Page MenuHomePhabricator

[libc++][test] Add const and reference tests for enable_view. NFC.

Authored by jloser on Jan 21 2022, 12:08 PM.



As discussed in, there is missing test coverage
for the behavior of enable_view when given a const or reference qualified
type. Add such tests showing the current behavior.

Diff Detail

Event Timeline

jloser requested review of this revision.Jan 21 2022, 12:08 PM
jloser created this revision.
Herald added a project: Restricted Project. · View Herald TranscriptJan 21 2022, 12:08 PM
Herald added a reviewer: Restricted Project. · View Herald Transcript
philnik added inline comments.Jan 21 2022, 12:30 PM

This seems weird. It looks to me like the standard requires it that way, but I don't know why.


Thanks! This is an improvement as-is, but I think you should consider merging all your new tests from D117714 into this review. They should all work today, because view_interface<T> already inherits from view_base. Then your D117714 will become a code-only change, with no diffs to any of the tests except for the empty-base-sizeof test (and I guess if merging the tests over here discovers any bugs in the existing implementation!).


Yikes. Good catch. This reminds me of D116388, but unlike bind, this stuff is so new that I think this deserves an LWG issue. I'll request one.

jloser updated this revision to Diff 402077.Jan 21 2022, 12:48 PM
jloser retitled this revision from [libc++][test] Add const and reference tests for enable_view to [libc++][test] Add const and reference tests for enable_view. NFC..

Merge tests from LWG3549 PR into this review.

jloser marked 4 inline comments as done.Jan 21 2022, 12:48 PM
jloser added inline comments.

Exactly my thoughts as well. It's clear from a technical perspective, but surely this isn't what the user intended. They likely intended to disable cv-qual of their type as well and expected us to do it for them basically by opting in for their type only (and not aalso
for cv-qual of their type).

Generally speaking, this is likely to come up with other customization points in the standard library, but I'd have to search for other "broken" ones.

jloser marked an inline comment as done.Jan 21 2022, 12:49 PM
philnik accepted this revision as: philnik.Jan 21 2022, 1:00 PM
philnik accepted this revision.
This revision is now accepted and ready to land.Jan 21 2022, 1:00 PM

Initial reply from LWG is: "Who cares?"

The only person who should be querying enable_view<X> is concept view itself; no user-programmer should ever be querying it. And concept view will care about enable_view<X> only when movable<X>, which means

  • enable_view<T&> is never relevant
  • enable_view<const T> is relevant only when assignable_from<const V&, const V>; such a V must have begin() and end() and const V& operator=(const V&) const.

Here are the test cases from my attempted-LWG-issue-filing, for the record.

struct V1 : std::ranges::view_base {};
struct V2 : std::ranges::view_interface<V2> {};
struct V3 {}; template<> constexpr bool std::ranges::enable_view<V3> = true;
struct V4 : std::view_base {}; template<> constexpr bool std::ranges::enable_view<V4> = false;

static_assert(std::ranges::enable_view<V1> == true);
static_assert(std::ranges::enable_view<V2> == true);
static_assert(std::ranges::enable_view<V3> == true);
static_assert(std::ranges::enable_view<V4> == false);

static_assert(std::ranges::enable_view<const V1> == true);
LIBCPP_STATIC_ASSERT(std::ranges::enable_view<const V2> == true);
static_assert(std::ranges::enable_view<const V3> == false);
static_assert(std::ranges::enable_view<const V4> == true);