Implementation would be by modifying clang. Traverse the clang ast emitting the new language instead of llvm IR.
You wouldn't get idiomatic code out but with some effort you'd get rust/d/c/other which clang compiles to the same IR as the original.
How much refactoring is warranted afterwards would depend on how much effort you put in to recreating templates / header files / modules etc on the fly.
I'm not sure I'd choose to do this myself if I was in Google's position but it would be tempting.
you would end up with llvm-ir-like code that was only technically rust/d insofar as the compilers could handle it. it would not be human-readable or maintainable at all; indeed, it would be harder to convert that to idiomatic rust/d than hand-translating the c++ code. and really, all you would gain would be getting rid of the c++ compiler and ending up with worse code.
the point of carbon is that you can incrementally migrate your c++ program to it in place, and the migrated code will end up easier to maintain than the original c++.