Check range_begin is dereferenceable #3964
Conversation
Arghnews
force-pushed
the
improve_range_check
branch
from
f8d0c51
to
f860fc9
Compare
test/ranges-test.cc
Outdated
| FMT_BEGIN_NAMESPACE | ||
| template <> struct formatter<not_range> : formatter<string_view> { | ||
| auto format(const not_range&, format_context& ctx) const | ||
| -> format_context::iterator { | ||
| return formatter<string_view>::format("Dummy output", ctx); | ||
| } | ||
| }; | ||
| FMT_END_NAMESPACE | ||
| TEST(ranges_test, format_not_range_using_specialized_formatter) { | ||
| EXPECT_EQ("Dummy output", fmt::format("{}", not_range{})); | ||
| } |
There was a problem hiding this comment.
I think it's sufficient to check that is_formattable<not_range> compiles and returns false. We don't need to introduce a formatter here.
There was a problem hiding this comment.
If you're satisfied with that, then I'll remove the test and this can just have the static assert then. AFAIK there isn't a gtest way to static assert so will just use a regular one
There was a problem hiding this comment.
You can keep the test and make a dynamic check (e.g. EXPECT_FALSE) instead of static_assert. Then at least some of the failures might be recoverable. But it's not important.
There was a problem hiding this comment.
Had pushed before your comment, apologies, but seems like failing as early as possible ie. compile time rather than runtime is fine anyway, so will leave it without the test if you're ok with that, can change it back if not
Fixes issue fmtlib#3839 An Eigen 3.4 2x2 matrix has a begin member function that returns void Be more strict checking that the result of calling *begin() is valid See input_or_output_iterator concept notes about void
Arghnews
force-pushed
the
improve_range_check
branch
from
f860fc9
to
32d7c43
Compare
Fixes issue #3839
An Eigen 3.4 2x2 matrix has a begin member function that returns void and a static_assert that evaluates to false (as you're not meant to call begin on a non-1d matrix).
However this doesn't play nicely with the SFINAE checks at the moment in fmt as they simply perform a SFINAE check that the expression
T{}.begin()is valid, which evaluating tovoidis (even though it's of course not a valid type for an iterator):fmt/include/fmt/ranges.h
Lines 100 to 104 in 75e8924
To model a range more strictly to eliminate this case:
If we look at the c++20 concept for a range, it requires
ranges::beginon the type, and that requires the iterator type satisfy input_or_output_iterator, which specifically notes thatbeginreturns:The exposition-only concept /*can-reference*/ is satisfied if and only if the type is referenceable (in particular, not void).So just check we can deref the result of that type to fix this case, and not incorrectly identify a type with
void begin/endfunctions as a range. Note we don't want to do this for the type returned byend, as (in c++ >= 20) it may be a different sentinel type that is not dereferenceable