mirror/openssl
2
0
Fork 0
mirror of https://github.com/openssl/openssl.git synced 2024-11-27 05:21:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

Bug-fix in CBC encrypt tail processing and commentary section update.

Browse Source
This commit is contained in:
Andy Polyakov 2005-01-20 10:33:37 +00:00
parent a963395a7b
commit bac252a5e3
1 changed files with 29 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
crypto/aes/asm/aes-586.pl
View File

@ -6,7 +6,7 @@
# forms are granted according to the OpenSSL license.
# ====================================================================
#
# Version 3.0.
# Version 3.1.
#
# You might fail to appreciate this module performance from the first
# try. If compared to "vanilla" linux-ia32-icc target, i.e. considered
@ -46,23 +46,27 @@
# Instruction Level Parallelism, and it indeed resulted in up to 15%
# better performance on most recent µ-archs...
#
# Current ECB performance numbers for 128-bit key in cycles per byte
# [measure commonly used by AES benchmarkers] are:
# Current ECB performance numbers for 128-bit key in CPU cycles per
# processed byte [measure commonly used by AES benchmarkers] are:
#
# small footprint fully unrolled
# P4[-3] 23[24] 22[23]
# AMD K8 19 18
# PIII 26(*) 23
# PIII 26 23
# Pentium 63(*) 52
#
# (*) Performance difference between small footprint code and fully
# unrolled in more commonly used CBC mode is not as big, 7% for
# PIII and 15% for Pentium, which I consider tolerable.
# unrolled in more commonly used CBC mode is not as big, 4% for
# for Pentium. PIII's ~13% difference [in both cases in 3rd
# version] is considered tolerable...
#
# Third version adds AES_cbc_encrypt implementation, which resulted in
# up to 40% performance imrovement of CBC benchmark results [on most
# recent µ-archs]. CBC performance is virtually as good as ECB now and
# sometimes even better, because function prologues and epilogues are
# up to 40% performance imrovement of CBC benchmark results. 40% was
# observed on P4 core, where "overall" imrovement coefficient, i.e. if
# compared to PIC generated by GCC and in CBC mode, was observed to be
# as large as 4x:-) CBC performance is virtually identical to ECB now
# and on some platforms even better, e.g. 56 "small" cycles/byte on
# senior Pentium, because certain function prologues and epilogues are
# effectively taken out of the loop...
push(@INC,"perlasm","../../perlasm");
@ -79,8 +83,9 @@ $acc="esi";
$small_footprint=1; # $small_footprint=1 code is ~5% slower [on
# recent µ-archs], but ~5 times smaller!
# I favor compact code, because it minimizes
# cache contention...
# I favor compact code to minimize cache
# contention and in hope to "collect" 5% back
# in real-life applications...
$vertical_spin=0; # shift "verticaly" defaults to 0, because of
# its proof-of-concept status...
@ -1296,12 +1301,18 @@ sub declast()
&push ($key eq "edi" ? $key : ""); # push ivp
&pushf ();
&mov ($key,&wparam(1)); # load out
&xor ($s0,$s0);
&mov (&DWP(0,$key),$s0); # zero output
&mov (&DWP(4,$key),$s0);
&mov (&DWP(8,$key),$s0);
&mov (&DWP(12,$key),$s0);
&mov ($s1,16);
&sub ($s1,$s2);
&cmp ($key,$acc); # compare with inp
&je (&label("enc_in_place"));
&data_word(0x90A4F3FC); # cld; rep movsb; nop # copy input
&jmp (&label("enc_skip_in_place"));
&set_label("enc_in_place");
&lea ($key,&DWP(0,$key,$s2));
&set_label("enc_skip_in_place");
&mov ($s2,$s1);
&xor ($s0,$s0);
&data_word(0x90AAF3FC); # cld; rep stosb; nop # zero tail
&popf ();
&pop ($key); # pop ivp
@ -1456,6 +1467,8 @@ sub declast()
&pushf ();
&data_word(0x90A4F3FC); # cld; rep movsb; nop # restore tail
&popf ();
&align (4);
&set_label("dec_out");
&stack_pop(5);
&function_end("AES_cbc_encrypt");