[NativeAOT] Initial osx-arm64 bring up (#75264)

* Enable NativeAOT osx-arm64 builds

* Restrict alignment in VirtualReserveInner

* Fix generation and handling of compact unwinding info on osx-arm64

* Fix PC check in findFDE in DWARF CFI unwinding

* Handle MAP_JIT for P/Invoke on osx-arm64

* Handle P/Invoke with MAP_JIT on osx-arm64

* Fix incorrect OS_PAGE_SIZE definition

* Fix TLS register trashing

* Fix memory trashing caused by incorrect PREPARE_EXTERNAL_VAR_INDIRECT[_W] definitions

* Ignore ESRCH in PalHijack (thread is already gone)

* Make CompactUnwinder_* parametrized with registry class

* Remove custom CompactUnwinder

* Update llvm-libunwind-version.txt

* Fix initial alignment for __module_initializer. ARM64 requires absolute symbol relocations to be aligned to native pointer size.

* Allow publishing with the osx-arm64 RID

* Use pointer sized alignment for fat function pointers since they contain absolute relocations. ld64 seem to silently corrupt them if they are not aligned to pointer size.

Co-authored-by: Adeel Mujahid <[email protected]>
Co-authored-by: Jan Kotas <[email protected]>

Author: 3 people

eng/Subsets.props

@@ -99,7 +99,7 @@
 
   <PropertyGroup>
     <!-- CLR NativeAot only builds in a subset of the matrix -->
-    <NativeAotSupported Condition="('$(TargetOS)' == 'windows' or '$(TargetOS)' == 'linux' or '$(TargetOS)' == 'OSX') and ('$(TargetArchitecture)' == 'x64' or '$(TargetArchitecture)' == 'arm64') and ('$(TargetOS)' != 'OSX' or '$(TargetArchitecture)' != 'arm64')">true</NativeAotSupported>
+    <NativeAotSupported Condition="('$(TargetOS)' == 'windows' or '$(TargetOS)' == 'linux' or '$(TargetOS)' == 'OSX') and ('$(TargetArchitecture)' == 'x64' or '$(TargetArchitecture)' == 'arm64')">true</NativeAotSupported>
 
     <!-- If we're building clr.nativeaotlibs and not building the CLR runtime, compile libraries against NativeAOT CoreLib -->
     <UseNativeAotCoreLib Condition="$(_subset.Contains('+clr.nativeaotlibs+')) and !$(_subset.Contains('+clr.native+')) and !$(_subset.Contains('+clr.runtime+'))">true</UseNativeAotCoreLib>

src/coreclr/CMakeLists.txt

@@ -139,7 +139,7 @@ add_subdirectory(tools/aot/jitinterface)
 
 if(NOT CLR_CROSS_COMPONENTS_BUILD)
   # NativeAOT only buildable for a subset of CoreCLR-supported configurations
-  if((CLR_CMAKE_HOST_LINUX OR CLR_CMAKE_HOST_OSX OR CLR_CMAKE_HOST_WIN32) AND (CLR_CMAKE_HOST_ARCH_ARM64 OR CLR_CMAKE_HOST_ARCH_AMD64) AND NOT (CLR_CMAKE_HOST_OSX AND CLR_CMAKE_HOST_ARCH_ARM64))
+  if((CLR_CMAKE_HOST_LINUX OR CLR_CMAKE_HOST_OSX OR CLR_CMAKE_HOST_WIN32) AND (CLR_CMAKE_HOST_ARCH_ARM64 OR CLR_CMAKE_HOST_ARCH_AMD64))
     add_subdirectory(nativeaot)
   endif()
 endif(NOT CLR_CROSS_COMPONENTS_BUILD)

src/coreclr/gc/unix/gcenv.unix.cpp

@@ -633,7 +633,7 @@ void GCToOSInterface::YieldThread(uint32_t switchCount)
 static void* VirtualReserveInner(size_t size, size_t alignment, uint32_t flags, uint32_t hugePagesFlag = 0)
 {
     assert(!(flags & VirtualReserveFlags::WriteWatch) && "WriteWatch not supported on Unix");
-    if (alignment == 0)
+    if (alignment < OS_PAGE_SIZE)
     {
         alignment = OS_PAGE_SIZE;
     }

src/coreclr/nativeaot/BuildIntegration/Microsoft.NETCore.Native.Publish.targets

@@ -61,8 +61,7 @@
     <Error Condition="'$(DisableUnsupportedError)' != 'true' and '$(OS)' != 'Windows_NT' and $(RuntimeIdentifier.StartsWith('win'))"
       Text="Cross-OS native compilation is not supported." />
 
-    <Error Condition="'$(DisableUnsupportedError)' != 'true' and !($(RuntimeIdentifier.EndsWith('x64')) or
-      ($(RuntimeIdentifier.EndsWith('arm64')) and !$(RuntimeIdentifier.StartsWith('osx'))))"
+    <Error Condition="'$(DisableUnsupportedError)' != 'true' and !($(RuntimeIdentifier.EndsWith('x64')) or $(RuntimeIdentifier.EndsWith('arm64')))"
       Text="Native compilation does not support targeting $(RuntimeIdentifier) yet." />
 
