This mainly adds comments explaining the situation and avoids an extra
graph walk.

This works around a problem with modeling control flow for calls to
setjmp and similar functions.

This reverts commit 3d1d2c43.

Creating a Bad node does not imply that it has users or is reachable
within an irg. While clearing the property is certainly correct as a
conservative approximation, it should not be necessary.

fc_val_from_ieee754 simply does not work correctly for denormalized
input values if the desired result is smaller than a long double.
However, converting to an fp_value using the mantissa/characteristic
information of a long double and later using fc_cast to convert to
the desired target format works.

This node was used by OO code in the past, but newer versions of liboo
have their own node and don't need it anymore.

CopyB is normally used to transfer data on the stack or to enable compound
value calling conventions. Exceptions don't happen in these cases.

I don't really understand what it was supposed to do (because nodes
using/producing should be pretty much immovable anyway because of their
dependencies, so an additional pinning type appears unnecessary).
In practice there was no code differentiating between mem_pinned and exc_pinned.

These SymConst nodes have an attached entity whose owner type is not a segment.

- This avoids missing cases when new arithmetic modes are introduced.
- This also fixes a wrong case in mode_overflow_on_unary_Minus()

- The backends can simply perform the finish_graph() steps in the emit()
  phase.

- You can now construct the lookup table separately and perform the
  lowering selectively on irgs (instead of the whole program at once)
  while reusing the lookup table.
- Simplified API a bit.
- Remove i_mapper_RuntimeCall. All users are simpler by doing the
  transformations directly instead of filling in runtime_rt
  structures...

Use the same style as the SPARC 64bit lowering, making the additional
handle_intrinsics step unnecessary now.

before_abi() callback was removed, you can simply put that code at the end
of the init() callback. custom_abi flag was removed, if a backend needs the
be_abi_introduce() stuff (formerly custom_abi == false) then it can call
it at the end of the init() callback.

The code was only working on ia32 and contains several ia32 specific
bits, so keep it in the ia32 backend for now. At least the amd64 backend
will require a different implementation of PIC.

The result of gs_matrix_gauss_seidel() never is negative.