[CIR] Upstream support for cir.get_global (#135095)

This adds basic support for referencing global variables from within
functions via the cir.get_global operation.

Author: andykaylor

clang/include/clang/CIR/Dialect/IR/CIROps.td

@@ -1279,6 +1279,37 @@ def GlobalOp : CIR_Op<"global"> {
   let hasVerifier = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// GetGlobalOp
+//===----------------------------------------------------------------------===//
+
+def GetGlobalOp : CIR_Op<"get_global",
+    [Pure, DeclareOpInterfaceMethods<SymbolUserOpInterface>]> {
+  let summary = "Get the address of a global variable";
+  let description = [{
+    The `cir.get_global` operation retrieves the address pointing to a
+    named global variable. If the global variable is marked constant, writing
+    to the resulting address (such as through a `cir.store` operation) is
+    undefined. The resulting type must always be a `!cir.ptr<...>` type with the
+    same address space as the global variable.
+
+    Example:
+    ```mlir
+    %x = cir.get_global @gv : !cir.ptr<i32>
+    ```
+  }];
+
+  let arguments = (ins FlatSymbolRefAttr:$name);
+  let results = (outs Res<CIR_PointerType, "", []>:$addr);
+
+  let assemblyFormat = [{
+    $name `:` qualified(type($addr)) attr-dict
+  }];
+
+  // `GetGlobalOp` is fully verified by its traits.
+  let hasVerifier = 0;
+}
+
 //===----------------------------------------------------------------------===//
 // FuncOp
 //===----------------------------------------------------------------------===//

clang/include/clang/CIR/MissingFeatures.h

@@ -35,6 +35,7 @@ struct MissingFeatures {
   static bool opGlobalThreadLocal() { return false; }
   static bool opGlobalConstant() { return false; }
   static bool opGlobalAlignment() { return false; }
+  static bool opGlobalWeakRef() { return false; }
 
   static bool supportIFuncAttr() { return false; }
   static bool supportVisibility() { return false; }
@@ -136,6 +137,10 @@ struct MissingFeatures {
   static bool objCGC() { return false; }
   static bool weakRefReference() { return false; }
   static bool hip() { return false; }
+  static bool setObjCGCLValueClass() { return false; }
+  static bool mangledNames() { return false; }
+  static bool setDLLStorageClass() { return false; }
+  static bool openMP() { return false; }
 
   // Missing types
   static bool dataMemberType() { return false; }

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -183,6 +183,43 @@ void CIRGenFunction::emitStoreThroughLValue(RValue src, LValue dst,
   emitStoreOfScalar(src.getScalarVal(), dst, isInit);
 }
 
+static LValue emitGlobalVarDeclLValue(CIRGenFunction &cgf, const Expr *e,
+                                      const VarDecl *vd) {
+  QualType T = e->getType();
+
+  // If it's thread_local, emit a call to its wrapper function instead.
+  assert(!cir::MissingFeatures::opGlobalThreadLocal());
+  if (vd->getTLSKind() == VarDecl::TLS_Dynamic)
+    cgf.cgm.errorNYI(e->getSourceRange(),
+                     "emitGlobalVarDeclLValue: thread_local variable");
+
+  // Check if the variable is marked as declare target with link clause in
+  // device codegen.
+  if (cgf.getLangOpts().OpenMP)
+    cgf.cgm.errorNYI(e->getSourceRange(), "emitGlobalVarDeclLValue: OpenMP");
+
+  // Traditional LLVM codegen handles thread local separately, CIR handles
+  // as part of getAddrOfGlobalVar.
+  mlir::Value v = cgf.cgm.getAddrOfGlobalVar(vd);
+
+  assert(!cir::MissingFeatures::addressSpace());
+  mlir::Type realVarTy = cgf.convertTypeForMem(vd->getType());
+  cir::PointerType realPtrTy = cgf.getBuilder().getPointerTo(realVarTy);
+  if (realPtrTy != v.getType())
+    v = cgf.getBuilder().createBitcast(v.getLoc(), v, realPtrTy);
+
+  CharUnits alignment = cgf.getContext().getDeclAlign(vd);
+  Address addr(v, realVarTy, alignment);
+  LValue lv;
+  if (vd->getType()->isReferenceType())
+    cgf.cgm.errorNYI(e->getSourceRange(),
+                     "emitGlobalVarDeclLValue: reference type");
+  else
+    lv = cgf.makeAddrLValue(addr, T, AlignmentSource::Decl);
+  assert(!cir::MissingFeatures::setObjCGCLValueClass());
+  return lv;
+}
+
 void CIRGenFunction::emitStoreOfScalar(mlir::Value value, Address addr,
                                        bool isVolatile, QualType ty,
                                        bool isInit, bool isNontemporal) {
@@ -288,7 +325,7 @@ LValue CIRGenFunction::emitDeclRefLValue(const DeclRefExpr *e) {
 
