Avoid repeatedly evaluating subexpressions when checking for unsequenced

operations in the case where evaluating a subexpression fails. No functionality
change, but test/Sema/many-logical-ops.c gets ~100x faster with this change.

llvm-svn: 184489