The generator stuff is not speed critical and this avoids polluting our
source directories.

This way we are independent of the type of specfile.

Otherwise the jFirm binding generator has a hard time determining the
class hierarchies because there is no __base__ field.

Also improve the comment a bit

* Consider the number of transformed inputs instead of original inputs.
  This differs, when immediate inputs are present.
* Do not consider the memory input.
* Only scan the outputs, if they are stricly fewer than the inputs.
  This avoids adding AnyVal inputs in case of equality.

This corrects x86code/asm_immediate.c.

This corrects x86code/asm_test16.c.

There might be multiple constraints for the same input position, not just output.

This fixes opt/fehler288.c.

Setting the register class got lost.

ia32: Remove strange case, which allowed the whole register class, when actually no register is allowed.

This lets the register allocator fail instead of silently generating invalid code.

This fixes opt/fehler287.c.

We should reach all blocks via control flow or keep alive edges.

tarval_neg() flips the sign of the nan, so you do not get the same nan out.

It was just copied from test_binop_nan_() and now tests the same thing n times.

Now the mantissa is correctly transferred, i.e. it is aligned at the most significant bit.

All unreachable blocks are cut off by apply_cf() already and all Phis adjusted accordingly.
Therefore no nodes in unreachable blocks can be reached anymore.

This fixes opt/combo11.c.