First it is unclear what is a good way to deconstruct a program into extended
basic blocks, also none of our optimizations works on extended basic
blocks.

It probably was true when the start block had itself as predecessor.

This was used as a hack to mark unreachable code by some people, but it
is prefered to simply use a new Block without predecessors, or set NULL
if you have to.

This commit removes the strange differentiation between
ir_visibility_external and ir_visibility_default. We now only have
ir_visibility_external for all symbols visible across compilation units.
You may or may not attach graphs/initializers to them.

(This happens when you break on a Sel creation and the firm debugger
 tries to tell you about the created node)

- The direct call did not respect the get_opt_constant_folding() flag

(otherwise CSE merges temporary Phi nodes without inputs)

(or multiple ir_init/ir_finishs will fail)

Types "own" entities, so they should also free them when they get freed.

It was a member of ir_prog before but not correctly handled.

The first block in a new ir_graph is not an immature block anymore. The
end block is matured in irg_finalize_cons() now (since maturing blocks
twice doesn't hurt this shouldn't break existing code).

The opcode list was a member of irprog before which wasn't really
handled consistently. Also make sure opcodes are properly freed at
ir_finish().

unknown_type, code_type, none_type, unknown_entity reference are hold in
the irprog now. This makes handling more consistent since now all types
and entities are equally part of irprog.

Replace with equivalent pset_ptr stuff

It should only be used by construction algorithms

public functions are already documented in their headers, no need to
repeat the comment at the definition site.

They're pointless with the new switch representation introduced a while
ago.

memops are nodes that have memory inputs, you can generically query them
for their memory input. We can also get rid of get_fragile_op_mem in
favor of get_memop_mem.