     <Error Condition="'$(DisableUnsupportedError)' != 'true' and !('$(IlcHostArch)' == 'x64' or '$(IlcHostArch)' == 'arm64')"

src/coreclr/nativeaot/Runtime.Base/src/System/Runtime/ThunkPool.cs

@@ -40,9 +40,14 @@ namespace System.Runtime
 {
     internal static class Constants
     {
+#if TARGET_ARM64 && TARGET_OSX
+        public const uint PageSize = 0x4000;                   // 16k
+        public const nuint PageSizeMask = 0x3FFF;
+#else
         public const uint PageSize = 0x1000;                   // 4k
-        public const uint AllocationGranularity = 0x10000;     // 64k
         public const nuint PageSizeMask = 0xFFF;
+#endif
+        public const uint AllocationGranularity = 0x10000;     // 64k
         public const nuint AllocationGranularityMask = 0xFFFF;
 
         public static readonly int ThunkDataSize = 2 * IntPtr.Size;

src/coreclr/nativeaot/Runtime/CommonMacros.h

@@ -155,8 +155,12 @@ inline bool IS_ALIGNED(T* val, uintptr_t alignment);
 
 #define DATA_ALIGNMENT  8
 #ifndef OS_PAGE_SIZE
+#ifdef HOST_OSX
+#define OS_PAGE_SIZE    0x4000
+#else
 #define OS_PAGE_SIZE    0x1000
 #endif
+#endif
 
 #elif defined(HOST_WASM)
 

src/coreclr/nativeaot/Runtime/ThunksMapping.cpp

@@ -121,6 +121,9 @@ EXTERN_C NATIVEAOT_API void* __cdecl RhAllocateThunksMapping()
         return NULL;
     }
 
+#if defined(HOST_OSX) && defined(HOST_ARM64)
+    pthread_jit_write_protect_np(0);
+#endif
 #endif
 
     int numBlocksPerMap = RhpGetNumThunkBlocksPerMapping();
@@ -223,11 +226,15 @@ EXTERN_C NATIVEAOT_API void* __cdecl RhAllocateThunksMapping()
         }
     }
 
+#if defined(HOST_OSX) && defined(HOST_ARM64)
+    pthread_jit_write_protect_np(1);
+#else
     if (!PalVirtualProtect(pThunksSection, THUNKS_MAP_SIZE, PAGE_EXECUTE_READ))
     {
         PalVirtualFree(pNewMapping, 0, MEM_RELEASE);
         return NULL;
     }
+#endif
 
     PalFlushInstructionCache(pThunksSection, THUNKS_MAP_SIZE);
 

src/coreclr/nativeaot/Runtime/unix/PalRedhawkUnix.cpp

@@ -776,8 +776,16 @@ REDHAWK_PALEXPORT _Ret_maybenull_ _Post_writable_byte_size_(size) void* REDHAWK_
         static const size_t Alignment = 64 * 1024;
 
         size_t alignedSize = size + (Alignment - OS_PAGE_SIZE);
+        int flags = MAP_ANON | MAP_PRIVATE;
 
-        void * pRetVal = mmap(pAddress, alignedSize, unixProtect, MAP_ANON | MAP_PRIVATE, -1, 0);
+#if defined(HOST_OSX) && defined(HOST_ARM64)
+        if (unixProtect & PROT_EXEC)
+        {
+            flags |= MAP_JIT;
+        }
+#endif
+
+        void * pRetVal = mmap(pAddress, alignedSize, unixProtect, flags, -1, 0);
 
         if (pRetVal != NULL)
         {
@@ -1057,7 +1065,7 @@ REDHAWK_PALEXPORT void REDHAWK_PALAPI PalHijack(HANDLE hThread, _In_opt_ void* p
     }
 #endif
 
-    if ((status != 0) && (status != EAGAIN))
+    if ((status != 0) && (status != EAGAIN) && (status != ESRCH))
     {
         // Failure to send the signal is fatal. There are only two cases when sending
         // the signal can fail. First, if the signal ID is invalid and second,

src/coreclr/nativeaot/Runtime/unix/UnwindHelpers.cpp

@@ -28,10 +28,12 @@
 
 #if defined(TARGET_AMD64)
 using libunwind::Registers_x86_64;
+using libunwind::CompactUnwinder_x86_64;
 #elif defined(TARGET_ARM)
 using libunwind::Registers_arm;
 #elif defined(TARGET_ARM64)
 using libunwind::Registers_arm64;
+using libunwind::CompactUnwinder_arm64;
 #elif defined(TARGET_X86)
 using libunwind::Registers_x86;
 #else
@@ -501,6 +503,8 @@ struct Registers_REGDISPLAY : REGDISPLAY
     uint64_t    getIP() const         { return IP;}
     void        setIP(uint64_t value, uint64_t location)
     { IP = value; pIP = (PTR_UIntNative)location; }
+    uint64_t    getFP() const         { return *pFP;}
+    void        setFP(uint64_t value, uint64_t location) { pFP = (PTR_UIntNative)location;}
 };
 
 inline bool Registers_REGDISPLAY::validRegister(int num) const {
@@ -772,7 +776,37 @@ bool DoTheStep(uintptr_t pc, UnwindInfoSections uwInfoSections, REGDISPLAY *regs
 #endif
 
