The common_transform stuff was meant to be reused by the PBQP code
selection which is not maintained anymore.

It returns a null pointer instead of a ia32_NoReg_GP.

This avoids the custom emitter, which just duplicated what the generic machinery does.

Some callers did not handle it.
They cannot be permuted (e.g. Jcc).

ia32: Use arch_get_irn_register() instead of arch_get_irn_register_in() where the operand is already available.

The noreg is always different to any output register.

This properly models the data dependencies and avoids keep edges.

ia32: Merge the attributes use_frame, need_stackent, need_32bit_stackent and need_64bit_stackent into one attribute.

* Out of the 16 possible combinations of these four boolean attributes, only four are sensible.
* This correct the latent bug, that not all of these attributes were compared in ia32_attrs_equal_().

* No \n at end; panic() adds one
* No punctuation mark at end
* Lowercase first letter

All other structs holding a register class call it cls, too.

ia32: Directly put the bitfields in ia32_attr_t instead of having another struct needlessly wrapping them.

All backends emit nothing for Phis and be_Keeps, so set this centrally.

The previous approach put the irg link into the block attributes, so
while saving memory an extra indirection was necessary. Some testing
with valgrind massif+callgrind revealed that a direct link increses peak
memory usage by ~1% while also increasing runtime by ~1%. More
importantly the code gets a bit simpler and some strange corner cases can be
removed with the direct link.

This should fix strange overly long immediate values emitted in the
assembler sometimes.