Vorbis - Aug 2001 - Serious GCC/EGCS 2.91 bug found (bites rc2)

Hi folks,

The folks on RedHat who have been complaining of poor-quality encodes
prompted us to track down the trouble, and it's a compiler bug.  We
have the asm .s files to prove it :-)

The summary: EGCS (gcc) versions through March 1999, including up to
at least gcc/egcs 2.91.66 have a serious floating point optimization
bug that hits Vorbis.  This is a fairly old gcc version, but RedHat
shipped with it until at least RedHat 6.2.  The broken compile will
produce Ogg files, but the audio quality will have from obvious to
severe artifacts.

Also, because we wanted our RPMs to be compatable with redhat 6 and 7,
we built the RPMs using RedHat 6.2.  That means our RPMs are also
broken.  We'll have an updated, correct build on the vorbis.com site
ASAP.

No other builds on the vorbis.com site are affected.  Building from
source is only affected if the compiler used is egcs 2.91 (or older).
gcc --version will report the compiler version.

Amusingly enough, this EGCS bug was fixed approximately two weeks
after it shipped, but the broken version continued to be shipped for
over a year.

Monty

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

Hi vorbis/ogg developers alltogether,

On Thu, 16 Aug 2001, Monty wrote:
> Hi folks,
> 
> The folks on RedHat who have been complaining of poor-quality encodes
> prompted us to track down the trouble, and it's a compiler bug.  We
> have the asm .s files to prove it :-)
> 
In summary, all the math in GCC seems not to be of high quality.
I have had many troubles compiling some kind of code (runs good
using Watcom C for DOS/Win32/OS2) for Linux. I have tested
three different Linux versions resp. compilers:
-Red Hat 6.1
-SuSE 7.0
-SuSE 7.2 shipped with gcc version 2.95.3!!!
All they have corrupted my code, even with optimization switched off!
The atanh function seems to be buggy. It gives correct results in
a stand-alone environment. But it behaves in some kind illegal,
it corruptes the whole program. By masking atanh by
#undef atanh
#define atanh(x) myatanh(x)
and setting up my own myatanh, the program runs good.

Because I think ogg/vorbis does extensive uses math functions, I
give this warning to you.

J"org Bergmann, Dresden, Germany
email@jbergmann.de

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

----------
> Van: Dr.Joerg Bergmann <bergmann@rcs.urz.tu-dresden.de>
> Aan: vorbis@xiph.org
> Onderwerp: Re: [vorbis] Serious GCC/EGCS 2.91 bug found (bites rc2)
> Datum: donderdag 16 augustus 2001 9:27
> 
> Hi vorbis/ogg developers alltogether,
> 
> On Thu, 16 Aug 2001, Monty wrote:
> 
> > Hi folks,
> > 
> > The folks on RedHat who have been complaining of poor-quality encodes
> > prompted us to track down the trouble, and it's a compiler bug. 
We
> > have the asm .s files to prove it :-)
> > 
> In summary, all the math in GCC seems not to be of high quality.
> [snip]
> All they have corrupted my code, even with optimization switched off!
Wait a minute... if EGCS and GCC 2.95.x are both broken, then doesn't that
mean that
nearly all Linux distributions (except RH 7.x and Mandrake 8.x) have broken
compilers? o_0

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

----------
> Van: ndrw mchl grnbrg <agreenbu@nyx.net>
> Aan: vorbis@xiph.org
> Onderwerp: Re: [vorbis] Serious GCC/EGCS 2.91 bug found (bites rc2)
> Datum: donderdag 16 augustus 2001 17:04
> : Wait a minute... if EGCS and GCC 2.95.x are both broken, then doesn't
that
> : mean that
> : nearly all Linux distributions (except RH 7.x and Mandrake 8.x) have
broken
> : compilers? o_0
> 
> Mandrake 8 comes with a bad compiler, too (2.96).
So all Linux compilers are broken? o_0

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

Hongli Lai (hongli@telekabel.nl) wrote :
> ---------- 
> > Van: ndrw mchl grnbrg <agreenbu@nyx.net> 
> > Aan: vorbis@xiph.org 
> > Onderwerp: Re: [vorbis] Serious GCC/EGCS 2.91 bug found (bites rc2) 
> > Datum: donderdag 16 augustus 2001 17:04 
> > : Wait a minute... if EGCS and GCC 2.95.x are both broken, then
doesn't that
> > : mean that 
> > : nearly all Linux distributions (except RH 7.x and Mandrake 8.x) have
broken
> > : compilers? o_0 
> > 
> > Mandrake 8 comes with a bad compiler, too (2.96). 
> 
> So all Linux compilers are broken? o_0 
Every software has bugs. All of them.
Even the 10 liners.


