Delete !FEATURE_IMPLICIT_TLS (#14398)

Linux and Windows arm64 are using the regular C/C++ thread local statics. This change unifies the remaining Windows architectures to be on the same plan.

8 files changed, 40 insertions, 378 deletions

diff --git a/src/vm/i386/asmconstants.h b/src/vm/i386/asmconstants.h
index f7d5f709dc..aa420428f7 100644
--- a/src/vm/i386/asmconstants.h
+++ b/src/vm/i386/asmconstants.h
@@ -317,9 +317,6 @@ ASMCONSTANTS_C_ASSERT(ASM__VTABLE_SLOTS_PER_CHUNK == VTABLE_SLOTS_PER_CHUNK)
 #define ASM__VTABLE_SLOTS_PER_CHUNK_LOG2 3
 ASMCONSTANTS_C_ASSERT(ASM__VTABLE_SLOTS_PER_CHUNK_LOG2 == VTABLE_SLOTS_PER_CHUNK_LOG2)
 
-#define TLS_GETTER_MAX_SIZE_ASM DBG_FRE(0x20, 0x10)
-ASMCONSTANTS_C_ASSERT(TLS_GETTER_MAX_SIZE_ASM == TLS_GETTER_MAX_SIZE)
-
 #define JIT_TailCall_StackOffsetToFlags       0x08
 
 #define CallDescrData__pSrc                0x00
diff --git a/src/vm/i386/asmhelpers.asm b/src/vm/i386/asmhelpers.asm
index 9df33219ac..17d521fb92 100644
--- a/src/vm/i386/asmhelpers.asm
+++ b/src/vm/i386/asmhelpers.asm
@@ -47,23 +47,10 @@ endif ; FEATURE_COMINTEROP
 EXTERN __alloca_probe:PROC
 EXTERN _NDirectImportWorker@4:PROC
 EXTERN _UMThunkStubRareDisableWorker@8:PROC
-ifndef FEATURE_IMPLICIT_TLS
-ifdef ENABLE_GET_THREAD_GENERIC_FULL_CHECK
-; This is defined in C (threads.cpp) and enforces EE_THREAD_NOT_REQUIRED contracts
-GetThreadGenericFullCheck EQU ?GetThreadGenericFullCheck@@YGPAVThread@@XZ
-EXTERN  GetThreadGenericFullCheck:PROC
-endif ; ENABLE_GET_THREAD_GENERIC_FULL_CHECK
-
-EXTERN _gThreadTLSIndex:DWORD
-EXTERN _gAppDomainTLSIndex:DWORD
-endif ; FEATURE_IMPLICIT_TLS
 
 EXTERN _VarargPInvokeStubWorker@12:PROC
 EXTERN _GenericPInvokeCalliStubWorker@12:PROC
 
-EXTERN _GetThread@0:PROC
-EXTERN _GetAppDomain@0:PROC
-
 ifdef MDA_SUPPORTED
 EXTERN _PInvokeStackImbalanceWorker@8:PROC
 endif
@@ -716,56 +703,6 @@ doRet:
 FASTCALL_ENDFUNC HelperMethodFrameRestoreState
 