 #if _LIBUNWIND_SUPPORT_DWARF_UNWIND
-    bool retVal = uc.getInfoFromDwarfSection(pc, uwInfoSections, 0 /* fdeSectionOffsetHint */);
+    uint32_t dwarfOffsetHint = 0;
+
+#if _LIBUNWIND_SUPPORT_COMPACT_UNWIND
+    // If there is a compact unwind encoding table, look there first.
+    if (uwInfoSections.compact_unwind_section != 0 && uc.getInfoFromCompactEncodingSection(pc, uwInfoSections)) {
+        unw_proc_info_t procInfo;
+        uc.getInfo(&procInfo);
+
+#if defined(TARGET_ARM64)
+        if ((procInfo.format & UNWIND_ARM64_MODE_MASK) != UNWIND_ARM64_MODE_DWARF) {
+            CompactUnwinder_arm64<LocalAddressSpace, Registers_REGDISPLAY> compactInst;
+            int stepRet = compactInst.stepWithCompactEncoding(procInfo.format, pc, _addressSpace, *(Registers_REGDISPLAY*)regs);
+            return stepRet == UNW_STEP_SUCCESS;
+        } else {
+            dwarfOffsetHint = procInfo.format & UNWIND_ARM64_DWARF_SECTION_OFFSET;
+        }
+#elif defined(TARGET_AMD64)
+        if ((procInfo.format & UNWIND_X86_64_MODE_MASK) != UNWIND_X86_64_MODE_DWARF) {
+            CompactUnwinder_x86_64<LocalAddressSpace, Registers_REGDISPLAY> compactInst;
+            int stepRet = compactInst.stepWithCompactEncoding(procInfo.format, pc, _addressSpace, *(Registers_REGDISPLAY*)regs);
+            return stepRet == UNW_STEP_SUCCESS;
+        } else {
+            dwarfOffsetHint = procInfo.format & UNWIND_X86_64_DWARF_SECTION_OFFSET;
+        }
+#else
+        PORTABILITY_ASSERT("DoTheStep");
+#endif
+    }
+#endif
+
+    bool retVal = uc.getInfoFromDwarfSection(pc, uwInfoSections, dwarfOffsetHint);
     if (!retVal)
     {
         return false;

src/coreclr/nativeaot/Runtime/unix/unixasmmacrosarm64.inc

@@ -73,6 +73,7 @@ C_FUNC(\Name):
 #if defined(__APPLE__)
         adrp \HelperReg, C_FUNC(\Name)@GOTPAGE
         ldr  \HelperReg, [\HelperReg, C_FUNC(\Name)@GOTPAGEOFF]
+        ldr  \HelperReg, [\HelperReg]
 #else
         adrp \HelperReg, C_FUNC(\Name)
         ldr  \HelperReg, [\HelperReg, :lo12:C_FUNC(\Name)]
@@ -82,7 +83,8 @@ C_FUNC(\Name):
 .macro PREPARE_EXTERNAL_VAR_INDIRECT_W Name, HelperReg
 #if defined(__APPLE__)
         adrp x\HelperReg, C_FUNC(\Name)@GOTPAGE
-        ldr  w\HelperReg, [x\HelperReg, C_FUNC(\Name)@GOTPAGEOFF]
+        ldr  x\HelperReg, [x\HelperReg, C_FUNC(\Name)@GOTPAGEOFF]
+        ldr  w\HelperReg, [x\HelperReg]
 #else
         adrp x\HelperReg, C_FUNC(\Name)
         ldr  w\HelperReg, [x\HelperReg, :lo12:C_FUNC(\Name)]
@@ -182,6 +184,10 @@ C_FUNC(\Name):
     .endif
 
     stp     x0, lr, [sp,#-0x10]!
+#if defined(__APPLE__)
+    // Apple's tls_get_var trashes xip0 (and possibly xip1)
+    stp     xip0, xip1, [sp,#-0x10]!
+#endif
 
     // This sequence of instructions is recognized and potentially patched
     // by the linker (GD->IE/LE relaxation).
@@ -206,6 +212,9 @@ C_FUNC(\Name):
     add     \target, \target, x0
 #endif
 
+#if defined(__APPLE__)
+    ldp     xip0, xip1, [sp],#0x10
+#endif
     ldp     x0, lr, [sp],#0x10
 .endm