================
@@ -1061,15 +1061,36 @@ BuildDeductionGuideForTypeAlias(Sema &SemaRef,
SmallVector<DeducedTemplateArgument> DeduceResults(
F->getTemplateParameters()->size());
+ // We don't have to deduce against the alias template specialization,
+ // if the source template is a synthesized alias deduction guide. This allows
+ // us to utilize the default template arguments from alias declaration.
+ //
+ // template <class T>
+ // using Foo = A<A<T>>;
+ //
+ // template <class U = int>
+ // using Bar = Foo<U>;
+ //
+ // In terms of Bar, we want U to appear in the synthesized deduction guide,
+ // but U would remain undeduced if we deduce against A<T> instead of T.
+ // Also note that since the deduced results are only used for synthesizing
+ // template parameters, they should not introduce unintended behavior in
+ // theory.
+ ArrayRef<TemplateArgument> Ps = FReturnType->template_arguments();
+ if (auto *DG = dyn_cast<CXXDeductionGuideDecl>(F->getTemplatedDecl());
+ DG && DG->getSourceDeductionGuideKind() ==
+ CXXDeductionGuideDecl::SourceDeductionGuideKind::Alias)
+ Ps = F->getInjectedTemplateArgs(Context);
----------------
zyn0217 wrote:
> Also, wouldn't we always be wanting to deduce against the injected template
> parameters of F anyway, so you could instead just change
> FReturnType->template_arguments() below to
> F->getInjectedTemplateArgs(Context)?
Ugh actually I ran into an assertion failure even without your suggestion
```cpp
template<typename T, typename...Us>
struct A{
template<typename V> requires __is_same(V, int)
A(V);
};
template<typename...TS>
using AA = A<int, TS...>;
template<typename...TS>
using BB = AA<TS...>;
BB a{0};
```
The (synthesized) template parameters are not guaranteed to be syntactially
correct, so we ended up having
```cpp
template <typename... Us, typename V>
AA(V) -> A<int, Us...>;
```
for AA, and using `<Us..., V>` instead of `<int, Us...>` would violate the
assumption in `hasTemplateArgumentForDeduction`: `Pack not at the end of
argument list?`
https://github.com/llvm/llvm-project/pull/147675
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
