[CIR] Add support for accessing members of base classes (llvm#143195)

This change adds the support for accessing a member of a base class from
a derived class object.

Author: andykaylor

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

@@ -278,6 +278,11 @@ class CIRBaseBuilderTy : public mlir::OpBuilder {
     return createCast(loc, cir::CastKind::bitcast, src, newTy);
   }
 
+  mlir::Value createPtrBitcast(mlir::Value src, mlir::Type newPointeeTy) {
+    assert(mlir::isa<cir::PointerType>(src.getType()) && "expected ptr src");
+    return createBitcast(src, getPointerTo(newPointeeTy));
+  }
+
   //===--------------------------------------------------------------------===//
   // Binary Operators
   //===--------------------------------------------------------------------===//

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

@@ -2275,4 +2275,47 @@ def VecTernaryOp : CIR_Op<"vec.ternary",
   let hasFolder = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// BaseClassAddrOp
+//===----------------------------------------------------------------------===//
+
+def BaseClassAddrOp : CIR_Op<"base_class_addr"> {
+  let summary = "Get the base class address for a class/struct";
+  let description = [{
+    The `cir.base_class_addr` operaration gets the address of a particular
+    non-virtual base class given a derived class pointer. The offset in bytes
+    of the base class must be passed in, since it is easier for the front end
+    to calculate that than the MLIR passes. The operation contains a flag for
+    whether or not the operand may be nullptr. That depends on the context and
+    cannot be known by the operation, and that information affects how the
+    operation is lowered.
+
+    Example:
+    ```c++
+    struct Base { };
+    struct Derived : Base { };
+    Derived d;
+    Base& b = d;
+    ```
+    will generate
+    ```mlir
+    %3 = cir.base_class_addr %1 : !cir.ptr<!rec_Derived> nonnull [0] -> !cir.ptr<!rec_Base>
+    ```
+  }];
+
+  // The validity of the relationship of derived and base cannot yet be
+  // verified, currently not worth adding a verifier.
+  let arguments = (ins
+    Arg<CIR_PointerType, "derived class pointer", [MemRead]>:$derived_addr,
+    IndexAttr:$offset, UnitAttr:$assume_not_null);
+
+  let results = (outs Res<CIR_PointerType, "">:$base_addr);
+
+  let assemblyFormat = [{
+      $derived_addr `:` qualified(type($derived_addr))
+      (`nonnull` $assume_not_null^)?
+      ` ` `[` $offset `]` `->` qualified(type($base_addr)) attr-dict
+  }];
+}
+
 #endif // CLANG_CIR_DIALECT_IR_CIROPS_TD

clang/include/clang/CIR/MissingFeatures.h

@@ -151,6 +151,11 @@ struct MissingFeatures {
   static bool cxxabiAppleARM64CXXABI() { return false; }
   static bool cxxabiStructorImplicitParam() { return false; }
 
+  // Address class
+  static bool addressOffset() { return false; }
+  static bool addressIsKnownNonNull() { return false; }
+  static bool addressPointerAuthInfo() { return false; }
+
   // Misc
   static bool cirgenABIInfo() { return false; }
   static bool abiArgInfo() { return false; }

clang/lib/CIR/CodeGen/Address.h

@@ -21,6 +21,9 @@
 
 namespace clang::CIRGen {
 
+// Forward declaration to avoid a circular dependency
+class CIRGenBuilderTy;
+
 class Address {
 
   // The boolean flag indicates whether the pointer is known to be non-null.
@@ -65,11 +68,22 @@ class Address {
     return pointerAndKnownNonNull.getPointer() != nullptr;
   }
 
+  /// Return address with different element type, a bitcast pointer, and
+  /// the same alignment.
+  Address withElementType(CIRGenBuilderTy &builder, mlir::Type ElemTy) const;
+
   mlir::Value getPointer() const {
     assert(isValid());
     return pointerAndKnownNonNull.getPointer();
   }
 
