Make sure that calls to broadcast4 are 16 bytes aligned

Eigen/src/Core/arch/SSE/PacketMath.h

diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index ad935d5..6912f3b 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -486,7 +486,7 @@
 pbroadcast4<Packet4f>(const float *a,
                       Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
 {
-  a3 = ploadu<Packet4f>(a);
+  a3 = pload<Packet4f>(a);
   a0 = vec4f_swizzle1(a3, 0,0,0,0);
   a1 = vec4f_swizzle1(a3, 1,1,1,1);
   a2 = vec4f_swizzle1(a3, 2,2,2,2);
@@ -502,10 +502,10 @@
   a2 = _mm_loaddup_pd(a+2);
   a3 = _mm_loaddup_pd(a+3);
 #else
-  a1 = ploadu<Packet2d>(a);
+  a1 = pload<Packet2d>(a);
   a0 = vec2d_swizzle1(a1, 0,0);
   a1 = vec2d_swizzle1(a1, 1,1);
-  a3 = ploadu<Packet2d>(a+2);
+  a3 = pload<Packet2d>(a+2);
   a2 = vec2d_swizzle1(a3, 0,0);
   a3 = vec2d_swizzle1(a3, 1,1);
 #endif

Eigen/src/Core/products/TriangularMatrixMatrix.h

diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index 62575af..8088aa6 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -300,6 +300,7 @@
       Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc;
 
       Scalar* geb = blockB+ts*ts;
+      geb = geb + internal::first_aligned(geb,EIGEN_ALIGN_BYTES/sizeof(Scalar));
 
       pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, actual_kc, rs);