A few more tweaks to the blocks AST representation:

  - BlockDeclRefExprs always store VarDecls
  - BDREs no longer store copy expressions
  - BlockDecls now store a list of captured variables, information about
    how they're captured, and a copy expression if necessary
    
With that in hand, change IR generation to use the captures data in       
blocks instead of walking the block independently.        

Additionally, optimize block layout by emitting fields in descending
alignment order, with a heuristic for filling in words when alignment
of the end of the block header is insufficient for the most aligned
field.



git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@125005 91177308-0d34-0410-b5e6-96231b3b80d8

11 files changed, 1038 insertions, 937 deletions

diff --git a/lib/CodeGen/CGBlocks.cpp b/lib/CodeGen/CGBlocks.cpp
index a594a0bb9e..bdbc9d3b55 100644
--- a/lib/CodeGen/CGBlocks.cpp
+++ b/lib/CodeGen/CGBlocks.cpp
@@ -24,141 +24,84 @@
 using namespace clang;
 using namespace CodeGen;
 
-CGBlockInfo::CGBlockInfo(const char *N)
-  : Name(N), CXXThisRef(0), NeedsObjCSelf(false),
-    HasCXXObject(false) {
+CGBlockInfo::CGBlockInfo(const BlockExpr *blockExpr, const char *N)
+  : Name(N), CXXThisIndex(0), CanBeGlobal(false), NeedsCopyDispose(false),
+    HasCXXObject(false), HasWeakBlockVariable(false),
+    StructureType(0), Block(blockExpr) {
     
   // Skip asm prefix, if any.
   if (Name && Name[0] == '\01')
     ++Name;
 }
 
+/// Build the given block as a global block.
+static llvm::Constant *buildGlobalBlock(CodeGenModule &CGM,
+                                        const CGBlockInfo &blockInfo,
+                                        llvm::Constant *blockFn);
 
-llvm::Constant *CodeGenFunction::
-BuildDescriptorBlockDecl(const BlockExpr *BE, const CGBlockInfo &Info,
-                         const llvm::StructType* Ty,
-                         llvm::Constant *BlockVarLayout,
-                         std::vector<HelperInfo> *NoteForHelper) {
-  CharUnits Size = Info.BlockSize;
-  const llvm::Type *UnsignedLongTy
-    = CGM.getTypes().ConvertType(getContext().UnsignedLongTy);
-  llvm::Constant *C;
-  std::vector<llvm::Constant*> Elts;
+/// Build the helper function to copy a block.
+static llvm::Constant *buildCopyHelper(CodeGenModule &CGM,
+                                       const CGBlockInfo &blockInfo) {
+  return CodeGenFunction(CGM).GenerateCopyHelperFunction(blockInfo);
+}
+
+/// Build the helper function to dipose of a block.
+static llvm::Constant *buildDisposeHelper(CodeGenModule &CGM,
+                                          const CGBlockInfo &blockInfo) {
+  return CodeGenFunction(CGM).GenerateDestroyHelperFunction(blockInfo);
+}
+
+/// Build the block descriptor constant for a block.
+static llvm::Constant *buildBlockDescriptor(CodeGenModule &CGM,
+                                            const CGBlockInfo &blockInfo) {
+  ASTContext &C = CGM.getContext();
+
+  const llvm::Type *ulong = CGM.getTypes().ConvertType(C.UnsignedLongTy);
+  const llvm::Type *i8p = CGM.getTypes().ConvertType(C.VoidPtrTy);
+
+  llvm::SmallVector<llvm::Constant*, 6> elements;
 
   // reserved
-  C = llvm::ConstantInt::get(UnsignedLongTy, 0);
-  Elts.push_back(C);
+  elements.push_back(llvm::ConstantInt::get(ulong, 0));
 
   // Size
   // FIXME: What is the right way to say this doesn't fit?  We should give
   // a user diagnostic in that case.  Better fix would be to change the
   // API to size_t.
-  C = llvm::ConstantInt::get(UnsignedLongTy, Size.getQuantity());
-  Elts.push_back(C);
+  elements.push_back(llvm::ConstantInt::get(ulong,
+                                            blockInfo.BlockSize.getQuantity()));
 
-  // optional copy/dispose helpers
-  if (Info.BlockHasCopyDispose) {
+  // Optional copy/dispose helpers.
+  if (blockInfo.NeedsCopyDispose) {
     // copy_func_helper_decl
-    Elts.push_back(BuildCopyHelper(Ty, NoteForHelper));
+    elements.push_back(buildCopyHelper(CGM, blockInfo));
 
     // destroy_func_decl
-    Elts.push_back(BuildDestroyHelper(Ty, NoteForHelper));
+    elements.push_back(buildDisposeHelper(CGM, blockInfo));
   }
 
