# Crash Course: core functionalities # Table of Contents * [Introduction](#introduction) * [Any as in any type](#any-as-in-any-type) * [Small buffer optimization](#small-buffer-optimization) * [Alignment requirement](#alignment-requirement) * [Compressed pair](#compressed-pair) * [Enum as bitmask](#enum-as-bitmask) * [Hashed strings](#hashed-strings) * [Wide characters](wide-characters) * [Conflicts](#conflicts) * [Memory](#memory) * [Power of two and fast modulus](#power-of-two-and-fast-modulus) * [Allocator aware unique pointers](#allocator-aware-unique-pointers) * [Monostate](#monostate) * [Type support](#type-support) * [Built-in RTTI support](#built-in-rtti-support) * [Type info](#type-info) * [Almost unique identifiers](#almost-unique-identifiers) * [Type traits](#type-traits) * [Size of](#size-of) * [Is applicable](#is-applicable) * [Constness as](#constness-as) * [Member class type](#member-class-type) * [Integral constant](#integral-constant) * [Tag](#tag) * [Type list and value list](#type-list-and-value-list) * [Unique sequential identifiers](#unique-sequential-identifiers) * [Compile-time generator](#compile-time-generator) * [Runtime generator](#runtime-generator) * [Utilities](#utilities) # Introduction `EnTT` comes with a bunch of core functionalities mostly used by the other parts of the library itself.
Hardly users will include these features in their code, but it's worth describing what `EnTT` offers so as not to reinvent the wheel in case of need. # Any as in any type `EnTT` comes with its own `any` type. It may seem redundant considering that C++17 introduced `std::any`, but it is not (hopefully).
First of all, the _type_ returned by an `std::any` is a const reference to an `std::type_info`, an implementation defined class that's not something everyone wants to see in a software. Furthermore, there is no way to connect it with the type system of the library and therefore with its integrated RTTI support.
Note that this class is largely used internally by the library itself. The API is very similar to that of its most famous counterpart, mainly because this class serves the same purpose of being an opaque container for any type of value.
Instances of `any` also minimize the number of allocations by relying on a well known technique called _small buffer optimization_ and a fake vtable. Creating an object of the `any` type, whether empty or not, is trivial: ```cpp // an empty container entt::any empty{}; // a container for an int entt::any any{0}; // in place construction entt::any in_place{std::in_place_type, 42}; ``` Alternatively, the `make_any` function serves the same purpose but requires to always be explicit about the type: ```cpp entt::any any = entt::make_any(42); ``` In both cases, the `any` class takes the burden of destroying the contained element when required, regardless of the storage strategy used for the specific object.
Furthermore, an instance of `any` isn't tied to an actual type. Therefore, the wrapper is reconfigured when it's assigned a new object of a type other than the one it contains. There exists also a way to directly assign a value to the variable contained by an `entt::any`, without necessarily replacing it. This is especially useful when the object is used in _aliasing mode_, as described below: ```cpp entt::any any{42}; entt::any value{3}; // assigns by copy any.assign(value); // assigns by move any.assign(std::move(value)); ``` The `any` class will also perform a check on the type information and whether or not the original type was copy or move assignable, as appropriate.
In all cases, the `assign` function returns a boolean value to indicate the success or failure of the operation. When in doubt about the type of object contained, the `type` member function of `any` returns a const reference to the `type_info` associated with its element, or `type_id()` if the container is empty. The type is also used internally when comparing two `any` objects: ```cpp if(any == empty) { /* ... */ } ``` In this case, before proceeding with a comparison, it's verified that the _type_ of the two objects is actually the same.
Refer to the `EnTT` type system documentation for more details about how `type_info` works and on possible risks of a comparison. A particularly interesting feature of this class is that it can also be used as an opaque container for const and non-const references: ```cpp int value = 42; entt::any any{std::in_place_type(value)}; entt::any cany = entt::make_any(value); entt::any fwd = entt::forward_as_any(value); any.emplace(value); ``` In other words, whenever `any` is explicitly told to construct an _alias_, it acts as a pointer to the original instance rather than making a copy of it or moving it internally. The contained object is never destroyed and users must ensure that its lifetime exceeds that of the container.
Similarly, it's possible to create non-owning copies of `any` from an existing object: ```cpp // aliasing constructor entt::any ref = other.as_ref(); ``` In this case, it doesn't matter if the original container actually holds an object or acts already as a reference for unmanaged elements, the new instance thus created won't create copies and will only serve as a reference for the original item.
