Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[WIP] Fast codegen-free generic object instantiation + improve Activator.CreateInstance perf #1

Closed
wants to merge 193 commits into from

Conversation

GrabYourPitchforks
Copy link
Owner

(Migrated from dotnet#32520 so I can iterate here without distracting the folks in the main runtime repo. See the PR feedback there for discussion.)


Contributes to dotnet#23716.

This is a very rough prototype of ObjectFactory and UninitializedObjectFactory, which allow fast codegen-free instantiation of arbitrary objects. It's primarily suited for DI and other scenarios where developers need to instantiate objects whose types aren't known until runtime but are required to do so with as minimal overhead as possible.

Historically, the way to instantiate such objects is to call Activator.CreateInstance or ConstructorInfo.Invoke. However, there is some overhead to this, as checks need to take place on each reflection API invocation. Some developers who need to eke out extra performance end up creating a DynamicMethod and hand-rolling a newobj instruction into the IL stream. This provides generally acceptable results, but it is tricky to get right, there's significant per-instantiation overhead due to spinning up the JIT, it's complex to debug issues in the code gen, and it doesn't work properly in environments where code gen is unavailable.

Note: These APIs have not yet been approved, as this is just an experimental PR to gauge general interest and to solicit feedback on the design and functionality.

General philosophy

Historically, most reflection APIs like Activator.CreateInstance and ConstructorInfo.Invoke perform checks on each method invocation. We can try to reduce the cost of these checks, but due to the nature of these APIs we'll never be able to remove the checks entirely.

The UninitializedObjectFactory and ObjectFactory APIs, on the other hand, perform all of their checks at construction time, not at invocation time. This means that it may be somewhat expensive to create the factory / dispenser itself, but once the factory is created it's very cheap to request a new instance from the factory. The intent is that these factories aren't used for one-off scenarios - callers should continue to use Activator.CreateInstance and ConstructorInfo.Invoke in such code paths. But in scenarios where the same type of object will be instantiated over and over again, it's very efficient to cache a single instance of the factory per-type and to query the cached instance whenever a new T instance is needed.

UninitializedObjectFactory

The UninitializedObjectFactory type is roughly equivalent to RuntimeHelpers.GetUninitializedObject. This instantiates the type without running any instance constructors. (Static constructors are still run.) Most applications should never need this, but it is useful for library authors building their own serialization or DI frameworks on top of this. The general pattern such library authors should follow is to use UninitializedObjectFactory to create a not-yet-constructed instance of target type T, then manually invoke the appropriate constructor on the newly-allocated instance.

ObjectFactory

The ObjectFactory type is roughly equivalent to Activator.CreateInstance<T>(). It allocates the type then calls the public, parameterless instance constructor. There are two main differences between ObjectFactory.CreateInstance and Activator.CreateInstance<T>():

  • If object construction fails, ObjectFactory.CreateInstance will not wrap the thrown exception within a TargetInvocationException. Activator.CreateInstance<T>() will wrap exceptions.
  • If T is a value type, ObjectFactory.CreateInstance will always return default(T). Activator.CreateInstance<T>() will call T's parameterless constructor if one exists; otherwise it returns default(T).

Potential consumers

System.Text.Json

https://github.com/dotnet/runtime/blob/9d85b82e3ad856068459c8f8bba8509e038afb40/src/libraries/System.Text.Json/src/System/Text/Json/Serialization/ReflectionEmitMemberAccessor.cs#L39-L51

ASP.NET Core

dotnet/aspnetcore#14615 (though they're using TypeBuilder to work around this right now)

Other runtime + libraries

https://github.com/dotnet/runtime/blob/9d85b82e3ad856068459c8f8bba8509e038afb40/src/libraries/System.ComponentModel.Composition/src/System/ComponentModel/Composition/MetadataViewGenerator.cs#L371-L380

Inner workings

A newobj instruction is logically two operations: (a) allocate a block of zero-inited memory, then (b) call the type's instance ctor over this block of memory. (UninitializedObjectFactory skips this second step.)

// C#
object o = new object();
// MSIL
newobj instance void [System.Private.CoreLib]System.Object::.ctor()
mov rcx, 0x12345678 ; method table pointer for System.Object
call <allocator> ; when this returns, rax := zero-inited object instance
mov rcx, rax
call System.Object::.ctor() ; essentially, a static method with signature (object @this) -> void

When using Activator.CreateInstance, the runtime uses a generic allocator that can handle many different scenarios for allocating an object whose type is known only at runtime. And while this generic allocator is flexible, there's significant overhead due to all the different checks it needs to perform.

On the other hand, when JITting a newobj instruction, since the JIT knows the destination type statically, it's able to choose from a variety of allocators whose implementations are optimized for various different scenarios. For example, there exists an allocator tailored for small, non-finalizable types. There exists a different allocator tailored for finalizable types. And there exists yet another allocator that's used when instrumentation is enabled.

