[Clang] `AttributedType`s are stripped from lambda function types

[Sirraide]

Consider

struct S {
  void g() const [[clang::annotate_type("foo", "bar", 1)]] {}
};

void f() {
  auto l = [] () [[clang::annotate_type("foo", "bar", 1)]] {};
}

AST dump prints 'void () const [[clang::annotate_type(...)]]':'void () const' for the type of S::g(), but void () const for the type of l, which is problematic if someone actually wants to use this information for something (#84983). We should fix this unless there’s a good reason why we’re desugaring the function type here.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (Sirraide)

Consider ```c++ struct S { void g() const [[clang::annotate_type("foo", "bar", 1)]] {} };

void f() {
auto l = [] () [[clang::annotate_type("foo", "bar", 1)]] {};
}

AST dump prints `'void () const [[clang::annotate_type(...)]]':'void () const'` for the type of `S::g()`, but `void () const` for the type of `l`, which is problematic if someone actually wants to use this information for something (#<!-- -->84983). We should fix this unless there’s a good reason why we’re desugaring the function type here.
</details>

[shafik]

CC @erichkeane @AaronBallman

[Sirraide]

So it seems this is happening because adjustDeducedFunctionResultType() doesn’t care about sugar—it just creates a new FunctionProtoType; I’ll try and see if changing that helps (if we even get to that point because we’re currently crashing before we ever get there because of #85154, which I’m currently looking into).