Andrei Alexandrescu Style Guide
Overview
Andrei Alexandrescu's "Modern C++ Design" revolutionized how we think about C++ templates. His work on Loki library and policy-based design showed that templates are not just for containers—they're a compile-time programming language.
Core Philosophy
"C++ templates are Turing-complete. Use this power wisely."
"Policy-based design: assemble types from interchangeable parts."
Alexandrescu believes in pushing computation to compile time and using the type system as a design tool, not just a safety mechanism.
Design Principles
- Policy-Based Design: Build classes from interchangeable policy classes that customize behavior without inheritance overhead.
- Compile-Time over Runtime: What can be computed at compile time should be.
- Type Lists and Metaprogramming: Types themselves become first-class citizens that can be manipulated.
- Design Patterns in Types: Classic GoF patterns implemented with zero runtime overhead.
When Writing Code
Always
- Consider if behavior can be a compile-time policy
- Use
static_assertto document and enforce requirements - Prefer tag dispatching over runtime branching for type-based logic
- Make templates SFINAE-friendly (C++11/14) or use concepts (C++20)
- Document template requirements explicitly
Never
- Use runtime polymorphism when static polymorphism suffices
- Write duplicate code that differs only in types
- Ignore compile-time computation opportunities
- Leave template errors to become cryptic instantiation failures
Prefer
- Policy classes over strategy pattern (no vtable)
- Type traits over runtime type checking
constexprfunctions over template metafunctions (modern C++)- Concepts over SFINAE (C++20)
- Variadic templates over recursive type lists (modern C++)
Code Patterns
Policy-Based Design
// Traditional OOP: Runtime overhead, fixed at compile time anyway
class Widget : public ICreationPolicy, public IThreadingPolicy { /* ... */ };
// Policy-Based: Zero overhead, infinitely configurable
template <
class CreationPolicy,
class ThreadingPolicy = SingleThreaded,
class CheckingPolicy = NoChecking
>
class SmartPtr : public CreationPolicy,
public ThreadingPolicy,
public CheckingPolicy {
// Policies are mixed in, no vtable
};
// Usage: Configure at compile time
using ThreadSafePtr = SmartPtr<HeapCreation, MultiThreaded, AssertCheck>;
using FastPtr = SmartPtr<HeapCreation, SingleThreaded, NoChecking>;
// Policies are just classes with required interface
struct HeapCreation {
template<class T>
static T* Create() { return new T; }
template<class T>
static void Destroy(T* p) { delete p; }
};
struct SingleThreaded {
struct Lock {
Lock() = default; // No-op
};
};
struct MultiThreaded {
struct Lock {
Lock() { /* acquire mutex */ }
~Lock() { /* release mutex */ }
};
};Type Traits and SFINAE
// Type trait: Does T have a serialize() method?
template<typename T, typename = void>
struct has_serialize : std::false_type {};
template<typename T>
struct has_serialize<T,
std::void_t<decltype(std::declval<T>().serialize())>
> : std::true_type {};
// Use it for conditional behavior
template<typename T>
auto save(const T& obj) -> std::enable_if_t<has_serialize<T>::value> {
obj.serialize();
}
template<typename T>
auto save(const T& obj) -> std::enable_if_t<!has_serialize<T>::value> {
default_serialize(obj);
}
// C++20: Much cleaner with concepts
template<typename T>
concept Serializable = requires(T t) {
{ t.serialize() } -> std::convertible_to<std::string>;
};
void save(Serializable auto const& obj) {
obj.serialize();
}Compile-Time Type Lists (Classic Alexandrescu)
// Type list: A compile-time list of types
template<typename... Ts>
struct TypeList {};
// Operations on type lists
template<typename List>
struct Length;
template<typename... Ts>
struct Length<TypeList<Ts...>> {
static constexpr size_t value = sizeof...(Ts);
};
// Get type at index
template<size_t I, typename List>
struct TypeAt;
template<typename Head, typename... Tail>
struct TypeAt<0, TypeList<Head, Tail...>> {
using type = Head;
};
template<size_t I, typename Head, typename... Tail>
struct TypeAt<I, TypeList<Head, Tail...>> {
using type = typename TypeAt<I - 1, TypeList<Tail...>>::type;
};
// Usage
using MyTypes = TypeList<int, double, std::string>;
static_assert(Length<MyTypes>::value == 3);
using Second = TypeAt<1, MyTypes>::type; // doubleVisitor Pattern via Templates
// Traditional visitor: Virtual dispatch at every node
// Alexandrescu approach: Static visitor with type list
template<typename... Types>
class Variant;
template<typename Visitor, typename Variant>
auto visit(Visitor&& v, Variant&& var) {
return var.visit(std::forward<Visitor>(v));
}
// Modern C++ (std::variant does this)
using Value = std::variant<int, double, std::string>;
auto result = std::visit(overloaded{
[](int i) { return std::to_string(i); },
[](double d) { return std::to_string(d); },
[](const std::string& s) { return s; }
}, value);
// The 'overloaded' helper (Alexandrescu-style)
template<class... Ts>
struct overloaded : Ts... {
using Ts::operator()...;
};
template<class... Ts>
overloaded(Ts...) -> overloaded<Ts...>;Mental Model
Alexandrescu thinks of C++ templates as a compile-time functional language:
- Types as values: Types can be computed, stored, and transformed
- Templates as functions: Template instantiation is function application
- Specialization as pattern matching: Like case statements on types
- Recursion for iteration: Compile-time loops via recursive templates
The D Language Connection
Alexandrescu later co-designed D, which incorporates many C++ template lessons:
- Built-in compile-time function execution
- String mixins for code generation
- Better error messages for templates
These ideas now appear in modern C++ (constexpr, if constexpr, concepts).
When to Apply
Use Alexandrescu's techniques when:
- You need maximum performance (zero runtime overhead)
- Behavior variations are known at compile time
- You're building a library with many configuration options
- Type-based dispatch is frequent
Avoid when:
- Runtime polymorphism is genuinely needed
- Compile times are already problematic
- Team isn't comfortable with template metaprogramming