+  mlir::Value getBasePointer() const {
+    // TODO(cir): Remove the version above when we catchup with OG codegen on
+    // ptr auth.
+    assert(isValid() && "pointer isn't valid");
+    return getPointer();
+  }
+
   mlir::Type getType() const {
     assert(mlir::cast<cir::PointerType>(
                pointerAndKnownNonNull.getPointer().getType())

clang/lib/CIR/CodeGen/CIRGenBuilder.cpp

@@ -38,3 +38,15 @@ mlir::Value CIRGenBuilderTy::getArrayElement(mlir::Location arrayLocBegin,
   const mlir::Type flatPtrTy = basePtr.getType();
   return create<cir::PtrStrideOp>(arrayLocEnd, flatPtrTy, basePtr, idx);
 }
+
+// This can't be defined in Address.h because that file is included by
+// CIRGenBuilder.h
+Address Address::withElementType(CIRGenBuilderTy &builder,
+                                 mlir::Type elemTy) const {
+  assert(!cir::MissingFeatures::addressOffset());
+  assert(!cir::MissingFeatures::addressIsKnownNonNull());
+  assert(!cir::MissingFeatures::addressPointerAuthInfo());
+
+  return Address(builder.createPtrBitcast(getBasePointer(), elemTy), elemTy,
+                 getAlignment());
+}

clang/lib/CIR/CodeGen/CIRGenBuilder.h

@@ -309,6 +309,18 @@ class CIRGenBuilderTy : public cir::CIRBaseBuilderTy {
     return create<cir::BinOp>(loc, cir::BinOpKind::Div, lhs, rhs);
   }
 
+  Address createBaseClassAddr(mlir::Location loc, Address addr,
+                              mlir::Type destType, unsigned offset,
+                              bool assumeNotNull) {
+    if (destType == addr.getElementType())
+      return addr;
+
+    auto ptrTy = getPointerTo(destType);
+    auto baseAddr = create<cir::BaseClassAddrOp>(
+        loc, ptrTy, addr.getPointer(), mlir::APInt(64, offset), assumeNotNull);
+    return Address(baseAddr, destType, addr.getAlignment());
+  }
+
   cir::LoadOp createLoad(mlir::Location loc, Address addr,
                          bool isVolatile = false) {
     mlir::IntegerAttr align = getAlignmentAttr(addr.getAlignment());

clang/lib/CIR/CodeGen/CIRGenClass.cpp

@@ -0,0 +1,65 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This contains code dealing with C++ code generation of classes
+//
+//===----------------------------------------------------------------------===//
+
+#include "CIRGenFunction.h"
+
+#include "clang/AST/RecordLayout.h"
+#include "clang/CIR/MissingFeatures.h"
+
+using namespace clang;
+using namespace clang::CIRGen;
+
+Address CIRGenFunction::getAddressOfBaseClass(
+    Address value, const CXXRecordDecl *derived,
+    llvm::iterator_range<CastExpr::path_const_iterator> path,
+    bool nullCheckValue, SourceLocation loc) {
+  assert(!path.empty() && "Base path should not be empty!");
+
+  if ((*path.begin())->isVirtual()) {
+    // The implementation here is actually complete, but let's flag this
+    // as an error until the rest of the virtual base class support is in place.
+    cgm.errorNYI(loc, "getAddrOfBaseClass: virtual base");
+    return Address::invalid();
+  }
+
+  // Compute the static offset of the ultimate destination within its
+  // allocating subobject (the virtual base, if there is one, or else
+  // the "complete" object that we see).
+  CharUnits nonVirtualOffset =
+      cgm.computeNonVirtualBaseClassOffset(derived, path);
+
+  // Get the base pointer type.
+  mlir::Type baseValueTy = convertType((path.end()[-1])->getType());
+  assert(!cir::MissingFeatures::addressSpace());
+
+  // The if statement here is redundant now, but it will be needed when we add
+  // support for virtual base classes.
+  // If there is no virtual base, use cir.base_class_addr.  It takes care of
+  // the adjustment and the null pointer check.
+  if (nonVirtualOffset.isZero()) {
+    assert(!cir::MissingFeatures::sanitizers());
+    return builder.createBaseClassAddr(getLoc(loc), value, baseValueTy, 0,
+                                       /*assumeNotNull=*/true);
+  }
+
+  assert(!cir::MissingFeatures::sanitizers());
+
+  // Apply the offset
+  value = builder.createBaseClassAddr(getLoc(loc), value, baseValueTy,
+                                      nonVirtualOffset.getQuantity(),
+                                      /*assumeNotNull=*/true);
+
+  // Cast to the destination type.
+  value = value.withElementType(builder, baseValueTy);
+
+  return value;
+}