-  // Signature.  non-optional ObjC-style method descriptor @encode sequence
-  std::string BlockTypeEncoding;
-  CGM.getContext().getObjCEncodingForBlock(BE, BlockTypeEncoding);
-
-  Elts.push_back(llvm::ConstantExpr::getBitCast(
-          CGM.GetAddrOfConstantCString(BlockTypeEncoding), PtrToInt8Ty));
+  // Signature.  Mandatory ObjC-style method descriptor @encode sequence.
+  std::string typeAtEncoding =
+    CGM.getContext().getObjCEncodingForBlock(blockInfo.getBlockExpr());
+  elements.push_back(llvm::ConstantExpr::getBitCast(
+                          CGM.GetAddrOfConstantCString(typeAtEncoding), i8p));
   
-  // Layout.
-  C = BlockVarLayout;
-    
-  Elts.push_back(C);
-
-  C = llvm::ConstantStruct::get(VMContext, Elts, false);
-
-  C = new llvm::GlobalVariable(CGM.getModule(), C->getType(), true,
-                               llvm::GlobalValue::InternalLinkage,
-                               C, "__block_descriptor_tmp");
-  return C;
-}
-
-static void CollectBlockDeclRefInfo(const Stmt *S, CGBlockInfo &Info) {
-  for (Stmt::const_child_iterator I = S->child_begin(), E = S->child_end();
-       I != E; ++I)
-    if (*I)
-      CollectBlockDeclRefInfo(*I, Info);
-
-  // We want to ensure we walk down into block literals so we can find
-  // all nested BlockDeclRefExprs.
-  if (const BlockExpr *BE = dyn_cast<BlockExpr>(S)) {
-    Info.InnerBlocks.insert(BE->getBlockDecl());
-    CollectBlockDeclRefInfo(BE->getBody(), Info);
-  }
-
-  else if (const BlockDeclRefExpr *BDRE = dyn_cast<BlockDeclRefExpr>(S)) {
-    const ValueDecl *D = BDRE->getDecl();
-    // FIXME: Handle enums.
-    if (isa<FunctionDecl>(D))
-      return;
-
-    if (isa<ImplicitParamDecl>(D) &&
-        isa<ObjCMethodDecl>(D->getDeclContext()) &&
-        cast<ObjCMethodDecl>(D->getDeclContext())->getSelfDecl() == D) {
-      Info.NeedsObjCSelf = true;
-      return;
-    }
-
-    // Only Decls that escape are added.
-    if (!Info.InnerBlocks.count(D->getDeclContext())) {
-      if (BDRE->getCopyConstructorExpr())
-        Info.HasCXXObject = true;
-      Info.DeclRefs.push_back(BDRE);
-    }
-  }
+  // GC layout.
+  if (C.getLangOptions().ObjC1)
+    elements.push_back(CGM.getObjCRuntime().BuildGCBlockLayout(CGM, blockInfo));
+  else
+    elements.push_back(llvm::Constant::getNullValue(i8p));
 
-  // Make sure to capture implicit 'self' references due to super calls.
-  else if (const ObjCMessageExpr *E = dyn_cast<ObjCMessageExpr>(S)) {
-    if (E->getReceiverKind() == ObjCMessageExpr::SuperClass || 
-        E->getReceiverKind() == ObjCMessageExpr::SuperInstance)
-      Info.NeedsObjCSelf = true;
-  }
-  else if (const ObjCPropertyRefExpr *PE = dyn_cast<ObjCPropertyRefExpr>(S)) {
-    // Getter/setter uses may also cause implicit super references,
-    // which we can check for with:
-    if (PE->isSuperReceiver())
-      Info.NeedsObjCSelf = true;
-  }
-  else if (isa<CXXThisExpr>(S))
-    Info.CXXThisRef = cast<CXXThisExpr>(S);
-}
+  llvm::Constant *init =
+    llvm::ConstantStruct::get(CGM.getLLVMContext(), elements.data(),
+                              elements.size(), false);
 
-/// CanBlockBeGlobal - Given a CGBlockInfo struct, determines if a block can be
-/// declared as a global variable instead of on the stack.
-static bool CanBlockBeGlobal(const CGBlockInfo &Info) {
-  return Info.DeclRefs.empty();
-}
+  llvm::GlobalVariable *global =
+    new llvm::GlobalVariable(CGM.getModule(), init->getType(), true,
+                             llvm::GlobalValue::InternalLinkage,
+                             init, "__block_descriptor_tmp");
 
-/// AllocateAllBlockDeclRefs - Preallocate all nested BlockDeclRefExprs to
-/// ensure we can generate the debug information for the parameter for the block
-/// invoke function.
-static void AllocateAllBlockDeclRefs(CodeGenFunction &CGF, CGBlockInfo &Info) {
-  if (Info.CXXThisRef)
-    CGF.AllocateBlockCXXThisPointer(Info.CXXThisRef);
-
-  for (size_t i = 0; i < Info.DeclRefs.size(); ++i)
-    CGF.AllocateBlockDecl(Info.DeclRefs[i]);
-
-  if (Info.NeedsObjCSelf) {
-    ValueDecl *Self = cast<ObjCMethodDecl>(CGF.CurFuncDecl)->getSelfDecl();
-    BlockDeclRefExpr *BDRE =
-      new (CGF.getContext()) BlockDeclRefExpr(Self, Self->getType(), VK_RValue,
-                                              SourceLocation(), false);
-    Info.DeclRefs.push_back(BDRE);
-    CGF.AllocateBlockDecl(BDRE);
-  }
+  return llvm::ConstantExpr::getBitCast(global, CGM.getBlockDescriptorType());
 }
 
 static unsigned computeBlockFlag(CodeGenModule &CGM,
@@ -176,256 +119,506 @@ static unsigned computeBlockFlag(CodeGenModule &CGM,
   return flags;
 }
 
