Extended Description

We've almost upgraded to LLVM 6 in rust-lang/rust 1 but we seem to have hit what is hopefully our last blocker, a segfault when optimizing IR on MSVC. The attached IR will non-deterministically segfault in opt during the LICM passes. I've reproduced this with:

for i in `seq 1 1000`; do echo $i opt -O2 -o /dev/null foo.ll || break done 

Sometimes it takes up to 100 runs to crash, but I've been able to crash on a debug and a release build of LLVM, both on Linux and on Windows.