clang fails to diagnose initializer list narrowing in constructor call during template instantiation

[efriedma-quic]

Bugzilla Link	45331
Version	unspecified
OS	Windows NT
CC	@tambry,@zygoloid

Extended Description

struct A { signed char a; constexpr A(signed char a) : a(a) {} };
template<int x> constexpr A f(int y, int (*z)[A{x}.a] = 0) { return A(1); }
template<int x> constexpr A f(char* y) { return A(2); }
static_assert(f<257>(0).a == 2, "Should pick second overload");

gcc accepts, clang rejects. I think gcc is right?

[ec04fc15-fa35-46f2-80e1-5d271f2ef708]

Reduced slightly:

struct A { A(char); };
template<int x> A a = A{x};
template A a<257>; // ill-formed, but clang accepts

[xgupta]

godbolt link - https://godbolt.org/z/fbWscjsGx.

clang produces a warning for int to char conversion while gcc gives an error for narrowing conversion for the above test case. CheckImplicitConversion of SemaChecking.cpp might be relevent here?

[shafik]

We still pick the first overload: https://godbolt.org/z/EPf3Gqzcv

but @zygoloid example we no longer accept.

[zygoloid]

but @zygoloid example we no longer accept.

I'm not seeing that: https://godbolt.org/z/18Eos4G3M (we should reject under -pedantic-errors at least; and we do outside of template instantiation).