When an [Uninitialized]ObjectFactory<T> instance is created, it queries the runtime for the address of the allocator that would have been used for an object of type T. All allocators share the signature (MethodTable*) -> O, so once the allocator address is determined we can perform the equivalent of the following IL in order to quickly allocate the object.

ldarg.0 // assume arg 0 = MethodTable* of destination type
ldarg.1 // assume arg1 = allocator function pointer
calli object(native int)
ret

If we need to fully construct the object, then we further query the runtime for the address of the code that corresponds to the public parameterless ctor of type T. All public parameterless ctors on reference types share the signature (object) -> void, so once the ctor address is determined we can perform the equivalent of the following IL in order to quickly call the ctor.

ldarg.0 // assume arg 0 = return value from UninitializedObjectFactory.CreateUninitializedInstance()
ldarg.1 // assume arg 1 = public parameterless ctor function pointer
calli void(object)
ret

Putting this all together, we get the following pseudo-codegen for ObjectFactory<T>.CreateInstance.

mov rcx, qword ptr [@this + <offset_to_methodtableptr_field>]
call qword ptr [@this + <offset_to_allocatorptr_field>]
mov rcx, rax
call qword ptr [@this + <offset_to_ctorptr_field>]
ret

Performance

Below are preliminary numbers for various "create an object" scenarios. Calling new object() directly is the baseline.

Method Mean Ratio Scenario
DirectNewobj 4.011 ns 1.00 Calling new object() directly
DynamicMethodNewobj 5.527 ns 1.31 Calling a cached DynamicMethod whose IL stream reads newobj [object]; ret;
ObjectFactoryCreateInstance 4.937 ns 1.23 Calling a cached ObjectFactory<object>'s CreateInstance method

I've also experimented with plumbing this through the existing Activator.CreateInstance pipeline. The perf numbers are not as good as calling ObjectFactory.CreateInstance directly because there's still overhead to querying the static cache and performing any necessary checks, and there's some compat behavior such as TargetInvocationException wrapping that we can't avoid, but they still show a significant improvement over the baseline (master branch) Activator.CreateInstance numbers.

Method Toolchain Mean Error StdDev Ratio RatioSD
ActivatorCreateInstance master 40.55 ns 0.889 ns 1.156 ns 1.00 0.00
ActivatorCreateInstance objectfactory 14.54 ns 0.362 ns 0.303 ns 0.36 0.01
ActivatorCreateInstanceOfT master 42.17 ns 0.917 ns 1.581 ns 1.00 0.00
ActivatorCreateInstanceOfT objectfactory 16.86 ns 0.409 ns 0.613 ns 0.40 0.02

And for RuntimeHelpers.GetUninitializedObject(typeof(object)) vs. UninitializedObjectFactory<object>.CreateUninitializedInstance():

Method Mean Ratio
RH_GetUninitializedObject 40.377 ns 1.00
UninitializedObjectFactory 3.588 ns 0.09

Remaining work

  • Continue with code cleanup and testing edge cases like trying to instantiate arrays.
  • Support parameterful ctors; e.g., new T(IServiceProvider).
  • Convert the code to use calli support when it's natively available in C#.
  • Full API review.
  • Support for Mono + CoreRT.

GrabYourPitchforks pushed a commit that referenced this pull request Apr 13, 2020
This adds support in the JIT emitter for Vector Load/Store structure instructions (C3.2.10 - Arm
Architecture Reference Manual):

- LD1 (1-4 registers)
- LD2
- LD3
- LD4
- LD1R
- LD2R
- LD3R
- LD4R
- ST1 (1-4 registers)
- ST2
- ST3
- ST4

in the following addressing modes:

- Base register only
- Post-indexed by a 64-bit register
- Post-indexed by an immediate, equal to the number of bytes transferred

Also adds support in JitDump for printing of

* A SIMD vector register list.
  For example, ld1     {v5.16b, v6.16b, v7.16b, v8.16b}, [x9]

* A SIMD vector element list. 
  For example, st1     {v0.b}[3], [x1],#1
- Eliminates ObjectFactory<T> except when caller explicitly wants a Func<T>
- Uses C# 9.0 function pointers to avoid introducing new JIT intrinsics
@GrabYourPitchforks
Copy link
Owner Author

A status update - I'm still working on this but keep running into issues where the linker either crashes or trims a method which is clearly statically reachable. Trying to work around those before sending up a new iteration. Meanwhile I've reduced the scope of the work so that it's not adding any new APIs in the 5.0 timeframe.

