Debugging OSCAR Code

Pitfalls: Mutable objects in OSCAR code

Suppose you are having the following difficulties. Your code is exhibiting inexplicable behaviour and values that should not be changing are changing in seemingly random locations. To get to the bottom of these kind of issues it is necessary to be familiar with mutable objects in Julia and some of the relevant conventions in place in OSCAR. This section discusses these informal rules as well as some of the exceptions to these rules.

In Julia, objects that can change after construction are declared with the mutable struct keywords and satisfy the ismutable predicate. These objects can be linked together into an arbitrary dependency graph, and a change to one object may therefore have unintended consequences on another object in the system.

The simplest example is the creation of a polynomial ring. If we mutate the array of symbols used for printing, we have effectively changed the ring.

julia> v = [:x, :y, :z]; R = PolynomialRing(QQ, v)[1]
Multivariate Polynomial Ring in x, y, z over Rational Field

julia> v[3] = :w; R
Multivariate Polynomial Ring in x, y, w over Rational Field

In this example, the modification of v is unexpected and may in fact corrupt the internal data structures used by the polynomial ring. As such, this modification of v has to be considered illegal. Upon creation of the array called v, we have full rights over the object and can mutate at will. However, after passing it to the function PolynomialRing, we have given up ownership of the array and are no longer free to modify it.

General OSCAR Principle (GOP):

Code should be expected to behave as if all objects are immutable.

Ramifications:

  1. This means that the polynomial ring constructor is allowed to expect that v is never mutated for the remaining duration of its life. In return, the constructor is guaranteed not to modify the array, so that v is still [:x, :y, :z] after PolynomialRing returns.
  2. In general this means that all functions should be expected to take ownership of their arguments: the user is safest never modifying an existing object that has been passed to an unknown Julia function. Note that assignments such as a[i] = b or a.foo = b usually mutate the object a. See Ownership of function arguments
  3. For reasons of efficiency, it is sometimes desirable to defy this principle and modify an existing object. The fact that a given function may modify a preexisting object is usually communicated via coding conventions on the name - either a ! or a _unsafe in the name of the function. See Unsafe arithmetic with OSCAR objects

Ownership of function arguments

In this example we construct the factored element x = 2^3 and then change the 2 to a 1. The GOP says this modification of a on line 3 is illegal.

julia> a = fmpz(2)
2

julia> x = FacElem([a], [fmpz(3)]); evaluate(x)
8

julia> a = one!(a)  # illegal in-place assignment of a to 1
1

julia> evaluate(x)  # x has been changed and possibly corrupted
1

In the previous example, the link between the object x and the object a can be broken by passing a deepcopy of a to the FacElem function.

julia> a = fmpz(2)
2

julia> x = FacElem([deepcopy(a)], [fmpz(3)]); evaluate(x)
8

julia> a = one!(a)  # we still own a, so modification is legal
1

julia> evaluate(x)  # x is now unchanged
8

It is of course not true that all Julia functions take ownership of their arguments, but the GOP derives from the fact that this decision is an implementation detail with performance consequences. The behaviour of a function may be inconsistent across different types and versions of OSCAR. In the following two snippets, the GOP says both modifications of a are illegal since they have since been passed to a function. If K = QQ, the two mutations turn out to be legal currently, while they are illegal if K = quadratic_field(-1)[1]. Only with special knowledge of the types can the GOP be safely ignored.

R = PolynomialRing(K, [:x, :y])[1]
a = one(K)
p = R([a], [[0,0]])
@show p
a = add!(a, a, a)       # legal? (does a += a in-place)
@show p
R = PolynomialRing(K, :x)[1]
a = [one(K), one(K)]
p = R(a)
@show (p, degree(p))
a[2] = zero(K)          # legal?
@show (p, degree(p))

Ownership of function return values

The nuances of who is allowed to modify an object returned by a function is best left to the next section Unsafe arithmetic with OSCAR objects. The GOP says of course you should not do it, but there are cases where it can be more efficient. However, there is another completely different issue of return values that can arise in certain interfaces.

First, we create the Gaussian rationals and the two primes above 5.

julia> K, i = quadratic_field(-1)
(Imaginary quadratic field defined by x^2 + 1, sqrt(-1))

julia> m = Hecke.modular_init(K, 5)
modular environment for p=5, using 2 ideals

The function modular_project returns the projection of an element of K into each of the residue fields.

julia> a = Hecke.modular_proj(1+2*i, m)
2-element Vector{fq_nmod}:
 2
 0

While the function has produced the correct answer, if we run it again on a different input, we will find that a has changed.

julia> b = Hecke.modular_proj(2+3*i, m)
2-element Vector{fq_nmod}:
 1
 3

julia> a
2-element Vector{fq_nmod}:
 1
 3

The preceding behaviour of the function modular_proj is an artifact of internal efficiency and may be desirable in certain circumstances. In other circumstances, the following deepcopys may be necessary for your code to function correctly.

julia> a = deepcopy(Hecke.modular_proj(1+2*i, m));
julia> b = deepcopy(Hecke.modular_proj(2+3*i, m));
julia> (a, b)
(fq_nmod[2, 0], fq_nmod[1, 3])

Unsafe arithmetic with OSCAR objects

Particularly with integers (BigInt and fmpz) - but also to a lesser extent with polynomials - the cost of basic arithmetic operations can easily be dominated by the cost of allocating space for the answer. For this reason, OSCAR offers an interface for in-place arithmetic operations.

Instead of writing x = a + b to compute a sum, one writes x = add!(x, a, b) with the idea that the object to which x is pointing is modified instead of having x point to a newly allocated object. In order for this to work, x must point to a fully independent object, that is, an object whose modification through the interface Unsafe operators will not change the values of other existing objects. The actual definition of "fully independent" is left to the implementation of the ring element type. For example, there is no distinction for immutables.

It is generally not safe to mutate the return of a function. However, the basic arithmetic operations +, -, *, and ^ are guaranteed to return a fully independent object regardless of the status of their inputs. As such, the following implementation of ^ is illegal by this guarantee.

function ^(a::RingElem, n::Int)
  if n == 1
    return a    # must be return deepcopy(a)
  else
    ...
  end
end

In general, if you are not sure if your object is fully independent, a deepcopy should always do the job.