[CIR] initial support for pointer-to-data-member type (llvm#401)

This patch adds initial support for the pointer-to-data-member type.
Specifically, this commit includes:

- New ops, types, and attributes:
- CodeGen for pointer-to-data-member types and values
  - Lower C++ pointer-to-member type
  - Lower C++ expression `&C::D`
  - Lower C++ expression `c.*p` and `c->*p`

This patch only includes an initial support. The following stuff related
to pointer-to-member types are not supported yet:

- Pointer to member function;
- Conversion from `T Base::*` to `T Derived::*`;
- LLVMIR lowering.

Author: Lancern

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

@@ -264,6 +264,40 @@ def ConstPtrAttr : CIR_Attr<"ConstPtr", "ptr", [TypedAttrInterface]> {
   let hasCustomAssemblyFormat = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// DataMemberAttr
+//===----------------------------------------------------------------------===//
+
+def DataMemberAttr : CIR_Attr<"DataMember", "data_member",
+                              [TypedAttrInterface]> {
+  let summary = "Holds a constant data member pointer value";
+  let parameters = (ins AttributeSelfTypeParameter<
+                            "", "mlir::cir::DataMemberType">:$type,
+                        OptionalParameter<
+                            "std::optional<size_t>">:$memberIndex);
+  let description = [{
+    A data member attribute is a literal attribute that represents a constant
+    pointer-to-data-member value.
+
+    The `memberIndex` parameter represents the index of the pointed-to member
+    within its containing struct. It is an optional parameter; lack of this
+    parameter indicates a null pointer-to-data-member value.
+
+    Example:
+    ```
+    #ptr = #cir.data_member<1> : !cir.data_member<!s32i in !ty_22Point22>
+
+    #null = #cir.data_member<null> : !cir.data_member<!s32i in !ty_22Point22>
+    ```
+  }];
+
+  let genVerifyDecl = 1;
+
+  let assemblyFormat = [{
+    `<` ($memberIndex^):(`null`)? `>`
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // SignedOverflowBehaviorAttr
 //===----------------------------------------------------------------------===//

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

@@ -1770,6 +1770,63 @@ def GetMemberOp : CIR_Op<"get_member"> {
   let hasVerifier = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// GetRuntimeMemberOp
+//===----------------------------------------------------------------------===//
+
+def GetRuntimeMemberOp : CIR_Op<"get_runtime_member"> {
+  let summary = "Get the address of a member of a struct";
+  let description = [{
+    The `cir.get_runtime_member` operation gets the address of a member from
+    the input record. The target member is given by a value of type
+    `!cir.data_member` (i.e. a pointer-to-data-member value).
+
+    This operation differs from `cir.get_member` in when the target member can
+    be determined. For the `cir.get_member` operation, the target member is
+    specified as a constant index so the member it returns access to is known
+    when the operation is constructed. For the `cir.get_runtime_member`
+    operation, the target member is given through a pointer-to-data-member
+    value which is unknown until the program being compiled is executed. In
+    other words, `cir.get_member` represents a normal member access through the
+    `.` operator in C/C++:
+
+    ```cpp
+    struct Foo { int x; };
+    Foo f;
+    (void)f.x;  // cir.get_member
+    ```
+
+    And `cir.get_runtime_member` represents a member access through the `.*` or
+    the `->*` operator in C++:
+
+    ```cpp
+    struct Foo { int x; }
+    Foo f;
+    Foo *p;
+    int Foo::*member;
+
+    (void)f.*member;   // cir.get_runtime_member
+    (void)f->*member;  // cir.get_runtime_member
+    ```
+
+    This operation expects a pointer to the base record as well as the pointer
+    to the target member.
+  }];
+
+  let arguments = (ins
+    Arg<StructPtr, "address of the struct object", [MemRead]>:$addr,
+    Arg<CIR_DataMemberType, "pointer to the target member">:$member);
+
+  let results = (outs Res<CIR_PointerType, "">:$result);
+
+  let assemblyFormat = [{
+    $addr `[` $member `:` qualified(type($member)) `]` attr-dict
+    `:` qualified(type($addr)) `->` qualified(type($result))
+  }];
+
+  let hasVerifier = 1;
+}
+
 //===----------------------------------------------------------------------===//
 // VecInsertOp
 //===----------------------------------------------------------------------===//

clang/include/clang/CIR/Dialect/IR/CIRTypes.h

@@ -20,13 +20,6 @@
 
 #include "clang/CIR/Interfaces/ASTAttrInterfaces.h"
 
