[CIR] Function calls with aggregate arguments and return values

This patch updates cir.call operation and allows function calls with aggregate
arguments and return values.

It seems that C++ class support is still at a minimum now. I try to make a call
to a C++ function with an argument of aggregate type but it failed because
the initialization of C++ class / struct is NYI. Thus, tests for calling
functions with aggregate arguments are added only for C for now.

Author: Lancern

clang/include/clang/CIR/MissingFeatures.h

@@ -171,6 +171,10 @@ struct MissingFeatures {
   static bool astVarDeclInterface() { return false; }
   static bool stackSaveOp() { return false; }
   static bool aggValueSlot() { return false; }
+  static bool aggValueSlotVolatile() { return false; }
+  static bool aggValueSlotDestructedFlag() { return false; }
+  static bool aggValueSlotAlias() { return false; }
+  static bool aggValueSlotGC() { return false; }
   static bool generateDebugInfo() { return false; }
   static bool pointerOverflowSanitizer() { return false; }
   static bool fpConstraints() { return false; }
@@ -226,6 +230,7 @@ struct MissingFeatures {
   static bool implicitConstructorArgs() { return false; }
   static bool intrinsics() { return false; }
   static bool attributeNoBuiltin() { return false; }
+  static bool lowerAggregateLoadStore() { return false; }
 
   // Missing types
   static bool dataMemberType() { return false; }

clang/lib/CIR/CodeGen/CIRGenBuilder.h

@@ -332,6 +332,18 @@ class CIRGenBuilderTy : public cir::CIRBaseBuilderTy {
     return Address(baseAddr, destType, addr.getAlignment());
   }
 
+  /// Cast the element type of the given address to a different type,
+  /// preserving information like the alignment.
+  Address createElementBitCast(mlir::Location loc, Address addr,
+                               mlir::Type destType) {
+    if (destType == addr.getElementType())
+      return addr;
+
+    auto ptrTy = getPointerTo(destType);
+    return Address(createBitcast(loc, addr.getPointer(), ptrTy), destType,
+                   addr.getAlignment());
+  }
+
   cir::LoadOp createLoad(mlir::Location loc, Address addr,
                          bool isVolatile = false) {
     mlir::IntegerAttr align = getAlignmentAttr(addr.getAlignment());

clang/lib/CIR/CodeGen/CIRGenCall.cpp

@@ -60,6 +60,23 @@ CIRGenCallee CIRGenCallee::prepareConcreteCallee(CIRGenFunction &cgf) const {
   return *this;
 }
 
+void CIRGenFunction::emitAggregateStore(mlir::Value value, Address dest) {
+  // In classic codegen:
+  // Function to store a first-class aggregate into memory. We prefer to
+  // store the elements rather than the aggregate to be more friendly to
+  // fast-isel.
+  // In CIR codegen:
+  // Emit the most simple cir.store possible (e.g. a store for a whole
+  // record), which can later be broken down in other CIR levels (or prior
+  // to dialect codegen).
+
+  // Stored result for the callers of this function expected to be in the same
+  // scope as the value, don't make assumptions about current insertion point.
+  mlir::OpBuilder::InsertionGuard guard(builder);
+  builder.setInsertionPointAfter(value.getDefiningOp());
+  builder.createStore(*currSrcLoc, value, dest);
+}
+
 /// Returns the canonical formal type of the given C++ method.
 static CanQual<FunctionProtoType> getFormalType(const CXXMethodDecl *md) {
   return md->getType()
@@ -399,8 +416,48 @@ RValue CIRGenFunction::emitCall(const CIRGenFunctionInfo &funcInfo,
       assert(!cir::MissingFeatures::opCallBitcastArg());
       cirCallArgs[argNo] = v;
     } else {
-      assert(!cir::MissingFeatures::opCallAggregateArgs());
-      cgm.errorNYI("emitCall: aggregate function call argument");
+      Address src = Address::invalid();
+      if (!arg.isAggregate())
+        cgm.errorNYI(loc, "emitCall: non-aggregate call argument");
+      else
+        src = arg.hasLValue() ? arg.getKnownLValue().getAddress()
+                              : arg.getKnownRValue().getAggregateAddress();
+
+      // Fast-isel and the optimizer generally like scalar values better than
+      // FCAs, so we flatten them if this is safe to do for this argument.
+      auto argRecordTy = cast<cir::RecordType>(argType);
+      mlir::Type srcTy = src.getElementType();
+      // FIXME(cir): get proper location for each argument.
+      mlir::Location argLoc = loc;
+
+      // If the source type is smaller than the destination type of the
+      // coerce-to logic, copy the source value into a temp alloca the size
+      // of the destination type to allow loading all of it. The bits past
+      // the source value are left undef.
+      // FIXME(cir): add data layout info and compare sizes instead of
+      // matching the types.
+      //
+      // uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
+      // uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
+      // if (SrcSize < DstSize) {
+      if (srcTy != argRecordTy) {
+        cgm.errorNYI(loc, "emitCall: source type does not match argument type");
+      } else {
+        // FIXME(cir): this currently only runs when the types are different,
+        // but should be when alloc sizes are different, fix this as soon as
+        // datalayout gets introduced.
+        src = builder.createElementBitCast(argLoc, src, argRecordTy);
+      }
+
+      // assert(NumCIRArgs == STy.getMembers().size());
+      // In LLVMGen: Still only pass the struct without any gaps but mark it
+      // as such somehow.
+      //
+      // In CIRGen: Emit a load from the "whole" struct,
+      // which shall be broken later by some lowering step into multiple
+      // loads.
+      assert(!cir::MissingFeatures::lowerAggregateLoadStore());
+      cirCallArgs[argNo] = builder.createLoad(argLoc, src);
     }
   }
 
