https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121080
Bug ID: 121080
Summary: GCC seems to bypass the alias template CTAD deduction
guides if the template-id is simple enough
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: zyn7109 at gmail dot com
Target Milestone: ---
(The example comes from Clang's bug reports, for which GCC seems to behave
different with the other 3 compilers https://godbolt.org/z/64Me6T4dn and hence
the bug report)
Consider the following example
```
template <class T>
struct traits { using type = T; };
template <class Key1, class U1 = Key1>
class unordered_set {
public:
template <class T>
unordered_set(T) {}
};
template <class Undeduced, class U = Undeduced>
unordered_set(Undeduced, U = U())
-> unordered_set<typename traits<Undeduced>::type, U>; // #2
template <class Key2, class U2 = Key2>
using my_unordered_set =
unordered_set<Key2, U2>; // #1
my_unordered_set s(0); // #3
```
To deduce the type of s, one should synthesize a CTAD guide from #1; and one of
the deduction guide that is synthesized against #2 is formed in ways like
1. We deduce #2's <Undeduced, U> against <typename traits<Undeduced>::type, U>
using #1's <Key2, U2>
2. U can be deduced as U2, Undeduced cannot be deduced because of non-deduced
context, and since U2 has a default argument, the synthesized deduction guide
ends up like
template <class Key2, class U2 = Key2, class Undeduced>
// is_deducible constriant is omitted
unordered_set(Undeduced, U2 = U2()) -> unordered_set<typename
traits<Undeduced>::type, U2>;
>From which we could see that Key2 cannot be deduced, so at least this deduction
guide shouldn't help us compile #3.
However, GCC seems to compile #3 without going through such transform, as can
be seen at https://godbolt.org/z/esP7K4fP1
If we somehow make #1 a bit complicated, e.g. if we write
template <class Key2, class U2 = Key2>
using my_unordered_set =
unordered_set<typename traits<Key2>::type, U2>;
then GCC rejects such as well, in the way I wrote above.
Is the behavior conforming?
