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]

[r23138]

  Proj by UnknownM, this would destroy the graph, so skip the
  predecessor in that case

[r23136]

- remove by_all_const: always saturate the type

[r23132]

  solution seems to be fixing the type of the Proj(Cond)

[r23130]

[r23129]

[r23123]

[r23120]

  this reduces the graph and fixes the problem of dead cf to dead blocks
  (which assert else)

[r23114]

- removed Cond hack: it does not help. Instead cond nodes
  are automatically handled AFTER all data nodes by
  placing them on an extra list
- BugFix: apply_cf() now remove bad cf even if the
  resulting block has only one predecessor and
  cannot be combined with its predecessor
- BugFix: data nodes computing Top must be replaced
  by Unknown, not Bad

[r23111]

[r23104]

[r23083]

- Hopefully fixed handling of unknown nodes for Cmp and Cond:
  - let cmp nodes calculate Top like all othe data nodes: this would let
    Mux nodes to calculate Unknown instead of taking the true result
  - Let Cond(Top) always select FALSE/default: This is tricky. Nodes are only reavaluated
    IFF the predecessor changed its type. Because nodes are initialized with Top
    this never happens, let all Proj(Cond) be unreachable.
    We avoid this condition by the same way we work around Phi: whenever a Block
    node is placed on the list, place its Cond nodes (and because they are Tuple
    all its Proj-nodes either n the cprop list)
- BugFix: Proj-Nodes are also pinned if theire predecessor is pinned
- while combo is running, add the vcg dump hook, so dumped graphes contains the
  partition and the type of a node
- put simpler condition first in an if-cascade

[r23082]

[r23058]

[r23051]

  newly created edge, so add a panic if this happens
- put freed edges into a free-list, reducing memory count used for edges

[r23050]

[r23047]

  kills Return(Top) nodes ...

[r23035]

- do not replace Phi(Top, ..., x, ..., Top) -> x, this might produce non-strict programs
  (does NOt happen as long WITH_UNKNOWN is undefined)

[r23033]

- compute: check the control input only for pinned nodes (prepare for
  global_cse run)

[r23031]

Fixed bug in conv-conv optimization: We have to assure, that we don't change the mode when deleting the conv.

[r23009]

[r23005]

Added trivial test case where optimize_reassociation will never get to a fixpoint (result will always be non-zero)

[r23004]

[r22985]

[r22984]

[r22983]

[r22982]

[r22981]

[r22976]

[r22975]

[r22974]

[r22963]

  when splitting a partition containing followers, these
  might stay on the "looser" side and must be constant-propagated
  then to allow them to split out.
  This might be a fix for a self-made problem. We do the follower ->
  leader transition on both sides while the original algo "might"
  do than on the winner side only...

[r22962]

[r22941]

[r22939]

[r22938]

[r22937]

Remove unused attribute const arch_env_t* arch_env from struct TEMPLATE_code_gen_t (no actual backend has it either).

[r22935]

[r22934]

[r22932]