Source Generator & Analyser Best Practices

Preview sourcegeneratorframework Reviewed 2026-09-12 purview-dev/sourcegeneratorframework

Source Generator & Analyser Best Practices

Practical guidance for writing Roslyn analysers and incremental source generators that remain fast, deterministic, cache-friendly, IDE-compatible, and safe to distribute.

1. Core Principles

The most important rules are:

Use an analyser to validate user code.
Use an incremental generator to generate code.
Use ForAttributeWithMetadataName for attribute-driven generators.
Remove Roslyn objects from the incremental pipeline as early as possible.
Every value crossing a pipeline boundary should have meaningful value equality.
Prefer many small incremental stages over one large transform.
Keep broad inputs such as Compilation away from downstream generation.
Generate deterministic output.
Compile against the oldest Roslyn API version you actually need.
Test caching, not just generated text.

The guiding principle for an incremental generator is:

Extract semantic information once, convert it into a small value model, and make everything downstream operate only on that value model.

2. Analyser or Source Generator?

Analysers and generators have different responsibilities.

A DiagnosticAnalyser answers:

Is the source code valid according to this library’s rules?

An IIncrementalGenerator answers:

Given valid source code, what source should be generated?

Decision Table

Requirement	Prefer	Reason
Require a class to be `partial`	Analyser	User-code contract violation
Require an attribute on a declaration	Analyser	User-code contract violation
Validate a method signature	Analyser	Semantic validation
Reject unsupported property types	Analyser	Better IDE feedback
Detect invalid attribute arguments	Analyser	Natural diagnostic
Detect unsupported API usage	Analyser	Operation analysis
Offer an automatic fix	Analyser + `CodeFixProvider`	Code fixes operate on diagnostics
Generate members for a marked class	Incremental generator	Generation
Generate serializers/validators/mappers	Incremental generator	Generation
Generate a registry from discovered types	Incremental generator	Generation
Read an additional schema file and generate C#	Incremental generator	Generation
Report an unexpected internal generation failure	Generator diagnostic	Generation-specific failure
Validate generator-only external input	Generator diagnostic may be appropriate	Analyser may not see equivalent input

Prefer:

User Source
    │
    ├── Analyser
    │     ├── discovers relevant source
    │     ├── validates the contract
    │     ├── reports diagnostics
    │     └── optionally provides code fixes
    │
    └── Incremental Generator
          ├── discovers relevant source
          ├── extracts semantic values
          ├── creates equatable models
          └── generates deterministic source

A useful shorthand is:

Analysers protect the contract. Generators implement the contract.

3. Choosing an Analyser Action

Use the narrowest analyser API that directly represents the thing being analysed.

Do not start with a broad syntax scan if Roslyn already exposes the concept as a symbol or operation.

Analyser Action Decision Matrix

API	Use when	Examples	Recommendation
`RegisterSyntaxNodeAction`	Exact source syntax matters	Modifier presence, declaration form	Use for syntax rules
`RegisterSymbolAction`	A declaration’s semantic meaning matters	Accessibility, attributes, implemented interfaces	Preferred for declaration rules
`RegisterOperationAction`	Executable behaviour/API usage matters	Invocation, assignment, conversion, object creation	Preferred for semantic usage rules
`RegisterOperationBlockStartAction`	Multiple operations inside one body need shared state	Track resource use throughout a method	Use for stateful method analysis
`RegisterOperationBlockAction`	Whole executable body must be analysed	Final body-level validation	Prefer narrower actions where possible
`RegisterCodeBlockStartAction`	Syntax-oriented body analysis needs state	Stateful syntax rules	Less common than operation-block analysis
`RegisterCodeBlockAction`	Entire syntax code block matters	Body-level syntax rules	Use sparingly
`RegisterSymbolStartAction`	A symbol and its members must be analysed together	Type-wide analysis across members	Powerful but relatively expensive
Symbol end action	Result depends on all child/member analysis	Report once for entire type	Register from symbol-start
`RegisterCompilationStartAction`	Expensive semantic setup should happen once	Resolve known framework/library symbols	Good initialization boundary
Compilation end action	A result genuinely depends on the entire compilation	Global collision/aggregate rule	Avoid unless necessary
`RegisterAdditionalFileAction`	Analyze `AdditionalFiles`	Config/schema validation	Correct abstraction
`RegisterSemanticModelAction`	Analysis genuinely applies to an entire semantic model	Rare tree-wide semantic rule	Usually too broad
`RegisterSyntaxTreeAction`	Entire raw syntax tree matters	File header / file-level syntax	Prefer node actions when possible

4. Syntax vs Symbol vs Operation

The most common analyser design decision is choosing between:

syntax;
symbols;
operations.

Quick Decision

Does exact source spelling/structure matter?
        │
        ├── Yes ──► Syntax
        │
        └── No
             │
             ├── Is this a declaration?
             │       └── Yes ──► Symbol
             │
             └── Is this executable behaviour?
                     └── Yes ──► Operation

Syntax

Use syntax when the literal structure of the user’s source matters.

Examples:

is partial explicitly present?
did the user write a primary constructor?
is the namespace file-scoped?
was an explicit modifier specified?
is a declaration syntactically structured in a particular way?

Example:

context.RegisterSyntaxNodeAction(
    AnalyzeClass,
    SyntaxKind.ClassDeclaration
);

Syntax is often the cheapest solution when no semantic information is required.

Do not ask the semantic model a question that can be answered directly from syntax.

Symbols

Use symbols when analysing declarations semantically.

Examples:

does this type implement IDisposable?
does this member have a particular attribute?
what is the method return type?
what is the property’s accessibility?
what generic type arguments are present?
is this type abstract?
which containing namespace owns this type?