This means that, starting from the example above, both `ref` and `other` will point to the same object, whether it's initially contained in `other` or already an unmanaged element. As a side note, it's worth mentioning that, while everything works transparently when it comes to non-const references, there are some exceptions when it comes to const references.
In particular, the `data` member function invoked on a non-const instance of `any` that wraps a const reference will return a null pointer in all cases. To cast an instance of `any` to a type, the library offers a set of `any_cast` functions in all respects similar to their most famous counterparts.
The only difference is that, in the case of `EnTT`, these won't raise exceptions but will only trigger an assert in debug mode, otherwise resulting in undefined behavior in case of misuse in release mode. ## Small buffer optimization The `any` class uses a technique called _small buffer optimization_ to reduce the number of allocations where possible.
The default reserved size for an instance of `any` is `sizeof(double[2])`. However, this is also configurable if needed. In fact, `any` is defined as an alias for `basic_any`, where `Len` is the size above.
Users can easily set a custom size or define their own aliases: ```cpp using my_any = entt::basic_any; ``` This feature, in addition to allowing the choice of a size that best suits the needs of an application, also offers the possibility of forcing dynamic creation of objects during construction.
In other terms, if the size is 0, `any` avoids the use of any optimization and always dynamically allocates objects (except for aliasing cases). Note that the size of the internal storage as well as the alignment requirements are directly part of the type and therefore contribute to define different types that won't be able to interoperate with each other. ## Alignment requirement The alignment requirement is optional and by default the most stringent (the largest) for any object whose size is at most equal to the one provided.
The `basic_any` class template inspects the alignment requirements in each case, even when not provided and may decide not to use the small buffer optimization in order to meet them. The alignment requirement is provided as an optional second parameter following the desired size for the internal storage: ```cpp using my_any = entt::basic_any; ``` Note that the alignment requirements as well as the size of the internal storage are directly part of the type and therefore contribute to define different types that won't be able to interoperate with each other. # Compressed pair Primarily designed for internal use and far from being feature complete, the `compressed_pair` class does exactly what it promises: it tries to reduce the size of a pair by exploiting _Empty Base Class Optimization_ (or _EBCO_).
This class **is not** a drop-in replacement for `std::pair`. However, it offers enough functionalities to be a good alternative for when reducing memory usage is more important than having some cool and probably useless feature. Although the API is very close to that of `std::pair` (apart from the fact that the template parameters are inferred from the constructor and therefore there is no` entt::make_compressed_pair`), the major difference is that `first` and `second` are functions for implementation needs: ```cpp entt::compressed_pair pair{0, 3.}; pair.first() = 42; ``` There isn't much to describe then. It's recommended to rely on documentation and intuition. At the end of the day, it's just a pair and nothing more. # Enum as bitmask Sometimes it's useful to be able to use enums as bitmasks. However, enum classes aren't really suitable for the purpose out of the box. Main problem is that they don't convert implicitly to their underlying type.
All that remains is to make a choice between using old-fashioned enums (with all their problems that I don't want to discuss here) or writing _ugly_ code. Fortunately, there is also a third way: adding enough operators in the global scope to treat enum classes as bitmask transparently.
The ultimate goal is to be able to write code like the following (or maybe something more meaningful, but this should give a grasp and remain simple at the same time): ```cpp enum class my_flag { unknown = 0x01, enabled = 0x02, disabled = 0x04 }; const my_flag flags = my_flag::enabled; const bool is_enabled = !!(flags & my_flag::enabled); ``` The problem with adding all operators to the global scope is that these will come into play even when not required, with the risk of introducing errors that are difficult to deal with.
However, C++ offers enough tools to get around this problem. In particular, the library requires users to register all enum classes for which bitmask support should be enabled: ```cpp template<> struct entt::enum_as_bitmask : std::true_type {}; ``` This is handy when dealing with enum classes defined by third party libraries and over which the users have no control. However, it's also verbose and can be avoided by adding a specific value to the enum class itself: ```cpp enum class my_flag { unknown = 0x01, enabled = 0x02, disabled = 0x04, _entt_enum_as_bitmask }; ``` In this case, there is no need to specialize the `enum_as_bitmask` traits, since `EnTT` will automatically detect the flag and enable the bitmask support.
