WebGPU IPC texel-copy layout, adapter options and limits serialization (refactor)
Component: WebKit WebGPU IPC | cfba6db
Source/WebKit/Shared/WebGPU/WebGPURequestAdapterOptions.h
Source/WebKit/GPUProcess/graphics/WebGPU/RemoteGPU.cpp
WebGPU runs out-of-process: WebContent proxies send IPC to Remote* objects in the GPU process wrapping the real Metal-backed ::WebGPU:: API. Previously each descriptor had a parallel WebKit::WebGPU:: struct copied field-by-field with convertFromBacking, and for these types the converters always succeeded. ImageDataLayout, RequestAdapterOptions and SupportedLimits now alias ::WebGPU::TexelCopyBufferLayout, ::WebGPU::RequestAdapterOptions and ::WebGPU::Limits, decoded directly; three converter files are deleted, along with the dead failure path in RemoteGPU::requestAdapter and the layout conversion in RemoteQueue::writeTexture.
Significance
Three duplicate IPC structs and their converters disappear, so the wire types track the C++ API automatically while validation still rests on generated decoders and the backend. Behavior is preserved because the old converters were plain copies.
Audit directions
Narrow: these structs are now declared on the C++ API types with CustomHeader, and ImageCopyBuffer inherits ::WebGPU::TexelCopyBufferLayout in its .serialization.in; check that generated decoders still validate enum fields such as powerPreference and that ::WebGPU::Limits field types and widths match the old encoding. The offset + bytesPerRow * rows extent was never checked in the converter, so verify it in the backing writeTexture against the data span — an existing target, not a regression. Wider: confirm the backend handles every decodable requestAdapter option combination without release-compiled-out ASSERTs, and since any field added to Limits is now serialized automatically, check how requiredLimits flows through device creation (DeviceDescriptor.cpp was touched).