1.1.2 (and newer): link-stage alignment bugs on alphaev6-unknown-linux-gnu

[abel at bfr dot co dot il (Alexander L. Belikoff)]

Hello everybody -

I'm trying to build our project on an Alpha DP264 running RedHat Linux 
5.2. The project has been in use for a long time on Alpha LX164 w/
RedHat 4.2 and gcc 2.7.x.

And it doesn't look good:

1. Egcs 1.0.3 which is shipped as a default C/C++ compiler for RedHat
   5.2 crashes with the infamous 'unknown insn' message on some files,
   when building a debugging version (-g). When building an optimized
   version (-g -O2), it works fine and the resulting program works
   fine.

2. I built egcs 1.1.2 on the machine and tried to sue it. It
   does compile the optimized version. However, when I try to run the
   program, it crashes. The crash is associated with an 'unaligned
   trap' message in the system logs and further SIGFPE, because some
   values get insane. The same happens when using the March 21
   snapshot of egcs.

Building egcs 1.1.2 on a LX164 running RedHat 5.2
(alphaev56-unknown-linux-gnu) and moving it to 21264 didn't help
either - it still builds the same bad binary. This makes me assume, it 
is something about egcs...

Further investigation shows that it all boils down to the linking
stage. If I build the program using egcs 1.0.3 and then I remove the
executable, switch to egcs 1.1.2 and relink the binary, then I get
this bug.

I'd love to create an isolated example for this problem, but I simply
cannot - the program is HUGE, it includes FORTRAN code translated to C 
by f2c and Pascal code, translated by p2c as well as C code and the
bug is not visibly associated with any of the files - it looks like a
linking stage bug.

My wild guess is that egcs misaligns something and things go wild.

So, did anybody see something similar? Maybe a test case?..

Regards,

-- 
Alexander L. Belikoff
Bloomberg L.P. / BFM Financial Research Ltd.
abel@vallinor4.com, abel@bfr.co.il