-// FIXME: Push most into CGM, passing down a few bits, like current function
-// name.
-llvm::Value *CodeGenFunction::BuildBlockLiteralTmp(const BlockExpr *BE) {
-  std::string Name = CurFn->getName();
-  CGBlockInfo Info(Name.c_str());
-  Info.InnerBlocks.insert(BE->getBlockDecl());
-  CollectBlockDeclRefInfo(BE->getBody(), Info);
+/*
+  Purely notional variadic template describing the layout of a block.
+
+  template <class _ResultType, class... _ParamTypes, class... _CaptureTypes>
+  struct Block_literal {
+    /// Initialized to one of:
+    ///   extern void *_NSConcreteStackBlock[];
+    ///   extern void *_NSConcreteGlobalBlock[];
+    ///
+    /// In theory, we could start one off malloc'ed by setting
+    /// BLOCK_NEEDS_FREE, giving it a refcount of 1, and using
+    /// this isa:
+    ///   extern void *_NSConcreteMallocBlock[];
+    struct objc_class *isa;
+
+    /// These are the flags (with corresponding bit number) that the
+    /// compiler is actually supposed to know about.
+    ///  25. BLOCK_HAS_COPY_DISPOSE - indicates that the block
+    ///   descriptor provides copy and dispose helper functions
+    ///  26. BLOCK_HAS_CXX_OBJ - indicates that there's a captured
+    ///   object with a nontrivial destructor or copy constructor
+    ///  28. BLOCK_IS_GLOBAL - indicates that the block is allocated
+    ///   as global memory
+    ///  29. BLOCK_USE_STRET - indicates that the block function
+    ///   uses stret, which objc_msgSend needs to know about
+    ///  30. BLOCK_HAS_SIGNATURE - indicates that the block has an
+    ///   @encoded signature string
+    /// And we're not supposed to manipulate these:
+    ///  24. BLOCK_NEEDS_FREE - indicates that the block has been moved
+    ///   to malloc'ed memory
+    ///  27. BLOCK_IS_GC - indicates that the block has been moved to
+    ///   to GC-allocated memory
+    /// Additionally, the bottom 16 bits are a reference count which
+    /// should be zero on the stack.
+    int flags;
+
+    /// Reserved;  should be zero-initialized.
+    int reserved;
+
+    /// Function pointer generated from block literal.
+    _ResultType (*invoke)(Block_literal *, _ParamTypes...);
+
+    /// Block description metadata generated from block literal.
+    struct Block_descriptor *block_descriptor;
+
+    /// Captured values follow.
+    _CapturesTypes captures...;
+  };
+ */
+
+/// The number of fields in a block header.
+const unsigned BlockHeaderSize = 5;
+
+namespace {
+  /// A chunk of data that we actually have to capture in the block.
+  struct BlockLayoutChunk {
+    CharUnits Alignment;
+    CharUnits Size;
+    const BlockDecl::Capture *Capture; // null for 'this'
+    const llvm::Type *Type;
+
+    BlockLayoutChunk(CharUnits align, CharUnits size,
+                     const BlockDecl::Capture *capture,
+                     const llvm::Type *type)
+      : Alignment(align), Size(size), Capture(capture), Type(type) {}
+
+    /// Tell the block info that this chunk has the given field index.
+    void setIndex(CGBlockInfo &info, unsigned index) {
+      if (!Capture)
+        info.CXXThisIndex = index;
+      else
+        info.Captures[Capture->getVariable()]
+          = CGBlockInfo::Capture::makeIndex(index);
+    }
+  };
 
-  // Check if the block can be global.
-  // FIXME: This test doesn't work for nested blocks yet.  Longer term, I'd like
-  // to just have one code path.  We should move this function into CGM and pass
-  // CGF, then we can just check to see if CGF is 0.
-  if (0 && CanBlockBeGlobal(Info))
-    return CGM.GetAddrOfGlobalBlock(BE, Name.c_str());
+  /// Order by descending alignment.
+  bool operator<(const BlockLayoutChunk &left, const BlockLayoutChunk &right) {
+    return left.Alignment > right.Alignment;
+  }
+}
 
