reporting bad coverage for code in unevaluated context

[hanickadot]

Dear LLVM developers! I noticed bad code-coverage for code in unevaluated context. Like when constexpr function is called in decltype(...) or even in concepts:

Such code shouldn't be marked uncovered.

[hanickadot]

Sorry for not proper bug report. I should be better. There is all you need:

template <typename T, template <typename...> typename Wrapper> 
concept is_base_of = requires(const T & obj) {
  []<typename... Y>(const Wrapper<Y...> &){ }(obj);
};

template <typename...> struct variant { };

struct foo: variant<int, bool> { };

static_assert(is_base_of<foo, variant>);

int main() { }

These commands needs to be run:

clang++ -fprofile-instr-generate -fcoverage-mapping test.cpp -std=c++20 -o test

./test

llvm-profdata merge default.profraw -o coverage.profdata

llvm-cov show test -instr-profile=coverage.profdata

And the resulting coverage is:

(notice the uncovered branch in lambda inside concepts)

[hanickadot]

And I'm using clang 17.0.6. Don't know about other versions as this is not possible (AFAIK) on CE.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Hana Dusíková (hanickadot)

Dear LLVM developers! I noticed bad code-coverage for code in unevaluated context. Like when constexpr function is called in `decltype(...)` or even in concepts:

<img width="1081" alt="Screenshot 2023-12-17 at 22 16 51" src="https://github.com/llvm/llvm-project/assets/6557263/10b6884a-b7fd-4e52-9973-82e5e51efbb0">

Such code shouldn't be marked uncovered.

[hanickadot]

Fixed in #78033