[CIR] Add support for member initialization from constructors (#144583)

Upstream the code to handle member variable initialization in a
constructor. At this point only simple scalar values (including members
of anonymous unions) are handled.

Author: andykaylor

clang/include/clang/CIR/MissingFeatures.h

@@ -242,6 +242,7 @@ struct MissingFeatures {
   static bool builtinCallMathErrno() { return false; }
   static bool nonFineGrainedBitfields() { return false; }
   static bool armComputeVolatileBitfields() { return false; }
+  static bool ctorMemcpyizer() { return false; }
 
   // Missing types
   static bool dataMemberType() { return false; }

clang/lib/CIR/CodeGen/CIRGenClass.cpp

@@ -53,19 +53,133 @@ bool CIRGenFunction::isConstructorDelegationValid(
   return true;
 }
 
+static void emitLValueForAnyFieldInitialization(CIRGenFunction &cgf,
+                                                CXXCtorInitializer *memberInit,
+                                                LValue &lhs) {
+  FieldDecl *field = memberInit->getAnyMember();
+  if (memberInit->isIndirectMemberInitializer()) {
+    // If we are initializing an anonymous union field, drill down to the field.
+    IndirectFieldDecl *indirectField = memberInit->getIndirectMember();
+    for (const auto *nd : indirectField->chain()) {
+      auto *fd = cast<clang::FieldDecl>(nd);
+      lhs = cgf.emitLValueForFieldInitialization(lhs, fd, fd->getName());
+    }
+  } else {
+    lhs = cgf.emitLValueForFieldInitialization(lhs, field, field->getName());
+  }
+}
+
+static void emitMemberInitializer(CIRGenFunction &cgf,
+                                  const CXXRecordDecl *classDecl,
+                                  CXXCtorInitializer *memberInit,
+                                  const CXXConstructorDecl *constructor,
+                                  FunctionArgList &args) {
+  assert(memberInit->isAnyMemberInitializer() &&
+         "Mush have member initializer!");
+  assert(memberInit->getInit() && "Must have initializer!");
+
+  assert(!cir::MissingFeatures::generateDebugInfo());
+
+  // non-static data member initializers
+  FieldDecl *field = memberInit->getAnyMember();
+  QualType fieldType = field->getType();
+
+  mlir::Value thisPtr = cgf.loadCXXThis();
+  QualType recordTy = cgf.getContext().getTypeDeclType(classDecl);
+
+  // If a base constructor is being emitted, create an LValue that has the
+  // non-virtual alignment.
+  LValue lhs = (cgf.curGD.getCtorType() == Ctor_Base)
+                   ? cgf.makeNaturalAlignPointeeAddrLValue(thisPtr, recordTy)
+                   : cgf.makeNaturalAlignAddrLValue(thisPtr, recordTy);
+
+  emitLValueForAnyFieldInitialization(cgf, memberInit, lhs);
+
+  // Special case: If we are in a copy or move constructor, and we are copying
+  // an array off PODs or classes with trivial copy constructors, ignore the AST
+  // and perform the copy we know is equivalent.
+  // FIXME: This is hacky at best... if we had a bit more explicit information
+  // in the AST, we could generalize it more easily.
+  const ConstantArrayType *array =
+      cgf.getContext().getAsConstantArrayType(fieldType);
+  if (array && constructor->isDefaulted() &&
+      constructor->isCopyOrMoveConstructor()) {
+    QualType baseElementTy = cgf.getContext().getBaseElementType(array);
+    // NOTE(cir): CodeGen allows record types to be memcpy'd if applicable,
+    // whereas ClangIR wants to represent all object construction explicitly.
+    if (!baseElementTy->isRecordType()) {
+      cgf.cgm.errorNYI(memberInit->getSourceRange(),
+                       "emitMemberInitializer: array of non-record type");
+      return;
+    }
+  }
+
+  cgf.emitInitializerForField(field, lhs, memberInit->getInit());
+}
+
 /// This routine generates necessary code to initialize base classes and
 /// non-static data members belonging to this constructor.
 void CIRGenFunction::emitCtorPrologue(const CXXConstructorDecl *cd,
                                       CXXCtorType ctorType,
                                       FunctionArgList &args) {
-  if (cd->isDelegatingConstructor())
-    return emitDelegatingCXXConstructorCall(cd, args);
+  if (cd->isDelegatingConstructor()) {
+    emitDelegatingCXXConstructorCall(cd, args);
+    return;
+  }
+
+  // If there are no member initializers, we can just return.
+  if (cd->getNumCtorInitializers() == 0)
+    return;
 
