[NativeAOT] ThreadStatics part 2

src/coreclr/nativeaot/Runtime/RuntimeInstance.cpp

@@ -267,25 +267,6 @@ RuntimeInstance::TypeManagerList& RuntimeInstance::GetTypeManagerList()
     return m_TypeManagerList;
 }
 
-TypeManager* RuntimeInstance::GetSingleTypeManager()
-{
-    auto head = m_TypeManagerList.GetHead();
-    if (head != NULL && head->m_pNext == NULL)
-    {
-        return head->m_pTypeManager;
-    }
-
-    return NULL;
-}
-
-COOP_PINVOKE_HELPER(TypeManagerHandle, RhGetSingleTypeManager, ())
-{
-    TypeManager* typeManager = GetRuntimeInstance()->GetSingleTypeManager();
-    ASSERT(typeManager != NULL);
-
-    return TypeManagerHandle::Create(typeManager);
-}
-
 // static
 bool RuntimeInstance::Initialize(HANDLE hPalInstance)
 {

src/coreclr/nativeaot/Runtime/RuntimeInstance.h

@@ -99,7 +99,6 @@ class RuntimeInstance
 
     bool RegisterTypeManager(TypeManager * pTypeManager);
     TypeManagerList& GetTypeManagerList();
-    TypeManager* GetSingleTypeManager();
     OsModuleList* GetOsModuleList();
 
     bool RegisterUnboxingStubs(PTR_VOID pvStartRange, uint32_t cbRange);

src/coreclr/nativeaot/Runtime/amd64/MiscStubs.S

@@ -44,22 +44,3 @@ LOCAL_LABEL(ProbeLoop):
         RESET_FRAME_WITH_RBP
         ret
 NESTED_END RhpStackProbe, _TEXT
-
-#ifndef TARGET_ANDROID
-NESTED_ENTRY RhpGetInlinedThreadStaticBase, _TEXT, NoHandler
-        // On exit:
-        //   rax - the thread static base for the given type
-
-        // rdi = &tls_InlinedThreadStatics
-        INLINE_GET_TLS_VAR tls_InlinedThreadStatics
-        mov     rdi, rax
-
-        // get per-thread storage
-        mov     rax, [rdi]
-        test    rax, rax
-        jz      C_FUNC(RhpGetInlinedThreadStaticBaseSlow)    // rdi contains the storage ref
-
-        // return it
-        ret
-NESTED_END RhpGetInlinedThreadStaticBase, _TEXT
-#endif

src/coreclr/nativeaot/Runtime/amd64/MiscStubs.asm

@@ -3,8 +3,6 @@
 
 include AsmMacros.inc
 
-EXTERN RhpGetInlinedThreadStaticBaseSlow : PROC
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; The following helper will access ("probe") a word on each page of the stack
 ; starting with the page right beneath rsp down to the one pointed to by r11.
@@ -39,20 +37,4 @@ ProbeLoop:
 
 LEAF_END RhpStackProbe, _TEXT
 
-LEAF_ENTRY RhpGetInlinedThreadStaticBase, _TEXT
-        ; On exit:
-        ;   rax - the thread static base for the given type
-
-        ;; rcx = &tls_InlinedThreadStatics, TRASHES r8
-        INLINE_GET_TLS_VAR rcx, r8, tls_InlinedThreadStatics
-
-        ;; get per-thread storage
-        mov     rax, [rcx]
-        test    rax, rax
-        jz      RhpGetInlinedThreadStaticBaseSlow   ;; rcx contains the storage ref
-
-        ;; return it
-        ret
-LEAF_END RhpGetInlinedThreadStaticBase, _TEXT
-
 end

