Avoid caching compilation data and use value equality for SyntaxNodes #79051
Conversation
ghost
assigned 
|
Tagging subscribers to this area: @dotnet/interop-contrib Issue DetailsFixes #78242. Creates separate types for InteropAttributeData: one that holds compilation data (InteropAttributeCompilationData), and one that holds only the data necessary for the model to create the generated code (InteropAttributeModelData). This uncovered some issues with record equality in records that use SyntaxNode. For those, we need to override Equals or wrap the SyntaxNode in a type that overrides Equals to use IsEquivalentTo on the SyntaxNode. There are probably more places where we use SyntaxNode that aren't caught in the current tests. To make sure every record has the right equality, I wasn't sure if it would be better to override Equals for each of the records, or create a wrapper record struct for each SyntaxNode that implements the equality we want (and implicit casts to and from the SyntaxNode). Then we wouldn't have to explicitly override the equality in each record that has a SyntaxNode.
|
jtschuster
requested review from
jkoritzinsky,
elinor-fung and
AaronRobinsonMSFT
jtschuster
added
the
source-generator
...ystem.Runtime.InteropServices/gen/Microsoft.Interop.SourceGeneration/InteropAttributeData.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Runtime.InteropServices/gen/LibraryImportGenerator/LibraryImportData.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Runtime.InteropServices/gen/LibraryImportGenerator/LibraryImportData.cs
Outdated
Show resolved
Hide resolved
...ystem.Runtime.InteropServices/gen/Microsoft.Interop.SourceGeneration/InteropAttributeData.cs
Outdated
Show resolved
Hide resolved
...untime.InteropServices/gen/Microsoft.Interop.SourceGeneration/ManualTypeMarshallingHelper.cs
Show resolved
Hide resolved
...ies/System.Runtime.InteropServices/gen/Microsoft.Interop.SourceGeneration/ManagedTypeInfo.cs
Outdated
Show resolved
Hide resolved
- Add type doc comments for new types - Don't use SyntaxNode in GetHashCode implementation - Add comment on Dictionary equality check
There was a problem hiding this comment.
This uncovered some issues with record equality in records that use SyntaxNode. For those, we need to override Equals or wrap the SyntaxNode in a type that overrides Equals to use IsEquivalentTo on the SyntaxNode.
There are probably more places where we use SyntaxNode that aren't caught in the current tests.
Do you think it is worth filing an issue to do a pass over this?
| @@ -32,19 +31,27 @@ public record InteropAttributeData | |||
| public InteropAttributeMember IsUserDefined { get; init; } | |||
| public bool SetLastError { get; init; } | |||
| public StringMarshalling StringMarshalling { get; init; } | |||
| } | |||
There was a problem hiding this comment.
I'd add remarks to the InteropAttributeData doc about how it should not include any symbols / types that would keep a compilation alive.
| @@ -18,6 +15,19 @@ public abstract record ManagedTypeInfo(string FullTypeName, string DiagnosticFor | |||
| private TypeSyntax? _syntax; | |||
| public TypeSyntax Syntax => _syntax ??= SyntaxFactory.ParseTypeName(FullTypeName); | |||
|
|
|||
| public virtual bool Equals(ManagedTypeInfo other) | |||
There was a problem hiding this comment.
Should ManagedTypeInfo? be the parameter? Asking because the implementation checks for null.
There was a problem hiding this comment.
There shouldn't be a null check, thanks for pointing that out!
|
Only failure is #79874 |
Fixes #78242.
Creates separate types for InteropAttributeData: one that holds compilation data (InteropAttributeCompilationData), and one that holds only the data necessary for the model to create the generated code (InteropAttributeModelData).
This uncovered some issues with record equality in records that use SyntaxNode. For those, we need to override Equals or wrap the SyntaxNode in a type that overrides Equals to use IsEquivalentTo on the SyntaxNode.
There are probably more places where we use SyntaxNode that aren't caught in the current tests.
To make sure every record has the right equality, I wasn't sure if it would be better to override Equals for each of the records, or create a wrapper record struct for each SyntaxNode that implements the equality we want (and implicit casts to and from the SyntaxNode). Then we wouldn't have to explicitly override the equality in each record that has a SyntaxNode. I also overrode both Equals and GetHashCode, but I'm not confident in my GetHashCode implementation. It could also be done with IEquatable.Equals without needing GetHashCode, but that would require implementing the TypeSyntax equality for every type that inherits from ManagedTypeInfo.