X-Git-Url: https://git.openssl.org/gitweb/?p=openssl.git;a=blobdiff_plain;f=crypto%2Frc4%2Fasm%2Frc4-x86_64.pl;h=9ccb70eeb47b98e0f61ce560adbd2739951fd22a;hp=36a9429ef73dfa664124b15762fd10dc2e0c2d16;hb=2dfb52d3968e838876406d47861488324d5990b4;hpb=9babf3929bf1f546aa646d9e1e2a934ccfe0b333 diff --git a/crypto/rc4/asm/rc4-x86_64.pl b/crypto/rc4/asm/rc4-x86_64.pl index 36a9429ef7..9ccb70eeb4 100755 --- a/crypto/rc4/asm/rc4-x86_64.pl +++ b/crypto/rc4/asm/rc4-x86_64.pl @@ -1,11 +1,21 @@ -#!/usr/bin/env perl +#! /usr/bin/env perl +# Copyright 2005-2016 The OpenSSL Project Authors. All Rights Reserved. +# +# Licensed under the OpenSSL license (the "License"). You may not use +# this file except in compliance with the License. You can obtain a copy +# in the file LICENSE in the source distribution or at +# https://www.openssl.org/source/license.html + # # ==================================================================== # Written by Andy Polyakov for the OpenSSL -# project. Rights for redistribution and usage in source and binary -# forms are granted according to the OpenSSL license. +# project. The module is, however, dual licensed under OpenSSL and +# CRYPTOGAMS licenses depending on where you obtain it. For further +# details see http://www.openssl.org/~appro/cryptogams/. # ==================================================================== # +# July 2004 +# # 2.22x RC4 tune-up:-) It should be noted though that my hand [as in # "hand-coded assembler"] doesn't stand for the whole improvement # coefficient. It turned out that eliminating RC4_CHAR from config @@ -18,6 +28,8 @@ # to operate on partial registers, it turned out to be the best bet. # At least for AMD... How IA32E would perform remains to be seen... +# November 2004 +# # As was shown by Marc Bevand reordering of couple of load operations # results in even higher performance gain of 3.3x:-) At least on # Opteron... For reference, 1x in this case is RC4_CHAR C-code @@ -25,6 +37,8 @@ # Latter means that if you want to *estimate* what to expect from # *your* Opteron, then multiply 54 by 3.3 and clock frequency in GHz. +# November 2004 +# # Intel P4 EM64T core was found to run the AMD64 code really slow... # The only way to achieve comparable performance on P4 was to keep # RC4_CHAR. Kind of ironic, huh? As it's apparently impossible to @@ -32,23 +46,29 @@ # on either AMD and Intel platforms, I implement both cases. See # rc4_skey.c for further details... -# P4 EM64T core appears to be "allergic" to 64-bit inc/dec. Replacing +# April 2005 +# +# P4 EM64T core appears to be "allergic" to 64-bit inc/dec. Replacing # those with add/sub results in 50% performance improvement of folded # loop... +# May 2005 +# # As was shown by Zou Nanhai loop unrolling can improve Intel EM64T # performance by >30% [unlike P4 32-bit case that is]. But this is # provided that loads are reordered even more aggressively! Both code -# pathes, AMD64 and EM64T, reorder loads in essentially same manner +# paths, AMD64 and EM64T, reorder loads in essentially same manner # as my IA-64 implementation. On Opteron this resulted in modest 5% # improvement [I had to test it], while final Intel P4 performance # achieves respectful 432MBps on 2.8GHz processor now. For reference. # If executed on