[r27418]

  have to construct explicit Keep nodes when killing barriers.
  This should fix the no-omit-frame-pointer missing IncSP problems.

[r27415]

[r27412]

[r27411]

combo, just combo's evaluators must be synced between compute_Cmp() and
compute_Proj_Cmp().

[r27410]

[r27408]

[r27407]

[r27402]

[r27401]

[r27395]

[r27394]

Create the in array before the node is constructed instead of copying the in array an set the operands afterwards.

[r27393]

[r27392]

[r27391]

- extende beabi so you can specify different registers at the call site
  and the beginning of a called function
- adapt all backends for API changes

[r27390]

[r27389]

Loop inversion: Fixed bug in find_condition_chain and construct_ssa, which prevented SPEC/perlbmk to compile.

[r27388]

[r27387]

- use Phi list to speed up algo
- initially clear the links of all nodes (block walker visits only initially reachable ones)
- track resources
- invalidate analysis info at end

[r27386]

[r27385]

[r27384]

[r27383]

[r27381]

[r27377]

...

[r27374]

[r27373]

[r27372]

[r27371]

[r27370]

[r27369]

[r27368]

This fixes testsuite/backend/floatset.c ...

[r27367]

[r27366]

fix consconfirm missing nodes in its statistics and worse not invalidating out-edges because of that

[r27365]

Partly undo r27277 and add a prototype for (formally undocumented) find_value(), needed by the edgjfe FE.

[r27363]

[r27362]

[r27360]

[r27359]

[r27356]

[r27352]