From 1b793a127c7dedc23e6e3d9d673bd8985587988c Mon Sep 17 00:00:00 2001
From: Erik Hofman <erik@ehofman.com>
Date: Thu, 12 Jan 2017 10:18:39 +0100
Subject: [PATCH] Be more explicit about alignment

---
 simgear/math/simd.hxx    | 38 ++++++++++++++++++++++++++++++++++----
 simgear/math/simd4x4.hxx |  8 ++++----
 2 files changed, 38 insertions(+), 8 deletions(-)

diff --git a/simgear/math/simd.hxx b/simgear/math/simd.hxx
index 3074e342..4f3ed3ad 100644
--- a/simgear/math/simd.hxx
+++ b/simgear/math/simd.hxx
@@ -317,8 +317,23 @@ inline simd4_t<T,N> operator*(simd4_t<T,N> v, T f) {
 #  include <pmmintrin.h>
 # endif
 
+ALIGN16
+class simd_aligned16
+{
+public:
+    simd_aligned16() {}
+    ~simd_aligned16() {}
+
+    void *operator new (size_t size) {
+        return _mm_malloc(size, 16);
+    }
+    void operator delete (void *p) {
+        _mm_free(p);
+    }
+};
+
 template<int N>
-class simd4_t<float,N>
+class simd4_t<float,N> : public simd_aligned16
 {
 private:
    typedef float  __vec4f_t[N];
@@ -509,8 +524,23 @@ inline simd4_t<float,N>abs(simd4_t<float,N> v) {
 #  include <immintrin.h>
 // #  include "avxintrin-emu.h"
 
+ALIGN32
+class simd_aligned32
+{
+public:
+    simd_aligned32() {}
+    ~simd_aligned32() {}
+
+    void *operator new (size_t size) {
+        return _mm_malloc(size, 32);
+    }
+    void operator delete (void *p) {
+        _mm_free(p);
+    }
+};
+
 template<int N>
-class simd4_t<double,N>
+class simd4_t<double,N> : public simd_aligned32
 {
 private:
    typedef double  __vec4d_t[N];
@@ -683,7 +713,7 @@ inline simd4_t<double,N>abs(simd4_t<double,N> v) {
 #  include <emmintrin.h>
 
 template<int N>
-class simd4_t<double,N>
+class simd4_t<double,N> : public simd_aligned16
 {
 private:
    typedef double  __vec4d_t[N];
@@ -886,7 +916,7 @@ inline simd4_t<double,N>abs(simd4_t<double,N> v) {
 #  endif
 
 template<int N>
-class simd4_t<int,N>
+class simd4_t<int,N>  : public simd_aligned16
 {
 private:
    typedef int  __vec4i_t[N];
diff --git a/simgear/math/simd4x4.hxx b/simgear/math/simd4x4.hxx
index bbf6908b..0f355fe3 100644
--- a/simgear/math/simd4x4.hxx
+++ b/simgear/math/simd4x4.hxx
@@ -298,7 +298,7 @@ inline simd4x4_t<T,N> operator*(const simd4x4_t<T,N>& m1, const simd4x4_t<T,N>&
 #  include <xmmintrin.h>
 
 template<>
-class simd4x4_t<float,4>
+class simd4x4_t<float,4>  : public simd_aligned16
 {
 private:
     typedef float  __mtx4f_t[4][4];
@@ -520,7 +520,7 @@ inline simd4_t<float,3> transform<float>(const simd4x4_t<float,4>& m, const simd
 // #  include "avxintrin-emu.h"
 
 template<>
-class simd4x4_t<double,4>
+class simd4x4_t<double,4> : public simd_aligned32
 {
 private:
     typedef double  __mtx4d_t[4][4];
@@ -758,7 +758,7 @@ inline simd4_t<double,3> transform<double>(const simd4x4_t<double,4>& m, const s
 #  include <emmintrin.h>
 
 template<>
-class simd4x4_t<double,4>
+class simd4x4_t<double,4> : public simd_aligned16
 {
 private:
     typedef double  __mtx4d_t[4][4];
@@ -1052,7 +1052,7 @@ inline simd4_t<double,3> transform<double>(const simd4x4_t<double,4>& m, const s
 #  include <xmmintrin.h>
 
 template<>
-class simd4x4_t<int,4>
+class simd4x4_t<int,4> : public simd_aligned16
 {
 private:
     typedef int  __mtx4i_t[4][4];