This reverts commit d0f23ce9.

(was already included for the benchmark runs for the M2 report)

This reverts commit a0119e59.

Fix applied at wrong place

When code motion past side exits is NOT combined with general register
renaming, the renaming necessary to make it safe to duplicate
instructions may cause final restoration code to be inserted.
Final restoration code is trated specially by its own function.

Before, downward scheduling would sometimes give an unverifiable result
because instructions are duplicated before restoration instructions and
aliasing are applied. If for example a loop increment was moved below
multiple side-exits, the restoration code would still mess up the
correct duplication of the instruction because the aliasing logic for
later side exits would overwrite this change.
As a work-around, this commit ensures that the aliasing logic for
side-exits is applied in-order. This may leave aliasing code that is
utterly useless, which should be cleaned up by DCE.

in debug mode.
Compiling glpk-4.65 exhibited this problem but not when using the
ILP-based instruction scheduling.

time

with regard to scheduling constraints necessitated by code outside of
the superblock. This behavior more closely matches that of the other
dependency calculations giving a more fair comparison.

information.
Correct in the sense that they match the expected scheduling
constraints.

If the Icond is the last instruction of the path, the live_renames
includes all the live renames for all successors (s_output_regs).
However not all of those need to be stricly restored before the Icond.
Some can be restored afterwards.

This becomes necessary once we allow register renaming for the
originally last instruction of the path.