Debugify is generally used to detect cases where debug info has gone missing, though it also picks up a lot of false positives - @ntesic made some effort to remove those false positives a while back, but it’s not in trunk right now. In any case, it doesn’t have a way of generating a test case from a known bug.
What I’d recommend for generating a test case where you know the source code responsible would be to add an assert at the point of the broken code, either just assert(false) or using any necessary preconditions for the bug (e.g. assert(SI->getDebugLoc()) in this case, to ensure only cases where there’s an existing debug loc get detected), and then just compile some test programs until the assert is hit - a build of clang itself or of the LLVM test suite usually works. Once you’ve hit that error, you can use creduce or llvm-reduce to reduce the input until you have your reduced test case.
More broadly regarding missing line numbers, my brief dip into this issue indicates that there are a lot of places where line numbers are dropped unnecessarily, usually due to a new instruction being created to replace an old without the debug line being copied across; this is a fairly common pattern. I’ll hopefully put out a more comprehensive post later, but right now I believe that debug lines should probably be a required parameter for creating instructions - if an instruction is going to have no debug line, it should be explicitly stated rather than the default.