@@ -439,6 +496,7 @@ RValue CIRGenFunction::emitCall(const CIRGenFunctionInfo &funcInfo,
 
   assert(!cir::MissingFeatures::opCallAttrs());
 
+  mlir::Location callLoc = loc;
   cir::CIRCallOpInterface theCall = emitCallLikeOp(
       *this, loc, indirectFuncTy, indirectFuncVal, directFuncOp, cirCallArgs);
 
@@ -452,6 +510,19 @@ RValue CIRGenFunction::emitCall(const CIRGenFunctionInfo &funcInfo,
   if (isa<cir::VoidType>(retCIRTy))
     return getUndefRValue(retTy);
   switch (getEvaluationKind(retTy)) {
+  case cir::TEK_Aggregate: {
+    Address destPtr = returnValue.getValue();
+
+    if (!destPtr.isValid())
+      destPtr = createMemTemp(retTy, callLoc, getCounterAggTmpAsString());
+
+    mlir::ResultRange results = theCall->getOpResults();
+    assert(results.size() <= 1 && "multiple returns from a call");
+
+    SourceLocRAIIObject loc{*this, callLoc};
+    emitAggregateStore(results[0], destPtr);
+    return RValue::getAggregate(destPtr);
+  }
   case cir::TEK_Scalar: {
     mlir::ResultRange results = theCall->getOpResults();
     assert(results.size() == 1 && "unexpected number of returns");
@@ -468,7 +539,6 @@ RValue CIRGenFunction::emitCall(const CIRGenFunctionInfo &funcInfo,
     return RValue::get(results[0]);
   }
   case cir::TEK_Complex:
-  case cir::TEK_Aggregate:
     cgm.errorNYI(loc, "unsupported evaluation kind of function call result");
     return getUndefRValue(retTy);
   }
@@ -487,10 +557,21 @@ void CIRGenFunction::emitCallArg(CallArgList &args, const clang::Expr *e,
 
   bool hasAggregateEvalKind = hasAggregateEvaluationKind(argType);
 
-  if (hasAggregateEvalKind) {
-    assert(!cir::MissingFeatures::opCallAggregateArgs());
-    cgm.errorNYI(e->getSourceRange(),
-                 "emitCallArg: aggregate function call argument");
+  // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
+  // However, we still have to push an EH-only cleanup in case we unwind before
+  // we make it to the call.
+  if (argType->isRecordType() &&
+      argType->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee()) {
+    assert(!cir::MissingFeatures::msabi());
+    cgm.errorNYI(e->getSourceRange(), "emitCallArg: msabi is NYI");
+  }
+
+  if (hasAggregateEvalKind && isa<ImplicitCastExpr>(e) &&
+      cast<CastExpr>(e)->getCastKind() == CK_LValueToRValue) {
+    LValue lv = emitLValue(cast<CastExpr>(e)->getSubExpr());
+    assert(lv.isSimple());
+    args.addUncopiedAggregate(lv, argType);
+    return;
   }
 
   args.add(emitAnyExprToTemp(e), argType);
@@ -511,12 +592,13 @@ QualType CIRGenFunction::getVarArgType(const Expr *arg) {
 /// Similar to emitAnyExpr(), however, the result will always be accessible
 /// even if no aggregate location is provided.
 RValue CIRGenFunction::emitAnyExprToTemp(const Expr *e) {
-  assert(!cir::MissingFeatures::opCallAggregateArgs());
+  AggValueSlot aggSlot = AggValueSlot::ignored();
 
   if (hasAggregateEvaluationKind(e->getType()))
-    cgm.errorNYI(e->getSourceRange(), "emit aggregate value to temp");
+    aggSlot = createAggTemp(e->getType(), getLoc(e->getSourceRange()),
+                            getCounterAggTmpAsString());
 
