Merge branch 'dotnet:main' into AOT_delegates

dotnet · May 15, 2023 · 175dae9 · 175dae9
2 parents 3144e6e + 9dd59af
commit 175dae9
Show file tree

Hide file tree

Showing 51 changed files with 644 additions and 282 deletions.
diff --git a/docs/workflow/debugging/coreclr/debugging-runtime.md b/docs/workflow/debugging/coreclr/debugging-runtime.md
@@ -41,6 +41,7 @@ If for some reason `System.Private.CoreLib.dll` is missing, you can rebuild it w
 
 Visual Studio's capabilities as a full IDE provide a lot of help making the runtime debugging more amiable.
 
+0. Run `.\build.cmd clr.nativeprereqs -a <architecture> -c <configuration>`. This will build some of the tools requiremented for the native build. This step only needs to be run once as long you don't clean the `artifacts` directory.
 1. Open the CoreCLR solution _(coreclr.sln)_ in Visual Studio.
    * _Method 1_: Use the build scripts to open the solution:
       1. Run `.\build.cmd -vs coreclr.sln -a <architecture> -c <configuration>`. This will create and launch the CoreCLR solution in VS for the specified architecture and configuration. By default, this will be `x64 Debug`.

diff --git a/eng/native/configureplatform.cmake b/eng/native/configureplatform.cmake
@@ -461,10 +461,12 @@ if(CLR_CMAKE_TARGET_UNIX)
     else()
         clr_unknown_arch()
     endif()
-else()
-    set(CLR_CMAKE_TARGET_WIN32 1)
 endif(CLR_CMAKE_TARGET_UNIX)
 
+if(CLR_CMAKE_TARGET_OS STREQUAL windows)
+    set(CLR_CMAKE_TARGET_WIN32 1)
+endif()
+
 # check if host & target os/arch combination are valid
 if (NOT (CLR_CMAKE_TARGET_OS STREQUAL CLR_CMAKE_HOST_OS) AND NOT CLR_CMAKE_TARGET_WASI)
     if(NOT (CLR_CMAKE_HOST_OS STREQUAL windows))

diff --git a/src/coreclr/inc/corinfo.h b/src/coreclr/inc/corinfo.h
@@ -548,6 +548,7 @@ enum CorInfoHelpFunc
     CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR,
     CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_DYNAMICCLASS,
     CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_DYNAMICCLASS,
+    CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED,
     CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED,
 
     /* Debugger */
@@ -1731,6 +1732,7 @@ struct CORINFO_THREAD_STATIC_BLOCKS_INFO
     uint32_t offsetOfThreadLocalStoragePointer;
     uint32_t offsetOfMaxThreadStaticBlocks;
     uint32_t offsetOfThreadStaticBlocks;
+    uint32_t offsetOfGCDataPointer;
 };
 
 //----------------------------------------------------------------------------
@@ -2733,11 +2735,13 @@ class ICorStaticInfo
                                CORINFO_FIELD_INFO    *pResult
                               ) = 0;
 
+    // Returns the index against which the field's thread static block in stored in TLS.
     virtual uint32_t getThreadLocalFieldInfo (
-                        CORINFO_FIELD_HANDLE  field) = 0;
+                        CORINFO_FIELD_HANDLE  field, bool isGCType) = 0;
 
+    // Returns the thread static block information like offsets, etc. from current TLS.
     virtual void getThreadLocalStaticBlocksInfo (
-                        CORINFO_THREAD_STATIC_BLOCKS_INFO* pInfo) = 0;
+                        CORINFO_THREAD_STATIC_BLOCKS_INFO* pInfo, bool isGCType) = 0;
 
