Revert partial changes for thread_local convergence.

src/coreclr/debug/ee/debugger.h

@@ -3835,7 +3835,7 @@ HANDLE OpenWin32EventOrThrow(
 #define SENDIPCEVENT_RAW_END SENDIPCEVENT_RAW_END_EX
 
 // Suspend-aware SENDIPCEVENT macros:
-// Check whether __thread has been suspended by the debugger via SetDebugState().
+// Check whether thread has been suspended by the debugger via SetDebugState().
 // If this thread has been suspended, it shouldn't send any event to the RS because the
 // debugger may not be expecting it.  Instead, just leave the lock and retry.
 // When we leave, we'll enter coop mode first and get suspended if a suspension is in progress.

src/coreclr/nativeaot/Runtime/CommonMacros.h

@@ -335,12 +335,6 @@ extern uint64_t g_startupTimelineEvents[NUM_STARTUP_TIMELINE_EVENTS];
 #define C_ASSERT(e) static_assert(e, #e)
 #endif // C_ASSERT
 
-#ifdef _MSC_VER
-#define DECLSPEC_THREAD __declspec(thread)
-#else // _MSC_VER
-#define DECLSPEC_THREAD __thread
-#endif // !_MSC_VER
-
 // PAL Numbers
 // Used to ensure cross-compiler compatibility when declaring large
 // integer constants. 64-bit integer constants should be wrapped in the

src/coreclr/nativeaot/Runtime/GCHelpers.cpp

@@ -405,7 +405,7 @@ FCIMPLEND
 
 // The MethodTable is remembered in some slow-path allocation paths. This value is used in event tracing.
 // It may statistically correlate with the most allocated type on the given stack/thread.
-DECLSPEC_THREAD
+static PLATFORM_THREAD_LOCAL
 MethodTable* tls_pLastAllocationEEType = NULL;
 
 MethodTable* GetLastAllocEEType()

src/coreclr/nativeaot/Runtime/eventpipe/ep-rt-aot.cpp

@@ -50,7 +50,7 @@ ep_rt_aot_walk_managed_stack_for_thread (
     EP_ASSERT (thread != NULL);
     EP_ASSERT (stack_contents != NULL);
 
-    StackFrameIterator frameIterator(thread, thread->GetTransitionFrameForSampling()); 
+    StackFrameIterator frameIterator(thread, thread->GetTransitionFrameForSampling());
 
     while (frameIterator.IsValid())
     {
@@ -223,11 +223,7 @@ ep_rt_aot_diagnostics_command_line_get (void)
 
 namespace
 {
-    #ifdef TARGET_UNIX
-    __thread EventPipeThreadHolder* eventpipe_tls_instance;
-    #else
-    thread_local EventPipeThreadHolder* eventpipe_tls_instance;
-    #endif
+    PLATFORM_THREAD_LOCAL EventPipeThreadHolder* eventpipe_tls_instance;
 
     void free_thread_holder ()
     {

src/coreclr/nativeaot/Runtime/threadstore.cpp

@@ -418,7 +418,7 @@ FCIMPLEND
 C_ASSERT(sizeof(Thread) == sizeof(RuntimeThreadLocals));
 
 #ifndef _MSC_VER
-__thread RuntimeThreadLocals tls_CurrentThread;
+PLATFORM_THREAD_LOCAL RuntimeThreadLocals tls_CurrentThread;
 #endif
 
 EXTERN_C RuntimeThreadLocals* RhpGetThread()

src/coreclr/nativeaot/Runtime/threadstore.inl

@@ -1,12 +1,12 @@
 // Licensed to the .NET Foundation under one or more agreements.
 // The .NET Foundation licenses this file to you under the MIT license.
 
