Add basic Serde serialization capabilities to Stable MIR by sskeirik · Pull Request #126963 · rust-lang/rust (original) (raw)

This PR adds basic Serde serialization capabilities to Stable MIR. It is intentionally minimal (just wrapping all stable MIR types with a Serde derive), so that any important design decisions can be discussed before going further. A simple test is included with this PR to validate that JSON can actually be emitted.

Notes

When I wrapped the Stable MIR error types in compiler/stable_mir/src/error.rs, it caused test failures (though I'm not sure why) so I backed those out.

Future Work

So, this PR will support serializing basic stable MIR, but it does not support serializing interned values beneath Tys and AllocIds, etc... My current thinking about how to handle this is as follows:

  1. Add new visited_X fields to the Tables struct for each interned category of interest.
  2. As serialization is occuring, serialize interned values as usual and also record the interned value we referenced in visited_X.
    (Possibly) In addition, if an interned value recursively references other interned values, record those interned values as well.
  3. Teach the stable MIR Context how to access the visited_X values and expose them with wrappers in stable_mir/src/lib.rs to users (e.g. to serialize and/or further analyze them).

Pros

This approach does not commit to any specific serialization format regarding interned values or other more complex cases, which avoids us locking into any behaviors that may not be desired long-term.

Cons

The user will need to manually handle serializing interned values.

Alternatives

  1. We can directly provide access to the underlying Tables maps for interned values; the disadvantage of this approach is that it either requires extra processing for users to filter out to only use the values that they need or users may serialize extra values that they don't need. The advantage is that the implementation is even simpler. The other pros/cons are similar to the above.
  2. We can directly serialize interned values by expanding them in-place. The pro is that this may make some basic inputs easier to consume. However, the cons are that there will need to be special provisions for dealing with cyclical values on both the producer and consumer and global values will possibly need to be de-duplicated on the consumer side.