[CIR] Upstream support for accessing structure members

andykaylor · andykaylor · commit 17800eddb70a · 2025-04-18T15:37:17.000-07:00
This adds ClangIR support for accessing structure members. Access to
union members is deferred to a later change.
diff --git a/clang/include/clang/CIR/Dialect/Builder/CIRBaseBuilder.h b/clang/include/clang/CIR/Dialect/Builder/CIRBaseBuilder.h
@@ -191,6 +191,12 @@ class CIRBaseBuilderTy : public mlir::OpBuilder {
     return create<cir::StoreOp>(loc, val, dst);
   }
 
+  cir::GetMemberOp createGetMember(mlir::Location loc, mlir::Type resultTy,
+                                   mlir::Value base, llvm::StringRef name,
+                                   unsigned index) {
+    return create<cir::GetMemberOp>(loc, resultTy, base, name, index);
+  }
+
   mlir::Value createDummyValue(mlir::Location loc, mlir::Type type,
                                clang::CharUnits alignment) {
     auto addr = createAlloca(loc, getPointerTo(type), type, {},
diff --git a/clang/include/clang/CIR/Dialect/IR/CIROps.td b/clang/include/clang/CIR/Dialect/IR/CIROps.td
@@ -1310,6 +1310,70 @@ def GetGlobalOp : CIR_Op<"get_global",
   let hasVerifier = 0;
 }
 
+//===----------------------------------------------------------------------===//
+// GetMemberOp
+//===----------------------------------------------------------------------===//
+
+def GetMemberOp : CIR_Op<"get_member"> {
+  let summary = "Get the address of a member of a record";
+  let description = [{
+    The `cir.get_member` operation gets the address of a particular named
+    member from the input record.
+
+    It expects a pointer to the base record as well as the name of the member
+    and its field index.
+
+    Example:
+    ```mlir
+    // Suppose we have a record with multiple members.
+    !s32i = !cir.int<s, 32>
+    !s8i = !cir.int<s, 32>
+    !ty_B = !cir.record<"struct.B" {!s32i, !s8i}>
+
+    // Get the address of the member at index 1.
+    %1 = cir.get_member %0[1] {name = "i"} : (!cir.ptr<!ty_B>) -> !cir.ptr<!s8i>
+    ```
+  }];
+
+  let arguments = (ins
+    Arg<CIR_PointerType, "the address to load from", [MemRead]>:$addr,
+    StrAttr:$name,
+    IndexAttr:$index_attr);
+
+  let results = (outs Res<CIR_PointerType, "">:$result);
+
+  let assemblyFormat = [{
+    $addr `[` $index_attr `]` attr-dict
+    `:` qualified(type($addr)) `->` qualified(type($result))
+  }];
+
+  let builders = [
+    OpBuilder<(ins "mlir::Type":$type,
+                   "mlir::Value":$value,
+                   "llvm::StringRef":$name,
+                   "unsigned":$index),
+    [{
+      mlir::APInt fieldIdx(64, index);
+      build($_builder, $_state, type, value, name, fieldIdx);
+    }]>
+  ];
+
+  let extraClassDeclaration = [{
+    /// Return the index of the record member being accessed.
+    uint64_t getIndex() { return getIndexAttr().getZExtValue(); }
+
+    /// Return the record type pointed by the base pointer.
+    cir::PointerType getAddrTy() { return getAddr().getType(); }
+
+    /// Return the result type.
+    cir::PointerType getResultTy() {
+      return mlir::cast<cir::PointerType>(getResult().getType());
+    }
+  }];
+
+  let hasVerifier = 1;
+}
+
 //===----------------------------------------------------------------------===//
 // FuncOp
 //===----------------------------------------------------------------------===//
diff --git a/clang/include/clang/CIR/Dialect/IR/CIRTypes.td b/clang/include/clang/CIR/Dialect/IR/CIRTypes.td
@@ -503,6 +503,9 @@ def CIR_RecordType : CIR_Type<"Record", "record",
     void complete(llvm::ArrayRef<mlir::Type> members, bool packed,
                   bool isPadded);
 