src/coreclr/nativeaot/Runtime/arm64/MiscStubs.S

@@ -3,24 +3,3 @@
 
 #include <unixasmmacros.inc>
 #include "AsmOffsets.inc"
-
-#ifndef TARGET_ANDROID
-NESTED_ENTRY RhpGetInlinedThreadStaticBase, _TEXT, NoHandler
-        // On exit:
-        //   x0 - the thread static base for the given type
-
-        // x1 = GetThread()
-        INLINE_GET_TLS_VAR x1, C_FUNC(tls_InlinedThreadStatics)
-
-        // get per-thread storage
-        ldr     x0, [x1]
-        cbnz    x0, HaveValue
-        mov     x0, x1
-        b       C_FUNC(RhpGetInlinedThreadStaticBaseSlow)
-
-HaveValue:
-        // return it
-        ret
-
-NESTED_END RhpGetInlinedThreadStaticBase, _TEXT
-#endif

src/coreclr/nativeaot/Runtime/arm64/MiscStubs.asm

@@ -3,25 +3,6 @@
 
 #include "AsmMacros.h"
 
-    EXTERN RhpGetInlinedThreadStaticBaseSlow
-
     TEXTAREA
 
-;; On exit:
-;;   x0 - the thread static base for the given type
-    LEAF_ENTRY RhpGetInlinedThreadStaticBase
-        ;; x1 = &tls_InlinedThreadStatics, TRASHES x2
-        INLINE_GET_TLS_VAR x1, x2, tls_InlinedThreadStatics
-
-        ;; get per-thread storage
-        ldr     x0, [x1]
-        cbnz    x0, HaveValue
-        mov     x0, x1
-        b       RhpGetInlinedThreadStaticBaseSlow
-
-HaveValue
-        ;; return it
-        ret
-    LEAF_END RhpGetInlinedThreadStaticBase
-
     end

src/coreclr/nativeaot/Runtime/threadstore.cpp

@@ -430,23 +430,13 @@ C_ASSERT(sizeof(Thread) == sizeof(ThreadBuffer));
 
 #ifndef _MSC_VER
 __thread ThreadBuffer tls_CurrentThread;
-
-// the root of inlined threadstatics storage
-// there is only one now,
-// eventually this will be emitted by ILC and we may have more than one such variable
-__thread InlinedThreadStaticRoot tls_InlinedThreadStatics;
 #endif
 
 EXTERN_C ThreadBuffer* RhpGetThread()
 {
     return &tls_CurrentThread;
 }
 
-COOP_PINVOKE_HELPER(Object**, RhGetInlinedThreadStaticStorage, ())
-{
-    return &tls_InlinedThreadStatics.m_threadStaticsBase;
-}
-
 #endif // !DACCESS_COMPILE
 
 #ifdef _WIN32

src/coreclr/nativeaot/Runtime/threadstore.inl

@@ -4,14 +4,8 @@
 #ifdef _MSC_VER
 // a workaround to prevent tls_CurrentThread from becoming dynamically checked/initialized.
 EXTERN_C __declspec(selectany) __declspec(thread) ThreadBuffer tls_CurrentThread;
-
-// the root of inlined threadstatics storage
-// there is only one now,
-// eventually this will be emitted by ILC and we may have more than one such variable
-EXTERN_C __declspec(selectany) __declspec(thread) InlinedThreadStaticRoot tls_InlinedThreadStatics;
 #else
 EXTERN_C __thread ThreadBuffer tls_CurrentThread;
-EXTERN_C __thread InlinedThreadStaticRoot tls_InlinedThreadStatics;
 #endif
 
 // static

src/coreclr/nativeaot/Runtime/unix/unixasmmacrosamd64.inc

@@ -288,9 +288,12 @@ C_FUNC(\Name):
         movq    _\Var@TLVP(%rip), %rdi
         callq   *(%rdi)
 #else
-        leaq    \Var@TLSLD(%rip), %rdi
+        .byte 0x66  // data16 prefix - padding to have space for linker relaxations
+        leaq    \Var@TLSGD(%rip), %rdi
+        .byte 0x66  //
+        .byte 0x66  //
+        .byte 0x48  // rex.W prefix, also for padding
         callq   __tls_get_addr@PLT
-        addq    $\Var@DTPOFF, %rax
 #endif
        .intel_syntax noprefix
 .endm

src/coreclr/nativeaot/System.Private.CoreLib/src/Internal/Runtime/ThreadStatics.cs