-  return emitAnyExpr(e);
+  return emitAnyExpr(e, aggSlot);
 }
 
 void CIRGenFunction::emitCallArgs(

clang/lib/CIR/CodeGen/CIRGenCall.h

@@ -133,8 +133,16 @@ struct CallArg {
   CallArg(RValue rv, clang::QualType ty)
       : rv(rv), hasLV(false), isUsed(false), ty(ty) {}
 
+  CallArg(LValue lv, clang::QualType ty)
+      : lv(lv), hasLV(true), isUsed(false), ty(ty) {}
+
   bool hasLValue() const { return hasLV; }
 
+  LValue getKnownLValue() const {
+    assert(hasLV && !isUsed);
+    return lv;
+  }
+
   RValue getKnownRValue() const {
     assert(!hasLV && !isUsed);
     return rv;
@@ -147,6 +155,10 @@ class CallArgList : public llvm::SmallVector<CallArg, 8> {
 public:
   void add(RValue rvalue, clang::QualType type) { emplace_back(rvalue, type); }
 
+  void addUncopiedAggregate(LValue lvalue, clang::QualType type) {
+    emplace_back(lvalue, type);
+  }
+
   /// Add all the arguments from another CallArgList to this one. After doing
   /// this, the old CallArgList retains its list of arguments, but must not
   /// be used to emit a call.
@@ -162,7 +174,15 @@ class CallArgList : public llvm::SmallVector<CallArg, 8> {
 
 /// Contains the address where the return value of a function can be stored, and
 /// whether the address is volatile or not.
-class ReturnValueSlot {};
+class ReturnValueSlot {
+  Address addr = Address::invalid();
+
+public:
+  ReturnValueSlot() = default;
+  ReturnValueSlot(Address addr) : addr(addr) {}
+
+  Address getValue() const { return addr; }
+};
 
 } // namespace clang::CIRGen
 

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -1010,16 +1010,20 @@ LValue CIRGenFunction::emitBinaryOperatorLValue(const BinaryOperator *e) {
 
 /// Emit code to compute the specified expression which
 /// can have any type.  The result is returned as an RValue struct.
-RValue CIRGenFunction::emitAnyExpr(const Expr *e) {
+RValue CIRGenFunction::emitAnyExpr(const Expr *e, AggValueSlot aggSlot) {
   switch (CIRGenFunction::getEvaluationKind(e->getType())) {
   case cir::TEK_Scalar:
     return RValue::get(emitScalarExpr(e));
   case cir::TEK_Complex:
     cgm.errorNYI(e->getSourceRange(), "emitAnyExpr: complex type");
     return RValue::get(nullptr);
-  case cir::TEK_Aggregate:
-    cgm.errorNYI(e->getSourceRange(), "emitAnyExpr: aggregate type");
-    return RValue::get(nullptr);
+  case cir::TEK_Aggregate: {
+    if (aggSlot.isIgnored())
+      aggSlot = createAggTemp(e->getType(), getLoc(e->getSourceRange()),
+                              getCounterAggTmpAsString());
+    emitAggExpr(e, aggSlot);
+    return aggSlot.asRValue();
+  }
   }
   llvm_unreachable("bad evaluation kind");
 }

clang/lib/CIR/CodeGen/CIRGenExprAggregate.cpp

@@ -28,6 +28,15 @@ class AggExprEmitter : public StmtVisitor<AggExprEmitter> {
   CIRGenFunction &cgf;
   AggValueSlot dest;
 
+  // Calls `fn` with a valid return value slot, potentially creating a temporary
+  // to do so. If a temporary is created, an appropriate copy into `Dest` will
+  // be emitted, as will lifetime markers.
+  //
+  // The given function should take a ReturnValueSlot, and return an RValue that
+  // points to said slot.
+  void withReturnValueSlot(const Expr *e,
+                           llvm::function_ref<RValue(ReturnValueSlot)> fn);
+
   AggValueSlot ensureSlot(mlir::Location loc, QualType t) {
     if (!dest.isIgnored())
       return dest;
@@ -40,16 +49,28 @@ class AggExprEmitter : public StmtVisitor<AggExprEmitter> {
   AggExprEmitter(CIRGenFunction &cgf, AggValueSlot dest)
       : cgf(cgf), dest(dest) {}
 
+  /// Given an expression with aggregate type that represents a value lvalue,
+  /// this method emits the address of the lvalue, then loads the result into
+  /// DestPtr.
+  void emitAggLoadOfLValue(const Expr *e);
+
   void emitArrayInit(Address destPtr, cir::ArrayType arrayTy, QualType arrayQTy,
                      Expr *exprToVisit, ArrayRef<Expr *> args,
                      Expr *arrayFiller);
 
