Implement __attribute__((objc_direct)), __attribute__((objc_direct_members))

__attribute__((objc_direct)) is an attribute on methods declaration, and
__attribute__((objc_direct_members)) on implementation, categories or
extensions.

A `direct` property specifier is added (@Property(direct) type name)

These attributes / specifiers cause the method to have no associated
Objective-C metadata (for the property or the method itself), and the
calling convention to be a direct C function call.

The symbol for the method has enforced hidden visibility and such direct
calls are hence unreachable cross image. An explicit C function must be
made if so desired to wrap them.

The implicit `self` and `_cmd` arguments are preserved, however to
maintain compatibility with the usual `objc_msgSend` semantics,
3 fundamental precautions are taken:

1) for instance methods, `self` is nil-checked. On arm64 backends this
   typically adds a single instruction (cbz x0, <closest-ret>) to the
   codegen, for the vast majority of the cases when the return type is a
   scalar.

2) for class methods, because the class may not be realized/initialized
   yet, a call to `[self self]` is emitted. When the proper deployment
   target is used, this is optimized to `objc_opt_self(self)`.

   However, long term we might want to emit something better that the
   optimizer can reason about. When inlining kicks in, these calls
   aren't optimized away as the optimizer has no idea that a single call
   is really necessary.

3) the calling convention for the `_cmd` argument is changed: the caller
   leaves the second argument to the call undefined, and the selector is
   loaded inside the body when it's referenced only.

As far as error reporting goes, the compiler refuses:
- making any overloads direct,
- making an overload of a direct method,
- implementations marked as direct when the declaration in the
  interface isn't (the other way around is allowed, as the direct
  attribute is inherited from the declaration),
- marking methods required for protocol conformance as direct,
- messaging an unqualified `id` with a direct method,
- forming any @selector() expression with only direct selectors.

As warnings:
- any inconsistency of direct-related calling convention when
  @selector() or messaging is used,
- forming any @selector() expression with a possibly direct selector.

Lastly an `objc_direct_members` attribute is added that can decorate
`@implementation` blocks and causes methods only declared there (and in
no `@interface`) to be automatically direct. When decorating an
`@interface` then all methods and properties declared in this block are
marked direct.

Radar-ID: rdar://problem/2684889
Differential Revision: https://reviews.llvm.org/D69991
Reviewed-By: John McCall

Author: MadCoder

clang/include/clang/AST/DeclObjC.h

@@ -410,7 +410,7 @@ class ObjCMethodDecl : public NamedDecl, public DeclContext {
   /// \return the type for \c self and set \arg selfIsPseudoStrong and
   /// \arg selfIsConsumed accordingly.
   QualType getSelfType(ASTContext &Context, const ObjCInterfaceDecl *OID,
-                       bool &selfIsPseudoStrong, bool &selfIsConsumed);
+                       bool &selfIsPseudoStrong, bool &selfIsConsumed) const;
 
   ImplicitParamDecl * getSelfDecl() const { return SelfDecl; }
   void setSelfDecl(ImplicitParamDecl *SD) { SelfDecl = SD; }
@@ -476,6 +476,9 @@ class ObjCMethodDecl : public NamedDecl, public DeclContext {
     ObjCMethodDeclBits.HasSkippedBody = Skipped;
   }
 
+  /// True if the method is tagged as objc_direct
+  bool isDirectMethod() const;
+
   /// Returns the property associated with this method's selector.
   ///
   /// Note that even if this particular method is not marked as a property
@@ -757,13 +760,14 @@ class ObjCPropertyDecl : public NamedDecl {
     /// property attribute rather than a type qualifier.
     OBJC_PR_nullability = 0x1000,
     OBJC_PR_null_resettable = 0x2000,
-    OBJC_PR_class = 0x4000
+    OBJC_PR_class = 0x4000,
+    OBJC_PR_direct = 0x8000
     // Adding a property should change NumPropertyAttrsBits
   };
 
   enum {
     /// Number of bits fitting all the property attributes.
-    NumPropertyAttrsBits = 15
+    NumPropertyAttrsBits = 16
   };
 
   enum SetterKind { Assign, Retain, Copy, Weak };
@@ -886,6 +890,7 @@ class ObjCPropertyDecl : public NamedDecl {
 
   bool isInstanceProperty() const { return !isClassProperty(); }
   bool isClassProperty() const { return PropertyAttributes & OBJC_PR_class; }
+  bool isDirectProperty() const { return PropertyAttributes & OBJC_PR_direct; }
 
