C++: Prevent an Expr from stepping to itself in IR dataflow

[MathiasVP]

While investigating some query changes, I noticed that it can be the case that a single local step can step from a node n1 representing the an expression, to a node n2 that represents the expression's first conversion. Because n.asExpr then resolves to the unconverted expression, the result will be an expression e such that localExprFlowStep(e, e) holds 😱.

The fix is to add another level of recursion that ensures that we always step to a different expression. This is done in a328565. 52bf39b fixes a performance shown by the join-order detection in DCA. It turns out the * in the localExprFlow relation wasn't turned into a fastTC by the compiler, and it resulted in quite a few iterations that looked like:

Evaluated recursive predicate #DataFlowUtil#47741e1f::ExprFlowCached::localExprFlowStep#2Plus#bf@aafb49jn in 916ms on iteration 2 (delta size: 844629).
Evaluated relational algebra for predicate #DataFlowUtil#47741e1f::ExprFlowCached::localExprFlowStep#2Plus#bf@aafb49jn on iteration 2 running pipeline standard with tuple counts:
    923848     ~0%    {2} r1 = SCAN #DataFlowUtil#47741e1f::ExprFlowCached::localExprFlowStep#2Plus#bf#prev_delta OUTPUT In.1, In.0
  23112681  ~1725%    {2} r2 = JOIN r1 WITH DataFlowUtil#47741e1f::ExprFlowCached::localExprFlowStep#2#ff ON FIRST 1 OUTPUT Lhs.1, Rhs.1
    928088    ~14%    {2} r3 = r2 AND NOT #DataFlowUtil#47741e1f::ExprFlowCached::localExprFlowStep#2Plus#bf#prev(Lhs.0, Lhs.1)
                      return r3

So the last commit transforms that * into a fastTC manually.

[jketema]

LGTM assuming DCA is happy.

[MathiasVP]

DCA looks fine 🎉

[jketema]

More than fine 😄