[clang] [serialization] no transitive decl change (PR #92083)

[Ilya Biryukov via cfe-commits]

ilya-biryukov wrote:

Oh, I didn't realize you were eager to land this in Clang 19, thanks for sharing that.
In that case, I think a specialiazed hash function for `GlobalDeclID` is indeed the way to go.

I was also worried a little there are other performance implications of this change that would block us, but we won't know until we run a full release testing cycle, which may take a week or more. Hopefully it will be okay, as my observations around the increase in PCM sizes align with the numbers from this patch and performance seems to be on par as well, if we change the hash function.
However, I still wanted to mention it so that we won't come out completely of the blue.

I'll try to help reviewing the two patches you posted, thanks!

https://github.com/llvm/llvm-project/pull/92083