ilya-biryukov wrote: > A special handling for `PreferredName` looks slightly add-hoc to me. What I > thought is to delay the deserialization of `attributes` after the > deserialization of the declaration. If this is too optimistic, I am wondering > if we can delay reading of some sorts of attributes only, which is more > generic.
I don't have concrete examples in my head, but I would be cautious of doing this for **all** attributes. Intuitively, it feels that there must be existing attributes that we somehow would depend on in the code during serialization. However, it might be worth a try to actually find real examples from running tests (even if only to prove that my intuition is wrong). That being said, having a concept for "sort of attributes that we can delay reading" seems like a really good idea. I am now highly suspicious of any attribute that stores `TypeSourceInfo`, they should probably hit the same bug in certain circumstances; there's no nothing specific to `preferred_name` itself other than the fact that it actually tends to end up recursively referencing its own type through a typedef more often than others. https://github.com/llvm/llvm-project/pull/122250 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