     // Check if this is a global variable
     if (vd->hasLinkage() || vd->isStaticDataMember())
-      cgm.errorNYI(vd->getSourceRange(), "emitDeclRefLValue: global variable");
+      return emitGlobalVarDeclLValue(*this, e, vd);
 
     Address addr = Address::invalid();
 
@@ -299,7 +336,7 @@ LValue CIRGenFunction::emitDeclRefLValue(const DeclRefExpr *e) {
     } else {
       // Otherwise, it might be static local we haven't emitted yet for some
       // reason; most likely, because it's in an outer function.
-      cgm.errorNYI(vd->getSourceRange(), "emitDeclRefLValue: static local");
+      cgm.errorNYI(e->getSourceRange(), "emitDeclRefLValue: static local");
     }
 
     return makeAddrLValue(addr, ty, AlignmentSource::Type);

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -202,6 +202,102 @@ void CIRGenModule::emitGlobalFunctionDefinition(clang::GlobalDecl gd,
   curCGF = nullptr;
 }
 
+mlir::Operation *CIRGenModule::getGlobalValue(StringRef name) {
+  return mlir::SymbolTable::lookupSymbolIn(theModule, name);
+}
+
+/// If the specified mangled name is not in the module,
+/// create and return an mlir GlobalOp with the specified type (TODO(cir):
+/// address space).
+///
+/// TODO(cir):
+/// 1. If there is something in the module with the specified name, return
+/// it potentially bitcasted to the right type.
+///
+/// 2. If \p d is non-null, it specifies a decl that correspond to this.  This
+/// is used to set the attributes on the global when it is first created.
+///
+/// 3. If \p isForDefinition is true, it is guaranteed that an actual global
+/// with type \p ty will be returned, not conversion of a variable with the same
+/// mangled name but some other type.
+cir::GlobalOp
+CIRGenModule::getOrCreateCIRGlobal(StringRef mangledName, mlir::Type ty,
+                                   LangAS langAS, const VarDecl *d,
+                                   ForDefinition_t isForDefinition) {
+  // Lookup the entry, lazily creating it if necessary.
+  cir::GlobalOp entry;
+  if (mlir::Operation *v = getGlobalValue(mangledName)) {
+    if (!isa<cir::GlobalOp>(v))
+      errorNYI(d->getSourceRange(), "global with non-GlobalOp type");
+    entry = cast<cir::GlobalOp>(v);
+  }
+
+  if (entry) {
+    assert(!cir::MissingFeatures::addressSpace());
+    assert(!cir::MissingFeatures::opGlobalWeakRef());
+
+    assert(!cir::MissingFeatures::setDLLStorageClass());
+    assert(!cir::MissingFeatures::openMP());
+
+    if (entry.getSymType() == ty)
+      return entry;
+
+    // If there are two attempts to define the same mangled name, issue an
+    // error.
+    //
+    // TODO(cir): look at mlir::GlobalValue::isDeclaration for all aspects of
+    // recognizing the global as a declaration, for now only check if
+    // initializer is present.
+    if (isForDefinition && !entry.isDeclaration()) {
+      errorNYI(d->getSourceRange(), "global with conflicting type");
+    }
+
+    // Address space check removed because it is unnecessary because CIR records
+    // address space info in types.
+
+    // (If global is requested for a definition, we always need to create a new
+    // global, not just return a bitcast.)
+    if (!isForDefinition)
+      return entry;
+  }
+
+  errorNYI(d->getSourceRange(), "reference of undeclared global");
+}
+
+cir::GlobalOp
+CIRGenModule::getOrCreateCIRGlobal(const VarDecl *d, mlir::Type ty,
+                                   ForDefinition_t isForDefinition) {
+  assert(d->hasGlobalStorage() && "Not a global variable");
+  QualType astTy = d->getType();
+  if (!ty)
+    ty = getTypes().convertTypeForMem(astTy);
+
+  assert(!cir::MissingFeatures::mangledNames());
+  return getOrCreateCIRGlobal(d->getIdentifier()->getName(), ty,
+                              astTy.getAddressSpace(), d, isForDefinition);
+}
+
+/// Return the mlir::Value for the address of the given global variable. If
+/// \p ty is non-null and if the global doesn't exist, then it will be created
+/// with the specified type instead of whatever the normal requested type would
+/// be. If \p isForDefinition is true, it is guaranteed that an actual global
+/// with type \p ty will be returned, not conversion of a variable with the same
+/// mangled name but some other type.
+mlir::Value CIRGenModule::getAddrOfGlobalVar(const VarDecl *d, mlir::Type ty,
+                                             ForDefinition_t isForDefinition) {
+  assert(d->hasGlobalStorage() && "Not a global variable");
+  QualType astTy = d->getType();
+  if (!ty)
+    ty = getTypes().convertTypeForMem(astTy);
+
+  assert(!cir::MissingFeatures::opGlobalThreadLocal());
+
+  cir::GlobalOp g = getOrCreateCIRGlobal(d, ty, isForDefinition);
+  mlir::Type ptrTy = builder.getPointerTo(g.getSymType());
+  return builder.create<cir::GetGlobalOp>(getLoc(d->getSourceRange()), ptrTy,
+                                          g.getSymName());
+}
+
 void CIRGenModule::emitGlobalVarDefinition(const clang::VarDecl *vd,
                                            bool isTentative) {
   const QualType astTy = vd->getType();
@@ -507,6 +603,7 @@ cir::FuncOp CIRGenModule::getAddrOfFunction(clang::GlobalDecl gd,
     funcType = convertType(fd->getType());
   }
 
