There are cases where we can normalize in 2 directions and both have
their merits. Just do both in this case. The frontend should at some
point set IRG_GRAPH_STATE_NORMALISATION2 to trigger the 2nd direction.

Fixes opt/ioet.c

* set_irg_outs_inconsistent is unnecessary as exchange does that already
* "useless if optimization" was a confusing comment
* more precise type: int -> block_flags_t

The remove_empty_blocks function must be used for a pre-walker, otherwise
it might be called on blocks, which were optimized away. Consider infinite
loops and the Keep edges:

  \ ___
   B   |
    \  /
    Empty
    /
   / Keep Edge!

In this example, we must not call remove_empty_blocks on Empty,
after we called it on B, but a post walker would do that.

- renamed clear_link() to clear_link_and_mark_blocks_removable()
- renamed optimize_conds() to optimize_ifs()
- simplified has_*() implementations
- made some walker POST walker
- invalidate out edges
- and add some doxygen docu
- fixed some typos
- adapt to firm indentation style

The get_n in the foreach loop is a macro and not a use of the function.

merge start-block successor with start-block in backend (necessary after equivalent_node_Block is gone)

Since Bads have modes now the check for mode_X initialized the bitinfo,
which results in incorrect reachability information.

Broken with csmith 2.0.0 seed 909842313
in libFirm 1953ce88