-  size_t BlockFields = 5;
+/// It is illegal to modify a const object after initialization.
+/// Therefore, if a const object has a constant initializer, we don't
+/// actually need to keep storage for it in the block; we'll just
+/// rematerialize it at the start of the block function.  This is
+/// acceptable because we make no promises about address stability of
+/// captured variables.
+static llvm::Constant *tryCaptureAsConstant(CodeGenModule &CGM,
+                                            const VarDecl *var) {
+  QualType type = var->getType();
+
+  // We can only do this if the variable is const.
+  if (!type.isConstQualified()) return 0;
+
+  // Furthermore, in C++ we can't do this for classes.  TODO: we might
+  // actually be able to get away with it for classes with a trivial
+  // destructor and a trivial copy constructor and no mutable fields.
+  if (CGM.getLangOptions().CPlusPlus &&
+      type->getBaseElementTypeUnsafe()->isRecordType())
+    return 0;
+
+  // If the variable doesn't have any initializer (shouldn't this be
+  // invalid?), it's not clear what we should do.  Maybe capture as
+  // zero?
+  const Expr *init = var->getInit();
+  if (!init) return 0;
+
+  return CGM.EmitConstantExpr(init, var->getType());
+}
 
-  std::vector<llvm::Constant*> Elts(BlockFields);
+/// Get the low bit of a nonzero character count.  This is the
+/// alignment of the nth byte if the 0th byte is universally aligned.
+static CharUnits getLowBit(CharUnits v) {
+  return CharUnits::fromQuantity(v.getQuantity() & (~v.getQuantity() + 1));
+}
 
-  llvm::Constant *C;
-  llvm::Value *V;
+static void initializeForBlockHeader(CodeGenModule &CGM, CGBlockInfo &info,
+                                std::vector<const llvm::Type*> &elementTypes) {
+  ASTContext &C = CGM.getContext();
 
-  bool hasWeakBlockVariable = false;
+  // The header is basically a 'struct { void *; int; int; void *; void *; }'.
+  CharUnits ptrSize, ptrAlign, intSize, intAlign;
+  llvm::tie(ptrSize, ptrAlign) = C.getTypeInfoInChars(C.VoidPtrTy);
+  llvm::tie(intSize, intAlign) = C.getTypeInfoInChars(C.IntTy);
 
-  {
-    llvm::Constant *BlockVarLayout;
-    // C = BuildBlockStructInitlist();
-    unsigned int flags = BLOCK_HAS_SIGNATURE;
+  // Are there crazy embedded platforms where this isn't true?
+  assert(intSize <= ptrSize && "layout assumptions horribly violated");
 
-    // We run this first so that we set BlockHasCopyDispose from the entire
-    // block literal.
-    // __invoke
-    llvm::Function *Fn
-      = CodeGenFunction(CGM).GenerateBlockFunction(CurGD, BE, Info, CurFuncDecl,
-                                                   BlockVarLayout,
-                                                   LocalDeclMap);
-    SynthesizeCopyDisposeHelpers |= Info.BlockHasCopyDispose;
-    Elts[3] = Fn;
-
-    // FIXME: Don't use BlockHasCopyDispose, it is set more often then
-    // necessary, for example: { ^{ __block int i; ^{ i = 1; }(); }(); }
-    if (Info.BlockHasCopyDispose)
-      flags |= BLOCK_HAS_COPY_DISPOSE;
-    // This block may import a c++ object.
-    if (Info.HasCXXObject)
-      flags |= BLOCK_HAS_CXX_OBJ;
-
-    // __isa
-    C = CGM.getNSConcreteStackBlock();
-    C = llvm::ConstantExpr::getBitCast(C, PtrToInt8Ty);
-    Elts[0] = C;
-
-    // __flags
-    flags = computeBlockFlag(CGM, BE, flags);
-    const llvm::IntegerType *IntTy = cast<llvm::IntegerType>(
-      CGM.getTypes().ConvertType(CGM.getContext().IntTy));
-    C = llvm::ConstantInt::get(IntTy, flags);
-    Elts[1] = C;
-
-    // __reserved
-    C = llvm::ConstantInt::get(IntTy, 0);
-    Elts[2] = C;
-
-    if (Info.BlockLayout.empty()) {
-      // __descriptor
-      C = llvm::Constant::getNullValue(PtrToInt8Ty);
-      Elts[4] = BuildDescriptorBlockDecl(BE, Info, 0, C, 0);
-
-      // Optimize to being a global block.
-      Elts[0] = CGM.getNSConcreteGlobalBlock();
-      
-      Elts[1] = llvm::ConstantInt::get(IntTy, flags|BLOCK_IS_GLOBAL);
-
-      C = llvm::ConstantStruct::get(VMContext, Elts, false);
-
-      C = new llvm::GlobalVariable(CGM.getModule(), C->getType(), true,
-                                   llvm::GlobalValue::InternalLinkage, C,
-                                   "__block_holder_tmp_" +
-                                   llvm::Twine(CGM.getGlobalUniqueCount()));
-      QualType BPT = BE->getType();
-      C = llvm::ConstantExpr::getBitCast(C, ConvertType(BPT));
-      return C;
-    }
+  CharUnits headerSize = ptrSize;
+  if (2 * intSize < ptrAlign) headerSize += ptrSize;
+  else headerSize += 2 * intSize;
+  headerSize += 2 * ptrSize;
 
