY 121 movl $TX[0]#d,($dat,$YY,4) 122 cmove $TX[0],$TX[1] 123 movl $TY#d,($dat,$XX[0],4) 124 add $TX[0]#b,$TY#b 125 movb ($dat,$TY,4),%al 126___ 127push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers 128} 129$code.=<<___; 130 ror \$8,%rax 131 sub \$8,$len 132 133 xor ($inp),%rax 134 add \$8,$inp 135 mov %rax,($out) 136 add \$8,$out 137 138 test \$-8,$len 139 jnz .Lloop8 140 cmp \$0,$len 141 jne .Lloop1 142 jmp .Lexit 143 144.align 16 145.Lloop1: 146 add $TX[0]#b,$YY#b 147 movl ($dat,$YY,4),$TY#d 148 movl $TX[0]#d,($dat,$YY,4) 149 movl $TY#d,($dat,$XX[0],4) 150 add $TY#b,$TX[0]#b 151 inc $XX[0]#b 152 movl ($dat,$TX[0],4),$TY#d 153 movl ($dat,$XX[0],4),$TX[0]#d 154 xorb ($inp),$TY#b 155 inc $inp 156 movb $TY#b,($out) 157 inc $out 158 dec $len 159 jnz .Lloop1 160 jmp .Lexit 161 162.align 16 163.LRC4_CHAR: 164 add \$1,$XX[0]#b 165 movzb ($dat,$XX[0]),$TX[0]#d 166 test \$-8,$len 167 jz .Lcloop1 168 cmpl \$0,260($dat) 169 jnz .Lcloop1 170 jmp .Lcloop8 171.align 16 172.Lcloop8: 173 mov ($inp),%eax 174 mov 4($inp),%ebx 175___ 176# unroll 2x4-wise, because 64-bit rotates kill Intel P4... 177for ($i=0;$i<4;$i++) { 178$code.=<<___; 179 add $TX[0]#b,$YY#b 180 lea 1($XX[0]),$XX[1] 181 movzb ($dat,$YY),$TY#d 182 movzb $XX[1]#b,$XX[1]#d 183 movzb ($dat,$XX[1]),$TX[1]#d 184 movb $TX[0]#b,($dat,$YY) 185 cmp $XX[1],$YY 186 movb $TY#b,($dat,$XX[0]) 187 jne .Lcmov$i # Intel cmov is sloooow... 188 mov $TX[0],$TX[1] 189.Lcmov$i: 190 add $TX[0]#b,$TY#b 191 xor ($dat,$TY),%al 192 ror \$8,%eax 193___ 194push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers 195} 196for ($i=4;$i<8;$i++) { 197$code.=<<___; 198 add $TX[0]#b,$YY#b 199 lea 1($XX[0]),$XX[1] 200 movzb ($dat,$YY),$TY#d 201 movzb $XX[1]#b,$XX[1]#d 202 movzb ($dat,$XX[1]),$TX[1]#d 203 movb $TX[0]#b,($dat,$YY) 204 cmp $XX[1],$YY 205 movb $TY#b,($dat,$XX[0]) 206 jne .Lcmov$i # Intel cmov is sloooow... 207 mov $TX[0],$TX[1] 208.Lcmov$i: 209 add $TX[0]#b,$TY#b 210 xor ($dat,$TY),%bl 211 ror \$8,%ebx 212___ 213push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers 214} 215$code.=<<___; 216 lea -8($len),$len 217 mov %eax,($out) 218 lea 8($inp),$inp 219 mov %ebx,4($out) 220 lea 8($out),$out 221 222 test \$-8,$len 223 jnz .Lcloop8 224 cmp \$0,$len 225 jne .Lcloop1 226 jmp .Lexit 227___ 228$code.=<<___; 229.align 16 230.Lcloop1: 231 add $TX[0]#b,$YY#b 232 movzb ($dat,$YY),$TY#d 233 movb $TX[0]#b,($dat,$YY) 234 movb $TY#b,($dat,$XX[0]) 235 add $TX[0]#b,$TY#b 236 add \$1,$XX[0]#b 237 movzb $TY#b,$TY#d 238 movzb $XX[0]#b,$XX[0]#d 239 movzb ($dat,$TY),$TY#d 240 movzb ($dat,$XX[0]),$TX[0]#d 241 xorb ($inp),$TY#b 242 lea 1($inp),$inp 243 movb $TY#b,($out) 244 lea 1($out),$out 245 sub \$1,$len 246 jnz .Lcloop1 247 jmp .Lexit 248 249.align 16 250.Lexit: 251 sub \$1,$XX[0]#b 252 movl $XX[0]#d,-8($dat) 253 movl $YY#d,-4($dat) 254 255 mov (%rsp),%r13 256 mov 8(%rsp),%r12 257 mov 16(%rsp),%rbx 258 add \$24,%rsp 259.Lepilogue: 260 ret 261.size RC4,.