GrabYourPitchforks and others added 6 commits July 7, 2020 22:35
- Remove new public APIs
- Remove most new native APIs
- Plumb GetUninitializedObject atop new native code paths
- Use actual builds from arcade
- Don't be so aggressive about running cctors
- Some minor code cleanup
- Don't swallow OOMs
benaadams and others added 14 commits November 18, 2020 13:25
* Change SuperPMI collection to not use altjit mechanism

The SuperPMI collection process interposes a "shim" JIT between the JIT and EE.
As it is inconvenient to physically replace the existing JIT, currently this
is done by enabling altjit compilation and setting `COMPlus_AltJitName` to the
name of the shim JIT. This creates other inconvenience, especially with the
new way of saving the altjit flag in the JIT flags, by requiring us to
force the altjit flags bit to not be set, and force unset all the altjit flags.
It also makes it inconvenient to collect and/or replay with an actual altjit.

Change this collection mechanism to use the newly restored `COMPlus_JitName`
variable to allow specifying the JIT, and use that to specify the SuperPMI shim.

In addition, do not record in the MC file the `COMPlus_EnableExtraSuperPmiQueries`
variable. This is only used during collections to attempt to make replays more
flexible, but we never want to tell the JIT during replay that this is set.

* Enable `COMPlus_JitName` for crossgen as well

* Fix issue with empty collection args
…ude right base type (dotnet#44774)

* Fix IComponent Designer attribute that was missing base designer type

* Fix IRootDesigner references to use assembly qualified name

* Fix build failures
…tnet#44532)

For the RuntimeEventSource, this removes around 30K of allocation, though that's only ~3% of what gets allocated.
Signed-off-by: Jeremy Koritzinsky <jekoritz@microsoft.com>
I also made a slight optimization to CallSiteFactory to use ToArray instead of ToList.
…tnet#44853)

* optimize constructor of ElapsedEventArgs

* delete the old ctor
The --raw command-line option used to dump bytes prefixed with their RVAs for all structures except GC blobs, which were dumped with file offsets instead. This change fixes that inconsistency. It also fixes the size reported by x86.GcInfo, which determines how many bytes to dump. In addition I simplified ReadyToRunMethod's code to store just the GC blob's RVA and avoid the delegate allocation.
* Fix illumos managed build
* Fix CA1823 (unused private field) in `NetworkChange` partial for
  `UnknownUnix`.
* Use official casing `illumos` in MSBuild property names (as done for
  iOS).
* Fix Solaris version in test with SDK's PlatformDetection.
  * only major version is needed.

* Implement Enviornment.WorkingSet for SunOS
Difference between Linux and SunOS procfs is that files in latter
contain binary data, so we need `read(2)` and cast into corresponding
struct. Redeclaring system structs in managed code is not ideal, as
they do change across the major versions of OS, which inevitably
requires recompilation of binaries and replicating them in C# as is
means additional/unnecessary maintenance of code.

* Address CR feedback
stephentoub and others added 13 commits November 25, 2020 18:05
* Reduce allocation from OptionsCache's concurrent dictionary

This type is primarily used for getting and rarely mutated after startup; we don't need to pay for lots of lock objects to optimize for mutation.

* Avoid closure/delegate allocations in `OptionsManager<T>.Value`

* Update src/libraries/Microsoft.Extensions.Options/src/OptionsCache.cs

Co-authored-by: David Fowler <davidfowl@gmail.com>

Co-authored-by: David Fowler <davidfowl@gmail.com>
…01125.2 (dotnet#45216)

Microsoft.DotNet.XHarness.CLI , Microsoft.DotNet.XHarness.TestRunners.Xunit
 From Version 1.0.0-prerelease.20574.2 -> To Version 1.0.0-prerelease.20575.2

Co-authored-by: dotnet-maestro[bot] <dotnet-maestro[bot]@users.noreply.github.com>
Co-authored-by: Jan Kotas <jkotas@microsoft.com>
@GrabYourPitchforks
Copy link
Owner Author

Closing since this is going into the real runtime branch imminently.

@GrabYourPitchforks GrabYourPitchforks deleted the fast_createinstance branch November 26, 2020 07:33
GrabYourPitchforks pushed a commit that referenced this pull request May 18, 2021
…tnet#52769)

Transition to GC Unsafe mode on every MONO_RT_EXTERNAL_ONLY function in
reflection.c

In particular, fix mono_reflection_type_from_name which is used in
https://github.com/xamarin/xamarin-android/blob/681887ebdbd192ce7ce1cd02221d4939599ba762/src/monodroid/jni/embedded-assemblies.cc#L350