@@ -17,35 +17,35 @@ namespace Internal.Runtime
     /// </summary>
     internal static class ThreadStatics
     {
+        [ThreadStatic]
+        private static object t_inlinedThreadStaticBase;
+
         /// <summary>
         /// This method is called from a ReadyToRun helper to get base address of thread
         /// static storage for the given type.
         /// </summary>
         internal static unsafe object GetThreadStaticBaseForType(TypeManagerSlot* pModuleData, int typeTlsIndex)
         {
-            if (typeTlsIndex >= 0)
-                return GetUninlinedThreadStaticBaseForType(pModuleData, typeTlsIndex);
-
-            ref object? threadStorage = ref RuntimeImports.RhGetInlinedThreadStaticStorage();
-            if (threadStorage != null)
-                return threadStorage;
+            if (typeTlsIndex < 0)
+                return t_inlinedThreadStaticBase;
 
-            return GetInlinedThreadStaticBaseSlow(ref threadStorage);
+            return GetUninlinedThreadStaticBaseForType(pModuleData, typeTlsIndex);
         }
 
-        [RuntimeExport("RhpGetInlinedThreadStaticBaseSlow")]
         internal static unsafe object GetInlinedThreadStaticBaseSlow(ref object? threadStorage)
         {
             Debug.Assert(threadStorage == null);
             // Allocate an object that will represent a memory block for all thread static fields
-            TypeManagerHandle typeManager = RuntimeImports.RhGetSingleTypeManager();
+            TypeManagerHandle typeManager = (new object()).GetMethodTable()->TypeManager;
             object threadStaticBase = AllocateThreadStaticStorageForType(typeManager, 0);
 
             // register the storage location with the thread for GC reporting.
             RuntimeImports.RhRegisterInlinedThreadStaticRoot(ref threadStorage);
 
             // assign the storage block to the storage variable and return
             threadStorage = threadStaticBase;
+            t_inlinedThreadStaticBase = threadStaticBase;
+
             return threadStaticBase;
         }
 
@@ -55,13 +55,13 @@ internal static unsafe object GetUninlinedThreadStaticBaseForType(TypeManagerSlo
             int moduleIndex = pModuleData->ModuleIndex;
             Debug.Assert(moduleIndex >= 0);
 
-            object[][] threadStorage = RuntimeImports.RhGetThreadStaticStorage();
-            if (threadStorage != null && threadStorage.Length > moduleIndex)
+            object[][] perThreadStorage = RuntimeImports.RhGetThreadStaticStorage();
+            if (perThreadStorage != null && perThreadStorage.Length > moduleIndex)
             {
-                object[] moduleStorage = threadStorage[moduleIndex];
-                if (moduleStorage != null && moduleStorage.Length > typeTlsIndex)
+                object[] perModuleStorage = perThreadStorage[moduleIndex];
+                if (perModuleStorage != null && perModuleStorage.Length > typeTlsIndex)
                 {
-                    object threadStaticBase = moduleStorage[typeTlsIndex];
+                    object threadStaticBase = perModuleStorage[typeTlsIndex];
                     if (threadStaticBase != null)
                     {
                         return threadStaticBase;

src/coreclr/nativeaot/System.Private.CoreLib/src/System/Runtime/RuntimeImports.cs

@@ -560,10 +560,6 @@ internal static IntPtr RhGetModuleSection(TypeManagerHandle module, ReadyToRunSe
         [RuntimeImport(RuntimeLibrary, "RhGetThreadStaticStorage")]
         internal static extern ref object[][] RhGetThreadStaticStorage();
 
