From b1bd53aa6bec39c53de475c90987eece86c206d2 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sun, 1 May 2016 23:25:06 +0200
Subject: [PATCH] Fix performance regression: with AVX, unaligned stores were
 emitted instead of aligned ones for fixed size assignement.

---
 Eigen/src/Core/AssignEvaluator.h | 15 +++++++++++----
 Eigen/src/Core/CoreEvaluators.h  |  4 ++--
 2 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h
index 9d4b315a0..b1193e421 100644
--- a/Eigen/src/Core/AssignEvaluator.h
+++ b/Eigen/src/Core/AssignEvaluator.h
@@ -256,12 +256,13 @@ struct copy_using_evaluator_innervec_CompleteUnrolling
   enum {
     outer = Index / DstXprType::InnerSizeAtCompileTime,
     inner = Index % DstXprType::InnerSizeAtCompileTime,
-    JointAlignment = Kernel::AssignmentTraits::JointAlignment
+    JointAlignment = Kernel::AssignmentTraits::JointAlignment,
+    DefaultAlignment = unpacket_traits<PacketType>::alignment
   };
 
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
-    kernel.template assignPacketByOuterInner<Aligned, JointAlignment, PacketType>(outer, inner);
+    kernel.template assignPacketByOuterInner<DefaultAlignment, JointAlignment, PacketType>(outer, inner);
     enum { NextIndex = Index + unpacket_traits<PacketType>::size };
     copy_using_evaluator_innervec_CompleteUnrolling<Kernel, NextIndex, Stop>::run(kernel);
   }
@@ -277,9 +278,12 @@ template<typename Kernel, int Index_, int Stop>
 struct copy_using_evaluator_innervec_InnerUnrolling
 {
   typedef typename Kernel::PacketType PacketType;
+  enum {
+    DefaultAlignment = unpacket_traits<PacketType>::alignment
+  };
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
   {
-    kernel.template assignPacketByOuterInner<Aligned, Aligned, PacketType>(outer, Index_);
+    kernel.template assignPacketByOuterInner<DefaultAlignment, DefaultAlignment, PacketType>(outer, Index_);
     enum { NextIndex = Index_ + unpacket_traits<PacketType>::size };
     copy_using_evaluator_innervec_InnerUnrolling<Kernel, NextIndex, Stop>::run(kernel, outer);
   }
@@ -433,6 +437,9 @@ template<typename Kernel>
 struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, NoUnrolling>
 {
   typedef typename Kernel::PacketType PacketType;
+  enum {
+    DefaultAlignment = unpacket_traits<PacketType>::alignment
+  };
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     const Index innerSize = kernel.innerSize();
@@ -440,7 +447,7 @@ struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, NoUnrolling>
     const Index packetSize = unpacket_traits<PacketType>::size;
     for(Index outer = 0; outer < outerSize; ++outer)
       for(Index inner = 0; inner < innerSize; inner+=packetSize)
-        kernel.template assignPacketByOuterInner<Aligned, Aligned, PacketType>(outer, inner);
+        kernel.template assignPacketByOuterInner<DefaultAlignment, DefaultAlignment, PacketType>(outer, inner);
   }
 };
 
diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 388805f0d..932178f53 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -850,14 +850,14 @@ struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBa
   template<int StoreMode, typename PacketType>
   EIGEN_STRONG_INLINE
   void writePacket(Index row, Index col, const PacketType& x) 
-  { 
+  {
     return m_argImpl.template writePacket<StoreMode,PacketType>(m_startRow.value() + row, m_startCol.value() + col, x); 
   }
   
   template<int StoreMode, typename PacketType>
   EIGEN_STRONG_INLINE
   void writePacket(Index index, const PacketType& x) 
-  { 
+  {
     return writePacket<StoreMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
                                              RowsAtCompileTime == 1 ? index : 0,
                                              x);