     // Returns true iff "fldHnd" represents a static field.
     virtual bool isFieldStatic(CORINFO_FIELD_HANDLE fldHnd) = 0;

diff --git a/src/coreclr/inc/icorjitinfoimpl_generated.h b/src/coreclr/inc/icorjitinfoimpl_generated.h
@@ -411,10 +411,12 @@ void getFieldInfo(
           CORINFO_FIELD_INFO* pResult) override;
 
 uint32_t getThreadLocalFieldInfo(
-          CORINFO_FIELD_HANDLE field) override;
+          CORINFO_FIELD_HANDLE field,
+          bool isGCtype) override;
 
 void getThreadLocalStaticBlocksInfo(
-          CORINFO_THREAD_STATIC_BLOCKS_INFO* pInfo) override;
+          CORINFO_THREAD_STATIC_BLOCKS_INFO* pInfo,
+          bool isGCType) override;
 
 bool isFieldStatic(
           CORINFO_FIELD_HANDLE fldHnd) override;

diff --git a/src/coreclr/inc/jiteeversionguid.h b/src/coreclr/inc/jiteeversionguid.h
@@ -43,11 +43,11 @@ typedef const GUID *LPCGUID;
 #define GUID_DEFINED
 #endif // !GUID_DEFINED
 
-constexpr GUID JITEEVersionIdentifier = { /* 50246073-be42-4227-b9e1-edc585672820 */
-    0x50246073,
-    0xbe42,
-    0x4227,
-    {0xb9, 0xe1, 0xed, 0xc5, 0x85, 0x67, 0x28, 0x20}
+constexpr GUID JITEEVersionIdentifier = { /* f63c2964-bae9-448f-baaf-9c9f2d4292f2  */
+    0xf63c2964,
+    0xbae9,
+    0x448f,
+    {0xba, 0xaf, 0x9c, 0x9f, 0x2d, 0x42, 0x92, 0xf2}
   };
 
 //////////////////////////////////////////////////////////////////////////////////////////////////////////

diff --git a/src/coreclr/inc/jithelpers.h b/src/coreclr/inc/jithelpers.h
@@ -203,8 +203,9 @@
     JITHELPER(CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE,              JIT_GetSharedNonGCThreadStaticBase, CORINFO_HELP_SIG_REG_ONLY)
     JITHELPER(CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR,          JIT_GetSharedGCThreadStaticBase, CORINFO_HELP_SIG_REG_ONLY)
     JITHELPER(CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR,       JIT_GetSharedNonGCThreadStaticBase, CORINFO_HELP_SIG_REG_ONLY)
-    JITHELPER(CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_DYNAMICCLASS,    JIT_GetSharedGCThreadStaticBaseDynamicClass, CORINFO_HELP_SIG_REG_ONLY)
+    JITHELPER(CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_DYNAMICCLASS,    JIT_GetSharedGCThreadStaticBaseDynamicClass, CORINFO_HELP_SIG_REG_ONLY)    
     JITHELPER(CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_DYNAMICCLASS, JIT_GetSharedNonGCThreadStaticBaseDynamicClass, CORINFO_HELP_SIG_REG_ONLY)
+    JITHELPER(CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED,    JIT_GetSharedGCThreadStaticBaseOptimized, CORINFO_HELP_SIG_REG_ONLY)
     JITHELPER(CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED,       JIT_GetSharedNonGCThreadStaticBaseOptimized, CORINFO_HELP_SIG_REG_ONLY)
 
     // Debugger

diff --git a/src/coreclr/jit/ICorJitInfo_wrapper_generated.hpp b/src/coreclr/jit/ICorJitInfo_wrapper_generated.hpp
@@ -972,19 +972,21 @@ void WrapICorJitInfo::getFieldInfo(
 }
 
 uint32_t WrapICorJitInfo::getThreadLocalFieldInfo(
-          CORINFO_FIELD_HANDLE field)
+          CORINFO_FIELD_HANDLE field,
+          bool isGCtype)
 {
     API_ENTER(getThreadLocalFieldInfo);
-    uint32_t temp = wrapHnd->getThreadLocalFieldInfo(field);
+    uint32_t temp = wrapHnd->getThreadLocalFieldInfo(field, isGCtype);
     API_LEAVE(getThreadLocalFieldInfo);
     return temp;
 }
 
 void WrapICorJitInfo::getThreadLocalStaticBlocksInfo(
-          CORINFO_THREAD_STATIC_BLOCKS_INFO* pInfo)
+          CORINFO_THREAD_STATIC_BLOCKS_INFO* pInfo,
+          bool isGCType)
 {
     API_ENTER(getThreadLocalStaticBlocksInfo);
-    wrapHnd->getThreadLocalStaticBlocksInfo(pInfo);
+    wrapHnd->getThreadLocalStaticBlocksInfo(pInfo, isGCType);
     API_LEAVE(getThreadLocalStaticBlocksInfo);
 }
 

diff --git a/src/coreclr/jit/compiler.h b/src/coreclr/jit/compiler.h
@@ -7818,7 +7818,6 @@ class Compiler
     void eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
                         CORINFO_ACCESS_FLAGS    flags,
                         CORINFO_FIELD_INFO*     pResult);
-    uint32_t eeGetThreadLocalFieldInfo(CORINFO_FIELD_HANDLE field);
 