-  if (cd->getNumCtorInitializers() != 0) {
-    // There's much more to do here.
-    cgm.errorNYI(cd->getSourceRange(), "emitCtorPrologue: any initializer");
+  const CXXRecordDecl *classDecl = cd->getParent();
+
+  // This code doesn't use range-based iteration because we may need to emit
+  // code between the virtual base initializers and the non-virtual base or
+  // between the non-virtual base initializers and the member initializers.
+  CXXConstructorDecl::init_const_iterator b = cd->init_begin(),
+                                          e = cd->init_end();
+
+  // Virtual base initializers first, if any. They aren't needed if:
+  // - This is a base ctor variant
+  // - There are no vbases
+  // - The class is abstract, so a complete object of it cannot be constructed
+  //
+  // The check for an abstract class is necessary because sema may not have
+  // marked virtual base destructors referenced.
+  bool constructVBases = ctorType != Ctor_Base &&
+                         classDecl->getNumVBases() != 0 &&
+                         !classDecl->isAbstract();
+  if (constructVBases) {
+    cgm.errorNYI(cd->getSourceRange(), "emitCtorPrologue: virtual base");
     return;
   }
+
+  if ((*b)->isBaseInitializer()) {
+    cgm.errorNYI(cd->getSourceRange(),
+                 "emitCtorPrologue: non-virtual base initializer");
+    return;
+  }
+
+  if (classDecl->isDynamicClass()) {
+    cgm.errorNYI(cd->getSourceRange(),
+                 "emitCtorPrologue: initialize vtable pointers");
+    return;
+  }
+
+  // Finally, initialize class members.
+  FieldConstructionScope fcs(*this, loadCXXThisAddress());
+  // Classic codegen uses a special class to attempt to replace member
+  // initializers with memcpy. We could possibly defer that to the
+  // lowering or optimization phases to keep the memory accesses more
+  // explicit. For now, we don't insert memcpy at all.
+  assert(!cir::MissingFeatures::ctorMemcpyizer());
+  for (; b != e; b++) {
+    CXXCtorInitializer *member = (*b);
+    assert(!member->isBaseInitializer());
+    assert(member->isAnyMemberInitializer() &&
+           "Delegating initializer on non-delegating constructor");
+    emitMemberInitializer(*this, cd->getParent(), member, cd, args);
+  }
 }
 
 Address CIRGenFunction::loadCXXThisAddress() {
@@ -84,6 +198,33 @@ Address CIRGenFunction::loadCXXThisAddress() {
   return Address(loadCXXThis(), cxxThisAlignment);
 }
 
+void CIRGenFunction::emitInitializerForField(FieldDecl *field, LValue lhs,
+                                             Expr *init) {
+  QualType fieldType = field->getType();
+  switch (getEvaluationKind(fieldType)) {
+  case cir::TEK_Scalar:
+    if (lhs.isSimple())
+      emitExprAsInit(init, field, lhs, false);
+    else
+      cgm.errorNYI(field->getSourceRange(),
+                   "emitInitializerForField: non-simple scalar");
+    break;
+  case cir::TEK_Complex:
+    cgm.errorNYI(field->getSourceRange(), "emitInitializerForField: complex");
+    break;
+  case cir::TEK_Aggregate: {
+    cgm.errorNYI(field->getSourceRange(), "emitInitializerForField: aggregate");
+    break;
+  }
+  }
+
+  // Ensure that we destroy this object if an exception is thrown later in the
+  // constructor.
+  QualType::DestructionKind dtorKind = fieldType.isDestructedType();
+  (void)dtorKind;
+  assert(!cir::MissingFeatures::requiresCleanups());
+}
+
 void CIRGenFunction::emitDelegateCXXConstructorCall(
     const CXXConstructorDecl *ctor, CXXCtorType ctorType,
     const FunctionArgList &args, SourceLocation loc) {