+EXTERN_C
 #ifdef _MSC_VER
 // a workaround to prevent tls_CurrentThread from becoming dynamically checked/initialized.
-EXTERN_C __declspec(selectany) __declspec(thread) RuntimeThreadLocals tls_CurrentThread;
-#else
-EXTERN_C __thread RuntimeThreadLocals tls_CurrentThread;
+__declspec(selectany)
 #endif
+PLATFORM_THREAD_LOCAL RuntimeThreadLocals tls_CurrentThread;
 
 // static
 inline Thread * ThreadStore::RawGetCurrentThread()

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

@@ -481,7 +481,7 @@ thread_local TlsDestructionMonitor tls_destructionMonitor;
 #endif
 
 // This thread local variable is used for delegate marshalling
-DECLSPEC_THREAD intptr_t tls_thunkData;
+PLATFORM_THREAD_LOCAL intptr_t tls_thunkData;
 
 #ifdef FEATURE_EMULATED_TLS
 EXTERN_C intptr_t* RhpGetThunkData()
@@ -1061,48 +1061,45 @@ int32_t PalGetProcessCpuCount()
     return g_RhNumberOfProcessors;
 }
 
-__thread void* pStackHighOut = NULL;
-__thread void* pStackLowOut = NULL;
-
 // Retrieves the entire range of memory dedicated to the calling thread's stack.  This does
 // not get the current dynamic bounds of the stack, which can be significantly smaller than
 // the maximum bounds.
 bool PalGetMaximumStackBounds(_Out_ void** ppStackLowOut, _Out_ void** ppStackHighOut)
 {
-    if (pStackHighOut == NULL)
-    {
+    void* pStackHighOut = NULL;
+    void* pStackLowOut = NULL;
+
 #ifdef __APPLE__
-        // This is a Mac specific method
-        pStackHighOut = pthread_get_stackaddr_np(pthread_self());
-        pStackLowOut = ((uint8_t *)pStackHighOut - pthread_get_stacksize_np(pthread_self()));
+    // This is a Mac specific method
+    pStackHighOut = pthread_get_stackaddr_np(pthread_self());
+    pStackLowOut = ((uint8_t *)pStackHighOut - pthread_get_stacksize_np(pthread_self()));
 #else // __APPLE__
-        pthread_attr_t attr;
-        size_t stackSize;
-        int status;
+    pthread_attr_t attr;
+    size_t stackSize;
+    int status;
 
-        pthread_t thread = pthread_self();
+    pthread_t thread = pthread_self();
 
-        status = pthread_attr_init(&attr);
-        ASSERT_MSG(status == 0, "pthread_attr_init call failed");
+    status = pthread_attr_init(&attr);
+    ASSERT_MSG(status == 0, "pthread_attr_init call failed");
 
 #if HAVE_PTHREAD_ATTR_GET_NP
-        status = pthread_attr_get_np(thread, &attr);
+    status = pthread_attr_get_np(thread, &attr);
 #elif HAVE_PTHREAD_GETATTR_NP
-        status = pthread_getattr_np(thread, &attr);
+    status = pthread_getattr_np(thread, &attr);
 #else
 #error Dont know how to get thread attributes on this platform!
 #endif
-        ASSERT_MSG(status == 0, "pthread_getattr_np call failed");
+    ASSERT_MSG(status == 0, "pthread_getattr_np call failed");
 
-        status = pthread_attr_getstack(&attr, &pStackLowOut, &stackSize);
-        ASSERT_MSG(status == 0, "pthread_attr_getstack call failed");
+    status = pthread_attr_getstack(&attr, &pStackLowOut, &stackSize);
+    ASSERT_MSG(status == 0, "pthread_attr_getstack call failed");
 
-        status = pthread_attr_destroy(&attr);
-        ASSERT_MSG(status == 0, "pthread_attr_destroy call failed");
+    status = pthread_attr_destroy(&attr);
+    ASSERT_MSG(status == 0, "pthread_attr_destroy call failed");
 
-        pStackHighOut = (uint8_t*)pStackLowOut + stackSize;
+    pStackHighOut = (uint8_t*)pStackLowOut + stackSize;
 #endif // __APPLE__
-    }
 
