[CIR] Upstream aggregate binary assign handling

This upstreams the implementation for handling binary assignment involving
aggregate types.

Author: andykaylor

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

@@ -303,8 +303,9 @@ class CIRBaseBuilderTy : public mlir::OpBuilder {
   }
 
   /// Create a copy with inferred length.
-  cir::CopyOp createCopy(mlir::Value dst, mlir::Value src) {
-    return cir::CopyOp::create(*this, dst.getLoc(), dst, src);
+  cir::CopyOp createCopy(mlir::Value dst, mlir::Value src,
+                         bool isVolatile = false) {
+    return cir::CopyOp::create(*this, dst.getLoc(), dst, src, isVolatile);
   }
 
   cir::StoreOp createStore(mlir::Location loc, mlir::Value val, mlir::Value dst,

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

@@ -2719,6 +2719,8 @@ def CIR_CopyOp : CIR_Op<"copy",[
     type of `src` and `dst` must match and both must implement the
     `DataLayoutTypeInterface`.
 
+    The `volatile` keyword indicates that the operation is volatile.
+
     Examples:
 
     ```mlir
@@ -2729,10 +2731,11 @@ def CIR_CopyOp : CIR_Op<"copy",[
 
   let arguments = (ins
       Arg<CIR_PointerType, "", [MemWrite]>:$dst,
-      Arg<CIR_PointerType, "", [MemRead]>:$src
+      Arg<CIR_PointerType, "", [MemRead]>:$src,
+      UnitAttr:$is_volatile
   );
 
-  let assemblyFormat = [{$src `to` $dst
+  let assemblyFormat = [{$src `to` $dst (`volatile` $is_volatile^)?
                         attr-dict `:` qualified(type($dst))
   }];
   let hasVerifier = 1;

clang/include/clang/CIR/MissingFeatures.h

@@ -170,9 +170,10 @@ struct MissingFeatures {
   static bool atomicInfo() { return false; }
   static bool atomicInfoGetAtomicPointer() { return false; }
   static bool atomicInfoGetAtomicAddress() { return false; }
-  static bool atomicUseLibCall() { return false; }
   static bool atomicScope() { return false; }
   static bool atomicSyncScopeID() { return false; }
+  static bool atomicTypes() { return false; }
+  static bool atomicUseLibCall() { return false; }
 
   // Global ctor handling
   static bool globalCtorLexOrder() { return false; }

clang/lib/CIR/CodeGen/CIRGenDecl.cpp

@@ -740,6 +740,16 @@ struct CallStackRestore final : EHScopeStack::Cleanup {
 };
 } // namespace
 
+/// Push the standard destructor for the given type as
+/// at least a normal cleanup.
+void CIRGenFunction::pushDestroy(QualType::DestructionKind dtorKind,
+                                 Address addr, QualType type) {
+  assert(dtorKind && "cannot push destructor for trivial type");
+
+  CleanupKind cleanupKind = getCleanupKind(dtorKind);
+  pushDestroy(cleanupKind, addr, type, getDestroyer(dtorKind));
+}
+
 void CIRGenFunction::pushDestroy(CleanupKind cleanupKind, Address addr,
                                  QualType type, Destroyer *destroyer) {
   pushFullExprCleanup<DestroyObject>(cleanupKind, addr, type, destroyer);