 
-ifndef FEATURE_IMPLICIT_TLS
-;---------------------------------------------------------------------------
-; Portable GetThread() function: used if no platform-specific optimizations apply.
-; This is in assembly code because we count on edx not getting trashed on calls
-; to this function.
-;---------------------------------------------------------------------------
-; Thread* __stdcall GetThreadGeneric(void);
-GetThreadGeneric PROC stdcall public USES ecx edx
-
-ifdef _DEBUG
-    cmp         dword ptr [_gThreadTLSIndex], -1
-    jnz         @F
-    int         3
-@@:
-endif
-ifdef ENABLE_GET_THREAD_GENERIC_FULL_CHECK
-    ; non-PAL, debug-only GetThreadGeneric should defer to GetThreadGenericFullCheck
-    ; to do extra contract enforcement.  (See GetThreadGenericFullCheck for details.)
-    ; This code is intentionally not added to asmhelper.s, as this enforcement is only
-    ; implemented for non-PAL builds.
-    call        GetThreadGenericFullCheck
-else
-    push        dword ptr [_gThreadTLSIndex]
-    call        dword ptr [__imp__TlsGetValue@4]
-endif
-    ret
-GetThreadGeneric ENDP
-
-;---------------------------------------------------------------------------
-; Portable GetAppdomain() function: used if no platform-specific optimizations apply.
-; This is in assembly code because we count on edx not getting trashed on calls
-; to this function.
-;---------------------------------------------------------------------------
-; Appdomain* __stdcall GetAppDomainGeneric(void);
-GetAppDomainGeneric PROC stdcall public USES ecx edx
-
-ifdef _DEBUG
-    cmp         dword ptr [_gAppDomainTLSIndex], -1
-    jnz         @F
-    int         3
-@@:
-endif
-
-    push        dword ptr [_gAppDomainTLSIndex]
-    call        dword ptr [__imp__TlsGetValue@4]
-    ret
-GetAppDomainGeneric ENDP
-endif
-
-
 ifdef FEATURE_HIJACK
 
 ; A JITted method's return address was hijacked to return to us here.  
diff --git a/src/vm/i386/cgencpu.h b/src/vm/i386/cgencpu.h
index 5360b3eb0e..ffdfb82b14 100644
--- a/src/vm/i386/cgencpu.h
+++ b/src/vm/i386/cgencpu.h
@@ -108,14 +108,6 @@ BOOL Runtime_Test_For_SSE2();
 #define ENREGISTERED_RETURNTYPE_INTEGER_MAXSIZE 4
 #define CALLDESCR_ARGREGS                       1   // CallDescrWorker has ArgumentRegister parameter
 
-// Max size of patched TLS helpers
-#ifdef _DEBUG
-// Debug build needs extra space for last error trashing
-#define TLS_GETTER_MAX_SIZE 0x20
-#else
-#define TLS_GETTER_MAX_SIZE 0x10
-#endif
-
 //=======================================================================
 // IMPORTANT: This value is used to figure out how much to allocate
 // for a fixed array of FieldMarshaler's. That means it must be at least
@@ -558,6 +550,12 @@ inline BOOL ClrFlushInstructionCache(LPCVOID pCodeAddr, size_t sizeOfCode)
     return TRUE;
 }
 
+//
+// JIT HELPER ALIASING FOR PORTABILITY.
+//
+// Create alias for optimized implementations of helpers provided on this platform
+//
+
 // optimized static helpers generated dynamically at runtime
 // #define JIT_GetSharedGCStaticBase
 // #define JIT_GetSharedNonGCStaticBase
@@ -573,4 +571,5 @@ inline BOOL ClrFlushInstructionCache(LPCVOID pCodeAddr, size_t sizeOfCode)
 #define JIT_NewCrossContext         JIT_NewCrossContext
 #define JIT_Stelem_Ref              JIT_Stelem_Ref
 #endif // FEATURE_PAL
+
 #endif // __cgenx86_h__
diff --git a/src/vm/i386/cgenx86.cpp b/src/vm/i386/cgenx86.cpp
index 3b523c9c28..c91a716137 100644
--- a/src/vm/i386/cgenx86.cpp
+++ b/src/vm/i386/cgenx86.cpp
@@ -1076,7 +1076,7 @@ Stub *GenerateInitPInvokeFrameHelper()
     unsigned negSpace = FrameInfo.offsetOfFrameVptr;
 
     // mov esi, GetThread()
-    psl->X86EmitCurrentThreadFetch(kESI, (1<<kEDI)|(1<<kEBX)|(1<<kECX)|(1<<kEDX));
+    psl->X86EmitCurrentThreadFetch(kESI);
 
     // mov [edi + FrameInfo.offsetOfGSCookie], GetProcessGSCookie()
     psl->X86EmitOffsetModRM(0xc7, (X86Reg)0x0, kEDI, FrameInfo.offsetOfGSCookie - negSpace);
diff --git a/src/vm/i386/jithelp.asm b/src/vm/i386/jithelp.asm
index 85e824040a..5d64b18c94 100644
--- a/src/vm/i386/jithelp.asm
+++ b/src/vm/i386/jithelp.asm
@@ -2,11 +2,6 @@
 ; The .NET Foundation licenses this file to you under the MIT license.
 ; See the LICENSE file in the project root for more information.
 