   ObjCPropertyQueryKind getQueryKind() const {
     return isClassProperty() ? ObjCPropertyQueryKind::OBJC_PR_query_class :

clang/include/clang/Basic/Attr.td

@@ -1861,6 +1861,20 @@ def ObjCDesignatedInitializer : Attr {
   let Documentation = [Undocumented];
 }
 
+def ObjCDirect : Attr {
+  let Spellings = [Clang<"objc_direct">];
+  let Subjects = SubjectList<[ObjCMethod], ErrorDiag>;
+  let LangOpts = [ObjC];
+  let Documentation = [ObjCDirectDocs];
+}
+
+def ObjCDirectMembers : Attr {
+  let Spellings = [Clang<"objc_direct_members">];
+  let Subjects = SubjectList<[ObjCImpl, ObjCCategory], ErrorDiag>;
+  let LangOpts = [ObjC];
+  let Documentation = [ObjCDirectMembersDocs];
+}
+
 def ObjCRuntimeName : Attr {
   let Spellings = [Clang<"objc_runtime_name">];
   let Subjects = SubjectList<[ObjCInterface, ObjCProtocol], ErrorDiag>;

clang/include/clang/Basic/AttrDocs.td

@@ -3912,6 +3912,104 @@ overheads associated with defining and calling such a method.
   }];
 }
 
+def ObjCDirectDocs : Documentation {
+  let Category = DocCatDecl;
+  let Content = [{
+The ``objc_direct`` attribute can be used to mark an Objective-C method as
+being *direct*.  A direct method is treated statically like an ordinary method,
+but dynamically it behaves more like a C function.  This lowers some of the costs
+associated with the method but also sacrifices some of the ordinary capabilities
+of Objective-C methods.
+
+A message send of a direct method calls the implementation directly, as if it
+were a C function, rather than using ordinary Objective-C method dispatch. This
+is substantially faster and potentially allows the implementation to be inlined,
+but it also means the method cannot be overridden in subclasses or replaced
+dynamically, as ordinary Objective-C methods can.
+
+Furthermore, a direct method is not listed in the class's method lists. This
+substantially reduces the code-size overhead of the method but also means it
+cannot be called dynamically using ordinary Objective-C method dispatch at all;
+in particular, this means that it cannot override a superclass method or satisfy
+a protocol requirement.
+
+Because a direct method cannot be overridden, it is an error to perform
+a ``super`` message send of one.
+
+Although a message send of a direct method causes the method to be called
+directly as if it were a C function, it still obeys Objective-C semantics in other
+ways:
+
+- If the receiver is ``nil``, the message send does nothing and returns the zero value
+  for the return type.
+
+- A message send of a direct class method will cause the class to be initialized,
+  including calling the ``+initialize`` method if present.
+
+- The implicit ``_cmd`` parameter containing the method's selector is still defined.
+ In order to minimize code-size costs, the implementation will not emit a reference
+ to the selector if the parameter is unused within the method.
+
+Symbols for direct method implementations are implicitly given hidden
+visibility, meaning that they can only be called within the same linkage unit.
+
+It is an error to do any of the following:
+
+- declare a direct method in a protocol,
+- declare an override of a direct method with a method in a subclass,
+- declare an override of a non-direct method with a direct method in a subclass,
+- declare a method with different directness in different class interfaces, or
+- implement a non-direct method (as declared in any class interface) with a direct method.
+
+If any of these rules would be violated if every method defined in an
+``@implementation`` within a single linkage unit were declared in an
+appropriate class interface, the program is ill-formed with no diagnostic
+required.  If a violation of this rule is not diagnosed, behavior remains
+well-defined; this paragraph is simply reserving the right to diagnose such
+conflicts in the future, not to treat them as undefined behavior.
+
+Additionally, Clang will warn about any ``@selector`` expression that
+names a selector that is only known to be used for direct methods.
+
+For the purpose of these rules, a "class interface" includes a class's primary
+``@interface`` block, its class extensions, its categories, its declared protocols,
+and all the class interfaces of its superclasses.
+
+An Objective-C property can be declared with the ``direct`` property
+attribute.  If a direct property declaration causes an implicit declaration of
+a getter or setter method (that is, if the given method is not explicitly
+declared elsewhere), the method is declared to be direct.
+
+Some programmers may wish to make many methods direct at once.  In order
+to simplify this, the ``objc_direct_members`` attribute is provided; see its
+documentation for more information.
+  }];
+}
+
+def ObjCDirectMembersDocs : Documentation {
+  let Category = DocCatDecl;
+  let Content = [{
+The ``objc_direct_members`` attribute can be placed on an  Objective-C
+``@interface`` or ``@implementation`` to mark that methods declared
+therein should be considered direct by default.  See the documentation
+for ``objc_direct`` for more information about direct methods.
+
+When ``objc_direct_members`` is placed on an ``@interface`` block, every
+method in the block is considered to be declared as direct.  This includes any
+implicit method declarations introduced by property declarations.  If the method
+redeclares a non-direct method, the declaration is ill-formed, exactly as if the
+method was annotated with the ``objc_direct`` attribute.  ``objc_direct_members``
+cannot be placed on the primary interface of a class, only on category or class
+extension interfaces.
+
+When ``objc_direct_members`` is placed on an ``@implementation`` block,
+methods defined in the block are considered to be declared as direct unless
+they have been previously declared as non-direct in any interface of the class.
+This includes the implicit method definitions introduced by synthesized
+properties, including auto-synthesized properties.
+  }];
+}
+
 def SelectAnyDocs : Documentation {
   let Category = DocCatDecl;
   let Content = [{