     *ppStackLowOut = pStackLowOut;
     *ppStackHighOut = pStackHighOut;

src/coreclr/vm/amd64/AsmMacros.inc

@@ -467,7 +467,7 @@ OFFSETOF__ee_alloc_context  EQU OFFSETOF__RuntimeThreadLocals__ee_alloc_context
 ; Pushes a TransitionBlock on the stack without saving the argument registers. See
 ; the PROLOG_WITH_TRANSITION_BLOCK macro for the stack layout.
 PUSH_COOP_PINVOKE_FRAME macro target
-        
+
         __PWTB_StackAlloc = SIZEOF_MAX_OUTGOING_ARGUMENT_HOMES + 8 ; alignment to make the stack 16b aligned
         __PWTB_TransitionBlock = __PWTB_StackAlloc
 

src/coreclr/vm/ceemain.cpp

@@ -1664,7 +1664,7 @@ static uint32_t g_flsIndex = FLS_OUT_OF_INDEXES;
 #define FLS_STATE_ARMED 1
 #define FLS_STATE_INVOKED 2
 
-static thread_local byte t_flsState;
+static PLATFORM_THREAD_LOCAL byte t_flsState;
 
 // This is called when each *fiber* is destroyed. When the home fiber of a thread is destroyed,
 // it means that the thread itself is destroyed.

src/coreclr/vm/common.h

@@ -207,18 +207,6 @@ EXTERN_C AppDomain* STDCALL GetAppDomain();
 
 extern BOOL isMemoryReadable(const TADDR start, unsigned len);
 
-#ifndef memcpyUnsafe_f
-#define memcpyUnsafe_f
-
-// use this when you want to memcpy something that contains GC refs
-FORCEINLINE void* memcpyUnsafe(void *dest, const void *src, size_t len)
-{
-    WRAPPER_NO_CONTRACT;
-    return memcpy(dest, src, len);
-}
-
-#endif // !memcpyUnsafe_f
-
 FORCEINLINE void* memcpyNoGCRefs(void * dest, const void * src, size_t len)
 {
     WRAPPER_NO_CONTRACT;

src/coreclr/vm/jithelpers.cpp

@@ -344,15 +344,11 @@ HCIMPLEND
 //
 //========================================================================
 
-// Define the t_ThreadStatics variable here, so that these helpers can use
-// the most optimal TLS access pattern for the platform when inlining the
-// GetThreadLocalStaticBaseIfExistsAndInitialized function.
 // Using compiler specific thread local storage directives due to linkage issues.
 #ifdef _MSC_VER
-__declspec(selectany) __declspec(thread) ThreadLocalData t_ThreadStatics;
-#else
-__thread ThreadLocalData t_ThreadStatics;
+__declspec(selectany)
 #endif // _MSC_VER
+PLATFORM_THREAD_LOCAL ThreadLocalData t_ThreadStatics;
 
 extern "C" void QCALLTYPE GetThreadStaticsByMethodTable(QCall::ByteRefOnStack refHandle, MethodTable* pMT, bool gcStatic)
 {

src/coreclr/vm/threads.cpp

@@ -72,7 +72,7 @@ TailCallTls::TailCallTls()
 }
 
 #ifndef _MSC_VER
-__thread RuntimeThreadLocals t_runtime_thread_locals;
+PLATFORM_THREAD_LOCAL RuntimeThreadLocals t_runtime_thread_locals;
 #endif
 
 Thread* STDCALL GetThreadHelper()
@@ -345,7 +345,7 @@ bool Thread::DetectHandleILStubsForDebugger()
 }
 
 #ifndef _MSC_VER
-__thread ThreadLocalInfo t_CurrentThreadInfo;
+PLATFORM_THREAD_LOCAL ThreadLocalInfo t_CurrentThreadInfo;
 #endif // _MSC_VER
 
