.NET 7 exception on Android with EnableLLVM=true #73003

jonathanpeppers · 2022-07-28T14:35:05Z

Description

Run with LLVM on an arm64 device, and you can run into:

07-22 09:59:55.062  4374  4374 F mono-rt : [ERROR] FATAL UNHANDLED EXCEPTION: System.PlatformNotSupportedException: Arg_PlatformNotSupported
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.ASCIIUtility.GetIndexOfFirstNonAsciiByte_Intrinsified(Byte* , UIntPtr )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.Unicode.Utf8Utility.GetPointerToFirstInvalidByte(Byte* , Int32 , Int32& , Int32& )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.UTF8Encoding.GetCharCount(Byte* , Int32 )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.String.CreateStringFromEncoding(Byte* , Int32 , Encoding )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.Encoding.GetString(Byte* , Int32 )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.String.CreateStringForSByteConstructor(Byte* , Int32 )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.String.Ctor(SByte* value)
07-22 09:59:55.062  4374  4374 F mono-rt :    at Java.Interop.TypeManager.GetClassName(IntPtr )
07-22 09:59:55.062  4374  4374 F mono-rt :    at Android.Runtime.JNIEnv.RegisterJniNatives(IntPtr , Int32 , IntPtr , IntPtr , Int32 )

Somewhere here:

runtime/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs

Line 88 in b31a48a

    
           return (Sse2.IsSupported || (AdvSimd.Arm64.IsSupported && BitConverter.IsLittleEndian))

So I think AdvSimd.Arm64.IsSupported is true, but because GetIndexOfFirstNonAsciiByte_Intrinsified() is not in the AOT profile (it's running under JIT), we get the above exception. It's possible that the above code should also work under JIT.

/cc @fanyang-mono

Reproduction Steps

dotnet new android
dotnet build -t:Run -c Release -p:EnableLLVM=true to run on an arm64 device (in my case a Pixel 5)

Expected behavior

You can use Profiled AOT + LLVM together on Android.

Actual behavior

We get an exception when trying to use Profiled AOT + LLVM together on Android.

Regression?

I don't hit this same exception in .NET 6, but maybe it's just luck?

Known Workarounds

You can turn off profiled AOT, such as -p:AndroidEnableProfiledAot=false or don't use LLVM.

Configuration

.NET 7.0.100-rc.1.22374.1

Other information

No response

The text was updated successfully, but these errors were encountered:

dotnet-issue-labeler · 2022-07-28T14:35:15Z

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

jonathanpeppers · 2022-07-28T14:38:26Z

We also tried passing -O=-simd to the AOT compiler, but ran into a different issue.

But that might be solved by: #72982

ghost · 2022-07-28T17:41:36Z

Tagging subscribers to 'arch-android': @steveisok, @akoeplinger
See info in area-owners.md if you want to be subscribed.

Issue Details

Description

Run with LLVM on an arm64 device, and you can run into:

07-22 09:59:55.062  4374  4374 F mono-rt : [ERROR] FATAL UNHANDLED EXCEPTION: System.PlatformNotSupportedException: Arg_PlatformNotSupported
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.ASCIIUtility.GetIndexOfFirstNonAsciiByte_Intrinsified(Byte* , UIntPtr )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.Unicode.Utf8Utility.GetPointerToFirstInvalidByte(Byte* , Int32 , Int32& , Int32& )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.UTF8Encoding.GetCharCount(Byte* , Int32 )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.String.CreateStringFromEncoding(Byte* , Int32 , Encoding )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.Text.Encoding.GetString(Byte* , Int32 )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.String.CreateStringForSByteConstructor(Byte* , Int32 )
07-22 09:59:55.062  4374  4374 F mono-rt :    at System.String.Ctor(SByte* value)
07-22 09:59:55.062  4374  4374 F mono-rt :    at Java.Interop.TypeManager.GetClassName(IntPtr )
07-22 09:59:55.062  4374  4374 F mono-rt :    at Android.Runtime.JNIEnv.RegisterJniNatives(IntPtr , Int32 , IntPtr , IntPtr , Int32 )

Somewhere here:

runtime/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs

Line 88 in b31a48a

    
           return (Sse2.IsSupported || (AdvSimd.Arm64.IsSupported && BitConverter.IsLittleEndian))

So I think AdvSimd.Arm64.IsSupported is true, but because GetIndexOfFirstNonAsciiByte_Intrinsified() is not in the AOT profile (it's running under JIT), we get the above exception. It's possible that the above code should also work under JIT.

