Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
488 lines (362 sloc) 22.5 KB

Guide for porting .NET to a new processor architecture

This document is broken up into 2 major sections.

  1. The various porting stages of porting the .NET Runtime

  2. A technical discussion of the major components affected by a port to a new architecture

Porting stages and steps

Porting the .NET Runtime to a new architecture typically follows along the following path.

As engineering continues along the development path, it is best if the logic can be placed into the master repository of the runtime as soon as possible. This will have 2 major effects.

  1. Individual commits are easier to review.

  2. Not all approaches for fixing problems will always be considered acceptable. It is plausible that a change may not ever be acceptable to take into the upstream git repo, and discovering such issues early can avoid large amounts of sunk cost.

  3. When some change is made which breaks other platforms, it will be relatively simple to identify the break. If changes are held until after all changes are complete and the product is fully functional, this work is likely to be much more difficult.

Stage 1 Initial Bring Up

Porting .NET to a new platform starts with porting CoreCLR to a new architecture.

The process follows the following strategy

  • Add a new target architecture to the build environment, and make it build.

  • Determine if there is sufficient incentive to bring up the interpreter, or if simply making the jit handle the new architecture is cheaper. The interpreter in the CLR is currently only used for bring up scenarios, and is not maintained as generally working. It is expected that the interpreter will take 1-2 months to enable for an engineer familiar with the CoreCLR codebase. A functional interpreter allows the porting team to have a set of engineers which focus exclusively on the JIT and a set which focusses on the VM portion of the runtime.

  • Build up a set of scripts that will run the coreclr tests. The normal routine for running coreclr tests is XUnit, which is only suitable once the framework is mostly functional. These scripts will evolve during the development effort to support ever increasing needs of development. This set of scripts will be expected to do the following tasks.

    • Run a subset of the tests. Tests are arranged in a directory structure by category, so this subsetting mechanism will only need to be a directory structure system.

    • Some set of tests will need to be excluded on a test by test basis. Once the product is ready to ship, most of these disabled tests will need to have been re-enabled, but there are tests which will be disabled for months/years as the product is brought up to quality.

    • Produce crash or core dumps. The failure mode of many tests during this phase will be a crash. A test running tool that captures core dumps will make these issues easier to diagnose.

    • Produce bucketized lists of failures. Generally the approach is to group by assertion, and if there is a crash, group by callstack of crash.

  • The first test category to focus on is the JIT category, to bring up the general ability to run .NET code. Most of these tests are very simple, but getting some code to work is a prerequisite for handling more complex scenarios. When doing initial bringup, configuring the Gen0 budget of the GC to be a large number so that the GC does not attempt to run during most tests is very useful. (Set COMPlus_GCgen0size=99999999)

  • Once basic code is executing, the focus shifts to enabling the GC to work. In this initial phase, the correct choice is to enable conservative GC tracking via the FEATURE_CONSERVATIVE_GC macro. This feature will make garbage collection largely function correctly, but it is not suitable for production use of .NET, and can under certain circumstances trigger unbounded memory use.

  • Once basic GC works, and basic JIT functionality is present, work can fan out into all of the various features of the runtime. Of particular interest to engineers porting the runtime are the EH, stackwalking, and interop portions of the test suite.

  • During this phase, porting the SOS plugin from the https://github.com/dotnet/diagnostics will be very useful. The various commands available via that tool such as dumpmt, dumpdomain and such are regularly useful to developers attempting to port the runtime.

Stage 2 Expand scenario coverage

  • Once the coreclr tests are largely passing, the next step is to enable XUnit. At this time the clr is probably mostly capable of running XUnit tests, and adding testing using the libraries tests will require XUnit to work well.

  • Once XUnit is functional, bring up the libraries set of tests. There is quite a lot of the CoreCLR codebase that is largely only tested by the libraries test suites.

  • Engineers should also begin to attempt real scenario tests at this point, such as ASP.NET Core applications. If the libraries test suites work, then ASP.NET Core should as well.

Stage 3 Focus on performance

  • Throughput performance at this time is likely to be not that great. There are three major opportunities to improve performance at this stage.

    • Replace conservative GC with precise GC.

    • Tune the assembly stubs to be high performance on the platform, and implement optional assembly stubs where hand-written assembly would be faster than the equivalent C++ code.

    • Improve the code generated by the JIT.

  • Up until this point, engineers have probably been using the JIT for all code instead of bringing the Ready To Run compiler (crossgen/crossgen2) into usage on the platform. Implementing the ahead of the time compiler starts to be useful at this time to improve startup performance.

