Thread Safety Analysis: major update to thread safety TIL.

Numerous changes, including:
  * Changed the way variables and instructions are handled in basic blocks to
    be more efficient.
  * Eliminated SExprRef.
  * Simplified futures.
  * Fixed documentation.
  * Compute dominator and post dominator trees.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@217556 91177308-0d34-0410-b5e6-96231b3b80d8

include/clang/Analysis/Analyses/ThreadSafetyCommon.h

@@ -477,9 +477,9 @@ private:
                                            // Indexed by clang BlockID.
 
   LVarDefinitionMap CurrentLVarMap;
-  std::vector<til::Variable*> CurrentArguments;
-  std::vector<til::Variable*> CurrentInstructions;
-  std::vector<til::Variable*> IncompleteArgs;
+  std::vector<til::Phi*>   CurrentArguments;
+  std::vector<til::SExpr*> CurrentInstructions;
+  std::vector<til::Phi*>   IncompleteArgs;
   til::BasicBlock *CurrentBB;
   BlockInfo *CurrentBlockInfo;
 };

include/clang/Analysis/Analyses/ThreadSafetyLogical.h

@@ -41,13 +41,13 @@ private:
 };
 
 class Terminal : public LExpr {
-  til::SExprRef Expr;
+  til::SExpr *Expr;
 
 public:
   Terminal(til::SExpr *Expr) : LExpr(LExpr::Terminal), Expr(Expr) {}
 
-  const til::SExpr *expr() const { return Expr.get(); }
-  til::SExpr *expr() { return Expr.get(); }
+  const til::SExpr *expr() const { return Expr; }
+  til::SExpr *expr() { return Expr; }
 
   static bool classof(const LExpr *E) { return E->kind() == LExpr::Terminal; }
 };

include/clang/Analysis/Analyses/ThreadSafetyOps.def

@@ -44,8 +44,11 @@ TIL_OPCODE_DEF(Cast)
 TIL_OPCODE_DEF(SCFG)
 TIL_OPCODE_DEF(BasicBlock)
 TIL_OPCODE_DEF(Phi)
+
+// Terminator instructions
 TIL_OPCODE_DEF(Goto)
 TIL_OPCODE_DEF(Branch)
+TIL_OPCODE_DEF(Return)
 
 // pseudo-terms
 TIL_OPCODE_DEF(Identifier)

include/clang/Analysis/Analyses/ThreadSafetyTraverse.h

@@ -58,11 +58,16 @@ public:
   // Traverse an expression -- returning a result of type R_SExpr.
   // Override this method to do something for every expression, regardless
   // of which kind it is.
-  typename R::R_SExpr traverse(SExprRef &E, typename R::R_Ctx Ctx) {
-    return traverse(E.get(), Ctx);
+  // E is a reference, so this can be use for in-place updates.
+  // The type T must be a subclass of SExpr.
+  template <class T>
+  typename R::R_SExpr traverse(T* &E, typename R::R_Ctx Ctx) {
+    return traverseSExpr(E, Ctx);
   }
 
-  typename R::R_SExpr traverse(SExpr *E, typename R::R_Ctx Ctx) {
+  // Override this method to do something for every expression.
+  // Does not allow in-place updates.
+  typename R::R_SExpr traverseSExpr(SExpr *E, typename R::R_Ctx Ctx) {
     return traverseByCase(E, Ctx);
   }
 
@@ -75,6 +80,7 @@ public:
 #include "ThreadSafetyOps.def"
 #undef TIL_OPCODE_DEF
     }
+    return self()->reduceNull();
   }
 
 // Traverse e, by static dispatch on the type "X" of e.
@@ -92,10 +98,10 @@ public:
 class SimpleReducerBase {
 public:
   enum TraversalKind {
-    TRV_Normal,
-    TRV_Decl,
-    TRV_Lazy,
-    TRV_Type
+    TRV_Normal,   // ordinary subexpressions
+    TRV_Decl,     // declarations (e.g. function bodies)
+    TRV_Lazy,     // expressions that require lazy evaluation
+    TRV_Type      // type expressions
   };
 
   // R_Ctx defines a "context" for the traversal, which encodes information
@@ -147,153 +153,6 @@ protected:
 };
 
