[r13990]

[r13985]

[r13983]

[r13982]

[r13981]

[r13980]

[r13979]

[r13975]

[r13973]

[r13972]

[r13971]

[r13970]

[r13969]

[r13968]

[r13964]

[r13963]

[r13962]

[r13961]

[r13960]

[r13959]

[r13958]

- pred was used to store two values :-(

[r13957]

[r13956]

[r13955]

[r13954]

[r13953]

[r13951]

__get_irn_generic_attr() added to speed up backend

[r13950]

[r13949]

- we cannot remove keep-alives here because only blocks are visited, removing KA requires node visiting
- now its clear why always moving Phis from pred block was a bad idea (and was fixed with the strange replace). It must be only done if pred dominated the current block.

[r13930]

Removed the old "defer" code, this code was wrong (and the old replacement code as well). After that, the original code was revived.
However, now ALL Phi nodes are copied even those unused.

Added keep-alive optimizer killing useless keep-alives.
This kills some created useless nodes.
It does not help with endless loops :-(

Possible fix: use back edges to check if there are users ...

[r13929]

[r13927]

[r13926]

[r13916]

emit execfreqs for all blocks. ignore the am_flavour setting and look at the actual values of the AM variables

[r13915]

[r13914]

[r13913]

[r13912]

[r13911]