 #ifndef DACCESS_COMPILE

src/coreclr/vm/threads.h

@@ -331,13 +331,13 @@ void    DestroyThread(Thread *th);
 
 DWORD GetRuntimeId();
 
-#define CREATETHREAD_IF_NULL_FAILFAST(__thread, __msg)                  \
+#define CREATETHREAD_IF_NULL_FAILFAST(thread__, msg__)                  \
 {                                                                       \
-    HRESULT __ctinffhr;                                                 \
-    __thread = SetupThreadNoThrow(&__ctinffhr);                         \
-    if (__thread == NULL)                                               \
+    HRESULT ctinffhr__;                                                 \
+    thread__ = SetupThreadNoThrow(&ctinffhr__);                         \
+    if (thread__ == NULL)                                               \
     {                                                                   \
-        EEPOLICY_HANDLE_FATAL_ERROR_WITH_MESSAGE(__ctinffhr, __msg);    \
+        EEPOLICY_HANDLE_FATAL_ERROR_WITH_MESSAGE(ctinffhr__, msg__);    \
         UNREACHABLE();                                                  \
     }                                                                   \
 }
@@ -433,11 +433,11 @@ struct RuntimeThreadLocals
 
 #ifdef _MSC_VER
 // use selectany to avoid initialization de-optimization issues in the compiler
-__declspec(selectany) thread_local
+__declspec(selectany)
 #else
-extern __thread
+extern
 #endif
-RuntimeThreadLocals t_runtime_thread_locals;
+PLATFORM_THREAD_LOCAL RuntimeThreadLocals t_runtime_thread_locals;
 
 typedef DPTR(struct RuntimeThreadLocals) PTR_RuntimeThreadLocals;
 typedef DPTR(struct gc_alloc_context) PTR_gc_alloc_context;

src/coreclr/vm/threads.inl

@@ -27,11 +27,11 @@ EXTERN_C UINT32 _tls_index;
 #endif
 
 #ifdef _MSC_VER
-__declspec(selectany) thread_local
+__declspec(selectany)
 #else
-EXTERN_C __thread
+EXTERN_C
 #endif
-ThreadLocalInfo t_CurrentThreadInfo;
+PLATFORM_THREAD_LOCAL ThreadLocalInfo t_CurrentThreadInfo;
 
 inline Thread* GetThreadNULLOk()
 {

src/coreclr/vm/threadstatics.h

@@ -87,11 +87,11 @@ typedef DPTR(ThreadLocalData) PTR_ThreadLocalData;
 
 // Using compiler specific thread local storage directives due to linkage issues.
 #ifndef DACCESS_COMPILE
+extern
 #ifdef _MSC_VER
-extern __declspec(selectany) __declspec(thread) ThreadLocalData t_ThreadStatics;
-#else
-extern __thread ThreadLocalData t_ThreadStatics;
+__declspec(selectany)
 #endif // _MSC_VER
+PLATFORM_THREAD_LOCAL ThreadLocalData t_ThreadStatics;
 #endif // !DACCESS_COMPILE
 
 #define NUMBER_OF_TLSOFFSETS_NOT_USED_IN_NONCOLLECTIBLE_ARRAY 2

src/native/minipal/utils.h

@@ -44,6 +44,17 @@
 #define LIBC_CALLBACK
 #endif
 
+// Define a macro to declare TLS variables. There are cases
+// where we want to use the platform specific thread-local
+// storage mechanism:
+//  * Better performance characteristics compared to C++'s thread_local keyword.
+//  * Makes consuming TLS variables from assembly code easier.
+#if defined(_MSC_VER)
+#define PLATFORM_THREAD_LOCAL __declspec(thread)
+#else
+#define PLATFORM_THREAD_LOCAL __thread
+#endif
+
 #if defined(_MSC_VER)
 #  if defined(__SANITIZE_ADDRESS__)
 #    define HAS_ADDRESS_SANITIZER