Xeon, current RC4_CHAR code-path is 2.7x faster than # RC4_INT code-path. While if executed on Opteron, it's only 25% -# slower than the RC4_INT one [meaning that if CPU µ-arch detection +# slower than the RC4_INT one [meaning that if CPU µ-arch detection # is not implemented, then this final RC4_CHAR code-path should be # preferred, as it provides better *all-round* performance]. +# March 2007 +# # Intel Core2 was observed to perform poorly on both code paths:-( It # apparently suffers from some kind of partial register stall, which # occurs in 64-bit mode only [as virtually identical 32-bit loop was @@ -57,21 +77,63 @@ # fit for Core2 and therefore the code was modified to skip cloop8 on # this CPU. -$output=shift; -open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output"; +# May 2010 +# +# Intel Westmere was observed to perform suboptimally. Adding yet +# another movzb to cloop1 improved performance by almost 50%! Core2 +# performance is improved too, but nominally... + +# May 2011 +# +# The only code path that was not modified is P4-specific one. Non-P4 +# Intel code path optimization is heavily based on submission by Maxim +# Perminov, Maxim Locktyukhin and Jim Guilford of Intel. I've used +# some of the ideas even in attempt to optmize the original RC4_INT +# code path... Current performance in cycles per processed byte (less +# is better) and improvement coefficients relative to previous +# version of this module are: +# +# Opteron 5.3/+0%(*) +# P4 6.5 +# Core2 6.2/+15%(**) +# Westmere 4.2/+60% +# Sandy Bridge 4.2/+120% +# Atom 9.3/+80% +# VIA Nano 6.4/+4% +# Ivy Bridge 4.1/+30% +# Bulldozer 4.5/+30%(*) +# +# (*) But corresponding loop has less instructions, which should have +# positive effect on upcoming Bulldozer, which has one less ALU. +# For reference, Intel code runs at 6.8 cpb rate on Opteron. +# (**) Note that Core2 result is ~15% lower than corresponding result +# for 32-bit code, meaning that it's possible to improve it, +# but more than likely at the cost of the others (see rc4-586.pl +# to get the idea)... + +$flavour = shift; +$output = shift; +if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } + +$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); + +$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; +( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or +( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or +die "can't locate x86_64-xlate.pl"; + +open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""; +*STDOUT=*OUT; $dat="%rdi"; # arg1 $len="%rsi"; # arg2 $inp="%rdx"; # arg3 $out="%rcx"; # arg4 -@XX=("%r8","%r10"); -@TX=("%r9","%r11"); -$YY="%r12"; -$TY="%r13"; - +{ $code=<<___; .text +.extern OPENSSL_ia32cap_P .globl RC4 .type RC4,\@function,4 @@ -80,62 +142,185 @@ RC4: or $len,$len jne .Lentry ret .Lentry: +.cfi_startproc + push %rbx +.cfi_push %rbx push %r12 +.cfi_push %r12 push %r13 +.cfi_push %r13 +.Lprologue: + mov $len,%r11 + mov $inp,%r12 + mov $out,%r13 +___ +my $len="%r11"; # reassign input arguments +my $inp="%r12"; +my $out="%r13"; + +my @XX=("%r10","%rsi"); +my @TX=("%rax","%rbx"); +my $YY="%rcx"; +my $TY="%rdx"; - add \$8,$dat - movl -8($dat),$XX[0]#d - movl -4($dat),$YY#d +$code.