-RC4 262___ 263 264$idx="%r8"; 265$ido="%r9"; 266 267$code.=<<___; 268.extern OPENSSL_ia32cap_P 269.globl RC4_set_key 270.type RC4_set_key,\@function,3 271.align 16 272RC4_set_key: 273 lea 8($dat),$dat 274 lea ($inp,$len),$inp 275 neg $len 276 mov $len,%rcx 277 xor %eax,%eax 278 xor $ido,$ido 279 xor %r10,%r10 280 xor %r11,%r11 281 282 mov OPENSSL_ia32cap_P(%rip),$idx#d 283 bt \$20,$idx#d 284 jnc .Lw1stloop 285 bt \$30,$idx#d 286 setc $ido#b 287 mov $ido#d,260($dat) 288 jmp .Lc1stloop 289 290.align 16 291.Lw1stloop: 292 mov %eax,($dat,%rax,4) 293 add \$1,%al 294 jnc .Lw1stloop 295 296 xor $ido,$ido 297 xor $idx,$idx 298.align 16 299.Lw2ndloop: 300 mov ($dat,$ido,4),%r10d 301 add ($inp,$len,1),$idx#b 302 add %r10b,$idx#b 303 add \$1,$len 304 mov ($dat,$idx,4),%r11d 305 cmovz %rcx,$len 306 mov %r10d,($dat,$idx,4) 307 mov %r11d,($dat,$ido,4) 308 add \$1,$ido#b 309 jnc .Lw2ndloop 310 jmp .Lexit_key 311 312.align 16 313.Lc1stloop: 314 mov %al,($dat,%rax) 315 add \$1,%al 316 jnc .Lc1stloop 317 318 xor $ido,$ido 319 xor $idx,$idx 320.align 16 321.Lc2ndloop: 322 mov ($dat,$ido),%r10b 323 add ($inp,$len),$idx#b 324 add %r10b,$idx#b 325 add \$1,$len 326 mov ($dat,$idx),%r11b 327 jnz .Lcnowrap 328 mov %rcx,$len 329.Lcnowrap: 330 mov %r10b,($dat,$idx) 331 mov %r11b,($dat,$ido) 332 add \$1,$ido#b 333 jnc .Lc2ndloop 334 movl \$-1,256($dat) 335 336.align 16 337.Lexit_key: 338 xor %eax,%eax 339 mov %eax,-8($dat) 340 mov %eax,-4($dat) 341 ret 342.size RC4_set_key,.-RC4_set_key 343 344.globl RC4_options 345.type RC4_options,\@abi-omnipotent 346.align 16 347RC4_options: 348 lea .Lopts(%rip),%rax 349 mov OPENSSL_ia32cap_P(%rip),%edx 350 bt \$20,%edx 351 jnc .Ldone 352 add \$12,%rax 353 bt \$30,%edx 354 jnc .Ldone 355 add \$13,%rax 356.Ldone: 357 ret 358.align 64 359.Lopts: 360.asciz "rc4(8x,int)" 361.asciz "rc4(8x,char)" 362.asciz "rc4(1x,char)" 363.asciz "RC4 for x86_64, CRYPTOGAMS by <appro\@openssl.org>" 364.align 64 365.size RC4_options,.-RC4_options 366___ 367 368# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, 369# CONTEXT *context,DISPATCHER_CONTEXT *disp) 370if ($win64) { 371$rec="%rcx"; 372$frame="%rdx"; 373$context="%r8"; 374$disp="%r9"; 375 376$code.