-; ==++==
-; 
-
-; 
-; ==--==
 ; ***********************************************************************
 ; File: JIThelp.asm
 ;
@@ -70,13 +65,11 @@ endif
 EXTERN _g_TailCallFrameVptr:DWORD
 EXTERN @JIT_FailFast@0:PROC
 EXTERN _s_gsCookie:DWORD
+EXTERN _GetThread@0:PROC
 EXTERN @JITutil_IsInstanceOfInterface@8:PROC
 EXTERN @JITutil_ChkCastInterface@8:PROC
 EXTERN @JITutil_IsInstanceOfAny@8:PROC
 EXTERN @JITutil_ChkCastAny@8:PROC
-ifdef FEATURE_IMPLICIT_TLS
-EXTERN _GetThread@0:PROC
-endif
 
 ifdef WRITE_BARRIER_CHECK 
 ; Those global variables are always defined, but should be 0 for Server GC
@@ -963,13 +956,15 @@ NewArgs         equ 20
 ; extra space is incremented as we push things on the stack along the way
 ExtraSpace      = 0
 
-        call    _GetThread@0; eax = Thread*
-        push    eax         ; Thread*
+        push    0           ; Thread*
 
         ; save ArgumentRegisters
         push    ecx
         push    edx
 
+        call    _GetThread@0; eax = Thread*
+        mov     [esp + 8], eax
+
 ExtraSpace      = 12    ; pThread, ecx, edx
 
 ifdef FEATURE_HIJACK
@@ -1247,44 +1242,8 @@ _JIT_PatchedCodeStart@0 proc public
 ret
 _JIT_PatchedCodeStart@0 endp
 
-;
-; Optimized TLS getters
-;
-
-            ALIGN 4
-            
-ifndef FEATURE_IMPLICIT_TLS
-_GetThread@0 proc public
-            ; This will be overwritten at runtime with optimized GetThread implementation
-            jmp short _GetTLSDummy@0
-            ; Just allocate space that will be filled in at runtime
-            db (TLS_GETTER_MAX_SIZE_ASM - 2) DUP (0CCh)
-_GetThread@0 endp
-
-            ALIGN 4
-
-_GetAppDomain@0 proc public
-            ; This will be overwritten at runtime with optimized GetAppDomain implementation
-            jmp short _GetTLSDummy@0
-            ; Just allocate space that will be filled in at runtime
-            db (TLS_GETTER_MAX_SIZE_ASM - 2) DUP (0CCh)
-_GetAppDomain@0 endp
-
-_GetTLSDummy@0 proc public
-            xor eax,eax
-            ret
-_GetTLSDummy@0 endp
-
             ALIGN 4
 
-_ClrFlsGetBlock@0 proc public
-            ; This will be overwritten at runtime with optimized ClrFlsGetBlock implementation
-            jmp short _GetTLSDummy@0
-            ; Just allocate space that will be filled in at runtime
-            db (TLS_GETTER_MAX_SIZE_ASM - 2) DUP (0CCh)
-_ClrFlsGetBlock@0 endp
-endif
-
 ;**********************************************************************
 ; Write barriers generated at runtime
 
diff --git a/src/vm/i386/jitinterfacex86.cpp b/src/vm/i386/jitinterfacex86.cpp
index 8463b084ac..1b470c9533 100644
--- a/src/vm/i386/jitinterfacex86.cpp
+++ b/src/vm/i386/jitinterfacex86.cpp
@@ -539,7 +539,7 @@ void JIT_TrialAlloc::EmitCore(CPUSTUBLINKER *psl, CodeLabel *noLock, CodeLabel *
                  && "EAX should contain size for allocation and it doesnt!!!");
 
         // Fetch current thread into EDX, preserving EAX and ECX
-        psl->X86EmitCurrentThreadFetch(kEDX, (1<<kEAX)|(1<<kECX));
+        psl->X86EmitCurrentThreadFetch(kEDX);
 
         // Try the allocation.
 