/cc @fanyang-mono

Reproduction Steps

dotnet new android
dotnet build -t:Run -c Release -p:EnableLLVM=true to run on an arm64 device (in my case a Pixel 5)

Expected behavior

You can use Profiled AOT + LLVM together on Android.

Actual behavior

We get an exception when trying to use Profiled AOT + LLVM together on Android.

Regression?

I don't hit this same exception in .NET 6, but maybe it's just luck?

Known Workarounds

You can turn off profiled AOT, such as -p:AndroidEnableProfiledAot=false or don't use LLVM.

Configuration

.NET 7.0.100-rc.1.22374.1

Other information

No response

Author:	jonathanpeppers
Assignees:	-
Labels:	`os-android`, `untriaged`, `area-Codegen-AOT-mono`
Milestone:	-

steveisok · 2022-07-28T17:43:07Z

@vargaz Will #72982 solve this?

vargaz · 2022-07-28T17:46:40Z

Hopefully.

steveisok · 2022-07-28T18:11:42Z

@jonathanpeppers Any chance you can help us validate?

jonathanpeppers · 2022-07-28T22:22:57Z

I triggered Maestro just now, but I don't see the change there yet: dotnet/android#7217

Maybe we can try it tomorrow.

fanyang-mono · 2022-08-02T14:20:57Z

This could be a dup of #60792

`dotnet new android` apps would crash on startup when built with `-c Release -p:EnableLLVM=true`: 07-20 08:55:44.642 2983 2983 F monodroid-assembly: Internal p/invoke symbol 'java-interop @ java_interop_jvm_list' (hash: 0x58c48fc8b89cb484) not found in compile-time map. ... 07-20 08:55:44.834 3004 3004 F DEBUG : signal 6 (SIGABRT), code -1 (SI_QUEUE), fault addr -------- 07-20 08:55:44.834 3004 3004 F DEBUG : Abort message: 'Internal p/invoke symbol 'java-interop @ java_interop_jvm_list' (hash: 0x58c48fc8b89cb484) not found in compile-time map.' `java_interop_jvm_list` should *never* be called on Android, it is the result of `new AndroidRuntime()` having not been called yet? The problem being that the static `Android.App.Application.cctor` was somehow running *before* `JNIEnv.Initialize()` was complete? And this only happens with LLVM? Reviewing the code: [UnmanagedCallersOnly] internal static unsafe void Initialize (JnienvInitializeArgs* args) { //... SynchronizationContext.SetSynchronizationContext (Android.App.Application.SynchronizationContext); } We do indeed access `Android.App.Application`... To fix this, we can move this to a new method decorated with `MethodImplOptions.NoInlining`. Apps built with LLVM now launch for me. However, we can't really enable the LLVM `Mono.Android-Tests` quite yet. Read on! ~~ Known Issues ~~ * dotnet/runtime#73003 Using Profiled AOT and LLVM at the same time can hit: [ERROR] FATAL UNHANDLED EXCEPTION: System.PlatformNotSupportedException: Arg_PlatformNotSupported at System.Text.ASCIIUtility.GetIndexOfFirstNonAsciiByte_Intrinsified(Byte* , UIntPtr ) at System.Text.Unicode.Utf8Utility.GetPointerToFirstInvalidByte(Byte* , Int32 , Int32& , Int32& ) at System.Text.UTF8Encoding.GetCharCount(Byte* , Int32 ) at System.String.CreateStringFromEncoding(Byte* , Int32 , Encoding ) at System.Text.Encoding.GetString(Byte* , Int32 ) at System.String.CreateStringForSByteConstructor(Byte* , Int32 ) at System.String.Ctor(SByte* value) at Java.Interop.TypeManager.GetClassName(IntPtr ) at Android.Runtime.JNIEnv.RegisterJniNatives(IntPtr , Int32 , IntPtr , IntPtr , Int32 ) For now we can set `AndroidEnableProfiledAot=false` for the LLVM tests. * dotnet/runtime#73304 Some Java.Interop unit tests fail with: System.InvalidOperationException : InvalidOperation_EnumEnded at System.ArrayEnumerator.get_Current() at Cadenza.Collections.Tests.CollectionContract`1[[System.Int16, System.Private.CoreLib, Version=7.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].DisposeCollection(IEnumerable c) at Cadenza.Collections.Tests.CollectionContract`1[[System.Int16, System.Private.CoreLib, Version=7.