Describe the OpenMP 2.0 constraints that follow from building with MSVC /openmp, and why an orphaned work sharing construct can deadlock the application when the master thread is also the Qt GUI thread. Also document the pattern of collecting results in per thread buffers and merging after the parallel region, the convention of naming critical sections, and the rule that exceptions must not escape a structured block.
9.6 KiB
Coding Style Guidelines
This document describes the coding style and formatting conventions for ResInsight.
Code Formatting
clang-format
ResInsight uses clang-format for C++ code formatting:
- Configuration:
.clang-formatfile in repository root - Version: Use clang-format-19 to enforce style
Python Formatting
Python code should be formatted using ruff:
# Format source code
python -m ruff format test_polygons.py
# Check code style
python -m ruff check --fix test_polygons.py
Copyright Headers
- New files must use the current year (the year the file is created) in the copyright header
- Never change the copyright year in existing files — the year reflects when the file was originally created
- Example header for a file created in 2026:
/////////////////////////////////////////////////////////////////////////////////
//
// Copyright (C) 2026- Equinor ASA
//
Language Standards
- C++: C++23 standard
- Python: Python 3.11+
Best Practices
General Guidelines
- Minimal Changes: Make the smallest possible changes to achieve the goal
- Preserve Formatting: Do not reformat unrelated code
- Comments: Match the style of existing comments in the file
- Libraries: Use existing libraries whenever possible; only add new libraries or update versions if absolutely necessary
Code Quality
- Always validate changes don't break existing behavior
- Always validate changes don't introduce security vulnerabilities
- Fix any vulnerabilities related to your changes
- Run linters, builds and tests before making code changes to understand any existing issues
- Always try to lint, build and test code changes as soon as possible after making them
Testing
- Only run linters, builds and tests that already exist
- Do not add new linting, building or testing tools unless necessary to fix the issue
- It is unacceptable to remove or edit unrelated tests
- Documentation changes do not need to be linted, built or tested unless there are specific tests for documentation
Header / Implementation Split (C++)
Put function bodies in the .cpp file, not inline in the header.
- Declare member functions in the header; define them in the matching
.cpp. - This applies even to one-line bodies (
return false;,return m_field;, trivial forwarders). - Exceptions: function templates that must stay in the header, and
constexprfunctions where the compiler requires the definition to be visible.
// Bad – inline body in the header
class RimFoo
{
public:
bool canAddSubCollection() const override { return false; }
};
// Good – declared in the header, defined in the .cpp
// RimFoo.h
class RimFoo
{
public:
bool canAddSubCollection() const override;
};
// RimFoo.cpp
bool RimFoo::canAddSubCollection() const
{
return false;
}
Why: keeps headers light (faster builds, smaller include surface), keeps the implementation file as the single source of truth for behavior, and matches the existing style across ApplicationLibCode.
Logging
Use std::format to build log messages — not QString(...).arg(...).
RiaLogging::info/warning/errortakestd::string_view. Passing aQStringrequires a trailing.toStdString();std::formatavoids that and reads better.QStringarguments can be passed directly tostd::format— includeRiaQStringFormatter.h, which provides astd::formatter<QString>specialization.
// Bad – QString with positional placeholders, manual conversion
RiaLogging::info( QString( "Exported %1 cells for well '%2'." ).arg( count ).arg( wellName ).toStdString() );
// Good – std::format, QString passed directly
RiaLogging::info( std::format( "Exported {} cells for well '{}'.", count, wellName ) );
Asserts
Use CAF_ASSERT (cafAssert.h). Do not add new CVF_ASSERT, CVF_ASSERT_MSG, CVF_FAIL_MSG or CVF_TIGHT_ASSERT — those are legacy and are being migrated away from.
An assert documents an invariant that holds whenever the program is correct. It is not error handling. Conditions that can legitimately occur at run time — a missing file, a malformed input deck, a failed network call, a user selecting nothing — must be handled with a return value, std::optional, std::expected or a logged error message, never with an assert.
// Bad – a missing file is a run-time condition, not a broken invariant
CAF_ASSERT( QFile::exists( fileName ) );
// Good – handled and reported
if ( !QFile::exists( fileName ) )
{
RiaLogging::error( std::format( "File not found: {}", fileName ) );
return {};
}
// Good – a broken invariant, unreachable unless the code is wrong
CAF_ASSERT( index >= 0 && index < m_values.size() );
CAF_ASSERT follows the semantics of the standard assert():
- Debug builds: active. A failure prints file, line and expression, then calls
std::abort(). - Optimized builds (Release, RelWithDebInfo): compiled out. The expression is still type checked but never evaluated, so it must be free of side effects — never put work inside the assert that the surrounding code depends on.
