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

[Chuanqi Xu via cfe-commits]

ChuanqiXu9 wrote:

> > > As far as I can tell from [#76774 (comment)](https://github.com/llvm/llvm-project/pull/76774#issuecomment-1914177330) above, the last push only changed the default value of `LoadExternalSpecializationsLazily`. In that case, my test results from [#76774 (comment)](https://github.com/llvm/llvm-project/pull/76774#issuecomment-1912182171) remain fully valid, especially
> > > > Switching the new `LoadExternalSpecializationsLazily` to disabled by default (somehow the argument didn't work on its own) instead crashes, most of the cases involving `MultiOnDiskHashTable`. I suspect some kind of memory error maybe?
> > 
> > 
> > No, this newest patch changes besides changing the default value. For example, [22c9d11#diff-125f472e690aa3d973bc42aa3c5d580226c5c47661551aca2889f960681aa64dR232-R237](https://github.com/llvm/llvm-project/commit/22c9d1145eb57d9c2cb2ef490b7c474598dd5d12#diff-125f472e690aa3d973bc42aa3c5d580226c5c47661551aca2889f960681aa64dR232-R237) now we won't write specializations into the hash table, while the original patch tries to load all the decls from the consumers' side only. This patch touches the producer's side.
> > Maybe I shouldn't push force to make the history clear..
> 
> If you want I can set you up with the ROOT project so that you can debug the failures easier.

Thanks. Although I don't understand how can it be, it will be better if things can be easier.

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