samba-tng: Reproducible, incomprehensible segfaults

[Steve Langasek]

This is very, very frustrating.

For the past three weeks, I've been unable to get samba-tng CVS to work
correctly on my machine.  Anything making use of libsmb segfaults.

Linux/x86 2.2.13, glibc 2.1.2, egcs 1.1.2.  (readline is present, if you're
interested, although this problem is unrelated to readline and was not solved
by the recent fixes posted.)

Then, the other night, I noticed that everything built just fine on another
machine.  I could even copy the binaries over from the second machine to the
first, and everything would work.

A few upgrades later... trading object files back and forth across the
network...

The only discernable difference between the compile environments is that one
machine is an i586, and the other is an i686.
The problematic code is somewhere in libsmb/smbencrypt.c, and the problem
happens sometime during the compiling stage.  Preprocessor output is identical
on both machines, and if I copy the compiled output from the happy machine to
the unhappy machine, the assembler produces properly working code.

Very frustrating indeed.

Short of a complete reinstall of the system, does anyone have any ideas?  I've
compiled any number of applications on this system without difficulty, and I
have no idea why I should suddenly have problems with this one chunk of
code...  Is anyone else seeing similar behavior with SAMBA_TNG?

Steve Langasek
postmodern programmer