rat128

Fixed-precision rational numbers for Go. This module has no dependencies and is fairly straightforward to use:

Construct new values with var x rat128.N, x := rat128.New(m, n), or x, err := rat128.Try(m, n).
Retrieve the numerator and denominator with x.Num() and x.Den(), respectively.
Perform arithmetic with panicking x.Add(y), x.Mul(y) or error-returning x.TryAdd(y), x.TryMul(y), etc.
Convert from/to float64 with FromFloat64 and x.Float64() respectively.
Convert from/to decimal strings ("12.34") with ParseDecimalString and x.DecimalString(digits) respectively.

Design Goals

First and foremost: value semantics. Unlike big.Rat in the standard library, rat128.N is a value type like int64 or float32 and is safe to copy and compare with == and !=. Methods return values instead of mutating their receiver.
Mathematical operations should return correct results. If they don't, the incorrect behavior is a bug and fixing this bug will be treated as a non-breaking change. Do not rely on incorrect results.
Widened integer arithmetic is used where possible to avoid overflow of intermediate values used in basic arithmetic operations (add, subtract, multiply, and divide). The finite precision is more limiting than big.Rat, however, and panics from overflow of the numerator or denominator in the final result are possible.
Valid values are always in reduced form, to simplify operations and reduce the risk of overflow.
Converting to and from floating-point values should be exact wherever it is possible to do so, and approximation should be explicit.
There are panicking and panic-free versions of most operations. This means there are actually three ways to use this library:
- Use the panicking operations (New, Add, etc.)
- Use the panic-free operations (Try, TryAdd, etc.) and check for errors
- Use the panic-free operations and ignore the errors; this is not recommended but will be fastest

File bug reports, feature requests, etc. through GitHub Issues on this repository.

The following are always considered issues:

Incorrect mathematical results
Violation of the design goals; though fixing these may require breaking changes and thus a new major version
Unexpected behavior; for example, operations on otherwise valid values returning invalid results instead of valid results, a panic, or an error
Panics in Try* functions/methods; these return errors, so ordinary arithmetic exceptions should not cause panics

The following will be evaluated on a case-by-case basis:

Optimization; in particular, micro-optimizations that reduce readability of the code for little measurable benefit may be rejected
(Un-)marshaling/(de-)serializing in various formats
Undefined behavior; for example, the results of operations on invalid values created through unsafe code, reflection, unsychronized sharing across goroutines, cgo hacks, memory corruption, etc.
Converting to formats outside of the core language and standard library
Performance regressions; in general, this library should outperform big.Rat on an apples-to-apples basis on 64-bit machines, but the exact performance characteristics are not guaranteed