This patch fixes ICE-on-valid regressions 22013 and 22098 (plus
duplicate 22050) in the manner I suggested in
<http://gcc.gnu.org/ml/gcc-patches/2005-06/msg01463.html>: adding a
langhook used by recompute_tree_invarant_for_addr_expr to go from a
COMPOUND_LITERAL_EXPR to the decl within and removing special-case
code in build_unary_op which always treats the address of a compound
literal as TREE_INVARIANT and TREE_CONSTANT.
Bootstrapped with no regressions on x86_64-unknown-linux-gnu, mainline
and 4.0 branch. OK to commit to mainline? OK for 4.0.2?