Reloc failures with NativeAOT on Apple Silicon

[am11]

I am trying to enable NativeAOT on OSX arm64. With this patch main...am11:feature/nativeaot/osx-arm64 (tested with both @GOTPAGE and @PAGE assembler directives), it builds the nupkg. Consuming that package results in the following errors during the ilc step:

# with `<add key="TestSource" value="/Users/am11/projects/runtime/artifacts/packages/Release/Shipping" />`
# in NuGet.config
$ dotnet nuget locals all --clear && rm -rf obj bin && dotnet publish --use-current-runtime -v:diag ...
... snip ...
21:06:05.007   1:7>Target "IlcCompile: (TargetId:181)" in file "/Users/am11/.nuget/packages/microsoft.dotnet.ilcompiler/7.0.0-dev/build/Microsoft.NETCore.Native.targets" from project "/Users/am11/projects/naot1/naot1.csproj" (target "LinkNative" depends on it):
                   Building target "IlcCompile" completely.
                   Output file "obj/release/net7.0/osx-arm64/native/naot1.o" does not exist.
                   Task "Message" skipped, due to false condition; ($(_BuildingInCompatibleMode) != 'true') was evaluated as (true != 'true').
                   Task "Message" (TaskId:126)
                     Task Parameter:Text=Generating compatible native code. To optimize for size or speed, visit https://aka.ms/OptimizeCoreRT (TaskId:126)
                     Task Parameter:Importance=high (TaskId:126)
                     Generating compatible native code. To optimize for size or speed, visit https://aka.ms/OptimizeCoreRT (TaskId:126)
                   Done executing task "Message". (TaskId:126)
                   Task "Exec" (TaskId:127)
                     Task Parameter:Command="/Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/tools/ilc" @"obj/release/net7.0/osx-arm64/native/naot1.ilc.rsp" (TaskId:127)
                     "/Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/tools/ilc" @"obj/release/net7.0/osx-arm64/native/naot1.ilc.rsp" (TaskId:127)
                     <unknown>:0: error: ADR/ADRP relocations must be GOT relative (TaskId:127)
                     <unknown>:0: error: unknown AArch64 fixup kind! (TaskId:127)
                     <unknown>:0: error: unknown AArch64 fixup kind! (TaskId:127)
                     <unknown>:0: error: fixup value out of range (TaskId:127)
                     <unknown>:0: error: ADR/ADRP relocations must be GOT relative (TaskId:127)
                     <unknown>:0: error: unknown AArch64 fixup kind! (TaskId:127)
                     <unknown>:0: error: unknown AArch64 fixup kind! (TaskId:127)
                     <unknown>:0: error: fixup value out of range (TaskId:127)
... repeats 1000s of times ...

somewhere after the objwriter has succeeded:

runtime/src/coreclr/tools/aot/ILCompiler.Compiler/Compiler/DependencyAnalysis/ObjectWriter.cs

Line 1183 in 071e772

succeeded = true;

and before the clang command is executed. While the ilc task does not fail, MSBuild fails on the clang step:

                 Set Property: _IgnoreLinkerWarnings=false
                   Set Property: _IgnoreLinkerWarnings=true
                   Task "Exec" (TaskId:129)
                     Task Parameter:IgnoreStandardErrorWarningFormat=True (TaskId:129)
                     Task Parameter:Command=clang "obj/release/net7.0/osx-arm64/native/naot1.o" -o "bin/release/net7.0/osx-arm64/native/naot1" /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libbootstrapper.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libRuntime.WorkstationGC.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Globalization.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.IO.Compression.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Net.Security.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Security.Cryptography.Native.Apple.a -g -Wl,-rpath,'@executable_path' -lstdc++ -ldl -lm -lz -licucore -framework CoreFoundation -framework Foundation -framework Security -framework GSS (TaskId:129)
                     clang "obj/release/net7.0/osx-arm64/native/naot1.o" -o "bin/release/net7.0/osx-arm64/native/naot1" /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libbootstrapper.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libRuntime.WorkstationGC.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Globalization.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.IO.Compression.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Net.Security.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Security.Cryptography.Native.Apple.a -g -Wl,-rpath,'@executable_path' -lstdc++ -ldl -lm -lz -licucore -framework CoreFoundation -framework Foundation -framework Security -framework GSS (TaskId:129)
                     ld: malformed __LD,__compact_unwind section, bad length file 'obj/release/net7.0/osx-arm64/native/naot1.o' (TaskId:129)
                     clang: error: linker command failed with exit code 1 (use -v to see invocation) (TaskId:129)
21:06:12.873   1:7>/Users/am11/.nuget/packages/microsoft.dotnet.ilcompiler/7.0.0-dev/build/Microsoft.NETCore.Native.targets(337,5): error MSB3073: The command "clang "obj/release/net7.0/osx-arm64/native/naot1.o" -o "bin/release/net7.0/osx-arm64/native/naot1" /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libbootstrapper.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libRuntime.WorkstationGC.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Globalization.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.IO.Compression.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Net.Security.Native.a /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Security.Cryptography.Native.Apple.a -g -Wl,-rpath,'@executable_path' -lstdc++ -ldl -lm -lz -licucore -framework CoreFoundation -framework Foundation -framework Security -framework GSS" exited with code 1. [/Users/am11/projects/naot1/naot1.csproj]
                   Done executing task "Exec" -- FAILED. (TaskId:129)
21:06:12.873   1:7>Done building target "LinkNative" in project "naot1.csproj" -- FAILED.: (TargetId:182)

