Generated serialization should use JsonMarshal.GetRawUtf8Value extension for better perf on net9.0+

[m-nash]

The generated serializers currently emit code like this for nested model deserialization (example from AvailabilitySetData.Serialization.cs#L184 in azure-sdk-for-net):

systemData = ModelReaderWriter.Read<SystemData>(
    new BinaryData(Encoding.UTF8.GetBytes(property.Value.GetRawText())),
    ModelSerializationExtensions.WireOptions,
    AzureResourceManagerComputeContext.Default);

This pattern does two extra allocations/copies per property:

1. GetRawText() allocates a transcoded UTF-16 string from the underlying UTF-8 bytes.
2. Encoding.UTF8.GetBytes(...) then re-encodes that string back into a new UTF-8 byte array.

On net9.0+ we can avoid both by using JsonMarshal.GetRawUtf8Value(element), which returns the raw UTF-8 span directly from the JsonDocument backing buffer. An extension method already exists in the repo demonstrating the pattern:

https://github.com/Azure/azure-sdk-for-net/blob/main/sdk/core/System.ClientModel/tests/client/ModelReaderWriter/JsonElementExtensions.cs

public static BinaryData GetUtf8Bytes(this JsonElement element)
{
#if NET9_0_OR_GREATER
    return new BinaryData(JsonMarshal.GetRawUtf8Value(element).ToArray());
#else
    return BinaryData.FromString(element.GetRawText());
#endif
}

Ask

Update the emitter so the generated deserialization code uses this extension (e.g. property.Value.GetUtf8Bytes()) instead of the Encoding.UTF8.GetBytes(GetRawText()) pattern. This will reduce allocations and improve deserialization throughput on the modern TFMs we now target.

[github-actions]

🎯 Issue Triage

A feature request to optimize the C# client emitter's generated deserialization code by replacing the inefficient Encoding.UTF8.GetBytes(GetRawText()) double-allocation pattern with JsonMarshal.GetRawUtf8Value() on net9.0+.

Category: Feature Request
Area: emitter:client:csharp

🔍 Analysis

• Affected component: http-client-csharp emitter — specifically the serialization code generation for nested model deserialization
• Root cause: The current codegen uses property.Value.GetRawText() (UTF-8 → UTF-16 string) followed by Encoding.UTF8.GetBytes(...) (UTF-16 string → UTF-8 bytes), performing two unnecessary transcoding allocations
• Proposed fix: Use a GetUtf8Bytes() extension method backed by JsonMarshal.GetRawUtf8Value(element) on net9.0+, which reads the raw UTF-8 span directly from the JsonDocument buffer — zero extra allocations
• Precedent: An equivalent extension method already exists in azure-sdk-for-net (JsonElementExtensions.cs), so the pattern is validated
• Impact: Reduced allocations and improved deserialization throughput for all generated models with nested model properties, on modern TFMs

📝 Suggested next steps

• Add or reference the GetUtf8Bytes extension method in the emitter's runtime helpers (with #if NET9_0_OR_GREATER guard)
• Update the serialization code generator to emit property.Value.GetUtf8Bytes() instead of new BinaryData(Encoding.UTF8.GetBytes(property.Value.GetRawText()))
• Add/update generated snapshot tests to reflect the new pattern

Generated by Agentic Triage