Error in UTF8ToString when calling wgpuDeviceCreateShaderModule

[Ravbug]

When calling wgpuDeviceCreateShaderModule, occasionally the call will fail with
Invalid UTF-8 leading byte 0x00000090 encountered when deserializing a UTF-8 string in wasm memory to a JS string!

Code that's causing the error:

void ShaderLibraryWG::InitFromShaderSource(decltype(owningDevice) owningDevice, const std::string_view shadercode){
        WGPUShaderModuleWGSLDescriptor shaderCodeDesc;
        shaderCodeDesc.chain.next = nullptr;
        shaderCodeDesc.chain.sType = WGPUSType_ShaderModuleWGSLDescriptor;
        shaderCodeDesc.code = shadercode.data();
        WGPUShaderModuleDescriptor moduleDesc;
        moduleDesc.nextInChain = &shaderCodeDesc.chain;
        shaderModule = wgpuDeviceCreateShaderModule(owningDevice->device, &moduleDesc);
    }

The shader string:

diagnostic(off, derivative_uniformity);

var<private> outColor : vec4f;

var<private> fragColor : vec3f;

fn main_1() {
  let x_14 : vec3f = fragColor;
  outColor = vec4f(x_14.x, x_14.y, x_14.z, 1.0f);
  return;
}

struct main_out {
  @location(0)
  outColor_1 : vec4f,
}

@fragment
fn main(@location(0) fragColor_param : vec3f) -> main_out {
  fragColor = fragColor_param;
  main_1();
  return main_out(outColor);
}

Call stack:

Maybe the issue is related to #20124 and that the shader source string in WGPUShaderModuleWGSLDescriptor is non-owning? Sometimes the code works without issues.

Version of emscripten/emsdk:

emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 3.1.48-git (5b63268c904dd0e542f9c45e2b136c3d4a0bd638)
clang version 18.0.0 (https://github.com/llvm/llvm-project 25da11541c0740950e812a85d332a6d5cf4a5f87)
Target: wasm32-unknown-emscripten
Thread model: posix
InstalledDir: C:\Users\Admin\Documents\emsdk\upstream\bin

[kripken]

cc @kainino0x

[kainino0x]

The string in WGPUShaderModuleWGSLDescriptor is indeed non-owning. But it should certainly be valid for duration of the wgpuDeviceCreateShaderModule call, and proxying would block the calling thread to maintain that guarantee. I'm guessing that's what the 'sync' means, in:

      LibraryWebGPU[key + '__proxy'] = 'sync';

It seems more likely the issue is in your code. Actually, looking more closely, I think the problem is that WGPUShaderModuleWGSLDescriptor requires a null-terminated string, but std::string_view::data() does not return one. (You need std::string::c_str(), which isn't available on std::string_view.) So unless you're manually null-terminating your string before it's passed here, we (Emscripten) will read garbage past the end of your string.