 
-// Implements a traversal that makes a deep copy of an SExpr.
-// The default behavior of reduce##X(...) is to create a copy of the original.
-// Subclasses can override reduce##X to implement non-destructive rewriting
-// passes.
-template<class Self>
-class CopyReducer : public Traversal<Self, CopyReducerBase>,
-                    public CopyReducerBase {
-public:
-  CopyReducer(MemRegionRef A) : CopyReducerBase(A) {}
-
-public:
-  R_SExpr reduceNull() {
-    return nullptr;
-  }
-  // R_SExpr reduceFuture(...)  is never used.
-
-  R_SExpr reduceUndefined(Undefined &Orig) {
-    return new (Arena) Undefined(Orig);
-  }
-  R_SExpr reduceWildcard(Wildcard &Orig) {
-    return new (Arena) Wildcard(Orig);
-  }
-
-  R_SExpr reduceLiteral(Literal &Orig) {
-    return new (Arena) Literal(Orig);
-  }
-  template<class T>
-  R_SExpr reduceLiteralT(LiteralT<T> &Orig) {
-    return new (Arena) LiteralT<T>(Orig);
-  }
-  R_SExpr reduceLiteralPtr(LiteralPtr &Orig) {
-    return new (Arena) LiteralPtr(Orig);
-  }
-
-  R_SExpr reduceFunction(Function &Orig, Variable *Nvd, R_SExpr E0) {
-    return new (Arena) Function(Orig, Nvd, E0);
-  }
-  R_SExpr reduceSFunction(SFunction &Orig, Variable *Nvd, R_SExpr E0) {
-    return new (Arena) SFunction(Orig, Nvd, E0);
-  }
-  R_SExpr reduceCode(Code &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) Code(Orig, E0, E1);
-  }
-  R_SExpr reduceField(Field &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) Field(Orig, E0, E1);
-  }
-
-  R_SExpr reduceApply(Apply &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) Apply(Orig, E0, E1);
-  }
-  R_SExpr reduceSApply(SApply &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) SApply(Orig, E0, E1);
-  }
-  R_SExpr reduceProject(Project &Orig, R_SExpr E0) {
-    return new (Arena) Project(Orig, E0);
-  }
-  R_SExpr reduceCall(Call &Orig, R_SExpr E0) {
-    return new (Arena) Call(Orig, E0);
-  }
-
-  R_SExpr reduceAlloc(Alloc &Orig, R_SExpr E0) {
-    return new (Arena) Alloc(Orig, E0);
-  }
-  R_SExpr reduceLoad(Load &Orig, R_SExpr E0) {
-    return new (Arena) Load(Orig, E0);
-  }
-  R_SExpr reduceStore(Store &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) Store(Orig, E0, E1);
-  }
-  R_SExpr reduceArrayIndex(ArrayIndex &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) ArrayIndex(Orig, E0, E1);
-  }
-  R_SExpr reduceArrayAdd(ArrayAdd &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) ArrayAdd(Orig, E0, E1);
-  }
-  R_SExpr reduceUnaryOp(UnaryOp &Orig, R_SExpr E0) {
-    return new (Arena) UnaryOp(Orig, E0);
-  }
-  R_SExpr reduceBinaryOp(BinaryOp &Orig, R_SExpr E0, R_SExpr E1) {
-    return new (Arena) BinaryOp(Orig, E0, E1);
-  }
-  R_SExpr reduceCast(Cast &Orig, R_SExpr E0) {
-    return new (Arena) Cast(Orig, E0);
-  }
-
-  R_SExpr reduceSCFG(SCFG &Orig, Container<BasicBlock *> &Bbs) {
-    return nullptr;  // FIXME: implement CFG rewriting
-  }
-  R_BasicBlock reduceBasicBlock(BasicBlock &Orig, Container<Variable *> &As,
-                                Container<Variable *> &Is, R_SExpr T) {
-    return nullptr;  // FIXME: implement CFG rewriting
-  }
-  R_SExpr reducePhi(Phi &Orig, Container<R_SExpr> &As) {
-    return new (Arena) Phi(Orig, std::move(As.Elems));
-  }
-  R_SExpr reduceGoto(Goto &Orig, BasicBlock *B) {
-    return new (Arena) Goto(Orig, B, 0);  // FIXME: set index
-  }
-  R_SExpr reduceBranch(Branch &O, R_SExpr C, BasicBlock *B0, BasicBlock *B1) {
-    return new (Arena) Branch(O, C, B0, B1, 0, 0);  // FIXME: set indices
-  }
-
-  R_SExpr reduceIdentifier(Identifier &Orig) {
-    return new (Arena) Identifier(Orig);
-  }
-  R_SExpr reduceIfThenElse(IfThenElse &Orig, R_SExpr C, R_SExpr T, R_SExpr E) {
-    return new (Arena) IfThenElse(Orig, C, T, E);
-  }
-  R_SExpr reduceLet(Let &Orig, Variable *Nvd, R_SExpr B) {
-    return new (Arena) Let(Orig, Nvd, B);
-  }
-
-  // Create a new variable from orig, and push it onto the lexical scope.
-  Variable *enterScope(Variable &Orig, R_SExpr E0) {
-    return new (Arena) Variable(Orig, E0);
-  }
-  // Exit the lexical scope of orig.
-  void exitScope(const Variable &Orig) {}
-
-  void enterCFG(SCFG &Cfg) {}
-  void exitCFG(SCFG &Cfg) {}
-  void enterBasicBlock(BasicBlock &BB) {}
-  void exitBasicBlock(BasicBlock &BB) {}
-
-  // Map Variable references to their rewritten definitions.
-  Variable *reduceVariableRef(Variable *Ovd) { return Ovd; }
-
-  // Map BasicBlock references to their rewritten definitions.
-  BasicBlock *reduceBasicBlockRef(BasicBlock *Obb) { return Obb; }
-};
-
-
-class SExprCopier : public CopyReducer<SExprCopier> {
-public:
-  typedef SExpr *R_SExpr;
-
-  SExprCopier(MemRegionRef A) : CopyReducer(A) { }
-
-  // Create a copy of e in region a.
-  static SExpr *copy(SExpr *E, MemRegionRef A) {
-    SExprCopier Copier(A);
-    return Copier.traverse(E, TRV_Normal);
-  }
-};
-
-
-
 // Base class for visit traversals.
 class VisitReducerBase : public SimpleReducerBase {
 public:
@@ -368,8 +227,8 @@ public:
   R_SExpr reduceSCFG(SCFG &Orig, Container<BasicBlock *> Bbs) {
     return Bbs.Success;
   }
-  R_BasicBlock reduceBasicBlock(BasicBlock &Orig, Container<Variable *> &As,
-                                Container<Variable *> &Is, R_SExpr T) {
+  R_BasicBlock reduceBasicBlock(BasicBlock &Orig, Container<R_SExpr> &As,
+                                Container<R_SExpr> &Is, R_SExpr T) {
     return (As.Success && Is.Success && T);
   }
   R_SExpr reducePhi(Phi &Orig, Container<R_SExpr> &As) {
@@ -381,6 +240,9 @@ public:
   R_SExpr reduceBranch(Branch &O, R_SExpr C, BasicBlock *B0, BasicBlock *B1) {
     return C;
   }
+  R_SExpr reduceReturn(Return &O, R_SExpr E) {
+    return E;
+  }
 
