C# 12: Primary constructors.

[michaelnebel]

In this PR we introduce library and data flow support for primary constructors.
Below we see an example of class C1 with a primary constructor defined on line 1.

public class C1(object o, string s)
{
    public object Obj => o;
    public string Str => s;

    public C1(object o) : this(o, "default") { }

    public void SetStr(string s1)
    {
        s = s1;
    }
}

The QL library now contains a class called PrimaryConstructor which contains exactly the primary constructors defined in source code. If it turns our that this implementation is insufficient we can extend the extractor to specifically extract whether a constructor is a primary constructor by inspecting the declaring syntax references using the the same "hack" as roslyn.

Before discussing the details of the data flow implementation for primary constructors is is worth noting that the example class declaration of C1 is roughly equivalent to (this is also what you will approximately get if you compile and de-compile the code):

public class C1
{
    private object vObj;
    public object Obj => vObj;

    private string vStr;
    public string Str => vStr;

    public C1(object o, string s)
    {
        vObj = o;
        vStr = s;
    }

    public C1(object o) : this(o, "default") { }

    public void SetStr(string s1)
    {
        vStr = s1;
    }
}

That is, we can think of class with a primary constructor as having a backing field for each primary constructor parameter and that the primary constructor as having a (synthetic) body that assigns the provided arguments to the backing fields.
Furthermore, each primary constructor parameter access (read/write) corresponds to reading or writing the corresponding backing field.
We use these observations to model data flow for primary constructors by introducing the following new content and dataflow node types

Content

• PrimaryConstructorParameterContent: As we don't have a "real" backing field for each primary constructor parameter we introduce a new content type that enables us to create "synthetic" backing fields for each primary constructor parameter.

The data-flow implementation introduces two new types of data-flow nodes:

• PrimaryConstructorThisAccessNode: Used for this access and synthesises the body of the primary constructor (that is, dataflow from the primary constructor parameters into their synthetic backing fields).
• InstanceParameterAccessNode: Used to synthesise reads and writes of the parameter synthetic backing fields. That is, for all primary constructor parameter accesses (reads and writes) within a class declaration we need to do the corresponding update to the synthetic backing fields.

The dataflow implementation has some limitations and doesn't cover e.g.

• Subtyping.

Huge thanks to @hvitved for outlining and helping with the implementation.

[michaelnebel]

DCA was uneventful; Even though it is a bit discouraging that we didn't find more results, doing improvements like this should help us avoid our dataflow analysis to deteriorate over time when the new language feature becomes more widely adopted.

[hvitved]

Overall LGTM, great work! Some (mostly cosmetic) suggestions.

[hvitved]

I think we should just inline this into the charpred of PrimaryConstructor.

[hvitved]

Alternatively

Suggested change

	pa instanceof ParameterRead and
	result = pa.getAControlFlowNode()
	result = pa.(ParameterRead).getAControlFlowNode()

[hvitved]

Rename to getAPrimaryConstructorParameterCfn.

[hvitved]

Gets a

[hvitved]

This only takes direct assignments into account, and e.g. not assignments via out arguments (perhaps we should add a test for that). AssignableDefinitions are meant to cover all cases:

Suggested change

	pa instanceof ParameterWrite and
	exists(AssignExpr ae | pa = ae.getLValue() and result = ae.getAControlFlowNode())
	pa = any(AssignableDefinition def | result = def.getAControlFlowNode()).getTargetAccess()

[hvitved]

Suggested change

	exists(InstanceParameterAccessNode n | n = node1 |
	n.getUnderlyingControlFlowNode() = node2.(ExprNode).getControlFlowNode() and
	n.getParameter() = c.(PrimaryConstructorParameterContent).getParameter()
	) and
	node1 = any(InstanceParameterAccessNode n |
	n.getUnderlyingControlFlowNode() = node2.(ExprNode).getControlFlowNode() and
	n.getParameter() = c.(PrimaryConstructorParameterContent).getParameter() and
	n.isPreUpdate()
	) and

[hvitved]

Suggested change

	exists(Node n1 |
	primaryConstructorParameterStore(_, c, n1) and n = n1.(PostUpdateNode).getPreUpdateNode()
	)
	n = any(PostUpdateNode n1 | primaryConstructorParameterStore(_, c, n1)).getPreUpdateNode()

[hvitved]

cfn

[hvitved]

I would just say result = "parameter " + p.getName()

[hvitved]

No need to add this.

[hvitved]

Missing backticks.

[hvitved]

Missing backticks.

[hvitved]

Missing backticks.

[hvitved]

Missing backticks.

[michaelnebel]

@hvitved : Your changes look plausible to me.

[michaelnebel]

DCA still looks good.