fix bug #195: fast unaligned load for integer using _mm_load_sd failed when the value interpreted as a NaN

1 files changed, 5 insertions, 9 deletions

diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index bbe784523..8401efe0b 100644
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -237,7 +237,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { E
     #endif
   }
   template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); }
-  template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*    from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); }
 #else
 // Fast unaligned loads. Note that here we cannot directly use intrinsics: this would
 // require pointer casting to incompatible pointer types and leads to invalid code
@@ -261,16 +260,13 @@ template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
   res = _mm_loadh_pd(res,from+1);
   return res;
 }
-template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
-{
-  EIGEN_DEBUG_UNALIGNED_LOAD
-  __m128d res;
-  res =  _mm_load_sd((const double*)(from)) ;
-  res =  _mm_loadh_pd(res, (const double*)(from+2)) ;
-  return _mm_castpd_si128(res);
-}
 #endif
 
+// bug 195: we used to have an optimized ploadu using _mm_load_sd/_mm_loadh_pd but that gave wrong results when some 64bit value,
+// interpreted as double, was a NaN
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*    from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); }
+
+
 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
 {
   return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd((const double*)from)), 0, 0, 1, 1);