Stage 4 Focus on stress

  • Stress testing the system is necessary to provide confidence that the system really works.

  • See the various test passes done in CI, but most critically GCStress testing is needed. See documentation around use of the ComPlus_GCStress environment variable.

Stage 5 productization

  • Productization is about making the runtime able to run shipped effectively on a platform.

  • This document does not attempt to list out the work here as it is largely specific to the platform in use and the opinions of numerous stakeholders.

Design issues

These large architecture specific design issues will have substantial impact on both the JIT and VM.

  1. Calling convention rules – Caller pop vs Callee pop, HFA arguments, structure argument passing rules, etc. CoreCLR is designed to utilize a broadly similar ABI to the OS api. Managed to managed calls typically have a small set of tweaks or extensions to the ABI for VM efficiency purposes, but it is generally intended that the ABI of managed code and the ABI of native code are very similar. (This is not a hard requirement, and on Windows X86 the runtime supports a managed to managed abi as well as 3 separate native abis for interop, but this scheme is generally not recommended.) See the CLR-ABI document for how the existing architectures work. Ensure that the CLR-ABI document is updated with all the requisite details and special cases of the new platform. When defining the behavior of a new processor architecture abi for CoreCLR, we must maintain that:

    1. The this pointer is always passed in the same register regardless of other parameters.

    2. Various stub types will require an extra "secret" parameter. Perf details typically drive exactly where these are placed.

    3. When executing managed code it must be possible to hijack the return address. Current implementations require that the return address always be on the stack to do so, although this is a known performance deficiency for RISC platforms on arm64.

  2. Architecture specific relocation information (to represent generation of relocations for use by load, store, jmp and call instructions) See https://docs.microsoft.com/en-us/windows/win32/debug/pe-format#coff-relocations-object-only for the sort of details that need to be defined.

  3. Behavior and accessibility of processor single step features from within a process. On Unix the CLR debugger uses an in process thread to single step through functions.

  4. Unwind information. CoreCLR uses Windows style unwind data internally, even on Unix platforms. A Windows style unwind structure must be defined. In addition, it is possible to enable generation of DWARF data for exposure through the GDB JIT https://sourceware.org/gdb/onlinedocs/gdb/JIT-Interface.html . This support is conditional on an #ifdef, but has been used in the past to support bring up of new platforms.

  5. EH Funclets. .NET requires a 2 pass exception model in order to properly support exception filters. This substantially differs from the typical Itanium ABI model which is used on most Linux architectures

  6. OS behavior with Signals. Especially exactly where the reported instruction pointer is located.

  7. Little vs big endian. While .NET runtimes have been ported to big endian in the past (notable examples include Mono support various game consoles, and POWER, and XNA support on Xbox360) there are no current ports of CoreCLR to a big endian platform.

Components of the Runtime affected by a port to a new architecture

This a list of the notable architecture specific components of the .NET runtime. The list is not complete, but covers most of the areas where work will need to be done.

Notable components

  1. The JIT. The jit maintains the largest concentration of architecture specific logic in the stack. This is not surprising. See Porting RyuJit for guidance.

  2. The CLR PAL. When porting to a non-Windows OS, the PAL will be the first component that needs to be ported.

  3. The CLR VM. The VM is a mix of completely architecture neutral logic, and very machine specific paths.

  4. The unwinder. The unwinder is used to unwind stacks on non-Windows platforms. It is located in https://github.com/dotnet/runtime/tree/master/src/coreclr/src/unwinder.

  5. System.Private.CoreLib/System.Reflection. There is little to no architecture specific work here that is necessary for bringup. Nice-to-have work involves adding support for the architecture in the System.Reflection.ImageFileMachine enum, and the ProcessorArchitecture enum, and logic that manipulates it.

  6. PE File format changes to add a new architecture. Also, the C# compiler likely also needs a new switch to generate machine specific code for the new architecture.

  7. Crossgen/Crossgen2 - As the AOT compilers that produce machine specific logic from general purpose MSIL, these will be needed to improve startup performance.

  8. R2RDump - This allows diagnosing issues in pre-compiled code.

  9. coredistools - Necessary for GCStress (if determining instruction boundaries is non-trivial), as well as for SuperPMI asm diffs for JIT development.

  10. debug and diagnostics components - The managed debugger and profiler are beyond the scope of this document.

CLR PAL

The PAL provides a similar to Win32 api as the CLR codebase was originally designed to run on Windows platforms. Mostly the PAL is concerned with OS independence, but there are also architecture specific components.