+  /// Perform the final copy to DestPtr, if desired.
+  void emitFinalDestCopy(QualType type, const LValue &src);
+
   void emitInitializationToLValue(Expr *e, LValue lv);
 
   void emitNullInitializationToLValue(mlir::Location loc, LValue lv);
 
   void Visit(Expr *e) { StmtVisitor<AggExprEmitter>::Visit(e); }
 
+  void VisitCallExpr(const CallExpr *e);
+
+  void VisitDeclRefExpr(DeclRefExpr *e) { emitAggLoadOfLValue(e); }
+
   void VisitInitListExpr(InitListExpr *e);
   void VisitCXXConstructExpr(const CXXConstructExpr *e);
 
@@ -80,6 +101,17 @@ static bool isTrivialFiller(Expr *e) {
   return false;
 }
 
+/// Given an expression with aggregate type that represents a value lvalue, this
+/// method emits the address of the lvalue, then loads the result into DestPtr.
+void AggExprEmitter::emitAggLoadOfLValue(const Expr *e) {
+  LValue lv = cgf.emitLValue(e);
+
+  // If the type of the l-value is atomic, then do an atomic load.
+  assert(!cir::MissingFeatures::opLoadStoreAtomic());
+
+  emitFinalDestCopy(e->getType(), lv);
+}
+
 void AggExprEmitter::emitArrayInit(Address destPtr, cir::ArrayType arrayTy,
                                    QualType arrayQTy, Expr *e,
                                    ArrayRef<Expr *> args, Expr *arrayFiller) {
@@ -182,6 +214,18 @@ void AggExprEmitter::emitArrayInit(Address destPtr, cir::ArrayType arrayTy,
   }
 }
 
+/// Perform the final copy to destPtr, if desired.
+void AggExprEmitter::emitFinalDestCopy(QualType type, const LValue &src) {
+  // If dest is ignored, then we're evaluating an aggregate expression
+  // in a context that doesn't care about the result.  Note that loads
+  // from volatile l-values force the existence of a non-ignored
+  // destination.
+  if (dest.isIgnored())
+    return;
+
+  cgf.cgm.errorNYI("emitFinalDestCopy: non-ignored dest is NYI");
+}
+
 void AggExprEmitter::emitInitializationToLValue(Expr *e, LValue lv) {
   const QualType type = lv.getType();
 
@@ -246,6 +290,44 @@ void AggExprEmitter::emitNullInitializationToLValue(mlir::Location loc,
   cgf.emitNullInitialization(loc, lv.getAddress(), lv.getType());
 }
 
+void AggExprEmitter::VisitCallExpr(const CallExpr *e) {
+  if (e->getCallReturnType(cgf.getContext())->isReferenceType()) {
+    cgf.cgm.errorNYI(e->getSourceRange(), "reference return type");
+    return;
+  }
+
+  withReturnValueSlot(
+      e, [&](ReturnValueSlot slot) { return cgf.emitCallExpr(e, slot); });
+}
+
+void AggExprEmitter::withReturnValueSlot(
+    const Expr *e, llvm::function_ref<RValue(ReturnValueSlot)> fn) {
+  QualType retTy = e->getType();
+
+  assert(!cir::MissingFeatures::aggValueSlotDestructedFlag());
+  bool requiresDestruction =
+      retTy.isDestructedType() == QualType::DK_nontrivial_c_struct;
+  if (requiresDestruction)
+    cgf.cgm.errorNYI(
+        e->getSourceRange(),
+        "withReturnValueSlot: return value requiring destruction is NYI");
+
+  // If it makes no observable difference, save a memcpy + temporary.
+  //
+  // We need to always provide our own temporary if destruction is required.
+  // Otherwise, fn will emit its own, notice that it's "unused", and end its
+  // lifetime before we have the chance to emit a proper destructor call.
+  assert(!cir::MissingFeatures::aggValueSlotAlias());
+  assert(!cir::MissingFeatures::aggValueSlotGC());
+
+  Address retAddr = dest.getAddress();
+  assert(!cir::MissingFeatures::emitLifetimeMarkers());
+
+  assert(!cir::MissingFeatures::aggValueSlotVolatile());
+  assert(!cir::MissingFeatures::aggValueSlotDestructedFlag());
+  fn(ReturnValueSlot(retAddr));
+}
+
 void AggExprEmitter::VisitInitListExpr(InitListExpr *e) {
   if (e->hadArrayRangeDesignator())
     llvm_unreachable("GNU array range designator extension");