[Clang] Wrong constraint evaluation for lambda in unevaluated context

[MagentaTreehouse]

The following should be valid C++20 code:

template <class T>
using F = decltype([] (auto a) requires (sizeof(T) > sizeof a) {});

int main() {
    F<int> f;
    f(char{});
}

Clang rejects with:

<source>:6:5: error: no matching function for call to object of type 'F<int>' (aka '(lambda at <source>:2:20)')
    6 |     f(char{});
      |     ^
<source>:2:20: note: candidate template ignored: constraints not satisfied [with a:auto = char]
    2 | using F = decltype([] (auto a) requires (sizeof(T) > sizeof a) {});
      |                    ^
<source>:2:42: note: because 'sizeof(char) > sizeof a' (1 > 1) evaluated to false
    2 | using F = decltype([] (auto a) requires (sizeof(T) > sizeof a) {});
      |                                          ^

Somehow it got char substituted into T in sizeof(T)?

See https://compiler-explorer.com/z/YvnzozWvx.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (MagentaTreehouse)

The following should be valid C++20 code: ```c++ template <class T> using F = decltype([] (auto a) requires (sizeof(T) > sizeof a) {});

int main() {
F<int> f;
f(char{});
}

Clang rejects with:
```console
<source>:6:5: error: no matching function for call to object of type 'F<int>' (aka '(lambda at <source>:2:20)')
    6 |     f(char{});
      |     ^
<source>:2:20: note: candidate template ignored: constraints not satisfied [with a:auto = char]
    2 | using F = decltype([] (auto a) requires (sizeof(T) > sizeof a) {});
      |                    ^
<source>:2:42: note: because 'sizeof(char) > sizeof a' (1 > 1) evaluated to false
    2 | using F = decltype([] (auto a) requires (sizeof(T) > sizeof a) {});
      |                                          ^

Somehow it got char substituted into T in sizeof(T)?

See https://compiler-explorer.com/z/YvnzozWvx.

[zyn0217]

@mizvekov

[shafik]

This looks like a clang-16 regression: https://compiler-explorer.com/z/bsq9cT134

@zyn0217 do you think you know which PR regressed this?

@MagentaTreehouse in general it is super helpful to see how far back an issue goes. This is super helpful especially if it is a recent regression.