Fixes stack traces like

```
05-14 08:06:12.848 31274 31274 F DEBUG   :       #00 pc 00000b99  [vdso] (__kernel_vsyscall+9)
05-14 08:06:12.848 31274 31274 F DEBUG   :       #1 pc 0005ad68  /apex/com.android.runtime/lib/bionic/libc.so (syscall+40) (BuildId: 6e3a0180fa6637b68c0d181c343e6806)
05-14 08:06:12.848 31274 31274 F DEBUG   :       #2 pc 00076511  /apex/com.android.runtime/lib/bionic/libc.so (abort+209) (BuildId: 6e3a0180fa6637b68c0d181c343e6806)
05-14 08:06:12.848 31274 31274 F DEBUG   :       #3 pc 0002afcd  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonodroid.so (xamarin::android::internal::MonodroidRuntime::mono_log_handler(char const*, char const*, char const*, int, void*)+141) (BuildId: 9726f32ad5f8fa5e7c5762baf2f6e3294da41cc1)
05-14 08:06:12.848 31274 31274 F DEBUG   :       #4 pc 00112c5d  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (eglib_log_adapter+141) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #5 pc 00020fdf  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (monoeg_g_logv+175) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #6 pc 0002113a  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (monoeg_g_log+42) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #7 pc 00128892  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_threads_transition_do_blocking+258) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #8 pc 0012a406  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_threads_enter_gc_safe_region_unbalanced_with_info+134) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #9 pc 0012a27e  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_threads_enter_gc_safe_region_internal+46) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #10 pc 000799a7  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_loader_lock+71) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #11 pc 000447a1  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_class_create_from_typedef+129) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #12 pc 0003c073  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_class_get_checked+99) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #13 pc 0003cc0f  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_class_from_name_checked_aux+735) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #14 pc 00037989  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_class_from_name_checked+73) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #15 pc 000cc5f4  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_reflection_get_type_internal+132) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #16 pc 000c9bce  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_reflection_get_type_with_rootimage+126) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #17 pc 000ca204  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (_mono_reflection_get_type_from_info+292) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #18 pc 000ca06e  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_reflection_type_from_name_checked+334) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #19 pc 000c9f01  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonosgen-2.0.so (mono_reflection_type_from_name+49) (BuildId: b67e93dd750dafdd6f65f408b021b6a3a74868ac)
05-14 08:06:12.849 31274 31274 F DEBUG   :       #20 pc 0001b40b  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonodroid.so (xamarin::android::internal::EmbeddedAssemblies::typemap_java_to_managed(char const*)+427) (BuildId: 9726f32ad5f8fa5e7c5762baf2f6e3294da41cc1)
05-14 08:06:12.849 31274 31274 F DEBUG   :       dotnet#21 pc 0001b551  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonodroid.so (xamarin::android::internal::EmbeddedAssemblies::typemap_java_to_managed(_MonoString*)+113) (BuildId: 9726f32ad5f8fa5e7c5762baf2f6e3294da41cc1)
05-14 08:06:12.849 31274 31274 F DEBUG   :       dotnet#22 pc 000211a7  /data/app/~~rMrkpKmVPaBpM5jKb8fPAg==/com.microsoft.maui-JfRo8RWSDJaNtJuBa0y7_Q==/lib/x86/libmonodroid.so (xamarin::android::internal::MonodroidRuntime::typemap_java_to_managed(_MonoString*)+39) (BuildId: 9726f32ad5f8fa5e7c5762baf2f6e3294da41cc1)
```
GrabYourPitchforks pushed a commit that referenced this pull request May 18, 2021
…2915)

* [build] Define NO_UNALIGNED_ACCESS for 32-bit arm platforms

Possibly related to crashes on Android like this:

```
05-18 10:59:07.466 17076 17076 F libc    : Fatal signal 7 (SIGBUS), code 1 (BUS_ADRALN), fault addr 0xb9c95a41 in tid 17076 (simplehellomaui), pid 17076 (simplehellomaui)
05-18 10:59:07.501 17104 17104 I crash_dump32: obtaining output fd from tombstoned, type: kDebuggerdTombstone
05-18 10:59:07.502   989   989 I tombstoned: received crash request for pid 17076
05-18 10:59:07.503 17104 17104 I crash_dump32: performing dump of process 17076 (target tid = 17076)
05-18 10:59:07.512 17104 17104 F DEBUG   : *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***
05-18 10:59:07.512 17104 17104 F DEBUG   : Build fingerprint: 'google/crosshatch/crosshatch:11/RQ2A.210405.005/7181113:user/release-keys'
05-18 10:59:07.512 17104 17104 F DEBUG   : Revision: 'MP1.0'
05-18 10:59:07.512 17104 17104 F DEBUG   : ABI: 'arm'
05-18 10:59:07.515 17104 17104 F DEBUG   : Timestamp: 2021-05-18 10:59:07+0200
05-18 10:59:07.515 17104 17104 F DEBUG   : pid: 17076, tid: 17076, name: simplehellomaui  >>> com.microsoft.simplehellomaui <<<
05-18 10:59:07.515 17104 17104 F DEBUG   : uid: 10364
05-18 10:59:07.515 17104 17104 F DEBUG   : signal 7 (SIGBUS), code 1 (BUS_ADRALN), fault addr 0xb9c95a41
05-18 10:59:07.515 17104 17104 F DEBUG   :     r0  bb4a5cd0  r1  b9c95a49  r2  00000000  r3  e94c7520
05-18 10:59:07.515 17104 17104 F DEBUG   :     r4  0000000c  r5  00000000  r6  ff843c50  r7  ff843e70
05-18 10:59:07.515 17104 17104 F DEBUG   :     r8  b69547f8  r9  e99eac50  r10 00000000  r11 00000021
05-18 10:59:07.515 17104 17104 F DEBUG   :     ip  e94c74f0  sp  ff843c48  lr  bb31e0dd  pc  bb3a4d24
05-18 10:59:07.531   709   709 E Layer   : [Surface(name=Task=1)/@0x52e6b1a - animation-leash#0] No local sync point found
05-18 10:59:07.532   709   709 E Layer   : [Surface(name=Task=1571)/@0x9c90165 - animation-leash#0] No local sync point found
05-18 10:59:07.706 17104 17104 F DEBUG   : backtrace:
05-18 10:59:07.707 17104 17104 F DEBUG   :       #00 pc 000ddd24  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (mono_method_to_ir+9232) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #1 pc 000d7777  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (inline_method+622) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #2 pc 000ec0a3  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (mono_method_to_ir+67470) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #3 pc 000cda6d  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (mini_method_compile+2264) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #4 pc 000cf413  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (mono_jit_compile_method_inner+50) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #5 pc 000d1d7f  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (mono_jit_compile_method_with_opt+1766) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #6 pc 0012d94d  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (common_call_trampoline+832) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #7 pc 0012d5cb  /data/app/~~J4DFQ3c1v2YGrEurX7TNjg==/com.microsoft.simplehellomaui-_jGGPiZpZ3yT-QCTNDcgvQ==/lib/arm/libmonosgen-2.0.so (mono_magic_trampoline+62) (BuildId: d0a4e41a500357a621884b64f6ca8533b62a664b)
05-18 10:59:07.707 17104 17104 F DEBUG   :       #8 pc 0000006a <anonymous:b7986000>
```

