From 5fb91f26ed5ca964a105868fffa377efbb2a2b27 Mon Sep 17 00:00:00 2001
From: Matthias Hochsteger <matthias.hochsteger@tuwien.ac.at>
Date: Wed, 10 Jul 2019 13:21:57 +0200
Subject: [PATCH] User AlignedAlloc for over-aligned types

MacOS < 10.13 doesn't support new with alignment>16 bytes
---
 CMakeLists.txt              |  1 +
 libsrc/core/taskmanager.hpp |  4 ++--
 libsrc/core/utils.hpp       | 29 +++++++++++++++++++++++++++
 libsrc/general/ngarray.hpp  | 12 ++++++------
 libsrc/general/ngsimd.hpp   | 39 +++----------------------------------
 5 files changed, 41 insertions(+), 44 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2e899ac1..d573504d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -115,6 +115,7 @@ endif (ADDITIONAL_PATHS)
 #######################################################################
 # build options
 include_directories ("${PROJECT_SOURCE_DIR}/include")
+include_directories ("${PROJECT_SOURCE_DIR}/libsrc")
 include_directories ("${PROJECT_SOURCE_DIR}/libsrc/include")
 include_directories ("${PROJECT_BINARY_DIR}")
 
diff --git a/libsrc/core/taskmanager.hpp b/libsrc/core/taskmanager.hpp
index e9056659..059a1678 100644
--- a/libsrc/core/taskmanager.hpp
+++ b/libsrc/core/taskmanager.hpp
@@ -40,7 +40,7 @@ namespace ngcore
   {
 //     PajeTrace *trace;
 
-    class alignas(64) NodeData //: public AlignedAlloc<NodeData>
+    class alignas(64) NodeData : public AlignedAlloc<NodeData>
     {
     public:
       atomic<int> start_cnt{0};
@@ -390,7 +390,7 @@ public:
 
 
    
-   class alignas(4096) AtomicRange //: public AlignedAlloc<AtomicRange>
+   class alignas(4096) AtomicRange : public AlignedAlloc<AtomicRange>
   {
     atomic<size_t> begin;
     atomic<size_t> end;
diff --git a/libsrc/core/utils.hpp b/libsrc/core/utils.hpp
index 4dca6c8b..35877d42 100644
--- a/libsrc/core/utils.hpp
+++ b/libsrc/core/utils.hpp
@@ -74,6 +74,35 @@ namespace ngcore
       b = std::move(temp);
   }
 
+  template <typename T>
+  class AlignedAlloc
+  {
+    protected:
+      static void * aligned_malloc(size_t s)
+      {
+        // Assume 16 byte alignment of standard library
+        if(alignof(T)<=16)
+            return malloc(s);
+        else
+            return  _mm_malloc(s, alignof(T));
+      }
+
+      static void aligned_free(void *p)
+      {
+        if(alignof(T)<=16)
+            free(p);
+        else
+            _mm_free(p);
+      }
+
+  public:
+    void * operator new (size_t s, void *p) { return p; }
+    void * operator new (size_t s) { return aligned_malloc(s); }
+    void * operator new[] (size_t s) { return aligned_malloc(s); }
+    void operator delete (void * p) { aligned_free(p); }
+    void operator delete[] (void * p) { aligned_free(p); }
+  };
+
 } // namespace ngcore
 
 #endif // NETGEN_CORE_UTILS_HPP
diff --git a/libsrc/general/ngarray.hpp b/libsrc/general/ngarray.hpp
index 2c4aed21..4b66d65d 100644
--- a/libsrc/general/ngarray.hpp
+++ b/libsrc/general/ngarray.hpp
@@ -388,10 +388,10 @@ namespace netgen
 
     NgArray & operator= (NgArray && a2)
     {
-      Swap (data, a2.data);
-      Swap (size, a2.size);
-      Swap (allocsize, a2.allocsize);
-      Swap (ownmem, a2.ownmem);
+      ngcore::Swap (data, a2.data);
+      ngcore::Swap (size, a2.size);
+      ngcore::Swap (allocsize, a2.allocsize);
+      ngcore::Swap (ownmem, a2.ownmem);
       return *this;
     }
 
@@ -745,8 +745,8 @@ namespace netgen
       
 	if (i <= j)
 	  {
-	    Swap (data[i], data[j]);
-	    Swap (slave[i], slave[j]);
+            ngcore::Swap (data[i], data[j]);
+            ngcore::Swap (slave[i], slave[j]);
 	    i++; j--;
 	  }
       }
diff --git a/libsrc/general/ngsimd.hpp b/libsrc/general/ngsimd.hpp
index 97e92052..feba2523 100644
--- a/libsrc/general/ngsimd.hpp
+++ b/libsrc/general/ngsimd.hpp
@@ -13,6 +13,8 @@
 #include <string>
 #include <type_traits>
 
+#include <core/utils.hpp>
+
 #ifdef WIN32
 #ifndef AVX_OPERATORS_DEFINED
 #define AVX_OPERATORS_DEFINED
@@ -48,6 +50,7 @@ NG_INLINE __m256d operator/= (__m256d &a, __m256d b) { return a = a/b; }
 
 namespace ngsimd
 {
+  using ngcore::AlignedAlloc;
 
   // MSVC does not define SSE. It's always present on 64bit cpus
 #if (defined(_M_AMD64) || defined(_M_X64) || defined(__AVX__))
@@ -121,42 +124,6 @@ namespace ngsimd
     NG_INLINE SIMD<double,N> operator/ (SIMD<double,N> a, T b) { return a / SIMD<double,N>(b); }
 
 
-#ifdef __AVX__
-  template <typename T>
-  class AlignedAlloc
-  {
-    protected:
-      static void * aligned_malloc(size_t s)
-      {
-        // Assume 16 byte alignment of standard library
-        if(alignof(T)<=16)
-            return malloc(s);
-        else
-            return  _mm_malloc(s, alignof(T));
-      }
-
-      static void aligned_free(void *p)
-      {
-        if(alignof(T)<=16)
-            free(p);
-        else
-            _mm_free(p);
-      }
-
-  public:
-    void * operator new (size_t s, void *p) { return p; }
-    void * operator new (size_t s) { return aligned_malloc(s); }
-    void * operator new[] (size_t s) { return aligned_malloc(s); }
-    void operator delete (void * p) { aligned_free(p); }
-    void operator delete[] (void * p) { aligned_free(p); }
-  };
-#else
-  // it's only a dummy without AVX
-  template <typename T>
-  class AlignedAlloc { ; };
-
-#endif
-
 using std::sqrt;
 using std::fabs;