https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120760
Bug ID: 120760
Summary: Deducing array bound with a small unsigned integral
type silently overflows
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: mital at mitalashok dot co.uk
Target Milestone: ---
The following fails to compile with `-std=c++11 -fsyntax-only`
<https://godbolt.org/z/E9ffPfTev> (You can make a version that fails in C++98
too):
template<long long Val, long long X>
struct hard_error_on {
static_assert(X != Val, "");
using type = void;
};
template<bool i> typename hard_error_on<false, i>::type
f1(int(&)[i]);
void f1(...);
template<unsigned char i> typename hard_error_on<1, i>::type
f2(int(&)[i]);
void f2(...);
int main() {
int a[2];
int b[257];
f1(a);
f2(b);
}
It fails both static_asserts, showing it tries to instantiate f1<false> and
f2<1>.
If the type is signed, it ends up working because signed overflow isn't a
constant expression.
It makes no sense to deduce `0` and `1` as the sizes of `int[2]` and `int[257]`
respectively.
<https://wg21.link/temp.deduct#footnote-120> also explicitly says:
> 120) Although the template-argument corresponding to a template parameter of
> type bool can be deduced from an array bound, the resulting value will always
> be true because the array bound will be nonzero.
But GCC gives `false` to the constant parameter
