Resolution of mixing rlibs and dylibs
When producing a final artifact, such as a dynamic library, the compiler has a choice between linking an rlib or linking a dylib of all upstream dependencies. The linking phase must guarantee, however, that a library only show up once in the object file. For example, it is illegal for library A to be statically linked to B and C in separate dylibs, and then link B and C into a crate D (because library A appears twice).
The job of this module is to calculate what format each upstream crate should be used when linking each output type requested in this session. This generally follows this set of rules:
- Each library must appear exactly once in the output.
- Each rlib contains only one library (it’s just an object file)
- Each dylib can contain more than one library (due to static linking), and can also bring in many dynamic dependencies.
With these constraints in mind, it’s generally a very difficult problem to find a solution that’s not “all rlibs” or “all dylibs”. I have suspicions that NP-ness may come into the picture here…
The current selection algorithm below looks mostly similar to:
- If static linking is required, then require all upstream dependencies to be available as rlibs. If not, generate an error.
- If static linking is requested (generating an executable), then attempt to use all upstream dependencies as rlibs. If any are not found, bail out and continue to step 3.
- Static linking has failed, at least one library must be dynamically linked. Apply a heuristic by greedily maximizing the number of dynamically linked libraries.
- Each upstream dependency available as a dynamic library is registered. The dependencies all propagate, adding to a map. It is possible for a dylib to add a static library as a dependency, but it is illegal for two dylibs to add the same static library as a dependency. The same dylib can be added twice. Additionally, it is illegal to add a static dependency when it was previously found as a dylib (and vice versa)
- After all dynamic dependencies have been traversed, re-traverse the remaining dependencies and add them statically (if they haven’t been added already).
While not perfect, this algorithm should help support use-cases such as leaf dependencies being static while the larger tree of inner dependencies are all dynamic. This isn’t currently very well battle tested, so it will likely fall short in some use cases.
Currently, there is no way to specify the preference of linkage with a particular library (other than a global dynamic/static switch). Additionally, the algorithm is geared towards finding any solution rather than finding a number of solutions (there are normally quite a few).