Example:

context.RegisterSymbolAction(
    AnalyzeNamedType,
    SymbolKind.NamedType
);

When comparing symbols:

SymbolEqualityComparer.Default.Equals(left, right)

should normally be used rather than reference equality.

Operations

Use IOperation when analysing executable semantics.

Examples:

method invocation;
constructor invocation;
assignment;
conversion;
property access;
field access;
argument passing;
return values;
await;
binary/unary operations.

For example:

context.RegisterOperationAction(
    AnalyzeInvocation,
    OperationKind.Invocation
);

Then:

static void AnalyzeInvocation(OperationAnalysisContext context)
{
    var invocation = (IInvocationOperation)context.Operation;

    var method = invocation.TargetMethod;

    // Semantic method information is already available.
}

This is normally preferable to:

context.RegisterSyntaxNodeAction(
    AnalyzeInvocation,
    SyntaxKind.InvocationExpression
);

followed by:

context.SemanticModel.GetSymbolInfo(...)

for every invocation.

Common Operation Kinds

Requirement	Operation
Method invocation	`OperationKind.Invocation`
Constructor invocation	`OperationKind.ObjectCreation`
Assignment	`OperationKind.SimpleAssignment`
Compound assignment	compound assignment operation kinds
Argument validation	`OperationKind.Argument`
Property access	`OperationKind.PropertyReference`
Field access	`OperationKind.FieldReference`
Conversion	`OperationKind.Conversion`
Return expression	`OperationKind.Return`
Await	`OperationKind.Await`
Binary expression	`OperationKind.Binary`
Unary expression	`OperationKind.Unary`

Use the semantic operation rather than reconstructing equivalent information from syntax whenever possible.

5. Analyser Best Practices

Enable Concurrent Execution

Analysers should normally enable concurrent execution:

public override void Initialize(AnalysisContext context)
{
    context.EnableConcurrentExecution();

    context.ConfigureGeneratedCodeAnalysis(
        GeneratedCodeAnalysisFlags.None
    );

    // Register actions.
}

Analyser callbacks can execute concurrently.

Avoid shared mutable state.

Configure Generated Code Explicitly

Do not leave generated-code handling implicit.

For most library contract analysers:

context.ConfigureGeneratedCodeAnalysis(
    GeneratedCodeAnalysisFlags.None
);

is appropriate.

Only inspect generated code if the analyser explicitly needs to.

Resolve Known Types Once

If an analyser needs to repeatedly compare against known framework or library types, resolve them during compilation start.

context.RegisterCompilationStartAction(static context =>
{
    var targetType =
        context.Compilation.GetTypeByMetadataName(
            "MyLibrary.SomeType"
        );

    if (targetType is null)
        return;

    context.RegisterOperationAction(
        c => AnalyzeInvocation(c, targetType),
        OperationKind.Invocation
    );
});

This is a good reason to use CompilationStart.

Do not use CompilationStart merely to enumerate the entire compilation.

Prefer Narrow Registration

Prefer:

context.RegisterOperationAction(
    AnalyzeInvocation,
    OperationKind.Invocation
);

over an action that sees every operation.

Prefer:

context.RegisterSyntaxNodeAction(
    AnalyzeClass,
    SyntaxKind.ClassDeclaration
);

over scanning an entire SyntaxTree.

6. Incremental Generator Golden Rules

Implement:

IIncrementalGenerator

rather than the legacy:

ISourceGenerator

But simply implementing IIncrementalGenerator does not make a generator meaningfully incremental.

Incrementally depends on the equality behaviour of the values flowing through the pipeline.

The Golden Rule

Pipeline values must be immutable and value-equatable.

Roslyn needs to determine:

Did this pipeline stage produce the same logical value as last time?

If the answer is yes, Roslyn can stop executing downstream stages and reuse cached results.

7. Pipeline Value Equality

Pipeline Red List

These should not normally survive into your generator model.

Type	Verdict	Why
`ISymbol`	❌ Never retain	Not suitable for pipeline equality; can retain old compilations
`INamedTypeSymbol`	❌ Never retain	Same problem as `ISymbol`
`IMethodSymbol`	❌ Never retain	Same problem as `ISymbol`
`IPropertySymbol`	❌ Never retain	Same problem as `ISymbol`
`Compilation`	❌ Do not propagate	Huge semantic graph and broad invalidation source
`SemanticModel`	❌ Do not propagate	Bound to compilation/tree
`IOperation`	❌ Do not propagate	Compiler semantic graph object
`SyntaxTree`	❌ Do not propagate	Changes with source edits
`SyntaxNode`	⚠ Remove ASAP	Usually loses equality after edits to its tree
`Location`	⚠ Remove ASAP	Same incrementally problem as syntax
`AdditionalText`	⚠ Project immediately	Host/compiler input object
`T[]`	❌ Avoid in models	Reference equality
`List<T>`	❌ Avoid in models	Mutable and reference equality
`ImmutableArray<T>`	⚠ Wrap/compare explicitly	Immutable but not sequence-value-equatable for model equality
Mutable class	❌ Avoid	Reference equality unless explicitly implemented

Good Model

internal sealed record TypeModel(
    string Namespace,
    string Name,
    string FullyQualifiedName,
    Accessibility Accessibility,
    EquatableArray<PropertyModel> Properties
);

internal sealed record PropertyModel(
    string Name,
    string FullyQualifiedTypeName,
    bool IsNullable
);

Bad Model

internal sealed record TypeModel(
    INamedTypeSymbol Symbol,
    Compilation Compilation,
    Location Location,
    ImmutableArray<IPropertySymbol> Properties
);

Making the outer object a record does not magically make its members suitable for incremental equality.

`ImmutableArray<T>` Is Not Enough

ImmutableArray<T> solves:

Can this collection be mutated?

