<ranges>: Questionable conditional noexcept

[CaseyCarter]

The range adaptors defined in <ranges> strengthen noexcept more aggressively than much of the STL. Upon noticing that a couple of conditional noexcepts in transform_view were incorrect for non-zero IDL modes, I threw together some preprocessor stuffs to "turn off" a conditional noexcept in non-zero IDL:

STL/stl/inc/ranges

Lines 1016 to 1020 in be0441c

	#if _ITERATOR_DEBUG_LEVEL == 0
	#define _NOEXCEPT_IDL0(...) noexcept(__VA_ARGS__)
	#else
	#define _NOEXCEPT_IDL0(...)
	#endif // _ITERATOR_DEBUG_LEVEL == 0

So that precondition-checking machinery that calls user-provided functions:

STL/stl/inc/ranges

Lines 1064 to 1067 in be0441c

	constexpr void _Check_dereference() const {
	_STL_VERIFY(_Parent != nullptr, "cannot dereference value-initialized transform_view iterator");
	_STL_VERIFY(_Current != _RANGES end(_Parent->_Range), "cannot dereference end transform_view iterator");
	}

is usable from functions that want to provide conditional noexcept:

STL/stl/inc/ranges

Lines 1112 to 1120 in be0441c

	_NODISCARD constexpr decltype(auto) operator*() const
	_NOEXCEPT_IDL0(noexcept(_STD invoke(*_Parent->_Fun, *_Current))) /* strengthened */ {
	#if _ITERATOR_DEBUG_LEVEL != 0
	_Check_dereference();
	_STL_VERIFY(
	_Parent->_Fun, "Cannot dereference iterator into transform_view with no transformation function");
	#endif // _ITERATOR_DEBUG_LEVEL != 0
	return _STD invoke(*_Parent->_Fun, *_Current);
	}

This is a bit of a hack, and it was received as such. I've since observed that other range adaptors have the same problem and simply slam into noexcept:

STL/stl/inc/ranges

Lines 876 to 882 in be0441c

	_NODISCARD friend constexpr range_rvalue_reference_t<_Vw> iter_move(const _Iterator& _It) noexcept(
	noexcept(_RANGES iter_move(_It._Current))) {
	#if _ITERATOR_DEBUG_LEVEL != 0
	_It._Check_dereference();
	#endif // _ITERATOR_DEBUG_LEVEL != 0
	return _RANGES iter_move(_It._Current);
	}

(Note that conditional noexcept on iter_move and iter_swap overloads is not strengthened, but Standard-mandated.)

We need to devise policy for how we want to handle conditional noexcept on functions that call user-provided in "debug" modes only, including both strengthened and non-strengthened cases.

[AlexGuteniev]

I don't think that there's much point in this conditional noexcept.

I mean simple inline functions that call noexcept functions.

Compiler can propagate the observation of non-throwing into inlined code. So that even if it would not be noexcept from the type system point of view, it would be noexcept for optimization.

Mostly I see the point for noexcept in exports, where compiler does not see the implementation. But those noexcept are never conditional.

[AdamBucior]

Another solution would be to silently catch the exceptions that go out of debug machinery. I still think that terminating is the best option but silent catching is not a bad idea.

@AlexGuteniev noexcept is not only a hint to the compiler, it's also a hint to the user. Conditional noexcept tells the user, under what conditions the function does not throw exceptions, so he does not have to check definitions of all functions the function calls and see if they throw any exceptions.

[CaseyCarter]

I've come to the conclusion over the last couple of years that debug checks should affect neither noexcept nor ABI for ranges.

We discussed at the weekly maintainer meeting, and have consensus for that principle. Casey will submit a PR to remove the conditional noexcept that current exists in <ranges>.