   R_SExpr reduceIdentifier(Identifier &Orig) {
     return true;
@@ -433,7 +295,7 @@ public:
 #include "ThreadSafetyOps.def"
 #undef TIL_OPCODE_DEF
     }
-    llvm_unreachable("invalid enum");
+    return false;
   }
 };
 
@@ -514,9 +376,9 @@ public:
 
 
 
-inline std::ostream& operator<<(std::ostream& SS, llvm::StringRef R) {
-  return SS.write(R.data(), R.size());
-}
+// inline std::ostream& operator<<(std::ostream& SS, StringRef R) {
+//   return SS.write(R.data(), R.size());
+// }
 
 // Pretty printer for TIL expressions
 template <typename Self, typename StreamType>
@@ -587,6 +449,7 @@ protected:
       case COP_Phi:        return Prec_Atom;
       case COP_Goto:       return Prec_Atom;
       case COP_Branch:     return Prec_Atom;
+      case COP_Return:     return Prec_Other;
 
       case COP_Identifier: return Prec_Atom;
       case COP_IfThenElse: return Prec_Other;
@@ -595,22 +458,29 @@ protected:
     return Prec_MAX;
   }
 
-  void printBlockLabel(StreamType & SS, const BasicBlock *BB, unsigned index) {
+  void printBlockLabel(StreamType & SS, const BasicBlock *BB, int index) {
     if (!BB) {
       SS << "BB_null";
       return;
     }
     SS << "BB_";
     SS << BB->blockID();
-    SS << ":";
-    SS << index;
+    if (index >= 0) {
+      SS << ":";
+      SS << index;
+    }
   }
 
-  void printSExpr(const SExpr *E, StreamType &SS, unsigned P) {
+
+  void printSExpr(const SExpr *E, StreamType &SS, unsigned P, bool Sub=true) {
     if (!E) {
       self()->printNull(SS);
       return;
     }
+    if (Sub && E->block() && E->opcode() != COP_Variable) {
+      SS << "_x" << E->id();
+      return;
+    }
     if (self()->precedence(E) > P) {
       // Wrap expr in () if necessary.
       SS << "(";
@@ -740,20 +610,11 @@ protected:
     SS << E->clangDecl()->getNameAsString();
   }
 
-  void printVariable(const Variable *V, StreamType &SS, bool IsVarDecl = false) {
-    if (!IsVarDecl && Cleanup) {
-      const SExpr* E = getCanonicalVal(V);
-      if (E != V) {
-        printSExpr(E, SS, Prec_Atom);
-        return;
-      }
-    }
-    if (V->kind() == Variable::VK_LetBB)
-      SS << V->name() << V->getBlockID() << "_" << V->getID();
-    else if (CStyle && V->kind() == Variable::VK_SFun)
+  void printVariable(const Variable *V, StreamType &SS, bool IsVarDecl=false) {
+    if (CStyle && V->kind() == Variable::VK_SFun)
       SS << "this";
     else
-      SS << V->name() << V->getID();
+      SS << V->name() << V->id();
   }
 
   void printFunction(const Function *E, StreamType &SS, unsigned sugared = 0) {
@@ -927,32 +788,38 @@ protected:
     newline(SS);
   }
 
