Two problems with gimplify_init_constructor

[Richard Kenner]

    I'm thinking now, actually.  It also interferes with a patch I have
    pending to once again share stack slots.  Something that's important
    because we can no longer build kernels with 4k stacks with 3.5.

I really worry about pessimizing code.  It should at least be kept for
objects that are real constants (e.g., never assigned) and can be
safely kept for scalars.