=<<___; + xor $XX[0],$XX[0] + xor $YY,$YY + + lea 8($dat),$dat + mov -8($dat),$XX[0]#b + mov -4($dat),$YY#b cmpl \$-1,256($dat) je .LRC4_CHAR + mov OPENSSL_ia32cap_P(%rip),%r8d + xor $TX[1],$TX[1] inc $XX[0]#b + sub $XX[0],$TX[1] + sub $inp,$out movl ($dat,$XX[0],4),$TX[0]#d - test \$-8,$len + test \$-16,$len jz .Lloop1 - jmp .Lloop8 + bt \$30,%r8d # Intel CPU? + jc .Lintel + and \$7,$TX[1] + lea 1($XX[0]),$XX[1] + jz .Loop8 + sub $TX[1],$len +.Loop8_warmup: + add $TX[0]#b,$YY#b + movl ($dat,$YY,4),$TY#d + movl $TX[0]#d,($dat,$YY,4) + movl $TY#d,($dat,$XX[0],4) + add $TY#b,$TX[0]#b + inc $XX[0]#b + movl ($dat,$TX[0],4),$TY#d + movl ($dat,$XX[0],4),$TX[0]#d + xorb ($inp),$TY#b + movb $TY#b,($out,$inp) + lea 1($inp),$inp + dec $TX[1] + jnz .Loop8_warmup + + lea 1($XX[0]),$XX[1] + jmp .Loop8 .align 16 -.Lloop8: +.Loop8: ___ for ($i=0;$i<8;$i++) { +$code.=<<___ if ($i==7); + add \$8,$XX[1]#b +___ $code.=<<___; add $TX[0]#b,$YY#b - mov $XX[0],$XX[1] movl ($dat,$YY,4),$TY#d - ror \$8,%rax # ror is redundant when $i=0 - inc $XX[1]#b - movl ($dat,$XX[1],4),$TX[1]#d - cmp $XX[1],$YY movl $TX[0]#d,($dat,$YY,4) - cmove $TX[0],$TX[1] - movl $TY#d,($dat,$XX[0],4) + movl `4*($i==7?-1:$i)`($dat,$XX[1],4),$TX[1]#d + ror \$8,%r8 # ror is redundant when $i=0 + movl $TY#d,4*$i($dat,$XX[0],4) add $TX[0]#b,$TY#b - movb ($dat,$TY,4),%al + movb ($dat,$TY,4),%r8b ___ -push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers +push(@TX,shift(@TX)); #push(@XX,shift(@XX)); # "rotate" registers } $code.=<<___; - ror \$8,%rax + add \$8,$XX[0]#b + ror \$8,%r8 sub \$8,$len - xor ($inp),%rax - add \$8,$inp - mov %rax,($out) - add \$8,$out + xor ($inp),%r8 + mov %r8,($out,$inp) + lea 8($inp),$inp test \$-8,$len - jnz .Lloop8 + jnz .Loop8 cmp \$0,$len jne .Lloop1 + jmp .Lexit + +.align 16 +.Lintel: + test \$-32,$len + jz .Lloop1 + and \$15,$TX[1] + jz .Loop16_is_hot + sub $TX[1],$len +.Loop16_warmup: + add $TX[0]#b,$YY#b + movl ($dat,$YY,4),$TY#d + movl $TX[0]#d,($dat,$YY,4) + movl $TY#d,($dat,$XX[0],4) + add $TY#b,$TX[0]#b + inc $XX[0]#b + movl ($dat,$TX[0],4),$TY#d + movl ($dat,$XX[0],4),$TX[0]#d + xorb ($inp),$TY#b + movb $TY#b,($out,$inp) + lea 1($inp),$inp + dec $TX[1] + jnz .Loop16_warmup + + mov $YY,$TX[1] + xor $YY,$YY + mov $TX[1]#b,$YY#b + +.Loop16_is_hot: + lea ($dat,$XX[0],4),$XX[1] ___ +sub RC4_loop { + my $i=shift; + my $j=$i<0?0:$i; + my $xmm="%xmm".($j&1); + + $code.=" add \$16,$XX[0]#b\n" if ($i==15); + $code.=" movdqu ($inp),%xmm2\n" if ($i==15); + $code.=" add $TX[0]#b,$YY#b\n" if ($i<=0); + $code.=" movl ($dat,$YY,4),$TY#d\n"; + $code.=" pxor %xmm0,%xmm2\n" if ($i==0); + $code.=" psllq \$8,%xmm1\n" if ($i==0); + $code.=" pxor $xmm,$xmm\n" if ($i<=1); + $code.=" movl $TX[0]#d,($dat,$YY,4)\n"; + $code.=" add $TY#b,$TX[0]#b\n"; + $code.=" movl `4*($j+1)`($XX[1]),$TX[1]#d\n" if ($i<15); + $code.=" movz $TX[0]#b,$TX[0]#d\n"; + $code.=" movl $TY#d,4*$j($XX[1])\n"; + $code.=" pxor %xmm1,%xmm2\n" if ($i==0); + $code.=" lea ($dat,$XX[0],4),$XX[1]\n" if ($i==15); + $code.=" add $TX[1]#b,$YY#b\n" if ($i<15); + $code.=" pinsrw \$`($j>>1)&7`,($dat,$TX[0],4),$xmm\n"; + $code.=" movdqu %xmm2,($out,$inp)\n" if ($i==0); + $code.=" lea 16($inp),$inp\n" if ($i==0); + $code.=" movl ($XX[1]),$TX[1]#d\n" if ($i==15); +} + RC4_loop(-1); $code.