Once the enum class has been registered (in one way or the other) all the most common operators will be available, such as `&`, `|` but also `&=` and `|=`. Refer to the official documentation for the full list of operators. # Hashed strings A hashed string is a zero overhead unique identifier. Users can use human-readable identifiers in the codebase while using their numeric counterparts at runtime, thus without affecting performance.
The class has an implicit `constexpr` constructor that chews a bunch of characters. Once created, all what one can do with it is getting back the original string through the `data` member function or converting the instance into a number.
The good part is that a hashed string can be used wherever a constant expression is required and no _string-to-number_ conversion will take place at runtime if used carefully. Example of use: ```cpp auto load(entt::hashed_string::hash_type resource) { // uses the numeric representation of the resource to load and return it } auto resource = load(entt::hashed_string{"gui/background"}); ``` There is also a _user defined literal_ dedicated to hashed strings to make them more user-friendly: ```cpp using namespace entt::literals; constexpr auto str = "text"_hs; ``` To use it, remember that all user defined literals in `EnTT` are enclosed in the `entt::literals` namespace. Therefore, the entire namespace or selectively the literal of interest must be explicitly included before each use, a bit like `std::literals`.
Finally, in case users need to create hashed strings at runtime, this class also offers the necessary functionalities: ```cpp std::string orig{"text"}; // create a full-featured hashed string... entt::hashed_string str{orig.c_str()}; // ... or compute only the unique identifier const auto hash = entt::hashed_string::value(orig.c_str()); ``` This possibility shouldn't be exploited in tight loops, since the computation takes place at runtime and no longer at compile-time and could therefore impact performance to some degrees. ## Wide characters The hashed string has a design that is close to that of an `std::basic_string`. It means that `hashed_string` is nothing more than an alias for `basic_hashed_string`. For those who want to use the C++ type for wide character representation, there exists also the alias `hashed_wstring` for `basic_hashed_string`.
In this case, the user defined literal to use to create hashed strings on the fly is `_hws`: ```cpp constexpr auto str = L"text"_hws; ``` Note that the hash type of the `hashed_wstring` is the same of its counterpart. ## Conflicts The hashed string class uses internally FNV-1a to compute the numeric counterpart of a string. Because of the _pigeonhole principle_, conflicts are possible. This is a fact.
There is no silver bullet to solve the problem of conflicts when dealing with hashing functions. In this case, the best solution seemed to be to give up. That's all.
After all, human-readable unique identifiers aren't something strictly defined and over which users have not the control. Choosing a slightly different identifier is probably the best solution to make the conflict disappear in this case. # Memory There are a handful of tools within EnTT to interact with memory in one way or another.
Some are geared towards simplifying the implementation of (internal or external) allocator aware containers. Others, on the other hand, are designed to help the developer with everyday problems. The former are very specific and for niche problems. These are tools designed to unwrap fancy or plain pointers (`to_address`) or to help forget the meaning of acronyms like _POCCA_, _POCMA_ or _POCS_.
I won't describe them here in detail. Instead, I recommend reading the inline documentation to those interested in the subject. ## Power of two and fast modulus Finding out if a number is a power of two (`is_power_of_two`) or what the next power of two is given a random value (`next_power_of_two`) is very useful at times.
For example, it helps to allocate memory in pages having a size suitable for the fast modulus: ```cpp const std::size_t result = entt::fast_mod(value, modulus); ``` Where `modulus` is necessarily a power of two. Perhaps not everyone knows that this type of operation is far superior in terms of performance to the basic modulus and for this reason preferred in many areas. ## Allocator aware unique pointers A nasty thing in C++ (at least up to C++20) is the fact that shared pointers support allocators while unique pointers don't.
There is a proposal at the moment that also shows among the other things how this can be implemented without any compiler support. The `allocate_unique` function follows this proposal, making a virtue out of necessity: ```cpp std::unique_ptr> ptr = entt::allocate_unique(allocator, arguments); ``` Although the internal implementation is slightly different from what is proposed for the standard, this function offers an API that is a drop-in replacement for the same feature. # Monostate The monostate pattern is often presented as an alternative to a singleton based configuration system. This is exactly its purpose in `EnTT`. Moreover, this implementation is thread safe by design (hopefully).