+    uint64_t getElementOffset(const mlir::DataLayout &dataLayout,
+              unsigned idx) const;
+
   private:
     unsigned computeStructSize(const mlir::DataLayout &dataLayout) const;
     uint64_t computeStructAlignment(const mlir::DataLayout &dataLayout) const;
diff --git a/clang/include/clang/CIR/MissingFeatures.h b/clang/include/clang/CIR/MissingFeatures.h
@@ -157,6 +157,8 @@ struct MissingFeatures {
   static bool emitCheckedInBoundsGEP() { return false; }
   static bool preservedAccessIndexRegion() { return false; }
   static bool bitfields() { return false; }
+  static bool typeChecks() { return false; }
+  static bool lambdaFieldToName() { return false; }
 
   // Missing types
   static bool dataMemberType() { return false; }
diff --git a/clang/lib/CIR/CodeGen/CIRGenExpr.cpp b/clang/lib/CIR/CodeGen/CIRGenExpr.cpp
@@ -27,6 +27,38 @@ using namespace clang;
 using namespace clang::CIRGen;
 using namespace cir;
 
+/// Get the address of a zero-sized field within a record. The resulting address
+/// doesn't necessarily have the right type.
+Address CIRGenFunction::emitAddrOfFieldStorage(Address base,
+                                               const FieldDecl *field,
+                                               llvm::StringRef fieldName,
+                                               unsigned fieldIndex) {
+  if (field->isZeroSize(getContext())) {
+    cgm.errorNYI(field->getSourceRange(),
+                 "emitAddrOfFieldStorage: zero-sized field");
+    return Address::invalid();
+  }
+
+  mlir::Location loc = getLoc(field->getLocation());
+
+  mlir::Type fieldType = convertType(field->getType());
+  auto fieldPtr = cir::PointerType::get(builder.getContext(), fieldType);
+  // For most cases fieldName is the same as field->getName() but for lambdas,
+  // which do not currently carry the name, so it can be passed down from the
+  // CaptureStmt.
+  cir::GetMemberOp memberAddr = builder.createGetMember(
+      loc, fieldPtr, base.getPointer(), fieldName, fieldIndex);
+
+  // Retrieve layout information, compute alignment and return the final
+  // address.
+  const RecordDecl *rec = field->getParent();
+  const CIRGenRecordLayout &layout = cgm.getTypes().getCIRGenRecordLayout(rec);
+  unsigned idx = layout.getCIRFieldNo(field);
+  CharUnits offset = CharUnits::fromQuantity(
+      layout.getCIRType().getElementOffset(cgm.getDataLayout().layout, idx));
+  return Address(memberAddr, base.getAlignment().alignmentAtOffset(offset));
+}
+
 /// Given an expression of pointer type, try to
 /// derive a more accurate bound on the alignment of the pointer.
 Address CIRGenFunction::emitPointerWithAlignment(const Expr *expr,
@@ -264,6 +296,66 @@ mlir::Value CIRGenFunction::emitStoreThroughBitfieldLValue(RValue src,
   return {};
 }
 