@@ -1253,7 +1253,7 @@ FastPrimitiveArrayAllocatorFuncPtr fastPrimitiveArrayAllocator = UnframedAllocat
 
 // "init" should be the address of a routine which takes an argument of
 // the module domain ID, the class domain ID, and returns the static base pointer
-void EmitFastGetSharedStaticBase(CPUSTUBLINKER *psl, CodeLabel *init, bool bCCtorCheck, bool bGCStatic, bool bSingleAppDomain)
+void EmitFastGetSharedStaticBase(CPUSTUBLINKER *psl, CodeLabel *init, bool bCCtorCheck, bool bGCStatic)
 {
     STANDARD_VM_CONTRACT;
 
@@ -1267,35 +1267,6 @@ void EmitFastGetSharedStaticBase(CPUSTUBLINKER *psl, CodeLabel *init, bool bCCto
     psl->Emit8(0x89);
     psl->Emit8(0xc8);
 
-    if(!bSingleAppDomain)
-    {
-        // Check tag
-        CodeLabel *cctorCheck = psl->NewCodeLabel();
-
-
-        // test eax, 1
-        psl->Emit8(0xa9);
-        psl->Emit32(1);
-
-        // jz cctorCheck
-        psl->X86EmitCondJump(cctorCheck, X86CondCode::kJZ);
-
-        // mov eax GetAppDomain()
-        psl->X86EmitCurrentAppDomainFetch(kEAX, (1<<kECX)|(1<<kEDX));
-
-        // mov eax [eax->m_sDomainLocalBlock.m_pModuleSlots]
-        psl->X86EmitIndexRegLoad(kEAX, kEAX, (__int32) AppDomain::GetOffsetOfModuleSlotsPointer());
-
-        // Note: weird address arithmetic effectively does:
-        // shift over 1 to remove tag bit (which is always 1), then multiply by 4.
-        // mov eax [eax + ecx*2 - 2]
-        psl->X86EmitOp(0x8b, kEAX, kEAX, -2, kECX, 2);
-
-        // cctorCheck:
-        psl->EmitLabel(cctorCheck);
-
-    }
-
     if (bCCtorCheck)
     {
         // test [eax + edx + offsetof(DomainLocalModule, m_pDataBlob], ClassInitFlags::INITIALIZED_FLAG       // Is class inited
@@ -1356,7 +1327,7 @@ void EmitFastGetSharedStaticBase(CPUSTUBLINKER *psl, CodeLabel *init, bool bCCto
 
 }
 
-void *GenFastGetSharedStaticBase(bool bCheckCCtor, bool bGCStatic, bool bSingleAppDomain)
+void *GenFastGetSharedStaticBase(bool bCheckCCtor, bool bGCStatic)
 {
     STANDARD_VM_CONTRACT;
 
@@ -1372,7 +1343,7 @@ void *GenFastGetSharedStaticBase(bool bCheckCCtor, bool bGCStatic, bool bSingleA
         init = sl.NewExternalCodeLabel((LPVOID)JIT_GetSharedNonGCStaticBase);
     }
 
-    EmitFastGetSharedStaticBase(&sl, init, bCheckCCtor, bGCStatic, bSingleAppDomain);
+    EmitFastGetSharedStaticBase(&sl, init, bCheckCCtor, bGCStatic);
 
     Stub *pStub = sl.Link(SystemDomain::GetGlobalLoaderAllocator()->GetExecutableHeap());
 