* move to host/target sections
GrabYourPitchforks pushed a commit that referenced this pull request Jul 8, 2021
…et#53792)

I have expanded the PerfMap format produced by Crossgen2 and
R2RDump to produce metadata in form of pseudo-symbol records with
high addresses. In this version I have implemented four metadata
entries - output GUID, target OS, target architecture and perfmap
format version number.  I have verified for System.Private.CoreLib
and for the composite framework that Crossgen2 and R2RDump
produce identical metadata.

To facilitate a smooth transition to the new perfmap format, in
accordance with Juan's suggestion I have introduced a new command-line
option to explicitly specify the perfmap format revision. As of today,
0 corresponds to the legacy Crossgen1-style output where the
perfmap file name includes the {MVID} section, perfmap format #1
corresponds to current Crossgen2 with its new naming scheme.
As of today there are no differences in the file content.

Thanks

Tomas
GrabYourPitchforks pushed a commit that referenced this pull request Mar 30, 2022
…otnet#63598)

* Fix native frame unwind in syscall on arm64 for VS4Mac crash report.

Add arm64 version of StepWithCompactNoEncoding for syscall leaf node wrappers that have compact encoding of 0.

Fix ReadCompactEncodingRegister so it actually decrements the addr.

Change StepWithCompactEncodingArm64 to match what MacOS libunwind does for framed and frameless stepping.

arm64 can have frames with the same SP (but different IPs). Increment SP for this condition so createdump's unwind
loop doesn't break out on the "SP not increasing" check and the frames are added to the thread frame list in the
correct order.

Add getting the unwind info for tail called functions like this:

__ZL14PROCEndProcessPvji:
   36630:       f6 57 bd a9     stp     x22, x21, [sp, #-48]!
   36634:       f4 4f 01 a9     stp     x20, x19, [sp, #16]
   36638:       fd 7b 02 a9     stp     x29, x30, [sp, dotnet#32]
   3663c:       fd 83 00 91     add     x29, sp, dotnet#32
...
   367ac:       e9 01 80 52     mov     w9, #15
   367b0:       7f 3e 02 71     cmp     w19, dotnet#143
   367b4:       20 01 88 1a     csel    w0, w9, w8, eq
   367b8:       2e 00 00 94     bl      _PROCAbort
_TerminateProcess:
-> 367bc:       22 00 80 52     mov     w2, #1
   367c0:       9c ff ff 17     b       __ZL14PROCEndProcessPvji

The IP (367bc) returns the (incorrect) frameless encoding with nothing on the stack (uses an incorrect LR to unwind). To fix this
get the unwind info for PC -1 which points to PROCEndProcess with the correct unwind info. This matches how lldb unwinds this frame.

Always address module segment to IP lookup list instead of checking the module regions.

Strip pointer authentication bits on PC/LR.
GrabYourPitchforks pushed a commit that referenced this pull request May 26, 2022
This adds support for EnC on arm64. A couple of notes on the
implementation compared to x64:
- On x64 we get the fixed stack size from unwind info. However, for the
  frames we set up on arm64 for EnC it is not possible to extract the
  frame size from there because their prologs generally look like

  stp fp, lr, [sp,#-16]!
  mov fp, sp
  sub sp, sp, dotnet#144

  with unwind codes like the following:

  set_fp; mov fp, sp

  save_fplr_x #1 (0x01); tp fp, lr, [sp, #-16]!

