VS2019 and template specialisation

James Lothian

unread,

Mar 23, 2021, 7:22:19 PM3/23/21

to

Given the following code:

template <int L, int M>
class Unit
{};

template <class U1, class U2>
struct Mul
{};

template <int... L, int... R>
struct Mul<Unit<L...>, Unit<R...> >
{
typedef Unit<L + R...> Result;
};

typedef Unit<1, 0> Length;
typedef Unit<0, 1> Mass;

// Error here
typedef typename Mul<Length, Mass>::Result LM;

int main()
{
LM lm;
return 0;
}

Visual Studio 2019 and 2015 both say:
C:\Users\James\Documents\SpecTest\SpecTest.cpp(21,37): error C2039:
'Result': is not a member of 'Mul<Length,Mass>'
C:\Users\James\Documents\SpecTest\SpecTest.cpp(21): message : see
declaration of 'Mul<Length,Mass>'

Apparently, VC++ picks the base case of the template rather than the
specialisation.
Every online c++ compiler I've managed to find (presumably using gcc)
accepts this code. So, who's right? Am I doing something stupid here?

Thanks,
James

Alf P. Steinbach

unread,

Mar 24, 2021, 10:21:18 AM3/24/21

to

On 24.03.2021 00:22, James Lothian wrote:
> template <int L, int M>
> class Unit
> {};
>
> template <class U1, class U2>
> struct Mul
> {};
>
> template <int... L, int... R>
> struct Mul<Unit<L...>, Unit<R...> >
> {
>     typedef Unit<L + R...> Result;
> };
>
> typedef Unit<1, 0> Length;
> typedef Unit<0, 1> Mass;
>
> // Error here
> typedef typename Mul<Length, Mass>::Result LM;
>
> int main()
> {
>     LM lm;
>     return 0;
> }

Looks like a Visual C++ bug. And it looks like a bug related to
parameter pack handling, and not related to (lack of) two-phase template
compilation. I think you should report it.

Workaround:

---------------------------------------------

template< int L, int M >
class Unit {};

template< class U1, class U2 >
struct Mul;

template< int L1, int L2, int R1, int R2 >
struct Mul<Unit<L1, L2>, Unit<R1, R2> >
{
using Result = Unit<L1 + R1, L2 + R2>;
};

using Length = Unit<1, 0>;
using Mass = Unit<0, 1>;

using LM = typename Mul<Length, Mass>::Result;

auto main() -> int
{
LM lm;
(void) lm;
}
---------------------------------------------

However, to my eyes the units computation doesn't seem correct. I don't
know if Boost Units does this correctly, but I think it must. So I
recommend using Boost Units instead.

<url: https://www.boost.org/doc/libs/1_65_0/doc/html/boost_units.html>

- Alf

James Lothian

unread,

Mar 24, 2021, 3:25:37 PM3/24/21

to

Alf P. Steinbach wrote:

> Looks like a Visual C++ bug. And it looks like a bug related to
> parameter pack handling, and not related to (lack of) two-phase template
> compilation. I think you should report it.
>

Thank you, that's pretty much what I thought.
> Workaround:

This is a much distilled-down version of something a good bit
bigger. I know I can work round this by avoiding the parameter pack and
fixing the number of dimensions in a Unit, but this means that when I
add another dimension to my units, I have to modify Mul (and Div, and
Exp, and so on). I've found that the following creeping horror works:

template <class U1, class U2>
struct Mul
{

private:

template <int... L, int... R>

static auto mulGuts(Unit<L...>, Unit<R...>)
{
return Unit<L + R...>();
}

public:
typedef decltype(mulGuts(U1(), U2())) Result;
};

but it does creep and it is horrible.

>
> However, to my eyes the units computation doesn't seem correct. I don't
> know if Boost Units does this correctly, but I think it must. So I
> recommend using Boost Units instead.
>
> <url: https://www.boost.org/doc/libs/1_65_0/doc/html/boost_units.html>

But where's the fun in that :-)

Thanks,
James

MrSpook_t...@8ihqvjzp9.gov

unread,

Mar 25, 2021, 4:58:58 AM3/25/21

to

On Wed, 24 Mar 2021 19:25:24 +0000
James Lothian <jamesl...@gmail.com> wrote:
>template <class U1, class U2>
>struct Mul
>{
>private:
> template <int... L, int... R>
> static auto mulGuts(Unit<L...>, Unit<R...>)
> {
> return Unit<L + R...>();
> }
>
>public:
> typedef decltype(mulGuts(U1(), U2())) Result;
>};
>
>but it does creep and it is horrible.

I can't help thinking that if you need syntax that mangled perhaps you should
revisit your design.

James Lothian

unread,

Mar 25, 2021, 9:57:18 AM3/25/21

to

I wouldn't need syntax as mangled as this if I weren't working around a
deficiency in VS2019.

James

Andrey Tarasevich

unread,

Apr 23, 2021, 3:17:07 PM4/23/21

to

On 3/23/2021 4:22 PM, James Lothian wrote:
> Every online c++ compiler I've managed to find (presumably using gcc)
> accepts this code. So, who's right? Am I doing something stupid here?

Yes, it appears to be a bug in MSVC++. It appears that if you separate
the first parameter from each pack, it serves as a workaround and MSVC++
choses the proper specialization

template <int L1, int... L, int R1, int... R>
struct Mul<Unit<L1, L...>, Unit<R1, R...> >
{
typedef Unit<L1 + R1, L + R...> Result;
};

--
Best regards,
Andrey Tarasevich

James Lothian

unread,

Apr 27, 2021, 7:28:31 AM4/27/21

to

Thanks for this -- it's less horrible than the workaround I've been using.

James