[Linux/arm32 R2R] Cross-Bitness Support for CrossGen

[echesakov]

Currently, when build Linux/arm coreclr on Linux/x64 dev machine the following two sets of user environments must be installed:

1. x64/arm compiler for native build;
2. x64/x84 for cross-components used for crossgen/AOT to improve performance.

Purpose: Adapting crossgen and coreclr to be able to generate 32-bit ReadyToRun images on a 64-bit host without requiring to install x64/x84 Linux environment.

Backlog:

• Replace expessions like sizeof(void*) with target specific macro TARGET_POINTER_SIZE ([ToFDisable] Port tests for: System.Security.Cryptography.Primitives #16514 Replace sizeof(TADDR) with TARGET_POINTER_SIZE in Zap coreclr#18028)
• Replace usages of macro _WIN64 with _TARGET_64BIT_ (as we do in RyuJIT) (Remove problematic ArrayPool usage #16531 Use _TARGET_64BIT_ macro coreclr#16660)
• Use emitter::code_t rather than ssize_t and size_t in emitarm.cpp and emitarm.h (PostAsync coreclr send different request compare to PostAsync clr #16617)
• Initial cross-bitness support for RyuJIT (WinHttp buffer supplied to the function was too small error should not be exposed #17005)
• Enable building of Ubuntu/arm crossarch components in CI and official build (Unix build breaks with package errors on recent buildtools update #16729 Disable fragile NGen for System.Private.CoreLib.dll on ARM coreclr#16839 Include Linux/cross-arch components into NuGet packages coreclr#16859 Publish x86_arm/crossgen during Linux/arm build coreclr#16953 Enable building of Ubuntu/arm crossarch components in netci coreclr#17375)
• Replace sizeof(TADDR) with TARGET_POINTER_SIZE in Zapper (Add API T:System.Xml.Serialization.XmlSerializationGeneratedCode  #18028)
• Replace sizeof(ObjHeader) sizeof(Object) etc with target-specific constants OBJHEADER_SIZE OBJECT_SIZE (Release these reserved names: COM1-COM9, LPT1-LPT9, CON, AUX, PRN, and NUL #18245)
• Use target-specific OFFSETOF__DelegateObject__target OFFSETOF__DelegateObject__methodPtr constants etc in JitInterface (LambdaCompiler could reduce Closure.Constants size #18329)
• Use target-specific constants during instance fields placement (Building System.Private.Xml with TargetGroup=netcore50aot Failed #18366)
• Use target-specific constants during regular and thread static fields placement (Port the rest of HttpWebRequest/Response APIs #18636)
• Use TARGET_POINTER_TYPE TARGET_POINTER_SIZE during ZapWriteReloc ([Ping] Consider adding synchronous path to Ping #18665)
• Code cleanup in ARM32 and ARM64 emitters (Remove temporary project references after changes to the referenced project are published #18878)
• Handle MovRelocatableImmediate as a special case (IF_T2_N3) in JIT (Create implementation of HandleCount on Unix #19013)
• Stop using size_t in GcInfoEncoder::Build() (Move types from S.Security.Permissions that are part of netstandard 2.0 #19206)
• Initial support for cross-bitness crossgen validation against native crossgen ([WinHttp] WebSocket.ReceiveAsync throws internal exception type WinHttpException #19514)
• Return target-specific NaN during constant folding in JIT when target ARM32/ARM64 (Build for x86 on Windows runs tests using 64-bit runtime #19395)
• Disable constant folding in JIT resulting in NaN, infinities etc
  (CI Failure System.Data.SqlClient.Tests.SqlConnectionBasicTests.ConnectionTest #19483)
• Split CodeGen::instGen_Set_Reg_To_Imm into relocatable immediate/non-relocatable immediate cases (CI failure in DateTimeFormatInfoGetInstance.GetInstance_ExpectedCurrent #19091)
• Add support in CMakeLists.txt, src/pal/src/CMakeLists.txt, build.sh and build.cmd for building x64_arm crosscomponents
• Enable some definitions and disable unreachable code causing compilation warnings/errors, do upcasts to suppress such warnings during cross-bitness crossgen compilation (ConcurrentDictionary performance issue #18707):

• For example, RtlUnwindEx RtlVirtualUnwind are needed during x86_arm on Windows and on x64_arm on Linux but not on x64_arm on Windows
• Many functions in src/inc/regdisp.h never reachable from crossgen but causes compilation errors

• Consider splitting GenTreeIntCon into "a tree node that holds handles (i.e. pointers to host functions)" and "a tree node that never holds host pointers - i.e. pure integer constant"

• Handling VNForPtrSizeIntCon and genTreeHashAdd in JIT in crossbitness scenario

• Even after these some code in JIT would cause compilation warnings - requires either explicit casting or refactoring (e.g. roundUp)

• EEClass::AddChunk must append methods (not prepend them as it now) in order to produce same output in both same-bitness and cross-bitness scenarios

[BredPet]

@dotnet/arm32-contrib

[AndyAyersMS]

Is this something we want to track for 2.1?

[RussKeldorph]

@AndyAyersMS I'd like to get this in for 2.1 if possible. For preview 2, we are going to have to use a workaround (using 32-bit tools and new Docker images), but it could still make the release. Since we can validate by simple binary output comparison on a fixed set of inputs (the framework binaries), it should be low risk.

[RussKeldorph]

Getting late for 2.1, and we should be unblocked with the x86 work, so moving out. If it becomes ready, we may consider for 2.1.