[CIR] Support for static variables (#143980)

This adds support for emitting static variables and their initializers.

Author: andykaylor

clang/include/clang/CIR/Dialect/Builder/CIRBaseBuilder.h

@@ -185,11 +185,25 @@ class CIRBaseBuilderTy : public mlir::OpBuilder {
                                     global.getSymName());
   }
 
+  mlir::Value createGetGlobal(cir::GlobalOp global) {
+    return createGetGlobal(global.getLoc(), global);
+  }
+
   cir::StoreOp createStore(mlir::Location loc, mlir::Value val, mlir::Value dst,
                            mlir::IntegerAttr align = {}) {
     return create<cir::StoreOp>(loc, val, dst, align);
   }
 
+  [[nodiscard]] cir::GlobalOp createGlobal(mlir::ModuleOp mlirModule,
+                                           mlir::Location loc,
+                                           mlir::StringRef name,
+                                           mlir::Type type,
+                                           cir::GlobalLinkageKind linkage) {
+    mlir::OpBuilder::InsertionGuard guard(*this);
+    setInsertionPointToStart(mlirModule.getBody());
+    return create<cir::GlobalOp>(loc, name, type, linkage);
+  }
+
   cir::GetMemberOp createGetMember(mlir::Location loc, mlir::Type resultTy,
                                    mlir::Value base, llvm::StringRef name,
                                    unsigned index) {

clang/lib/CIR/CodeGen/CIRGenBuilder.h

@@ -24,6 +24,7 @@ namespace clang::CIRGen {
 class CIRGenBuilderTy : public cir::CIRBaseBuilderTy {
   const CIRGenTypeCache &typeCache;
   llvm::StringMap<unsigned> recordNames;
+  llvm::StringMap<unsigned> globalsVersioning;
 
 public:
   CIRGenBuilderTy(mlir::MLIRContext &mlirContext, const CIRGenTypeCache &tc)
@@ -371,6 +372,23 @@ class CIRGenBuilderTy : public cir::CIRBaseBuilderTy {
   /// pointed to by \p arrayPtr.
   mlir::Value maybeBuildArrayDecay(mlir::Location loc, mlir::Value arrayPtr,
                                    mlir::Type eltTy);
+
+  /// Creates a versioned global variable. If the symbol is already taken, an ID
+  /// will be appended to the symbol. The returned global must always be queried
+  /// for its name so it can be referenced correctly.
+  [[nodiscard]] cir::GlobalOp
+  createVersionedGlobal(mlir::ModuleOp module, mlir::Location loc,
+                        mlir::StringRef name, mlir::Type type,
+                        cir::GlobalLinkageKind linkage) {
+    // Create a unique name if the given name is already taken.
+    std::string uniqueName;
+    if (unsigned version = globalsVersioning[name.str()]++)
+      uniqueName = name.str() + "." + std::to_string(version);
+    else
+      uniqueName = name.str();
+
+    return createGlobal(module, loc, uniqueName, type, linkage);
+  }
 };
 
 } // namespace clang::CIRGen

clang/lib/CIR/CodeGen/CIRGenDecl.cpp

@@ -208,8 +208,25 @@ void CIRGenFunction::emitVarDecl(const VarDecl &d) {
   if (d.hasExternalStorage())
     return;
 
-  if (d.getStorageDuration() != SD_Automatic)
-    cgm.errorNYI(d.getSourceRange(), "emitVarDecl automatic storage duration");
+  if (d.getStorageDuration() != SD_Automatic) {
+    // Static sampler variables translated to function calls.
+    if (d.getType()->isSamplerT()) {
+      // Nothing needs to be done here, but let's flag it as an error until we
+      // have a test. It requires OpenCL support.
+      cgm.errorNYI(d.getSourceRange(), "emitVarDecl static sampler type");
+      return;
+    }
+
+    cir::GlobalLinkageKind linkage =
+        cgm.getCIRLinkageVarDefinition(&d, /*IsConstant=*/false);
+
+    // FIXME: We need to force the emission/use of a guard variable for
+    // some variables even if we can constant-evaluate them because
+    // we can't guarantee every translation unit will constant-evaluate them.
+
+    return emitStaticVarDecl(d, linkage);
+  }
+
   if (d.getType().getAddressSpace() == LangAS::opencl_local)
     cgm.errorNYI(d.getSourceRange(), "emitVarDecl openCL address space");
 
@@ -219,6 +236,233 @@ void CIRGenFunction::emitVarDecl(const VarDecl &d) {
   return emitAutoVarDecl(d);
 }
 