     // Get the flags
 
@@ -9394,11 +9393,19 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
             return IsInstrumented() && jitFlags->IsSet(JitFlags::JIT_FLAG_BBOPT);
         }
 
-        bool CanBeInstrumentedOrIsOptimized() const
+        bool DoEarlyBlockMerging() const
         {
-            return IsInstrumented() || (jitFlags->IsSet(JitFlags::JIT_FLAG_TIER0) &&
-                                        jitFlags->IsSet(JitFlags::JIT_FLAG_BBINSTR_IF_LOOPS)) ||
-                   jitFlags->IsSet(JitFlags::JIT_FLAG_BBOPT);
+            if (jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_EnC) || jitFlags->IsSet(JitFlags::JIT_FLAG_DEBUG_CODE))
+            {
+                return false;
+            }
+
+            if (jitFlags->IsSet(JitFlags::JIT_FLAG_MIN_OPT) && !jitFlags->IsSet(JitFlags::JIT_FLAG_TIER0))
+            {
+                return false;
+            }
+
+            return true;
         }
 
         // true if we should use the PINVOKE_{BEGIN,END} helpers instead of generating

diff --git a/src/coreclr/jit/compiler.hpp b/src/coreclr/jit/compiler.hpp
@@ -3518,6 +3518,7 @@ inline bool Compiler::IsSharedStaticHelper(GenTree* tree)
         helper == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE ||
         helper == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE ||
         helper == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR ||
+        helper == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED ||
         helper == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR ||
         helper == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED ||
         helper == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_DYNAMICCLASS ||

diff --git a/src/coreclr/jit/ee_il_dll.hpp b/src/coreclr/jit/ee_il_dll.hpp
@@ -44,12 +44,6 @@ void Compiler::eeGetFieldInfo(CORINFO_RESOLVED_TOKEN* pResolvedToken,
     info.compCompHnd->getFieldInfo(pResolvedToken, info.compMethodHnd, accessFlags, pResult);
 }
 
-FORCEINLINE
-uint32_t Compiler::eeGetThreadLocalFieldInfo(CORINFO_FIELD_HANDLE field)
-{
-    return info.compCompHnd->getThreadLocalFieldInfo(field);
-}
-
 /*****************************************************************************
  *
  *          VOS info, method sigs, etc

diff --git a/src/coreclr/jit/fgbasic.cpp b/src/coreclr/jit/fgbasic.cpp
@@ -1840,7 +1840,7 @@ void Compiler::fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, Fixed
 
                 if ((jmpDist == 0) &&
                     (opcode == CEE_LEAVE || opcode == CEE_LEAVE_S || opcode == CEE_BR || opcode == CEE_BR_S) &&
-                    opts.CanBeInstrumentedOrIsOptimized())
+                    opts.DoEarlyBlockMerging())
                 {
                     break; /* NOP */
                 }
@@ -2989,7 +2989,7 @@ unsigned Compiler::fgMakeBasicBlocks(const BYTE* codeAddr, IL_OFFSET codeSize, F
 
                 jmpDist = (sz == 1) ? getI1LittleEndian(codeAddr) : getI4LittleEndian(codeAddr);
 
