Building the whole Debian archive with GCC 4.1: a summary

[Martin Michlmayr <tbm at cyrius dot com>]

Over the last 2.5 weeks I have built the complete Debian archive
on a quad-core MIPS machine donated by Broadcom using the recently
released version 4.1 of GCC.  In parallel, I have done the same
on an EM64T box donated to Debian by Intel.

The purpose of this exercise was three-fold:
 - Find out about compiler problems in GCC 4.1 itself as well
   as in packages that may fail with the new version *before*
   GCC 4.1 is uploaded to Debian.
 - Find out about MIPS specific problems in GCC 4.1 and to answer
   the question of Matthias Klose, Debian's GCC maintainer, [1]
   as to which platforms can move to GCC 4.1 as the default compiler.
 - Find MIPS specific assembler warnings and create a list of all
   users of xgot (a MIPS specific toolchain problem).

I thought I'd summarize my findings for the GCC developers so you know
what kind of broken code is out there in the wild.  Given that Debian
is a fairly large distribution, we probably have a better coverage of
the "typical free software" project (i.e. not just the successful
projects everyone talks about) than most other distros.


Executive summary
-----------------

GCC 4.1 itself appears to be very stable, both on MIPS and AMD64.
There are, however, a large number of packages using code (especially
C++) which GCC 4.1 treats as errors.  Fortunately, most of them are
trivial to fix.  By compiling about 6200 software packages (see [2]
for the list), over 500 new bugs have been discovered and submitted:
280 of which are specific to the increased strictness of GCC 4.1.
Patches for 2/3 of those GCC 4.1 specific bugs have been submitted.


Detailed summary of problems
----------------------------

Basically, it all boils down to broken C++ code.  There were a few
bugs in C code, but the majority was in C++.   The most common
errors I found (and some *approximate* numbers) are:

 - extra qualification: about 187 bugs
 - reliance on friend injection: 26 bugs
 - wrong escape characters (e.g. "\."; most commonly seen in regular
   expressions): 6 bugs
 - iterator problems (such as assigning 0 or NULL to an iterator): 3
 - template specialisation in wrong namespace
 - template reliance on a function declared later
 - use of template's base class members, unqualified, where the base class
   is dependent
 - use of "assert" without #include <cassert>: 5 bugs
 - dereferencing type-punned pointer will break strict-aliasing rules
   together with -Werror: 6 bugs
 - the use of "mips" as a variable name when GCC on mips defines it
   as a constant (arg!) and you end up with: const 1 = -1: 1 bug

A list of all GCC 4.1 specific bugs in packages can be found at [3].
The single biggest syntax error found is the use of extra qualification,
which fortunately are trivial to fix.


Compiler bugs in GCC 4.1
------------------------

I've found a number of GCC bugs in this process (the # bug numbers refer
to bugs in the Debian bug tracker, please use http://bugs.debian.org/nnn):

 - gcc ICE: #356231 (when building gnubik_2.2-5)
   http://gcc.gnu.org/PR26672 - fixed

 - g++ ICE: #356569 (when building 3ddesktop_0.2.9-5.1)
   http://gcc.gnu.org/PR26757 - test case needs further reduction

 - gcc bug: #356574 (address of register requested)
   http://gcc.gnu.org/PR26004 - fixed

 - g++ bug: #356548 (cannot bind packed field)
   http://gcc.gnu.org/PR26670 - not yet assigned but related to known bugs

 - g++ bug: #356435 (no constructor)  [package bug: #356245]
   http://gcc.gnu.org/PR26755 - waiting
   This is a binary incompatibility!  4.0 has a different behaviour to
   3.3 and 4.1, which means that linking 4.1 code against 4.0 libraries
   may fail.  4.0 and 4.1 were supposed to be binary compatible so I'm not
   sure what will be done about this.

 - gcc bug: #356896 (miscompiles code, leading to segfault in program)
   http://gcc.gnu.org/PR26763 - confirmed

 - gcc or autoconf bug: #358644 (change in behaviour on mips regarding
   -fstack-protector)
   In the past, the use of -fstack-protector on MIPS would result in
   an error.  Now it succeeds (with a warning) but then fails to link.
   autoconf expects it to fail, however.

   It's not clear right now whether this behaviour change was correct and
   autoconf should be changed or whether GCC should revert that change.
   Fortunately, this change only breaks 2 packages out of ~6200.

   Maybe other people can look at http://bugs.debian.org/358644 and
   comment.


Musing from the point of view of a distro
-----------------------------------------

I think GCC is a really solid tool and I was surprised not to see more
problems related specifically to MIPS given that it's not one of the
mainstream platforms.  As someone involved in a major Linux distribution,
I'm not terribly happy about the way errors are handled.  I'm specifically
talking about the extra qualification error here, which to the best of
my knowledge was never treated as a warning: -pedantic in 4.0 would cause
an error and now 4.1 causes an error by default.

While I appreciate G++ becoming more standards compliant, I think you
should do it in steps: issue a warning in one release (unconditionally,
i.e. without the need for -pedantic) and *then* in the next release make
it an error.  This gives developers a change to fix their bugs before
they become errors.  Granted the extra qualification bugs are trivial
to fix, but as a major distro we now have almost 200 packages that will
fail because of this error and this slows down our adoption of GCC 4.1.

I'd like to propose that you explicitly make it part of your process to
generate a warning in one release of GCC and only in the next one to
change it to an error.  Please give developer some advance warning.  As
you can see from my examples above, there is *lots* of broken code out
there.


Further work
------------

Given that I have fast MIPS hardware (thanks to Broadcom) I will try to
test GCC snapshots on MIPS from time to time and report problems to you.
If someone needs access to a fast MIPS machine for some testing, I may
be able to arrange something on a temporary basis.  Please contact me
privately if this is the case.


Acknowledgements
----------------

 - Google for supporting my PhD, thereby allowing me to spend two weeks
   compiling the archive with GCC 4.1 and sorting out failures.
 - Intel for supporting some PhD work which led to this experiment.
 - Broadcom for supporting Debian's MIPS port through the donation of
   hardware to the project and individual developers.
 - Ben Hutchings for explaining many of the C++ bugs I found to me.
   I've learned more about C++ in these two weeks than I ever wanted to
   know. ;-)  Ben also submitted a number of patches for tricky bugs I
   couldn't fix.
 - Intel for donating an EM64T machine to Debian.
 - The GCC project for creating a great compiler, and in particular
   Andrew Pinski for doing lots of bug triaging work.


References
----------

[1] http://lists.debian.org/debian-gcc/2006/02/msg00173.html
[2] http://people.debian.org/~tbm/logs/gcc-4.1/PACKAGES
[3] http://bugs.debian.org/cgi-bin/pkgreport.cgi?which=submitter&includesubj=4.1&data=tbm%40cyrius.com&archive=no&version=&dist=unstable

-- 
Martin Michlmayr
http://www.cyrius.com/