+
+  void printBBInstr(const SExpr *E, StreamType &SS) {
+    bool Sub = false;
+    if (E->opcode() == COP_Variable) {
+      auto *V = cast<Variable>(E);
+      SS << "let " << V->name() << V->id() << " = ";
+      E = V->definition();
+      Sub = true;
+    }
+    else if (E->opcode() != COP_Store) {
+      SS << "let _x" << E->id() << " = ";
+    }
+    self()->printSExpr(E, SS, Prec_MAX, Sub);
+    SS << ";";
+    newline(SS);
+  }
+
   void printBasicBlock(const BasicBlock *E, StreamType &SS) {
     SS << "BB_" << E->blockID() << ":";
     if (E->parent())
       SS << " BB_" << E->parent()->blockID();
     newline(SS);
-    for (auto *A : E->arguments()) {
-      SS << "let ";
-      self()->printVariable(A, SS, true);
-      SS << " = ";
-      self()->printSExpr(A->definition(), SS, Prec_MAX);
-      SS << ";";
-      newline(SS);
-    }
-    for (auto *I : E->instructions()) {
-      if (I->definition()->opcode() != COP_Store) {
-        SS << "let ";
-        self()->printVariable(I, SS, true);
-        SS << " = ";
-      }
-      self()->printSExpr(I->definition(), SS, Prec_MAX);
-      SS << ";";
-      newline(SS);
-    }
+
+    for (auto *A : E->arguments())
+      printBBInstr(A, SS);
+
+    for (auto *I : E->instructions())
+      printBBInstr(I, SS);
+
     const SExpr *T = E->terminator();
     if (T) {
-      self()->printSExpr(T, SS, Prec_MAX);
+      self()->printSExpr(T, SS, Prec_MAX, false);
       SS << ";";
       newline(SS);
     }
@@ -983,9 +850,14 @@ protected:
     SS << "branch (";
     self()->printSExpr(E->condition(), SS, Prec_MAX);
     SS << ") ";
-    printBlockLabel(SS, E->thenBlock(), E->thenIndex());
+    printBlockLabel(SS, E->thenBlock(), -1);
     SS << " ";
-    printBlockLabel(SS, E->elseBlock(), E->elseIndex());
+    printBlockLabel(SS, E->elseBlock(), -1);
+  }
+
+  void printReturn(const Return *E, StreamType &SS) {
+    SS << "return ";
+    self()->printSExpr(E->returnValue(), SS, Prec_Other);
   }
 
   void printIdentifier(const Identifier *E, StreamType &SS) {

include/clang/Analysis/Analyses/ThreadSafetyUtil.h

@@ -142,20 +142,35 @@ public:
     assert(i < Size && "Array index out of bounds.");
     return Data[i];
   }
+  T &back() {
+    assert(Size && "No elements in the array.");
+    return Data[Size - 1];
+  }
+  const T &back() const {
+    assert(Size && "No elements in the array.");
+    return Data[Size - 1];
+  }
 
   iterator begin() { return Data; }
+  iterator end()   { return Data + Size; }
+
   const_iterator begin() const { return Data; }
-  iterator end() { return Data + Size; }
-  const_iterator end() const { return Data + Size; }
+  const_iterator end()   const { return Data + Size; }
 
   const_iterator cbegin() const { return Data; }
-  const_iterator cend() const { return Data + Size; }
+  const_iterator cend()   const { return Data + Size; }
 
   void push_back(const T &Elem) {
     assert(Size < Capacity);
     Data[Size++] = Elem;
   }
 
+  // drop last n elements from array
+  void drop(unsigned n = 0) {
+    assert(Size > n);
+    Size -= n;
+  }
+
   void setValues(unsigned Sz, const T& C) {
     assert(Sz <= Capacity);
     Size = Sz;
@@ -173,6 +188,37 @@ public:
     return J - Osz;
   }
 
+  // An adaptor to reverse a simple array
+  class ReverseAdaptor {
+   public:
+    ReverseAdaptor(SimpleArray &Array) : Array(Array) {}
+    // A reverse iterator used by the reverse adaptor
+    class Iterator {
+     public:
+      Iterator(T *Data) : Data(Data) {}
+      T &operator*() { return *Data; }
+      const T &operator*() const { return *Data; }
+      Iterator &operator++() {
+        --Data;
+        return *this;
+      }
+      bool operator!=(Iterator Other) { return Data != Other.Data; }
+
+     private:
+      T *Data;
+    };
+    Iterator begin() { return Array.end() - 1; }
+    Iterator end() { return Array.begin() - 1; }
+    const Iterator begin() const { return Array.end() - 1; }
+    const Iterator end() const { return Array.begin() - 1; }
+
+   private:
+    SimpleArray &Array;
+  };
+
+  const ReverseAdaptor reverse() const { return ReverseAdaptor(*this); }
+  ReverseAdaptor reverse() { return ReverseAdaptor(*this); }
+
 private:
   // std::max is annoying here, because it requires a reference,
   // thus forcing InitialCapacity to be initialized outside the .h file.
@@ -187,6 +233,7 @@ private:
   size_t Capacity;
 };
 
+
 }  // end namespace til
 
 
@@ -312,6 +359,12 @@ private:
 };
 
 
+inline std::ostream& operator<<(std::ostream& ss, const StringRef str) {
+  ss << str.data();
+  return ss;
+}
+
+
 } // end namespace threadSafety
 } // end namespace clang
 

lib/Analysis/ThreadSafetyCommon.cpp

@@ -63,11 +63,9 @@ std::string getSourceLiteralString(const clang::Expr *CE) {
 namespace til {
 
 // Return true if E is a variable that points to an incomplete Phi node.
-static bool isIncompleteVar(const SExpr *E) {
-  if (const auto *V = dyn_cast<Variable>(E)) {
-    if (const auto *Ph = dyn_cast<Phi>(V->definition()))
-      return Ph->status() == Phi::PH_Incomplete;
-  }
+static bool isIncompletePhi(const SExpr *E) {
+  if (const auto *Ph = dyn_cast<Phi>(E))
+    return Ph->status() == Phi::PH_Incomplete;
   return false;
 }
 
