The specific problem here is that I am going to be adding some code to
SemanticARCOpts that eliminates reborrows and may need to create new phi
arguments and thus add arguments to edges. The weird thing about this is that
doing so actually requires us to create a new terminator!
This means that subtle pointer invalidation issues can occur here. To work
around that we store our terminators as SILBasicBlock, operand number since we
can always immediately find a terminator from its basic block. If we do not have
a terminator, we keep on just storing the SILInstruction itself.
NOTE: This only saves us from additive changes. Deletions are still an issue.
This makes it easier to understand conceptually why a ValueOwnershipKind with
Any ownership is invalid and also allowed me to explicitly document the lattice
that relates ownership constraints/value ownership kinds.
This patch moves state from the main SemanticARCOptVisitor struct to instead be
on a context object. Sub-transformations should not need to know about the
visitor since how it processes things is orthogonal from the transformations
themselves.