Keys are represented by hashed strings, values are basic types like `int`s or `bool`s. Values of different types can be associated to each key, even more than one at a time. Because of this, users must pay attention to use the same type both during an assignment and when they try to read back their data. Otherwise, they will probably incur in unexpected results. Example of use: ```cpp entt::monostate{} = true; entt::monostate<"mykey"_hs>{} = 42; // ... const bool b = entt::monostate<"mykey"_hs>{}; const int i = entt::monostate{}; ``` # Type support `EnTT` provides some basic information about types of all kinds.
It also offers additional features that are not yet available in the standard library or that will never be. ## Built-in RTTI support Runtime type identification support (or RTTI) is one of the most frequently disabled features in the C++ world, especially in the gaming sector. Regardless of the reasons for this, it's often a shame not to be able to rely on opaque type information at runtime.
The library tries to fill this gap by offering a built-in system that doesn't serve as a replacement but comes very close to being one and offers similar information to that provided by its counterpart. Basically, the whole system relies on a handful of classes. In particular: * The unique sequential identifier associated with a given type: ```cpp auto index = entt::type_index::value(); ``` The returned value isn't guaranteed to be stable across different runs. However, it can be very useful as index in associative and unordered associative containers or for positional accesses in a vector or an array. So as not to conflict with the other tools available, the `family` class isn't used to generate these indexes. Therefore, the numeric identifiers returned by the two tools may differ.
On the other hand, this leaves users with full powers over the `family` class and therefore the generation of custom runtime sequences of indices for their own purposes, if necessary. An external generator can also be used if needed. In fact, `type_index` can be specialized by type and is also _sfinae-friendly_ in order to allow more refined specializations such as: ```cpp template struct entt::type_index> { static entt::id_type value() ENTT_NOEXCEPT { return Type::index(); } }; ``` Note that indexes **must** still be generated sequentially in this case.
The tool is widely used within `EnTT`. Generating indices not sequentially would break an assumption and would likely lead to undesired behaviors. * The hash value associated with a given type: ```cpp auto hash = entt::type_hash::value(); ``` In general, the `value` function exposed by `type_hash` is also `constexpr` but this isn't guaranteed for all compilers and platforms (although it's valid with the most well-known and popular ones). This function **can** use non-standard features of the language for its own purposes. This makes it possible to provide compile-time identifiers that remain stable across different runs.
In all cases, users can prevent the library from using these features by means of the `ENTT_STANDARD_CPP` definition. In this case, there is no guarantee that identifiers remain stable across executions. Moreover, they are generated at runtime and are no longer a compile-time thing. As for `type_index`, also `type_hash` is a _sfinae-friendly_ class that can be specialized in order to customize its behavior globally or on a per-type or per-traits basis. * The name associated with a given type: ```cpp auto name = entt::type_name::value(); ``` The name associated with a type is extracted from some information generally made available by the compiler in use. Therefore, it may differ depending on the compiler and may be empty in the event that this information isn't available.
For example, given the following class: ```cpp struct my_type { /* ... */ }; ``` The name is `my_type` when compiled with GCC or CLang and `struct my_type` when MSVC is in use.
Most of the time the name is also retrieved at compile-time and is therefore always returned through an `std::string_view`. Users can easily access it and modify it as needed, for example by removing the word `struct` to standardize the result. `EnTT` won't do this for obvious reasons, since it requires copying and creating a new string potentially at runtime. This function **can** use non-standard features of the language for its own purposes. Users can prevent the library from using non-standard features by means of the `ENTT_STANDARD_CPP` definition. In this case, the name will be empty by default. As for `type_index`, also `type_name` is a _sfinae-friendly_ class that can be specialized in order to customize its behavior globally or on a per-type or per-traits basis. These are then combined into utilities that aim to offer an API that is somewhat similar to that offered by the language. ### Type info The `type_info` class isn't a drop-in replacement for `std::type_info` but can provide similar information which are not implementation defined and don't require to enable RTTI.
Therefore, they can sometimes be even more reliable than those obtained otherwise. Its type defines an opaque class that is also copyable and movable.
Objects of this type are generally returned by the `type_id` functions: ```cpp // by type auto info = entt::type_id(); // by value auto other = entt::type_id(42); ``` All elements thus received are nothing more than const references to instances of `type_info` with static storage duration.