=<<___; -.Lexit: - sub \$1,$XX[0]#b - movl $XX[0]#d,-8($dat) - movl $YY#d,-4($dat) + jmp .Loop16_enter +.align 16 +.Loop16: +___ + +for ($i=0;$i<16;$i++) { + $code.=".Loop16_enter:\n" if ($i==1); + RC4_loop($i); + push(@TX,shift(@TX)); # "rotate" registers +} +$code.=<<___; + mov $YY,$TX[1] + xor $YY,$YY # keyword to partial register + sub \$16,$len + mov $TX[1]#b,$YY#b + test \$-16,$len + jnz .Loop16 + + psllq \$8,%xmm1 + pxor %xmm0,%xmm2 + pxor %xmm1,%xmm2 + movdqu %xmm2,($out,$inp) + lea 16($inp),$inp + + cmp \$0,$len + jne .Lloop1 + jmp .Lexit - pop %r13 - pop %r12 - ret .align 16 .Lloop1: add $TX[0]#b,$YY#b @@ -147,9 +332,8 @@ $code.=<<___; movl ($dat,$TX[0],4),$TY#d movl ($dat,$XX[0],4),$TX[0]#d xorb ($inp),$TY#b - inc $inp - movb $TY#b,($out) - inc $out + movb $TY#b,($out,$inp) + lea 1($inp),$inp dec $len jnz .Lloop1 jmp .Lexit @@ -160,14 +344,11 @@ $code.=<<___; movzb ($dat,$XX[0]),$TX[0]#d test \$-8,$len jz .Lcloop1 - cmp \$0,260($dat) - jnz .Lcloop1 - push %rbx jmp .Lcloop8 .align 16 .Lcloop8: - mov ($inp),%eax - mov 4($inp),%ebx + mov ($inp),%r8d + mov 4($inp),%r9d ___ # unroll 2x4-wise, because 64-bit rotates kill Intel P4... for ($i=0;$i<4;$i++) { @@ -184,8 +365,8 @@ $code.=<<___; mov $TX[0],$TX[1] .Lcmov$i: add $TX[0]#b,$TY#b - xor ($dat,$TY),%al - ror \$8,%eax + xor ($dat,$TY),%r8b + ror \$8,%r8d ___ push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers } @@ -203,21 +384,20 @@ $code.=<<___; mov $TX[0],$TX[1] .Lcmov$i: add $TX[0]#b,$TY#b - xor ($dat,$TY),%bl - ror \$8,%ebx + xor ($dat,$TY),%r9b + ror \$8,%r9d ___ push(@TX,shift(@TX)); push(@XX,shift(@XX)); # "rotate" registers } $code.=<<___; lea -8($len),$len - mov %eax,($out) + mov %r8d,($out) lea 8($inp),$inp - mov %ebx,4($out) + mov %r9d,4($out) lea 8($out),$out test \$-8,$len jnz .Lcloop8 - pop %rbx cmp \$0,$len jne .Lcloop1 jmp .Lexit @@ -226,6 +406,7 @@ $code.=<<___; .align 16 .Lcloop1: add $TX[0]#b,$YY#b + movzb $YY#b,$YY#d movzb ($dat,$YY),$TY#d movb $TX[0]#b,($dat,$YY) movb $TY#b,($dat,$XX[0]) @@ -242,14 +423,32 @@ $code.=<<___; sub \$1,$len jnz .Lcloop1 jmp .Lexit + +.align 16 +.Lexit: + sub \$1,$XX[0]#b + movl $XX[0]#d,-8($dat) + movl $YY#d,-4($dat) + + mov (%rsp),%r13 +.cfi_restore %r13 + mov 8(%rsp),%r12 +.cfi_restore %r12 + mov 16(%rsp),%rbx +.cfi_restore %rbx + add \$24,%rsp +.cfi_adjust_cfa_offset -24 +.Lepilogue: + ret +.cfi_endproc .size RC4,.-RC4 ___ +} $idx="%r8"; $ido="%r9"; $code.=<<___; -.extern OPENSSL_ia32cap_P .globl RC4_set_key .type RC4_set_key,\@function,3 .align 16 @@ -264,12 +463,9 @@ RC4_set_key: xor %r11,%r11 mov OPENSSL_ia32cap_P(%rip),$idx#d - bt \$20,$idx#d - jnc .Lw1stloop - bt \$30,$idx#d - setc $ido#b - mov $ido#d,260($dat) - jmp .Lc1stloop + bt \$20,$idx#d # RC4_CHAR? + jc .Lc1stloop + jmp .Lw1stloop .align 16 .Lw1stloop: @@ -326,31 +522,174 @@ RC4_set_key: .size RC4_set_key,.-RC4_set_key .globl RC4_options -.type RC4_options,\@function,0 +.type RC4_options,\@abi-omnipotent .align 16 RC4_options: - .picmeup %rax - lea .Lopts-.(%rax),%rax + lea .Lopts(%rip),%rax mov OPENSSL_ia32cap_P(%rip),%edx bt \$20,%edx - jnc .Ldone - add \$12,%rax + jc .L8xchar bt \$30,%edx jnc .Ldone - add \$13,%rax + add \$25,%rax + ret +.L8xchar: + add \$12,%rax .Ldone: ret .align 64 .Lopts: .asciz "rc4(8x,int)" .asciz "rc4(8x,char)" -.asciz "rc4(1x,char)" -.asciz "RC4 for x86_64, OpenSSL project" +.asciz "rc4(16x,int)" +.asciz "RC4 for x86_64, CRYPTOGAMS by " .align 64 .size RC4_options,.