-- 
David Balazic
--------------
"Be excellent to each other." - Bill & Ted
- - - - - - - - - - - - - - - - - - - - - -

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

Wow.  Huge difference.  I tried Ogg earlier this week and was quite 
disappointed.  I was using said RPMs.  I couldn't figure out why lots of 
people were using Ogg, it was slow and didn't sound great.  ;-) Building
from
source fixed that.  Good work guys.  I'm glad this was caught before 1.0.  

-Scott Wheeler

On Thursday 16 August 2001 02:14 am, you wrote:
> Hi folks,
>
> The folks on RedHat who have been complaining of poor-quality encodes
> prompted us to track down the trouble, and it's a compiler bug.  We
> have the asm .s files to prove it :-)
>
> The summary: EGCS (gcc) versions through March 1999, including up to
> at least gcc/egcs 2.91.66 have a serious floating point optimization
> bug that hits Vorbis.  This is a fairly old gcc version, but RedHat
> shipped with it until at least RedHat 6.2.  The broken compile will
> produce Ogg files, but the audio quality will have from obvious to
> severe artifacts.
>
> Also, because we wanted our RPMs to be compatable with redhat 6 and 7,
> we built the RPMs using RedHat 6.2.  That means our RPMs are also
> broken.  We'll have an updated, correct build on the vorbis.com site
> ASAP.
>
> No other builds on the vorbis.com site are affected.  Building from
> source is only affected if the compiler used is egcs 2.91 (or older).
> gcc --version will report the compiler version.
>
> Amusingly enough, this EGCS bug was fixed approximately two weeks
> after it shipped, but the broken version continued to be shipped for
> over a year.
>
> Monty
>
> --- >8 ----
> List archives:  http://www.xiph.org/archives/
> Ogg project homepage: http://www.xiph.org/ogg/
> To unsubscribe from this list, send a message to
> 'vorbis-dev-request@xiph.org' containing only the word
'unsubscribe' in the
> body.  No subject is needed. Unsubscribe messages sent to the list will be
> ignored/filtered.
--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

>Hi folks,
>
>The folks on RedHat who have been complaining of poor-quality encodes
>prompted us to track down the trouble, and it's a compiler bug.  We
>have the asm .s files to prove it :-)
>
>The summary: EGCS (gcc) versions through March 1999, including up to
>at least gcc/egcs 2.91.66 have a serious floating point optimization
Is there a quick and easy test for this bug for those of use 
with gcc version numbers like 2.9-beos-991026. This is
from someone with cloth ears and cheap speakers so a listening 
test is right out of it :)

Dave

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

>On Thu, Aug 16, 2001 at 08:19:09PM +0000, David Burnett wrote:
>> Is there a quick and easy test for this bug for those of use 
>
>The difference is pretty obvious, but you can tell because if you compile
>-O1 you will get much lower bitrate results then -O2 +
-O1 and -O20 both give the same average bitrate according to ogginfo 
so I  guess I'm in the clear.

Dave

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

"Dr.Joerg Bergmann" <bergmann@Rcs1.urz.tu-dresden.de> writes:
> OK, my words were incorrect. The compiler is OK, but the libc is
> buggy. But: A compiler without library is nonsens, at least for
> all non-system (non-kernel) development. E.g. the JAVA distributions
> must contain a huge set of libraries, otherwise no one would call them
> as "compiler+runtime".
> 
> I'm not so familiar with the Linux world, I'm very familiar with
> JAVA and commercial C compilers like the Watcom. Therefore, in my 
> opinion the libraries are part of the compiler distribution (as it is 
> e.g. for the Watcom C compiler). 
Simply not true for the Unix world, in general.  Libc (and libm)
generally comes with the system rather than the compiler.  If atan()
is buggy on a Solaris program compiled with GCC, it's Sun's fault, not
GNU's.

Of course, on Linux, both packages (compiler and library) come with
the system, but they are maintained by completely separate groups of
developers.

Java is somewhat of a different beast, but even there you can run code
compiled with IBM's Jikes compiler on a Sun JVM/libraries (though it
might or might not be a good idea).

</pedant>

-Doug

-- 
Free Dmitry Sklyarov! 
http://www.freesklyarov.org/ 

We will return to our regularly scheduled signature shortly.

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.