This is convenient for saving the entire object aside for the cost of a pointer. However, nothing prevents from constructing `type_info` objects directly: ```cpp entt::type_info info{std::in_place_type}; ``` These are the information made available by `type_info`: * The index associated with a given type: ```cpp auto idx = entt::type_id().index(); ``` This is also an alias for the following: ```cpp auto idx = entt::type_index>>::value(); ``` * The hash value associated with a given type: ```cpp auto hash = entt::type_id().hash(); ``` This is also an alias for the following: ```cpp auto hash = entt::type_hash>>::value(); ``` * The name associated with a given type: ```cpp auto name = entt::type_id().name(); ``` This is also an alias for the following: ```cpp auto name = entt::type_name>>::value(); ``` Where all accessed features are available at compile-time, the `type_info` class is also fully `constexpr`. However, this cannot be guaranteed in advance and depends mainly on the compiler in use and any specializations of the classes described above. ### Almost unique identifiers Since the default non-standard, compile-time implementation of `type_hash` makes use of hashed strings, it may happen that two types are assigned the same hash value.
In fact, although this is quite rare, it's not entirely excluded. Another case where two types are assigned the same identifier is when classes from different contexts (for example two or more libraries loaded at runtime) have the same fully qualified name. In this case, also `type_name` will return the same value for the two types.
Fortunately, there are several easy ways to deal with this: * The most trivial one is to define the `ENTT_STANDARD_CPP` macro. Runtime identifiers don't suffer from the same problem in fact. However, this solution doesn't work well with a plugin system, where the libraries aren't linked. * Another possibility is to specialize the `type_name` class for one of the conflicting types, in order to assign it a custom identifier. This is probably the easiest solution that also preserves the feature of the tool. * A fully customized identifier generation policy (based for example on enum classes or preprocessing steps) may represent yet another option. These are just some examples of possible approaches to the problem but there are many others. As already mentioned above, since users have full control over their types, this problem is in any case easy to solve and should not worry too much.
In all likelihood, it will never happen to run into a conflict anyway. ## Type traits A handful of utilities and traits not present in the standard template library but which can be useful in everyday life.
This list **is not** exhaustive and contains only some of the most useful classes. Refer to the inline documentation for more information on the features offered by this module. ### Size of The standard operator `sizeof` complains when users provide it for example with function or incomplete types. On the other hand, it's guaranteed that its result is always nonzero, even if applied to an empty class type.
This small class combines the two and offers an alternative to `sizeof` that works under all circumstances, returning zero if the type isn't supported: ```cpp const auto size = entt::size_of_v; ``` ### Is applicable The standard library offers the great `std::is_invocable` trait in several forms. This takes a function type and a series of arguments and returns true if the condition is satisfied.
Moreover, users are also provided with `std::apply`, a tool for combining invocable elements and tuples of arguments. It would therefore be a good idea to have a variant of `std::is_invocable` that also accepts its arguments in the form of a tuple-like type, so as to complete the offer: ```cpp constexpr bool result = entt::is_applicable>; ``` This trait is built on top of `std::is_invocable` and does nothing but unpack a tuple-like type and simplify the code at the call site. ### Constness as An utility to easily transfer the constness of a type to another type: ```cpp // type is const dst_type because of the constness of src_type using type = entt::constness_as_t; ``` The trait is subject to the rules of the language. Therefore, for example, transferring constness between references won't give the desired effect. ### Member class type The `auto` template parameter introduced with C++17 made it possible to simplify many class templates and template functions but also made the class type opaque when members are passed as template arguments.
The purpose of this utility is to extract the class type in a few lines of code: ```cpp template using clazz = entt::member_class_t; ``` ### Integral constant Since `std::integral_constant` may be annoying because of its form that requires to specify both a type and a value of that type, there is a more user-friendly shortcut for the creation of integral constants.
This shortcut is the alias template `entt::integral_constant`: ```cpp constexpr auto constant = entt::integral_constant<42>; ``` Among the other uses, when combined with a hashed string it helps to define tags as human-readable _names_ where actual types would be required otherwise: ```cpp constexpr auto enemy_tag = entt::integral_constant<"enemy"_hs>; registry.emplace(entity); ``` ### Tag Since `id_type` is very important and widely used in `EnTT`, there is a more user-friendly shortcut for the creation of integral constants based on it.