  1. pal.h - Contains architecture specific details for handling unwinding scenarios such as CONTEXT / _KNONVOLATILE_CONTEXT_POINTERS/ _RUNTIME_FUNCTION.

  2. Unwinding support in seh-unwind.cpp

  3. context.cpp - Which manipulates and captures register contexts

  4. jitsupport.cpp - Depending on how the features of the CPU are exposed, there may need to be code to call OS apis to gather information about CPU features.

  5. pal arch directory - https://github.com/dotnet/runtime/tree/master/src/coreclr/src/pal/src/arch This directory primarily contains assembly stubs for architecture specific handling of signals and exceptions.

In addition to the PAL source code, there is a comprehensive set of PAL tests located in https://github.com/dotnet/runtime/tree/master/src/coreclr/src/pal/tests.

CLR VM

The VM support for architecture specific logic is encoded in a variety of different ways.

  1. Entirely architecture specific components. These are held in an architecture specific folder.

  2. Features which are only enabled on certain architectures. E.g. FEATURE_HFA.

  3. Ad-hoc #if blocks used for specific architectures. As needed these are added. The general goal is to keep these to a minimum, but difficulty here is primarily driven by what special behavior the processor architecture requires.

My recommendation would be to look at how Arm64 is implemented in the VM for the most up to date model of how to implement a CPU architecture.

Architecture Specific Components

There are a variety of architecture specific components that all architectures must implement.

  1. Assembly Stubs

  2. cgencpu.h (CPU specific header defining stubs and miscellaneous other CPU specific details.)

  3. VSD call stub generation (virtualcallstubcpu.hpp and associated logic)

  4. Precode/Prestub/Jumpstub generation

  5. callingconventions.h/argdestination.h Provides an implementation of the ABI used by VM components. The implementation made architecture specific via a long series of C preprocessor macros.

  6. gcinfodecoder.h The GC info format is archictecture specific as it holds information about which specific registers hold GC data. The implementation is generally simplified to be defined in terms of register numbers, but if the architecture has more registers available for use than existing architectures then the format will need extension.

Assembly Stubs

There are many reasons for which the runtime requires various assembly stubs. Here is an annotated list of the stubs implemented for Unix on Arm64.

  1. Only Performance. Some stubs have alternative implementations in C++ code which are used if there isn't an assembly stub. As compilers have gotten better, it has become more reasonable to just use the C++ versions. Often the biggest performance cost/win is due to fast paths being written that do not require setting up a stack frame. Most of the casting helpers fall in this category.

    1. JIT_Stelem_Ref – very slightly faster version of JIT_Stelem_Ref_Portable.
  2. General purpose correctness. Some helpers adjust the abi of whatever they call in interesting ways, manipulate/parse the "secret" arguments, or do other not quite compilable to standardized C concepts.

    1. CallDescrWorkerInternal – Needed to support VM to managed function calls. Necessary for all applications as this is how the main method is called.

    2. LazyMachStateCaptureState/HelperMethodFrameRestoreState – Needed to support a GC occurring with an FCALL or HCALL on the stack. (Incorrect implementations will cause unpredictable crashes during or after garbage collection)

    3. NDirectImportThunk – Needed to support saving off a set of arguments to a p/invoke so that the runtime can find the actual target. Also uses one of the secret arguments (Used by all p/invoke methods)

    4. PrecodeFixupThunk – Needed to convert the secret argument from a FixupPrecode* to a MethodDesc*. This function exists to reduce the code size of FixupPrecodes as there are (Used by many managed methods)

    5. ThePreStub - Needed to support saving off a set of arguments to the stack so that the runtime can find or jit the right target method. (Needed for any jitted method to execute Used by all managed methods)

    6. ThePreStubPatch – Exists to provide a reliable spot for the managed debugger to put a breakpoint.

    7. GC Write Barriers – These are used to provide the GC with information about what memory is being updated. The existing implementations of these are all complex, and there are a number of controls where the runtime can adjust to tweak the behavior of the barrier in various ways. Some of these adjustments involve modifying the code to inject constants, or even wholesale replacements of various bits and pieces. To achieve high performance, all of these features must work; however, to achieve bringup supporting a simple GC, focus on the case of the single heap workstation GC. Additionally, the FEATURE_MANUALLY_MANAGED_CARD_BUNDLES and FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP can be implemented as performance needs require.