-//===----------------------------------------------------------------------===//
-// CIR Dialect Tablegen'd Types
-//===----------------------------------------------------------------------===//
-
-#define GET_TYPEDEF_CLASSES
-#include "clang/CIR/Dialect/IR/CIROpsTypes.h.inc"
-
 //===----------------------------------------------------------------------===//
 // CIR StructType
 //
@@ -184,4 +177,11 @@ class StructType
 } // namespace cir
 } // namespace mlir
 
+//===----------------------------------------------------------------------===//
+// CIR Dialect Tablegen'd Types
+//===----------------------------------------------------------------------===//
+
+#define GET_TYPEDEF_CLASSES
+#include "clang/CIR/Dialect/IR/CIROpsTypes.h.inc"
+
 #endif // MLIR_DIALECT_CIR_IR_CIRTYPES_H_

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

@@ -145,6 +145,28 @@ def CIR_PointerType : CIR_Type<"Pointer", "ptr",
   let hasCustomAssemblyFormat = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// DataMemberType
+//===----------------------------------------------------------------------===//
+
+def CIR_DataMemberType : CIR_Type<"DataMember", "data_member",
+    [DeclareTypeInterfaceMethods<DataLayoutTypeInterface>]> {
+
+  let summary = "CIR type that represents pointer-to-data-member type in C++";
+  let description = [{
+    `cir.member_ptr` models the pointer-to-data-member type in C++. Values of
+    this type are essentially offsets of the pointed-to member within one of
+    its containing struct.
+  }];
+
+  let parameters = (ins "mlir::Type":$memberTy,
+                        "mlir::cir::StructType":$clsTy);
+
+  let assemblyFormat = [{
+    `<` $memberTy `in` $clsTy `>`
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // BoolType
 //
@@ -309,6 +331,15 @@ def VoidPtr : Type<
       "mlir::cir::VoidType::get($_builder.getContext()))"> {
 }
 
+// Pointer to struct
+def StructPtr : Type<
+    And<[
+      CPred<"$_self.isa<::mlir::cir::PointerType>()">,
+      CPred<"$_self.cast<::mlir::cir::PointerType>()"
+            ".getPointee().isa<::mlir::cir::StructType>()">,
+    ]>, "!cir.struct*"> {
+}
+
 // Pointers to exception info
 def ExceptionInfoPtr : Type<
     And<[
@@ -351,8 +382,9 @@ def CIR_StructType : Type<CPred<"$_self.isa<::mlir::cir::StructType>()">,
 //===----------------------------------------------------------------------===//
 
 def CIR_AnyType : AnyTypeOf<[
-  CIR_IntType, CIR_PointerType, CIR_BoolType, CIR_ArrayType, CIR_VectorType,
-  CIR_FuncType, CIR_VoidType, CIR_StructType, CIR_ExceptionInfo, CIR_AnyFloat,
+  CIR_IntType, CIR_PointerType, CIR_DataMemberType, CIR_BoolType, CIR_ArrayType,
+  CIR_VectorType, CIR_FuncType, CIR_VoidType, CIR_StructType, CIR_ExceptionInfo,
+  CIR_AnyFloat,
 ]>;
 
 #endif // MLIR_CIR_DIALECT_CIR_TYPES

clang/lib/CIR/CodeGen/CIRGenBuilder.h

@@ -221,6 +221,16 @@ class CIRGenBuilderTy : public CIRBaseBuilderTy {
     return mlir::cir::TypeInfoAttr::get(anonStruct.getType(), fieldsAttr);
   }
 
+  mlir::cir::DataMemberAttr getDataMemberAttr(mlir::cir::DataMemberType ty,
+                                              size_t memberIndex) {
+    return mlir::cir::DataMemberAttr::get(getContext(), ty, memberIndex);
+  }
+
+  mlir::cir::DataMemberAttr
+  getNullDataMemberAttr(mlir::cir::DataMemberType ty) {
+    return mlir::cir::DataMemberAttr::get(getContext(), ty, std::nullopt);
+  }
+
   mlir::TypedAttr getZeroInitAttr(mlir::Type ty) {
     if (ty.isa<mlir::cir::IntType>())
       return mlir::cir::IntAttr::get(ty, 0);
@@ -551,6 +561,12 @@ class CIRGenBuilderTy : public CIRBaseBuilderTy {
     return create<mlir::cir::ConstantOp>(loc, ty, getConstPtrAttr(ty, 0));
   }
 
