Attribute noexcept_verify

Today on Slack, people were discussing the ever-popular noexcept(auto). The idea is that just as C++14 let us stop writing

template<class T, class U>
auto plus(T t, U u)
    -> decltype(t + u)
{
    return t + u;
}

and start writing simply

template<class T, class U>
auto plus(T t, U u)
{
    return t + u;
}

then wouldn’t it be nice if some future version of C++ let us stop writing

template<class T, class U>
auto plus(T t, U u)
    noexcept(noexcept(t + u))
{
    return t + u;
}

(remember, the number of the noexcepts shall be two) and start writing simply the following?

template<class T, class U>
auto plus(T t, U u) noexcept(auto)
{
    return t + u;
}

The upside

The nice thing about noexcept(auto) (which is not a real active proposal at the moment; it’s just something people have been talking about for about a decade now) is that it would save us from rogue calls to std::terminate. Consider the C++17 code:

template<class T>
auto frobnicate(T t)
    noexcept(std::is_nothrow_copy_constructible_v<T>)
{
    return t;
}

static_assert(noexcept(frobnicate(42)));
static_assert(not noexcept(frobnicate(std::string{})));

Looks sane, right? But consider calling it like this:

struct Evil {
    Evil() = default;
    Evil(const Evil&) = default;
    Evil(Evil&) { throw "ha"; }
};

int main()
{
    frobnicate(Evil{});
}

It just calls std::terminate… because we got our noexcept condition wrong! We meant to ask whether noexcept(T(t)). (And sure, that bugfix looks simple in hindsight. But remember, this was a simple contrived example. Getting it right in real code, and getting it right 100% of the time, is not only difficult but practically a fool’s game.)


UPDATE: David Stone points out that even the “expert” (me) got it wrong! noexcept(T(t)) tests the noexceptness of an explicit conversion, whereas what we actually did in return x; was an implicit conversion. Here’s an EvenMoreEvil that still calls std::terminate even after our supposed “fix.”


If we were able to write this instead:

template<class T>
auto frobnicate(T t) noexcept(auto)
{
    return t;
}

then frobnicate would be noexcept when it could be, and non-noexcept when it needed to be. Nice!

The downside

While I love the idea of noexcept(auto), I would oppose any serious attempt to push it into the C++ Standard until we’ve seen that it can actually be implemented and used — correctly. Consider this simple code:

template<class T>
auto frotz(T t) noexcept(auto)
{
    puts("testing");
    return t;
}

Under which circumstances is this function noexcept?

Well, it turns out that it’s never noexcept. Because the entire C library is nothrow, but none of it is noexcept! (And no, spelling it std::puts will not change anything in this department.)

So, adding a single puts or printf — or memcpy or strlen or abort — to your noexcept(auto) function can instantly render it “non-noexcept” from the compiler’s point of view. Which might cause library pessimizations (such as how vector<T> reallocation falls off a cliff if you omit noexcept from T’s move-constructor) or might even cause static_assert failures in far-removed client code.

And the problem is not limited to the C library: there are actually lots of APIs in the C++ standard library that you might naïvely expect to be noexcept, but they’re not. For example:

template<class T>
struct Holder {
    T m_data;
    T exchange(T value) const noexcept(auto) {
        return std::exchange(m_data, value);
    }
};
Holder<int> h;
static_assert(not noexcept(h.exchange(42)));  // yikes!

I don’t think we’ll ever get noexcept(auto) in the C++ Standard unless somebody does the hard work to show concrete evidence (not just papers, but actual real-world usage) that this problem is surmountable or mitigable somehow.

The bright idea

The problem with noexcept(auto), outlined above, boils down to: noexcept(auto) instructs the compiler to take the wheel. For this to work, the compiler has to be a dang good driver — it has to guess correctly for each function whether we intended it to be noexcept or not. As we saw in the first example, humans are terrible at this game — but as we saw in the second example, compilers seem also to be terrible at it. We can’t hand the wheel to the compiler unless we know the compiler is going to drive right.

So, trust, but verify? Suppose we had a vendor-specific attribute — let’s for the sake of argument call it [[clang::noexcept_verify]] — which would instruct the compiler to compute the same exact thing as noexcept(auto), but then, instead of applying it blindly to our function’s signature, the compiler would merely look for an existing (possibly defaulted) noexcept specification on our function and verify that the computed noexcept-ness matched the noexcept-ness expressed by our existing specification!

The compiled program would still use the noexcept-ness expressed by the programmer. The computed noexcept-ness would be used only for diagnostics; and the diagnostics would not necessarily have to be fatal. (They could be dialed down to warnings.)

This would turn our “rogue std::terminate” example above into a compile-time failure.

template<class T>
[[clang::noexcept_verify]]
auto frobnicate(T t)
    noexcept(std::is_nothrow_copy_constructible_v<T>)
{
    return t;
}

int main()
{
    frobnicate(Evil{});
}

warning: computed noexcept-specification 'noexcept(false)' does not match
explicit noexcept-specification 'noexcept(true)' [-Wnoexcept-verify]
    frobnicate(Evil{});
    ^
note: previous noexcept-specification is here
    noexcept(std::is_nothrow_copy_constructible_v<T>)
    ^

Such an attribute might see significant adoption — or at least experimentation. It would still suffer from the same problem with C library functions; but now the symptom of that problem would be bogus diagnostics at compile-time, instead of quietly turning functions noexcept(false) when we didn’t intend it. This would remove the risk of asking for the compiler’s help; and so it should lead to quick adoption, especially by template libraries.

Finally, another benefit of the “trust but verify” approach is that it would quickly yield empirical data on how often human programmers actually do get their noexcept conditions right. For non-constant conditions (i.e. excluding cases of literal noexcept(true) and noexcept(false)), I bet the answer is “less than 20% of the time.”

I have not implemented this attribute. Frankly, I wouldn’t know where to start with it. But if you like the idea and want to give it a shot… please feel free!

Posted 2018-06-12