zyn0217 wrote:
Just get back to this again. So, I have explored the other case and found it
particular to explicit template arguments. In fact, we could slightly revise
the motivating case to
```cpp
template <class T>
concept C = false;
template <typename U, template <C> typename T>
int wow(T<U> ts);
template <typename T> struct S {};
int main() { return wow<int>(S<int>{}); }
```
with an *explicit* template argument `int`. This takes us to such the branch in
`DeduceTemplateArguments`,
https://github.com/llvm/llvm-project/blob/e0c554ad87d18dcbfcb9b6485d0da800ae1338d1/clang/lib/Sema/SemaTemplateDeduction.cpp#L4222-L4232
which would deduce the remaining template arguments from the given
`FunctionTemplate` and fall into our `getTemplateInstantiationArgs` at the end
of the day. It is also interesting to notice that the given
`TemplateTemplateParmDecl` originates from this `FunctionTemplate`, and the
surrounding `DeclContext`s for these `Decl`s has already been *redirected* to
its associated `FunctionDecl` at `AdoptTemplateParameterList`.
https://github.com/llvm/llvm-project/blob/59af659ee3c790d06cb5e2bf580e042547c24323/clang/lib/AST/DeclTemplate.cpp#L202-L206
Therefore, I think our assumption that the `TemplateTemplateParmDecl` always
resides in the `TranslationUnitDecl` is probably wrong. And, it is indeed murky
that the `DeclContext`s where these `Decl`s live depend on the callers to this
function...
Removing these checks works for the code above and, fortunately, doesn't fail
any tests.
@erichkeane WDYT?
https://github.com/llvm/llvm-project/pull/76811
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