+LValue CIRGenFunction::emitLValueForField(LValue base, const FieldDecl *field) {
+  LValueBaseInfo baseInfo = base.getBaseInfo();
+
+  if (field->isBitField()) {
+    cgm.errorNYI(field->getSourceRange(), "emitLValueForField: bitfield");
+    return LValue();
+  }
+
+  QualType fieldType = field->getType();
+  const RecordDecl *rec = field->getParent();
+  AlignmentSource baseAlignSource = baseInfo.getAlignmentSource();
+  LValueBaseInfo fieldBaseInfo(getFieldAlignmentSource(baseAlignSource));
+  assert(!cir::MissingFeatures::opTBAA());
+
+  Address addr = base.getAddress();
+  if (auto *classDef = dyn_cast<CXXRecordDecl>(rec)) {
+    cgm.errorNYI(field->getSourceRange(), "emitLValueForField: C++ class");
+    return LValue();
+  }
+
+  unsigned recordCVR = base.getVRQualifiers();
+  if (rec->isUnion()) {
+    cgm.errorNYI(field->getSourceRange(), "emitLValueForField: union");
+    return LValue();
+  }
+
+  assert(!cir::MissingFeatures::preservedAccessIndexRegion());
+  llvm::StringRef fieldName = field->getName();
+  const CIRGenRecordLayout &layout =
+      cgm.getTypes().getCIRGenRecordLayout(field->getParent());
+  unsigned fieldIndex = layout.getCIRFieldNo(field);
+
+  assert(!cir::MissingFeatures::lambdaFieldToName());
+
+  addr = emitAddrOfFieldStorage(addr, field, fieldName, fieldIndex);
+
+  // If this is a reference field, load the reference right now.
+  if (fieldType->isReferenceType()) {
+    cgm.errorNYI(field->getSourceRange(), "emitLValueForField: reference type");
+    return LValue();
+  }
+
+  if (field->hasAttr<AnnotateAttr>()) {
+    cgm.errorNYI(field->getSourceRange(), "emitLValueForField: AnnotateAttr");
+    return LValue();
+  }
+
+  LValue lv = makeAddrLValue(addr, fieldType, fieldBaseInfo);
+  lv.getQuals().addCVRQualifiers(recordCVR);
+
+  // __weak attribute on a field is ignored.
+  if (lv.getQuals().getObjCGCAttr() == Qualifiers::Weak) {
+    cgm.errorNYI(field->getSourceRange(),
+                 "emitLValueForField: __weak attribute");
+    return LValue();
+  }
+
+  return lv;
+}
+
 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.
@@ -608,6 +700,48 @@ CIRGenFunction::emitArraySubscriptExpr(const clang::ArraySubscriptExpr *e) {
   return lv;
 }
 
+LValue CIRGenFunction::emitMemberExpr(const MemberExpr *e) {
+  if (auto *vd = dyn_cast<VarDecl>(e->getMemberDecl())) {
+    cgm.errorNYI(e->getSourceRange(), "emitMemberExpr: VarDecl");
+    return LValue();
+  }
+
+  Expr *baseExpr = e->getBase();
+  // If this is s.x, emit s as an lvalue.  If it is s->x, emit s as a scalar.
+  LValue baseLV;
+  if (e->isArrow()) {
+    LValueBaseInfo baseInfo;
+    assert(!cir::MissingFeatures::opTBAA());
+    Address addr = emitPointerWithAlignment(baseExpr, &baseInfo);
+    QualType ptrTy = baseExpr->getType()->getPointeeType();
+    assert(!cir::MissingFeatures::typeChecks());
+    baseLV = makeAddrLValue(addr, ptrTy, baseInfo);
+  } else {
+    assert(!cir::MissingFeatures::typeChecks());
+    baseLV = emitLValue(baseExpr);
+  }
+
+  const NamedDecl *nd = e->getMemberDecl();
+  if (auto *field = dyn_cast<FieldDecl>(nd)) {
+    LValue lv = emitLValueForField(baseLV, field);
+    assert(!cir::MissingFeatures::setObjCGCLValueClass());
+    if (getLangOpts().OpenMP) {
+      // If the member was explicitly marked as nontemporal, mark it as
+      // nontemporal. If the base lvalue is marked as nontemporal, mark access
+      // to children as nontemporal too.
+      cgm.errorNYI(e->getSourceRange(), "emitMemberExpr: OpenMP");
+    }
+    return lv;
+  }
+
+  if (const auto *fd = dyn_cast<FunctionDecl>(nd)) {
+    cgm.errorNYI(e->getSourceRange(), "emitMemberExpr: FunctionDecl");
+    return LValue();
+  }
+
+  llvm_unreachable("Unhandled member declaration!");
+}
+
 LValue CIRGenFunction::emitBinaryOperatorLValue(const BinaryOperator *e) {
   // Comma expressions just emit their LHS then their RHS as an l-value.
   if (e->getOpcode() == BO_Comma) {
diff --git a/clang/lib/CIR/CodeGen/CIRGenExprScalar.cpp b/clang/lib/CIR/CodeGen/CIRGenExprScalar.cpp
@@ -171,6 +171,8 @@ class ScalarExprEmitter : public StmtVisitor<ScalarExprEmitter, mlir::Value> {
     return emitLoadOfLValue(e);
   }
 
