Redundant OutputComplete elilmination doesn't look like it's quite in the right place.

Moving the DxilMod init code was unnecessary, hlsl::OP can be retrieved from the m_pHLModule instead, as the others do.

See

I don't think this code should have been moved up here, above the comment that related to this step: // High-level metadata should now be turned into low-level metadata.
I think the code that gets rid of OutputComplete calls with zeroinit handle should probably be moved to some other (DCE-like) pass, later (since further optimization may produce this as well).