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -988,6 +988,22 @@ def warn_objc_boxing_invalid_utf8_string : Warning<
   "string is ill-formed as UTF-8 and will become a null %0 when boxed">,
   InGroup<ObjCBoxing>;
 
+def err_objc_direct_on_protocol : Error<
+  "'objc_direct' attribute cannot be applied to %select{methods|properties}0 "
+  "declared in an Objective-C protocol">;
+def err_objc_direct_missing_on_decl : Error<
+  "direct method implementation was previously declared not direct">;
+def err_objc_direct_on_override : Error<
+  "methods that %select{override superclass methods|implement protocol requirements}0 cannot be direct">;
+def err_objc_override_direct_method : Error<
+  "cannot override a method that is declared direct by a superclass">;
+def warn_objc_direct_ignored : Warning<
+  "%0 attribute isn't implemented by this Objective-C runtime">,
+  InGroup<IgnoredAttributes>;
+def warn_objc_direct_property_ignored : Warning<
+  "direct attribute on property %0 ignored (not implemented by this Objective-C runtime)">,
+  InGroup<IgnoredAttributes>;
+
 def warn_conflicting_overriding_ret_types : Warning<
   "conflicting return type in "
   "declaration of %0%diff{: $ vs $|}1,2">,
@@ -1073,6 +1089,7 @@ def warn_accessor_property_type_mismatch : Warning<
   "type of property %0 does not match type of accessor %1">;
 def note_conv_function_declared_at : Note<"type conversion function declared here">;
 def note_method_declared_at : Note<"method %0 declared here">;
+def note_direct_method_declared_at : Note<"direct method %0 declared here">;
 def note_property_attribute : Note<"property %0 is declared "
   "%select{deprecated|unavailable|partial}1 here">;
 def err_setter_type_void : Error<"type of setter must be void">;
@@ -1308,6 +1325,8 @@ def warn_multiple_selectors: Warning<
   "several methods with selector %0 of mismatched types are found "
   "for the @selector expression">,
   InGroup<SelectorTypeMismatch>, DefaultIgnore;
+def err_direct_selector_expression: Error<
+  "@selector expression formed with direct selector %0">;
 
 def err_objc_kindof_nonobject : Error<
   "'__kindof' specifier cannot be applied to non-object type %0">;
@@ -1321,6 +1340,12 @@ def err_objc_method_unsupported_param_ret_type : Error<
 def warn_messaging_unqualified_id : Warning<
   "messaging unqualified id">, DefaultIgnore,
   InGroup<DiagGroup<"objc-messaging-id">>;
+def err_messaging_unqualified_id_with_direct_method : Error<
+  "messaging unqualified id with a method that is possibly direct">;
+def err_messaging_super_with_direct_method : Error<
+  "messaging super with a direct method">;
+def err_messaging_class_with_direct_method : Error<
+  "messaging a Class with a method that is possibly direct">;
 
