Implement A.noalias() = B * C without temporaries

* Wrap expression inside EvalToTemp in copy_using_evaluators() if we
  assume aliasing for that expression (that is, for products)
* Remove temporary kludge of evaluating expression to temporary in
  AllAtOnce traversal
* Implement EvalToTemp expression object

3 files changed, 77 insertions, 8 deletions

diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h
index 9be00067d..9f5deb8e7 100644
--- a/Eigen/src/Core/AssignEvaluator.h
+++ b/Eigen/src/Core/AssignEvaluator.h
@@ -617,12 +617,8 @@ struct copy_using_evaluator_impl<DstXprType, SrcXprType, AllAtOnceTraversal, NoU
     SrcEvaluatorType srcEvaluator(src);
 
     // Evaluate rhs in temporary to prevent aliasing problems in a = a * a;
-    // TODO: Be smarter about this
     // TODO: Do not pass the xpr object to evalTo()
-    typename DstXprType::PlainObject tmp;
-    typename evaluator<typename DstXprType::PlainObject>::type tmpEvaluator(tmp);
-    srcEvaluator.evalTo(tmpEvaluator, tmp);
-    copy_using_evaluator(dst, tmp);
+    srcEvaluator.evalTo(dstEvaluator, dst);
   }
 };
 
@@ -640,11 +636,33 @@ const DstXprType& copy_using_evaluator(const NoAlias<DstXprType, StorageBase>& d
   return noalias_copy_using_evaluator(dst.expression(), src.derived());
 }
 
+template<typename XprType, int AssumeAliasing = evaluator_traits<XprType>::AssumeAliasing>
+struct AddEvalIfAssumingAliasing;
+
+template<typename XprType>
+struct AddEvalIfAssumingAliasing<XprType, 0>
+{
+  static const XprType& run(const XprType& xpr) 
+  {
+    return xpr;
+  }
+};
+
+template<typename XprType>
+struct AddEvalIfAssumingAliasing<XprType, 1>
+{
+  static const EvalToTemp<XprType> run(const XprType& xpr)
+  {
+    return EvalToTemp<XprType>(xpr);
+  }
+};
+
 template<typename DstXprType, typename SrcXprType>
 EIGEN_STRONG_INLINE
 const DstXprType& copy_using_evaluator(const EigenBase<DstXprType>& dst, const EigenBase<SrcXprType>& src)
 {
-  return noalias_copy_using_evaluator(dst.const_cast_derived(), src.derived());
+  return noalias_copy_using_evaluator(dst.const_cast_derived(), 
+				      AddEvalIfAssumingAliasing<SrcXprType>::run(src.derived()));
 }
 
 template<typename DstXprType, typename SrcXprType>
diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 808546ec1..906df2af7 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -40,6 +40,10 @@ struct evaluator_traits
   // 1 if evaluator_impl<T>::evalTo() exists
   // 0 if evaluator_impl<T> allows coefficient-based access
   static const int HasEvalTo = 0;
+
+  // 1 if assignment A = B assumes aliasing when B is of type T and thus B needs to be evaluated into a
+  // temporary; 0 if not.
+  static const int AssumeAliasing = 0;
 };
 
 // expression class for evaluating nested expression to a temporary
@@ -246,16 +250,61 @@ struct evaluator_impl<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
 // -------------------- EvalToTemp --------------------
 
 template<typename ArgType>
+struct traits<EvalToTemp<ArgType> >
+  : public traits<ArgType>
+{ };
+
+template<typename ArgType>
+class EvalToTemp
+  : public dense_xpr_base<EvalToTemp<ArgType> >::type
+{
+ public:
+ 
+  typedef typename dense_xpr_base<EvalToTemp>::type Base;
+  EIGEN_GENERIC_PUBLIC_INTERFACE(EvalToTemp)
+ 
+  EvalToTemp(const ArgType& arg)
+    : m_arg(arg)
+  { }
+ 
+  const ArgType& arg() const
+  {
+    return m_arg;
+  }
+
+  Index rows() const 
+  {
+    return m_arg.rows();
+  }
+
+  Index cols() const 
+  {
+    return m_arg.cols();
+  }
+
+ private:
+  const ArgType& m_arg;
+};
+ 
+template<typename ArgType>
 struct evaluator_impl<EvalToTemp<ArgType> >
   : evaluator_impl<typename ArgType::PlainObject>
 {
+  typedef EvalToTemp<ArgType> XprType;
   typedef typename ArgType::PlainObject PlainObject;
   typedef evaluator_impl<PlainObject> BaseType;
 
+  evaluator_impl(const XprType& xpr) 
+    : BaseType(m_result)
+  { 
+    noalias_copy_using_evaluator(m_result, xpr.arg());
+  };
+
+  // this constructor is used when nesting an EvalTo evaluator in another evaluator
   evaluator_impl(const ArgType& arg) 
     : BaseType(m_result)
   { 
-    copy_using_evaluator(m_result, arg);
+    noalias_copy_using_evaluator(m_result, arg);
   };
 
 protected:
diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index e814a4710..3804113f8 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -56,7 +56,9 @@ struct product_evaluator_traits_dispatcher;
 template<typename Lhs, typename Rhs>
 struct evaluator_traits<Product<Lhs, Rhs> >
   : product_evaluator_traits_dispatcher<Product<Lhs, Rhs>, typename ProductReturnType<Lhs, Rhs>::Type> 
-{ };
+{ 
+  static const int AssumeAliasing = 1;
+};
 
 // Case 1: Evaluate all at once
 //