The default of the dispatch switch statement was to branch to a BB that executed
authorBill Wendling <isanbard@gmail.com>
Mon, 11 Apr 2011 21:32:34 +0000 (21:32 +0000)
committerBill Wendling <isanbard@gmail.com>
Mon, 11 Apr 2011 21:32:34 +0000 (21:32 +0000)
commitcc88595bce813464f4dc0c8351f7311bb233a3e2
treeae5c3706f99b81550363e232be220ef70b7828cc
parent35563fee7b52f3bc9ef828abc139e8d9cb4ba2b9
The default of the dispatch switch statement was to branch to a BB that executed
the 'unwind' instruction. However, later on that instruction was converted into
a jump to the basic block it was located in, causing an infinite loop when we
get there.

It turns out, we get there if the _Unwind_Resume_or_Rethrow call returns (which
it's not supposed to do). It returns if it cannot find a place to unwind
to. Thus we would get what appears to be a "hang" when in reality it's just that
the EH couldn't be propagated further along.

Instead of infinitely looping (or calling `unwind', which none of our back-ends
support (it's lowered into nothing...)), call the @llvm.trap() intrinsic
instead. This may not conform to specific rules of a particular language, but
it's rather better than infinitely looping.

<rdar://problem/9175843&9233582>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@129302 91177308-0d34-0410-b5e6-96231b3b80d8
lib/CodeGen/SjLjEHPrepare.cpp