+  /// Create constant nullptr for pointer-to-data-member type ty.
+  mlir::cir::ConstantOp getNullDataMemberPtr(mlir::cir::DataMemberType ty,
+                                             mlir::Location loc) {
+    return create<mlir::cir::ConstantOp>(loc, ty, getNullDataMemberAttr(ty));
+  }
+
   // Creates constant null value for integral type ty.
   mlir::cir::ConstantOp getNullValue(mlir::Type ty, mlir::Location loc) {
     return create<mlir::cir::ConstantOp>(loc, ty, getZeroInitAttr(ty));
@@ -866,6 +882,19 @@ class CIRGenBuilderTy : public CIRBaseBuilderTy {
     }
   }
 
+  mlir::cir::GetRuntimeMemberOp createGetIndirectMember(mlir::Location loc,
+                                                        mlir::Value objectPtr,
+                                                        mlir::Value memberPtr) {
+    auto memberPtrTy = memberPtr.getType().cast<mlir::cir::DataMemberType>();
+
+    // TODO(cir): consider address space.
+    assert(!UnimplementedFeature::addressSpace());
+    auto resultTy = getPointerTo(memberPtrTy.getMemberTy());
+
+    return create<mlir::cir::GetRuntimeMemberOp>(loc, resultTy, objectPtr,
+                                                 memberPtr);
+  }
+
   mlir::Value createPtrIsNull(mlir::Value ptr) {
     return createNot(createPtrToBoolCast(ptr));
   }

clang/lib/CIR/CodeGen/CIRGenClass.cpp

@@ -1475,3 +1475,61 @@ mlir::Value CIRGenFunction::getVTablePtr(mlir::Location Loc, Address This,
 
   return VTable;
 }
+
+Address CIRGenFunction::buildCXXMemberDataPointerAddress(
+    const Expr *E, Address base, mlir::Value memberPtr,
+    const MemberPointerType *memberPtrType, LValueBaseInfo *baseInfo) {
+  assert(!UnimplementedFeature::cxxABI());
+
+  auto op = builder.createGetIndirectMember(getLoc(E->getSourceRange()),
+                                            base.getPointer(), memberPtr);
+
+  QualType memberType = memberPtrType->getPointeeType();
+  CharUnits memberAlign = CGM.getNaturalTypeAlignment(memberType, baseInfo);
+  memberAlign = CGM.getDynamicOffsetAlignment(
+      base.getAlignment(), memberPtrType->getClass()->getAsCXXRecordDecl(),
+      memberAlign);
+
+  return Address(op, convertTypeForMem(memberPtrType->getPointeeType()),
+                 memberAlign);
+}
+
+clang::CharUnits
+CIRGenModule::getDynamicOffsetAlignment(clang::CharUnits actualBaseAlign,
+                                        const clang::CXXRecordDecl *baseDecl,
+                                        clang::CharUnits expectedTargetAlign) {
+  // If the base is an incomplete type (which is, alas, possible with
+  // member pointers), be pessimistic.
+  if (!baseDecl->isCompleteDefinition())
+    return std::min(actualBaseAlign, expectedTargetAlign);
+
+  auto &baseLayout = getASTContext().getASTRecordLayout(baseDecl);
+  CharUnits expectedBaseAlign = baseLayout.getNonVirtualAlignment();
+
+  // If the class is properly aligned, assume the target offset is, too.
+  //
+  // This actually isn't necessarily the right thing to do --- if the
+  // class is a complete object, but it's only properly aligned for a
+  // base subobject, then the alignments of things relative to it are
+  // probably off as well.  (Note that this requires the alignment of
+  // the target to be greater than the NV alignment of the derived
+  // class.)
+  //
+  // However, our approach to this kind of under-alignment can only
+  // ever be best effort; after all, we're never going to propagate
+  // alignments through variables or parameters.  Note, in particular,
+  // that constructing a polymorphic type in an address that's less
+  // than pointer-aligned will generally trap in the constructor,
+  // unless we someday add some sort of attribute to change the
+  // assumed alignment of 'this'.  So our goal here is pretty much
+  // just to allow the user to explicitly say that a pointer is
+  // under-aligned and then safely access its fields and vtables.
+  if (actualBaseAlign >= expectedBaseAlign) {
+    return expectedTargetAlign;
+  }
+
+  // Otherwise, we might be offset by an arbitrary multiple of the
+  // actual alignment.  The correct adjustment is to take the min of
+  // the two alignments.
+  return std::min(actualBaseAlign, expectedTargetAlign);
+}

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -891,6 +891,30 @@ LValue CIRGenFunction::buildDeclRefLValue(const DeclRefExpr *E) {
   llvm_unreachable("Unhandled DeclRefExpr");
 }
 