-        [MethodImplAttribute(MethodImplOptions.InternalCall)]
-        [RuntimeImport(RuntimeLibrary, "RhGetInlinedThreadStaticStorage")]
-        internal static extern ref object? RhGetInlinedThreadStaticStorage();
-
         [MethodImplAttribute(MethodImplOptions.InternalCall)]
         [RuntimeImport(RuntimeLibrary, "RhRegisterInlinedThreadStaticRoot")]
         internal static extern void RhRegisterInlinedThreadStaticRoot(ref object? root);
@@ -592,10 +588,6 @@ internal static IntPtr RhGetModuleSection(TypeManagerHandle module, ReadyToRunSe
         [RuntimeImport(RuntimeLibrary, "RhGetTargetOfUnboxingAndInstantiatingStub")]
         public static extern IntPtr RhGetTargetOfUnboxingAndInstantiatingStub(IntPtr pCode);
 
-        [MethodImplAttribute(MethodImplOptions.InternalCall)]
-        [RuntimeImport(RuntimeLibrary, "RhGetSingleTypeManager")]
-        public static extern TypeManagerHandle RhGetSingleTypeManager();
-
         //
         // EH helpers
         //

src/coreclr/tools/Common/Compiler/DependencyAnalysis/ObjectDataBuilder.cs

@@ -291,6 +291,8 @@ public void EmitReloc(ISymbolNode symbol, RelocType relocType, int delta = 0)
                 case RelocType.IMAGE_REL_BASED_ABSOLUTE:
                 case RelocType.IMAGE_REL_BASED_HIGHLOW:
                 case RelocType.IMAGE_REL_SECREL:
+                case RelocType.IMAGE_REL_TLSGD:
+                case RelocType.IMAGE_REL_TPOFF:
                 case RelocType.IMAGE_REL_FILE_ABSOLUTE:
                 case RelocType.IMAGE_REL_BASED_ADDR32NB:
                 case RelocType.IMAGE_REL_SYMBOL_SIZE:
@@ -305,6 +307,14 @@ public void EmitReloc(ISymbolNode symbol, RelocType relocType, int delta = 0)
                 case RelocType.IMAGE_REL_BASED_ARM64_PAGEBASE_REL21:
                 case RelocType.IMAGE_REL_BASED_ARM64_PAGEOFFSET_12L:
                 case RelocType.IMAGE_REL_BASED_ARM64_PAGEOFFSET_12A:
+
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_ADR_PAGE21:
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_LD64_LO12:
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_ADD_LO12:
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_CALL:
+                case RelocType.IMAGE_REL_AARCH64_TLSLE_ADD_TPREL_HI12:
+                case RelocType.IMAGE_REL_AARCH64_TLSLE_ADD_TPREL_LO12_NC:
+
                 case RelocType.IMAGE_REL_BASED_LOONGARCH64_PC:
                 case RelocType.IMAGE_REL_BASED_LOONGARCH64_JIR:
                     Debug.Assert(delta == 0);

src/coreclr/tools/Common/Compiler/DependencyAnalysis/ObjectNodeSection.cs

@@ -46,7 +46,7 @@ public bool IsStandardSection
         public static readonly ObjectNodeSection ReadOnlyDataSection = new ObjectNodeSection("rdata", SectionType.ReadOnly);
         public static readonly ObjectNodeSection FoldableReadOnlyDataSection = new ObjectNodeSection("rdata", SectionType.ReadOnly);
         public static readonly ObjectNodeSection TextSection = new ObjectNodeSection("text", SectionType.Executable);
-        public static readonly ObjectNodeSection TLSSection = new ObjectNodeSection("TLS", SectionType.Writeable);
+        public static readonly ObjectNodeSection TLSSection = new ObjectNodeSection("tdata", SectionType.Writeable);
         public static readonly ObjectNodeSection BssSection = new ObjectNodeSection("bss", SectionType.Uninitialized);
         public static readonly ObjectNodeSection HydrationTargetSection = new ObjectNodeSection("hydrated", SectionType.Uninitialized);
         public static readonly ObjectNodeSection ManagedCodeWindowsContentSection = new ObjectNodeSection(".managedcode$I", SectionType.Executable);

src/coreclr/tools/Common/Compiler/DependencyAnalysis/Relocation.cs

@@ -23,12 +23,34 @@ public enum RelocType
                                                        // This is a special NGEN-specific relocation type
                                                        // for relative pointer (used to make NGen relocation
                                                        // section smaller)
-        IMAGE_REL_SECREL                     = 0x80,   // 32 bit offset from base of section containing target
 
