Skip to content

Latest commit

 

History

History
43 lines (34 loc) · 5.37 KB

CONTRIBUTING.md

File metadata and controls

43 lines (34 loc) · 5.37 KB

Contributing

Thank you for your interest in contributing to snarkOS! Below you can find some guidelines that the project strives to follow.

Pull requests

Please follow the instructions below when filing pull requests:

  • ensure that your branch is forked from the current master branch
  • run cargo fmt before you commit; we use the nightly version of rustfmt to format the code, so you'll need to have the nightly toolchain installed on your machine; there's a git hook that ensures proper formatting before any commits can be made, and .rustfmt.toml specifies some of the formatting conventions
  • run cargo clippy --all-targets --all-features to ensure that popular correctness and performance pitfalls are avoided

Coding conventions

snarkOS is a big project, so (non-)adherence to best practices related to performance can have a considerable impact; below are the rules we try to follow at all times in order to ensure high quality of the code:

Memory handling

  • if the final size is known, pre-allocate the collections (Vec, HashMap etc.) using with_capacity or reserve - this ensures that there are both fewer allocations (which involve system calls) and that the final allocated capacity is as close to the required size as possible
  • create the collections right before they are populated/used, as opposed to e.g. creating a few big ones at the beginning of a function and only using them later on; this reduces the amount of time they occupy memory
  • if an intermediate vector is avoidable, use an Iterator instead; most of the time this just amounts to omitting the call to .collect() if a single-pass iteraton follows afterwards, or returning an impl Iterator<Item = T> from a function when the caller only needs to iterate over that result once
  • when possible, fill/resize collections "in bulk" instead of pushing a single element in a loop; this is usually (but not always) detected by clippy, suggesting to create vectors containing a repeated value with vec![x; N] or extending them with .resize(N, x)
  • when a value is to eventually be consumed in a chain of function calls, pass it by value instead of by reference; this has the following benefits:
    • it makes the fact that the value is needed by value clear to the caller, who can then potentially reclaim it from the object afterwards if it is "heavy", limiting allocations
    • it often enables the value to be cloned fewer times (whenever it's no longer needed at the callsite)
    • when the value is consumed and is not needed afterwards, the memory it occupies is freed, improving memory utilization
  • if a slice may or may not be extended (which requires a promotion to a vector) and does not need to be consumed afterwards, consider using a Cow<'a, [T]> combined with Cow::to_mut instead to potentially avoid an extra allocation; an example in snarkOS could be conditional padding of bits
  • prefer arrays and temporary slices to vectors where possible; arrays are often a good choice if their final size is known in advance and isn't too great (as they are stack-bound), and a small temporary slice &[x, y, z] is preferable to a vec![x, y, z] if it's applicable
  • if a reference is sufficient, don't use .clone()/to_vec(), which is often the case with methods on structs that provide access to their contents; if they only need to be referenced, there's no need for the extra allocation
  • use into_iter() instead of iter().cloned() where possible, i.e. whenever the values being iterated over can be consumed altogether
  • if possible, reuse collections; an example would be a loop that needs a clean vector on each iteration: instead of creating and allocating it over and over, create it before the loop and use .clear() on every iteration instead
  • try to keep the sizes of enum variants uniform; use Box<T> on ones that are large

Misc. performance

  • avoid the format!() macro; if it is used only to convert a single value to a String, use .to_string() instead, which is also available to all the implementors of Display
  • don't check if an element belongs to a map (using contains or get) if you want to conditionally insert it too, as the return value of insert already indicates whether the value was present or not; use that or the Entry API instead
  • if a reference is sufficient as a function parameter, use:
    • &[T] instead of &Vec<T>
    • &str instead of &String
    • &Path instead of &PathBuf
  • if a lot of computational power is needed, consider parallelizing that workload with rayon - it's not always a viable solution, but can yield great performance improvements when used in the right context
  • for structs that can be compared/discerned based on some specific field(s), consider hand-written implementations of PartialEq and Hash (they must match) for faster comparison and hashing
  • if possible, ensure that the results of your changes are not detrimental to performance using criterion (for smaller, fine-grained adjustments) and valgrind --tool={cachegrind | massif} (for larger-scale changes)