It does not automatically solve:

Do two separately-created collections containing equivalent elements compare as the same sequence for my pipeline model?

These are different problems.

For incremental models, prefer something such as:

EquatableArray<T>

with sequence-based equality.

Conceptually:

internal readonly struct EquatableArray<T>
    : IEquatable<EquatableArray<T>>
{
    private readonly ImmutableArray<T> _items;

    public bool Equals(EquatableArray<T> other)
    {
        if (_items.Length != other._items.Length)
            return false;

        return _items
            .AsSpan()
            .SequenceEqual(other._items.AsSpan());
    }

    public override bool Equals(object? obj) =>
        obj is EquatableArray<T> other &&
        Equals(other);

    public override int GetHashCode()
    {
        var hash = new HashCode();

        foreach (var item in _items)
            hash.Add(item);

        return hash.ToHashCode();
    }
}

The precise implementation can vary.

The important requirement is:

same contents => equal pipeline value

8. Designing the Incremental Pipeline

Think of every transformation as a cache checkpoint.

Prefer:

Roslyn Input
    │
    ▼
Cheap discovery
    │
    ▼
Semantic extraction
    │
    ▼
Small equatable model
    │
    ▼
Validation/transformation
    │
    ▼
Generation model
    │
    ▼
Source output

Do not do:

Roslyn Input
    │
    ▼
Giant transform containing symbols + syntax + compilation
    │
    ▼
Generate everything

Project Early

The semantic transform should usually be the boundary where Roslyn objects disappear.

Example:

static TypeModel CreateModel(
    GeneratorAttributeSyntaxContext context,
    CancellationToken cancellationToken
)
{
    var symbol = (INamedTypeSymbol)context.TargetSymbol;

    return new TypeModel(
        Namespace:
            symbol.ContainingNamespace.ToDisplayString(),

        Name:
            symbol.Name,

        FullyQualifiedName:
            symbol.ToDisplayString(
                SymbolDisplayFormat.FullyQualifiedFormat
            ),

        Accessibility:
            symbol.DeclaredAccessibility
    );
}

Everything downstream should receive TypeModel, not INamedTypeSymbol.

Prefer Static Lambdas

Prefer:

.Select(static (value, cancellationToken) =>
{
    return Transform(value, cancellationToken);
});

Static callbacks prevent accidental capture of generator instance state.

Generator instances should not be treated as application services or state containers.

Honour Cancellation

For non-trivial transformations:

.Select(static (value, cancellationToken) =>
{
    cancellationToken.ThrowIfCancellationRequested();

    return Transform(value, cancellationToken);
});

Pass cancellation tokens into Roslyn APIs that accept them.

Split Transformations

Prefer:

Syntax
  ↓
Symbol projection
  ↓
Type model
  ↓
Property models
  ↓
Generation model
  ↓
Output

over:

Syntax
  ↓
Do absolutely everything
  ↓
Output

More meaningful boundaries give Roslyn more opportunities to short-circuit downstream processing.

9. Syntax Discovery

Prefer `ForAttributeWithMetadataName`

Attribute-driven generation should normally start with:

context.SyntaxProvider.ForAttributeWithMetadataName(
    fullyQualifiedMetadataName:
        "MyLibrary.GenerateAttribute",

    predicate:
        static (node, _) =>
            node is TypeDeclarationSyntax,

    transform:
        static (context, cancellationToken) =>
            CreateModel(context, cancellationToken)
);

Advantages include:

highly optimized discovery;
alias support;
direct TargetSymbol;
matching AttributeData;
obvious user intent;
easier analyser integration.

For marker-attribute generators, this should be the default.

Use `CreateSyntaxProvider` When Syntax Is Actually the Trigger

Use:

context.SyntaxProvider.CreateSyntaxProvider(...)

when there is no appropriate marker attribute.

Examples:

syntax-driven DSL;
a generator intentionally driven by a language construct;
a pattern that cannot reasonably use an attribute.

The predicate must be cheap.

Good:

predicate:
    static (node, _) =>
        node is ClassDeclarationSyntax
        {
            AttributeLists.Count: > 0
        }

Bad:

predicate:
    static (node, _) =>
    {
        // expensive walking
        // semantic work
        // allocations
        // string construction
        return true;
    }

The predicate runs extremely frequently.

Semantic work belongs in the transformation callback.

Avoid Indirect Discovery

Avoid designs that require discovering:

every indirect implementation of an interface;
every indirect subclass;
inherited marker attributes through arbitrary hierarchies;
every type in a compilation followed by manual filtering.

A change high in a type hierarchy can invalidate a large arbitrary portion of the compilation.

Prefer explicit intent:

[GenerateSchema]
partial class Customer
{
}

over:

Generate everything somewhere downstream of IBaseSchemaThing

10. `Collect`, `Combine`, and Invalidation

`Collect()`

Collect() transforms:

IncrementalValuesProvider<T>

into roughly:

IncrementalValueProvider<ImmutableArray<T>>

This changes invalidation scope.

Before:

A ──► output A
B ──► output B
C ──► output C

After collection:

A ─┐
B ─┼──► [A,B,C] ──► output
C ─┘

Changing B changes the aggregate [A,B,C].

Prefer Per-Item Output

Prefer:

context.RegisterSourceOutput(
    models,
    static (context, model) =>
        Emit(context, model)
);

instead of:

context.RegisterSourceOutput(
    models.Collect(),
    static (context, models) =>
    {
        foreach (var model in models)
            Emit(context, model);
    }
);

unless the generation genuinely requires the complete set.

Good Uses of `Collect()`

Use Collect() when generating something intrinsically global:

one registry containing every handler;
one lookup containing every generated type;
duplicate-name detection across all targets;
one aggregate switch;
one generated dependency map.

