[CIR] Add basic thread-local storage (TLS) support

clang/lib/CIR/CodeGen/CIRGenCXXABI.h

@@ -347,6 +347,10 @@ class CIRGenCXXABI {
   // directly or require access through a thread wrapper function.
   virtual bool usesThreadWrapperFunction(const VarDecl *VD) const = 0;
 
+  virtual LValue EmitThreadLocalVarDeclLValue(CIRGenFunction &CGF,
+                                              const VarDecl *VD,
+                                              QualType LValType) = 0;
+
   /// Emit the code to initialize hidden members required to handle virtual
   /// inheritance, if needed by the ABI.
   virtual void

clang/lib/CIR/CodeGen/CIRGenDecl.cpp

@@ -503,7 +503,7 @@ CIRGenModule::getOrCreateStaticVarDecl(const VarDecl &D,
     GV.setComdat(true);
 
   if (D.getTLSKind())
-    llvm_unreachable("TLS mode is NYI");
+    setTLSMode(GV, D);
 
   setGVProperties(GV, &D);
 

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -937,7 +937,7 @@ static LValue emitGlobalVarDeclLValue(CIRGenFunction &CGF, const Expr *E,
   // If it's thread_local, emit a call to its wrapper function instead.
   if (VD->getTLSKind() == VarDecl::TLS_Dynamic &&
       CGF.CGM.getCXXABI().usesThreadWrapperFunction(VD))
-    assert(0 && "not implemented");
+    return CGF.CGM.getCXXABI().EmitThreadLocalVarDeclLValue(CGF, VD, T);
 
   // Check if the variable is marked as declare target with link clause in
   // device codegen.
@@ -1095,8 +1095,13 @@ LValue CIRGenFunction::emitDeclRefLValue(const DeclRefExpr *E) {
     }
 
     // Handle threadlocal function locals.
-    if (VD->getTLSKind() != VarDecl::TLS_None)
-      llvm_unreachable("thread-local storage is NYI");
+    // ClangIR marks the GlobalOp with thread_local, which gets lowered to
+    // @llvm.threadlocal.address intrinsics during LLVM lowering. This approach
+    // provides correct TLS semantics without requiring wrapper functions in
+    // CIR.
+    if (VD->getTLSKind() != VarDecl::TLS_None) {
+      // Nothing additional needed - the GlobalOp is already thread_local.
+    }
 
     // Check for OpenMP threadprivate variables.
     if (getLangOpts().OpenMP && !getLangOpts().OpenMPSimd &&

clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp

@@ -318,6 +318,9 @@ class CIRGenItaniumCXXABI : public CIRGenCXXABI {
     return !isEmittedWithConstantInitializer(VD) || mayNeedDestruction(VD);
   }
 
+  LValue EmitThreadLocalVarDeclLValue(CIRGenFunction &CGF, const VarDecl *VD,
+                                      QualType LValType) override;
+
   bool doStructorsInitializeVPtrs(const CXXRecordDecl *VTableClass) override {
     return true;
   }
@@ -2925,3 +2928,28 @@ Address CIRGenARMCXXABI::initializeArrayCookie(CIRGenFunction &cgf,
   return Address(dataPtr, cgf.getBuilder().getUIntNTy(8),
                  newPtr.getAlignment());
 }
+
+LValue CIRGenItaniumCXXABI::EmitThreadLocalVarDeclLValue(CIRGenFunction &CGF,
+                                                         const VarDecl *VD,
+                                                         QualType LValType) {
+  // ClangIR's approach to thread-local variables differs from traditional
+  // CodeGen. Traditional CodeGen creates wrapper functions (e.g., _ZTW*) that
+  // handle dynamic initialization. ClangIR instead marks the GlobalOp with
+  // tls_dyn and relies on LLVM lowering to insert @llvm.threadlocal.address
+  // intrinsics, which achieves the same semantics without intermediate wrapper
+  // functions in CIR.
+
+  mlir::Value V = CGF.CGM.getAddrOfGlobalVar(VD);
+
+  auto RealVarTy = CGF.convertTypeForMem(VD->getType());
+  CharUnits Alignment = CGF.getContext().getDeclAlign(VD);
+  Address Addr(V, RealVarTy, Alignment);
+
+  LValue LV;
+  if (VD->getType()->isReferenceType())
+    LV = CGF.emitLoadOfReferenceLValue(Addr, CGF.getLoc(VD->getLocation()),
+                                       VD->getType(), AlignmentSource::Decl);
+  else
+    LV = CGF.makeAddrLValue(Addr, LValType, AlignmentSource::Decl);
+  return LV;
+}