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].CopyTo() at System.Reflection.MethodInvoker.InterpretedInvoke(Object , Span`1 , BindingFlags ) * https://github.com/xamarin/java.interop/blob/a3de91efdaec04b3927b755a07018e8d26cfdc71/tests/Java.Interop-Tests/Cadenza.Collections/CollectionContract.cs#L205-L221 * https://github.com/xamarin/java.interop/blob/a3de91efdaec04b3927b755a07018e8d26cfdc71/tests/Java.Interop-Tests/Cadenza.Collections/CollectionContract.cs#L85-L86 The `InetAccess` category causes the `Mono.Android-Tests` suite to crash with: 08-02 21:26:57.581 5154 5173 I NUnit : System.NetTests.ProxyTest : 646.269 ms 08-02 21:26:57.593 5154 5173 I droid.NET_Test: Explicit concurrent copying GC freed 20(95KB) AllocSpace objects, 0(0B) LOS objects, 59% free, 1054KB/2590KB, paused 1.469ms total 7.180ms 08-02 21:26:57.594 5154 5173 I NUnit : SslTest 08-02 21:26:57.594 5154 5173 I NUnit : HttpsShouldWork 08-02 21:26:57.858 5154 5181 D NetworkSecurityConfig: No Network Security Config specified, using platform default 08-02 21:26:58.585 5154 5164 I droid.NET_Test: Background young concurrent copying GC freed 39658(2920KB) AllocSpace objects, 0(0B) LOS objects, 22% free, 1999KB/2590KB, paused 69.607ms total 194.902ms 08-02 21:26:58.922 5154 5164 I droid.NET_Test: Background concurrent copying GC freed 12412(836KB) AllocSpace objects, 0(0B) LOS objects, 46% free, 6987KB/12MB, paused 2.127ms total 327.448ms 08-02 21:26:58.953 5154 5181 I monodroid-timing: [1/0] LZ4 decompression time for <assembly_store>; elapsed: 0:0::75000 08-02 21:26:58.954 5154 5181 W linker : Warning: "/data/app/Mono.Android.NET_Tests-QcXozuJGB3EpfK6bvDvvMQ==/split_config.x86_64.apk!/lib/x86_64/libaot-System.Formats.Asn1.dll.so" has no DT_SONAME (will use libaot-System.Formats.Asn1.dll.so instead) and will not work when the app moves to API level 23 or later (https://android.googlesource.com/platform/bionic/+/master/missing-soname-enforced-for-api-level-23) (allowing for now because this app's target API level is still 21) 08-02 21:26:59.037 5154 5181 I monodroid-timing: [1/0] LZ4 decompression time for <assembly_store>; elapsed: 0:0::129000 08-02 21:26:59.038 5154 5181 W linker : Warning: "/data/app/Mono.Android.NET_Tests-QcXozuJGB3EpfK6bvDvvMQ==/split_config.x86_64.apk!/lib/x86_64/libaot-System.Runtime.Numerics.dll.so" has no DT_SONAME (will use libaot-System.Runtime.Numerics.dll.so instead) and will not work when the app moves to API level 23 or later (https://android.googlesource.com/platform/bionic/+/master/missing-soname-enforced-for-api-level-23) (allowing for now because this app's target API level is still 21) 08-02 21:26:56.680 5154 5154 W Instrumentation: type=1400 audit(0.0:432): avc: denied { search } for name="run" dev="vdc" ino=32774 scontext=u:r:untrusted_app_25:s0:c512,c768 tcontext=u:object_r:varrun_file:s0 tclass=dir permissive=0 app=Mono.Android.NET_Tests 08-02 21:26:56.770 5154 5154 W Instrumentation: type=1400 audit(0.0:433): avc: denied { ioctl } for path="socket:[49468]" dev="sockfs" ino=49468 ioctlcmd=0x8946 scontext=u:r:untrusted_app_25:s0:c512,c768 tcontext=u:r:untrusted_app_25:s0:c512,c768 tclass=udp_socket permissive=0 app=Mono.Android.NET_Tests 08-02 21:26:56.780 5154 5154 W Instrumentation: type=1400 audit(0.0:434): avc: denied { ioctl } for path="socket:[49468]" dev="sockfs" ino=49468 ioctlcmd=0x8946 scontext=u:r:untrusted_app_25:s0:c512,c768 tcontext=u:r:untrusted_app_25:s0:c512,c768 tclass=udp_socket permissive=0 app=Mono.Android.NET_Tests 08-02 21:27:00.029 1540 1540 I Zygote : Process 5154 exited due to signal 11 (Segmentation fault) Not really any useful info here... Disabled this category for now. We can file an issue for this one, assuming it is different than the other issues.