Configure with -DRESINSIGHT_ENABLE_ASSERTS_IN_RELEASE=ON to keep the asserts active in an optimized build, which is useful when reproducing a problem in RelWithDebInfo with a debugger attached.
A failing assert aborts via SIGABRT, which the crash handler installed in RiaMain.cpp picks up: the failure is written to the log file and reported to OpenTelemetry with a stack trace.
Lambda Functions
Keep lambdas short and readable:
- 1–3 lines: keep inline
- 4+ lines or contains control flow (if/for/while): extract to a named method (e.g.
onApplyClicked()) and call it from a single-line lambda
// Good – short inline lambda
addNewButton( "Show Report", [this]() { showReport(); } );
// Good – multi-statement, no control flow, still inline
addNewButton( "Clear Data",
[this]()
{
clearData();
updateConnectedEditors();
} );
// Good – complex logic extracted to a named method
addNewButton( "Apply", [this]() { onApplyClicked(); } );
void MyClass::onApplyClicked() { /* complex logic here */ }
Context Menus
When adding or modifying context-menu items for a PDM object, override appendMenuItems( caf::CmdFeatureMenuBuilder& menuBuilder ) on that object and add the entries there.
- Do not modify
RimContextCommandBuilderto add object-specific menu items. It dispatches to each object'sappendMenuItems; keep object-specific logic in the object.
// Good – menu items added in the object's own appendMenuItems override
void RimFoo::appendMenuItems( caf::CmdFeatureMenuBuilder& menuBuilder ) const
{
menuBuilder << "RicSomeFeature";
}
OpenMP
ResInsight builds with MSVC /openmp, which is OpenMP 2.0. Loop indices must be a signed
integer type, and collapse and min/max reductions are not available.
Never write an orphaned work-sharing construct
A #pragma omp for must have a lexically enclosing #pragma omp parallel in the same function.
An orphaned #pragma omp for is not merely useless - it can deadlock the application.
The reason is that the OpenMP master thread is also the Qt GUI thread. When the GUI thread is
inside a parallel region and a progress update or a repaint runs, Qt work executes directly on
that thread, which pulls rendering code into the dynamic extent of an active parallel team. An
orphaned #pragma omp for reached that way is encountered by one thread of a team whose other
threads never reach it, and the implicit barrier at the end of the construct never completes.
// Bad - deadlocks if this function is reached from inside an active parallel region
void toTextureImageRegion( ... )
{
#pragma omp for
for ( int y = 0; y < sizeY; ++y )
Collect per thread, merge afterwards
Do not guard a push_back into a shared container with #pragma omp critical inside a hot loop.
Give each thread its own buffer and merge once the region has ended. Merging in thread order also
makes the result reproducible from run to run, because the default static schedule assigns
contiguous chunks to the threads.
// Good
const int numberOfThreads = RiaOpenMPTools::availableThreadCount();
std::vector<std::vector<size_t>> threadCells( numberOfThreads );
#pragma omp parallel
{
const int myThread = RiaOpenMPTools::currentThreadIndex();
// NB! We are inside a parallel section, do not use "parallel for" here
#pragma omp for
for ( int i = 0; i < cellCount; i++ )
{
if ( isIncluded( i ) ) threadCells[myThread].push_back( i );
}
}
for ( const auto& cellsForThread : threadCells )
{
cells.insert( cells.end(), cellsForThread.begin(), cellsForThread.end() );
}
No synchronization is needed when each iteration writes to its own element of a container that was
sized before the loop. Do not add critical for that case.
Name every critical section
All unnamed #pragma omp critical regions share one process-wide lock, so unrelated subsystems
serialize against each other. Use the critical_section_<purpose> convention instead.
Exceptions and the GUI
An exception that escapes an OpenMP structured block terminates the process on MSVC instead of
unwinding, so catch inside the loop body. Touch Qt widgets only from the thread owning them; use
RiaThreadSafeLogger or a queued connection to hand data back.
Commit Conventions
When creating commits:
- Use issue number at the start of the title:
#12773 Python: Add API for creating valve templates - Follow git conventions for commit messages
- Always run python formatting/check on changed files before commits