-    std::vector<const llvm::Type *> Types(BlockFields+Info.BlockLayout.size());
-    for (int i=0; i<4; ++i)
-      Types[i] = Elts[i]->getType();
-    Types[4] = PtrToInt8Ty;
-
-    for (unsigned i = 0, n = Info.BlockLayout.size(); i != n; ++i) {
-      const Expr *E = Info.BlockLayout[i];
-      const BlockDeclRefExpr *BDRE = dyn_cast<BlockDeclRefExpr>(E);
-      QualType Ty = E->getType();
-      if (BDRE && BDRE->isByRef()) {
-        if (BDRE->getDecl()->getType().isObjCGCWeak())
-          hasWeakBlockVariable = true;
-        Types[i+BlockFields] = 
-          llvm::PointerType::get(BuildByRefType(BDRE->getDecl()), 0);
-      } else if (BDRE && BDRE->getDecl()->getType()->isReferenceType()) {
-         Types[i+BlockFields] = llvm::PointerType::get(ConvertType(Ty), 0);
-      } else 
-        Types[i+BlockFields] = ConvertType(Ty);
-    }
+  info.BlockAlign = ptrAlign;
+  info.BlockSize = headerSize;
 
-    llvm::StructType *Ty = llvm::StructType::get(VMContext, Types, true);
+  assert(elementTypes.empty());
+  const llvm::Type *i8p = CGM.getTypes().ConvertType(C.VoidPtrTy);
+  const llvm::Type *intTy = CGM.getTypes().ConvertType(C.IntTy);
+  elementTypes.push_back(i8p);
+  elementTypes.push_back(intTy);
+  elementTypes.push_back(intTy);
+  elementTypes.push_back(i8p);
+  elementTypes.push_back(CGM.getBlockDescriptorType());
 
-    llvm::AllocaInst *A = CreateTempAlloca(Ty);
-    A->setAlignment(Info.BlockAlign.getQuantity());
-    V = A;
+  assert(elementTypes.size() == BlockHeaderSize);
+}
 
-    // Build layout / cleanup information for all the data entries in the
-    // layout, and write the enclosing fields into the type.
-    std::vector<HelperInfo> NoteForHelper(Info.BlockLayout.size());
-    unsigned NumHelpers = 0;
+/// Compute the layout of the given block.  Attempts to lay the block
+/// out with minimal space requirements.
+static void computeBlockInfo(CodeGenModule &CGM, CGBlockInfo &info) {
+  ASTContext &C = CGM.getContext();
+  const BlockDecl *block = info.getBlockDecl();
 
-    for (unsigned i=0; i<4; ++i)
-      Builder.CreateStore(Elts[i], Builder.CreateStructGEP(V, i, "block.tmp"));
+  std::vector<const llvm::Type*> elementTypes;
+  initializeForBlockHeader(CGM, info, elementTypes);
 
-    for (unsigned i=0; i < Info.BlockLayout.size(); ++i) {
-      const Expr *E = Info.BlockLayout[i];
+  if (!block->hasCaptures()) {
+    info.StructureType =
+      llvm::StructType::get(CGM.getLLVMContext(), elementTypes, true);
+    info.CanBeGlobal = true;
+    return;
+  }
 
-      // Skip padding.
-      if (isa<DeclRefExpr>(E)) continue;
+  // Collect the layout chunks.
+  llvm::SmallVector<BlockLayoutChunk, 16> layout;
+  layout.reserve(block->capturesCXXThis() +
+                 (block->capture_end() - block->capture_begin()));
 
-      llvm::Value* Addr = Builder.CreateStructGEP(V, i+BlockFields, "tmp");
-      HelperInfo &Note = NoteForHelper[NumHelpers++];
+  CharUnits maxFieldAlign;
 
-      Note.index = i+5;
+  // First, 'this'.
+  if (block->capturesCXXThis()) {
+    const DeclContext *DC = block->getDeclContext();
+    for (; isa<BlockDecl>(DC); DC = cast<BlockDecl>(DC)->getDeclContext())
+      ;
+    QualType thisType = cast<CXXMethodDecl>(DC)->getThisType(C);
 
-      if (isa<CXXThisExpr>(E)) {
-        Note.RequiresCopying = false;
-        Note.flag = BLOCK_FIELD_IS_OBJECT;
+    const llvm::Type *llvmType = CGM.getTypes().ConvertType(thisType);
+    std::pair<CharUnits,CharUnits> tinfo
+      = CGM.getContext().getTypeInfoInChars(thisType);
+    maxFieldAlign = std::max(maxFieldAlign, tinfo.second);
 
-        Builder.CreateStore(LoadCXXThis(), Addr);
-        continue;
-      }
+    layout.push_back(BlockLayoutChunk(tinfo.second, tinfo.first, 0, llvmType));
+  }
 