@@ -1521,16 +1492,14 @@ void InitJITHelpers1()
         //UnframedAllocateString;
     }
 
-    bool bSingleAppDomain = IsSingleAppDomain();
-
     // Replace static helpers with faster assembly versions
-    pMethodAddresses[6] = GenFastGetSharedStaticBase(true, true, bSingleAppDomain);
+    pMethodAddresses[6] = GenFastGetSharedStaticBase(true, true);
     SetJitHelperFunction(CORINFO_HELP_GETSHARED_GCSTATIC_BASE, pMethodAddresses[6]);
-    pMethodAddresses[7] = GenFastGetSharedStaticBase(true, false, bSingleAppDomain);
+    pMethodAddresses[7] = GenFastGetSharedStaticBase(true, false);
     SetJitHelperFunction(CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE, pMethodAddresses[7]);
-    pMethodAddresses[8] = GenFastGetSharedStaticBase(false, true, bSingleAppDomain);
+    pMethodAddresses[8] = GenFastGetSharedStaticBase(false, true);
     SetJitHelperFunction(CORINFO_HELP_GETSHARED_GCSTATIC_BASE_NOCTOR, pMethodAddresses[8]);
-    pMethodAddresses[9] = GenFastGetSharedStaticBase(false, false, bSingleAppDomain);
+    pMethodAddresses[9] = GenFastGetSharedStaticBase(false, false);
     SetJitHelperFunction(CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE_NOCTOR, pMethodAddresses[9]);
 
     ETW::MethodLog::StubsInitialized(pMethodAddresses, (PVOID *)pHelperNames, ETW_NUM_JIT_HELPERS);
diff --git a/src/vm/i386/stublinkerx86.cpp b/src/vm/i386/stublinkerx86.cpp
index b77609822b..3b376be26d 100644
--- a/src/vm/i386/stublinkerx86.cpp
+++ b/src/vm/i386/stublinkerx86.cpp
@@ -2320,7 +2320,7 @@ static const X86Reg c_argRegs[] = {
 
 #ifndef CROSSGEN_COMPILE
 
-#if defined(_DEBUG) && (defined(_TARGET_AMD64_) || defined(_TARGET_X86_)) && !defined(FEATURE_PAL)
+#if defined(_DEBUG) && !defined(FEATURE_PAL)
 void StubLinkerCPU::EmitJITHelperLoggingThunk(PCODE pJitHelper, LPVOID helperFuncCount)
 {
     STANDARD_VM_CONTRACT;
@@ -2358,181 +2358,36 @@ void StubLinkerCPU::EmitJITHelperLoggingThunk(PCODE pJitHelper, LPVOID helperFun
 #endif
     X86EmitTailcallWithSinglePop(NewExternalCodeLabel(pJitHelper), kECX);
 }
-#endif // _DEBUG && (_TARGET_AMD64_ || _TARGET_X86_) && !FEATURE_PAL
+#endif // _DEBUG && !FEATURE_PAL
 
-#ifndef FEATURE_IMPLICIT_TLS
-//---------------------------------------------------------------
-// Emit code to store the current Thread structure in dstreg
-// preservedRegSet is a set of registers to be preserved
-// TRASHES  EAX, EDX, ECX unless they are in preservedRegSet.
-// RESULTS  dstreg = current Thread
-//---------------------------------------------------------------
-VOID StubLinkerCPU::X86EmitTLSFetch(DWORD idx, X86Reg dstreg, unsigned preservedRegSet)
+#if !defined(FEATURE_STUBS_AS_IL)
+VOID StubLinkerCPU::X86EmitCurrentThreadFetch(X86Reg dstreg)
 {
     CONTRACTL
     {
         STANDARD_VM_CHECK;
 
-    // It doesn't make sense to have the destination register be preserved
-        PRECONDITION((preservedRegSet & (1<<dstreg)) == 0);
         AMD64_ONLY(PRECONDITION(dstreg < 8)); // code below doesn't support high registers
     }
     CONTRACTL_END;
 
