[NativeAOT-LLVM] Invalid Release code generation for `wasi-wasm`

[dicej]

I've been developing an AspNetCore example based on wasi-http. One of the URI handlers in that example uses System.Text.Json.Serialization and source generation to deserialize request bodies. When built using -c Debug, it all works fine. However, the -c Release build is incorrect, having replaced the code for deserializing a custom class with an infinite loop.

I've tried to reduce the issue to a minimal reproduction, e.g. as a CLI app with a Main method, but so far I've only managed to reproduce this issue in an AspNetCore web server. To reproduce:

• Clone this repo
• Install the prerequisites as indicated in the README.md of that repo
  • Note that you can skip installing Spin and use Wasmtime instead, which may make debugging easier. However, you'll need to build Wasmtime from source using this commit or later.
• Comment out this line in the example source code.
• Follow these instructions to build the example, but replace the spin build -u command with e.g. dotnet publish -o build App.csproj -c Release && wasmtime serve -O pooling-table-elements=60000 -O pooling-max-core-instance-size=2000000 -S cli --dir static::/ build/App.wasm
• In another terminal, use e.g. curl -i localhost:8080/data -H 'content-type: application/json' -d '{"Greeting":"Hola"}' to send a request to the problematic endpoint and note that the server loops forever without returning a response.
• Use e.g. wasm-tools print build/App.wasm | less to examine the Wasm file and search for App_SpinHttpWorld_wit_exports_wasi_http_v0_2_0_AppJsonSerializerContext__global__System_Text_Json_Serialization_Metadata_IJsonTypeInfoResolver_GetTypeInfo, noting the infinite loop generated by the compiler. Compare that with the correct, expected code.

Apologies for the complicated steps to reproduce this. Let me know if you have any trouble following them.

[dicej]

NB: I left out an important step above when I first opened the issue and just edited it in: "Comment out this line in the example source code."

[dicej]

@SingleAccretion helped narrow this down and find a solution. Turns out some of the warnings I was ignoring were important, and I needed <EnableRequestDelegateGenerator>true</EnableRequestDelegateGenerator>, which I had read about, but ignored because it's supposedly implied by <PublishTrimmed>true</PublishTrimmed>. Turns out that's not always the case. See the discussion following https://discord.com/channels/143867839282020352/1141126727028985877/1278474370842628116 for details.