I reduced ed, which has not worked for the longest time, into a test case.
It turned out to be kind of interesting: ed's main() has a loop involving a goto
and a sigsetjmp. Unfortunately, I am now pretty tired, so I'm going to make this
brief: the LLVM program which I am going to attach to this bug is supposed to
print "ok\n", and it loops infinitely instead if you apply opt -adce.

Created an attachment (id=23) [details]
test case

This is the test case, reduced beyond recognition from ed-0.2/main.c

This bug is caused by the code doing basically this:

...
  while (1) exit(0);
...

No post-dominance information exists for this loop because it cannot reach the
exit node of the CFG (a return or unwind instruction).  ADCE should obviously
not delete these blocks though.  I'm preparing a patch.

-Chris

Fixed like this (comments in the patch):
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20031110/009564.html

Thanks for taking the time to reduce this testcase.  This is a _bad_
miscompilation.  :(

-Chris

Oh yeah, the testcase is here:
llvm/test/Regression/Transforms/ADCE/2003-11-16-MissingPostDominanceInfo.ll