-    TLSACCESSMODE mode = GetTLSAccessMode(idx);
-
-#ifdef _DEBUG
-    {
-        static BOOL f = TRUE;
-        f = !f;
-        if (f)
-        {
-           mode = TLSACCESS_GENERIC;
-        }
-    }
-#endif
-
-    switch (mode)
-    {
-        case TLSACCESS_WNT: 
-            {
-                unsigned __int32 tlsofs = offsetof(TEB, TlsSlots) + (idx * sizeof(void*));
-#ifdef _TARGET_AMD64_                
-                BYTE code[] = {0x65,0x48,0x8b,0x04,0x25};    // mov dstreg, qword ptr gs:[IMM32]
-                static const int regByteIndex = 3;
-#elif defined(_TARGET_X86_)
-                BYTE code[] = {0x64,0x8b,0x05};              // mov dstreg, dword ptr fs:[IMM32]
-                static const int regByteIndex = 2;
-#endif 
-                code[regByteIndex] |= (dstreg << 3);
-
-                EmitBytes(code, sizeof(code));
-                Emit32(tlsofs);
-            }
-            break;
-
-        case TLSACCESS_GENERIC:
-
-            X86EmitPushRegs(preservedRegSet & ((1<<kEAX)|(1<<kEDX)|(1<<kECX)));
-
-            X86EmitPushImm32(idx);
 #ifdef _TARGET_AMD64_
-            X86EmitPopReg (kECX);       // arg in reg
-#endif
-
-            // call TLSGetValue
-            X86EmitCall(NewExternalCodeLabel((LPVOID) TlsGetValue), sizeof(void*));
-
-            // mov dstreg, eax
-            X86EmitMovRegReg(dstreg, kEAX);
-
-            X86EmitPopRegs(preservedRegSet & ((1<<kEAX)|(1<<kEDX)|(1<<kECX)));
-
-            break;
-
-        default:
-            _ASSERTE(0);
-    }
-
-#ifdef _DEBUG
-    // Trash caller saved regs that we were not told to preserve, and that aren't the dstreg.
-    preservedRegSet |= 1<<dstreg;
-    if (!(preservedRegSet & (1<<kEAX)))
-        X86EmitDebugTrashReg(kEAX);
-    if (!(preservedRegSet & (1<<kEDX)))
-        X86EmitDebugTrashReg(kEDX);
-    if (!(preservedRegSet & (1<<kECX)))
-        X86EmitDebugTrashReg(kECX);
-#endif
-
-}
-#endif // FEATURE_IMPLICIT_TLS
-
-VOID StubLinkerCPU::X86EmitCurrentThreadFetch(X86Reg dstreg, unsigned preservedRegSet)
-{
-    CONTRACTL
-    {
-        STANDARD_VM_CHECK;
-
-    // It doesn't make sense to have the destination register be preserved
-        PRECONDITION((preservedRegSet & (1<<dstreg)) == 0);
-        AMD64_ONLY(PRECONDITION(dstreg < 8)); // code below doesn't support high registers
-    }
-    CONTRACTL_END;
-
-#ifdef FEATURE_IMPLICIT_TLS
-
