Backwards analysis required for ForceDominance

[skrynski]

At:

rir/rir/src/compiler/opt/force_dominance.cpp

Line 112 in 0040223

if (a.isUnused(mk)) {

To determine whether a promise is unused anywhere else and therefore can be lowered, a forward analysis does not suffice. See file attached. If a promise is not used 'so far' within a particular branch, it does not mean it is not used in any other branches and, therefore, is not necessarily a candidate.
FDGVN.txt

Notice that, in the file, the Deopt BB comes from a dead CheckPoint. Maybe we should schedule a Checkpoint cleanup at the very end of ForceDominance.

Also, we should revisit the method isUnused.

This is the only issue I did not solve when working on #1250. The bug is exposed when the budget is lifted from the pass scheduler. Sometimes ForceDominance and GVN will fight each other while doing and undoing changes, getting stuck in a loop. For example, ForceDominance would push MkArg instructions down to exit BBs, and GNU would push them back up.
To expose the bug, run : PIR_WARMUP=2 PIR_DEOPT_CHAOS=50000 bin/R -f ../../external/custom-r/tests/Examples/utils-Ex.R