| 13 Dec 2024 |
 6pak | and even though it has the full library path, it registers the directory as a search path | 16:43:59 |
| 6pak | so if lib A has runtimes/linux-x64/native/liba.so and lib B has runtimes/linux-musl-x64/native/libb.so, both runtimes/linux-x64/native and runtimes/linux-musl-x64/native will be registered as native search paths | 16:47:02 |
 GGG | In reply to @6pak:matrix.org
so you could have just done JsonRuntimeFormat.ReadRuntimeGraph(<$(RuntimeIdentifierGraphPath) from msbuild>).ExpandRuntime("linux-x64") That's an interesting idea, however I wanted to have everything as optimized as possible, although I know most of the overhead will be in AsmResolver, I wanted to have little to no overhead on anything else so this doesn't slow down any builds
My long term objective with this is to use NativeAOT to have a really good cold start performance | 22:04:01 |
| GGG | Don't know if I'm overthinking this though | 22:04:14 |
| GGG | In reply to @6pak:matrix.org
and looking more into the runtimes/{rid}/native/{libname}.so pattern, seems like it's only a build-time nuget thing Yes, however my plan is to use autoPatchcil on the nuget restore step so nuget libraries with native tools and stuff work just fine
Long term I also want to find and transform any self-contained deployments into framework dependent ones so that they use the nixpkgs SDK/runtime instead of their own bundled one, that way we can avoid issues with ICU, openssl and libz | 22:06:56 |
| GGG | * Yes, however my plan is to use autoPatchcil on the nuget restore step so nuget libraries with native tools and stuff work just fine as they'll probably run on the build step (like the grpc package and entity framework)
Long term I also want to find and transform any self-contained deployments into framework dependent ones so that they use the nixpkgs SDK/runtime instead of their own bundled one, that way we can avoid issues with ICU, openssl and libz | 22:07:42 |
| GGG | In reply to @6pak:matrix.org
the runtime only looks at the .deps.json file to discover the library paths, which dotnet sdk populates from nuget info That's good to know, I'll look into it and try to add that logic into the command | 22:08:50 |
| GGG | Also, while running it in a dry run against avalonia, I ran into something rather unpleasant | 22:11:45 |
| GGG | There's `.dll` and `.dylib` imports in it as well, so it won't be as plug'n'play as `autoPatchelfHook`, I had to add quite a few thing to the ignores list like `*.dll`, `kernel32`, `user32` and more | 22:13:17 |
| GGG | It might be worth adding an option to exclude certain files on top of ignoring certain imports | 22:13:51 |
| 14 Dec 2024 |
| 6pak | corehost parses the json every launch so it would probably be fine? | 01:34:55 |
| 6pak | I meant that the runtime doesn't look at the runtimes/{rid}/native path, it's the nuget client that checks it in the packages and converts it into metadata that the .net sdk then puts into the final .deps.json | 01:37:31 |
| 6pak | but build time nuget packages requiring native libs themselves is a valid point | 01:38:40 |
| 6pak | wouldn't it make more sense to run (auto)patchcil as part of nuget overrides though? | 01:38:57 |
| 6pak | yeahh thats fair | 01:41:31 |
| 6pak | it would be nice if people annotated those with SupportedOSPlatformAttribute, but they probably don't | 01:41:54 |
| GGG | In reply to @6pak:matrix.org
corehost parses the json every launch so it would probably be fine? I mean, it probably wouldn't be much overhead, I'm probably just overthinking it, but since it's already done, I wouldn't just throw it away myself | 01:43:21 |
| GGG | In reply to @6pak:matrix.org
I meant that the runtime doesn't look at the runtimes/{rid}/native path, it's the nuget client that checks it in the packages and converts it into metadata that the .net sdk then puts into the final .deps.json Oh, I see, that's indeed a big issue, I'll have to remove those lines from the logic since it's nuget only | 01:44:19 |
| 6pak | it just so happens that it's the same path in the output usually | 01:44:49 |
| 6pak | but in theory the .deps.json could specify any path | 01:44:58 |
| GGG | In reply to @6pak:matrix.org
wouldn't it make more sense to run (auto)patchcil as part of nuget overrides though? Well, thinking about it, I don't think the tools in a nuget package use the native libraries in the runtime dir, they're probably packaged as a normal command line tool | 01:46:27 |
| GGG | Or there would be something in the `.deps.json` file pointing at it, so we shouldn't need any other special logic other than handling that file. | 01:47:03 |
| GGG | In reply to @6pak:matrix.org
it would be nice if people annotated those with SupportedOSPlatformAttribute, but they probably don't They'd probably only annotate the methods that call those pinvoke shims instead of the shims themselves, so worst case it'd need some blackbox control flow analysis which would most likely fail due to reflection, so I don't think it's possible | 01:48:11 |
| GGG | Best that can be done is exclude some extensions based on the target RID most likely | 01:48:26 |
| GGG | Like, if it's `linux-x64` we know it won't be needing any `.dll`s nor `.dylib`s | 01:49:05 |
| 6pak | you don't need to include the extension in dllimport library name though | 01:49:28 |
| GGG | True, but a few do, so we could save some headache by excluding those beforehand | 01:49:55 |
| GGG | I saw a dllimport with kernel32.dll in avalonia, so if we auto excluded those it'd already save the need to manually specify the exclude globs based on the target platform on the nix side | 01:50:55 |
| GGG | It was all fine and dandy when I was only targeting the Avalonia.X11 assembly, but once I targeted the whole directory avalonia-ilspy outputs then it became a huge mess | 01:53:22 |
| 6pak | actually, can .deps.json take absolute paths for managed dlls? | 01:56:43 |
