The optimization a + (-0.0f) -> a was being misapplied to a + (+0.0f) in the vector case (because

we weren't differntiating floating-point zeroinitializers from other zero-initializers)
which was causing problems for code relying upon a + (+0.0f) to, eg, flush denormals to
0. Make the scalar and vector cases have the same behaviour.

llvm-svn: 177279