A useful design is:

                    ┌──► Per-type source
Type Models ────────┤
                    │
                    └──► Collect()
                            │
                            ▼
                     Global registry

Only the registry should pay the global invalidation cost.

`Combine()`

Use Combine() when one output logically depends on two providers.

Example:

var generationInput =
    typeModels.Combine(generatorOptions);

That means:

type changed ───────┐
                    ├──► generation invalidated
option changed ─────┘

This is correct if either input should regenerate the output.

Be Very Careful Combining `CompilationProvider`

This:

models.Combine(context.CompilationProvider)

is often an incremental performance smell.

Almost any semantic change can replace the compilation.

If possible, project the compilation into the tiny fact you actually need:

var capabilities =
    context.CompilationProvider
        .Select(static (compilation, _) =>
            new CompilationCapabilities(
                HasRequiredType:
                    compilation.GetTypeByMetadataName(
                        "MyLibrary.RequiredType"
                    ) is not null
            )
        );

Then:

models.Combine(capabilities)

At least downstream equality can now short-circuit when the relevant capability did not change.

`WithComparer()`

Roslyn provides:

.WithComparer(...)

when default equality is insufficient.

Example:

provider.WithComparer(
    MyModelComparer.Instance
);

Use this when your logical equality differs from the default implementation.

Do not use it as a way to justify retaining large compiler objects inside the model.

This is suspicious:

record Model(
    INamedTypeSymbol Symbol,
    Compilation Compilation
);

followed by an elaborate comparer.

The better solution is usually to redesign Model.

11. Diagnostics

Prefer a Separate Analyser

Normal user validation belongs in a DiagnosticAnalyser.

Benefits include:

immediate IDE feedback;
independent execution from generation;
easier testing;
code-fix support;
simpler incremental generator pipelines.

Generator Diagnostics Are Still Valid

Generator diagnostics make sense for things such as:

malformed additional files;
invalid generator-only configuration;
conflicting generated output discovered only during generation;
failures that cannot naturally be expressed by a separate analyser.

Do not turn the generator pipeline into an analyser pipeline by default.

Blocking vs Non-Blocking Generator Diagnostics

GeneratorResult<T>.ShouldProcess decides whether the output stage runs for a target. It is true when the result carries a value and none of its carried ReportableDiagnostic diagnostics are blocking.

Whether a diagnostic blocks is an explicit, per-diagnostic decision (ReportableDiagnostic.IsBlocking), independent of its severity. An Error-severity diagnostic can still allow generation to continue when the generated code helps the developer fix the problem — for example, emitting an abstract base class alongside an error for a missing override, so the user can see what to implement:

var diagnostic = ReportableDiagnostic.Create(
    MissingOverride,
    isBlocking: false,          // report the error, but keep generating
    symbol,
    symbol.Name,
    "Execute"
);

return GeneratorResult<BaseModel>.Create(model, diagnostic);

A blocking diagnostic (isBlocking: true) stops generation for that target while still being reported. Prefer blocking diagnostics for genuine contract violations that would produce misleading output; prefer non-blocking diagnostics when the partial output is still useful.

Location Handling

Analysers should report diagnostics on the most useful user-authored Location.

Generators should avoid keeping Location inside long-lived pipeline models.

If a generator absolutely requires source position information, convert it into a value model:

internal readonly record struct SourceLocationModel(
    string FilePath,
    int Start,
    int Length
);

But even this should only be carried downstream if generation actually depends on the location.

12. Output Generation

Output Must Be Deterministic

For the same generator model:

input model
    ↓
identical generated source

Avoid:

current timestamps;
random GUIDs;
process IDs;
machine-specific paths;
unordered dictionary output;
machine environment variables;
current culture affecting generation.

Deterministic Hint Names

Good:

context.AddSource(
    $"{model.HintName}.g.cs",
    source
);

Bad:

context.AddSource(
    $"{Guid.NewGuid():N}.g.cs",
    source
);

Hint names must be:

deterministic;
unique within the generator;
stable when irrelevant source changes.

Prefer Text Generation

Do not build a complete Roslyn syntax tree merely to generate source unless there is a strong reason.

For generator output, a small code writer or structured string builder is generally easier and faster.

Avoid repeatedly doing:

syntax.NormalizeWhitespace().ToFullString()

for large generated trees.

Generated output may use C# 14 features — extension-member blocks (extension(...), via CodeWriter.ExtensionBlockScope), the field keyword, collection expressions — when the target compilation supports them. The framework is built against Roslyn 5.x, so its generators may emit C# 14 output; consumers need a matching compiler (.NET 10 SDK / Roslyn 5.0 or later) to compile it. Gate any newer-than-baseline features on GenerationSettings.LanguageVersion when a generator must also serve older hosts.

Post-Initialization Output

Use:

RegisterPostInitializationOutput

for source that is constant regardless of the user’s compilation.

Examples:

marker attributes;
fixed helper attributes;
static support types.

Example:

context.RegisterPostInitializationOutput(
    static context =>
    {
        context.AddSource(
            "GenerateAttribute.g.cs",
            SourceText.From(
                """
                // <auto-generated/>

                namespace MyLibrary;

                [global::System.AttributeUsage(
                    global::System.AttributeTargets.Class,
                    AllowMultiple = false,
                    Inherited = false)]
                internal sealed class GenerateAttribute
                    : global::System.Attribute
                {
                }
                """,
                Encoding.UTF8
            )
        );
    }
);

13. Testing Incrementally

Snapshot-testing generated source is not sufficient.

A generator can generate perfectly correct code while defeating almost all incremental caching.