         IMAGE_REL_BASED_ARM64_PAGEBASE_REL21 = 0x81,   // ADRP
         IMAGE_REL_BASED_ARM64_PAGEOFFSET_12A = 0x82,   // ADD/ADDS (immediate) with zero shift, for page offset
         IMAGE_REL_BASED_ARM64_PAGEOFFSET_12L = 0x83,   // LDR (indexed, unsigned immediate), for page offset
 
+        //
+        // Relocation operators related to TLS access
+        //
+
+        // Windows x64
+        IMAGE_REL_SECREL                     = 0x104,
+
+        // Linux x64
+        // GD model
+        IMAGE_REL_TLSGD                      = 0x105,
+        // LE model
+        IMAGE_REL_TPOFF                      = 0x106,
+
+        // Linux arm64
+        //    TLSDESC  (dynamic)
+        IMAGE_REL_AARCH64_TLSDESC_ADR_PAGE21 = 0x107,
+        IMAGE_REL_AARCH64_TLSDESC_LD64_LO12  = 0x108,
+        IMAGE_REL_AARCH64_TLSDESC_ADD_LO12   = 0x109,
+        IMAGE_REL_AARCH64_TLSDESC_CALL       = 0x10A,
+        //    LE model
+        IMAGE_REL_AARCH64_TLSLE_ADD_TPREL_HI12    = 0x10B,
+        IMAGE_REL_AARCH64_TLSLE_ADD_TPREL_LO12_NC = 0x10C,
+
         //
         // Relocations for R2R image production
         //
@@ -459,6 +481,8 @@ public static unsafe long ReadValue(RelocType relocType, void* location)
                 case RelocType.IMAGE_REL_BASED_REL32:
                 case RelocType.IMAGE_REL_BASED_RELPTR32:
                 case RelocType.IMAGE_REL_SECREL:
+                case RelocType.IMAGE_REL_TLSGD:
+                case RelocType.IMAGE_REL_TPOFF:
                 case RelocType.IMAGE_REL_FILE_ABSOLUTE:
                 case RelocType.IMAGE_REL_SYMBOL_SIZE:
                     return *(int*)location;
@@ -475,6 +499,20 @@ public static unsafe long ReadValue(RelocType relocType, void* location)
                     return GetArm64Rel21((uint*)location);
                 case RelocType.IMAGE_REL_BASED_ARM64_PAGEOFFSET_12A:
                     return GetArm64Rel12((uint*)location);
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_LD64_LO12:
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_ADD_LO12:
+                case RelocType.IMAGE_REL_AARCH64_TLSLE_ADD_TPREL_HI12:
+                case RelocType.IMAGE_REL_AARCH64_TLSLE_ADD_TPREL_LO12_NC:
+                    // TLS relocs do not have offsets
+                    Debug.Assert((GetArm64Rel12((uint*)location) & 0xFF) == 0);
+                    return 0;
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_ADR_PAGE21:
+                    // TLS relocs do not have offsets
+                    Debug.Assert((GetArm64Rel21((uint*)location) & 0xFF) == 0);
+                    return 0;
+                case RelocType.IMAGE_REL_AARCH64_TLSDESC_CALL:
+                    // TLS relocs do not have offsets
+                    return 0;
                 case RelocType.IMAGE_REL_BASED_LOONGARCH64_PC:
                     return (long)GetLoongArch64PC12((uint*)location);
                 case RelocType.IMAGE_REL_BASED_LOONGARCH64_JIR: