Two problems with gimplify_init_constructor

[Richard Kenner]

These come up with the ACATS test cc51006.

(1) When an operand is a STRING_CST, the count of nonzero and nonconstant
elements counts the STRING_CST as one operand, but count_type_elements
counts it as the size of the string.  This is easy to fix.

(2) In the "cleared" case, we can't have the result marked RTX_UNCHANGING_P
for the same reason we can't do it in the pre-tree-ssa code.  This has
to handle both the simple readonly cases and when there are readonly fields.
Should we just do the same thing as for RTL with the corresponding check?