-                if ((jmpDist == 0) && (opcode == CEE_BR || opcode == CEE_BR_S) && opts.CanBeInstrumentedOrIsOptimized())
+                if ((jmpDist == 0) && (opcode == CEE_BR || opcode == CEE_BR_S) && opts.DoEarlyBlockMerging())
                 {
                     continue; /* NOP */
                 }

diff --git a/src/coreclr/jit/flowgraph.cpp b/src/coreclr/jit/flowgraph.cpp
@@ -698,7 +698,8 @@ bool Compiler::fgIsCommaThrow(GenTree* tree, bool forFolding /* = false */)
 //    cls       - The class handle
 //    helper    - The helper function
 //    typeIndex - The static block type index. Used only for
-//                CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED to cache
+//                CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED or
+//                CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED to cache
 //                the static block in an array at index typeIndex.
 //
 // Return Value:
@@ -715,6 +716,7 @@ GenTreeCall* Compiler::fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfo
     switch (helper)
     {
         case CORINFO_HELP_GETSHARED_GCSTATIC_BASE_NOCTOR:
+        case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED:
             bNeedClassID = false;
             FALLTHROUGH;
 
@@ -795,7 +797,8 @@ GenTreeCall* Compiler::fgGetStaticsCCtorHelper(CORINFO_CLASS_HANDLE cls, CorInfo
 
         result = gtNewHelperCallNode(helper, type, opModuleIDArg, opClassIDArg);
     }
-    else if (helper == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED)
+    else if ((helper == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED) ||
+             (helper == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED))
     {
         result = gtNewHelperCallNode(helper, type, gtNewIconNode(typeIndex));
         result->SetExpTLSFieldAccess();

diff --git a/src/coreclr/jit/helperexpansion.cpp b/src/coreclr/jit/helperexpansion.cpp
@@ -410,7 +410,8 @@ bool Compiler::fgExpandRuntimeLookupsForCall(BasicBlock** pBlock, Statement* stm
 
 //------------------------------------------------------------------------------
 // fgExpandThreadLocalAccess: Inline the CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED
-//      helper. See fgExpandThreadLocalAccessForCall for details.
+//      or CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED helper. See
+//      fgExpandThreadLocalAccessForCall for details.
 //
 // Returns:
 //    PhaseStatus indicating what, if anything, was changed.
@@ -446,7 +447,7 @@ PhaseStatus Compiler::fgExpandThreadLocalAccess()
 
 //------------------------------------------------------------------------------
 // fgExpandThreadLocalAccessForCall : Expand the CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED
-//                             that access fields marked with [ThreadLocal].
+//  or CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED, that access fields marked with [ThreadLocal].
 //
 // Arguments:
 //    pBlock - Block containing the helper call to expand. If expansion is performed,
@@ -483,20 +484,31 @@ bool Compiler::fgExpandThreadLocalAccessForCall(BasicBlock** pBlock, Statement*
     assert(!"Unsupported scenario of optimizing TLS access on Arm32");
 #endif
 
+    JITDUMP("Expanding thread static local access for [%06d] in " FMT_BB ":\n", dspTreeID(call), block->bbNum);
+    DISPTREE(call);
+    JITDUMP("\n");
+    bool isGCThreadStatic =
+        eeGetHelperNum(call->gtCallMethHnd) == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED;
+
     CORINFO_THREAD_STATIC_BLOCKS_INFO threadStaticBlocksInfo;
-    info.compCompHnd->getThreadLocalStaticBlocksInfo(&threadStaticBlocksInfo);
-    JITDUMP("getThreadLocalStaticBlocksInfo\n:");
+    info.compCompHnd->getThreadLocalStaticBlocksInfo(&threadStaticBlocksInfo, isGCThreadStatic);
+
+    uint32_t offsetOfMaxThreadStaticBlocksVal = 0;
+    uint32_t offsetOfThreadStaticBlocksVal    = 0;
+
+    JITDUMP("getThreadLocalStaticBlocksInfo (%s)\n:", isGCThreadStatic ? "GC" : "Non-GC");
+    offsetOfMaxThreadStaticBlocksVal = threadStaticBlocksInfo.offsetOfMaxThreadStaticBlocks;
+    offsetOfThreadStaticBlocksVal    = threadStaticBlocksInfo.offsetOfThreadStaticBlocks;
+
     JITDUMP("tlsIndex= %u\n", (ssize_t)threadStaticBlocksInfo.tlsIndex.addr);
-    JITDUMP("offsetOfMaxThreadStaticBlocks= %u\n", threadStaticBlocksInfo.offsetOfMaxThreadStaticBlocks);
     JITDUMP("offsetOfThreadLocalStoragePointer= %u\n", threadStaticBlocksInfo.offsetOfThreadLocalStoragePointer);
-    JITDUMP("offsetOfThreadStaticBlocks= %u\n", threadStaticBlocksInfo.offsetOfThreadStaticBlocks);
+    JITDUMP("offsetOfMaxThreadStaticBlocks= %u\n", offsetOfMaxThreadStaticBlocksVal);
+    JITDUMP("offsetOfThreadStaticBlocks= %u\n", offsetOfThreadStaticBlocksVal);
+    JITDUMP("offsetOfGCDataPointer= %u\n", threadStaticBlocksInfo.offsetOfGCDataPointer);
 