-RC4_options ___ -$code =~ s/#([bwd])/$1/gm; +# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, +# CONTEXT *context,DISPATCHER_CONTEXT *disp) +if ($win64) { +$rec="%rcx"; +$frame="%rdx"; +$context="%r8"; +$disp="%r9"; + +$code.=<<___; +.extern __imp_RtlVirtualUnwind +.type stream_se_handler,\@abi-omnipotent +.align 16 +stream_se_handler: + push %rsi + push %rdi + push %rbx + push %rbp + push %r12 + push %r13 + push %r14 + push %r15 + pushfq + sub \$64,%rsp + + mov 120($context),%rax # pull context->Rax + mov 248($context),%rbx # pull context->Rip + + lea .Lprologue(%rip),%r10 + cmp %r10,%rbx # context->RipRsp + + lea .Lepilogue(%rip),%r10 + cmp %r10,%rbx # context->Rip>=epilogue label + jae .Lin_prologue + + lea 24(%rax),%rax + + mov -8(%rax),%rbx + mov -16(%rax),%r12 + mov -24(%rax),%r13 + mov %rbx,144($context) # restore context->Rbx + mov %r12,216($context) # restore context->R12 + mov %r13,224($context) # restore context->R13 + +.Lin_prologue: + mov 8(%rax),%rdi + mov 16(%rax),%rsi + mov %rax,152($context) # restore context->Rsp + mov %rsi,168($context) # restore context->Rsi + mov %rdi,176($context) # restore context->Rdi + + jmp .Lcommon_seh_exit +.size stream_se_handler,.-stream_se_handler + +.type key_se_handler,\@abi-omnipotent +.align 16 +key_se_handler: + push %rsi + push %rdi + push %rbx + push %rbp + push %r12 + push %r13 + push %r14 + push %r15 + pushfq + sub \$64,%rsp + + mov 152($context),%rax # pull context->Rsp + mov 8(%rax),%rdi + mov 16(%rax),%rsi + mov %rsi,168($context) # restore context->Rsi + mov %rdi,176($context) # restore context->Rdi + +.Lcommon_seh_exit: + + mov 40($disp),%rdi # disp->ContextRecord + mov $context,%rsi # context + mov \$154,%ecx # sizeof(CONTEXT) + .long 0xa548f3fc # cld; rep movsq + + mov $disp,%rsi + xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER + mov 8(%rsi),%rdx # arg2, disp->ImageBase + mov 0(%rsi),%r8 # arg3, disp->ControlPc + mov 16(%rsi),%r9 # arg4, disp->FunctionEntry + mov 40(%rsi),%r10 # disp->ContextRecord + lea 56(%rsi),%r11 # &disp->HandlerData + lea 24(%rsi),%r12 # &disp->EstablisherFrame + mov %r10,32(%rsp) # arg5 + mov %r11,40(%rsp) # arg6 + mov %r12,48(%rsp) # arg7 + mov %rcx,56(%rsp) # arg8, (NULL) + call *__imp_RtlVirtualUnwind(%rip) + + mov \$1,%eax # ExceptionContinueSearch + add \$64,%rsp + popfq + pop %r15 + pop %r14 + pop %r13 + pop %r12 + pop %rbp + pop %rbx + pop %rdi + pop %rsi + ret +.size key_se_handler,.-key_se_handler + +.section .pdata +.align 4 + .rva .LSEH_begin_RC4 + .rva .LSEH_end_RC4 + .rva .LSEH_info_RC4 + + .rva .LSEH_begin_RC4_set_key + .rva .LSEH_end_RC4_set_key + .rva .LSEH_info_RC4_set_key + +.section .xdata +.align 8 +.LSEH_info_RC4: + .byte 9,0,0,0 + .rva stream_se_handler +.LSEH_info_RC4_set_key: + .byte 9,0,0,0 + .rva key_se_handler +___ +} + +sub reg_part { +my ($reg,$conv)=@_; + if ($reg =~ /%r[0-9]+/) { $reg .= $conv; } + elsif ($conv eq "b") { $reg =~ s/%[er]([^x]+)x?/%$1l/; } + elsif ($conv eq "w") { $reg =~ s/%[er](.+)/%$1/; } + elsif ($conv eq "d") { $reg =~ s/%[er](.+)/%e$1/; } + return $reg; +} + +$code =~ s/(%[a-z0-9]+)#([bwd])/reg_part($1,$2)/gem; +$code =~ s/\`([^\`]*)\`/eval $1/gem; print $code;