-      const BlockDeclRefExpr *BDRE = cast<BlockDeclRefExpr>(E);
-      Note.RequiresCopying = BlockRequiresCopying(BDRE);
-      Note.cxxvar_import = 
-        BDRE->getCopyConstructorExpr() ? BDRE : 0;
-      const ValueDecl *VD = BDRE->getDecl();
-      QualType T = VD->getType();
-
-      if (BDRE->isByRef()) {
-        Note.flag = BLOCK_FIELD_IS_BYREF;
-        if (T.isObjCGCWeak())
-          Note.flag |= BLOCK_FIELD_IS_WEAK;
-      } else if (T->isBlockPointerType()) {
-        Note.flag = BLOCK_FIELD_IS_BLOCK;
-      } else {
-        Note.flag = BLOCK_FIELD_IS_OBJECT;
-      }
+  // Next, all the block captures.
+  for (BlockDecl::capture_const_iterator ci = block->capture_begin(),
+         ce = block->capture_end(); ci != ce; ++ci) {
+    const VarDecl *variable = ci->getVariable();
 
-      if (LocalDeclMap[VD]) {
-        if (BDRE->isByRef()) {
-          llvm::Value *Loc = LocalDeclMap[VD];
-          Loc = Builder.CreateStructGEP(Loc, 1, "forwarding");
-          Loc = Builder.CreateLoad(Loc);
-          Builder.CreateStore(Loc, Addr);
-          continue;
-        } else {
-          if (BDRE->getCopyConstructorExpr()) {
-            E = BDRE->getCopyConstructorExpr();
-            PushDestructorCleanup(E->getType(), Addr);
-          } else {
-            E = new (getContext()) DeclRefExpr(const_cast<ValueDecl*>(VD),
-                                 VD->getType().getNonReferenceType(),
-                                 Expr::getValueKindForType(VD->getType()),
-                                 SourceLocation());
-            if (VD->getType()->isReferenceType()) {
-              E = new (getContext())
-                UnaryOperator(const_cast<Expr*>(E), UO_AddrOf,
-                              getContext().getPointerType(E->getType()),
-                              VK_RValue, OK_Ordinary, SourceLocation());
-            }
-          }
+    if (ci->isByRef()) {
+      // We have to copy/dispose of the __block reference.
+      info.NeedsCopyDispose = true;
+
+      // Also note that it's weak for GC purposes.
+      if (variable->getType().isObjCGCWeak())
+        info.HasWeakBlockVariable = true;
+
+      // Just use void* instead of a pointer to the byref type.
+      QualType byRefPtrTy = C.VoidPtrTy;
+
+      const llvm::Type *llvmType = CGM.getTypes().ConvertType(byRefPtrTy);
+      std::pair<CharUnits,CharUnits> tinfo
+        = CGM.getContext().getTypeInfoInChars(byRefPtrTy);
+      maxFieldAlign = std::max(maxFieldAlign, tinfo.second);
+
+      layout.push_back(BlockLayoutChunk(tinfo.second, tinfo.first,
+                                        &*ci, llvmType));
+      continue;
+    }
+
+    // Otherwise, build a layout chunk with the size and alignment of
+    // the declaration.
+    if (llvm::Constant *constant = tryCaptureAsConstant(CGM, variable)) {
+      info.Captures[variable] = CGBlockInfo::Capture::makeConstant(constant);
+      continue;
+    }
+
+    // Block pointers require copy/dispose.
+    if (variable->getType()->isBlockPointerType()) {
+      info.NeedsCopyDispose = true;
+
+    // So do Objective-C pointers.
+    } else if (variable->getType()->isObjCObjectPointerType() ||
+               C.isObjCNSObjectType(variable->getType())) {
+      info.NeedsCopyDispose = true;
+
+    // So do types that require non-trivial copy construction.
+    } else if (ci->hasCopyExpr()) {
+      info.NeedsCopyDispose = true;
+      info.HasCXXObject = true;
+
+    // And so do types with destructors.
+    } else if (CGM.getLangOptions().CPlusPlus) {
+      if (const CXXRecordDecl *record =
+            variable->getType()->getAsCXXRecordDecl()) {
+        if (!record->hasTrivialDestructor()) {
+          info.HasCXXObject = true;
+          info.NeedsCopyDispose = true;
         }
       }
+    }
+
+    CharUnits size = C.getTypeSizeInChars(variable->getType());
+    CharUnits align = C.getDeclAlign(variable);
+    maxFieldAlign = std::max(maxFieldAlign, align);
+
+    const llvm::Type *llvmType =
+      CGM.getTypes().ConvertTypeForMem(variable->getType());
+
+    layout.push_back(BlockLayoutChunk(align, size, &*ci, llvmType));
+  }
 
