It's both harder and easier. There is a mechanical transformation, without un- or approximately translatable idioms like natural language. On the other hand, the dependency chain is much more complex - with something like link-time optimization, a change to one part of the code can completely change the result (for instance, if it suddenly allows inlining of a function everywhere). There is also the problem of, if not "idiom translation", "idiom generation" - people write code in a particular style that may not be captured by the generated output, even if it compiles the same.
Targeting something like Clang specifically, where you have access not only to the assembler & a potential source, but also a whole AST & intermediate data structures, would be pretty interesting.
Targeting something like Clang specifically, where you have access not only to the assembler & a potential source, but also a whole AST & intermediate data structures, would be pretty interesting.