[clang] [Serialization] Load Specializations Lazily (1/2) (PR #76774)

[Jonas Hahnfeld via cfe-commits]

================
@@ -1249,3 +1249,5 @@ void ODRHash::AddQualType(QualType T) {
 void ODRHash::AddBoolean(bool Value) {
   Bools.push_back(Value);
 }
+
+void ODRHash::AddInteger(unsigned Value) { ID.AddInteger(Value); }
----------------
hahnjo wrote:

The review related to `ODRHash` is this one: https://reviews.llvm.org/D153003

In short, my understanding is that `ODRHash` gives the following guarantee: If the hashes are different, there is guaranteed to be a ODR violation. In the other direction, if two hashes are the same, the declarations have to be compared in more detail, ie there may or may not be an ODR violation.

For the specializations, we need the opposite: If two template arguments are semantically the same (*), they *must* hash to the same value or otherwise we will not find the correct bucket. On the other hand, two different specialization arguments may have the same hash, that's fine for the map data structure.

Now the additional caveat (*) is that "semantically the same" is not the same congruence as "no ODR violation". In https://reviews.llvm.org/D153003 we discuss `using` declarations, but IIRC it's also possible to construct problematic cases with (nested) namespaces, top-level `::` prefixes, and template template parameters. Taken together, my conclusion from the discussion above is that `ODRHash` is simply not the right method to find template specialization parameters in a map.

https://github.com/llvm/llvm-project/pull/76774