Test both:

Correctness
+
Incrementally

Test Cases

At minimum test:

first execution produces expected output;
identical second execution is cached;
unrelated source changes remain cached;
changing one target only invalidates that target;
changing one property only invalidates dependent stages;
deleting a target removes its output;
renaming a target changes the expected hint/source;
changing global generator options invalidates appropriate output;
changing an additional file invalidates only dependent output;
global registry generation invalidates when expected.

Track Incremental Generator Steps

Create the generator driver with tracking enabled.

For example:

var driverOptions =
    new GeneratorDriverOptions(
        disabledOutputs:
            IncrementalGeneratorOutputKind.None,

        trackIncrementalGeneratorSteps:
            true
    );

Inspect tracked output reasons such as:

New
Modified
Unchanged
Cached
Removed

The exact reason expected depends on the stage and test scenario.

The important point is that tests should prove:

An unrelated edit does not rerun expensive downstream generation.

Framework support

The framework’s testing packages make step-cache tests first-class. SourceGeneratorTestRunner.RunIncrementalAsync runs one shared driver over a sequence of source sets with step tracking enabled, and every pipeline helper assigns a tracking name so tests can reference individual stages.

var result = await new SourceGeneratorTestRunner<ServiceRegistrationGenerator>().RunIncrementalAsync(
  [
    new IncrementalRunInput([firstSources]),
    new IncrementalRunInput([changedSources]),
  ],
  options,
  cancellationToken
);

await Assert.That(result.Runs[0]).AllStepsNew();
await Assert.That(result.Runs[1]).StepIsCached("ForAttribute_GenerateServiceAttribute");
await Assert.That(result.Runs[1]).StepIsModified("GetGenerationConfiguration");

Assertions on IncrementalCacheRun (AllStepsNew, AllStepsCachedOrUnchanged, StepIsCached, StepIsModified, HasStepReason) plus GetStepReasons() cover the golden matrix. See docs/step-cache-tests.md for the full walkthrough and the canonical StepCacheTests.cs sample in the ExampleGenerator unit tests.

14. Roslyn Version Compatibility

The most important packaging rule is:

The version of Microsoft.CodeAnalysis.* used to compile your analyser/generator establishes a minimum compiler-host API requirement.

The consumer’s:

<TargetFramework>...</TargetFramework>

does not determine analyser compatibility.

Analyser/generator code executes inside a compiler/IDE host.

Roslyn / Visual Studio Compatibility

Microsoft’s published compatibility baseline is:

Roslyn package	Minimum Visual Studio	Language / .NET generation
4.0.1	VS 2022 17.0	C# 10 / .NET 6
4.1	VS 2022 17.1	C# 10 / .NET 6
4.2	VS 2022 17.2	C# 10 / .NET 6
4.3.1	VS 2022 17.3	C# 10 / .NET 6
4.4	VS 2022 17.4	C# 11 / .NET 7
4.5	VS 2022 17.5	C# 11 / .NET 7
4.6	VS 2022 17.6	C# 11 / .NET 7
4.7	VS 2022 17.7	C# 11 / .NET 7
4.8	VS 2022 17.8	C# 12 / .NET 8
4.9.2	VS 2022 17.9	C# 12 / .NET 8
4.10	VS 2022 17.10	C# 12 / .NET 8
4.11	VS 2022 17.11	C# 12 / .NET 8
4.12	VS 2022 17.12	C# 13 / .NET 9
4.13	VS 2022 17.13	C# 13 / .NET 9
4.14	VS 2022 17.14	C# 13 / .NET 9
5.0	VS 2026 18.0	C# 14 / .NET 10

This table gives the minimum documented Visual Studio host.

This framework is built against Roslyn 5.0. The generator, analyzer, and testing assemblies in Purview.SourceGeneratorFramework* are compiled against Microsoft.CodeAnalysis 5.x, so compiler hosts that load them must be Roslyn 5.0 or later (.NET 10 SDK / Visual Studio 2026 18.0). The testing packages multi-target net8.0–net10.0; Roslyn 5.x ships net8.0/net9.0 package assets, so those test targets still load the test runner.

Do not interpret it as:

net8.0 application = Roslyn 4.8 analyser

That is incorrect.

Example

A project may target:

<TargetFramework>net8.0</TargetFramework>

while being compiled by:

Visual Studio 2026 / Roslyn 5.x

An analyser compiled against Roslyn 5.0 may therefore work.

The same net8.0 project opened in:

Visual Studio 2022 17.8 / Roslyn 4.8

cannot be assumed to load that Roslyn-5.0-based analyser.

The application TFM did not change.

The compiler host did.

15. Visual Studio, .NET SDK, and Rider

Safe Roslyn Baselines

Choose the oldest Roslyn version containing the APIs you require.

Typical baseline choices are:

Minimum tooling you intend to support	Maximum baseline you should normally compile against
VS 2022 17.8 / initial .NET 8 generation	Roslyn 4.8
VS 2022 17.10	Roslyn 4.10
VS 2022 17.12 / initial .NET 9 generation	Roslyn 4.12
VS 2022 17.14	Roslyn 4.14
VS 2026 18.0 / initial .NET 10 generation	Roslyn 5.0

If you compile against a later package, you have deliberately raised your minimum host requirement unless you have proven otherwise.

.NET SDK

The .NET SDK contains a compiler toolchain.

Broad release alignment is:

.NET 8  / C# 12 ──► Roslyn 4.8 generation
.NET 9  / C# 13 ──► Roslyn 4.12 generation
.NET 10 / C# 14 ──► Roslyn 5.0 generation

However, SDK servicing and feature bands can contain later compiler versions.

Therefore do not use:

TargetFramework == net10.0

as proof that a particular Roslyn API is available to your analyser.

Likewise:

$(TargetFramework)

should not be used to choose the analyser binary.

