If we don't, values that aren't precisely representable in f16 could
be used as-is in a promoted f32 operation, which would produce
incorrect results.
AArch64 had the correct behavior; add a focused test.
Fixes http://llvm.org/PR26871
llvm-svn: 268700