  As can be seen, it is not possible to get the fixed stack size from
  unwind info in this case. Instead we pass it through the GC info that
  already has a section for EnC data.

- On arm64 the JIT is required to place the PSPSym at the same offset
  from caller-SP for both the main function and for funclets. Due to
  this we try to allocate the PSPSym as early as possible in the main
  function and we must take some care in funclets.  However, this
  conflicts with the EnC frame header that the JIT uses to place values
  that must be preserved on EnC transitions. This is currently
  callee-saved registers and the MonitorAcquired boolean.

  Before this change we were allocating PSPSym above (before) the
  monitor acquired boolean, but we now have to allocate MonitorAcquired
  first, particularly because the size of the preserved header cannot
  change on EnC transitions, while the PSPSym can disappear or appear.
  This changes frame allocation slightly for synchronized functions.
GrabYourPitchforks pushed a commit that referenced this pull request Jun 10, 2022
These helpers are used to report names of things in warnings. The functional changes are:
* For method parameters, use the parameter name if available (and only if not fallback to the #1 notation)
* For property accessor methods, use the C# naming scheme, so for example Type.Property.get instead of Type.get_Property.

Both of these changes are in preparation to bring NativeAOT closer in behavior to ILLink and the trim analyzers.

For this I moved some of the helpers to the common shared code.

Some unrelated code cleanup as well.

Co-authored-by: Michal Strehovský <MichalStrehovsky@users.noreply.github.com>
GrabYourPitchforks pushed a commit that referenced this pull request Jun 30, 2022
* Initial implementation for contract customization

fix build errors

Move converter rooting to DefaultJsonTypeInfoResolver so that it can be used standalone

Fix ConfigurationList.IsReadOnly

Minor refactorings (#1)

* Makes the following changes:

* Move singleton initialization for DefaultTypeInfoResolver behind a static property.
* Consolidate JsonSerializerContext & IJsonTypeInfoResolver values to a single field.
* Move reflection fallback logic away from JsonSerializerContext and into JsonSerializerOptions

* Update src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializerOptions.cs

* remove testing of removed field

Simplify the JsonTypeInfo.CreateObject implemenetation (#2)

* Simplify the JsonTypeInfo.CreateObject implemenetation

* Update src/libraries/System.Text.Json/src/System/Text/Json/Serialization/Metadata/JsonTypeInfoOfT.cs

* Update src/libraries/System.Text.Json/src/System/Text/Json/Serialization/Metadata/JsonTypeInfoOfT.cs

Co-authored-by: Krzysztof Wicher <mordotymoja@gmail.com>

Co-authored-by: Krzysztof Wicher <mordotymoja@gmail.com>

Tests and fixes for JsonTypeInfoKind.None

TypeInfo type mismatch tests

Allow setting NumberHandling on JsonTypeInfoKind.None

test resolver returning wrong type of options

JsonTypeInfo/JsonPropertyInfo mutability tests

rename test file

Move default converter rooting responsibility behind DefaultJsonTypeInfoResolver (#3)

* Move default converter rooting responsibility behind DefaultJsonTypeInfoResolver

* address feedback

Add simple test for using JsonTypeInfo<T> with APIs directly taking it

fix and tests for untyped/typed CreateObject

uncomment test cases, remove todo

More tests and tiny fixes

Add a JsonTypeInfoResolver.Combine test for JsonSerializerContext (#4)

* Fix JsonTypeInfoResolver.Combine for JsonSerializerContext

* Break up failing test

Fix simple scenarios for combining contexts (#6)

* Fix simple scenarios for combining contexts

* feedback

JsonSerializerContext combine test with different camel casing

Remove unneeded virtual calls & branching when accessing Get & Set delegates (#7)

JsonPropertyInfo tests everything minus ShouldSerialize & NumberHandling

Update src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonConverterOfT.cs

Update src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonConverterOfT.cs

throw InvalidOperationException rather than ArgumentNullException for source gen when PropertyInfo.Name is assigned through JsonPropertyInfoValues

tests for duplicated property names and JsonPropertyInfo.NumberHandling

Add tests for NumberHandling and failing tests for ShouldSerialize

disable the failing test and add extra checks

disable remainder of the failing ShouldSerialize tests, fix working one

Fix ShouldSerialize and IgnoreCondition interop

Add failing tests for CreateObject + parametrized constructors

Fix CreateObject support for JsonConstructor types (#10)

* Fix CreateObject support for JsonConstructor types

* address feedback

Make contexts more combinator friendly (#9)

* Make contexts more combinator friendly

* remove converter cache

* redesign test to account for JsonConstructorAttribute

* Combine unit tests

* address feedback

* Add acceptance tests for DataContract attributes & Specified pattern (#11)

