Create managed version of llvm-objcopy

[jpobst]

Context: #9455

For each assembly/pdb/config/others file we wrap as a .so in order to place in the /lib directory of an APK, we shell out twice to llvm-objcopy:

llvm-objcopy.EXE --add-section payload=obj\Debug\net9.0-android\android\assets\arm64-v8a\myproject.dll obj\Debug\net9.0-android\android-arm64\wrapped\lib_myproject.dll.so
llvm-objcopy.EXE --set-section-flags payload=readonly,data --set-section-alignment payload=16384 obj\Debug\net9.0-android\android-arm64\wrapped\lib_myproject.dll.so

There is considerable overhead to this process.

In this example we spend ~42 seconds in this step:

dotnet new android
dotnet build -p:EmbedAssembliesIntoApk=true

A managed version that did not need to shell out to external processes of this would likely be faster.

Some potential options:

• https://github.com/xoofx/LibObjectFile (.NET 8+ 😢)
• Porting llvm-objcopy to managed code using managed LLVM bindings (potentially LLVMSharp)

[filipnavara]

Can you post an example of the input and output files? This seems to be simple enough operation that could likely be easily written by reusing the code from NativeAOT ELF ObjWriter.

[jpobst]

The inputs and outputs can be found in any .NET for Android project, ie:

• Inputs: obj\Debug\net10.0-android\android\assets\arm64-v8a\Microsoft.CSharp.dll
• Outputs: in the .apk: lib\arm64-v8a\lib_Microsoft.CSharp.dll.so

example.zip

[filipnavara]

Here's a rough implementation of the code to append the payload section to the template:
ElfAppendPayloadSection.zip

It compiles with both .NET 9.0 and .NET 4.8. Not everything is written as efficiently as it could be. I compared the output with the one you provided. It differs in the contents of .note.gnu.build-id section but I believe that's an artifact of not using the exact same template as the input.

[filipnavara]

Notably, you could also shave off some time if you create the payload section in the actual template (either through the same llvm-objcopy commands or a just adding it as assembly source code during the build). Then it would already have the right flags/alignment and you can reduce the number of invocations to half by using llvm-objcopy --update-section.

[jpobst]

can you, by any chance, share the libarchive-dso-stub.so that was used for
#9692 (comment)

It looks like these come from:

C:\Program Files\dotnet\packs\Microsoft.Android.Sdk.Windows\35.0.39\tools\dsostubs

Here is the android-arm64 one:

libarchive-dso-stub.zip

[filipnavara]

The build-id of libarchive-dso-stub.so differs for each build of the Android workload. The one you uploaded doesn't seem to match the previous build. That said, I run llvm-objcopy locally and verified that it doesn't do any rewriting of the build-id (unless told to do so). That makes the code linked above match the the llvm-objcopy output for this particular use case.

[jpobst]

A quick and dirty test of this code seems very promising:

Debug - EmbedAssembliesInApk - android

Scenario (WrapAssembliesAsSharedLibraries Task)	main	This PR
Full	40.6 s	14.37 s

Tested by:

• Converting ElfAppendPayloadSection.csproj to a netstandard2.0 .dll and adding it as a reference in Xamarin.Android.Build.Tasks.
• Applying patch.txt
• dotnet new android
• dotnet build -p:EmbedAssembliesIntoApk=true

[lukasjuhrich]

Out of curiosity, what causes the overhead? Is it just the subprocess spawning, or does llvm-objcopy do something in an inefficient way?

Thanks for taking the issue seriously by the way – this performance regression blocks our net9 upgrade.

[jpobst]

Out of curiosity, what causes the overhead? Is it just the subprocess spawning?

That is my hunch, yes. Spawning a process (on Windows at least) takes a non-trivial amount of time, and the default .NET for Android template references ~170 assemblies. Shelling out to llvm-objcopy at least twice per assembly is 340+ spawns, which add up.