With objdump, that __LD,__compact_unwind section looks like:

Disassembly of section __LD,__compact_unwind:

00000000003b2858 <ltmp8>:
  3b2858: 40 4b 00 00   udf     #19264
  3b285c: 00 00 00 00   udf     #0
  3b2860: 74 00 00 00   udf     #116
  3b2864: 00 00 00 03   <unknown>
                ...
  3b2874: c0 4b 00 00   udf     #19392
  3b2878: 00 00 00 00   udf     #0
  3b287c: 74 00 00 00   udf     #116
  3b2880: 00 00 00 03   <unknown>
                ...
  3b2890: 40 4c 00 00   udf     #19520
  3b2894: 00 00 00 00   udf     #0
  3b2898: 74 00 00 00   udf     #116
  3b289c: 00 00 00 03   <unknown>
                ...
  3b28ac: c0 4c 00 00   udf     #19648
  3b28b0: 00 00 00 00   udf     #0
  3b28b4: 74 00 00 00   udf     #116
  3b28b8: 00 00 00 03   <unknown>
... repeats ...

[MichalStrehovsky]

Ah, so those messages are still generated by the object writer in ILC.

For example here: https://github.com/dotnet/llvm-project/blob/f1120a92d05f1c57e75af7d16504012570ef3409/llvm/lib/Target/AArch64/MCTargetDesc/AArch64MachObjectWriter.cpp#L102-L103.

Looks like we need to decide what kind of relocation to generate when we're generating it.

[am11]

@MichalStrehovsky, thanks for the pointers. The managed part of objwriter API currently does not support VariantKind, so I updated objwriter: dotnet/llvm-project#185. With the current state of that PR (+ main...am11:feature/nativeaot/osx-arm64), these two ilc errors are vanished:

error: fixup value out of range
error: ADR/ADRP relocations must be GOT relative 

but this one remains:

error: unknown AArch64 fixup kind!

it is false return from https://github.com/dotnet/llvm-project/blob/f1120a92d05f1c57e75af7d16504012570ef3409/llvm/lib/Target/AArch64/MCTargetDesc/AArch64MachObjectWriter.cpp#L59. I noticed that in objwriter, we do not explicitly create a MCFixup for mach, looks like those are classified by AArch64AsmBackend. 💭

[MichalStrehovsky]

Really nice progress!

What is the reloc it's complaining about? Is it FK_PCRel_4 by any chance? We add special handling to it in our LLVM patch: dotnet/llvm-project@67f5503. Maybe having it take the same path as FK_Data_4 would work?

You'll know if the reloc got messed up if it ends up pointing at garbage after linking.

[am11]

Those were indeed all FK_PCRel_4. 😄

Pushed a commit to treat it as FK_Data_4. With that ilc msbuild task completes without producing any diagnostics.

clang task, however, continues to fail (still complaining about __LD,__compact_unwind section, bad length file). Interestingly, now the __LD,__compact_unwind section has disappeared when I run with objdump -d (as it was showing previously), but it shows up with objdump --full-contents -d.