* Add private field serialization acceptance test (#13)

* tests, PR feedback (#14)

* PR feedback and extra tests

* Shorten class name, remove incorrect check (not true for polimorphic cases)

* Make parameter matching for custom properties map property Name with parameter (#16)

* Test static initialization with JsonTypeInfo (#17)

* Fix test failures and proper fix this time (#18)

* Fix test failures and proper fix this time

* reinstate ActiveIssueAttribute

* PR feedback and adjust couple of tests which don't set TypeInfoResolver

* fix IAsyncEnumerable tests

* Lock JsonSerializerOptions in JsonTypeInfo.EnsureConfigured()

Co-authored-by: Eirik Tsarpalis <eirik.tsarpalis@gmail.com>

Co-authored-by: Eirik Tsarpalis <eirik.tsarpalis@gmail.com>
GrabYourPitchforks pushed a commit that referenced this pull request Jul 20, 2022
E.g.,

Update LSRA "Allocating Registers" table description.

Dump nodes added during resolution, e.g.:
```
   BB29 bottom (BB08->BB08): move V25 from STK to rdi (Critical)
N001 (  1,  1) [001174] ----------z                 t1174 =    LCL_VAR   int    V25 cse4          rdi REG rdi
```

Dump more data in the LSRA block sequence data:
```
-BB03( 16   )
-BB04(  4   )
+BB03 ( 16   ) critical-in critical-out
+BB04 (  4   ) critical-out
```

When dumping various flow bitvectors, annotate the bitvectors better:
```
-BB25 in gen out
-0000000000000000
-0000000000000003 CSE #1.c
-0000000000000003 CSE #1.c
+BB25
+ in: 0000000000000000
+gen: 0000000000000003 CSE #1.c
+out: 0000000000000003 CSE #1.c
```

Dump hoisting bitvectors using the sorting number:
```
-  USEDEF  (5)={V04 V00 V01 V02 V03}
+  USEDEF  (5)={V00 V01 V02 V03 V04}
```

Also, fix various typos and formatting.
GrabYourPitchforks pushed a commit that referenced this pull request Aug 29, 2022
* WIP: add gRPC tests

* Fix AOT and trimming

* WIP

* Implement IncludeNetworkSecurityConfig

* Use IncludeNetworkSecurityConfig

* Fix gRPC test

* Avoid git checkout

* Remove unnecessary code

* WIP: start working on CI configuration

* Remove WinHttpHandler

* Fix problem with SSL

* Change server host

* Setup CI (#1)

* Get Docker container building & exported via test build

* Changes

* Add missing pfx certificate

* changes

* cleanup

Co-authored-by: Simon Rozsival <simon@rozsival.com>

* Use tls

* Update yml

* Revert changes to the mono Android sample app

* Bump android image version

* Bump image version

* Enable TLS

* Remove hardcoded package versions

* Update package versions

* Update package versions

* Rename pipeline

* Move interop tests website dependencies versions to Versions.props

* Add cred scan supression for the interop test server private key

* Fix licenses

* Remove dependencies

* Fix path to Versions.props

* Remove unnecessary dependency version

* Fix building docker image

* Change pfx password

Co-authored-by: Jo Shields <directhex@apebox.org>
GrabYourPitchforks pushed a commit that referenced this pull request Feb 7, 2023
* switch to managed thread ID in Lock

* fattening the lock

* __declspec(selectany)

* few tweaks

* fairness

* more room for thread ids

* remove CurrentNativeThreadId

* couple fixes

* fix win-arm64 build

* win-arm64 build , another try

* Apply suggestions from code review

Co-authored-by: Jan Kotas <jkotas@microsoft.com>

* fix after renaming

* do not report successful spin if thread has waited

* keep extern and undo mangling of tls_CurrentThread in asm

* use SyncTable indexer in less perf-sensitive places.

* GetNewHashCode just delegate to shared random

* Apply suggestions from code review

Co-authored-by: Jan Kotas <jkotas@microsoft.com>

* unchecked const conversion

* some refactoring comments and typos

* min number of spins in the backoff

* moved CurrentManagedThreadIdUnchecked to ManagedThreadId

* Use `-1` to report success and allow using  element #1 in the SyncTable

* use threadstatic for managed thread ID

* check before calling RhGetProcessCpuCount

* use 0 as default thread ID

Co-authored-by: Jan Kotas <jkotas@microsoft.com>
GrabYourPitchforks pushed a commit that referenced this pull request Jul 18, 2023
…tnet#87189)

This fixes a startup crash on Big Sur:

> error: * Assertion at /Users/runner/work/1/s/src/mono/mono/utils/mono-hwcap-arm64.c:35, condition `res == 0' not met

Because sysctl can't find some of these options:

