Minor optimization to constant evaluation: don't bother computing expr source

locations for diagnostics we're not going to emit, and don't track the subobject
designator outside C++11 (since we're not going to use it anyway).

This seems to give about a 0.5% speedup on 403.gcc/combine.c, but the results
were sufficiently noisy that I can't reject the null hypothesis.

llvm-svn: 152761