This shortcut is the alias template `entt::tag`. If used in combination with hashed strings, it helps to use human-readable names where types would be required otherwise. As an example: ```cpp registry.emplace>(entity); ``` However, this isn't the only permitted use. Literally any value convertible to `id_type` is a good candidate, such as the named constants of an unscoped enum. ### Type list and value list There is no respectable library where the much desired _type list_ can be missing.
`EnTT` is no exception and provides (making extensive use of it internally) the `type_list` type, in addition to its `value_list` counterpart dedicated to non-type template parameters. Here is a (possibly incomplete) list of the functionalities that come with a type list: * `type_list_element[_t]` to get the N-th element of a type list. * `type_list_cat[_t]` and a handy `operator+` to concatenate type lists. * `type_list_unique[_t]` to remove duplicate types from a type list. * `type_list_contains[_v]` to know if a type list contains a given type. * `type_list_diff[_t]` to remove types from type lists. I'm also pretty sure that more and more utilities will be added over time as needs become apparent.
Many of these functionalities also exist in their version dedicated to value lists. We therefore have `value_list_element[_v]` as well as `value_list_cat[_t]`and so on. # Unique sequential identifiers Sometimes it's useful to be able to give unique, sequential numeric identifiers to types either at compile-time or runtime.
There are plenty of different solutions for this out there and I could have used one of them. However, I decided to spend my time to define a couple of tools that fully embraces what the modern C++ has to offer. ## Compile-time generator To generate sequential numeric identifiers at compile-time, `EnTT` offers the `identifier` class template: ```cpp // defines the identifiers for the given types using id = entt::identifier; // ... switch(a_type_identifier) { case id::type: // ... break; case id::type: // ... break; default: // ... } ``` This is all what this class template has to offer: a `type` inline variable that contains a numeric identifier for the given type. It can be used in any context where constant expressions are required. As long as the list remains unchanged, identifiers are also guaranteed to be stable across different runs. In case they have been used in a production environment and a type has to be removed, one can just use a placeholder to left the other identifiers unchanged: ```cpp template struct ignore_type {}; using id = entt::identifier< a_type_still_valid, ignore_type, another_type_still_valid >; ``` Perhaps a bit ugly to see in a codebase but it gets the job done at least. ## Runtime generator To generate sequential numeric identifiers at runtime, `EnTT` offers the `family` class template: ```cpp // defines a custom generator using id = entt::family; // ... const auto a_type_id = id::type; const auto another_type_id = id::type; ``` This is all what a _family_ has to offer: a `type` inline variable that contains a numeric identifier for the given type.
The generator is customizable, so as to get different _sequences_ for different purposes if needed. Please, note that identifiers aren't guaranteed to be stable across different runs. Indeed it mostly depends on the flow of execution. # Utilities It's not possible to escape the temptation to add utilities of some kind to a library. In fact, `EnTT` also provides a handful of tools to simplify the life of developers: * `entt::identity`: the identity function object that will be available with C++20. It returns its argument unchanged and nothing more. It's useful as a sort of _do nothing_ function in template programming. * `entt::overload`: a tool to disambiguate different overloads from their function type. It works with both free and member functions.
Consider the following definition: ```cpp struct clazz { void bar(int) {} void bar() {} }; ``` This utility can be used to get the _right_ overload as: ```cpp auto *member = entt::overload(&clazz::bar); ``` The line above is literally equivalent to: ```cpp auto *member = static_cast(&clazz::bar); ``` Just easier to read and shorter to type. * `entt::overloaded`: a small class template used to create a new type with an overloaded `operator()` from a bunch of lambdas or functors.
As an example: ```cpp entt::overloaded func{ [](int value) { /* ... */ }, [](char value) { /* ... */ } }; func(42); func('c'); ``` Rather useful when doing metaprogramming and having to pass to a function a callable object that supports multiple types at once. * `entt::y_combinator`: this is a C++ implementation of **the** _y-combinator_. If it's not clear what it is, there is probably no need for this utility.
Below is a small example to show its use: ```cpp entt::y_combinator gauss([](const auto &self, auto value) -> unsigned int { return value ? (value + self(value-1u)) : 0; }); const auto result = gauss(3u); ``` Maybe convoluted at a first glance but certainly effective. Unfortunately, the language doesn't make it possible to do much better. This is a rundown of the (actually few) utilities made available by `EnTT`. The list will probably grow over time but the size of each will remain rather small, as has been the case so far.