Set cost of conjugate to 0 (in practice it boils down to a no-op).

This is also important to make sure that A.conjugate() * B.conjugate() does not evaluate
its arguments into temporaries (e.g., if A and B are fixed and small, or * fall back to lazyProduct)

2 files changed, 11 insertions, 1 deletions

diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h
index 55994047e..1d5eb3678 100644
--- a/Eigen/src/Core/functors/UnaryFunctors.h
+++ b/Eigen/src/Core/functors/UnaryFunctors.h
@@ -117,7 +117,15 @@ template<typename Scalar>
 struct functor_traits<scalar_conjugate_op<Scalar> >
 {
   enum {
-    Cost = NumTraits<Scalar>::IsComplex ? NumTraits<Scalar>::AddCost : 0,
+    Cost = 0,
+    // Yes the cost is zero even for complexes because in most cases for which
+    // the cost is used, conjugation turns to be a no-op. Some examples:
+    //   cost(a*conj(b)) == cost(a*b)
+    //   cost(a+conj(b)) == cost(a+b)
+    //   <etc.
+    // If we don't set it to zero, then:
+    //   A.conjugate().lazyProduct(B.conjugate())
+    // will bake its operands. We definitely don't want that!
     PacketAccess = packet_traits<Scalar>::HasConj
   };
 };
diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp
index 7f169e6ae..8b6419d0c 100644
--- a/test/product_notemporary.cpp
+++ b/test/product_notemporary.cpp
@@ -134,7 +134,9 @@ template<typename MatrixType> void product_notemporary(const MatrixType& m)
   VERIFY_EVALUATION_COUNT( m3.noalias() = m1.block(r0,r0,r1,r1).template triangularView<UnitUpper>()  * m2.block(r0,c0,r1,c1), 1);
 
   // Zero temporaries for lazy products ...
+  m3.setRandom(rows,cols);
   VERIFY_EVALUATION_COUNT( Scalar tmp = 0; tmp += Scalar(RealScalar(1)) /  (m3.transpose().lazyProduct(m3)).diagonal().sum(), 0 );
+  VERIFY_EVALUATION_COUNT( m3.noalias() = m1.conjugate().lazyProduct(m2.conjugate()), 0);
 
   // ... and even no temporary for even deeply (>=2) nested products
   VERIFY_EVALUATION_COUNT( Scalar tmp = 0; tmp += Scalar(RealScalar(1)) /  (m3.transpose() * m3).diagonal().sum(), 0 );