[PATCH] Fix PR19195

[Richard Smith]

  This doesn't look right to me. If the initializer hasn't been parsed by the time we get to IR generation for the default constructor, something has gone *very* badly wrong.

  The right fix here is probably to implement the proposed resolution for CWG issue 1396 -- lazily instantiate default initializers for non-static data members as we do for default arguments.

http://reviews.llvm.org/D3243