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -1626,14 +1626,15 @@ 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, AggValueSlot aggSlot) {
+RValue CIRGenFunction::emitAnyExpr(const Expr *e, AggValueSlot aggSlot,
+                                   bool ignoreResult) {
   switch (CIRGenFunction::getEvaluationKind(e->getType())) {
   case cir::TEK_Scalar:
     return RValue::get(emitScalarExpr(e));
   case cir::TEK_Complex:
     return RValue::getComplex(emitComplexExpr(e));
   case cir::TEK_Aggregate: {
-    if (aggSlot.isIgnored())
+    if (!ignoreResult && aggSlot.isIgnored())
       aggSlot = createAggTemp(e->getType(), getLoc(e->getSourceRange()),
                               getCounterAggTmpAsString());
     emitAggExpr(e, aggSlot);
@@ -1869,8 +1870,7 @@ RValue CIRGenFunction::emitCallExpr(const clang::CallExpr *e,
 /// Emit code to compute the specified expression, ignoring the result.
 void CIRGenFunction::emitIgnoredExpr(const Expr *e) {
   if (e->isPRValue()) {
-    assert(!cir::MissingFeatures::aggValueSlot());
-    emitAnyExpr(e);
+    emitAnyExpr(e, AggValueSlot::ignored(), true);
     return;
   }
 

clang/lib/CIR/CodeGen/CIRGenExprAggregate.cpp

@@ -24,6 +24,73 @@ using namespace clang;
 using namespace clang::CIRGen;
 
 namespace {
+// FIXME(cir): This should be a common helper between CIRGen
+// and traditional CodeGen
+/// Is the value of the given expression possibly a reference to or
+/// into a __block variable?
+static bool isBlockVarRef(const Expr *e) {
+  // Make sure we look through parens.
+  e = e->IgnoreParens();
+
+  // Check for a direct reference to a __block variable.
+  if (const DeclRefExpr *dre = dyn_cast<DeclRefExpr>(e)) {
+    const VarDecl *var = dyn_cast<VarDecl>(dre->getDecl());
+    return (var && var->hasAttr<BlocksAttr>());
+  }
+
+  // More complicated stuff.
+
+  // Binary operators.
+  if (const BinaryOperator *op = dyn_cast<BinaryOperator>(e)) {
+    // For an assignment or pointer-to-member operation, just care
+    // about the LHS.
+    if (op->isAssignmentOp() || op->isPtrMemOp())
+      return isBlockVarRef(op->getLHS());
+
+    // For a comma, just care about the RHS.
+    if (op->getOpcode() == BO_Comma)
+      return isBlockVarRef(op->getRHS());
+
+    // FIXME: pointer arithmetic?
+    return false;
+
+    // Check both sides of a conditional operator.
+  } else if (const AbstractConditionalOperator *op =
+                 dyn_cast<AbstractConditionalOperator>(e)) {
+    return isBlockVarRef(op->getTrueExpr()) ||
+           isBlockVarRef(op->getFalseExpr());
+
+    // OVEs are required to support BinaryConditionalOperators.
+  } else if (const OpaqueValueExpr *op = dyn_cast<OpaqueValueExpr>(e)) {
+    if (const Expr *src = op->getSourceExpr())
+      return isBlockVarRef(src);
+
+    // Casts are necessary to get things like (*(int*)&var) = foo().
+    // We don't really care about the kind of cast here, except
+    // we don't want to look through l2r casts, because it's okay
+    // to get the *value* in a __block variable.
+  } else if (const CastExpr *cast = dyn_cast<CastExpr>(e)) {
+    if (cast->getCastKind() == CK_LValueToRValue)
+      return false;
+    return isBlockVarRef(cast->getSubExpr());
+
+    // Handle unary operators.  Again, just aggressively look through
+    // it, ignoring the operation.
+  } else if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(e)) {
+    return isBlockVarRef(uop->getSubExpr());
+
+    // Look into the base of a field access.
+  } else if (const MemberExpr *mem = dyn_cast<MemberExpr>(e)) {
+    return isBlockVarRef(mem->getBase());
+
+    // Look into the base of a subscript.
+  } else if (const ArraySubscriptExpr *sub = dyn_cast<ArraySubscriptExpr>(e)) {
+    return isBlockVarRef(sub->getBase());
+  }
+
+  return false;
+}
+
 class AggExprEmitter : public StmtVisitor<AggExprEmitter> {
 
