From 626942d9ddcc17c21c2d79a690537e54237275bc Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 28 Sep 2018 16:57:32 +0200
Subject: [PATCH] fix alignment issue in ploaddup for AVX512

---
 Eigen/src/Core/arch/AVX512/PacketMath.h | 4 +++-
 test/packetmath.cpp                     | 2 ++
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index 9fbb256a1..e3f8c1c87 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -466,7 +466,9 @@ EIGEN_STRONG_INLINE Packet16i ploadu<Packet16i>(const int* from) {
 // {a0, a0  a1, a1, a2, a2, a3, a3, a4, a4, a5, a5, a6, a6, a7, a7}
 template <>
 EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) {
-  __m256i low_half = _mm256_load_si256(reinterpret_cast<const __m256i*>(from));
+  // an unaligned load is required here as there is no requirement
+  // on the alignment of input pointer 'from'
+  __m256i low_half = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from));
   __m512 even_elements = _mm512_castsi512_ps(_mm512_cvtepu32_epi64(low_half));
   __m512 pairs = _mm512_permute_ps(even_elements, _MM_SHUFFLE(2, 2, 0, 0));
   return pairs;
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 2b0dda573..babb7c20e 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -227,6 +227,7 @@ template<typename Scalar> void packetmath()
 
   if(PacketSize>1)
   {
+    // apply different offsets to check that ploaddup is robust to unaligned inputs
     for(int offset=0;offset<4;++offset)
     {
       for(int i=0;i<PacketSize/2;++i)
@@ -238,6 +239,7 @@ template<typename Scalar> void packetmath()
 
   if(PacketSize>2)
   {
+    // apply different offsets to check that ploadquad is robust to unaligned inputs
     for(int offset=0;offset<4;++offset)
     {
       for(int i=0;i<PacketSize/4;++i)