The relevant variable is the compiler host.

Rider

Rider supports:

Roslyn analysers;
source generators;
generated-source navigation;
analyser diagnostics;
analyser quick fixes;
source generator execution.

However, JetBrains does not publish the same simple:

Rider Version => Maximum Microsoft.CodeAnalysis Version

matrix that Microsoft publishes for Visual Studio.

Therefore:

Do not invent a Rider/Roslyn version mapping.

If Rider support is part of your package contract:

choose a conservative Roslyn baseline;
test the oldest Rider version you support;
test dotnet build;
test Rider design-time generation;
test generated-source navigation;
test analysers and code fixes where applicable.

Build-time compiler compatibility and Rider IDE integration should be tested independently.

16. Multi-Version Roslyn Packaging

This area is frequently misunderstood.

NuGet Analyser Assets Are Not Normal TFM Assets

Normal runtime/library assets support selection such as:

lib/net8.0/
lib/net9.0/
lib/net10.0/

Analyser assets conventionally live under:

analysers/
    dotnet/
        cs/
            MyGenerator.dll

This is not a general-purpose:

Roslyn 4.8
Roslyn 4.14
Roslyn 5.0

selection mechanism.

Do not place multiple Roslyn-targeted implementations into the ordinary analyser folder and expect NuGet to automatically choose the correct one.

Strategy 1 — One Conservative Binary

Recommended Default

Compile against the oldest Roslyn version required by your implementation.

For example:

<PackageReference
    Include="Microsoft.CodeAnalysis.CSharp"
    Version="4.8.0"
    PrivateAssets="all" />

Package:

analysers/
    dotnet/
        cs/
            MyGenerator.dll

Advantages:

simplest;
predictable;
broadest compatibility;
works naturally with IDEs;
minimal packaging logic.

Disadvantage:

cannot statically call newer Roslyn APIs.

For most public generators, this is the correct approach.

Strategy 2 — Raise the Package Baseline

If a newer Roslyn feature materially improves the generator, it may be better to explicitly raise the minimum compiler version.

For example:

MyGenerator 2.x
    Roslyn >= 4.8

MyGenerator 3.x
    Roslyn >= 5.0

Document the minimum IDE/compiler requirement.

This is much easier for users to reason about than hidden runtime selection.

Strategy 3 — Separate Packages

For significantly different implementations:

MyGenerator
MyGenerator.Roslyn5

can be reasonable.

Advantages:

explicit;
predictable;
simple runtime behaviour.

Disadvantages:

more packages;
more maintenance;
users must select correctly.

Strategy 4 — MSBuild-Selected Binary

Advanced packages can store binaries outside the automatically discovered analyser directory:

analysers/
    roslyn4.8/
        MyGenerator.dll

    roslyn5.0/
        MyGenerator.dll

buildTransitive/
    MyGenerator.targets

Then a targets file can explicitly add exactly one:

<Analyser Include="..." />

depending on an intentionally selected compatibility band.

Conceptually:

<ItemGroup Condition="'$(MyGeneratorRoslynBand)' == '4.8'">
    <Analyser
        Include="$(MSBuildThisFileDirectory)..\analysers\roslyn4.8\MyGenerator.dll" />
</ItemGroup>

<ItemGroup Condition="'$(MyGeneratorRoslynBand)' == '5.0'">
    <Analyser
        Include="$(MSBuildThisFileDirectory)..\analysers\roslyn5.0\MyGenerator.dll" />
</ItemGroup>

The difficult question is:

How is MyGeneratorRoslynBand determined reliably?

There is no general NuGet analyser-asset negotiation equivalent to normal TFM selection.

Do not use:

$(TargetFramework)

for this.

It identifies the application runtime target, not the compiler host.

Using:

$(NETCoreSdkVersion)

may work for a deliberately SDK-bound support model but must not be treated as universally equivalent to the active Roslyn host.

Design-time builds, Visual Studio, Rider, CI, and explicit compiler toolsets all need testing.

Multi-Targeting the Generator Is Not Automatic Selection

This:

<TargetFrameworks>
    netstandard2.0;net8.0
</TargetFrameworks>

may produce two generator assemblies.

It does not mean NuGet will choose:

netstandard2.0 analyser for old compiler
net8.0 analyser for new compiler

for you.

Building multiple binaries and selecting analyser assets are separate problems.

Recommended Rule

Unless there is a compelling requirement:

Ship one netstandard2.0 analyser/generator binary compiled against the oldest Roslyn API version you need.

This remains the most robust distribution strategy.

17. `Microsoft.CodeAnalysis.Analysers`

Do not confuse:

Microsoft.CodeAnalysis.CSharp

with:

Microsoft.CodeAnalysis.Analysers

They serve different purposes.

`Microsoft.CodeAnalysis.CSharp`

Provides Roslyn compiler APIs used to implement your analyser/generator.

Examples:

IIncrementalGenerator
DiagnosticAnalyser
SyntaxNode
Compilation
ISymbol
IOperation

`Microsoft.CodeAnalysis.Analysers`

This is a meta-analyser package.

It analyses your analyser or source generator.

Its purpose is to detect incorrect or unsafe usage of Roslyn/compiler APIs.

It does not define your source-generator API baseline.

Current Package Version

As of August 2026, the current stable package is:

Microsoft.CodeAnalysis.Analysers 5.9.0

Do not assume its version must match:

Microsoft.CodeAnalysis.CSharp

For example, it is perfectly reasonable to have:

<PackageReference
    Include="Microsoft.CodeAnalysis.CSharp"
    Version="4.8.0"
    PrivateAssets="all" />

<PackageReference
    Include="Microsoft.CodeAnalysis.Analysers"
    Version="5.9.0"
    PrivateAssets="all" />

provided the meta-analyser version itself works with your build tooling.