Running ld command with -v shows nothing useful:

 % "/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/ld" \
       -demangle -lto_library /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/libLTO.dylib \
       -dynamic -arch arm64 -platform_version macos 12.0.0 12.1 \
       -syslibroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk \
       -o bin/release/net7.0/osx-arm64/native/naot1 \
       -L/usr/local/lib obj/release/net7.0/osx-arm64/native/naot1.o \
       /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libbootstrapper.a \
       /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/sdk/libRuntime.WorkstationGC.a \
       /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Native.a \
       /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Globalization.Native.a \
       /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.IO.Compression.Native.a \
       /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Net.Security.Native.a \
       /Users/am11/.nuget/packages/runtime.osx-arm64.microsoft.dotnet.ilcompiler/7.0.0-dev/framework/libSystem.Security.Cryptography.Native.Apple.a \
       -rpath @executable_path -lc++ -ldl -lm -lz -licucore \
       -framework CoreFoundation -framework Foundation -framework Security -framework GSS \
       -lSystem /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/13.0.0/lib/darwin/libclang_rt.osx.a \
       -v

@(#)PROGRAM:ld  PROJECT:ld64-711
BUILD 21:57:24 Nov 17 2021
configured to support archs: armv6 armv7 armv7s arm64 arm64e arm64_32 i386 x86_64 x86_64h armv6m armv7k armv7m armv7em
Library search paths:
	/usr/local/lib
	/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/lib
Framework search paths:
	/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/
ld: malformed __LD,__compact_unwind section, bad length file 'obj/release/net7.0/osx-arm64/native/naot1.o'

[MichalStrehovsky]

Digging in Apple's source code, the error seems to be:

https://github.com/apple-oss-distributions/ld64/blob/dbf8f7feb5579761f1623b004bd468bdea7c6225/src/ld/parsers/macho_relocatable_file.cpp#L5631-L5632

the size of __compact_unwind section is not divisible by the size of the unwind entry. That's odd because we don't generate the Apple weird thing, we generate DWARF CFI.

However, looking at LLVM source code, I think this is kicking in:

https://github.com/dotnet/llvm-project/blob/f1120a92d05f1c57e75af7d16504012570ef3409/llvm/lib/MC/MCObjectFileInfo.cpp#L30-L32

And LLVM does generate something on our behalf. Probably broken from the sound of it.

I would dig around that - can we still make do without a __compact_unwind section? If it's not present in the executable, maybe ld would still do the right thing and convert it from CFI to the compact unwind scheme for us.

Maybe the right thing would be to start generating compact unwinding because Apple tends to unceremoniously cut off things they don't like anymore after a couple years of supporting both the thing they stopped liking and the new shiny thing. Unwinding codes are currently generated in RyuJIT.

[am11]

Thanks, and nice detective work finding those links! (I was only searching across org:dotnet 😅)

Disabling compact unwind (dotnet/llvm-project@090a465) revealed some missing symbols errors. Those error messages were generous enough to point me to the missing C_FUNC(): am11@3bcd7b6c (I previously fixed couple of those in am11@b8cc922 which were failing the runtime build, but the new ones only show up when consuming Microsoft.DotNet.ILCompiler package 🤷)

Looks like we are getting there. Next error is:

  ld: in section __DATA,.corert_eh_table reloc 0: unknown relocation type 15 file 'obj/release/net7.0/osx-arm64/native/naot1.o'

Apparently, reloc 15 is assigned to PPC_RELOC_LOCAL_SECTDIFF (suggesting something is off the kilter.. 👀).

[MichalStrehovsky]

.corert_eh_table uses the 32-bit relative relocs that would be translated to FK_PCRel_4 I think - so something might be going wrong around that.

Can you try updating this:

runtime/src/coreclr/tools/Common/TypeSystem/Common/TargetDetails.cs

Lines 94 to 100 in 59bb459

	public bool SupportsRelativePointers
	{
	get
	{
	return (Abi != TargetAbi.CppCodegen) && (Architecture != TargetArchitecture.Wasm32);
	}
	}

and make it so that it returns false for ARM64 macOS? (use Architecture and OperatingSystem that is conveniently available on the class).

This will make the compiler avoid generating 32bit relative relocations in favor of full pointers. It will make the executables a bit bigger. We would want to try find a way to do 32bit relative relocs eventually, but let's first find out if that's really the problem.

[am11]

Having SupportsRelativePointers to return false had no effect, getting the ditto ld error. :(

[MichalStrehovsky]

Ah, there's an extra code path that is not active for Wasm or CppCodegen (for which SupportsRelativePointers was introduced) - replace IMAGE_REL_BASED_RELPTR32 with IMAGE_REL_BASED_DIR64 in src\coreclr\tools\aot\ILCompiler.Compiler\Compiler\DependencyAnalysis\ObjectWriter.cs. This is nothing but a hack (it's going to crash if we take a GC or exception at runtime), but it might help narrowing down the problem/make further progress.