3 files changed, 11 insertions, 15 deletions

diff --git a/intern/cycles/device/device_cpu.cpp b/intern/cycles/device/device_cpu.cpp
index 5d279ebb965..837a8186064 100644
--- a/intern/cycles/device/device_cpu.cpp
+++ b/intern/cycles/device/device_cpu.cpp
@@ -43,7 +43,6 @@
 #include "render/buffers.h"
 #include "render/coverage.h"
 
-#include "util/util_aligned_malloc.h"
 #include "util/util_debug.h"
 #include "util/util_foreach.h"
 #include "util/util_function.h"
@@ -166,7 +165,7 @@ class CPUDevice : public Device {
   bool need_texture_info;
 
 #ifdef WITH_OSL
-  OSLGlobals *osl_globals;
+  OSLGlobals osl_globals;
 #endif
 
   bool use_split_kernel;
@@ -283,9 +282,7 @@ class CPUDevice : public Device {
     }
 
 #ifdef WITH_OSL
-    /* Must use aligned malloc due to concurrent hash map. */
-    osl_globals = util_aligned_new<OSLGlobals>();
-    kernel_globals.osl = osl_globals;
+    kernel_globals.osl = &osl_globals;
 #endif
     use_split_kernel = DebugFlags().cpu.split_kernel;
     if (use_split_kernel) {
@@ -320,9 +317,6 @@ class CPUDevice : public Device {
 
   ~CPUDevice()
   {
-#ifdef WITH_OSL
-    util_aligned_delete(osl_globals);
-#endif
     task_pool.stop();
     texture_info.free();
   }
@@ -498,7 +492,7 @@ class CPUDevice : public Device {
   void *osl_memory()
   {
 #ifdef WITH_OSL
-    return osl_globals;
+    return &osl_globals;
 #else
     return NULL;
 #endif
@@ -987,7 +981,7 @@ class CPUDevice : public Device {
     KernelGlobals kg = kernel_globals;
 
 #ifdef WITH_OSL
-    OSLShader::thread_init(&kg, &kernel_globals, osl_globals);
+    OSLShader::thread_init(&kg, &kernel_globals, &osl_globals);
 #endif
     for (int sample = 0; sample < task.num_samples; sample++) {
       for (int x = task.shader_x; x < task.shader_x + task.shader_w; x++)
@@ -1059,7 +1053,7 @@ class CPUDevice : public Device {
     kg.decoupled_volume_steps_index = 0;
     kg.coverage_asset = kg.coverage_object = kg.coverage_material = NULL;
 #ifdef WITH_OSL
-    OSLShader::thread_init(&kg, &kernel_globals, osl_globals);
+    OSLShader::thread_init(&kg, &kernel_globals, &osl_globals);
 #endif
     return kg;
   }
diff --git a/intern/cycles/render/osl.cpp b/intern/cycles/render/osl.cpp
index 4bc18e53c9b..6f927bd5c42 100644
--- a/intern/cycles/render/osl.cpp
+++ b/intern/cycles/render/osl.cpp
@@ -30,6 +30,7 @@
 #  include "kernel/osl/osl_services.h"
 #  include "kernel/osl/osl_shader.h"
 
+#  include "util/util_aligned_malloc.h"
 #  include "util/util_foreach.h"
 #  include "util/util_logging.h"
 #  include "util/util_md5.h"
@@ -224,7 +225,8 @@ void OSLShaderManager::shading_system_init()
   thread_scoped_lock lock(ss_shared_mutex);
 
   if (ss_shared_users == 0) {
-    services_shared = new OSLRenderServices(ts_shared);
+    /* Must use aligned new due to concurrent hash map. */
+    services_shared = util_aligned_new<OSLRenderServices>(ts_shared);
 
     string shader_path = path_get("shader");
 #  ifdef _WIN32
@@ -293,7 +295,7 @@ void OSLShaderManager::shading_system_free()
     delete ss_shared;
     ss_shared = NULL;
 
-    delete services_shared;
+    util_aligned_delete(services_shared);
     services_shared = NULL;
   }
 
diff --git a/intern/cycles/util/util_aligned_malloc.h b/intern/cycles/util/util_aligned_malloc.h
index 7115e4cb0c6..df7d93c056d 100644
--- a/intern/cycles/util/util_aligned_malloc.h
+++ b/intern/cycles/util/util_aligned_malloc.h
@@ -31,10 +31,10 @@ void *util_aligned_malloc(size_t size, int alignment);
 void util_aligned_free(void *ptr);
 
 /* Aligned new operator. */
-template<typename T> T *util_aligned_new()
+template<typename T, typename... Args> T *util_aligned_new(Args... args)
 {
   void *mem = util_aligned_malloc(sizeof(T), alignof(T));
-  return new (mem) T();
+  return new (mem) T(args...);
 }
 
 template<typename T> void util_aligned_delete(T *t)