It impacts only the bitcode and not the final binary AFAIK. The issue is with debug-info not being stripped out completely as expected. It comes from a change from LLVM upstream that was made (in 3.9) to keep source locations for providing better optimization diagnostics. It interacts badly with the debug info stripping and apparently wasn't caught. It is used by the optimizer to tell you e.g. that a specific loop can't be vectorized.
See http://llvm.org/devmtg/2016-11/Slides/Nemet-Compiler-assiste... for more details about how it's used.