Exploit this-return of a callsite in a this-return function.

For constructors/desctructors that return 'this', if there exists a callsite
that returns 'this' and is immediately before the return instruction, make
sure we are using the return value from the callsite.

We don't need to keep 'this' alive through the callsite. It also enables
optimizations in the backend, such as tail call optimization.

Updated from r177211.
rdar://12818789


git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@177541 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/CGCXXABI.h b/lib/CodeGen/CGCXXABI.h
index cdc87b7..d0384ec 100644
--- a/lib/CodeGen/CGCXXABI.h
+++ b/lib/CodeGen/CGCXXABI.h
@@ -91,6 +91,10 @@
     return *MangleCtx;
   }
 
+  /// Returns true if the given instance method is one of the
+  /// kinds that the ABI says returns 'this'.
+  virtual bool HasThisReturn(GlobalDecl GD) const { return false; }
+
   /// Find the LLVM type used to represent the given member pointer
   /// type.
   virtual llvm::Type *
@@ -209,7 +213,8 @@
   /// Emit the ABI-specific prolog for the function.
   virtual void EmitInstanceFunctionProlog(CodeGenFunction &CGF) = 0;
 
-  virtual void EmitConstructorCall(CodeGenFunction &CGF,
+  /// Emit the constructor call. Return the function that is called.
+  virtual llvm::Value *EmitConstructorCall(CodeGenFunction &CGF,
                                    const CXXConstructorDecl *D,
                                    CXXCtorType Type, bool ForVirtualBase,
                                    bool Delegating,
diff --git a/lib/CodeGen/CGCall.cpp b/lib/CodeGen/CGCall.cpp
index 9e97bce..9318973 100644
--- a/lib/CodeGen/CGCall.cpp
+++ b/lib/CodeGen/CGCall.cpp
@@ -1614,6 +1614,18 @@
   return store;
 }
 
+/// Check whether 'this' argument of a callsite matches 'this' of the caller.
+static bool checkThisPointer(llvm::Value *ThisArg, llvm::Value *This) {
+  if (ThisArg == This)
+    return true;
+  // Check whether ThisArg is a bitcast of This.
+  llvm::BitCastInst *Bitcast;
+  if ((Bitcast = dyn_cast<llvm::BitCastInst>(ThisArg)) &&
+      Bitcast->getOperand(0) == This)
+    return true;
+  return false;
+}
+
 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI) {
   // Functions with no result always return void.
   if (ReturnValue == 0) {
@@ -1705,6 +1717,19 @@
     llvm_unreachable("Invalid ABI kind for return argument");
   }
 
+  // If this function returns 'this', the last instruction is a CallInst
+  // that returns 'this', and 'this' argument of the CallInst points to
+  // the same object as CXXThisValue, use the return value from the CallInst.
+  // We will not need to keep 'this' alive through the callsite. It also enables
+  // optimizations in the backend, such as tail call optimization.
+  if (CalleeWithThisReturn && CGM.getCXXABI().HasThisReturn(CurGD)) {
+    llvm::BasicBlock *IP = Builder.GetInsertBlock();
+    llvm::CallInst *Callsite;
+    if (!IP->empty() && (Callsite = dyn_cast<llvm::CallInst>(&IP->back())) &&
+        Callsite->getCalledFunction() == CalleeWithThisReturn &&
+        checkThisPointer(Callsite->getOperand(0), CXXThisValue))
+      RV = Builder.CreateBitCast(Callsite, RetAI.getCoerceToType());
+  }
   llvm::Instruction *Ret = RV ? Builder.CreateRet(RV) : Builder.CreateRetVoid();
   if (!RetDbgLoc.isUnknown())
     Ret->setDebugLoc(RetDbgLoc);
diff --git a/lib/CodeGen/CGClass.cpp b/lib/CodeGen/CGClass.cpp
index 287d164..2ececb0 100644
--- a/lib/CodeGen/CGClass.cpp
+++ b/lib/CodeGen/CGClass.cpp
@@ -1666,8 +1666,11 @@
   }
 