 // C++ declarations
 def err_static_assert_expression_is_not_constant : Error<

clang/include/clang/Basic/ObjCRuntime.h

@@ -446,6 +446,20 @@ class ObjCRuntime {
     llvm_unreachable("bad kind");
   }
 
+  /// Does this runtime supports direct dispatch
+  bool allowsDirectDispatch() const {
+    switch (getKind()) {
+    case FragileMacOSX: return false;
+    case MacOSX: return true;
+    case iOS: return true;
+    case WatchOS: return true;
+    case GCC: return false;
+    case GNUstep: return false;
+    case ObjFW: return false;
+    }
+    llvm_unreachable("bad kind");
+  }
+
   /// Try to parse an Objective-C runtime specification from the given
   /// string.
   ///

clang/include/clang/Sema/DeclSpec.h

@@ -833,7 +833,8 @@ class ObjCDeclSpec {
     DQ_PR_unsafe_unretained = 0x800,
     DQ_PR_nullability = 0x1000,
     DQ_PR_null_resettable = 0x2000,
-    DQ_PR_class = 0x4000
+    DQ_PR_class = 0x4000,
+    DQ_PR_direct = 0x8000,
   };
 
   ObjCDeclSpec()
@@ -903,7 +904,7 @@ class ObjCDeclSpec {
   unsigned objcDeclQualifier : 7;
 
   // NOTE: VC++ treats enums as signed, avoid using ObjCPropertyAttributeKind
-  unsigned PropertyAttributes : 15;
+  unsigned PropertyAttributes : 16;
 
   unsigned Nullability : 2;
 

clang/include/clang/Sema/Sema.h

@@ -8901,6 +8901,9 @@ class Sema {
     RTC_Unknown
   };
 
+  void CheckObjCMethodDirectOverrides(ObjCMethodDecl *method,
+                                      ObjCMethodDecl *overridden);
+
   void CheckObjCMethodOverrides(ObjCMethodDecl *ObjCMethod,
                                 ObjCInterfaceDecl *CurrentClass,
                                 ResultTypeCompatibilityKind RTC);

clang/lib/AST/DeclObjC.cpp

@@ -823,6 +823,10 @@ ObjCMethodDecl *ObjCMethodDecl::CreateDeserialized(ASTContext &C, unsigned ID) {
                                     Selector(), QualType(), nullptr, nullptr);
 }
 
+bool ObjCMethodDecl::isDirectMethod() const {
+  return hasAttr<ObjCDirectAttr>();
+}
+
 bool ObjCMethodDecl::isThisDeclarationADesignatedInitializer() const {
   return getMethodFamily() == OMF_init &&
       hasAttr<ObjCDesignatedInitializerAttr>();
@@ -1077,7 +1081,7 @@ ObjCMethodFamily ObjCMethodDecl::getMethodFamily() const {
 QualType ObjCMethodDecl::getSelfType(ASTContext &Context,
                                      const ObjCInterfaceDecl *OID,
                                      bool &selfIsPseudoStrong,
-                                     bool &selfIsConsumed) {
+                                     bool &selfIsConsumed) const {
   QualType selfTy;
   selfIsPseudoStrong = false;
   selfIsConsumed = false;

clang/lib/AST/DeclPrinter.cpp

@@ -1461,6 +1461,11 @@ void DeclPrinter::VisitObjCPropertyDecl(ObjCPropertyDecl *PDecl) {
       first = false;
     }
 
+    if (PDecl->getPropertyAttributes() & ObjCPropertyDecl::OBJC_PR_direct) {
+      Out << (first ? "" : ", ") << "direct";
+      first = false;
+    }
+
     if (PDecl->getPropertyAttributes() &
         ObjCPropertyDecl::OBJC_PR_nonatomic) {
       Out << (first ? "" : ", ") << "nonatomic";

clang/lib/AST/JSONNodeDumper.cpp

@@ -1017,6 +1017,7 @@ void JSONNodeDumper::VisitObjCPropertyDecl(const ObjCPropertyDecl *D) {
     attributeOnlyIfTrue("unsafe_unretained",
                         Attrs & ObjCPropertyDecl::OBJC_PR_unsafe_unretained);
     attributeOnlyIfTrue("class", Attrs & ObjCPropertyDecl::OBJC_PR_class);
+    attributeOnlyIfTrue("direct", Attrs & ObjCPropertyDecl::OBJC_PR_direct);
     attributeOnlyIfTrue("nullability",
                         Attrs & ObjCPropertyDecl::OBJC_PR_nullability);
     attributeOnlyIfTrue("null_resettable",