     assert(threadStaticBlocksInfo.tlsIndex.accessType == IAT_VALUE);
-    assert(eeGetHelperNum(call->gtCallMethHnd) == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED);
-
-    JITDUMP("Expanding thread static local access for [%06d] in " FMT_BB ":\n", dspTreeID(call), block->bbNum);
-    DISPTREE(call);
-    JITDUMP("\n");
+    assert((eeGetHelperNum(call->gtCallMethHnd) == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED) ||
+           (eeGetHelperNum(call->gtCallMethHnd) == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED));
 
     call->ClearExpTLSFieldAccess();
     assert(call->gtArgs.CountArgs() == 1);
@@ -508,6 +520,7 @@ bool Compiler::fgExpandThreadLocalAccessForCall(BasicBlock** pBlock, Statement*
     DebugInfo   debugInfo    = stmt->GetDebugInfo();
     block                    = fgSplitBlockBeforeTree(block, stmt, call, &newFirstStmt, &callUse);
     *pBlock                  = block;
+    var_types callType       = call->TypeGet();
     assert(prevBb != nullptr && block != nullptr);
 
     // Block ops inserted by the split need to be morphed here since we are after morph.
@@ -523,8 +536,8 @@ bool Compiler::fgExpandThreadLocalAccessForCall(BasicBlock** pBlock, Statement*
 
     // Grab a temp to store result (it's assigned from either fastPathBb or fallbackBb)
     unsigned threadStaticBlockLclNum         = lvaGrabTemp(true DEBUGARG("TLS field access"));
-    lvaTable[threadStaticBlockLclNum].lvType = TYP_I_IMPL;
-    threadStaticBlockLcl                     = gtNewLclvNode(threadStaticBlockLclNum, call->TypeGet());
+    lvaTable[threadStaticBlockLclNum].lvType = callType;
+    threadStaticBlockLcl                     = gtNewLclvNode(threadStaticBlockLclNum, callType);
 
     *callUse = gtClone(threadStaticBlockLcl);
 
@@ -564,8 +577,7 @@ bool Compiler::fgExpandThreadLocalAccessForCall(BasicBlock** pBlock, Statement*
     GenTree* tlsLclValueUse    = gtNewLclVarNode(tlsLclNum);
 
     // Create tree for "maxThreadStaticBlocks = tls[offsetOfMaxThreadStaticBlocks]"
-    GenTree* offsetOfMaxThreadStaticBlocks =
-        gtNewIconNode(threadStaticBlocksInfo.offsetOfMaxThreadStaticBlocks, TYP_I_IMPL);
+    GenTree* offsetOfMaxThreadStaticBlocks = gtNewIconNode(offsetOfMaxThreadStaticBlocksVal, TYP_I_IMPL);
     GenTree* maxThreadStaticBlocksRef =
         gtNewOperNode(GT_ADD, TYP_I_IMPL, gtCloneExpr(tlsLclValueUse), offsetOfMaxThreadStaticBlocks);
     GenTree* maxThreadStaticBlocksValue =
@@ -577,7 +589,7 @@ bool Compiler::fgExpandThreadLocalAccessForCall(BasicBlock** pBlock, Statement*
     maxThreadStaticBlocksCond = gtNewOperNode(GT_JTRUE, TYP_VOID, maxThreadStaticBlocksCond);
 
