Cleanup EscapeAnalysis::ConnectionGraph::initializePointsTo.

[atrick]

Remove cruft that I added in the previous commit. This eliminates
unnecessary cleverness so initializePointsTo is now very simple.

Add hand-coded SIL test cases to somewhat verify that the new
algorithm works.

[atrick]

@eeckstein I had added some junk to initializePointsTo in a hurry right before you reviewed it. I was trying to remove even more edges but there was no basis for doing that. This PR removes that junk. Both initializePointsTo and mergePointsTo should be utterly straightforward to read now.

This should address your review comments. You're right that this small loop is only relevant when defer-edge cycles occur:

    // For all nodes visited in step 1, if any node was not backward-reachable
    // from a pointsTo edge, create an edge for it and restart traversal. This
    // only happens when step 1 fails to find leaves in the defer web because
    // of defer edge cycles.
    while (!updatedNodes.empty()) {
      CGNode *node = updatedNodes.nodeVector.pop_back_val();
      if (!node->pointsTo) {
        pointsToEdges.push_back(node);
        break;
      }
    }

The original algorithm was problematic because it would add pointsTo edges even when no cycles exist in the graph. It would do so whenever connecting an uninitialized web to an already-initialized web.

[atrick]

@swift-ci test

[eeckstein]

thanks, LGTM!

[atrick]

@swift-ci smoke test