Each Bad node has a proper mode now.

Rematerialisations involving multiple nodes and Projs (e.g. Proj->Load) got miscounted.
Now a rematerialization always counts as 1.

Resetting a node's schedule info and not properly removing it from the schedule is questionable anyway.

The node was just created, so it is not scheduled.

Since the prologue is added late, there are no Save nodes during transformation.

Since the epilogue is added late, there are no Leave nodes during transformation.

This is the same correction as for ia32.
First construct all epilogues, then the prologue.
Check there, whether the IncSP has no users and add a keep edge in that case.

If the start node was deleted, something is very wrong.

There are no tests for this, so nobody noticed that the result is broken.

Now eat_immediate() simply returns, whether it was successful.
So there is less code (duplication) and only one pass is necessary instead of two.

Either it is not supported (-Address) or should not occur due to normalisation (Sub(x, Const)).

This hack had many unwanted effects where nodes following the IncSP were
unnecessary delayed for a long time. We should rather improve the ia32
Push/Pop peephole optimizations.