+static std::string getStaticDeclName(CIRGenModule &cgm, const VarDecl &d) {
+  if (cgm.getLangOpts().CPlusPlus)
+    return cgm.getMangledName(&d).str();
+
+  // If this isn't C++, we don't need a mangled name, just a pretty one.
+  assert(!d.isExternallyVisible() && "name shouldn't matter");
+  std::string contextName;
+  const DeclContext *dc = d.getDeclContext();
+  if (auto *cd = dyn_cast<CapturedDecl>(dc))
+    dc = cast<DeclContext>(cd->getNonClosureContext());
+  if (const auto *fd = dyn_cast<FunctionDecl>(dc))
+    contextName = std::string(cgm.getMangledName(fd));
+  else if (isa<BlockDecl>(dc))
+    cgm.errorNYI(d.getSourceRange(), "block decl context for static var");
+  else if (isa<ObjCMethodDecl>(dc))
+    cgm.errorNYI(d.getSourceRange(), "ObjC decl context for static var");
+  else
+    cgm.errorNYI(d.getSourceRange(), "Unknown context for static var decl");
+
+  contextName += "." + d.getNameAsString();
+  return contextName;
+}
+
+// TODO(cir): LLVM uses a Constant base class. Maybe CIR could leverage an
+// interface for all constants?
+cir::GlobalOp
+CIRGenModule::getOrCreateStaticVarDecl(const VarDecl &d,
+                                       cir::GlobalLinkageKind linkage) {
+  // In general, we don't always emit static var decls once before we reference
+  // them. It is possible to reference them before emitting the function that
+  // contains them, and it is possible to emit the containing function multiple
+  // times.
+  if (cir::GlobalOp existingGV = getStaticLocalDeclAddress(&d))
+    return existingGV;
+
+  QualType ty = d.getType();
+  assert(ty->isConstantSizeType() && "VLAs can't be static");
+
+  // Use the label if the variable is renamed with the asm-label extension.
+  if (d.hasAttr<AsmLabelAttr>())
+    errorNYI(d.getSourceRange(), "getOrCreateStaticVarDecl: asm label");
+
+  std::string name = getStaticDeclName(*this, d);
+
+  mlir::Type lty = getTypes().convertTypeForMem(ty);
+  assert(!cir::MissingFeatures::addressSpace());
+
+  if (d.hasAttr<LoaderUninitializedAttr>() || d.hasAttr<CUDASharedAttr>())
+    errorNYI(d.getSourceRange(),
+             "getOrCreateStaticVarDecl: LoaderUninitializedAttr");
+  assert(!cir::MissingFeatures::addressSpace());
+
+  mlir::Attribute init = builder.getZeroInitAttr(convertType(ty));
+
+  cir::GlobalOp gv = builder.createVersionedGlobal(
+      getModule(), getLoc(d.getLocation()), name, lty, linkage);
+  // TODO(cir): infer visibility from linkage in global op builder.
+  gv.setVisibility(getMLIRVisibilityFromCIRLinkage(linkage));
+  gv.setInitialValueAttr(init);
+  gv.setAlignment(getASTContext().getDeclAlign(&d).getAsAlign().value());
+
+  if (supportsCOMDAT() && gv.isWeakForLinker())
+    gv.setComdat(true);
+
+  assert(!cir::MissingFeatures::opGlobalThreadLocal());
+
+  setGVProperties(gv, &d);
+
+  // OG checks if the expected address space, denoted by the type, is the
+  // same as the actual address space indicated by attributes. If they aren't
+  // the same, an addrspacecast is emitted when this variable is accessed.
+  // In CIR however, cir.get_global already carries that information in
+  // !cir.ptr type - if this global is in OpenCL local address space, then its
+  // type would be !cir.ptr<..., addrspace(offload_local)>. Therefore we don't
+  // need an explicit address space cast in CIR: they will get emitted when
+  // lowering to LLVM IR.
+
+  // Ensure that the static local gets initialized by making sure the parent
+  // function gets emitted eventually.
+  const Decl *dc = cast<Decl>(d.getDeclContext());
+
+  // We can't name blocks or captured statements directly, so try to emit their
+  // parents.
+  if (isa<BlockDecl>(dc) || isa<CapturedDecl>(dc)) {
+    dc = dc->getNonClosureContext();
+    // FIXME: Ensure that global blocks get emitted.
+    if (!dc)
+      errorNYI(d.getSourceRange(), "non-closure context");
+  }
+
+  GlobalDecl gd;
+  if (isa<CXXConstructorDecl>(dc))
+    errorNYI(d.getSourceRange(), "C++ constructors static var context");
+  else if (isa<CXXDestructorDecl>(dc))
+    errorNYI(d.getSourceRange(), "C++ destructors static var context");
+  else if (const auto *fd = dyn_cast<FunctionDecl>(dc))
+    gd = GlobalDecl(fd);
+  else {
+    // Don't do anything for Obj-C method decls or global closures. We should
+    // never defer them.
+    assert(isa<ObjCMethodDecl>(dc) && "unexpected parent code decl");
+  }
+  if (gd.getDecl() && cir::MissingFeatures::openMP()) {
+    // Disable emission of the parent function for the OpenMP device codegen.
+    errorNYI(d.getSourceRange(), "OpenMP");
+  }
+
+  return gv;
+}
+
+/// Add the initializer for 'd' to the global variable that has already been
+/// created for it. If the initializer has a different type than gv does, this
+/// may free gv and return a different one. Otherwise it just returns gv.