-    X86EmitPushRegs(preservedRegSet & ((1<<kEAX)|(1<<kEDX)|(1<<kECX)));
-
-    //TODO: Inline the instruction instead of a call
-    // call GetThread
-    X86EmitCall(NewExternalCodeLabel((LPVOID) GetThread), sizeof(void*));
-
-    // mov dstreg, eax
-    X86EmitMovRegReg(dstreg, kEAX);
-
-    X86EmitPopRegs(preservedRegSet & ((1<<kEAX)|(1<<kEDX)|(1<<kECX)));
-
-#ifdef _DEBUG
-    // Trash caller saved regs that we were not told to preserve, and that aren't the dstreg.
-    preservedRegSet |= 1<<dstreg;
-    if (!(preservedRegSet & (1<<kEAX)))
-        X86EmitDebugTrashReg(kEAX);
-    if (!(preservedRegSet & (1<<kEDX)))
-        X86EmitDebugTrashReg(kEDX);
-    if (!(preservedRegSet & (1<<kECX)))
-        X86EmitDebugTrashReg(kECX);
-#endif // _DEBUG
-
-#else // FEATURE_IMPLICIT_TLS
-
-    X86EmitTLSFetch(GetThreadTLSIndex(), dstreg, preservedRegSet);
-    
-#endif // FEATURE_IMPLICIT_TLS
-
-}
-
-VOID StubLinkerCPU::X86EmitCurrentAppDomainFetch(X86Reg dstreg, unsigned preservedRegSet)
-{
-    CONTRACTL
-    {
-        STANDARD_VM_CHECK;
-
-    // It doesn't make sense to have the destination register be preserved
-        PRECONDITION((preservedRegSet & (1<<dstreg)) == 0);
-        AMD64_ONLY(PRECONDITION(dstreg < 8)); // code below doesn't support high registers
-    }
-    CONTRACTL_END;
-
-#ifdef FEATURE_IMPLICIT_TLS
-    X86EmitPushRegs(preservedRegSet & ((1<<kEAX)|(1<<kEDX)|(1<<kECX)));
-
-    //TODO: Inline the instruction instead of a call
-    // call GetThread
-    X86EmitCall(NewExternalCodeLabel((LPVOID) GetAppDomain), sizeof(void*));
-
-    // mov dstreg, eax
-    X86EmitMovRegReg(dstreg, kEAX);
-
-    X86EmitPopRegs(preservedRegSet & ((1<<kEAX)|(1<<kEDX)|(1<<kECX)));
-
-#ifdef _DEBUG
-    // Trash caller saved regs that we were not told to preserve, and that aren't the dstreg.
-    preservedRegSet |= 1<<dstreg;
-    if (!(preservedRegSet & (1<<kEAX)))
-        X86EmitDebugTrashReg(kEAX);
-    if (!(preservedRegSet & (1<<kEDX)))
-        X86EmitDebugTrashReg(kEDX);
-    if (!(preservedRegSet & (1<<kECX)))
-        X86EmitDebugTrashReg(kECX);
+    BYTE code[] = { 0x65,0x48,0x8b,0x04,0x25 };    // mov dstreg, qword ptr gs:[IMM32]
+    static const int regByteIndex = 3;
+#elif defined(_TARGET_X86_)
+    BYTE code[] = { 0x64,0x8b,0x05 };              // mov dstreg, dword ptr fs:[IMM32]
+    static const int regByteIndex = 2;
 #endif
+    code[regByteIndex] |= (dstreg << 3);
 
