The verifier asserts after tail duplication on the following testcase:

void %interpret() {
entry:
        br label %retry

retry:
        %tmp.8 = call int %interp() 
        switch uint 0, label %endif.4 [
                 uint 4294967271, label %return
                 uint 4294967280, label %return
        ]

endif.4:
        br bool false, label %return, label %retry

return:
        %result.0 = phi int [ %tmp.8, %retry ], [ %tmp.8, %endif.4 ]
        ret void
}

declare int %interp()

Err, here's the correct testcase:

void %interpret() {
entry:
        br label %retry

retry:
        %tmp.8 = call int %interp()
        switch uint 0, label %endif.4 [
                 uint 4294967271, label %return
                 uint 4294967280, label %return
        ]

endif.4:
        br bool false, label %return, label %retry

return:
        %result.0 = phi int [ %tmp.8, %retry ], [%tmp.8, %retry], [ %tmp.8,
%endif.4 ]
        ret void
}

declare int %interp()

Fixed, here's the patch:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20040329/013415.html

Testcase here: TailDup/2004-04-01-DemoteRegToStack.llx

I'm marking this as a low-priority bug because it never would have caused a
problem in gccas, it only crashed because I didn't run the mem2reg pass after
it.  It's still a bug, just not one that would realistically occur in the wild.

-Chris