Home > Articles

  • Print
  • + Share This
This chapter is from the book

23.2 The Dimensions of Reflective Metaprogramming

Previously, we described value metaprogramming based on constexpr evaluation and type metaprogramming based on recursive template instantiation. These two options, both available in modern C++, clearly involve different methods driving the computation. It turns out that value metaprogramming can also be driven in terms of recursive template instantiation, and, prior to the introduction of constexpr functions in C++11, that was the mechanism of choice. For example, the following code computes a square root of an integer using recursive instantiation:

// primary template to compute sqrt(N)
template<int N, int LO=1, int HI=N>
struct Sqrt {
  // compute the midpoint, rounded up
  static constexpr auto mid = (LO+HI+1)/2;

  // search a not too large value in a halved interval
  static constexpr auto value = (N<mid*mid) ? Sqrt<N,LO,mid-1>::value
                                            : Sqrt<N,mid,HI>::value;
// partial specialization for the case when LO equals HI
template<int N, int M>
struct Sqrt<N,M,M> {
  static constexpr auto value = M;

This metaprogram uses much the same algorithm as our integer square root constexpr function in Section 23.1.1 on page 529, successively halving an interval known to contain the square root. However, the input to the metafunction is a nontype template argument instead of a function argument, and the “local variables” tracking the bounds to the interval are also recast as nontype template arguments. Clearly, this is a far less friendly approach than the constexpr function, but we will nevertheless analyze this code later on to examine how it consumes compiler resources.

In any case, we can see that the computational engine of metaprogramming could potentially have many options. That is, however, not the only dimension in which such options may be considered. Instead, we like to think that a comprehensive metaprogramming solution for C++ must make choices along three dimensions:

  • Computation

  • Reflection

  • Generation

Reflection is the ability to programmatically inspect features of the program. Generation refers to the ability to generate additional code for the program.

We have seen two options for computation: recursive instantiation and constexpr evaluation. For reflection, we have found a partial solution in type traits (see Section 19.6.1 on page 431). Although available traits enable quite a few advanced template techniques, they are far from covering all that is wanted from a reflection facility in the language. For example, given a class type, many applications would like to programmatically explore the members of that class. Our current traits are based on template instantiation, and C++ could conceivably provide additional language facilities or “intrinsic” library components5 to produce class template instances that contain the reflected information at compile time. Such an approach is a good match for computations based on recursive template instantiations. Unfortunately, class template instances consume a lot of compiler storage and that storage cannot be released until the end of the compilation (trying to do otherwise would result in taking significant more compilation time). An alternative option, which is expected to pair well with the constexpr evaluation option for the “computation” dimension, is to introduce a new standard type to represent reflected information. Section 17.9 on page 363 discusses this option (which is now under active investigation by the C++ standardization committee).

Section 17.9 on page 363 also shows a potential future approach to powerful code generation. Creating a flexible, general, and programmer-friendly code generation mechanism within the existing C++ language remains a challenge that is being investigated by various parties. However, it is also true that instantiating templates has always been a “code generation” mechanism of sorts. In addition, compilers have become reliable enough at expanding calls to small functions in-line that that mechanism can be used as a vehicle for code generation. Those observations are exactly what underlies our DotProductT example above and, combined with more powerful reflection facilities, existing techniques can already achieve remarkable metaprogramming effects.

  • + Share This
  • 🔖 Save To Your Account