Allowed language features#
Most of these are not strict rules, but you should have a very good reason for deviating from them.
Most GROMACS files compile as C++17, but some files remain that compile as C99. C++ has a lot of features, but to keep the source code maintainable and easy to read, we will avoid using some of them in GROMACS code. The basic principle is to keep things as simple as possible.
MSVC supports only a subset of C99 and work-arounds are required in those cases.
We should be able to use virtually all C++17 features outside of OpenCL kernels (which compile as C), and for consistency also in CUDA kernels.
C++ Standard Library#
GROMACS code must support the lowest common denominator of C++17 standard library
features available on supported platforms.
Some modern features are useful enough to warrant back-porting.
Consistent and forward-compatible headers are provided in
as described in the Library documentation
As a baseline, GROMACS follows the C++ Core Guidelines c++ guidelines, unless our own more specific guidelines below say otherwise. We tend to be more restrictive in some areas, both because we depend on the code compiling with a lot of different C++ compilers, and because we want to increase readability. However, GROMACS is an advanced projects in constant development, and as our needs evolve we will both relax and tighten many of these points. Some of these changes happen naturally as part of agreements in code review, while major parts where we don’t agree should be pushed to a issue tracker thread. Large changes should be suggested early in the development cycle for each release so we avoid being hit by last-minute compiler bugs just before a release.
Use namespaces. Everything in
libgromacsshould be in a
gmxnamespace. Don’t use using in headers except possibly for aliasing some commonly-used names, and avoid file-level blanket
using namespace gmxand similar. If only a small number of
gmxnamespace symbols needed in a not-yet-updated file, consider importing just those symbols. See also here.
Use STL, but do not use iostreams outside of the unit tests. iostreams can have a negative impact on performance compared to other forms of string streams, depending on the use case. Also, they don’t always play well with using C
stdioroutines at the same time, which are used extensively in the current code. However, since Google tests rely on iostreams, you should use it in the unit test code.
Don’t use non-const references as function parameters. They make it impossible to tell whether a variable passed as a parameter may change as a result of a function call without looking up the prototype.
string_viewin cases where you want to only use a read-only-sequence of characters instead of using
const std::string &. See also here. Because null termination expected by some C APIs (e.g. fopen, fputs, fprintf) is not guaranteed, string_view should not be used in such cases.
optional<T>types in situations where there is exactly one, reason (that is clear to all parties) for having no value of type T, and where the lack of value is as natural as having any regular value of T, see here. Good examples include the return type of a function that parses an integer value from a string, searching for a matching element in a range, or providing an optional name for a residue type. Do use optional for lazy loading of resources, e.g., objects that have no default constructor and are hard to construct. Prefer other constructs when the logic requires an explanation of the reason why no regular value for T exists, e.g., do not use
optional<T>for error handling.
optional<T>“models an object, not a pointer, even though operator*() and operator->() are defined” (cppreference). No dynamic memory allocation ever takes place and forward declaration of objects stored in
optional<T>does not work. Thus refrain from optional when passing handles; in contrast to unique_ptr, optional has value semantics, not reference semantics.
Don’t use C-style casts; use
reinterpret_castas appropriate. See the point on RTTI for
dynamic_cast. For emphasizing type (e.g. intentional integer division) use constructor syntax. For creating real constants use the user-defined literal
Use signed integers for arithmetic (including loop indices). Use
ssize(available as free function and member of
ArrayRef) to avoid casting.
Avoid overloading functions unless all variants really do the same thing, just with different types. Instead, consider making the function names more descriptive.
Avoid using default function arguments. They can lead to the code being less readable than without (see here). If you think that your specific case improves readability (see here), you can justify their use.
Don’t overload operators before thorough consideration whether it really is the best thing to do. Never overload
||, or the comma operator, because it’s impossible to keep their original behavior with respect to evaluation order.
Try to avoid complex templates, complex template specialization or techniques like SFINAE as much as possible. If nothing else, they can make the code more difficult to understand.
Don’t use multiple inheritance. Inheriting from multiple pure interfaces is OK, as long as at most one base class (which should be the first base class) has any code. Please also refer to the explanation here and here.
Don’t write excessively deep inheritance graphs. Try to not inherit implementation just to save a bit of coding; follow the principle “inherit to be reused, not to reuse.” Also, you should not mix implementation and interface inheritance. For explanation please see here.
