Liveness analysis may not be adequate for SSA form

[dibyendumajumdar]

The discussion in the thread below indicates that the standard liveness calculation may not be handling Phi functions correctly.

Currently in our implementation, Phi's do not contribute uses (which is correct). But we do not do anything special with the values that feed into the Phi.

https://www.reddit.com/r/Compilers/comments/9qt31m/liveness_and_ssa/

[dibyendumajumdar]

The new approach is based on formula described in
Computing Liveness Sets for SSA-Form Programs
Florian Brandner, Benoit Boissinot, Alain Darte, Benoît Dupont de Dinechin, Fabrice Rastello

The implementation is the unoptimized simple one.
However, we have a modification to ensure that if we see a block
which loops to itself and has Phi cycles, then the Phi is only added to
PhiDefs.