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -1197,7 +1197,7 @@ CIRGenModule::getOrCreateCIRGlobal(StringRef mangledName, mlir::Type ty,
 
     if (d->getTLSKind()) {
       if (d->getTLSKind() == VarDecl::TLS_Dynamic)
-        llvm_unreachable("NYI");
+        CXXThreadLocals.push_back(d);
       setTLSMode(gv, *d);
     }
 
@@ -1622,7 +1622,7 @@ void CIRGenModule::emitGlobalVarDefinition(const clang::VarDecl *d,
 
   if (d->getTLSKind() && !gv.getTlsModelAttr()) {
     if (d->getTLSKind() == VarDecl::TLS_Dynamic)
-      llvm_unreachable("NYI");
+      CXXThreadLocals.push_back(d);
     setTLSMode(gv, *d);
   }
 

clang/lib/CIR/CodeGen/CIRGenModule.h

@@ -135,6 +135,9 @@ class CIRGenModule : public CIRGenTypeCache {
   /// for the same decl.
   llvm::DenseSet<clang::GlobalDecl> DiagnosedConflictingDefinitions;
 
+  /// thread_local variables defined or used in this TU.
+  std::vector<const clang::VarDecl *> CXXThreadLocals;
+
   /// -------
   /// Annotations
   /// -------

clang/test/CIR/CodeGen/thread-local.cpp

@@ -0,0 +1,65 @@
+// RUN: %clang_cc1 -std=c++20 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --check-prefix=CIR --input-file=%t.cir %s
+// RUN: %clang_cc1 -std=c++20 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o %t.ll
+// RUN: FileCheck --check-prefix=LLVM --input-file=%t.ll %s
+// RUN: %clang_cc1 -std=c++20 -triple x86_64-unknown-linux-gnu -emit-llvm %s -o %t.og.ll
+// RUN: FileCheck --check-prefix=OGCG --input-file=%t.og.ll %s
+
+// Test basic thread_local variable with constant initialization
+thread_local int tls_const = 42;
+// CIR: cir.global{{.*}}tls_dyn{{.*}}@tls_const = #cir.int<42> : !s32i
+// LLVM: @tls_const = thread_local global i32 42
+// OGCG: @tls_const = thread_local global i32 42
+
+// Test __thread (GNU-style) thread_local
+__thread int tls_gnu_style = 10;
+// CIR: cir.global{{.*}}tls_dyn{{.*}}@tls_gnu_style = #cir.int<10> : !s32i
+// LLVM: @tls_gnu_style = thread_local global i32 10
+// OGCG: @tls_gnu_style = thread_local global i32 10
+
+// Test thread_local function-local static (constant init)
+int get_tls_static() {
+  thread_local int tls_func_static = 100;
+  return ++tls_func_static;
+}
+// CIR-LABEL: cir.func{{.*}}@_Z14get_tls_staticv
+// CIR: cir.get_global{{.*}}@_ZZ14get_tls_staticvE15tls_func_static
+// LLVM-LABEL: @_Z14get_tls_staticv
+// LLVM: load{{.*}}@_ZZ14get_tls_staticvE15tls_func_static
+// OGCG-LABEL: @_Z14get_tls_staticv
+// OGCG: @llvm.threadlocal.address.p0(ptr{{.*}}@_ZZ14get_tls_staticvE15tls_func_static)
+
+// Test reading from thread_local variable
+int read_tls() {
+  return tls_const;
+}
+// CIR-LABEL: cir.func{{.*}}@_Z8read_tlsv
+// CIR: cir.get_global thread_local @tls_const
+// LLVM-LABEL: @_Z8read_tlsv
+// LLVM: @llvm.threadlocal.address.p0(ptr @tls_const)
+// OGCG-LABEL: @_Z8read_tlsv
+// OGCG: @llvm.threadlocal.address.p0(ptr{{.*}}@tls_const)
+
+// Test writing to thread_local variable
+void write_tls(int val) {
+  tls_const = val;
+}
+// CIR-LABEL: cir.func{{.*}}@_Z9write_tlsi
+// CIR: cir.get_global thread_local @tls_const
+// CIR: cir.store
+// LLVM-LABEL: @_Z9write_tlsi
+// LLVM: @llvm.threadlocal.address.p0(ptr @tls_const)
+// OGCG-LABEL: @_Z9write_tlsi
+// OGCG: @llvm.threadlocal.address.p0(ptr{{.*}}@tls_const)
+
+// Test extern thread_local
+extern thread_local int tls_extern;
+int use_extern_tls() {
+  return tls_extern;
+}
+// CIR-LABEL: cir.func{{.*}}@_Z14use_extern_tlsv
+// CIR: cir.get_global thread_local @tls_extern
+// LLVM-LABEL: @_Z14use_extern_tlsv
+// LLVM: @llvm.threadlocal.address.p0(ptr @tls_extern)
+// OGCG-LABEL: @_Z14use_extern_tlsv
+// OGCG: call ptr @_ZTW10tls_extern()