These represent separate concerns:

Microsoft.CodeAnalysis.CSharp
        │
        └── minimum Roslyn API used by your generator

Microsoft.CodeAnalysis.Analysers
        │
        └── rules used while developing the generator

Transitive Availability

Roslyn compiler packages already bring Microsoft.CodeAnalysis.Analysers into the dependency graph as development tooling.

You may nevertheless explicitly reference it when:

using Central Package Management;
deliberately pinning meta-analyser behaviour;
keeping analyser tooling versions consistent across a repository;
making the analyser-project configuration obvious.

`PrivateAssets="all"`

Roslyn development dependencies should normally use:

PrivateAssets="all"

Example:

<PackageReference
    Include="Microsoft.CodeAnalysis.CSharp"
    Version="$(RoslynVersion)"
    PrivateAssets="all" />

<PackageReference
    Include="Microsoft.CodeAnalysis.Analysers"
    Version="$(RoslynAnalyserVersion)"
    PrivateAssets="all" />

Your consumer should not gain ordinary runtime Roslyn package dependencies simply because it installed your source generator.

`EnforceExtendedAnalyserRules`

Analyser and generator projects should normally enable:

<EnforceExtendedAnalyserRules>
    true
</EnforceExtendedAnalyserRules>

These rules detect implementation patterns that are particularly dangerous inside compiler-hosted code.

Do not immediately suppress an RSxxxx diagnostic.

First determine what compiler-host invariant the rule is protecting.

RS1035

RS1035 bans APIs considered inappropriate for analysers.

A common example is direct access to environment-dependent state.

The underlying principle is:

Analyser/generator execution should not silently depend on machine-global environment state.

Configuration should normally arrive through explicit compiler inputs such as:

analyser config options;
additional files;
MSBuild properties exposed through analyser config;
source code;
metadata references.

RS2008

RS2008 relates to analyser diagnostic release tracking.

If your analyser publishes public diagnostic IDs, maintain release tracking files such as:

AnalyserReleases.Shipped.md
AnalyserReleases.Unshipped.md

This helps detect accidental changes to diagnostic contracts.

Diagnostic IDs are effectively part of your public API.

Treat Diagnostic Descriptors as Public Contracts

Changing:

ZS0001

to:

ZS0017

may break:

.editorconfig;
suppressions;
CI configuration;
documentation;
consumer tooling.

Similarly, changing:

default severity;
category;
diagnostic semantics;

should be treated as a compatibility decision.

18. Recommended Project Configuration

A broadly-compatible generator project might start with:

<Project Sdk="Microsoft.NET.Sdk">

  <PropertyGroup>
    <TargetFramework>netstandard2.0</TargetFramework>

    <LangVersion>latest</LangVersion>
    <Nullable>enable</Nullable>

    <IsPackable>true</IsPackable>
    <IncludeBuildOutput>false</IncludeBuildOutput>

    <EnforceExtendedAnalyserRules>true</EnforceExtendedAnalyserRules>

    <TreatWarningsAsErrors>true</TreatWarningsAsErrors>
  </PropertyGroup>

  <ItemGroup>

    <PackageReference
        Include="Microsoft.CodeAnalysis.CSharp"
        Version="$(RoslynVersion)"
        PrivateAssets="all" />

    <PackageReference
        Include="Microsoft.CodeAnalysis.Analysers"
        Version="$(RoslynAnalyserVersion)"
        PrivateAssets="all" />

  </ItemGroup>

  <ItemGroup>

    <None
        Include="$(TargetPath)"
        Pack="true"
        PackagePath="analysers/dotnet/cs"
        Visible="false" />

  </ItemGroup>

</Project>

Then centrally define:

<PropertyGroup>
  <RoslynVersion>4.8.0</RoslynVersion>
  <RoslynAnalyserVersion>5.9.0</RoslynAnalyserVersion>
</PropertyGroup>

The exact Roslyn baseline is a product-support decision.

ProjectReference During Development

A consuming project can reference the generator as:

<ProjectReference
    Include="..\MyGenerator\MyGenerator.csproj"
    OutputItemType="Analyser"
    ReferenceOutputAssembly="false" />

Separate Runtime Contracts From Compiler Tooling

Prefer:

MyLibrary.Abstractions
    │
    ├── public attributes
    ├── runtime contracts
    └── shared public APIs

MyLibrary.SourceGenerators
    │
    └── IIncrementalGenerator

MyLibrary.Analysers
    │
    └── DiagnosticAnalyser

MyLibrary.CodeFixes
    │
    └── CodeFixProvider

over mixing runtime APIs and compiler tooling into one assembly.

This prevents Roslyn dependencies leaking into runtime package assets.

Roslyn Component Discovery

The compiler host only loads a source generator, diagnostic analyser, or code fix provider when three conditions hold. Missing any one means the component is silently ignored, which is why “nothing shows up in Visual Studio” is usually a setup problem, not a code problem:

The type is public. Non-public component types cannot be instantiated by Roslyn (PSGFR27).
The type is decorated. A generator needs [Generator] (PSGFR26), an analyser needs [DiagnosticAnalyzer] (PSGFR25), and a code fix provider needs [ExportCodeFixProvider] (PSGFR24).
The assembly is loaded as an analyser. In a package the component assembly must be packed under analyzers/dotnet/cs/; in a project reference it must be referenced with OutputItemType="Analyser". A normal library reference never surfaces a component to Roslyn.

A code fix provider also only appears when the diagnostic ID in FixableDiagnosticIds is actually produced by an analyser that is loaded alongside it (PSGFR28). Visual Studio MEF-composes fix providers when the analyser set loads, so after adding or updating a fixer assembly you must restart Visual Studio or reload the project for the fixes to appear.