   CIRGenFunction &cgf;
@@ -41,9 +108,7 @@ class AggExprEmitter : public StmtVisitor<AggExprEmitter> {
   AggValueSlot ensureSlot(mlir::Location loc, QualType t) {
     if (!dest.isIgnored())
       return dest;
-
-    cgf.cgm.errorNYI(loc, "Slot for ignored address");
-    return dest;
+    return cgf.createAggTemp(t, loc, "agg.tmp.ensured");
   }
 
   void ensureDest(mlir::Location loc, QualType ty) {
@@ -89,6 +154,47 @@ class AggExprEmitter : public StmtVisitor<AggExprEmitter> {
     (void)cgf.emitCompoundStmt(*e->getSubStmt(), &retAlloca, dest);
   }
 
+  void VisitBinAssign(const BinaryOperator *e) {
+    // For an assignment to work, the value on the right has
+    // to be compatible with the value on the left.
+    assert(cgf.getContext().hasSameUnqualifiedType(e->getLHS()->getType(),
+                                                   e->getRHS()->getType()) &&
+           "Invalid assignment");
+
+    if (isBlockVarRef(e->getLHS()) &&
+        e->getRHS()->HasSideEffects(cgf.getContext())) {
+      cgf.cgm.errorNYI(e->getSourceRange(),
+                       "block var reference with side effects");
+      return;
+    }
+
+    LValue lhs = cgf.emitLValue(e->getLHS());
+
+    // If we have an atomic type, evaluate into the destination and then
+    // do an atomic copy.
+    assert(!cir::MissingFeatures::atomicTypes());
+
+    // Codegen the RHS so that it stores directly into the LHS.
+    assert(!cir::MissingFeatures::aggValueSlotGC());
+    AggValueSlot lhsSlot = AggValueSlot::forLValue(
+        lhs, AggValueSlot::IsDestructed, AggValueSlot::IsAliased,
+        AggValueSlot::MayOverlap);
+
+    // A non-volatile aggregate destination might have volatile member.
+    if (!lhsSlot.isVolatile() && cgf.hasVolatileMember(e->getLHS()->getType()))
+      lhsSlot.setVolatile(true);
+
+    cgf.emitAggExpr(e->getRHS(), lhsSlot);
+
+    // Copy into the destination if the assignment isn't ignored.
+    emitFinalDestCopy(e->getType(), lhs);
+
+    if (!dest.isIgnored() && !dest.isExternallyDestructed() &&
+        e->getType().isDestructedType() == QualType::DK_nontrivial_c_struct)
+      cgf.pushDestroy(QualType::DK_nontrivial_c_struct, dest.getAddress(),
+                      e->getType());
+  }
+
   void VisitDeclRefExpr(DeclRefExpr *e) { emitAggLoadOfLValue(e); }
 
   void VisitInitListExpr(InitListExpr *e);
@@ -195,9 +301,6 @@ class AggExprEmitter : public StmtVisitor<AggExprEmitter> {
     cgf.cgm.errorNYI(e->getSourceRange(),
                      "AggExprEmitter: VisitPointerToDataMemberBinaryOperator");
   }
-  void VisitBinAssign(const BinaryOperator *e) {
-    cgf.cgm.errorNYI(e->getSourceRange(), "AggExprEmitter: VisitBinAssign");
-  }
   void VisitBinComma(const BinaryOperator *e) {
     cgf.emitIgnoredExpr(e->getLHS());
     Visit(e->getRHS());
@@ -487,7 +590,8 @@ void AggExprEmitter::emitCopy(QualType type, const AggValueSlot &dest,
   LValue destLV = cgf.makeAddrLValue(dest.getAddress(), type);
   LValue srcLV = cgf.makeAddrLValue(src.getAddress(), type);
   assert(!cir::MissingFeatures::aggValueSlotVolatile());
-  cgf.emitAggregateCopy(destLV, srcLV, type, dest.mayOverlap());
+  cgf.emitAggregateCopy(destLV, srcLV, type, dest.mayOverlap(),
+                        dest.isVolatile() || src.isVolatile());
 }
 