-#else // FEATURE_IMPLICIT_TLS
+    EmitBytes(code, sizeof(code));
+    Emit32(offsetof(TEB, ThreadLocalStoragePointer));
 
-    X86EmitTLSFetch(GetAppDomainTLSIndex(), dstreg, preservedRegSet);
+    X86EmitIndexRegLoad(dstreg, dstreg, sizeof(void *) * (g_TlsIndex & 0xFFFF));
 
-#endif // FEATURE_IMPLICIT_TLS
+    X86EmitIndexRegLoad(dstreg, dstreg, (g_TlsIndex & 0x7FFF0000) >> 16);
 }
+#endif // !FEATURE_STUBS_AS_IL
 
 #if defined(_TARGET_X86_)
 
@@ -2861,56 +2716,7 @@ VOID StubLinkerCPU::EmitSetup(CodeLabel *pForwardRef)
 {
     STANDARD_VM_CONTRACT;
 
-#ifdef FEATURE_IMPLICIT_TLS
-    DWORD idx = 0;
-    TLSACCESSMODE mode = TLSACCESS_GENERIC;
-#else
-    DWORD idx = GetThreadTLSIndex();
-    TLSACCESSMODE mode = GetTLSAccessMode(idx);
-#endif
-
-#ifdef _DEBUG
-    {
-        static BOOL f = TRUE;
-        f = !f;
-        if (f)
-        {
-           mode = TLSACCESS_GENERIC;
-        }
-    }
-#endif
-
-    switch (mode)
-    {
-        case TLSACCESS_WNT: 
-#ifndef FEATURE_PAL
-            {
-                unsigned __int32 tlsofs = offsetof(TEB, TlsSlots) + (idx * sizeof(void*));
-
-                static const BYTE code[] = {0x64,0x8b,0x1d};              // mov ebx, dword ptr fs:[IMM32]
-                EmitBytes(code, sizeof(code));
-                Emit32(tlsofs);
-            }
-#else  // !FEATURE_PAL
-            _ASSERTE("TLSACCESS_WNT mode is not supported");
-#endif // !FEATURE_PAL
-            break;
-
-        case TLSACCESS_GENERIC:
-#ifdef FEATURE_IMPLICIT_TLS
-            X86EmitCall(NewExternalCodeLabel((LPVOID) GetThread), sizeof(void*));
-#else
-            X86EmitPushImm32(idx);
-
-            // call TLSGetValue
-            X86EmitCall(NewExternalCodeLabel((LPVOID) TlsGetValue), sizeof(void*));
-#endif
-            // mov ebx,eax
-            Emit16(0xc389);
-            break;
-        default:
-            _ASSERTE(0);
-    }
+    X86EmitCurrentThreadFetch(kEBX);
 
     // cmp ebx, 0
     static const BYTE b[] = { 0x83, 0xFB, 0x0};
@@ -3150,8 +2956,7 @@ VOID StubLinkerCPU::EmitMethodStubProlog(TADDR pFrameVptr, int transitionBlockOf
 #endif // _TARGET_X86_
 
     // ebx <-- GetThread()
-    // Trashes X86TLSFetch_TRASHABLE_REGS
-    X86EmitCurrentThreadFetch(kEBX, 0);
+    X86EmitCurrentThreadFetch(kEBX);
 
 #if _DEBUG
 
diff --git a/src/vm/i386/stublinkerx86.h b/src/vm/i386/stublinkerx86.h
index 76d1f95845..d523bb0461 100644
--- a/src/vm/i386/stublinkerx86.h
+++ b/src/vm/i386/stublinkerx86.h
@@ -219,11 +219,7 @@ class StubLinkerCPU : public StubLinker
         VOID X86EmitLeaRIP(CodeLabel *target, X86Reg reg);
 #endif
 
-        static const unsigned X86TLSFetch_TRASHABLE_REGS = (1<<kEAX) | (1<<kEDX) | (1<<kECX);
-        VOID X86EmitTLSFetch(DWORD idx, X86Reg dstreg, unsigned preservedRegSet);
-
-        VOID X86EmitCurrentThreadFetch(X86Reg dstreg, unsigned preservedRegSet);
-        VOID X86EmitCurrentAppDomainFetch(X86Reg dstreg, unsigned preservedRegSet);
+        VOID X86EmitCurrentThreadFetch(X86Reg dstreg);
         
         VOID X86EmitIndexRegLoad(X86Reg dstreg, X86Reg srcreg, __int32 ofs = 0);
         VOID X86EmitIndexRegStore(X86Reg dstreg, __int32 ofs, X86Reg srcreg);
@@ -421,7 +417,7 @@ class StubLinkerCPU : public StubLinker
         VOID EmitDebugBreak();
 #endif // !FEATURE_STUBS_AS_IL
 
-#if defined(_DEBUG) && (defined(_TARGET_AMD64_) || defined(_TARGET_X86_)) && !defined(FEATURE_PAL)
+#if defined(_DEBUG) && !defined(FEATURE_PAL)
         //===========================================================================
         // Emits code to log JITHelper access
         void EmitJITHelperLoggingThunk(PCODE pJitHelper, LPVOID helperFuncCount);