https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80667
--- Comment #1 from Ed Catmur <ed at catmur dot co.uk> ---
note: the rationale for using std::integral_constant rather than a T non-type
argument is CWG 1315.
Clang rejects in -std=c++1z:
<source>:22:63: error: ambiguous partial specializations of 'Impl<unsigned
char, std::integral_constant<unsigned char, '\x00'> >'
Impl<unsigned char, std::integral_constant<unsigned char, 0>> foo()
^
<source>:13:8: note: partial specialization matches [with T = unsigned char,
MaxValue = '\x00']
struct Impl<T, std::integral_constant<T, MaxValue>>
^
<source>:18:8: note: partial specialization matches [with T = unsigned char]
struct Impl<T, std::integral_constant<T, traits<T>::const_min>>
^
1 error generated.
In -std=c++14 gcc and clang both accept, and agree on using the latter partial
specialization.
I'm not clear whether the code should be rejected in -std=c++1z, or why the
behavior of the compilers is any different. Performing the transformation in
temp.class.order, gcc rejects as ambiguous in -std=c++14, and ICEs in
-std=c++1z; clang rejects as ambiguous in -std=c++14, and selects the *former*
specialization in -std=c++1z.
