Contributing to Arb

The Arb project welcomes new feature contributions in addition to patches for bugs and performance problems.

What are appropriate new features? Most of the numerical functionality that can be found in a general-purpose computer algebra system is certainly within scope (of course, the main restriction is that the algorithm must have a proof of correctness). However, if the functionality is easily accomplished by combining existing functions in Arb, consider whether it is worth the increase in code size, maintenance effort and test time. Prospective contributors are recommended to discuss their ideas on the mailing list or the issue tracker.

The process for actually submitting code is simple: anyone can submit a pull request on GitHub. If the patch looks good to the maintainer and the test code passes, the patch will get merged into the git master.

Code conventions

Four steps are needed to add a new function:

  • Add the function module_foo() in a new file module/foo.c.

  • Add a corresponding test program in a new file module/test/t-foo.c.

  • Add the function prototype to module.h.

  • Document the function in doc/source/module.rst.

The build system takes care of everything else automatically.

Test code (see below) can be omitted if module_foo() is a trivial helper function, but it should at least be tested indirectly via another function in that case. Auxiliary functions needed to implement module_foo() but which have no use elsewhere should be declared as static in module/foo.c. If module_foo() is very short, it can be declared inline directly in module.h with the MODULE_INLINE macro.

Use the following checklist regarding code style:

  • Try to keep names and function arguments consistent with existing code.

  • Follow the conventions regarding types, aliasing rules, etc. described in Technical conventions and potential issues and in code_conventions.txt in FLINT (https://github.com/wbhart/flint2/blob/trunk/code_conventions.txt).

  • Use basic FLINT constants, types and functions: FLINT_BITS, flint_malloc/flint_free, flint_abort, flint_printf, etc.

  • Complex macros should be avoided.

  • Indentation is four spaces.

  • Curly braces normally go on a new line.

  • Binary operators are surrounded by spaces (but parentheses and brackets are not).

  • Logically distinct chunks of code (variable declarations, initialization, precomputations, the main loop, cleanup, etc.) should be separated by a single blank line.

  • Lines are up to 79 characters long, but this rule can be broken if it helps readability.

  • Add correct copyright notices at the top of each file.

Test code

See Setup for instructions on running test code.

The easiest way to write a test program for a new function is to adapt the test code for an existing, similar function.

Most of the test code in Arb uses the strategy of computing the same mathematical quantity in two or more different ways (for example, using functional equations, interchanging the order of parameter, or varying the precision and other algorithm parameters) and verifying that the results are consistent. It is also a good idea to test that aliasing works. Input data is usually generated randomly, but in some cases including precomputed reference values also makes sense.

Faster test code is better. A single test program should not take more than 10 seconds to run, and preferably no more than 1 second. Most functions can be tested effectively in less than 0.1 seconds. Think of what the corner cases are and try to generate random input biased toward such cases. The randtest() functions attempt to generate corner cases automatically, but some thought may be needed to use them optimally. Try to ensure that the test code fails if you deliberately break the tested function in any way. It is also a good idea to run the test code once with ARB_TEST_MULTIPLIER=10.0 or higher. If a function’s input space is too large to probe effectively for corner cases with random input, that can be a hint that the function should be split into smaller logical parts that can be tested separately.

The test code must complete without errors when run with valgrind. The most common mistake leading to memory corruption or memory leaks is to miss or duplicate an init() or clear() call. Check that the init() and clear() calls exactly match the variable declarations in each code block, including the test code itself.

Profiling code is not needed in most cases, but it is often a good idea to run some benchmarks at least during the initial development of a new feature. The TIMEIT_START/TIMEIT_STOP and SHOW_MEMORY_USAGE macros in FLINT are useful for quick measurements.