-      if (BDRE->isByRef()) {
-        E = new (getContext())
-          UnaryOperator(const_cast<Expr*>(E), UO_AddrOf,
-                        getContext().getPointerType(E->getType()),
-                        VK_RValue, OK_Ordinary, SourceLocation());
+  // If that was everything, we're done here.
+  if (layout.empty()) {
+    info.StructureType =
+      llvm::StructType::get(CGM.getLLVMContext(), elementTypes, true);
+    info.CanBeGlobal = true;
+    return;
+  }
+
+  // Sort the layout by alignment.  We have to use a stable sort here
+  // to get reproducible results.  There should probably be an
+  // llvm::array_pod_stable_sort.
+  std::stable_sort(layout.begin(), layout.end());
+
+  CharUnits &blockSize = info.BlockSize;
+  info.BlockAlign = std::max(maxFieldAlign, info.BlockAlign);
+
+  // Assuming that the first byte in the header is maximally aligned,
+  // get the alignment of the first byte following the header.
+  CharUnits endAlign = getLowBit(blockSize);
+
+  // If the end of the header isn't satisfactorily aligned for the
+  // maximum thing, look for things that are okay with the header-end
+  // alignment, and keep appending them until we get something that's
+  // aligned right.  This algorithm is only guaranteed optimal if
+  // that condition is satisfied at some point; otherwise we can get
+  // things like:
+  //   header                 // next byte has alignment 4
+  //   something_with_size_5; // next byte has alignment 1
+  //   something_with_alignment_8;
+  // which has 7 bytes of padding, as opposed to the naive solution
+  // which might have less (?).
+  if (endAlign < maxFieldAlign) {
+    llvm::SmallVectorImpl<BlockLayoutChunk>::iterator
+      li = layout.begin() + 1, le = layout.end();
+
+    // Look for something that the header end is already
+    // satisfactorily aligned for.
+    for (; li != le && endAlign < li->Alignment; ++li)
+      ;
+
+    // If we found something that's naturally aligned for the end of
+    // the header, keep adding things...
+    if (li != le) {
+      llvm::SmallVectorImpl<BlockLayoutChunk>::iterator first = li;
+      for (; li != le; ++li) {
+        assert(endAlign >= li->Alignment);
+
+        li->setIndex(info, elementTypes.size());
+        elementTypes.push_back(li->Type);
+        blockSize += li->Size;
+        endAlign = getLowBit(blockSize);
+
+        // ...until we get to the alignment of the maximum field.
+        if (endAlign >= maxFieldAlign)
+          break;
       }
 