lambdageek · 2022-08-17T19:47:23Z

this is a nasty problem.

so the issue is.

Android uses profiled AOT pretty much by default now - for fast startup
There are functions that don't get AOTed (they're not early enough or they weren't used in the profile that Android uses)
when you turn on LLVM, the method that makes the decision "should we support intrinsics" gets inlined into something that is part of the AOT profile, so AdvSimd.IsSupported is hardcoded to "true"

runtime/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs

Line 81 in aae6e9b

public static unsafe nuint GetIndexOfFirstNonAsciiByte(byte* pBuffer, nuint bufferLength)
But at runtime the actual method that does the work is JITed, so when it checks AdvSimd.IsSupported it gets false

runtime/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs

Line 294 in aae6e9b

else if (AdvSimd.Arm64.IsSupported)

I believe the fix is that when we check for "jit_supported" instead of just checking "are we using LLVM", we need to check "and we don't have the JIT as a fallback". in other words, we have to be in FullAOT or HybridAOT mode.

runtime/src/mono/mono/mini/simd-intrinsics.c

Lines 721 to 723 in 8027c80

    
           // Hardware intrinsics are LLVM-only. 
        
           if (!COMPILE_LLVM (cfg) && !intrin_group->jit_supported) 
        
           	goto support_probe_complete;

In other words: profiled AOT cannot benefit from the arm64 intrinsics work we've done for LLVM.

akoeplinger · 2022-08-17T19:50:57Z

not a fix for the underlying issue but could we somehow force the intrinsics APIs to be included in the profile so that we still get the LLVM intrinsics handling by default?

lambdageek · 2022-08-17T19:54:50Z

not a fix for the underlying issue but could we somehow force the intrinsics APIs to be included in the profile so that we still get the LLVM intrinsics handling by default?

I'm not sure I understand the suggestion.

The issue is: the code that decides if we should use an optimized version of an algorithm sees different IsSupported values than the code that implements the specialized version of the algorithm.

The code that decides is compiled with LLVM AOT and therefore IsSupported is true. The code that implements the algorithm is JITed and locally sees IsSupported as false.

lambdageek · 2022-08-17T19:57:02Z

Basically the "fallback" execution engines need to return IsSupported == true if any accelerted execution engine returns IsSupported == true. Otherwise code is unsafe.

fanyang-mono · 2022-08-17T20:41:06Z

The reason why Android is seeing this is because Android uses this setup: LLVM AOT with non-LLVM JIT fallback. Currently, AdvSimd.IsSupported only returns true, when LLVM is enabled, no matter it is AOT or JIT.

akoeplinger · 2022-08-18T10:32:46Z

The issue is: the code that decides if we should use an optimized version of an algorithm sees different IsSupported values than the code that implements the specialized version of the algorithm.

My suggestion was to make sure we always LLVM AOT the implementation, regardless of whether it is used by the profile, so that no fallback to the JIT occurs.

lambdageek · 2022-08-18T11:27:40Z

The issue is: the code that decides if we should use an optimized version of an algorithm sees different IsSupported values than the code that implements the specialized version of the algorithm.

My suggestion was to make sure we always LLVM AOT the implementation, regardless of whether it is used by the profile, so that no fallback to the JIT occurs.

That's hard for a number of reasons:

We don't know what "the implementation" is. In this case it's called after the IsSupported check, but in general the decision about which algorithm to use could be far removed from the place where the intrinsics get used.
The implementation method could use something that we don't know how to AOT like exception filters.

