[clang] [Serialization] Fix lazy template loading (PR #133057)

[Vassil Vassilev via cfe-commits]

vgvassilev wrote:

> > I'd feel uncomfortable exposing a `noload_redecls` interface because it can potentially break the AST invariants. I think allowing non-modules developers to choose when to complete a redeclaration chain will introduce more bugs than it can fix ;)
> > I'd propose to move forward with `getMostRecentExistingDecl` if it does the job. We can have a separate PR discussing the ergonomics if there is a strong feeling about going this way.
> 
> I'd like it as: (1) We already have `noload_lookup` and other similar "unpublic" API (e.g., `getOwningModuleID` `invalidateCachedLinkage`) in DeclBase.h.

  Well, sure, `noload_lookup` is done to check what's the current state before completing something externally. In fact, I think we should keep these interfaces at minimum and only use them where we really have no other way out.

> (2) Merging Redecls is a major performance hurting point of using modules. 

  Only when the modules approach is not bottom up. The non-bottom up approach has bigger problems than performance at the moment and frankly we cannot reasonably support it right now...

> (3) In the past, there are many "fix" which just tried to mitigate the bug by loading all redecls without understanding the real problem.

  Does that need to happen in this PR? I believe it brings down the peak memory significantly for most of the module users and we should not wait too long to get this merged.


https://github.com/llvm/llvm-project/pull/133057