+cir::GlobalOp CIRGenFunction::addInitializerToStaticVarDecl(
+    const VarDecl &d, cir::GlobalOp gv, cir::GetGlobalOp gvAddr) {
+  ConstantEmitter emitter(*this);
+  mlir::TypedAttr init =
+      mlir::cast<mlir::TypedAttr>(emitter.tryEmitForInitializer(d));
+
+  // If constant emission failed, then this should be a C++ static
+  // initializer.
+  if (!init) {
+    cgm.errorNYI(d.getSourceRange(), "static var without initializer");
+    return gv;
+  }
+
+  // TODO(cir): There should be debug code here to assert that the decl size
+  // matches the CIR data layout type alloc size, but the code for calculating
+  // the type alloc size is not implemented yet.
+  assert(!cir::MissingFeatures::dataLayoutTypeAllocSize());
+
+  // The initializer may differ in type from the global. Rewrite
+  // the global to match the initializer.  (We have to do this
+  // because some types, like unions, can't be completely represented
+  // in the LLVM type system.)
+  if (gv.getSymType() != init.getType()) {
+    gv.setSymType(init.getType());
+
+    // Normally this should be done with a call to cgm.replaceGlobal(oldGV, gv),
+    // but since at this point the current block hasn't been really attached,
+    // there's no visibility into the GetGlobalOp corresponding to this Global.
+    // Given those constraints, thread in the GetGlobalOp and update it
+    // directly.
+    assert(!cir::MissingFeatures::addressSpace());
+    gvAddr.getAddr().setType(builder.getPointerTo(init.getType()));
+  }
+
+  bool needsDtor =
+      d.needsDestruction(getContext()) == QualType::DK_cxx_destructor;
+
+  assert(!cir::MissingFeatures::opGlobalConstant());
+  gv.setInitialValueAttr(init);
+
+  emitter.finalize(gv);
+
+  if (needsDtor) {
+    // We have a constant initializer, but a nontrivial destructor. We still
+    // need to perform a guarded "initialization" in order to register the
+    // destructor.
+    cgm.errorNYI(d.getSourceRange(), "C++ guarded init");
+  }
+
+  return gv;
+}
+
+void CIRGenFunction::emitStaticVarDecl(const VarDecl &d,
+                                       cir::GlobalLinkageKind linkage) {
+  // Check to see if we already have a global variable for this
+  // declaration.  This can happen when double-emitting function
+  // bodies, e.g. with complete and base constructors.
+  cir::GlobalOp globalOp = cgm.getOrCreateStaticVarDecl(d, linkage);
+  // TODO(cir): we should have a way to represent global ops as values without
+  // having to emit a get global op. Sometimes these emissions are not used.
+  mlir::Value addr = builder.createGetGlobal(globalOp);
+  auto getAddrOp = mlir::cast<cir::GetGlobalOp>(addr.getDefiningOp());
+
+  CharUnits alignment = getContext().getDeclAlign(&d);
+
+  // Store into LocalDeclMap before generating initializer to handle
+  // circular references.
+  mlir::Type elemTy = convertTypeForMem(d.getType());
+  setAddrOfLocalVar(&d, Address(addr, elemTy, alignment));
+
+  // We can't have a VLA here, but we can have a pointer to a VLA,
+  // even though that doesn't really make any sense.
+  // Make sure to evaluate VLA bounds now so that we have them for later.
+  if (d.getType()->isVariablyModifiedType()) {
+    cgm.errorNYI(d.getSourceRange(),
+                 "emitStaticVarDecl: variably modified type");
+  }
+
+  // Save the type in case adding the initializer forces a type change.
+  mlir::Type expectedType = addr.getType();
+
+  cir::GlobalOp var = globalOp;
+
+  assert(!cir::MissingFeatures::cudaSupport());
+
+  // If this value has an initializer, emit it.
+  if (d.getInit())
+    var = addInitializerToStaticVarDecl(d, var, getAddrOp);
+
+  var.setAlignment(alignment.getAsAlign().value());
+
+  // There are a lot of attributes that need to be handled here. Until
+  // we start to support them, we just report an error if there are any.
+  if (d.hasAttrs())
+    cgm.errorNYI(d.getSourceRange(), "static var with attrs");
+
+  if (cgm.getCodeGenOpts().KeepPersistentStorageVariables)
+    cgm.errorNYI(d.getSourceRange(), "static var keep persistent storage");
+
+  // From traditional codegen:
+  // We may have to cast the constant because of the initializer
+  // mismatch above.
+  //
+  // FIXME: It is really dangerous to store this in the map; if anyone
+  // RAUW's the GV uses of this constant will be invalid.
+  mlir::Value castedAddr =
+      builder.createBitcast(getAddrOp.getAddr(), expectedType);
+  localDeclMap.find(&d)->second = Address(castedAddr, elemTy, alignment);
+  cgm.setStaticLocalDeclAddress(&d, var);
+
+  assert(!cir::MissingFeatures::sanitizers());
+  assert(!cir::MissingFeatures::generateDebugInfo());
+}
+
 void CIRGenFunction::emitScalarInit(const Expr *init, mlir::Location loc,
                                     LValue lvalue, bool capturedByInit) {
   assert(!cir::MissingFeatures::objCLifetime());