@@ -320,6 +318,8 @@ til::SExpr *SExprBuilder::translateCXXThisExpr(const CXXThisExpr *TE,
 const ValueDecl *getValueDeclFromSExpr(const til::SExpr *E) {
   if (auto *V = dyn_cast<til::Variable>(E))
     return V->clangDecl();
+  if (auto *Ph = dyn_cast<til::Phi>(E))
+    return Ph->clangDecl();
   if (auto *P = dyn_cast<til::Project>(E))
     return P->clangDecl();
   if (auto *L = dyn_cast<til::LiteralPtr>(E))
@@ -641,14 +641,14 @@ SExprBuilder::translateDeclStmt(const DeclStmt *S, CallingContext *Ctx) {
 // If E is trivial returns E.
 til::SExpr *SExprBuilder::addStatement(til::SExpr* E, const Stmt *S,
                                        const ValueDecl *VD) {
-  if (!E || !CurrentBB || til::ThreadSafetyTIL::isTrivial(E))
+  if (!E || !CurrentBB || E->block() || til::ThreadSafetyTIL::isTrivial(E))
     return E;
-
-  til::Variable *V = new (Arena) til::Variable(E, VD);
-  CurrentInstructions.push_back(V);
+  if (VD)
+    E = new (Arena) til::Variable(E, VD);
+  CurrentInstructions.push_back(E);
   if (S)
-    insertStmt(S, V);
-  return V;
+    insertStmt(S, E);
+  return E;
 }
 
 
@@ -705,11 +705,11 @@ void SExprBuilder::makePhiNodeVar(unsigned i, unsigned NPreds, til::SExpr *E) {
   unsigned ArgIndex = CurrentBlockInfo->ProcessedPredecessors;
   assert(ArgIndex > 0 && ArgIndex < NPreds);
 
-  til::Variable *V = dyn_cast<til::Variable>(CurrentLVarMap[i].second);
-  if (V && V->getBlockID() == CurrentBB->blockID()) {
+  til::SExpr *CurrE = CurrentLVarMap[i].second;
+  if (CurrE->block() == CurrentBB) {
     // We already have a Phi node in the current block,
     // so just add the new variable to the Phi node.
-    til::Phi *Ph = dyn_cast<til::Phi>(V->definition());
+    til::Phi *Ph = dyn_cast<til::Phi>(CurrE);
     assert(Ph && "Expecting Phi node.");
     if (E)
       Ph->values()[ArgIndex] = E;
@@ -718,27 +718,26 @@ void SExprBuilder::makePhiNodeVar(unsigned i, unsigned NPreds, til::SExpr *E) {
 
   // Make a new phi node: phi(..., E)
   // All phi args up to the current index are set to the current value.
-  til::SExpr *CurrE = CurrentLVarMap[i].second;
   til::Phi *Ph = new (Arena) til::Phi(Arena, NPreds);
   Ph->values().setValues(NPreds, nullptr);
   for (unsigned PIdx = 0; PIdx < ArgIndex; ++PIdx)
     Ph->values()[PIdx] = CurrE;
   if (E)
     Ph->values()[ArgIndex] = E;
+  Ph->setClangDecl(CurrentLVarMap[i].first);
   // If E is from a back-edge, or either E or CurrE are incomplete, then
   // mark this node as incomplete; we may need to remove it later.
-  if (!E || isIncompleteVar(E) || isIncompleteVar(CurrE)) {
+  if (!E || isIncompletePhi(E) || isIncompletePhi(CurrE)) {
     Ph->setStatus(til::Phi::PH_Incomplete);
   }
 
   // Add Phi node to current block, and update CurrentLVarMap[i]
-  auto *Var = new (Arena) til::Variable(Ph, CurrentLVarMap[i].first);
-  CurrentArguments.push_back(Var);
+  CurrentArguments.push_back(Ph);
   if (Ph->status() == til::Phi::PH_Incomplete)
-    IncompleteArgs.push_back(Var);
+    IncompleteArgs.push_back(Ph);
 
   CurrentLVarMap.makeWritable();
-  CurrentLVarMap.elem(i).second = Var;
+  CurrentLVarMap.elem(i).second = Ph;
 }
 
 
@@ -812,15 +811,13 @@ void SExprBuilder::mergePhiNodesBackEdge(const CFGBlock *Blk) {
   unsigned ArgIndex = BBInfo[Blk->getBlockID()].ProcessedPredecessors;
   assert(ArgIndex > 0 && ArgIndex < BB->numPredecessors());
 
-  for (til::Variable *V : BB->arguments()) {
-    til::Phi *Ph = dyn_cast_or_null<til::Phi>(V->definition());
+  for (til::SExpr *PE : BB->arguments()) {
+    til::Phi *Ph = dyn_cast_or_null<til::Phi>(PE);
     assert(Ph && "Expecting Phi Node.");
     assert(Ph->values()[ArgIndex] == nullptr && "Wrong index for back edge.");
-    assert(V->clangDecl() && "No local variable for Phi node.");
 
-    til::SExpr *E = lookupVarDecl(V->clangDecl());
+    til::SExpr *E = lookupVarDecl(Ph->clangDecl());
     assert(E && "Couldn't find local variable for Phi node.");
-
     Ph->values()[ArgIndex] = E;
   }
 }
@@ -899,8 +896,8 @@ void SExprBuilder::enterCFGBlockBody(const CFGBlock *B) {
   // Push those arguments onto the basic block.
   CurrentBB->arguments().reserve(
     static_cast<unsigned>(CurrentArguments.size()), Arena);
-  for (auto *V : CurrentArguments)
-    CurrentBB->addArgument(V);
+  for (auto *A : CurrentArguments)
+    CurrentBB->addArgument(A);
 }
 