     // Create tree for "threadStaticBlockBase = tls[offsetOfThreadStaticBlocks]"
-    GenTree* offsetOfThreadStaticBlocks = gtNewIconNode(threadStaticBlocksInfo.offsetOfThreadStaticBlocks, TYP_I_IMPL);
+    GenTree* offsetOfThreadStaticBlocks = gtNewIconNode(offsetOfThreadStaticBlocksVal, TYP_I_IMPL);
     GenTree* threadStaticBlocksRef =
         gtNewOperNode(GT_ADD, TYP_I_IMPL, gtCloneExpr(tlsLclValueUse), offsetOfThreadStaticBlocks);
     GenTree* threadStaticBlocksValue =
@@ -642,6 +654,16 @@ bool Compiler::fgExpandThreadLocalAccessForCall(BasicBlock** pBlock, Statement*
         fgNewBBFromTreeAfter(BBJ_ALWAYS, threadStaticBlockNullCondBB, fallbackValueDef, debugInfo, true);
 
     // fastPathBb
+    if (isGCThreadStatic)
+    {
+        // Need to add extra indirection to access the data pointer.
+
+        threadStaticBlockBaseLclValueUse = gtNewIndir(callType, threadStaticBlockBaseLclValueUse, GTF_IND_NONFAULTING);
+        threadStaticBlockBaseLclValueUse =
+            gtNewOperNode(GT_ADD, callType, threadStaticBlockBaseLclValueUse,
+                          gtNewIconNode(threadStaticBlocksInfo.offsetOfGCDataPointer, TYP_I_IMPL));
+    }
+
     GenTree* fastPathValueDef =
         gtNewStoreLclVarNode(threadStaticBlockLclNum, gtCloneExpr(threadStaticBlockBaseLclValueUse));
     BasicBlock* fastPathBb = fgNewBBFromTreeAfter(BBJ_ALWAYS, fallbackBb, fastPathValueDef, debugInfo, true);

diff --git a/src/coreclr/jit/importer.cpp b/src/coreclr/jit/importer.cpp
@@ -3963,7 +3963,15 @@ GenTree* Compiler::impImportStaticFieldAccess(CORINFO_RESOLVED_TOKEN* pResolvedT
 
         case CORINFO_FIELD_STATIC_TLS_MANAGED:
 
-            typeIndex = info.compCompHnd->getThreadLocalFieldInfo(pResolvedToken->hField);
+            if (pFieldInfo->helper == CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED)
+            {
+                typeIndex = info.compCompHnd->getThreadLocalFieldInfo(pResolvedToken->hField, false);
+            }
+            else
+            {
+                assert(pFieldInfo->helper == CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED);
+                typeIndex = info.compCompHnd->getThreadLocalFieldInfo(pResolvedToken->hField, true);
+            }
 
             FALLTHROUGH;
         case CORINFO_FIELD_STATIC_SHARED_STATIC_HELPER:
@@ -7412,7 +7420,7 @@ void Compiler::impImportBlockCode(BasicBlock* block)
             case CEE_BR_S:
                 jmpDist = (sz == 1) ? getI1LittleEndian(codeAddr) : getI4LittleEndian(codeAddr);
 
-                if ((jmpDist == 0) && opts.CanBeInstrumentedOrIsOptimized())
+                if ((jmpDist == 0) && opts.DoEarlyBlockMerging())
                 {
                     break; /* NOP */
                 }

diff --git a/src/coreclr/jit/utils.cpp b/src/coreclr/jit/utils.cpp
@@ -1476,6 +1476,7 @@ void HelperCallProperties::init()
             case CORINFO_HELP_GETSHARED_GCSTATIC_BASE_NOCTOR:
             case CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE_NOCTOR:
             case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR:
+            case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED:
             case CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR:
             case CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED:
 

diff --git a/src/coreclr/jit/valuenum.cpp b/src/coreclr/jit/valuenum.cpp
@@ -12228,6 +12228,9 @@ VNFunc Compiler::fgValueNumberJitHelperMethodVNFunc(CorInfoHelpFunc helpFunc)
         case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR:
             vnf = VNF_GetsharedGcthreadstaticBaseNoctor;
             break;
+        case CORINFO_HELP_GETSHARED_GCTHREADSTATIC_BASE_NOCTOR_OPTIMIZED:
+            vnf = VNF_GetsharedGcthreadstaticBaseNoctorOptimized;
+            break;
         case CORINFO_HELP_GETSHARED_NONGCTHREADSTATIC_BASE_NOCTOR:
             vnf = VNF_GetsharedNongcthreadstaticBaseNoctor;
             break;