   // Non-trivial constructors are handled in an ABI-specific manner.
-  CGM.getCXXABI().EmitConstructorCall(*this, D, Type, ForVirtualBase,
-                                      Delegating, This, ArgBeg, ArgEnd);
+  llvm::Value *Callee = CGM.getCXXABI().EmitConstructorCall(*this, D, Type,
+                            ForVirtualBase, Delegating, This, ArgBeg, ArgEnd);
+  if (CGM.getCXXABI().HasThisReturn(CurGD) &&
+      CGM.getCXXABI().HasThisReturn(GlobalDecl(D, Type)))
+     CalleeWithThisReturn = Callee;
 }
 
 void
@@ -1756,9 +1759,12 @@
     EmitDelegateCallArg(DelegateArgs, param);
   }
 
+  llvm::Value *Callee = CGM.GetAddrOfCXXConstructor(Ctor, CtorType);
   EmitCall(CGM.getTypes().arrangeCXXConstructorDeclaration(Ctor, CtorType),
-           CGM.GetAddrOfCXXConstructor(Ctor, CtorType), 
-           ReturnValueSlot(), DelegateArgs, Ctor);
+           Callee, ReturnValueSlot(), DelegateArgs, Ctor);
+  if (CGM.getCXXABI().HasThisReturn(CurGD) &&
+      CGM.getCXXABI().HasThisReturn(GlobalDecl(Ctor, CtorType)))
+     CalleeWithThisReturn = Callee;
 }
 
 namespace {
@@ -1825,6 +1831,9 @@
   EmitCXXMemberCall(DD, SourceLocation(), Callee, ReturnValueSlot(), This,
                     VTT, getContext().getPointerType(getContext().VoidPtrTy),
                     0, 0);
+  if (CGM.getCXXABI().HasThisReturn(CurGD) &&
+      CGM.getCXXABI().HasThisReturn(GlobalDecl(DD, Type)))
+     CalleeWithThisReturn = Callee;
 }
 
 namespace {
diff --git a/lib/CodeGen/CodeGenFunction.cpp b/lib/CodeGen/CodeGenFunction.cpp
index 27ef65f..bd24435 100644
--- a/lib/CodeGen/CodeGenFunction.cpp
+++ b/lib/CodeGen/CodeGenFunction.cpp
@@ -564,6 +564,10 @@
   SourceRange BodyRange;
   if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
 
+  // CalleeWithThisReturn keeps track of the last callee inside this function
+  // that returns 'this'. Before starting the function, we set it to null.
+  CalleeWithThisReturn = 0;
+
   // Emit the standard function prologue.
   StartFunction(GD, ResTy, Fn, FnInfo, Args, BodyRange.getBegin());
 
@@ -615,6 +619,9 @@
 
   // Emit the standard function epilogue.
   FinishFunction(BodyRange.getEnd());
+  // CalleeWithThisReturn keeps track of the last callee inside this function
+  // that returns 'this'. After finishing the function, we set it to null.
+  CalleeWithThisReturn = 0;
 
   // If we haven't marked the function nothrow through other means, do
   // a quick pass now to see if we can.
diff --git a/lib/CodeGen/CodeGenFunction.h b/lib/CodeGen/CodeGenFunction.h
index 55c21e8..19a4d57 100644
--- a/lib/CodeGen/CodeGenFunction.h
+++ b/lib/CodeGen/CodeGenFunction.h
@@ -1131,6 +1131,10 @@
   CGDebugInfo *DebugInfo;
   bool DisableDebugInfo;
 
+  /// If the current function returns 'this', use the field to keep track of
+  /// the callee that returns 'this'.
+  llvm::Value *CalleeWithThisReturn;
+
   /// DidCallStackSave - Whether llvm.stacksave has been called. Used to avoid
   /// calling llvm.stacksave for multiple VLAs in the same scope.
   bool DidCallStackSave;
diff --git a/lib/CodeGen/ItaniumCXXABI.cpp b/lib/CodeGen/ItaniumCXXABI.cpp
index 8c7a759..17e83a1 100644
--- a/lib/CodeGen/ItaniumCXXABI.cpp
+++ b/lib/CodeGen/ItaniumCXXABI.cpp
@@ -112,7 +112,7 @@
 
   void EmitInstanceFunctionProlog(CodeGenFunction &CGF);
 