 
@@ -934,7 +931,7 @@ void SExprBuilder::exitCFGBlockBody(const CFGBlock *B) {
     til::BasicBlock *BB = *It ? lookupBlock(*It) : nullptr;
     // TODO: set index
     unsigned Idx = BB ? BB->findPredecessorIndex(CurrentBB) : 0;
-    til::SExpr *Tm = new (Arena) til::Goto(BB, Idx);
+    auto *Tm = new (Arena) til::Goto(BB, Idx);
     CurrentBB->setTerminator(Tm);
   }
   else if (N == 2) {
@@ -942,9 +939,8 @@ void SExprBuilder::exitCFGBlockBody(const CFGBlock *B) {
     til::BasicBlock *BB1 = *It ? lookupBlock(*It) : nullptr;
     ++It;
     til::BasicBlock *BB2 = *It ? lookupBlock(*It) : nullptr;
-    unsigned Idx1 = BB1 ? BB1->findPredecessorIndex(CurrentBB) : 0;
-    unsigned Idx2 = BB2 ? BB2->findPredecessorIndex(CurrentBB) : 0;
-    til::SExpr *Tm = new (Arena) til::Branch(C, BB1, BB2, Idx1, Idx2);
+    // FIXME: make sure these arent' critical edges.
+    auto *Tm = new (Arena) til::Branch(C, BB1, BB2);
     CurrentBB->setTerminator(Tm);
   }
 }
@@ -971,10 +967,9 @@ void SExprBuilder::exitCFGBlock(const CFGBlock *B) {
 
 
 void SExprBuilder::exitCFG(const CFGBlock *Last) {
-  for (auto *V : IncompleteArgs) {
-    til::Phi *Ph = dyn_cast<til::Phi>(V->definition());
-    if (Ph && Ph->status() == til::Phi::PH_Incomplete)
-      simplifyIncompleteArg(V, Ph);
+  for (auto *Ph : IncompleteArgs) {
+    if (Ph->status() == til::Phi::PH_Incomplete)
+      simplifyIncompleteArg(Ph);
   }
 
   CurrentArguments.clear();

lib/Analysis/ThreadSafetyTIL.cpp

@@ -48,12 +48,20 @@ StringRef getBinaryOpcodeString(TIL_BinaryOpcode Op) {
 }
 
 
+SExpr* Future::force() {
+  Status = FS_evaluating;
+  Result = compute();
+  Status = FS_done;
+  return Result;
+}
+
+
 unsigned BasicBlock::addPredecessor(BasicBlock *Pred) {
   unsigned Idx = Predecessors.size();
   Predecessors.reserveCheck(1, Arena);
   Predecessors.push_back(Pred);
-  for (Variable *V : Args) {
-    if (Phi* Ph = dyn_cast<Phi>(V->definition())) {
+  for (SExpr *E : Args) {
+    if (Phi* Ph = dyn_cast<Phi>(E)) {
       Ph->values().reserveCheck(1, Arena);
       Ph->values().push_back(nullptr);
     }
@@ -61,105 +69,73 @@ unsigned BasicBlock::addPredecessor(BasicBlock *Pred) {
   return Idx;
 }
 
+
 void BasicBlock::reservePredecessors(unsigned NumPreds) {
   Predecessors.reserve(NumPreds, Arena);
-  for (Variable *V : Args) {
-    if (Phi* Ph = dyn_cast<Phi>(V->definition())) {
+  for (SExpr *E : Args) {
+    if (Phi* Ph = dyn_cast<Phi>(E)) {
       Ph->values().reserve(NumPreds, Arena);
     }
   }
 }
 
-void BasicBlock::renumberVars() {
-  unsigned VID = 0;
-  for (Variable *V : Args) {
-    V->setID(BlockID, VID++);
-  }
-  for (Variable *V : Instrs) {
-    V->setID(BlockID, VID++);
-  }
-}
-
-void SCFG::renumberVars() {
-  for (BasicBlock *B : Blocks) {
-    B->renumberVars();
-  }
-}
-
-
 
 // If E is a variable, then trace back through any aliases or redundant
 // Phi nodes to find the canonical definition.
 const SExpr *getCanonicalVal(const SExpr *E) {
-  while (auto *V = dyn_cast<Variable>(E)) {
-    const SExpr *D;
-    do {
-      if (V->kind() != Variable::VK_Let)
-        return V;
-      D = V->definition();
-      auto *V2 = dyn_cast<Variable>(D);
-      if (V2)
-        V = V2;
-      else
-        break;
-    } while (true);
-
-    if (ThreadSafetyTIL::isTrivial(D))
-      return D;
-
-    if (const Phi *Ph = dyn_cast<Phi>(D)) {
+  while (true) {
+    if (auto *V = dyn_cast<Variable>(E)) {
+      if (V->kind() == Variable::VK_Let) {
+        E = V->definition();
+        continue;
+      }
+    }
+    if (const Phi *Ph = dyn_cast<Phi>(E)) {
       if (Ph->status() == Phi::PH_SingleVal) {
         E = Ph->values()[0];
         continue;
       }
     }
-    return V;
+    break;
   }
   return E;
 }
 
