john-users - Re: 64bit support on windows?

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 4 Jan 2006 23:13:24 +0300
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com
Subject: Re:  64bit support on windows?

On Wed, Jan 04, 2006 at 03:40:22PM +0000, Phantom wrote:
> I recently acquired an AMD64 bit CPU and was wondering why there is no makefile
> option for 64bit under windows? (using x86-64.h ..??)

That's for several reasons:

1. I am not using Windows myself; it is tricky enough for me to go to a
Windows box just to do plain Win32 builds of John right now.

(FWIW, I tried to install Cygwin on ReactOS running under QEMU running
under Openwall GNU/*/Linux, but that failed in the middle of the
installation process.  I suspect that it might work with a future
version of ReactOS eventually.)

2. I do not know whether there's Cygwin for Win64 already (is there?)

> And, if made, would it require Win64 for it to run?

Yes.

> And.... is there any significant performance increase in running the 64bit
> linux build over the default cygwin-mmx one? if so, how much?

This gives us a third reason to not hurry to support 64-bit Windows:

3. No.  MMX was essentially 64-bit already as far as this application is
concerned.  The availability of 16 general-purpose registers with x86-64
(as opposed to just 8 with x86/MMX) may provide perhaps a 20%
improvement at bitslice DES, but on the other hand the MMX code is
hand-optimized, whereas the x86-64 code is currently being generated by
a C compiler.

The benchmarks done on Linux systems so far show that the hand-optimized
MMX code provides almost the same performance that the compiler-generated
x86-64 code does.  On the particular benchmark results I have in my
collection (for an Opteron CPU), the following hash types performed
better when built for 32-bit x86 with MMX: Traditional DES, BSDI DES,
FreeBSD MD5, OpenBSD Blowfish.  The following hash types performed
better when built for x86-64 natively: Kerberos AFS DES, NT LM DES.  But
the differences in performance were within 20% for most hash types, with
the rarely-used Kerberos AFS DES being the only exception.  The results
may differ for other CPUs and other C compilers (or gcc versions).

To answer your next question: no, I am not planning to develop
hand-optimized x86-64 assembly routines for John, at least not any time
soon.  20% is not enough of a performance win to justify that work for
me right now.  (The performance improvement with assembly code was
_much_ bigger for x86/MMX since the architecture was register-starved
and since MMX was not available for use by a C compiler.  x86-64 is
rather good, so C compilers manage to do a good job, too.)

What I _might_ do at a later stage, though, is produce some hybrid
x86-64+MMX and/or x86-64+SSE code for a bigger performance win.

-- 
Alexander Peslyak <solar at openwall.com>
GPG key ID: B35D3598  fp: 6429 0D7E F130 C13E C929  6447 73C3 A290 B35D 3598
http://www.openwall.com - bringing security into open computing environments

Was I helpful?  Please give your feedback here: http://rate.affero.net/solar