    8. ComCallPreStub/ COMToCLRDispatchHelper /GenericComCallStub - not necessary for non-Windows platforms at this time

    9. TheUMEntryPrestub/ UMThunkStub - used to enter the runtime from non-managed code through entrypoints generated from the Marshal.GetFunctionPointerForDelagate api.

    10. OnHijackTripThread - needed for thread suspension to support GC + other suspension requiring events. This is typically not needed for very early stage bringup of the product, but will be needed for any decent size application

    11. CallEHFunclet – Used to call catch, finally and fault funclets. Behavior is specific to exactly how funclets are implemented. Only used if USE_FUNCLET_CALL_HELPER is set

    12. CallEHFilterFunclet – Used to call filter funclets. Behavior is specific to exactly how funclets are implemented. Only used if USE_FUNCLET_CALL_HELPER is set

    13. ResolveWorkerChainLookupAsmStub/ ResolveWorkerAsmStub Used for virtual stub dispatch (virtual call support for interface, and some virtual methods). These work in tandem with the logic in virtualcallstubcpu.h to implement the logic described in Virtual Stub Dispatch

    14. ProfileEnter/ ProfileeLeave/ ProfileTailcall – Used to call function entry/exit profile functions acquired through the ICorProfiler interface. Used in VERY rare circumstances. It is reasonable to wait to implement these until the final stages of productization. Most profilers do not use this functionality.

    15. JIT_PInvokeBegin/JIT_PInvokeEnd – Leave/enter the managed runtime state. Necessary for ReadyToRun pre-compiled pinvoke calls, so that they do not cause GC starvation

    16. VarargPInvokeStub/ GenericPInvokeCalliHelper Used to support calli pinvokes. It is expected that C# 8.0 will increase use of this feature. Today use of this feature on Unix requires hand-written IL. On Windows this feature is commonly used by C++/CLI

  3. EH Correctness. Some helpers are written in assembly to provide well known locations for NullReferenceExceptions to be generated out of a SIGSEGV signal.

    1. JIT_MemSet, and JIT_MemCpy have this requirement

cgencpu.h

This header is included by various code in the VM directory. It provides a large set of functionality that is architecture specific, including but not limited to

  1. Defines that are architecture specific specifying the sizes of various data structures the VM should create, and such

  2. Defines which specify which of various jit helpers should be replaced with asm functions instead of the portable C++ implementations

  3. The CalleeSavedRegisters, ArgumentRegisters, and FloatArgumentRegisters as needed to describe the calling convention for the platform

  4. The ClrFlushInstructionCache function. If the architecture doesn't actually need to manually flush the icache, then this function is empty.

  5. Various functions for decoding and manipulating jump instructions. These are used by various stub routines to predict where code will go, and to produce simple jump stubs.

  6. The StubLinkerCpu class for the architecture. Each Architecture defines its own StubLinkerCpu api surface and uses it to produce VM generated code. There is a small set of apis that are called from general purpose vm code (EmitComputedInstantiatingMethodStub, EmitShuffleThunkshared) across multiple architectures, and then there are the individual assembly instruction emission functions which are architecture specific. The StubLinker is used to generate complex stubs, where the set of assembly instructions emitted varies from stub to stub.

  7. Various stub data structures. Many very simple stubs are not emitted via an emission of a stream of bytes, but instead are exceptionally regular, and are effectively the same instructions for each different stub, only with slightly different data members. Instead of using the StubLinker mechanism, the VM instead has structures that represent the entirety of the stub and its associate data, and fill in the assembly instructions and data fields with a normal constructor call setting magic numbers. In addition to being executable, these stubs are often parsed to determine exactly what a given function is, what it is doing, where control flow will lead to, etc.

virtualcallstubcpu.h

This header is used to provide implementation of various stubs as used by virtual stub dispatch. These stubs are the lookup, resolver, and dispatch stubs as described in Virtual Stub Dispatch. These are maintained in a separate file from the rest of cgencpu.h for historical reasons, and for reasons of size (there is quite a lot of logic here.)

System.Private.CoreLib

Initial Bring up

In System.Private.CoreLib there is no work necessary for initial bring up.

Complete support

Complete support involves changing the publicly visible api surface of the product. Doing so is a process handled via public issues on GitHub and discussions with the api review board.

  • Adding support for the architecture to the System.Reflection.ImageFileMachine enum, and System.Reflection.ProcessorArchitecture enum as well as related logic

  • Adding support for architecture specific intrinsics such as SIMD instructions, or other non-standard api surface.

You can’t perform that action at this time.