 
-
 // If E is a variable, then trace back through any aliases or redundant
 // Phi nodes to find the canonical definition.
 // The non-const version will simplify incomplete Phi nodes.
 SExpr *simplifyToCanonicalVal(SExpr *E) {
-  while (auto *V = dyn_cast<Variable>(E)) {
-    SExpr *D;
-    do {
+  while (true) {
+    if (auto *V = dyn_cast<Variable>(E)) {
       if (V->kind() != Variable::VK_Let)
         return V;
-      D = V->definition();
-      auto *V2 = dyn_cast<Variable>(D);
-      if (V2)
-        V = V2;
-      else
-        break;
-    } while (true);
-
-    if (ThreadSafetyTIL::isTrivial(D))
-      return D;
-
-    if (Phi *Ph = dyn_cast<Phi>(D)) {
+      // Eliminate redundant variables, e.g. x = y, or x = 5,
+      // but keep anything more complicated.
+      if (til::ThreadSafetyTIL::isTrivial(V->definition())) {
+        E = V->definition();
+        continue;
+      }
+      return V;
+    }
+    if (auto *Ph = dyn_cast<Phi>(E)) {
       if (Ph->status() == Phi::PH_Incomplete)
-        simplifyIncompleteArg(V, Ph);
-
+        simplifyIncompleteArg(Ph);
+      // Eliminate redundant Phi nodes.
       if (Ph->status() == Phi::PH_SingleVal) {
         E = Ph->values()[0];
         continue;
       }
     }
-    return V;
+    return E;
   }
-  return E;
 }
 
 
-
 // Trace the arguments of an incomplete Phi node to see if they have the same
 // canonical definition.  If so, mark the Phi node as redundant.
 // getCanonicalVal() will recursively call simplifyIncompletePhi().
-void simplifyIncompleteArg(Variable *V, til::Phi *Ph) {
+void simplifyIncompleteArg(til::Phi *Ph) {
   assert(Ph && Ph->status() == Phi::PH_Incomplete);
 
   // eliminate infinite recursion -- assume that this node is not redundant.
@@ -168,18 +144,200 @@ void simplifyIncompleteArg(Variable *V, til::Phi *Ph) {
   SExpr *E0 = simplifyToCanonicalVal(Ph->values()[0]);
   for (unsigned i=1, n=Ph->values().size(); i<n; ++i) {
     SExpr *Ei = simplifyToCanonicalVal(Ph->values()[i]);
-    if (Ei == V)
+    if (Ei == Ph)
       continue;  // Recursive reference to itself.  Don't count.
     if (Ei != E0) {
       return;    // Status is already set to MultiVal.
     }
   }
   Ph->setStatus(Phi::PH_SingleVal);
-  // Eliminate Redundant Phi node.
-  V->setDefinition(Ph->values()[0]);
 }
 
