Should I follow the Epic coding standard?

This one will be controversial but I really wish someone told me all of this. When I initially started working with Unreal Engine I defaulted to following the official standards fully, not having known better. This remains excellent advice if you don’t but I do now, and maybe this will help inform your choices to not end up with some of the same regrets and technical debt that I now have to deal with.

There are a few situations where the answer is an obvious “yes”:

• You’re working on engine code intending to submit it to Epic
• You’re making a plugin for the Marketplace (at least on your public API)

There’s also clear value in having a common coding convention across multiple UE teams and different companies so deviating from it too far would also be silly. “Modern C++” still retains some controversy and Unreal Engine, like the majority of C++ projects disables exceptions and RTTI which neuters some of it when it comes to exception safety. UnrealHeaderTool also imposes some of its own limitations, most notably around namespaces.

With all that said, there is also a case to be made for not following these conventions fully. The main reason behind many of them is to be consistent with the existing heaps of legacy engine code, which doesn’t apply to your projects. Epic themselves are getting more lax when it comes to following their own conventions (seems to be different per module) and it was officially relaxed a little for UE5.

What this is not is some kind of “better” or alternative coding standard, or your defense if you want to do stuff like for (auto i{0};.... Please don’t, but take some time to determine if completely following the Epic standards is really the best for your team, or if deviating from it would be better. In our case there was a gradual slide away from the Epic standards which has led directly to bugfixes and cleaner code.

Not all of these items are from Epic’s official coding standard. Some are widespread and perhaps misguided/outdated practices that I’ve seen within the community that I’ll address without distinguishing between the two.

PascalCase everything

Epic seriously overuses PascalCase and doesn’t use camelCase at all… at least as far as the coding standard documentation goes. You can find multiple instances of them using it anyway, and there’s clear value in not leaving camelCase on the table.

Consider using it for some of: function parameters, locals, class fields, maybe only for non-properties or private, etc. Up to you where you split it. Local variables are probably the strongest contenders for it.

Pointers for UObjects

Reading engine code I soon got the impression that every UObject (including actors of course) has to be stored and passed around as a pointer (UObject* originally, then TObjectPtr<UObject>). Epic code often uses UObject&, and TArray::Sort even enforces it (but don’t use that one in particular, see below on std::sort).

int32

You’re supposed to use int32 for every int that may be serialized, e.g., on UPROPERTYs. This is correct to do, but also practically speaking you don’t need to do it. int is widely expected to remain exactly 32 bit on all platforms that run UE, forever.

If int32 is not used everywhere by default it stands out as explicitly marking something a 32-bit value, bringing extra attention to this which can help guide your reader.

auto

Epic has a very restrictive stance on auto that goes directly against established modern C++ practices, mainly driven by a misunderstanding of its motivation as a typing reduction tool instead of the correctness, performance, maintainability, and robustness improvement that it mainly provides. 1 2 3

Even if you disagree with one of the foremost experts on C++, there are situations where you can use it with no drawbacks (imagined or real). It’s important to distinguish between “decorated” and regular auto, see below.

I can personally attest to this: porting an auto-heavy UE4 codebase to UE5 highlighted and prevented a lot of implicit float-double conversions (performance) that would have had issues when converting large LWC coordinates beyond the old WORLD_MAX (correctness, maintainability), something that Epic has been manually tracking down and fixing for multiple 5.x releases.

auto is completely harmless when the type is obvious, and even on the same line:

auto* thing = Cast<UWhatCouldThisPossiblyBe>(object);

You can also consider using it when the type is not immediately visible, but it’s still obvious what the variable will be:

auto& LAM = GetWorld()->GetLatentActionManager();

Note that I was mainly using auto* and auto&, and that’s what I’d suggest using most of the time in code interacting with the engine, where UObjects don’t conform to RAII, and interfaces are not consistently designed, and TObjectPtr’s existence makes things even more complicated.
_{Exercise: guess what TMap::Find, TMap::FindChecked, and TMap::FindRef return. Did you get it right?}

const auto a = Cast<AActor>(Object); // This is not const AActor*!
auto b = GetSomething(); // Is this a copy? A reference?
auto&& c = GetSomething(); // This can be both an lvalue or an rvalue reference!

There’s also an auto-like syntax for returning a default value, return {}; that is less chatty. Otherwise {} is best limited to std::initializer_list overloads: even the STL itself is defined in many cases to construct values with T(params) instead of T{params}.

float literals

During UE4->UE5 LWC porting a lot of 0.0fs and 1.0fs needed to be fixed. 0 and 1 worked for both float and double without ambiguous overload errors and it’s what I would’ve used in hindsight, especially if it’s generic code that still needs to support both types. Obviously it doesn’t apply if you need to select a particular overload. Chances are you can also settle on doubles forever and use 0.0/1.0.

Structured bindings

Epic employees are not allowed to use this, but you can! It makes code a lot cleaner when you don’t have to create your function’s return values at every call site just to give to a T& out parameter.

If you need to reassign values to existing variables, there’s std::tie for std::tuple and Tie for TTuple (see below on STL usage).

auto [two, returns] = YayNoOutRefs();
std::tie(two, returns) = YayNoOutRefs();
Tie(two, returns) = YayNoOutRefs();

Use of namespace std

It’s hard to give a short, easy-to-remember guideline for this one. There are plenty of issues with the STL, but Epic’s replacements have their own issues as well. When to use which is mostly situational and based on experience and even engine version. The official conventions now state:

When there is a choice between a standard library feature instead of our own, prefer the option which gives superior results, but bear in mind that consistency is also valued highly.

