# is better) and improvement coefficients relative to previous
# version of this module are:
#
-# Opteron 5.3/+0%
+# Opteron 5.3/+0%(*)
# P4 6.5
-# Core2 6.2/+15%(*)
+# 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%(*)
#
-# (*) Note that Core2 result is ~15% lower than corresponding result
+# (*) 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)...
( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
die "can't locate x86_64-xlate.pl";
-open STDOUT,"| $^X $xlate $flavour $output";
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
$dat="%rdi"; # arg1
$len="%rsi"; # arg2
{
$code=<<___;
.text
+.extern OPENSSL_ia32cap_P
.globl RC4
.type RC4,\@function,4
$ido="%r9";
$code.=<<___;
-.extern OPENSSL_ia32cap_P
.globl RC4_set_key
.type RC4_set_key,\@function,3
.align 16