This should fix 403.gcc.
A SBB with AM could not fulfill its should_be_same contraint, so the AM gets split from the SBB.
The Load of the AM gets assigned the out register of the SBB, which violates the !in_r5 constraint of the SBB.
Therefore it's necessary to turn the SBB into NOT+ADC.

Correct inaccurate comment: The only corner case of wrong CF in NEG+ADD as replacement for SUB is if both operands are zero.

use a 'low-tech' solution for emitting 8- and 16-bit register names. This also fixes a recently introduced bug in ia32_build_8bit_reg_map_high

[r28168]

[r28139]

[r28127]

[r28038]

[r28037]

eliminate the unnecessary and especially confusing concept of a 'code_generator' an isa-interface is enough

[r28009]

[r27297]

- rework ir_phase API
(sorry for mixing these 2 things into 1 commit)

[r27285]

[r27257]

Correct a subtle bug in the ia32 backend: Sub(x, x) triggered that the Neg+Add trick was used, which resulted in incorrect code.

[r27256]

- cleanup ir_spec generation a bit

[r27251]

[r27153]

panic instead of wrong results for Set and CMov with float compare. Rename Set to Setcc and CMov to CMovcc

[r26950]

- Add a generic requirements+register+flag dumper to bearch; This avoids all
  backends/benode having their own slightly different dumpers
- Lots of cleanups in the backends along the way

[r26542]

[r26535]

[r26317]

[r26239]

- SymConst's are now ALWAYS placed in the start block

[r26236]

Generate new_bd_* instead of new_rd_* functions in the backend.  The nodes are always created on the current irg.

[r23535]

[r23534]

[r23474]

	Node flags and node registers are stored in a generic backend_info struct now
	instead of every part of the backend doing custom (and slow) stuff

[r23142]

[r22889]

[r22771]

[r22754]

[r22699]

[r22697]

[r22670]

It does not make any sense to remove AM, which loads a non-GP value to fix register conflicts with base and index.  This situation plain does not exist.  Base and index are always GP.

[r22669]

[r22666]

[r22658]

[r22503]

  (where this is specified?), so do NOT overwrite it (analog to
  turn_back_am())

[r22485]