Fixed the "naive and wrong regex stubs", so it has the right signature and returns the right value/error code now.

Simplify: if (!irn_visited(x)) { mark_irn_visited(x); ... }  ->  if (!irn_visited_else_mark(x)) { ... }.

This is asserted in the next statement anyways

* get_associated_type is unnecessary
* is_lowered_type is an internal inline function now
* set_lowered_type is an internal inline function now

This fixes opt/jumpthreading3.c.

* removed unnecessary call_tp variable
* removed shadowed tp variable (twice)

Similiar to 9d207c0 the call lowering should always use the current
entity type, instead of the associate type, which may or may not be
a lowered type.

- is_lowered_type is not helpful, because the flag may already be set
  by lower_calls before. This flag should probably be removed in general.
- hence lower_mtp must not check for is_lowered_type
- lower_Start must update the entity type, but not lower_irg,
  otherwise lower_Start can not decide, whether a lowering is necessary.

If there is an unreachable loop (i.e. there is no loop entry edge)
then there may is a cycle of blocks with only one predecessor.
Since the algorithm handle this case by returning the value of the predecessor,
this causes an endless recursion.
The fix uses the visited flag to detect such cycles.

* use a local variable for the walker function
* use setter function for block settings

The node walk used before was inefficient,
since everything but blocks was ignored.

Additionally, this fixes a bug, because ssa_cons_finish already used a
block walk already, such that unreachable blocks were immatured,
but not matured in the scalar-replace optimization. This bug could
also be avoided by ensuring cfopt before scalar-replace, but keeping
ssacons start and finish symmetric seems to be more robust.