Date:	Fri, 6 Aug 1999 02:46:26 -0300 (ADT)
From:	Peter Cordes <peter AT Cordes DOT Phys DOT Dal DOT Ca>
To:	pgcc AT delorie DOT com
Subject:	bzip2-mmx
Message-ID:	<Pine.LNX.4.10.9908060214320.16525-100000@Cordes.Phys.Dal.Ca>
MIME-Version:	1.0
Reply-To:	pgcc AT delorie DOT com

 I downloaded the 3 versions of bzip2 from http://goof.com/pcg/data/test/
I have no idea what they're supposed to be, other than that they are
mmx-enabled bzip2 binaries from a bleeding-edge pgcc.

 I noticed a couple problems with them though.

1.  mmx versions break when decompressing:

# on a PIII running RH6.0
# compress with system bzip2, no monkey business yet
cordes]~/tmp $ echo hello | bzip2 > foo.bz2 
cordes]~/tmp $ ./bzip2-nommx -dc foo.bz2 
hello
cordes]~/tmp $ ./bzip2-mmx -dc foo.bz2 

bzip2-mmx: Caught a SIGSEGV or SIGBUS whilst decompressing,
[snip]

cordes]~/tmp $ ./bzip2-mmxonly -dc foo.bz2 
[same error]

(strace indicates a SIGSEGV, BTW)

 These results are duplicated on a P200MMX running Stampede 0.89
Kernel version 2.2.10, compiled with egcs 1.1.2 the RH6.0 machine, and
with pgcc 1.1.1 (-O6 -fno-strength-reduce) on the Stampede machine.



2. MMX instructions are executed by the bzip2-mmx binary, even on non-mmx
machines.  (I'm guessing that it is supposed to figure out which you have
and run the appropriate code, so it is meant to work on non-mmx machines.
If not, then this is not a bug.)

 note: lever has an Intel Pentium @ 100MHz
 I ran bzip2-mmx with gdb, so I could find what instruction was bad.  
From the shell, it just dies from the SIGILL it gets when it tries to run
the emms instruction.  (BTW, all I know about mmx programming is what I
read in a Linux Journal article last year.)

peter AT lever]~ $ gdb -q bzip2-mmx
(no debugging symbols found)...(gdb) 
(gdb) run < /usr/dict/words > /dev/null
Starting program: /home/peter/bzip2-mmx < /usr/dict/words > /dev/null
warning: shared library handler failed to enable breakpoint

Program received signal SIGILL, Illegal instruction.
0x8052e7a in ?? ()
(gdb) disassemble 0x8052e7a 0x8052eaa
Dump of assembler code from 0x8052e7a to 0x8052eaa:
0x8052e7a:	emms   
0x8052e7c:	call   0x805dd50
0x8052e81:	addl   $0x4,%esp
0x8052e84:	testl  %eax,%eax
0x8052e86:	je     0x8052e90
0x8052e88:	emms   
0x8052e8a:	call   0x80572d8
0x8052e8f:	nop    
0x8052e90:	movl   0xc(%ebp),%edi
0x8052e93:	pushl  %edi
0x8052e94:	emms   
0x8052e96:	call   0x805dd50
0x8052e9b:	addl   $0x4,%esp
0x8052e9e:	testl  %eax,%eax
0x8052ea0:	je     0x8052eb0
0x8052ea2:	emms   
0x8052ea4:	call   0x80572d8
0x8052ea9:	leal   0x0(%esi,1),%esi
End of assembler dump.


 Keep up the good work, Marc + team :)

#define X(x,y) x##y
Peter Cordes ;  e-mail: X(peter AT cordes DOT phys. , dal.ca)

- Raw text -