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

[Chuanqi Xu via cfe-commits]

ChuanqiXu9 wrote:

> > With both PRs I'm getting at least this Clang assertion error on some of our code. There's more, and I'll try to create a reproducer for this one, but maybe the stack trace is enough to figure out what's wrong:
> > ```
> > assert.h assertion failed at clang/include/clang/AST/Redeclarable.h:262 in redecl_iterator &clang::Redeclarable<clang::TagDecl>::redecl_iterator::operator++() [de
> > cl_type = clang::TagDecl]: 0 && "Passed first decl twice, invalid redecl chain!"
> >     @     0x561251408044  __assert_fail
> >     @     0x56124c214b91  clang::ASTDeclMerger::MergeDefinitionData()
> >     @     0x56124c21722e  clang::ASTDeclReader::VisitClassTemplateSpecializationDeclImpl()
> >     @     0x56124c207953  clang::declvisitor::Base<>::Visit()
> >     @     0x56124c207135  clang::ASTDeclReader::Visit()
> >     @     0x56124d69ddcf  clang::StackExhaustionHandler::runWithSufficientStackSpace()
> > ```
> 
> Thanks for testing! "Luckily" it's the exact code that I touched in `MergeDefinitionData` in #170090. I guess it comes down to @ChuanqiXu9's comment in [#170090 (comment)](https://github.com/llvm/llvm-project/pull/170090#discussion_r2576320393) that `redecls()` can trigger deserialization and that may invalidate the iterator (?). Unfortunately, I don't have an easy idea to fix it - going back to the old code and just calling `getMostRecentDecl` at the beginning doesn't pass the `Modules/GH170084.cpp` test case...

I feel it may not be too hard to implement a `noload_redecls()`. You can get the most recent decl without calling `getMostRecentDecl ()` (that will trigger the deserialization of all redecls, I hate it) in Redeclarable.h internally. And you can simply traverse the redecls.

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