=<<___; 377.extern __imp_RtlVirtualUnwind 378.type stream_se_handler,\@abi-omnipotent 379.align 16 380stream_se_handler: 381 push %rsi 382 push %rdi 383 push %rbx 384 push %rbp 385 push %r12 386 push %r13 387 push %r14 388 push %r15 389 pushfq 390 sub \$64,%rsp 391 392 mov 120($context),%rax # pull context->Rax 393 mov 248($context),%rbx # pull context->Rip 394 395 lea .Lprologue(%rip),%r10 396 cmp %r10,%rbx # context->Rip<prologue label 397 jb .Lin_prologue 398 399 mov 152($context),%rax # pull context->Rsp 400 401 lea .Lepilogue(%rip),%r10 402 cmp %r10,%rbx # context->Rip>=epilogue label 403 jae .Lin_prologue 404 405 lea 24(%rax),%rax 406 407 mov -8(%rax),%rbx 408 mov -16(%rax),%r12 409 mov -24(%rax),%r13 410 mov %rbx,144($context) # restore context->Rbx 411 mov %r12,216($context) # restore context->R12 412 mov %r13,224($context) # restore context->R13 413 414.Lin_prologue: 415 mov 8(%rax),%rdi 416 mov 16(%rax),%rsi 417 mov %rax,152($context) # restore context->Rsp 418 mov %rsi,168($context) # restore context->Rsi 419 mov %rdi,176($context) # restore context->Rdi 420 421 jmp .Lcommon_seh_exit 422.size stream_se_handler,.-stream_se_handler 423 424.type key_se_handler,\@abi-omnipotent 425.align 16 426key_se_handler: 427 push %rsi 428 push %rdi 429 push %rbx 430 push %rbp 431 push %r12 432 push %r13 433 push %r14 434 push %r15 435 pushfq 436 sub \$64,%rsp 437 438 mov 152($context),%rax # pull context->Rsp 439 mov 8(%rax),%rdi 440 mov 16(%rax),%rsi 441 mov %rsi,168($context) # restore context->Rsi 442 mov %rdi,176($context) # restore context->Rdi 443 444.Lcommon_seh_exit: 445 446 mov 40($disp),%rdi # disp->ContextRecord 447 mov $context,%rsi # context 448 mov \$154,%ecx # sizeof(CONTEXT) 449 .long 0xa548f3fc # cld; rep movsq 450 451 mov $disp,%rsi 452 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER 453 mov 8(%rsi),%rdx # arg2, disp->ImageBase 454 mov 0(%rsi),%r8 # arg3, disp->ControlPc 455 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry 456 mov 40(%rsi),%r10 # disp->ContextRecord 457 lea 56(%rsi),%r11 # &disp->HandlerData 458 lea 24(%rsi),%r12 # &disp->EstablisherFrame 459 mov %r10,32(%rsp) # arg5 460 mov %r11,40(%rsp) # arg6 461 mov %r12,48(%rsp) # arg7 462 mov %rcx,56(%rsp) # arg8, (NULL) 463 call *__imp_RtlVirtualUnwind(%rip) 464 465 mov \$1,%eax # ExceptionContinueSearch 466 add \$64,%rsp 467 popfq 468 pop %r15 469 pop %r14 470 pop %r13 471 pop %r12 472 pop %rbp 473 pop %rbx 474 pop %rdi 475 pop %rsi 476 ret 477.size key_se_handler,.-key_se_handler 478 479.section .pdata 480.align 4 481 .rva .LSEH_begin_RC4 482 .rva .LSEH_end_RC4 483 .rva .LSEH_info_RC4 484 485 .rva .LSEH_begin_RC4_set_key 486 .rva .LSEH_end_RC4_set_key 487 .rva .LSEH_info_RC4_set_key 488 489.section .xdata 490.align 8 491.LSEH_info_RC4: 492 .byte 9,0,0,0 493 .rva stream_se_handler 494.LSEH_info_RC4_set_key: 495 .byte 9,0,0,0 496 .rva key_se_handler 497___ 498} 499 500$code =~ s/#([bwd])/$1/gm; 501 502print $code; 503 504close STDOUT; 505