-      RValue r = EmitAnyExpr(E, AggValueSlot::forAddr(Addr, false, true));
-      if (r.isScalar()) {
-        llvm::Value *Loc = r.getScalarVal();
-        const llvm::Type *Ty = Types[i+BlockFields];
-        if  (BDRE->isByRef()) {
-          // E is now the address of the value field, instead, we want the
-          // address of the actual ByRef struct.  We optimize this slightly
-          // compared to gcc by not grabbing the forwarding slot as this must
-          // be done during Block_copy for us, and we can postpone the work
-          // until then.
-          CharUnits offset = BlockDecls[BDRE->getDecl()];
-
-          llvm::Value *BlockLiteral = LoadBlockStruct();
-
-          Loc = Builder.CreateGEP(BlockLiteral,
-                     llvm::ConstantInt::get(Int64Ty, offset.getQuantity()),
-                                  "block.literal");
-          Ty = llvm::PointerType::get(Ty, 0);
-          Loc = Builder.CreateBitCast(Loc, Ty);
-          Loc = Builder.CreateLoad(Loc);
-          // Loc = Builder.CreateBitCast(Loc, Ty);
-        }
-        Builder.CreateStore(Loc, Addr);
-      } else if (r.isComplex())
-        // FIXME: implement
-        ErrorUnsupported(BE, "complex in block literal");
-      else if (r.isAggregate())
-        ; // Already created into the destination
+      // Don't re-append everything we just appended.
+      layout.erase(first, li);
+    }
+  }
+
+  // At this point, we just have to add padding if the end align still
+  // isn't aligned right.
+  if (endAlign < maxFieldAlign) {
+    CharUnits padding = maxFieldAlign - endAlign;
+
+    const llvm::Type *i8 = llvm::IntegerType::get(CGM.getLLVMContext(), 8);
+    elementTypes.push_back(llvm::ArrayType::get(i8, padding.getQuantity()));
+    blockSize += padding;
+
+    endAlign = getLowBit(blockSize);
+    assert(endAlign >= maxFieldAlign);
+  }
+
+  // Slam everything else on now.  This works because they have
+  // strictly decreasing alignment and we expect that size is always a
+  // multiple of alignment.
+  for (llvm::SmallVectorImpl<BlockLayoutChunk>::iterator
+         li = layout.begin(), le = layout.end(); li != le; ++li) {
+    assert(endAlign >= li->Alignment);
+    li->setIndex(info, elementTypes.size());
+    elementTypes.push_back(li->Type);
+    blockSize += li->Size;
+    endAlign = getLowBit(blockSize);
+  }
+
+  info.StructureType =
+    llvm::StructType::get(CGM.getLLVMContext(), elementTypes, true);
+}
+
+/// Emit a block literal expression in the current function.
+llvm::Value *CodeGenFunction::EmitBlockLiteral(const BlockExpr *blockExpr) {
+  std::string Name = CurFn->getName();
+  CGBlockInfo blockInfo(blockExpr, Name.c_str());
+
+  // Compute information about the layout, etc., of this block.
+  computeBlockInfo(CGM, blockInfo);
+
+  // Using that metadata, generate the actual block function.
+  llvm::Constant *blockFn
+    = CodeGenFunction(CGM).GenerateBlockFunction(CurGD, blockInfo,
+                                                 CurFuncDecl, LocalDeclMap);
+  blockFn = llvm::ConstantExpr::getBitCast(blockFn, PtrToInt8Ty);
+
+  // If there is nothing to capture, we can emit this as a global block.
+  if (blockInfo.CanBeGlobal)
+    return buildGlobalBlock(CGM, blockInfo, blockFn);
+
+  // Otherwise, we have to emit this as a local block.
+
+  llvm::Constant *isa = CGM.getNSConcreteStackBlock();
+  isa = llvm::ConstantExpr::getBitCast(isa, PtrToInt8Ty);
+
+  // Build the block descriptor.
+  llvm::Constant *descriptor = buildBlockDescriptor(CGM, blockInfo);
+
+  const llvm::Type *intTy = ConvertType(getContext().IntTy);
+
+  llvm::AllocaInst *blockAddr =
+    CreateTempAlloca(blockInfo.StructureType, "block");
+  blockAddr->setAlignment(blockInfo.BlockAlign.getQuantity());
+
+  // Compute the initial on-stack block flags.
+  unsigned int flags = BLOCK_HAS_SIGNATURE;
+  if (blockInfo.NeedsCopyDispose) flags |= BLOCK_HAS_COPY_DISPOSE;
+  if (blockInfo.HasCXXObject) flags |= BLOCK_HAS_CXX_OBJ;
+  flags = computeBlockFlag(CGM, blockInfo.getBlockExpr(), flags);
+
+  // Initialize the block literal.
+  Builder.CreateStore(isa, Builder.CreateStructGEP(blockAddr, 0, "block.isa"));
+  Builder.CreateStore(llvm::ConstantInt::get(intTy, flags),
+                      Builder.CreateStructGEP(blockAddr, 1, "block.flags"));
+  Builder.CreateStore(llvm::ConstantInt::get(intTy, 0),
+                      Builder.CreateStructGEP(blockAddr, 2, "block.reserved"));
+  Builder.CreateStore(blockFn, Builder.CreateStructGEP(blockAddr, 3,
+                                                       "block.invoke"));
+  Builder.CreateStore(descriptor, Builder.CreateStructGEP(blockAddr, 4,
+                                                          "block.descriptor"));
+
+  // Finally, capture all the values into the block.
+  const BlockDecl *blockDecl = blockInfo.getBlockDecl();
+
+  // First, 'this'.
+  if (blockDecl->capturesCXXThis()) {
+    llvm::Value *addr = Builder.CreateStructGEP(blockAddr,
+                                                blockInfo.CXXThisIndex,
+                                                "block.captured-this.addr");
+    Builder.CreateStore(LoadCXXThis(), addr);
+  }
+
+  // Next, captured variables.
+  for (BlockDecl::capture_const_iterator ci = blockDecl->capture_begin(),
+         ce = blockDecl->capture_end(); ci != ce; ++ci) {
+    const VarDecl *variable = ci->getVariable();
+    const CGBlockInfo::Capture &capture = blockInfo.getCapture(variable);
+
+    // Ignore constant captures.
+    if (capture.isConstant()) continue;
+
+    QualType type = variable->getType();
+
+    // This will be a [[type]]*, except that a byref entry will just be
+    // an i8**.
+    llvm::Value *blockField =
+      Builder.CreateStructGEP(blockAddr, capture.getIndex(),
+                              "block.captured");
+
+    // Compute the address of the thing we're going to move into the
+    // block literal.
+    llvm::Value *src;
+    if (ci->isNested()) {
+      // We need to use the capture from the enclosing block.
+      const CGBlockInfo::Capture &enclosingCapture =
+        BlockInfo->getCapture(variable);
+
+      // This is a [[type]]*, except that a byref entry wil just be an i8**.
+      src = Builder.CreateStructGEP(LoadBlockStruct(),
+                                    enclosingCapture.getIndex(),
+                                    "block.capture.addr");
+    } else {
+      // This is a [[type]]*.
+      src = LocalDeclMap[variable];
+    }
+
+    // For byrefs, we just write the pointer to the byref struct into
+    // the block field.  There's no need to chase the forwarding
+    // pointer at this point, since we're building something that will
+    // live a shorter life than the stack byref anyway.
+    if (ci->isByRef()) {
+      // Get an i8* that points to the byref struct.
+      if (ci->isNested())
+        src = Builder.CreateLoad(src, "byref.capture");
       else
-        assert (0 && "bad block variable");
-      // FIXME: Ensure that the offset created by the backend for
-      // the struct matches the previously computed offset in BlockDecls.
+        src = Builder.CreateBitCast(src, PtrToInt8Ty);
+
+      // Write that i8* into the capture field.
+      Builder.CreateStore(src, blockField);
+
+    // If we have a copy constructor, evaluate that into the block field.
+    } else if (const Expr *copyExpr = ci->getCopyExpr()) {
+      EmitSynthesizedCXXCopyCtor(blockField, src, copyExpr);
+
+    // If it's a reference variable, copy the reference into the block field.
+    } else if (type->isRefere