-  void EmitConstructorCall(CodeGenFunction &CGF,
+  llvm::Value *EmitConstructorCall(CodeGenFunction &CGF,
                            const CXXConstructorDecl *D,
                            CXXCtorType Type, bool ForVirtualBase,
                            bool Delegating,
@@ -177,11 +177,11 @@
   llvm::Value *readArrayCookieImpl(CodeGenFunction &CGF, llvm::Value *allocPtr,
                                    CharUnits cookieSize);
 
-private:
   /// \brief Returns true if the given instance method is one of the
   /// kinds that the ARM ABI says returns 'this'.
-  static bool HasThisReturn(GlobalDecl GD) {
-    const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
+  bool HasThisReturn(GlobalDecl GD) const {
+    const CXXMethodDecl *MD = dyn_cast_or_null<CXXMethodDecl>(GD.getDecl());
+    if (!MD) return false;
     return ((isa<CXXDestructorDecl>(MD) && GD.getDtorType() != Dtor_Deleting) ||
             (isa<CXXConstructorDecl>(MD)));
   }
@@ -834,7 +834,7 @@
     CGF.Builder.CreateStore(getThisValue(CGF), CGF.ReturnValue);
 }
 
-void ItaniumCXXABI::EmitConstructorCall(CodeGenFunction &CGF,
+llvm::Value *ItaniumCXXABI::EmitConstructorCall(CodeGenFunction &CGF,
                                         const CXXConstructorDecl *D,
                                         CXXCtorType Type, bool ForVirtualBase,
                                         bool Delegating,
@@ -849,6 +849,7 @@
   // FIXME: Provide a source location here.
   CGF.EmitCXXMemberCall(D, SourceLocation(), Callee, ReturnValueSlot(), This,
                         VTT, VTTTy, ArgBeg, ArgEnd);
+  return Callee;
 }
 
 RValue ItaniumCXXABI::EmitVirtualDestructorCall(CodeGenFunction &CGF,
diff --git a/lib/CodeGen/MicrosoftCXXABI.cpp b/lib/CodeGen/MicrosoftCXXABI.cpp
index fb6b86d..85d9260 100644
--- a/lib/CodeGen/MicrosoftCXXABI.cpp
+++ b/lib/CodeGen/MicrosoftCXXABI.cpp
@@ -55,7 +55,7 @@
 
   void EmitInstanceFunctionProlog(CodeGenFunction &CGF);
 
-  void EmitConstructorCall(CodeGenFunction &CGF,
+  llvm::Value *EmitConstructorCall(CodeGenFunction &CGF,
                            const CXXConstructorDecl *D,
                            CXXCtorType Type, bool ForVirtualBase,
                            bool Delegating,
@@ -238,7 +238,7 @@
   }
 }
 
-void MicrosoftCXXABI::EmitConstructorCall(CodeGenFunction &CGF,
+llvm::Value *MicrosoftCXXABI::EmitConstructorCall(CodeGenFunction &CGF,
                                           const CXXConstructorDecl *D,
                                           CXXCtorType Type, bool ForVirtualBase,
                                           bool Delegating,
@@ -259,6 +259,7 @@
   CGF.EmitCXXMemberCall(D, SourceLocation(), Callee, ReturnValueSlot(), This,
                         ImplicitParam, ImplicitParamTy,
                         ArgBeg, ArgEnd);
+  return Callee;
 }
 
 RValue MicrosoftCXXABI::EmitVirtualDestructorCall(CodeGenFunction &CGF,
diff --git a/test/CodeGenCXX/arm.cpp b/test/CodeGenCXX/arm.cpp
index 3d3b147..48f2f00 100644
--- a/test/CodeGenCXX/arm.cpp
+++ b/test/CodeGenCXX/arm.cpp
@@ -56,15 +56,15 @@
   // CHECK:   [[THIS:%.*]] = alloca [[A]]*, align 4
   // CHECK:   store [[A]]* {{.*}}, [[A]]** [[THIS]]
   // CHECK:   [[THIS1:%.*]] = load [[A]]** [[THIS]]
-  // CHECK:   call [[A]]* @_ZN5test11AC2Ei(
-  // CHECK:   ret [[A]]* [[THIS1]]
+  // CHECK:   [[THIS2:%.*]] = call [[A]]* @_ZN5test11AC2Ei(
+  // CHECK:   ret [[A]]* [[THIS2]]
 