clang/lib/AST/TextNodeDumper.cpp

@@ -1921,6 +1921,8 @@ void TextNodeDumper::VisitObjCPropertyDecl(const ObjCPropertyDecl *D) {
       OS << " unsafe_unretained";
     if (Attrs & ObjCPropertyDecl::OBJC_PR_class)
       OS << " class";
+    if (Attrs & ObjCPropertyDecl::OBJC_PR_direct)
+      OS << " direct";
     if (Attrs & ObjCPropertyDecl::OBJC_PR_getter)
       dumpDeclRef(D->getGetterMethodDecl(), "getter");
     if (Attrs & ObjCPropertyDecl::OBJC_PR_setter)

clang/lib/CodeGen/CGObjC.cpp

@@ -430,6 +430,20 @@ tryGenerateSpecializedMessageSend(CodeGenFunction &CGF, QualType ResultType,
   return None;
 }
 
+CodeGen::RValue CGObjCRuntime::GeneratePossiblySpecializedMessageSend(
+    CodeGenFunction &CGF, ReturnValueSlot Return, QualType ResultType,
+    Selector Sel, llvm::Value *Receiver, const CallArgList &Args,
+    const ObjCInterfaceDecl *OID, const ObjCMethodDecl *Method,
+    bool isClassMessage) {
+  if (Optional<llvm::Value *> SpecializedResult =
+          tryGenerateSpecializedMessageSend(CGF, ResultType, Receiver, Args,
+                                            Sel, Method, isClassMessage)) {
+    return RValue::get(SpecializedResult.getValue());
+  }
+  return GenerateMessageSend(CGF, Return, ResultType, Sel, Receiver, Args, OID,
+                             Method);
+}
+
 /// Instead of '[[MyClass alloc] init]', try to generate
 /// 'objc_alloc_init(MyClass)'. This provides a code size improvement on the
 /// caller side, as well as the optimized objc_alloc.
@@ -611,16 +625,9 @@ RValue CodeGenFunction::EmitObjCMessageExpr(const ObjCMessageExpr *E,
                                               method);
   } else {
     // Call runtime methods directly if we can.
-    if (Optional<llvm::Value *> SpecializedResult =
-            tryGenerateSpecializedMessageSend(*this, ResultType, Receiver, Args,
-                                              E->getSelector(), method,
-                                              isClassMessage)) {
-      result = RValue::get(SpecializedResult.getValue());
-    } else {
-      result = Runtime.GenerateMessageSend(*this, Return, ResultType,
-                                           E->getSelector(), Receiver, Args,
-                                           OID, method);
-    }
+    result = Runtime.GeneratePossiblySpecializedMessageSend(
+        *this, Return, ResultType, E->getSelector(), Receiver, Args, OID,
+        method, isClassMessage);
   }
 
   // For delegate init calls in ARC, implicitly store the result of
@@ -683,7 +690,13 @@ void CodeGenFunction::StartObjCMethod(const ObjCMethodDecl *OMD,
   llvm::Function *Fn = CGM.getObjCRuntime().GenerateMethod(OMD, CD);
 
   const CGFunctionInfo &FI = CGM.getTypes().arrangeObjCMethodDeclaration(OMD);
-  CGM.SetInternalFunctionAttributes(OMD, Fn, FI);
+  if (OMD->isDirectMethod()) {
+    Fn->setVisibility(llvm::Function::HiddenVisibility);
+    CGM.SetLLVMFunctionAttributes(OMD, FI, Fn);
+    CGM.SetLLVMFunctionAttributesForDefinition(OMD, Fn);
+  } else {
+    CGM.SetInternalFunctionAttributes(OMD, Fn, FI);
+  }
 
   args.push_back(OMD->getSelfDecl());
   args.push_back(OMD->getCmdDecl());
@@ -696,6 +709,14 @@ void CodeGenFunction::StartObjCMethod(const ObjCMethodDecl *OMD,
   StartFunction(OMD, OMD->getReturnType(), Fn, FI, args,
                 OMD->getLocation(), StartLoc);
 
+  if (OMD->isDirectMethod()) {
+    // This function is a direct call, it has to implement a nil check
+    // on entry.
+    //
+    // TODO: possibly have several entry points to elide the check
+    CGM.getObjCRuntime().GenerateDirectMethodPrologue(*this, Fn, OMD, CD);
+  }
+
   // In ARC, certain methods get an extra cleanup.
   if (CGM.getLangOpts().ObjCAutoRefCount &&
       OMD->isInstanceMethod() &&