 void AggExprEmitter::emitInitializationToLValue(Expr *e, LValue lv) {
@@ -788,7 +892,8 @@ void CIRGenFunction::emitAggExpr(const Expr *e, AggValueSlot slot) {
 }
 
 void CIRGenFunction::emitAggregateCopy(LValue dest, LValue src, QualType ty,
-                                       AggValueSlot::Overlap_t mayOverlap) {
+                                       AggValueSlot::Overlap_t mayOverlap,
+                                       bool isVolatile) {
   // TODO(cir): this function needs improvements, commented code for now since
   // this will be touched again soon.
   assert(!ty->isAnyComplexType() && "Unexpected copy of complex");
@@ -844,7 +949,7 @@ void CIRGenFunction::emitAggregateCopy(LValue dest, LValue src, QualType ty,
     cgm.errorNYI("emitAggregateCopy: GC");
 
   [[maybe_unused]] cir::CopyOp copyOp =
-      builder.createCopy(destPtr.getPointer(), srcPtr.getPointer());
+      builder.createCopy(destPtr.getPointer(), srcPtr.getPointer(), isVolatile);
 
   assert(!cir::MissingFeatures::opTBAA());
 }

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -853,6 +853,14 @@ class CIRGenFunction : public CIRGenTypeCache {
                      FunctionArgList args, clang::SourceLocation loc,
                      clang::SourceLocation startLoc);
 
+  /// returns true if aggregate type has a volatile member.
+  /// TODO(cir): this could be a common AST helper between LLVM / CIR.
+  bool hasVolatileMember(QualType t) {
+    if (const auto *rd = t->getAsRecordDecl())
+      return rd->hasVolatileMember();
+    return false;
+  }
+
   /// The cleanup depth enclosing all the cleanups associated with the
   /// parameters.
   EHScopeStack::stable_iterator prologueCleanupDepth;
@@ -1077,6 +1085,9 @@ class CIRGenFunction : public CIRGenTypeCache {
 
   static Destroyer destroyCXXObject;
 
+  void pushDestroy(QualType::DestructionKind dtorKind, Address addr,
+                   QualType type);
+
   void pushDestroy(CleanupKind kind, Address addr, QualType type,
                    Destroyer *destroyer);
 
@@ -1131,14 +1142,16 @@ class CIRGenFunction : public CIRGenTypeCache {
   ///        occupied by some other object. More efficient code can often be
   ///        generated if not.
   void emitAggregateCopy(LValue dest, LValue src, QualType eltTy,
-                         AggValueSlot::Overlap_t mayOverlap);
+                         AggValueSlot::Overlap_t mayOverlap,
+                         bool isVolatile = false);
 
   /// Emit code to compute the specified expression which can have any type. The
   /// result is returned as an RValue struct. If this is an aggregate
   /// expression, the aggloc/agglocvolatile arguments indicate where the result
   /// should be returned.
   RValue emitAnyExpr(const clang::Expr *e,
-                     AggValueSlot aggSlot = AggValueSlot::ignored());
+                     AggValueSlot aggSlot = AggValueSlot::ignored(),
+                     bool ignoreResult = false);
 
   /// Emits the code necessary to evaluate an arbitrary expression into the
   /// given memory location.

clang/lib/CIR/CodeGen/CIRGenValue.h

@@ -380,6 +380,15 @@ class AggValueSlot {
 
   clang::Qualifiers getQualifiers() const { return quals; }
 
+  bool isVolatile() const { return quals.hasVolatile(); }
+
+  void setVolatile(bool flag) {
+    if (flag)
+      quals.addVolatile();
+    else
+      quals.removeVolatile();
+  }
+
   Address getAddress() const { return addr; }
 
