Class template specialisation doesn't always require constraints to match

[cjdb]

GCC test case concepts-ttp.C fails on Clang.

// PR c++/99904
// RUN: %clang_cc1 -std=c++20 -fsyntax-only -verify %s

template<class... Ts> concept C = (Ts::value && ...);
template<class... Ts> requires C<Ts...> struct A;
template<class T> requires true struct B;
template<template<class... Ts> requires C<Ts...> class TT> struct S;
using ty1 = S<A>;
using ty2 = S<B>; // expected-error{{}}

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[cor3ntin]

@erichkeane I'm still confused by template template parameter constraints. Should they match?

[erichkeane]

Its not clear to me what the answer should be here. I THINK all we end up doing is checking 'C' at each case, which is true? But MSVC seems to agree with us: https://godbolt.org/z/Tve4Gd61c

[cor3ntin]

@mizvekov

[mizvekov]

Yeah this is one case I pointed out in a recent core reflector thread about divergence in the handling of constraints on the exact match fallback.

Technically clang is correct in ignoring the constraints on this fallback, but to me this is clearly undesirable.