Compile error in include/boost/stl_interfaces/iterator_interface.hpp when compiling with /std:c++latest

[NEIL-smtg]

Hi, I work on Microsoft Visual C++ testing, where we regularly build popular open-source projects, including Boost, with development builds of our compiler and libraries with some particular cl options such as /permissive- and /std:c++latest to detect and prevent shipping regressions that would affect you.

This issue is only occurring when building with /std:c++latest, if there are no plans to support /std:c++latest it in the future, feel free to close this issue.

Recently, as I build Boost from source and compiling with /std:c++latest, I encounter the following errors in https://github.com/boostorg/stl_interfaces/blob/develop/include/boost/stl_interfaces/iterator_interface.hpp:

.\boost/stl_interfaces/iterator_interface.hpp(657): error C2139: 'iterator<node>': an undefined class is not allowed as an argument to compiler intrinsic type trait '__is_base_of'
libs\stl_interfaces\test\compile_sfinae_path_mutable_iterator.cpp(19): note: see declaration of 'iterator<node>'
.\boost/stl_interfaces/iterator_interface.hpp(657): note: the template instantiation context (the oldest one first) is
libs\stl_interfaces\test\compile_sfinae_path_mutable_iterator.cpp(73): note: see reference to class template instantiation 'iterator<node>' being compiled
libs\stl_interfaces\test\compile_sfinae_path_mutable_iterator.cpp(57): note: see reference to class template instantiation 'boost::stl_interfaces::v3::iterator_interface<iterator<node>,std::forward_iterator_tag,Node,ValueType *,ptrdiff_t>' being compiled
        with
        [
            Node=node,
            ValueType=std::forward_iterator_tag
        ]
.\boost/stl_interfaces/iterator_interface.hpp(682): note: see reference to function template instantiation 'auto boost::stl_interfaces::v2::v2_dtl::category_tag<IteratorConcept,ReferenceType>(void)' being compiled
        with
        [
            IteratorConcept=iterator<node>,
            ReferenceType=node
        ]
.\boost/stl_interfaces/iterator_interface.hpp(657): note: see reference to variable template 'const bool is_base_of_v<std::forward_iterator_tag,iterator<node> >' being compiled
.\boost/stl_interfaces/iterator_interface.hpp(940): error C2976: 'boost::stl_interfaces::v2::v2_dtl::iterator_category_base': too few template arguments
.\boost/stl_interfaces/iterator_interface.hpp(683): note: see declaration of 'boost::stl_interfaces::v2::v2_dtl::iterator_category_base'
libs\stl_interfaces\test\compile_sfinae_path_mutable_iterator.cpp(57): error C3210: 'iterator_interface<iterator<node>,std::forward_iterator_tag,node,std::forward_iterator_tag *,__int64>': a member using-declaration can only be applied to a base class member

Repro Steps:

1. Open x64 native tools command prompt for VS 2022
2. git clone https://github.com/boostorg/boost.git
3. set _CL_= /std:c++latest /DBOOST_TIMER_ENABLE_DEPRECATED /Zc:equalityRewrite- /Zc:enumTypes %CL%
4. cd /d C:\gitP\boostorg\boost
5. .\bootstrap 2>&1
6. .\b2 headers variant=release --build-dir=C:\gitP\boostorg\boost\amd64 address-model=64 architecture=x86 2>&1
7. .\b2 variant=release --build-dir=C:\gitP\boostorg\boost\amd64 address-model=64 architecture=x86 2>&1
8. set _CL_=%_CL_% /Zc:__cplusplus
9. .\b2 -j16 variant=release --build-dir=C:\gitP\boostorg\boost\amd64 libs\stl_interfaces\test address-model=64 2>&1

Build log: Build.log
stl_interfaces Test log: Test.log.94.txt

[tzlaine]

Thanks very much for reporting this! I'll try to produce a fix.

[tzlaine]

@NEIL-smtg So, I had a closer look at this. It seems like it may be a compiler regression. I say this because this code:

stl_interfaces/include/boost/stl_interfaces/iterator_interface.hpp

Lines 924 to 945 in cefbd30

	#if defined(BOOST_STL_INTERFACES_DOXYGEN) || BOOST_STL_INTERFACES_USE_DEDUCED_THIS
	namespace boost { namespace stl_interfaces { BOOST_STL_INTERFACES_NAMESPACE_V3 {
	// clang-format off
	/** A base template that one may derive from to make defining iterators
	easier. */
	template<
	typename IteratorConcept,
	typename ValueType,
	typename Reference = ValueType &,
	typename Pointer = ValueType *,
	typename DifferenceType = std::ptrdiff_t>
	struct iterator_interface
	: v2::v2_dtl::iterator_category_base<IteratorConcept, Reference>
	{
	using iterator_concept = IteratorConcept;
	using value_type = std::remove_const_t<ValueType>;
	using reference = Reference;
	using pointer = detail::pointer_t<Pointer, iterator_concept>;

... is guarded by BOOST_STL_INTERFACES_USE_DEDUCED_THIS, as you can see at the top (BOOST_STL_INTERFACES_DOXYGEN is only defined in doc builds). Also note that the first template parameter is IteratorConcept, not the derived type D. Earlier versions of the template (in namespaces v1:: and v2::) use CRTP, and so the derived type D appears as the first template parameter to those templates.

Here is the code that exercises the failure:

stl_interfaces/test/compile_sfinae_path_mutable_iterator.cpp

Lines 18 to 26 in cefbd30

	template<typename Node>
	struct iterator
	: boost::stl_interfaces::iterator_interface<
	#if !BOOST_STL_INTERFACES_USE_DEDUCED_THIS
	iterator<Node>,
	#endif
	std::forward_iterator_tag, Node>
	{
	using value_type = typename Node::value_type;

That code is written to conditionally provide the derived type D as the first template parameter, based on the value of BOOST_STL_INTERFACES_USE_DEDUCED_THIS -- if it's 0, we provide D first, and if it's 1, we provide the iterator category first. This allows this test to work with the v1/v2 and v3 template parameter expectations I explained above.

The latest vN namespace that your compiler supports becomes inline -- that's why the .cpp code is using the version of iterator_interface from v3::.

It looks like in the same TU, we have different definitions of BOOST_STL_INTERFACES_USE_DEDUCED_THIS (1 in the iterator_interface.hpp code, and 0 later, in the .cpp file), unless I'm missing something. Note that this build passes with recent releases of VS2022, which is why it seems to me to be a VS2022 dev-build regression.

Let me know what you think of this analysis -- I may well be confused. :)