> Of cource, the "right way" is importing newer egcs sources
> to our source tree but it is too hard to do right now.

Moving to 2.95.2 would be my preferred solution, but it is by no means the 
only one and I understand why it hasn't been done.

> 
> I think a simple "workaround" is just to disable schedule-insns
> optimization, which would cause this bug.
> (i.e. add "-fno-schedule-insns" to COPTS in mk.conf or elsewhere)

That's a pretty draconian way to fix the problem :-(

Anyway, it sounds like it's the post-reload scheduling pass that you want 
to disable, since it is after reload that the death notes are incorrect.  
The flag to disable that pass is -fno-schedule-insns2.

R.