 
+// Renumbers the arguments and instructions to have unique, sequential IDs.
+int BasicBlock::renumberInstrs(int ID) {
+  for (auto *Arg : Args)
+    Arg->setID(this, ID++);
+  for (auto *Instr : Instrs)
+    Instr->setID(this, ID++);
+  TermInstr->setID(this, ID++);
+  return ID;
+}
+
+// Sorts the CFGs blocks using a reverse post-order depth-first traversal.
+// Each block will be written into the Blocks array in order, and its BlockID
+// will be set to the index in the array.  Sorting should start from the entry
+// block, and ID should be the total number of blocks.
+int BasicBlock::topologicalSort(SimpleArray<BasicBlock*>& Blocks, int ID) {
+  if (Visited) return ID;
+  Visited = 1;
+  for (auto *Block : successors())
+    ID = Block->topologicalSort(Blocks, ID);
+  // set ID and update block array in place.
+  // We may lose pointers to unreachable blocks.
+  assert(ID > 0);
+  BlockID = --ID;
+  Blocks[BlockID] = this;
+  return ID;
+}
+
+// Performs a reverse topological traversal, starting from the exit block and
+// following back-edges.  The dominator is serialized before any predecessors,
+// which guarantees that all blocks are serialized after their dominator and
+// before their post-dominator (because it's a reverse topological traversal).
+// ID should be initially set to 0.
+//
+// This sort assumes that (1) dominators have been computed, (2) there are no
+// critical edges, and (3) the entry block is reachable from the exit block
+// and no blocks are accessable via traversal of back-edges from the exit that
+// weren't accessable via forward edges from the entry.
+int BasicBlock::topologicalFinalSort(SimpleArray<BasicBlock*>& Blocks, int ID) {
+  // Visited is assumed to have been set by the topologicalSort.  This pass
+  // assumes !Visited means that we've visited this node before.
+  if (!Visited) return ID;
+  Visited = 0;
+  if (DominatorNode.Parent)
+    ID = DominatorNode.Parent->topologicalFinalSort(Blocks, ID);
+  for (auto *Pred : Predecessors)
+    ID = Pred->topologicalFinalSort(Blocks, ID);
+  assert(ID < Blocks.size());
+  BlockID = ID++;
+  Blocks[BlockID] = this;
+  return ID;
+}
+
+// Computes the immediate dominator of the current block.  Assumes that all of
+// its predecessors have already computed their dominators.  This is achieved
+// by visiting the nodes in topological order.
+void BasicBlock::computeDominator() {
+  BasicBlock *Candidate = nullptr;
+  // Walk backwards from each predecessor to find the common dominator node.
+  for (auto *Pred : Predecessors) {
+    // Skip back-edges
+    if (Pred->BlockID >= BlockID) continue;
+    // If we don't yet have a candidate for dominator yet, take this one.
+    if (Candidate == nullptr) {
+      Candidate = Pred;
+      continue;
+    }
+    // Walk the alternate and current candidate back to find a common ancestor.
+    auto *Alternate = Pred;
+    while (Alternate != Candidate) {
+      if (Candidate->BlockID > Alternate->BlockID)
+        Candidate = Candidate->DominatorNode.Parent;
+      else
+        Alternate = Alternate->DominatorNode.Parent;
+    }
+  }
+  DominatorNode.Parent = Candidate;
+  DominatorNode.SizeOfSubTree = 1;
+}
+
+// Computes the immediate post-dominator of the current block.  Assumes that all
+// of its successors have already computed their post-dominators.  This is
+// achieved visiting the nodes in reverse topological order.
+void BasicBlock::computePostDominator() {
+  BasicBlock *Candidate = nullptr;
+  // Walk back from each predecessor to find the common post-dominator node.
+  for (auto *Succ : successors()) {
+    // Skip back-edges
+    if (Succ->BlockID <= BlockID) continue;
+    // If we don't yet have a candidate for post-dominator yet, take this one.
+    if (Candidate == nullptr) {
+      Candidate = Succ;
+      continue;
+    }
+    // Walk the alternate and current candidate back to find a common ancestor.
+    auto *Alternate = Succ;
+    while (Alternate != Candidate) {
+      if (Candidate->BlockID < Alternate->BlockID)
+        Candidate = Candidate->PostDominatorNode.Parent;
+      else
+        Alternate = Alternate->PostDominatorNode.Parent;
+    }
+  }
+  PostDominatorNode.Parent = Candidate;
+  PostDominatorNode.SizeOfSubTree = 1;
+}
+
+
+// Renumber instructions in all blocks
+void SCFG::renumberInstrs() {
+  int InstrID = 0;
+  for (auto *Block : Blocks)
+    InstrID = Block->renumberInstrs(InstrID);
+}
+
+
+static inline void computeNodeSize(BasicBlock *B,
+                                   BasicBlock::TopologyNode BasicBlock::*TN) {
+  BasicBlock::TopologyNode *N = &(B->*TN);
+  if (N->Parent) {
+    BasicBlock::TopologyNode *P = &(N->Parent->*TN);
+    // Initially set ID relative to the (as yet uncomputed) parent ID
+    N->NodeID = P->SizeOfSubTree;
+    P->SizeOfSubTree += N->SizeOfSubTree;
+  }
+}
+
+static inline void computeNodeID(BasicBlock *B,
+                                 BasicBlock::TopologyNode BasicBlock::*TN) {
+  BasicBlock::TopologyNode *N = &(B->*TN);
+  if (N->Parent) {
+    BasicBlock::TopologyNode *P = &(N->Parent->*TN);
+    N->NodeID += P->NodeID;    // Fix NodeIDs relative to starting node.
+  }
+}
+
+
+// Normalizes a CFG.  Normalization has a few major components:
+// 1) Removing unreachable blocks.
+// 2) Computing dominators and post-dominators
+// 3) Topologically sorting the blocks into the "Blocks" array.
+void SCFG::computeNormalForm() {
+  // Topologically sort the blocks starting from the entry block.
+  int NumUnreachableBlocks = Entry->topologicalSort(Blocks, Blocks.size());
+  if (NumUnreachableBlocks > 0) {
+    // If there were unreachable blocks shift everything down, and delete them.
+    for (size_t I = NumUnreachableBlocks, E = Blocks.size(); I < E; ++I) {
+      size_t NI = I - NumUnreachableBlocks;
+      Blocks[NI] = Blocks[I];
+      Blocks[NI]->BlockID = NI;
+      // FIXME: clean up predecessor pointers to unreachable blocks?
+    }
+    Blocks.drop(NumUnreachableBlocks);
+  }
+
+  // Compute dominators.
+  for (auto *Block : Blocks)
+    Block->computeDominator();
+
+  // Once dominators have been computed, the final sort may be performed.
+  int NumBlocks = Exit->topologicalFinalSort(Blocks, 0);
+  assert(NumBlocks == Blocks.size());
+  (void) NumBlocks;
+
+  // Renumber the instructions now that we have a final sort.
+  renumberInstrs();
+
+  // Compute post-dominators and compute the sizes of each node in the
+  // dominator tree.
+  for (auto *Block : Blocks.reverse()) {
+    Block->computePostDominator();
+    computeNodeSize(Block, &BasicBlock::DominatorNode);
+  }
+  // Compute the sizes of each node in the post-dominator tree and assign IDs in
+  // the dominator tree.
+  for (auto *Block : Blocks) {
+    computeNodeID(Block, &BasicBlock::DominatorNode);
+    computeNodeSize(Block, &BasicBlock::PostDominatorNode);
+  }
+  // Assign IDs in the post-dominator tree.
+  for (auto *Block : Blocks.reverse()) {
+    computeNodeID(Block, &BasicBlock::PostDominatorNode);
+  }
+}
+
 }  // end namespace til
 }  // end namespace threadSafety
 }  // end namespace clang