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -469,6 +469,10 @@ class CIRGenFunction : public CIRGenTypeCache {
   /// compare the result against zero, returning an Int1Ty value.
   mlir::Value evaluateExprAsBool(const clang::Expr *e);
 
+  cir::GlobalOp addInitializerToStaticVarDecl(const VarDecl &d,
+                                              cir::GlobalOp gv,
+                                              cir::GetGlobalOp gvAddr);
+
   /// Set the address of a local variable.
   void setAddrOfLocalVar(const clang::VarDecl *vd, Address addr) {
     assert(!localDeclMap.count(vd) && "Decl already exists in LocalDeclMap!");
@@ -955,6 +959,8 @@ class CIRGenFunction : public CIRGenTypeCache {
   void emitScalarInit(const clang::Expr *init, mlir::Location loc,
                       LValue lvalue, bool capturedByInit = false);
 
+  void emitStaticVarDecl(const VarDecl &d, cir::GlobalLinkageKind linkage);
+
   void emitStoreOfScalar(mlir::Value value, Address addr, bool isVolatile,
                          clang::QualType ty, bool isInit = false,
                          bool isNontemporal = false);

clang/lib/CIR/CodeGen/CIRGenModule.h

@@ -113,8 +113,21 @@ class CIRGenModule : public CIRGenTypeCache {
 
   mlir::Operation *lastGlobalOp = nullptr;
 
+  llvm::DenseMap<const Decl *, cir::GlobalOp> staticLocalDeclMap;
+
   mlir::Operation *getGlobalValue(llvm::StringRef ref);
 
+  cir::GlobalOp getStaticLocalDeclAddress(const VarDecl *d) {
+    return staticLocalDeclMap[d];
+  }
+
+  void setStaticLocalDeclAddress(const VarDecl *d, cir::GlobalOp c) {
+    staticLocalDeclMap[d] = c;
+  }
+
+  cir::GlobalOp getOrCreateStaticVarDecl(const VarDecl &d,
+                                         cir::GlobalLinkageKind linkage);
+
   /// 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,

clang/test/CIR/CodeGen/static-vars.c

@@ -0,0 +1,37 @@
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --input-file=%t.cir %s
+
+void func1(void) {
+  // Should lower default-initialized static vars.
+  static int i;
+  // CHECK-DAG: cir.global "private" internal dsolocal @func1.i = #cir.int<0> : !s32i
+
+  // Should lower constant-initialized static vars.
+  static int j = 1;
+  // CHECK-DAG: cir.global "private" internal dsolocal @func1.j = #cir.int<1> : !s32i
+
+  // Should properly shadow static vars in nested scopes.
+  {
+    static int j = 2;
+    // CHECK-DAG: cir.global "private" internal dsolocal @func1.j.1 = #cir.int<2> : !s32i
+  }
+  {
+    static int j = 3;
+    // CHECK-DAG: cir.global "private" internal dsolocal @func1.j.2 = #cir.int<3> : !s32i
+  }
+
+  // Should lower basic static vars arithmetics.
+  j++;
+  // CHECK-DAG: %[[#V2:]] = cir.get_global @func1.j : !cir.ptr<!s32i>
+  // CHECK-DAG: %[[#V3:]] = cir.load{{.*}} %[[#V2]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-DAG: %[[#V4:]] = cir.unary(inc, %[[#V3]]) nsw : !s32i, !s32i
+  // CHECK-DAG: cir.store{{.*}} %[[#V4]], %[[#V2]] : !s32i, !cir.ptr<!s32i>
+}
+
+// Should shadow static vars on different functions.
+void func2(void) {
+  static char i;
+  // CHECK-DAG: cir.global "private" internal dsolocal @func2.i = #cir.int<0> : !s8i
+  static float j;
+  // CHECK-DAG: cir.global "private" internal dsolocal @func2.j = #cir.fp<0.000000e+00> : !cir.float
+}