+  mlir::Value VisitMemberExpr(MemberExpr *e);
+
   mlir::Value VisitExplicitCastExpr(ExplicitCastExpr *e) {
     return VisitCastExpr(e);
   }
@@ -1529,6 +1531,19 @@ mlir::Value ScalarExprEmitter::VisitCallExpr(const CallExpr *e) {
   return v;
 }
 
+mlir::Value ScalarExprEmitter::VisitMemberExpr(MemberExpr *e) {
+  // TODO(cir): The classic codegen calls tryEmitAsConstant() here. Folding
+  // constants sound like work for MLIR optimizers, but we'll keep an assertion
+  // for now.
+  assert(!cir::MissingFeatures::tryEmitAsConstant());
+  Expr::EvalResult result;
+  if (e->EvaluateAsInt(result, cgf.getContext(), Expr::SE_AllowSideEffects)) {
+    cgf.cgm.errorNYI(e->getSourceRange(), "Constant interger member expr");
+    // Fall through to emit this as a non-constant access.
+  }
+  return emitLoadOfLValue(e);
+}
+
 mlir::Value CIRGenFunction::emitScalarConversion(mlir::Value src,
                                                  QualType srcTy, QualType dstTy,
                                                  SourceLocation loc) {
diff --git a/clang/lib/CIR/CodeGen/CIRGenFunction.cpp b/clang/lib/CIR/CodeGen/CIRGenFunction.cpp
@@ -513,6 +513,8 @@ LValue CIRGenFunction::emitLValue(const Expr *e) {
     return emitArraySubscriptExpr(cast<ArraySubscriptExpr>(e));
   case Expr::UnaryOperatorClass:
     return emitUnaryOpLValue(cast<UnaryOperator>(e));
+  case Expr::MemberExprClass:
+    return emitMemberExpr(cast<MemberExpr>(e));
   case Expr::BinaryOperatorClass:
     return emitBinaryOperatorLValue(cast<BinaryOperator>(e));
   case Expr::ParenExprClass:
diff --git a/clang/lib/CIR/CodeGen/CIRGenFunction.h b/clang/lib/CIR/CodeGen/CIRGenFunction.h
@@ -423,6 +423,10 @@ class CIRGenFunction : public CIRGenTypeCache {
                               clang::CharUnits alignment);
 
 public:
+  Address emitAddrOfFieldStorage(Address base, const FieldDecl *field,
+                                 llvm::StringRef fieldName,
+                                 unsigned fieldIndex);
+
   mlir::Value emitAlloca(llvm::StringRef name, mlir::Type ty,
                          mlir::Location loc, clang::CharUnits alignment,
                          bool insertIntoFnEntryBlock,
@@ -551,6 +555,9 @@ class CIRGenFunction : public CIRGenTypeCache {
   /// of the expression.
   /// FIXME: document this function better.
   LValue emitLValue(const clang::Expr *e);
+  LValue emitLValueForField(LValue base, const clang::FieldDecl *field);
+
+  LValue emitMemberExpr(const MemberExpr *e);
 
   /// Given an expression with a pointer type, emit the value and compute our
   /// best estimate of the alignment of the pointee.
diff --git a/clang/lib/CIR/CodeGen/CIRGenModule.h b/clang/lib/CIR/CodeGen/CIRGenModule.h
@@ -19,6 +19,7 @@
 #include "CIRGenValue.h"
 
 #include "clang/AST/CharUnits.h"
+#include "clang/CIR/Dialect/IR/CIRDataLayout.h"
 #include "clang/CIR/Dialect/IR/CIRDialect.h"
 