If we had some attribute [ForceAOT(AdvSimd.IsSupported)] then maybe we could address (1), and then (2) could be an error and we could educate users to only use the supported constructs or refactor their code

ghost added the untriaged New issue has not been triaged by the area owner label Jul 28, 2022

jonathanpeppers added the area-Codegen-AOT-mono label Jul 28, 2022

jonathanpeppers changed the title ~~.NET 7 crash on Android with EnableLLVM=true~~ .NET 7 exception on Android with EnableLLVM=true Jul 28, 2022

lewing added the os-android label Jul 28, 2022

steveisok removed the untriaged New issue has not been triaged by the area owner label Jul 28, 2022

steveisok added this to the 7.0.0 milestone Jul 28, 2022

steveisok assigned vargaz Jul 28, 2022

jonathanpeppers mentioned this issue Aug 3, 2022

[Mono.Android] fix crash on startup with EnableLLVM dotnet/android#7188

Merged

SamMonoRT added the blocking-release label Aug 10, 2022

jonathanpeppers mentioned this issue Aug 17, 2022

AOT+LLVM throws for Enum.GetUnderlyingType() on Android in .NET 7 #68914

Closed

lambdageek self-assigned this Aug 17, 2022

lambdageek mentioned this issue Aug 17, 2022

SIMD problems when mixing llvm and non-llvm code #73454

Closed

fanyang-mono mentioned this issue Aug 20, 2022

[Mono] Only enable Arm intrinsics for full aot mode #74301

Merged

ghost added the in-pr There is an active PR which will close this issue when it is merged label Aug 20, 2022

lambdageek mentioned this issue Aug 22, 2022

[release/7.0-rc1] [Mono] Only enable Arm intrinsics for full aot mode #74357

Merged

SamMonoRT closed this as completed in #74301 Aug 23, 2022

ghost removed the in-pr There is an active PR which will close this issue when it is merged label Aug 23, 2022

ghost locked as resolved and limited conversation to collaborators Sep 22, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

.NET 7 exception on Android with EnableLLVM=true #73003

.NET 7 exception on Android with EnableLLVM=true #73003

jonathanpeppers commented Jul 28, 2022 •

edited

Loading

dotnet-issue-labeler bot commented Jul 28, 2022

jonathanpeppers commented Jul 28, 2022

ghost commented Jul 28, 2022

Description

Reproduction Steps

Expected behavior

Actual behavior

Regression?

Known Workarounds

Configuration

Other information

steveisok commented Jul 28, 2022

vargaz commented Jul 28, 2022

steveisok commented Jul 28, 2022

jonathanpeppers commented Jul 28, 2022

fanyang-mono commented Aug 2, 2022

lambdageek commented Aug 17, 2022

akoeplinger commented Aug 17, 2022

lambdageek commented Aug 17, 2022

lambdageek commented Aug 17, 2022

fanyang-mono commented Aug 17, 2022

akoeplinger commented Aug 18, 2022

lambdageek commented Aug 18, 2022

.NET 7 exception on Android with EnableLLVM=true #73003

.NET 7 exception on Android with EnableLLVM=true #73003

Comments

jonathanpeppers commented Jul 28, 2022 • edited Loading

Description

Reproduction Steps

Expected behavior

Actual behavior

Regression?

Known Workarounds

Configuration

Other information

dotnet-issue-labeler bot commented Jul 28, 2022

jonathanpeppers commented Jul 28, 2022

ghost commented Jul 28, 2022

Description

Reproduction Steps

Expected behavior

Actual behavior

Regression?

Known Workarounds

Configuration

Other information

steveisok commented Jul 28, 2022

vargaz commented Jul 28, 2022

steveisok commented Jul 28, 2022

jonathanpeppers commented Jul 28, 2022

fanyang-mono commented Aug 2, 2022

lambdageek commented Aug 17, 2022

akoeplinger commented Aug 17, 2022

lambdageek commented Aug 17, 2022

lambdageek commented Aug 17, 2022

fanyang-mono commented Aug 17, 2022

akoeplinger commented Aug 18, 2022

lambdageek commented Aug 18, 2022

jonathanpeppers commented Jul 28, 2022 •

edited

Loading