+  assert(!cir::MissingFeatures::mangledNames());
   cir::FuncOp func = getOrCreateCIRFunction(
       cast<NamedDecl>(gd.getDecl())->getIdentifier()->getName(), funcType, gd,
       forVTable, dontDefer, /*isThunk=*/false, isForDefinition);

clang/lib/CIR/CodeGen/CIRGenModule.h

@@ -91,6 +91,31 @@ class CIRGenModule : public CIRGenTypeCache {
   const clang::LangOptions &getLangOpts() const { return langOpts; }
   mlir::MLIRContext &getMLIRContext() { return *builder.getContext(); }
 
+  /// -------
+  /// Handling globals
+  /// -------
+
+  mlir::Operation *getGlobalValue(llvm::StringRef ref);
+
+  /// If the specified mangled name is not in the module, create and return an
+  /// mlir::GlobalOp value
+  cir::GlobalOp getOrCreateCIRGlobal(llvm::StringRef mangledName, mlir::Type ty,
+                                     LangAS langAS, const VarDecl *d,
+                                     ForDefinition_t isForDefinition);
+
+  cir::GlobalOp getOrCreateCIRGlobal(const VarDecl *d, mlir::Type ty,
+                                     ForDefinition_t isForDefinition);
+
+  /// Return the mlir::Value for the address of the given global variable.
+  /// If Ty is non-null and if the global doesn't exist, then it will be created
+  /// with the specified type instead of whatever the normal requested type
+  /// would be. If IsForDefinition is true, it is guaranteed that an actual
+  /// global with type Ty will be returned, not conversion of a variable with
+  /// the same mangled name but some other type.
+  mlir::Value
+  getAddrOfGlobalVar(const VarDecl *d, mlir::Type ty = {},
+                     ForDefinition_t isForDefinition = NotForDefinition);
+
   /// Helpers to convert the presumed location of Clang's SourceLocation to an
   /// MLIR Location.
   mlir::Location getLoc(clang::SourceLocation cLoc);

clang/lib/CIR/Dialect/IR/CIRDialect.cpp

@@ -848,6 +848,41 @@ parseGlobalOpTypeAndInitialValue(OpAsmParser &parser, TypeAttr &typeAttr,
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// GetGlobalOp
+//===----------------------------------------------------------------------===//
+
+LogicalResult
+cir::GetGlobalOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
+  // Verify that the result type underlying pointer type matches the type of
+  // the referenced cir.global or cir.func op.
+  mlir::Operation *op =
+      symbolTable.lookupNearestSymbolFrom(*this, getNameAttr());
+  if (op == nullptr || !(isa<GlobalOp>(op) || isa<FuncOp>(op)))
+    return emitOpError("'")
+           << getName()
+           << "' does not reference a valid cir.global or cir.func";
+
+  mlir::Type symTy;
+  if (auto g = dyn_cast<GlobalOp>(op)) {
+    symTy = g.getSymType();
+    assert(!cir::MissingFeatures::addressSpace());
+    assert(!cir::MissingFeatures::opGlobalThreadLocal());
+  } else if (auto f = dyn_cast<FuncOp>(op)) {
+    symTy = f.getFunctionType();
+  } else {
+    llvm_unreachable("Unexpected operation for GetGlobalOp");
+  }
+
+  auto resultType = dyn_cast<PointerType>(getAddr().getType());
+  if (!resultType || symTy != resultType.getPointee())
+    return emitOpError("result type pointee type '")
+           << resultType.getPointee() << "' does not match type " << symTy
+           << " of the global @" << getName();
+
+  return success();
+}
+
 //===----------------------------------------------------------------------===//
 // FuncOp
 //===----------------------------------------------------------------------===//

clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp

@@ -828,6 +828,26 @@ mlir::LogicalResult CIRToLLVMFuncOpLowering::matchAndRewrite(
   return mlir::LogicalResult::success();
 }
 