    $ sysctl hw.optional.armv8_crc32
    hw.optional.armv8_crc32: 1
    $ sysctl hw.optional.arm.FEAT_RDM
    sysctl: unknown oid 'hw.optional.arm.FEAT_RDM'
    $ sysctl hw.optional.arm.FEAT_DotProd
    sysctl: unknown oid 'hw.optional.arm.FEAT_DotProd'
    $ sysctl hw.optional.arm.FEAT_SHA1
    sysctl: unknown oid 'hw.optional.arm.FEAT_SHA1'
    $ sysctl hw.optional.arm.FEAT_SHA256
    sysctl: unknown oid 'hw.optional.arm.FEAT_SHA256'
    $ sysctl hw.optional.arm.FEAT_AES
    sysctl: unknown oid 'hw.optional.arm.FEAT_AES'

Full stack trace:

* thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 2.1
  * frame #0: 0x0000010ef37560 libmonosgen-2.0.dylib`monoeg_assertion_message
    frame #1: 0x0000010ef375cc libmonosgen-2.0.dylib`mono_assertion_message + 32
    frame #2: 0x0000010ef40d6c libmonosgen-2.0.dylib`mono_hwcap_arch_init + 544
    frame #3: 0x0000010ef54bd8 libmonosgen-2.0.dylib`mono_hwcap_init + 72
    frame #4: 0x0000010ee14dc0 libmonosgen-2.0.dylib`parse_optimizations + 52
    frame #5: 0x0000010edbed48 libmonosgen-2.0.dylib`mono_init
    frame #6: 0x0000010ee18968 libmonosgen-2.0.dylib`mono_jit_init_version
    frame #7: 0x0000010f48a300 libxamarin-dotnet-debug.dylib`xamarin_bridge_initialize + 216
    frame #8: 0x0000010f4900a4 libxamarin-dotnet-debug.dylib`xamarin_main + 376
GrabYourPitchforks pushed a commit that referenced this pull request Oct 17, 2023
* Fix the MacOS remote unwinder for VS4Mac

The wrong module was being passed to the remote unwinder because the load bias for shared modules
was being calculated incorrectly.

Issue: dotnet#63309

* Fix native frame unwind in syscall on arm64 for VS4Mac crash report

From PR in main: dotnet#63598

Add arm64 version of StepWithCompactNoEncoding for syscall leaf node wrappers that have compact encoding of 0.

Fix ReadCompactEncodingRegister so it actually decrements the addr.

Change StepWithCompactEncodingArm64 to match what MacOS libunwind does for framed and frameless stepping.

arm64 can have frames with the same SP (but different IPs). Increment SP for this condition so createdump's unwind
loop doesn't break out on the "SP not increasing" check and the frames are added to the thread frame list in the
correct order.

Add getting the unwind info for tail called functions like this:

__ZL14PROCEndProcessPvji:
   36630:       f6 57 bd a9     stp     x22, x21, [sp, #-48]!
   36634:       f4 4f 01 a9     stp     x20, x19, [sp, #16]
   36638:       fd 7b 02 a9     stp     x29, x30, [sp, dotnet#32]
   3663c:       fd 83 00 91     add     x29, sp, dotnet#32
...
   367ac:       e9 01 80 52     mov     w9, #15
   367b0:       7f 3e 02 71     cmp     w19, dotnet#143
   367b4:       20 01 88 1a     csel    w0, w9, w8, eq
   367b8:       2e 00 00 94     bl      _PROCAbort
_TerminateProcess:
-> 367bc:       22 00 80 52     mov     w2, #1
   367c0:       9c ff ff 17     b       __ZL14PROCEndProcessPvji

The IP (367bc) returns the (incorrect) frameless encoding with nothing on the stack (uses an incorrect LR to unwind). To fix this
get the unwind info for PC -1 which points to PROCEndProcess with the correct unwind info. This matches how lldb unwinds this frame.

Always address module segment to IP lookup list instead of checking the module regions.

Strip pointer authentication bits on PC/LR.
GrabYourPitchforks pushed a commit that referenced this pull request Apr 30, 2024
CodeQL flagged various places where we're dereferencing pointers that could be NULL, this PR systematically cleans some of them up via g_assert.
* g_assert result of g_build_path calls
* Allocation failure handling
* mono_class_inflate_generic_class_checked can return NULL
GrabYourPitchforks pushed a commit that referenced this pull request Oct 15, 2024
* bug #1: don't allow for values out of the SerializationRecordType enum range

* bug #2: throw SerializationException rather than KeyNotFoundException when the referenced record is missing or it points to a record of different type

* bug #3: throw SerializationException rather than FormatException when it's being thrown by BinaryReader (or sth else that we use)

* bug #4: document the fact that IOException can be thrown

* bug #5: throw SerializationException rather than OverflowException when parsing the decimal fails

* bug #6: 0 and 17 are illegal values for PrimitiveType enum

* bug #7: throw SerializationException when a surrogate character is read (so far an ArgumentException was thrown)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.