   // CHECK: define linkonce_odr [[A]]* @_ZN5test11AD1Ev([[A]]* %this) unnamed_addr
   // CHECK:   [[THIS:%.*]] = alloca [[A]]*, align 4
   // CHECK:   store [[A]]* {{.*}}, [[A]]** [[THIS]]
   // CHECK:   [[THIS1:%.*]] = load [[A]]** [[THIS]]
-  // CHECK:   call [[A]]* @_ZN5test11AD2Ev(
-  // CHECK:   ret [[A]]* [[THIS1]]
+  // CHECK:   [[THIS2:%.*]] = call [[A]]* @_ZN5test11AD2Ev(
+  // CHECK:   ret [[A]]* [[THIS2]]
 }
 
 // Awkward virtual cases.
diff --git a/test/CodeGenCXX/constructor-destructor-return-this.cpp b/test/CodeGenCXX/constructor-destructor-return-this.cpp
new file mode 100644
index 0000000..1ff922d
--- /dev/null
+++ b/test/CodeGenCXX/constructor-destructor-return-this.cpp
@@ -0,0 +1,60 @@
+//RUN: %clang_cc1 %s -emit-llvm -o - -triple=thumbv7-apple-ios3.0 -target-abi apcs-gnu | FileCheck %s
+
+// For constructors/desctructors that return 'this', if there exists a callsite
+// that returns 'this' and is immediately before the return instruction, make
+// sure we are using the return value from the callsite.
+// rdar://12818789
+
+// CHECK: define linkonce_odr [[A:%.*]] @_ZN11ObjectCacheC1Ev([[A]] %this) unnamed_addr
+// CHECK: [[THIS1:%.*]] = call [[A]] @_ZN11ObjectCacheC2Ev(
+// CHECK-NEXT: ret [[A]] [[THIS1]]
+
+// CHECK: define linkonce_odr [[A:%.*]] @_ZN5TimerI11ObjectCacheEC1EPS0_MS0_FvPS1_E([[A]] %this
+// CHECK: [[THIS1:%.*]] = call [[A]] @_ZN5TimerI11ObjectCacheEC2EPS0_MS0_FvPS1_E(
+// CHECK-NEXT: ret [[A]] [[THIS1]]
+
+// CHECK: define linkonce_odr [[A:%.*]] @_ZN5TimerI11ObjectCacheED1Ev([[A]] %this) unnamed_addr
+// CHECK: [[THIS1:%.*]] = call [[A]] @_ZN5TimerI11ObjectCacheED2Ev(
+// CHECK-NEXT: ret [[A]] [[THIS1]]
+
+// CHECK: define linkonce_odr [[A:%.*]] @_ZN5TimerI11ObjectCacheED2Ev([[A]] %this) unnamed_addr
+// CHECK: [[THIS1:%.*]] = call [[B:%.*]] @_ZN9TimerBaseD2Ev(
+// CHECK-NEXT: [[THIS2:%.*]] = bitcast [[B]] [[THIS1]] to [[A]]
+// CHECK-NEXT: ret [[A]] [[THIS2]]
+
+class TimerBase {
+public:
+    TimerBase();
+    virtual ~TimerBase();
+};
+
+template <typename TimerFiredClass> class Timer : public TimerBase {
+public:
+    typedef void (TimerFiredClass::*TimerFiredFunction)(Timer*);
+
+    Timer(TimerFiredClass* o, TimerFiredFunction f)
+        : m_object(o), m_function(f) { }
+
+private:
+    virtual void fired() { (m_object->*m_function)(this); }
+
+    TimerFiredClass* m_object;
+    TimerFiredFunction m_function;
+};
+
+class ObjectCache {
+public:
+    explicit ObjectCache();
+    ~ObjectCache();
+
+private:
+    Timer<ObjectCache> m_notificationPostTimer;
+};
+
+inline ObjectCache::ObjectCache() : m_notificationPostTimer(this, 0) { }
+inline ObjectCache::~ObjectCache() { }
+
+ObjectCache *test() {
+  ObjectCache *dd = new ObjectCache();
+  return dd;
+}