Sorry about the lateness. I want to take a look at this now but the patch does
not apply cleanly against tot. Can you supply a new patch?

new patch
Hi Evan, here's a refresh patch. Enjoy!

Note that the test CodeGen/Generic/2007-04-14-EHSelectorCrash.ll
is now marked XFAIL because it causes the following assertion to
fire in SelectionDAGISel (the assertion shouldn't really exist -
it is part of a hack to workaround an exception handling problem
that we don't yet have a good solution for):

#ifndef NDEBUG
assert(FuncInfo.CatchInfoFound.size() == FuncInfo.CatchInfoLost.size() &&
"Not all catch info was assigned to a landing pad!");
#endif

If you remove that assertion and un-XFAIL the test, you should see
the assertion in live-intervals that I mentioned in the initial
bugreport.

Working on #​2. Initial thoughts: LiveIntervals::computerIntervals() should
traverse the graph in df order. If the live-in eh registers are not used, then
they should be live-outs and propagated as live-ins to the successors.

If the live-in eh registers are not used

I guess you mean: if the live-in eh registers are not killed.

Nevermind. I've decided the proper fix is to simply add the EH register(s) as a
livein in the landing pad successor (and any block in the path from the landing
pad to the said block). I've also fixed live-interval to not to puke if a livein
is not used.

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070625/050807.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070625/050808.html

Let me know if you run into other issues.

2007-04-14-EHSelectorCrash.ll is still XFAIL due to the CatchInfo issue. Please
open another bug to track that (if there isn't one already).