GitLab

It is the job of the frontend to keep loops.
These hacks here only worked for some, but not all cases.
In particular a more clever SSA-construction algorithm might not place redundant PhiMs, which would render memory before memory-less loops unkept.

get_irn_op_() checks the same.

sched_add_after() checks this, too.

I wonder, how long this code was not executed.

Use edges_reroute_except() instead of edges_reroute() to avoid fiddling with the memory input of the inserted vfld.

- The src irn of a edge cannot be null.
- A node cannot be after itself in the schedule, so the later test will fail.

Always inserting before the cf node results in the same order as first inserting before the cf node and then inserting after the last inserted node.

Before it was calculated from op_idx.

It is the same as N_ia32_st_REGS - depth.

Copy it before x87_kill_deads() instead of after it, so x87_kill_deads() does not need to conditionally copy it.

Just flush the fp stack, only check for the correct number of fp values in debug builds.

Now more than one node can be pushed through a Perm.
After pushing the first node through a Perm, we next inspected the Perm itself (the new schedule predecessor of the moved node), which obviously cannot be pushed through itself.

Do not leave out the last (highest numbered) irg.

some inputs had "no requirement" set instead of the correct one

irverify already checks the dominance property