On 30/10/20 11:11 -0400, Patrick Palka via Libstdc++ wrote:
This works around a subtle issue where instantiating the begin()/end()
member of some views (as part of return type deduction) inadvertently
requires computing the satisfaction value of range<foo_view>.
This is problematic because the constraint range<foo_view> requires the
begin()/end() member to be callable. But it's not callable until we've
deduced its return type, so evaluation of range<foo_view> yields false
at this point. And if at any point after both members are instantiated
(and their return types deduced) we evaluate range<foo_view> again, this
time it will yield true since the begin()/end() members are now both
callable. This makes the program ill-formed according to
[temp.constr.atomic]/3:
If, at different points in the program, the satisfaction result is
different for identical atomic constraints and template arguments, the
program is ill-formed, no diagnostic required.
The views affected by this issue are those whose begin()/end() member
has a placeholder return type and that member initializes an _Iterator
or _Sentinel object from a reference to *this. The second condition is
relevant because it means explicit conversion functions are considered
during overload resolution (as per [over.match.copy], I think), and
therefore it causes g++ to check the constraints of the conversion
function view_interface<foo_view>::operator bool(). And this conversion
function's constraints indirectly require range<foo_view>.
This issue is observable on trunk only with basic_istream_view (as in
the testcase in the PR). But a pending patch that makes g++ memoize
constraint satisaction values indefinitely (it currently invalidates
the satisfaction cache on various events) causes many existing tests for
the other affected views to fail, because range<foo_view> then remains
false for the whole compilation.
This patch works around this issue by adjusting the constructors of the
_Iterator and _Sentinel types of the affected views to take their
foo_view argument by pointer instead of by reference, so that g++ no
longer considers explicit conversion functions when resolving the
direct-initialization inside these views' begin()/end() members.
Nice solution.
Tested on x86_64-pc-linux-gnu, and also verified that this fixes the
testsuite failures when combined with the mentioned frontend patch
(https://gcc.gnu.org/pipermail/gcc-patches/2020-October/557237.html).
Does this look OK for trunk?
Yes, thanks.