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -390,6 +390,34 @@ LValue CIRGenFunction::emitLValueForField(LValue base, const FieldDecl *field) {
   return lv;
 }
 
+LValue CIRGenFunction::emitLValueForFieldInitialization(
+    LValue base, const clang::FieldDecl *field, llvm::StringRef fieldName) {
+  QualType fieldType = field->getType();
+
+  if (!fieldType->isReferenceType())
+    return emitLValueForField(base, field);
+
+  const CIRGenRecordLayout &layout =
+      cgm.getTypes().getCIRGenRecordLayout(field->getParent());
+  unsigned fieldIndex = layout.getCIRFieldNo(field);
+
+  Address v =
+      emitAddrOfFieldStorage(base.getAddress(), field, fieldName, fieldIndex);
+
+  // Make sure that the address is pointing to the right type.
+  mlir::Type memTy = convertTypeForMem(fieldType);
+  v = builder.createElementBitCast(getLoc(field->getSourceRange()), v, memTy);
+
+  // TODO: Generate TBAA information that describes this access as a structure
+  // member access and not just an access to an object of the field's type. This
+  // should be similar to what we do in EmitLValueForField().
+  LValueBaseInfo baseInfo = base.getBaseInfo();
+  AlignmentSource fieldAlignSource = baseInfo.getAlignmentSource();
+  LValueBaseInfo fieldBaseInfo(getFieldAlignmentSource(fieldAlignSource));
+  assert(!cir::MissingFeatures::opTBAA());
+  return makeAddrLValue(v, fieldType, fieldBaseInfo);
+}
+
 mlir::Value CIRGenFunction::emitToMemory(mlir::Value value, QualType ty) {
   // Bool has a different representation in memory than in registers,
   // but in ClangIR, it is simply represented as a cir.bool value.

clang/lib/CIR/CodeGen/CIRGenFunction.cpp

@@ -550,6 +550,15 @@ LValue CIRGenFunction::makeNaturalAlignPointeeAddrLValue(mlir::Value val,
   return makeAddrLValue(Address(val, align), ty, baseInfo);
 }
 
+LValue CIRGenFunction::makeNaturalAlignAddrLValue(mlir::Value val,
+                                                  QualType ty) {
+  LValueBaseInfo baseInfo;
+  CharUnits alignment = cgm.getNaturalTypeAlignment(ty, &baseInfo);
+  Address addr(val, convertTypeForMem(ty), alignment);
+  assert(!cir::MissingFeatures::opTBAA());
+  return makeAddrLValue(addr, ty, baseInfo);
+}
+
 clang::QualType CIRGenFunction::buildFunctionArgList(clang::GlobalDecl gd,
                                                      FunctionArgList &args) {
   const auto *fd = cast<FunctionDecl>(gd.getDecl());

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -68,6 +68,10 @@ class CIRGenFunction : public CIRGenTypeCache {
   mlir::Value cxxThisValue = nullptr;
   clang::CharUnits cxxThisAlignment;
 
+  /// The value of 'this' to sue when evaluating CXXDefaultInitExprs within this
+  /// expression.
+  Address cxxDefaultInitExprThis = Address::invalid();
+
   // Holds the Decl for the current outermost non-closure context
   const clang::Decl *curFuncDecl = nullptr;
 
@@ -490,7 +494,26 @@ class CIRGenFunction : public CIRGenTypeCache {
   static bool
   isConstructorDelegationValid(const clang::CXXConstructorDecl *ctor);
 