 #include "TargetInfo.h"
@@ -95,6 +96,12 @@ class CIRGenModule : public CIRGenTypeCache {
   const clang::LangOptions &getLangOpts() const { return langOpts; }
   mlir::MLIRContext &getMLIRContext() { return *builder.getContext(); }
 
+  const cir::CIRDataLayout getDataLayout() const {
+    // FIXME(cir): instead of creating a CIRDataLayout every time, set it as an
+    // attribute for the CIRModule class.
+    return cir::CIRDataLayout(theModule);
+  }
+
   /// -------
   /// Handling globals
   /// -------
diff --git a/clang/lib/CIR/CodeGen/CIRGenRecordLayout.h b/clang/lib/CIR/CodeGen/CIRGenRecordLayout.h
@@ -42,10 +42,10 @@ class CIRGenRecordLayout {
   cir::RecordType getCIRType() const { return completeObjectType; }
 
   /// Return cir::RecordType element number that corresponds to the field FD.
-  unsigned getCIRFieldNo(const clang::FieldDecl *FD) const {
-    FD = FD->getCanonicalDecl();
-    assert(fieldInfo.count(FD) && "Invalid field for record!");
-    return fieldInfo.lookup(FD);
+  unsigned getCIRFieldNo(const clang::FieldDecl *fd) const {
+    fd = fd->getCanonicalDecl();
+    assert(fieldInfo.count(fd) && "Invalid field for record!");
+    return fieldInfo.lookup(fd);
   }
 };
 
diff --git a/clang/lib/CIR/CodeGen/CIRGenTypes.cpp b/clang/lib/CIR/CodeGen/CIRGenTypes.cpp
@@ -446,6 +446,27 @@ mlir::Type CIRGenTypes::convertTypeForMem(clang::QualType qualType,
   return convertedType;
 }
 
+/// Return record layout info for the given record decl.
+const CIRGenRecordLayout &
+CIRGenTypes::getCIRGenRecordLayout(const RecordDecl *rd) {
+  const auto *key = astContext.getTagDeclType(rd).getTypePtr();
+
+  // If we have already computed the layout, return it.
+  auto it = cirGenRecordLayouts.find(key);
+  if (it != cirGenRecordLayouts.end())
+    return *it->second;
+
+  // Compute the type information.
+  convertRecordDeclType(rd);
+
+  // Now try again.
+  it = cirGenRecordLayouts.find(key);
+
+  assert(it != cirGenRecordLayouts.end() &&
+         "Unable to find record layout information for type");
+  return *it->second;
+}
+
 bool CIRGenTypes::isZeroInitializable(clang::QualType t) {
   if (t->getAs<PointerType>())
     return astContext.getTargetNullPointerValue(t) == 0;
diff --git a/clang/lib/CIR/CodeGen/CIRGenTypes.h b/clang/lib/CIR/CodeGen/CIRGenTypes.h
@@ -108,6 +108,8 @@ class CIRGenTypes {
   std::string getRecordTypeName(const clang::RecordDecl *,
                                 llvm::StringRef suffix);
 
+  const CIRGenRecordLayout &getCIRGenRecordLayout(const clang::RecordDecl *rd);
+
   /// Convert type T into an mlir::Type. This differs from convertType in that
   /// it is used to convert to the memory representation for a type. For
   /// example, the scalar representation for bool is i1, but the memory
diff --git a/clang/lib/CIR/CodeGen/CIRGenValue.h b/clang/lib/CIR/CodeGen/CIRGenValue.h
diff --git a/clang/lib/CIR/Dialect/IR/CIRDialect.cpp b/clang/lib/CIR/Dialect/IR/CIRDialect.cpp
diff --git a/clang/lib/CIR/Dialect/IR/CIRTypes.cpp b/clang/lib/CIR/Dialect/IR/CIRTypes.cpp
diff --git a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
diff --git a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.h b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.h
diff --git a/clang/test/CIR/CodeGen/struct.c b/clang/test/CIR/CodeGen/struct.c