You don’t have to be consistent with Epic’s decades’ worth of code in your own project! :)

There was a strong argument to be made for not using the STL at all simply for consistency’s sake, but Epic themselves have started using the STL in some situations (see below), this is unfortunately (or fortunately?) getting weaker and weaker as time goes on: future Unreal code will look like a mixture between Epic and STL style anyway.

Containers

Containers are one of the weakest points of the STL, and you’re forced to use the TContainers for anything exposed to the reflection system anyway. The Unreal versions generally look and perform better, have a better syntax for custom allocators, but are sloppier when it comes to type safety.

For instance, if you try and push_back an immovable type into a std::vector, it will correctly not compile and prevent any bugs that might arise from this. TArray will happily memcpy you around without a warning. Keep this in mind when you’re working with types that have special or unknown needs (such as in template code). You might want to add a few static_asserts in this case.

Algo

On the other hand, <algorithm> is one of the best parts of the STL, and Epic’s Algo namespace has some questionable implementation choices and is often outperformed by STL algorithms (for instance std::sort). At this point I default to STL algorithms, deferring to Algo only in situations where the STL-like iterators that the UE containers provide are not enough.

The EASTL might work even better if you’re OK with adding the extra dependency.

std::ranges improves the begin/end syntax of regular <algorithm> (where Algo remains better) in C++20 but it’s more or less incompatible with UE containers.

You can get a lot of mileage out of this small adapter function:

#include <span>

auto AsSpan(auto& Container)
{
    return std::span(GetData(Container), GetNum(Container));
}

Smart pointers

TUniquePtr has some bizarre bugs in UE4 that were fixed in UE5, and is mostly a “just because” replacement that serves no additional function.

TSharedPtr performs similarly to std::shared_ptr when in its default, thread-safe mode. For game thread stuff, making it not thread safe significantly improves its performance (also applies to TWeakPtr/std::weak_ptr).

TSharedFromThis has some weird limitations compared to std::enable_shared_from_this. If you’re hitting mysterious assertions, consider using the STL version.

std::move / std::forward

MoveTemp iswas a great replacement for std::move because it enforces move semantics with a static_assert. On the other hand MoveTempIfPossible is exactly like std::move.

std::forward and Forward are equivalent.

For all of these, the STL versions use static_cast while the UE ones use C-style casts.

Update: MSVC in VS2022 17.5 has added intrinsics for std::move and std::forward that greatly improve performance in debug builds, making this choice less obvious.

Miscellaneous

TAtomic is an easy case. std::atomic is the officially-preferred atomic and should always be used instead.

The Epic type traits and SFINAE helpers (TEnableIf, etc.) are also officially deprecated in favor of their STL counterparts. If you’re using UE5, C++17 can help eliminate a lot of SFINAE with if constexpr and C++20 lets you go even further beyond with concepts if you can opt in.

The same applies to std::numeric_limits, etc. If an Epic reimplementation provides no benefit but it’s only there for the sake of it, it’s worth defaulting to the STL which receives more testing and focused dev attention.

TTuple and TVariant are incomplete reimplementations of their STL variants. They got better in UE5 but there are still STL-exclusive features (like a working operator==…) and they have started becoming STL hybrids themselves due to language features being defined in terms of std::tuple_size, (lower-case) get<i>(), etc. On the other hand, they have FArchive support, which means the best one to choose depends on usage.

TFunction has (or had) some bugs related to perfect forwarding, and who knows where else. I stopped looking when my crashes got fixed by nothing but changing it to std::function, which is an all-around better replacement.

TOptional only received BP support in 5.4, if you’re sticking with 5.3 due to the licensing changes, it’s pointless.

const locals and parameters

This is not really an Epic thing, but I still see it often, usually because some IDE or productivity tool recommends that you can do this. There’s an academic/theoretical benefit to this: it can prevent accidentally overwriting a local when it was not intended. In practice, this is not really an issue - C# doesn’t have a similar language feature and the bugs that are supposedly not prevented by it mostly don’t exist.

Overusing it adds noise and takes attention away from the real, important consts, as well as directly impacting performance: moving needs T&&, not const T&&, and many const method overloads return copies instead of references.

Of course, this doesn’t apply to pointers or references to const things.

Language or engine features

NULL

For some reason, NULL lives on in some questionable tutorials. You should always use nullptr instead. NULL is the integer constant 0 in C++ and its only purpose is to keep ancient code on life support.

typedef

typedef is considered legacy, the new syntax is using MyAlias = int;. It gets even better for function pointers and it can also be a template!

mutable and FORCEINLINE

Engine code seriously overuses these, they should be very rare. Even if you’re already at the point where you’re genuinely microoptimizing calls, you can outperform FORCEINLINE by moving some parts of a function to a wrapper caller macro.

Other, game-focused libraries, such as the EASTL, use these far less often than Epic.

Legacy … varargs

These are crash-prone, not type-safe, and all-around horrible. They don’t even know what float is. Use variadic templates instead. There’s a subtly-horrible syntactic difference between them, make sure you place the ... in the right place:

void Crashf(T x, ...); // C
void Crashf(T x...); // still C!
void Safef(T... x); // variadic template

GENERATED_…_BODY

GENERATED_UCLASS_BODY, GENERATED_USTRUCT_BODY, etc. are all obsolete. It’s always GENERATED_BODY() for everything. Your code might break if it relied on the legacy macros declaring its constructor, simply declare it to fix it.