19. Extension Class Conventions

Extension classes should form a coherent, discoverable shape so that “which type does this extend?” and “where does it live?” are answerable from the file path alone.

Placement: under an Extensions folder whose path mirrors the extended type’s namespace, e.g. Extensions/Microsoft/CodeAnalysis/..., Extensions/System/....
Namespace: the extended type’s namespace, so the folder, namespace, and receiver all align (IDE0130 enforces the folder↔namespace pairing).
Name: {Receiver}Extensions (plural suffix), one receiver type per class.
Style: prefer C# 14 extension(Receiver receiver) blocks over classic public static T Method(this Receiver receiver, ...) methods.
Metadata: [EditorBrowsable(EditorBrowsableState.Never)] on the class, so IntelliSense and documentation tooling treat them as framework plumbing rather than public API.

The framework’s analyzers enforce these rules:

Rule	What it enforces
`PSGFR34`	Prefer `extension(...)` blocks over classic `this`-parameter methods (gated on the language version supporting extension blocks).
`PSGFR35`	Class name matches the extended type (`{Receiver}Extensions`).
`PSGFR36`	Class lives in the extended type’s namespace.
`PSGFR37`	One receiver type per class.
`PSGFR38`	`[EditorBrowsable(EditorBrowsableState.Never)]` is present.

The ReorganizeExtensionClassCodeFixProvider converts a class with many disparate extensions into the coherent shape: it renames (PSGFR35), splits multi-receiver classes into per-type files (PSGFR37), moves the class under Extensions/{ReceiverNamespace}/ and updates referencing files (PSGFR36), and the ConvertToExtensionBlockCodeFixProvider converts classic methods to extension blocks (PSGFR34).

20. Review Checklist

Analyser

Incremental Generator

Packaging

Summary

The shortest version of this guide is:

Analyser for validation; generator for generation. Syntax for syntax, symbols for declarations, operations for executable semantics. Use ForAttributeWithMetadataName whenever possible. ISymbol, Compilation, SemanticModel, and IOperation do not belong in incremental pipeline models. Remove SyntaxNode and Location as soon as possible. Immutable does not mean equatable: arrays, lists, and ImmutableArray<T> require deliberate sequence equality. Use EquatableArray<T> or an equivalent value-equatable collection abstraction. Avoid Collect() until global knowledge is genuinely required. Never combine CompilationProvider into the pipeline merely because it is convenient. Compile against the oldest Roslyn API version containing the functionality you need. The consumer TFM does not determine analyser compatibility—the compiler host does. NuGet does not automatically choose between Roslyn-version-specific analyser binaries. Test incrementally and compatibility, not just generated source.

Source Generator & Analyser Best Practices

Source Generator & Analyser Best Practices

Contents

1. Core Principles

2. Analyser or Source Generator?

Decision Table

3. Choosing an Analyser Action

Analyser Action Decision Matrix

4. Syntax vs Symbol vs Operation

Quick Decision

Syntax

Symbols

Operations

Common Operation Kinds

5. Analyser Best Practices

Enable Concurrent Execution

Configure Generated Code Explicitly

Resolve Known Types Once

Prefer Narrow Registration

6. Incremental Generator Golden Rules

The Golden Rule

7. Pipeline Value Equality

Pipeline Red List

Good Model

Bad Model

ImmutableArray<T> Is Not Enough

8. Designing the Incremental Pipeline

Project Early

Prefer Static Lambdas

Honour Cancellation

Split Transformations

9. Syntax Discovery

Prefer ForAttributeWithMetadataName

Use CreateSyntaxProvider When Syntax Is Actually the Trigger

Avoid Indirect Discovery

10. Collect, Combine, and Invalidation

Collect()

Prefer Per-Item Output

Good Uses of Collect()

Combine()

Be Very Careful Combining CompilationProvider

WithComparer()

11. Diagnostics

Prefer a Separate Analyser

Generator Diagnostics Are Still Valid

Blocking vs Non-Blocking Generator Diagnostics

Location Handling

12. Output Generation

Output Must Be Deterministic

Deterministic Hint Names

Prefer Text Generation

Post-Initialization Output

13. Testing Incrementally

Test Cases

Track Incremental Generator Steps

Framework support

14. Roslyn Version Compatibility

Roslyn / Visual Studio Compatibility

Example

15. Visual Studio, .NET SDK, and Rider

Safe Roslyn Baselines

.NET SDK

Rider

16. Multi-Version Roslyn Packaging

NuGet Analyser Assets Are Not Normal TFM Assets

Strategy 1 — One Conservative Binary

Recommended Default

Strategy 2 — Raise the Package Baseline

Strategy 3 — Separate Packages

Strategy 4 — MSBuild-Selected Binary

Multi-Targeting the Generator Is Not Automatic Selection

Recommended Rule

17. Microsoft.CodeAnalysis.Analysers

Microsoft.CodeAnalysis.CSharp

Microsoft.CodeAnalysis.Analysers

Current Package Version

Transitive Availability

PrivateAssets="all"

EnforceExtendedAnalyserRules

RS1035

`ImmutableArray<T>` Is Not Enough

Prefer `ForAttributeWithMetadataName`

Use `CreateSyntaxProvider` When Syntax Is Actually the Trigger

10. `Collect`, `Combine`, and Invalidation

`Collect()`

Good Uses of `Collect()`

`Combine()`

Be Very Careful Combining `CompilationProvider`

`WithComparer()`

17. `Microsoft.CodeAnalysis.Analysers`

`Microsoft.CodeAnalysis.CSharp`

`Microsoft.CodeAnalysis.Analysers`

`PrivateAssets="all"`

`EnforceExtendedAnalyserRules`