Test for is_buffer_sequence: false positives

[karzhenkov]

Simple test cases indicate that the is_buffer_sequence template is implemented incorrectly:

• struct A1 with single member is recognized as a MutableBufferSequence when compile with the decltype available;
• struct B1 with inappropriate typedefs and no member functions is recognized as a MutableBufferSequence when the decltype is unavailable or disabled.

[raJeev-M]

Why this change?

[raJeev-M]

Now the fallback becomes char[2] not char, so should it not be !=2

[karzhenkov]

When called with single argument this is resolved to the fallback, since no default argument value in checker function declaration is specified. The fallback was char[2], the condition was true regardless of the type being tested.

The modified fallback for ASIO_HAS_DECLTYPE case is char, so != 1 is correct.

When ASIO_HAS_DECLTYPE is not defined, fallback is still char[2], but in this case checker function must be rejected for acceptable type; != 1 is correct again.

[raJeev-M]

why this change?

[karzhenkov]

Fallback is now char when ASIO_HAS_DECLTYPE is defined (see above).
For checker function we need char[2]

[raJeev-M]

struct A1 with single member is recognized as a MutableBufferSequence when compile with the decltype available;

I am not sure why do you say it, Have looked at your test-code, here is my observation and correct me if am wrong

so you have the below stuff

struct A1
{
  mutable_buffer* begin();

  // no "value_type" type
  // no "const_iterator" type
  // no "end" member function
};

ASIO_CHECK(!is_mutable_buffer_sequence<A1>::value);

I believe is_mutable_buffer_sequence<A1> returns true.. and it bombs you out because there is no
end(), yet, still the buffer_sequence_end_helper is not falling-back?

Here is my experiment on asio::buffer_sequence_end_helper, looks like, things are working fine as expected: https://godbolt.org/z/Ch_jCq . Do you think I am miss anything?

[karzhenkov]

When I add mutable_buffer* end(); to struct A1 in your sample, it still compile successfully. The second overload of helper (i. e. the checker) will always be rejected independently of declarations in struct A1. Condition will always be true.

To change this we can provide default value for its second argument (or add second argument when it is called from static_assert).

After such a change we need also invert condition, and write != 2, but this will break the logic for the case when ASIO_HAS_DECLTYPE is not defined.

So instead of != 2 we have to "swap" return types of the fallback and the checker.

[karzhenkov]

When ASIO_HAS_DECLTYPE	for acceptable type the checker must be	fallback return type before change	fallback return type after change
is defined	accepted (but was rejected for any type)	char[2]	char
is not defined	rejected	char[2]	char[2]

[karzhenkov]

ASIO_CHECK(!is_mutable_buffer_sequence<A1>::value);

I believe is_mutable_buffer_sequence<A1> returns true.. and it bombs you out because there is no
end(), ...

I would interpret this situation a bit differently.

The test (055c51f) is written with assumption that A1 should not be considered as MutableBufferSequence because it has no end(). The test failed, i. e. the program wrongly accepted A1 as a valid MutableBufferSequence. This indicates an error in the is_buffer_sequence implemenation.

Or perhaps the assumption used in the test is wrong?

[karzhenkov]

Obsoleted by #403