Allow AnalyzeDataFlow for ConstructorInitializerSyntax

[bernd5]

Background and Motivation

For analyzers and rewriters it is sometimes required to analyze the data flow of a constructor initializer which is very similar to an implicit first expression statement.
Unfortenately it is not an expression nor a statement and the public API allows only to analyze statements or expressions (specified in the comment).

Proposed API

The current public API accepts all SyntaxNode but throws expections if the node is not a statement nor an expression.
Therefore we don't need to change the public API. We should just change the comment.

using System.Collections.Immutable;
using System.Threading;

namespace Microsoft.CodeAnalysis
{
    public static class ModelExtensions
    {
        ...

        /// <summary>
        /// Analyze data-flow within a part of a method body.
+       /// note (for C#): ConstructorInitializerSyntax and PrimaryConstructorBaseTypeSyntax are treated by this API as regular statements
        /// </summary>
        public static DataFlowAnalysis AnalyzeDataFlow(this SemanticModel semanticModel, SyntaxNode statementOrExpression)
        {
            //impl
        }
    }
}

Optional API extension:

To make the support for ConstructorInitializerSyntax and PrimaryConstructorBaseTypeSyntax more obvious we could add the folowing extension methods:

namespace Microsoft.CodeAnalysis
{
    public static class CSharpExtensions
    {
+        /// <summary>
+        /// Analyze data-flow within a constructor initializer.
+        /// </summary>
+        public static DataFlowAnalysis? AnalyzeDataFlow(this SemanticModel? semanticModel, ConstructorInitializerSyntax constructorInitializer)
+        {
+            //impl
+        }
+
+        /// <summary>
+        /// Analyze data-flow within a primary constructor base type.
+        /// </summary>
+        public static DataFlowAnalysis? AnalyzeDataFlow(this SemanticModel? semanticModel, PrimaryConstructorBaseTypeSyntax primaryConstructorBaseType)
+        {
+            //impl
+        }
    }
}

Usage Examples

analysis = model.AnalyzeDataFlow(ctorInit);

Alternative Designs

Some user hacked around this missing feature and created a corresponding object creation expression for the analysis which produces a quite good result. The downside is just the mapping of symbols between two different models - which can become quite complicated and error prone.

Risks

No known risk

Implementation

#57576

[bernd5]

@AlekseyTs @333fred: could you have a look?

[bernd5]

@jaredpar

[jaredpar]

@333fred PTAL

@bernd5 could you please consider not pinging the issue multiple times in the same 24 hour period? Generally engineers look a couple of times a day and I generally triage all incoming issue once a day (although I do occasionally miss a day here and there). Multiple pings in the same day are unnecessary. If we have not responded after a couple of days then yes please do ping because we may have missed it.

[333fred]

There's no actual API change here, so I don't think there's really anything to look at from the public API side. I also don't see any problem with allowing the DataFlowAnalysis APIs to expose this information, so I think we should just review the linked PR.

[333fred]

@AlekseyTs pointed out that we do have a bit of public API for C#-specific syntax nodes here: https://sourceroslyn.io/#Microsoft.CodeAnalysis.CSharp/Compilation/SyntaxTreeSemanticModel.cs,2250. So we'll want to add a new overload that takes a constructor initializer syntax node there.

[bernd5]

In my PR I added a specific overload in CSharpSemanticModel and in SyntaxTreeSemanticModel similar how it is implemented for statements and expressions (and it works quite great).
Both types are internal so it is not public API.

[333fred]

@bernd5 we still need to update the proposal to include extension overloads for https://sourceroslyn.io/#Microsoft.CodeAnalysis.CSharp/CSharpExtensions.cs,1038. We also probably want to do the same thing for the AnalyzeControlFlow API, to be consistent across the board.

[bernd5]

@333fred is this API-issue now ready for review?

[AlekseyTs]

I think we would want to add support for initializers of primary constructors as well.

[AlekseyTs]

I know "primary constructors" only from records - but how can we define initializers for them?

Arguments for the base constructor invocation can be supplied on the base class. For example:

record Base(int x);
record Derived(int x, int y) : Base(x);

[bernd5]

Oh, yeah I forgot that - I'll add it, too.

[bernd5]

@AlekseyTs @333fred have you seen, that I updated the API proposal / PR?
Do you know when the next API review meeting is scheduled?

[333fred]

API review meets every two weeks. The next one is scheduled for next Thursday.

[333fred]

API Review

• We need to make sure that we add tests for Extract Method to make sure it doesn't have negative behavior introduced with this method.
• API shape is approved with the optional C#-specific overloads. The primary constructor overload needs to have its comments adjust to not imply that it's giving info about the base constructor itself, only the base call syntax.

[bernd5]

I would change the comment to: "Analyze data-flow within a primary-constructor-base-type initializer.", okay?

[bernd5]

@333fred can you help me with the Extract-Method test?

[333fred]

@333fred can you help me with the Extract-Method test?

Not really, but @CyrusNajmabadi should be able to :)

[bernd5]

@CyrusNajmabadi can you help here?

[CyrusNajmabadi]

Hi @bernd5 can you just go to src\EditorFeatures\CSharpTest\CodeActions\ExtractMethod\ExtractMethodTests.cs and add some tests that demonstrate the behavior of Extract-Method inside a constructor initializer. We should either see no change in behavior with your compiler change (so run the tests before/after to validate that), or we should see better behavior after your change. If things change and get worse, we'd need to discuss what was happening and what we should do about it. THanks!

[bernd5]

I can try - but I actually do not know how the tests work - or what they actually test.

What means for example:

[|b != true|]

or

|Rename:NewMethod|

I assume, [| describes the region start of the analysis - b != true is the analyzed expression - and |] terminates the analysis context.