Don't help the compiler!

Clang Frontend
#1

Hi all,
STL (Stephan T. Lavavej) in his talk at Going Native 2013, gave this
example of an attempt to "help the compiler", which is at best
superfluous:

    auto p = std::make_pair<int, double>(42, 3.1415);

and recommended

    auto p = std::make_pair (42, 3.1415);

I thought of writing a plugin to warn in the "helpful" case. However,
clang -std=c++11 -Xclang -ast-dump produced virtually identical output.

Is there a way to distinguish between these two situations?

Csaba

pair_expl.ast (2.73 KB)

pair_impl.ast (2.73 KB)

#2

You can check the hasExplicitTemplateArgs bit on the DeclRefExpr to std::make_pair. We also have a clang-tidy checker for this pattern in ExplicitMakePairCheck.cpp that you can use as an example.

- Ben

#3

The latter form seems less readable. The fact that a floating point literal is a double unless suffixed with f is something that is rarely important for developers to remember, so this extra piece of information increases the cognitive burden when reading the code.

The goal of the former form is not to 'help the compiler', it is to help the poor sap who has to maintain this code in a couple of years.

David

#4

If you really need to think about whether it is float or double, then the
unnecessary cognitive burden is somewhere else entirely.

-- J. Z.

#5

Hi.

The latter form seems less readable. The fact that a floating point literal is a double unless suffixed with f is something that is rarely important for developers to remember, so this extra piece of information increases the cognitive burden when reading the code.

The goal of the former form is not to 'help the compiler', it is to help the poor sap who has to maintain this code in a couple of years.

While this might be true, the point of std::make_pair is to deduce the type of the arguments.
If one wants to explicitly specify the types, he can simply do:

auto p = std::pair<int,double>(42, 3.14);

or

std::pair<int,double> p{42, 3.14};

which are also a few less keystrokes. It doesn’t need make_pair, which is perfectly redundant if
you don’t want it to deduce the types.

Also in general, explicitly specifying the types of arguments of a template function which is
designed with deduction in mind, seems to me bad practice, because it’s fragile if one
change the arguments but not the types of vice-versa, and could break or slowdown things
when perfect-forwarding comes in the mix.

This example exacerbates the issue:
auto p = std::make_unique<T, ArgT1, ArgT2, ArgT3>(arg1, arg2, arg3); // Please don’t.

So IMHO a check to tell the inexperienced user about this redundancy is very helpful.

David

Bye,
Nicola