   bool isIgnored() const { return !addr.isValid(); }

clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp

@@ -182,7 +182,7 @@ mlir::LogicalResult CIRToLLVMCopyOpLowering::matchAndRewrite(
       rewriter, op.getLoc(), rewriter.getI32Type(), op.getLength(layout));
   assert(!cir::MissingFeatures::aggValueSlotVolatile());
   rewriter.replaceOpWithNewOp<mlir::LLVM::MemcpyOp>(
-      op, adaptor.getDst(), adaptor.getSrc(), length, /*isVolatile=*/false);
+      op, adaptor.getDst(), adaptor.getSrc(), length, op.getIsVolatile());
   return mlir::success();
 }
 

clang/test/CIR/CodeGen/binassign.c

@@ -1,4 +1,4 @@
-// RUN: %clang_cc1 -std=c23 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir %s -o %t.cir
+// RUN: %clang_cc1 -std=c23 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir -mmlir --mlir-print-ir-before=cir-lowering-prepare %s -o %t.cir 2> %t-before-lp.cir
 // RUN: FileCheck --input-file=%t.cir %s -check-prefix=CIR
 // RUN: %clang_cc1 -std=c23 -triple x86_64-unknown-linux-gnu -Wno-unused-value -fclangir -emit-llvm %s -o %t-cir.ll
 // RUN: FileCheck --input-file=%t-cir.ll %s -check-prefix=LLVM
@@ -54,3 +54,49 @@ void binary_assign(void) {
 // OGCG:         store float 0x40091EB860000000, ptr %[[F_PTR]]
 // OGCG:         store i32 42, ptr %[[I_PTR]]
 // OGCG:         ret void
+
+struct S {
+  int a;
+  float b;
+};
+
+struct SV {
+  int a;
+  volatile float b;
+};
+
+struct S gs;
+struct SV gsv;
+
+void binary_assign_struct() {
+  // Test normal struct assignment
+  struct S ls;
+  ls = gs;
+
+  // Test assignment of a struct with a volatile member
+  struct SV lsv;
+  lsv = gsv;
+}
+
+// CIR: cir.func{{.*}} @binary_assign_struct() {
+// CIR:   %[[LS:.*]] = cir.alloca ![[REC_S:.*]], !cir.ptr<![[REC_S]]>, ["ls"]
+// CIR:   %[[LSV:.*]] = cir.alloca ![[REC_SV:.*]], !cir.ptr<![[REC_SV]]>, ["lsv"]
+// CIR:   %[[GS_PTR:.*]] = cir.get_global @gs : !cir.ptr<![[REC_S]]>
+// CIR:   cir.copy %[[GS_PTR]] to %[[LS]] : !cir.ptr<![[REC_S]]>
+// CIR:   %[[GSV_PTR:.*]] = cir.get_global @gsv : !cir.ptr<![[REC_SV]]>
+// CIR:   cir.copy %[[GSV_PTR]] to %[[LSV]] volatile : !cir.ptr<![[REC_SV]]>
+// CIR:   cir.return
+
+// LLVM: define {{.*}}void @binary_assign_struct() {
+// LLVM:   %[[LS_PTR:.*]] = alloca %struct.S
+// LLVM:   %[[LSV_PTR:.*]] = alloca %struct.SV
+// LLVM:   call void @llvm.memcpy.p0.p0.i32(ptr %[[LS_PTR]], ptr @gs, i32 8, i1 false)
+// LLVM:   call void @llvm.memcpy.p0.p0.i32(ptr %[[LSV_PTR]], ptr @gsv, i32 8, i1 true)
+// LLVM:   ret void
+
+// OGCG: define {{.*}}void @binary_assign_struct()
+// OGCG:   %[[LS_PTR:.*]] = alloca %struct.S
+// OGCG:   %[[LSV_PTR:.*]] = alloca %struct.SV
+// OGCG:   call void @llvm.memcpy.p0.p0.i64(ptr align 4 %[[LS_PTR]], ptr align 4 @gs, i64 8, i1 false)
+// OGCG:   call void @llvm.memcpy.p0.p0.i64(ptr align 4 %[[LSV_PTR]], ptr align 4 @gsv, i64 8, i1 true)
+// OGCG:   ret void