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -98,9 +98,14 @@ Address CIRGenFunction::emitPointerWithAlignment(const Expr *expr,
 
     case CK_UncheckedDerivedToBase:
     case CK_DerivedToBase: {
-      cgm.errorNYI(expr->getSourceRange(),
-                   "emitPointerWithAlignment: derived-to-base cast");
-      return Address::invalid();
+      assert(!cir::MissingFeatures::opTBAA());
+      assert(!cir::MissingFeatures::addressIsKnownNonNull());
+      Address addr = emitPointerWithAlignment(ce->getSubExpr(), baseInfo);
+      const CXXRecordDecl *derived =
+          ce->getSubExpr()->getType()->getPointeeCXXRecordDecl();
+      return getAddressOfBaseClass(addr, derived, ce->path(),
+                                   shouldNullCheckClassCastValue(ce),
+                                   ce->getExprLoc());
     }
 
     case CK_AnyPointerToBlockPointerCast:
@@ -824,8 +829,6 @@ LValue CIRGenFunction::emitCastLValue(const CastExpr *e) {
   case CK_NonAtomicToAtomic:
   case CK_AtomicToNonAtomic:
   case CK_Dynamic:
-  case CK_UncheckedDerivedToBase:
-  case CK_DerivedToBase:
   case CK_ToUnion:
   case CK_BaseToDerived:
   case CK_LValueBitCast:
@@ -864,6 +867,27 @@ LValue CIRGenFunction::emitCastLValue(const CastExpr *e) {
     return lv;
   }
 
+  case CK_UncheckedDerivedToBase:
+  case CK_DerivedToBase: {
+    const auto *derivedClassTy =
+        e->getSubExpr()->getType()->castAs<clang::RecordType>();
+    auto *derivedClassDecl = cast<CXXRecordDecl>(derivedClassTy->getDecl());
+
+    LValue lv = emitLValue(e->getSubExpr());
+    Address thisAddr = lv.getAddress();
+
+    // Perform the derived-to-base conversion
+    Address baseAddr =
+        getAddressOfBaseClass(thisAddr, derivedClassDecl, e->path(),
+                              /*NullCheckValue=*/false, e->getExprLoc());
+
+    // TODO: Support accesses to members of base classes in TBAA. For now, we
+    // conservatively pretend that the complete object is of the base class
+    // type.
+    assert(!cir::MissingFeatures::opTBAA());
+    return makeAddrLValue(baseAddr, e->getType(), lv.getBaseInfo());
+  }
+
   case CK_ZeroToOCLOpaqueType:
     llvm_unreachable("NULL to OpenCL opaque type lvalue cast is not valid");
   }

clang/lib/CIR/CodeGen/CIRGenFunction.cpp

@@ -16,6 +16,7 @@
 #include "CIRGenCall.h"
 #include "CIRGenValue.h"
 #include "mlir/IR/Location.h"
+#include "clang/AST/ExprCXX.h"
 #include "clang/AST/GlobalDecl.h"
 #include "clang/CIR/MissingFeatures.h"
 
@@ -629,4 +630,25 @@ void CIRGenFunction::emitNullInitialization(mlir::Location loc, Address destPtr,
   builder.createStore(loc, zeroValue, destPtr);
 }
 
+// TODO(cir): should be shared with LLVM codegen.
+bool CIRGenFunction::shouldNullCheckClassCastValue(const CastExpr *ce) {
+  const Expr *e = ce->getSubExpr();
+
+  if (ce->getCastKind() == CK_UncheckedDerivedToBase)
+    return false;
+
+  if (isa<CXXThisExpr>(e->IgnoreParens())) {
+    // We always assume that 'this' is never null.
+    return false;
+  }
+
+  if (const ImplicitCastExpr *ice = dyn_cast<ImplicitCastExpr>(ce)) {
+    // And that glvalue casts are never null.
+    if (ice->isGLValue())
+      return false;
+  }
+
+  return true;
+}
+
 } // namespace clang::CIRGen