+LValue
+CIRGenFunction::buildPointerToDataMemberBinaryExpr(const BinaryOperator *E) {
+  assert((E->getOpcode() == BO_PtrMemD || E->getOpcode() == BO_PtrMemI) &&
+         "unexpected binary operator opcode");
+
+  auto baseAddr = Address::invalid();
+  if (E->getOpcode() == BO_PtrMemD)
+    baseAddr = buildLValue(E->getLHS()).getAddress();
+  else
+    baseAddr = buildPointerWithAlignment(E->getLHS());
+
+  const auto *memberPtrTy = E->getRHS()->getType()->castAs<MemberPointerType>();
+
+  auto memberPtr = buildScalarExpr(E->getRHS());
+
+  LValueBaseInfo baseInfo;
+  // TODO(cir): add TBAA
+  assert(!UnimplementedFeature::tbaa());
+  auto memberAddr = buildCXXMemberDataPointerAddress(E, baseAddr, memberPtr,
+                                                     memberPtrTy, &baseInfo);
+
+  return makeAddrLValue(memberAddr, memberPtrTy->getPointeeType(), baseInfo);
+}
+
 LValue CIRGenFunction::buildBinaryOperatorLValue(const BinaryOperator *E) {
   // Comma expressions just emit their LHS then their RHS as an l-value.
   if (E->getOpcode() == BO_Comma) {
@@ -899,7 +923,7 @@ LValue CIRGenFunction::buildBinaryOperatorLValue(const BinaryOperator *E) {
   }
 
   if (E->getOpcode() == BO_PtrMemD || E->getOpcode() == BO_PtrMemI)
-    assert(0 && "not implemented");
+    return buildPointerToDataMemberBinaryExpr(E);
 
   assert(E->getOpcode() == BO_Assign && "unexpected binary l-value");
 

clang/lib/CIR/CodeGen/CIRGenExprConst.cpp

@@ -1761,6 +1761,21 @@ mlir::Attribute ConstantEmitter::tryEmitPrivate(const APValue &Value,
     return buildArrayConstant(CGM, Desired, CommonElementType, NumElements,
                               Elts, typedFiller);
   }
+  case APValue::MemberPointer: {
+    assert(!UnimplementedFeature::cxxABI());
+
+    const ValueDecl *memberDecl = Value.getMemberPointerDecl();
+    assert(!Value.isMemberPointerToDerivedMember() && "NYI");
+
+    if (const auto *memberFuncDecl = dyn_cast<CXXMethodDecl>(memberDecl))
+      assert(0 && "not implemented");
+
+    auto cirTy =
+        CGM.getTypes().ConvertType(DestType).cast<mlir::cir::DataMemberType>();
+
+    const auto *fieldDecl = cast<FieldDecl>(memberDecl);
+    return builder.getDataMemberAttr(cirTy, fieldDecl->getFieldIndex());
+  }
   case APValue::LValue:
     return ConstantLValueEmitter(*this, Value, DestType).tryEmit();
   case APValue::Struct:
@@ -1771,7 +1786,6 @@ mlir::Attribute ConstantEmitter::tryEmitPrivate(const APValue &Value,
   case APValue::ComplexFloat:
   case APValue::Vector:
   case APValue::AddrLabelDiff:
-  case APValue::MemberPointer:
     assert(0 && "not implemented");
   }
   llvm_unreachable("Unknown APValue kind");
@@ -1800,6 +1814,26 @@ mlir::Value CIRGenModule::buildNullConstant(QualType T, mlir::Location loc) {
   return {};
 }
 
+mlir::Value CIRGenModule::buildMemberPointerConstant(const UnaryOperator *E) {
+  assert(!UnimplementedFeature::cxxABI());
+
+  auto loc = getLoc(E->getSourceRange());
+
+  const auto *decl = cast<DeclRefExpr>(E->getSubExpr())->getDecl();
+
+  // A member function pointer.
+  // Member function pointer is not supported yet.
+  if (const auto *methodDecl = dyn_cast<CXXMethodDecl>(decl))
+    assert(0 && "not implemented");
+
+  auto ty = getCIRType(E->getType()).cast<mlir::cir::DataMemberType>();
+
+  // Otherwise, a member data pointer.
+  const auto *fieldDecl = cast<FieldDecl>(decl);
+  return builder.create<mlir::cir::ConstantOp>(
+      loc, ty, builder.getDataMemberAttr(ty, fieldDecl->getFieldIndex()));
+}
+
 mlir::Attribute ConstantEmitter::emitAbstract(const Expr *E,
                                               QualType destType) {
   auto state = pushAbstract();