+  /// A scope within which we are constructing the fields of an object which
+  /// might use a CXXDefaultInitExpr. This stashes away a 'this' value to use if
+  /// we need to evaluate the CXXDefaultInitExpr within the evaluation.
+  class FieldConstructionScope {
+  public:
+    FieldConstructionScope(CIRGenFunction &cgf, Address thisAddr)
+        : cgf(cgf), oldCXXDefaultInitExprThis(cgf.cxxDefaultInitExprThis) {
+      cgf.cxxDefaultInitExprThis = thisAddr;
+    }
+    ~FieldConstructionScope() {
+      cgf.cxxDefaultInitExprThis = oldCXXDefaultInitExprThis;
+    }
+
+  private:
+    CIRGenFunction &cgf;
+    Address oldCXXDefaultInitExprThis;
+  };
+
   LValue makeNaturalAlignPointeeAddrLValue(mlir::Value v, clang::QualType t);
+  LValue makeNaturalAlignAddrLValue(mlir::Value val, QualType ty);
 
   /// Construct an address with the natural alignment of T. If a pointer to T
   /// is expected to be signed, the pointer passed to this function must have
@@ -844,6 +867,9 @@ class CIRGenFunction : public CIRGenTypeCache {
 
   mlir::LogicalResult emitFunctionBody(const clang::Stmt *body);
 
+  void emitInitializerForField(clang::FieldDecl *field, LValue lhs,
+                               clang::Expr *init);
+
   mlir::Value emitPromotedScalarExpr(const Expr *e, QualType promotionType);
 
   /// Emit the computation of the specified expression of scalar type.
@@ -940,6 +966,13 @@ class CIRGenFunction : public CIRGenTypeCache {
   LValue emitLValue(const clang::Expr *e);
   LValue emitLValueForField(LValue base, const clang::FieldDecl *field);
 
+  /// Like emitLValueForField, excpet that if the Field is a reference, this
+  /// will return the address of the reference and not the address of the value
+  /// stored in the reference.
+  LValue emitLValueForFieldInitialization(LValue base,
+                                          const clang::FieldDecl *field,
+                                          llvm::StringRef fieldName);
+
   LValue emitMemberExpr(const MemberExpr *e);
 
   /// Given an expression with a pointer type, emit the value and compute our

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -68,6 +68,7 @@ CIRGenModule::CIRGenModule(mlir::MLIRContext &mlirContext,
 
   // Initialize cached types
   VoidTy = cir::VoidType::get(&getMLIRContext());
+  VoidPtrTy = cir::PointerType::get(VoidTy);
   SInt8Ty = cir::IntType::get(&getMLIRContext(), 8, /*isSigned=*/true);
   SInt16Ty = cir::IntType::get(&getMLIRContext(), 16, /*isSigned=*/true);
   SInt32Ty = cir::IntType::get(&getMLIRContext(), 32, /*isSigned=*/true);
@@ -94,6 +95,9 @@ CIRGenModule::CIRGenModule(mlir::MLIRContext &mlirContext,
   // TODO(CIR): Should be updated once TypeSizeInfoAttr is upstreamed
   const unsigned sizeTypeSize =
       astContext.getTypeSize(astContext.getSignedSizeType());
+  // In CIRGenTypeCache, UIntPtrTy and SizeType are fields of the same union
+  UIntPtrTy =
+      cir::IntType::get(&getMLIRContext(), sizeTypeSize, /*isSigned=*/false);
   PtrDiffTy =
       cir::IntType::get(&getMLIRContext(), sizeTypeSize, /*isSigned=*/true);
 

clang/lib/CIR/CodeGen/CIRGenTypeCache.h

@@ -22,7 +22,7 @@ namespace clang::CIRGen {
 /// during IR emission. It's initialized once in CodeGenModule's
 /// constructor and then copied around into new CIRGenFunction's.
 struct CIRGenTypeCache {
-  CIRGenTypeCache() = default;
+  CIRGenTypeCache() {}
 
   // ClangIR void type
   cir::VoidType VoidTy;
@@ -49,8 +49,17 @@ struct CIRGenTypeCache {
   cir::FP80Type FP80Ty;
   cir::FP128Type FP128Ty;
 
+  /// intptr_t, size_t, and ptrdiff_t, which we assume are the same size.
+  union {
+    mlir::Type UIntPtrTy;
+    mlir::Type SizeTy;
+  };
+
   mlir::Type PtrDiffTy;
 
+  /// void* in address space 0
+  cir::PointerType VoidPtrTy;
+
   /// The size and alignment of a pointer into the generic address space.
   union {
     unsigned char PointerAlignInBytes;