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -465,6 +465,8 @@ class CIRGenFunction : public CIRGenTypeCache {
     // TODO: Add symbol table support
   }
 
+  bool shouldNullCheckClassCastValue(const CastExpr *ce);
+
   LValue makeNaturalAlignPointeeAddrLValue(mlir::Value v, clang::QualType t);
 
   /// Construct an address with the natural alignment of T. If a pointer to T
@@ -480,6 +482,11 @@ class CIRGenFunction : public CIRGenTypeCache {
     return Address(ptr, convertTypeForMem(t), alignment);
   }
 
+  Address getAddressOfBaseClass(
+      Address value, const CXXRecordDecl *derived,
+      llvm::iterator_range<CastExpr::path_const_iterator> path,
+      bool nullCheckValue, SourceLocation loc);
+
   LValue makeAddrLValue(Address addr, QualType ty,
                         AlignmentSource source = AlignmentSource::Type) {
     return makeAddrLValue(addr, ty, LValueBaseInfo(source));

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -19,6 +19,7 @@
 #include "clang/AST/DeclBase.h"
 #include "clang/AST/DeclOpenACC.h"
 #include "clang/AST/GlobalDecl.h"
+#include "clang/AST/RecordLayout.h"
 #include "clang/Basic/SourceManager.h"
 #include "clang/CIR/Dialect/IR/CIRDialect.h"
 #include "clang/CIR/Interfaces/CIROpInterfaces.h"
@@ -1683,6 +1684,33 @@ bool CIRGenModule::verifyModule() const {
   return mlir::verify(theModule).succeeded();
 }
 
+// TODO(cir): this can be shared with LLVM codegen.
+CharUnits CIRGenModule::computeNonVirtualBaseClassOffset(
+    const CXXRecordDecl *derivedClass,
+    llvm::iterator_range<CastExpr::path_const_iterator> path) {
+  CharUnits offset = CharUnits::Zero();
+
+  const ASTContext &astContext = getASTContext();
+  const CXXRecordDecl *rd = derivedClass;
+
+  for (const CXXBaseSpecifier *base : path) {
+    assert(!base->isVirtual() && "Should not see virtual bases here!");
+
+    // Get the layout.
+    const ASTRecordLayout &layout = astContext.getASTRecordLayout(rd);
+
+    const auto *baseDecl = cast<CXXRecordDecl>(
+        base->getType()->castAs<clang::RecordType>()->getDecl());
+
+    // Add the offset.
+    offset += layout.getBaseClassOffset(baseDecl);
+
+    rd = baseDecl;
+  }
+
+  return offset;
+}
+
 DiagnosticBuilder CIRGenModule::errorNYI(SourceLocation loc,
                                          llvm::StringRef feature) {
   unsigned diagID = diags.getCustomDiagID(

clang/lib/CIR/CodeGen/CIRGenModule.h

@@ -141,6 +141,10 @@ class CIRGenModule : public CIRGenTypeCache {
   getAddrOfGlobalVar(const VarDecl *d, mlir::Type ty = {},
                      ForDefinition_t isForDefinition = NotForDefinition);
 
+  CharUnits computeNonVirtualBaseClassOffset(
+      const CXXRecordDecl *derivedClass,
+      llvm::iterator_range<CastExpr::path_const_iterator> path);
+
   /// Return a constant array for the given string.
   mlir::Attribute getConstantArrayFromStringLiteral(const StringLiteral *e);
 

clang/lib/CIR/CodeGen/CMakeLists.txt

@@ -10,6 +10,7 @@ add_clang_library(clangCIR
   CIRGenerator.cpp
   CIRGenBuilder.cpp
   CIRGenCall.cpp
+  CIRGenClass.cpp
   CIRGenCXXABI.cpp
   CIRGenCXXExpr.cpp
   CIRGenDecl.cpp