REQUIRE_THROWS_AS does not support class constructors

[jktjkt]

Description

I would like to use the REQUIRE_THROWS_AS macro with class constructors, not just freestanding functions. The following code results in a build failure for me (GCC 8.1.0, Doctest from the latest dev):

#define FAKE_READING(X) \
  // this is our internal macro which calls into Trompeloeil, it mocks
  // a virtual TTY as the variable named "serial"

DOCTEST_SUBCASE("nothing on the tty") {
    FAKE_READING("");
    REQUIRE_THROWS_AS(DriverUnderTest(serial), FramingError);
}

The error:

foo.cpp: In function ‘void _DOCTEST_ANON_FUNC_2()’:
foo.cpp:175:36: error: unnecessary parentheses in declaration of ‘serial’ [-Werror=parentheses]
         REQUIRE_THROWS_AS(DriverUnderTest(serial), FramingError);
                                          ^
foo.cpp:175:43: error: no matching function for call to ‘some::driver::DriverUnderTest::DriverUnderTest()’
         REQUIRE_THROWS_AS(DriverUnderTest(serial), FramingError);

I know that in this particular scneario I could workaround this via initializing via the uniform construction syntax with curly braces, i.e., DriverUnderTest{serial}, but that's not possible in general because Foo{x, y} and Foo(x, y) are different constructors (see std::vector::vector for example).

• doctest version: latest dev
• Operating System: Linux
• Compiler+version: GCC 8.1.0

[onqtam]

well... you could do what I suggested in this commit in a PR to json: nlohmann/json@5d511a6

basically: have a compatibility header which includes the actual doctest header and have the following in there:

#include <doctest/doctest.h>

#undef CHECK_THROWS
#undef CHECK_THROWS_AS
#undef CHECK_THROWS_WITH
#undef CHECK_NOTHROW

#undef REQUIRE_THROWS
#undef REQUIRE_THROWS_AS
#undef REQUIRE_THROWS_WITH
#undef REQUIRE_NOTHROW

// doctest allows multiple statements in these macros (even blocks of code) but json
// tests rely on passing single function/constructor calls which have a [[nodiscard]]
// attribute so here we static_cast to void - just like Catch does
#define CHECK_THROWS(expr) DOCTEST_CHECK_THROWS(static_cast<void>(expr))
#define CHECK_THROWS_AS(expr, e) DOCTEST_CHECK_THROWS_AS(static_cast<void>(expr), e)
#define CHECK_THROWS_WITH(expr, e) DOCTEST_CHECK_THROWS_WITH(static_cast<void>(expr), e)
#define CHECK_NOTHROW(expr) DOCTEST_CHECK_NOTHROW(static_cast<void>(expr))

#define REQUIRE_THROWS(expr) DOCTEST_REQUIRE_THROWS(static_cast<void>(expr))
#define REQUIRE_THROWS_AS(expr, e) DOCTEST_REQUIRE_THROWS_AS(static_cast<void>(expr), e)
#define REQUIRE_THROWS_WITH(expr, e) DOCTEST_REQUIRE_THROWS_WITH(static_cast<void>(expr), e)
#define REQUIRE_NOTHROW(expr) DOCTEST_REQUIRE_NOTHROW(static_cast<void>(expr))

I don't think REQUIRE_THROWS_AS() will be changed by default - perhaps a config option, or a separate macro, or just a link to this issue in the docs... not really sure what the right path is

[jktjkt]

Thanks for your quick reply. I have to admit that I do not fully understand how a [[nodiscard]] is relevant here. Neither my constructors nor the class itself are marked that way, and this isn't implicit for constructors as far as I can tell.

Looking further, this looks like an instance of the C++'s most vexing parse. And indeed, here's what clang says:

error: parentheses were disambiguated as redundant parentheses around declaration of variable named
      'serial' [-Werror,-Wvexing-parse]
        REQUIRE_THROWS_AS(FooModule(serial), OperationError);
                                   ^~~~~~~~

I wonder what your motivation for not casting is here. To me it is quite reasonable to assert that a particular construction fails...

[onqtam]

The idea is to support arbitrary code - like entire blocks or multiple statements:

REQUIRE_THROWS_AS(if (x) { foo(); bar(); }, ex_type);
REQUIRE_THROWS_AS(foo(); bar(), ex_type);

But when I think about it - users could just use a lambda - this was a concern when doctest supported C++98...

So if I change this it would be a breaking change........... And yes - nodiscard has nothing to do but the casting will do the trick, right?

[jktjkt]

Confirming that static_cast<void> is a workaround.

[onqtam]

6 days ago I added the DOCTEST_CONFIG_VOID_CAST_EXPRESSIONS config option - when you define it before the doctest header you would get the desired behavior. Currently only in the dev branch - will release an official version... soon (1-2 weeks probably).