Don’t include unnecessary headers. In header files, prefer to forward declare the names of types used only “in name” in the header file. This reduces compilation coupling and thus time. If a source file also only uses the type by name (e.g. passing a pointer received from the caller to a callee), then no include statements are needed!
Make liberal use of assertions to help document your intentions (but prefer to write the code such that no assertion is necessary).
assert()because the former permit you to add descriptive text.
Use gmx::Mutex rather than pthreads, std or raw thread-MPI mutexes.
Use proper enums for variable whose type can only contain one of a limited set of values. C++ is much better than C in catching errors in such code. Ideally, all enums should be typed enums, please see here.
When writing a new class, think whether it will be necessary to make copies of that class. If not, declare the copy constructor and the assignment operator as private and don’t define them, making any attempt to copy objects of that class fail. If you allow copies, either provide the copy constructor and the assignment operator, or write a clear comment that the compiler-generated ones will do (and make sure that they do what you want).
src/gromacs/utility/classhelpers.hhas some convenience macros for doing this well. You can also use deleted functions in this case.
Declare all constructors with one parameter as explicit unless you really know what you are doing. Otherwise, they can be used for implicit type conversions, which can make the code difficult to understand, or even hide bugs that would be otherwise reported by the compiler. For the same reason, don’t declare operators for converting your classes to other types without thorough consideration. For an explanation, please see here.
Write const-correct code (no
const_castunless absolutely necessary).
Avoid using RTTI (run-time type information, in practice
typeid) unless you really need it. The cost of RTTI is very high, both in binary size (which you always pay if you compile with it) and in execution time (which you pay only if you use it). If your problem seems to require RTTI, think about whether there would be an alternative design that wouldn’t. Such alternative designs are often better.
Don’t depend on compiler metadata propagation. struct elements and captured lambda parameters tend to have
restrictand alignment qualifiers discarded by compilers, so when you later define an instance of that structure or allocate memory to hold it, the data member might not be aligned at all.
Plan for code that runs in compute-sensitive kernels to have useful data layout for re-use, alignment for SIMD memory operations
Recognize that some parts of the code have different requirements - compute kernels, mdrun setup code, high-level MD-loop code, simulation setup tools, and analysis tools have different needs, and the trade-off point between correctness vs reviewer time vs developer time vs compile time vs run time will differ.
Implementing exceptions for error handling#
See Error handling for the approach to handling run-time errors, ie. use exceptions.
Write exception-safe code. All new code has to offer at least the basic or nothrow guarantee to make this feasible.
Use std (or custom) containers wherever possible.
Use smart pointers for memory management. By default, use
gmx::unique_cptrin assocation with any necessary raw
std::shared_ptrcan be used wherever responsibility for lifetime must be shared. Never use
Use RAII for managing resources (memory, mutexes, file handles, …).
It is preferable to avoid calling a function which might throw an exception from a legacy function which is not exception safe. However, we make the practical exception to permit the use of features such as
std::stringthat could throw
std::bad_allocwhen out of memory. In particular, GROMACS has a lot of old C-style memory handling that checking tools continue to issue valid warnings about as the tools acquire more functionality, and fixing these with old constructs is an inefficient use of developer time.
Functions / methods should be commented whether they are exception safe, whether they might throw an exception (even indirectly), and if so, which exception(s) they might throw.
Don’t use preprocessor defines for things other than directly related to configuring the build. Use templates or inline functions to generate code, and enums or const variables for constants.
Preprocessing variables used for configuring the build should be organized so that a valid value is always defined, i.e. we never test whether one of our preprocessor variables is defined, rather we test what value it has. This is much more robust under maintenance, because a compiler can tell you that the variable is undefined.
Avoid code with lengthy segments whose compilation depends on #if (or worse, #ifdef of symbols provided from outside GROMACS).
Prefer to organize the definition of a const variable at the top of the source code file, and use that in the code. This helps keep all compilation paths built in all configurations, which reduces the incidence of silent bugs.
Indent nested preprocessor conditions if nesting is necessary and the result looks clearer than without indenting.
Please strongly consider a comment repeating the preprocessor condition at the end of the region, if a lengthy region is necessary and benefits from that. For long regions this greatly helps in understanding and debugging the code.