+mlir::LogicalResult CIRToLLVMGetGlobalOpLowering::matchAndRewrite(
+    cir::GetGlobalOp op, OpAdaptor adaptor,
+    mlir::ConversionPatternRewriter &rewriter) const {
+  // FIXME(cir): Premature DCE to avoid lowering stuff we're not using.
+  // CIRGen should mitigate this and not emit the get_global.
+  if (op->getUses().empty()) {
+    rewriter.eraseOp(op);
+    return mlir::success();
+  }
+
+  mlir::Type type = getTypeConverter()->convertType(op.getType());
+  mlir::Operation *newop =
+      rewriter.create<mlir::LLVM::AddressOfOp>(op.getLoc(), type, op.getName());
+
+  assert(!cir::MissingFeatures::opGlobalThreadLocal());
+
+  rewriter.replaceOp(op, newop);
+  return mlir::success();
+}
+
 /// Replace CIR global with a region initialized LLVM global and update
 /// insertion point to the end of the initializer block.
 void CIRToLLVMGlobalOpLowering::setupRegionInitializedLLVMGlobalOp(
@@ -1418,6 +1438,7 @@ void ConvertCIRToLLVMPass::runOnOperation() {
                CIRToLLVMCmpOpLowering,
                CIRToLLVMConstantOpLowering,
                CIRToLLVMFuncOpLowering,
+               CIRToLLVMGetGlobalOpLowering,
                CIRToLLVMTrapOpLowering,
                CIRToLLVMUnaryOpLowering
       // clang-format on

clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.h

@@ -140,6 +140,16 @@ class CIRToLLVMFuncOpLowering : public mlir::OpConversionPattern<cir::FuncOp> {
                   mlir::ConversionPatternRewriter &) const override;
 };
 
+class CIRToLLVMGetGlobalOpLowering
+    : public mlir::OpConversionPattern<cir::GetGlobalOp> {
+public:
+  using mlir::OpConversionPattern<cir::GetGlobalOp>::OpConversionPattern;
+
+  mlir::LogicalResult
+  matchAndRewrite(cir::GetGlobalOp op, OpAdaptor,
+                  mlir::ConversionPatternRewriter &) const override;
+};
+
 class CIRToLLVMGlobalOpLowering
     : public mlir::OpConversionPattern<cir::GlobalOp> {
   const mlir::DataLayout &dataLayout;

clang/test/CIR/CodeGen/basic.c

@@ -144,3 +144,28 @@ void f5(void) {
 // OGCG:   br label %[[LOOP:.*]]
 // OGCG: [[LOOP]]:
 // OGCG:   br label %[[LOOP]]
+
+int gv;
+int f6(void) {
+  return gv;
+}
+
+//      CIR: cir.func @f6() -> !s32i
+// CIR-NEXT:   %[[RV:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["__retval"] {alignment = 4 : i64}
+// CIR-NEXT:   %[[GV_PTR:.*]] = cir.get_global @gv : !cir.ptr<!s32i>
+// CIR-NEXT:   %[[GV:.*]] = cir.load %[[GV_PTR]] : !cir.ptr<!s32i>, !s32i
+// CIR-NEXT:   cir.store %[[GV]], %[[RV]] : !s32i, !cir.ptr<!s32i>
+// CIR-NEXT:   %[[R:.*]] = cir.load %[[RV]] : !cir.ptr<!s32i>, !s32i
+// CIR-NEXT:   cir.return %[[R]] : !s32i
+
+// LLVM:      define i32 @f6()
+// LLVM-NEXT:   %[[RV_PTR:.*]] = alloca i32, i64 1, align 4
+// LLVM-NEXT:   %[[GV:.*]] = load i32, ptr @gv, align 4
+// LLVM-NEXT:   store i32 %[[GV]], ptr %[[RV_PTR]], align 4
+// LLVM-NEXT:   %[[RV:.*]] = load i32, ptr %[[RV_PTR]], align 4
+// LLVM-NEXT:   ret i32 %[[RV]]
+
+// OGCG:      define{{.*}} i32 @f6()
+// OGCG-NEXT: entry:
+// OGCG-NEXT:   %[[GV:.*]] = load i32, ptr @gv, align 4
+// OGCG-NEXT:   ret i32 %[[GV]]