`P0088R3_variant`: Investigate possible Clang bugs

[StephanTLavavej]

In #4713, I disabled these parts of P0088R3_variant, but never investigated:

STL/tests/std/tests/P0088R3_variant/test.cpp

Lines 1477 to 1489 in 26b7ca5

	#ifdef __clang__ // TRANSITION, not yet investigated
	#pragma clang diagnostic push
	#pragma clang diagnostic ignored "-Wundefined-inline"
	#endif // ^^^ workaround ^^^
	struct HasFullOrdering {
	constexpr bool operator==(const HasFullOrdering&) const;
	constexpr std::weak_ordering operator<=>(const HasFullOrdering&) const;
	};
	#ifdef __clang__ // TRANSITION, not yet investigated
	#pragma clang diagnostic pop
	#endif // ^^^ workaround ^^^

STL/tests/std/tests/P0088R3_variant/test.cpp

Lines 4781 to 4788 in 26b7ca5

	void test_no_narrowing_check_for_class_types() {
	#ifndef __clang__ // TRANSITION, not yet investigated
	using V = std::variant<int, BoomOnAnything>;
	V v(42);
	assert(v.index() == 0);
	assert(std::get<0>(v) == 42);
	#endif // ^^^ no workaround ^^^
	}

I commented that these appeared to be mirrored in libcxx/expected_results.txt, which are also unsolved mysteries:

STL/tests/libcxx/expected_results.txt

Lines 843 to 844 in 26b7ca5

	# Not analyzed. Clang instantiates BoomOnAnything during template argument substitution.
	std/utilities/variant/variant.variant/variant.ctor/T.pass.cpp:2 FAIL

STL/tests/libcxx/expected_results.txt

Lines 1058 to 1059 in 26b7ca5

	# Not analyzed. Clang emits a -Wundefined-inline warning.
	std/utilities/variant/variant.relops/three_way.pass.cpp:2 FAIL

[frederick-vs-ja]

For the HasFullOrdering case, I think Clang is emitting correct (but undesired) warnings. HasFullOrdering::operator<=> is actually inline but not defined, and it makes Clang warn in the noexcept specification of our operator<=> for variant. The constexpr also has some effects for Clang in instantiation.

We reuse compare_three_way, which has deduced return type and triggers more instantiations, for that operator<=>. Contrarily, libc++ uses an internal lambda with non-deduced trailing return type, which makes less things instantiated. I guess it's a possible improvement to instantiate a bit less things by not reusing compare_three_way.

Also, I think constexpr should be dropped from these undefined operators. The warning also disappears with constexpr dropped.

---

For the BoomOnAnything, the error disappears with constexpr dropped, which seems to be due to the side effect of constexpr function and brace-init-list, see also WG21-P0859R0. If there's a bug around constexpr, it might be tracked in LLVM-93364 - but Clang seems to be correct on this.

WG21-P3606R1 is possibly related. Clang seemingly performs overload resolution in copy/move construction in a more annoying (but possibly conforming) way.

We may also consider adding some workaround like llvm/llvm-project@194b337.

Edit: WG21-P3146 also seems related.

---

In all, it seems to me that

• Clang isn't really buggy for these cases, and
• the BoomOnAnything test case is possibly considered invalid, although it seems doable to accept that case.