Make sure the shader execution heap is cache aligned (#1159)

Author: lgritz

src/liboslexec/context.cpp

@@ -71,15 +71,10 @@ ShadingContext::execute_init (ShaderGroup &sgroup, ShaderGlobals &ssg, bool run)
 
     // Allocate enough space on the heap
     size_t heap_size_needed = sgroup.llvm_groupdata_size();
-    if (heap_size_needed > m_heap.size()) {
-        if (shadingsys().debug())
-            infof("  ShadingContext %p growing heap to %d",
-                  static_cast<void*>(this), heap_size_needed);
-        m_heap.resize (heap_size_needed);
-    }
+    reserve_heap(heap_size_needed);
     // Zero out the heap memory we will be using
     if (shadingsys().m_clearmemory)
-        memset (&m_heap[0], 0, heap_size_needed);
+        memset (m_heap.get(), 0, heap_size_needed);
 
     // Set up closure storage
     m_closure_pool.clear();
@@ -100,7 +95,7 @@ ShadingContext::execute_init (ShaderGroup &sgroup, ShaderGlobals &ssg, bool run)
         ssg.context = this;
         ssg.renderer = renderer();
         ssg.Ci = NULL;
-        run_func (&ssg, &m_heap[0]);
+        run_func (&ssg, m_heap.get());
     }
 
     if (profile)
@@ -124,7 +119,7 @@ ShadingContext::execute_layer (ShaderGlobals &ssg, int layernumber)
     if (! run_func)
         return false;
 
-    run_func (&ssg, &m_heap[0]);
+    run_func (&ssg, m_heap.get());
 
     if (profile)
         m_ticks += timer.ticks();
@@ -257,9 +252,9 @@ ShadingContext::symbol_data (const Symbol &sym) const
     if (! sgroup.optimized())
         return NULL;   // can't retrieve symbol if we didn't optimize it
 
-    if (sym.dataoffset() >= 0 && (int)m_heap.size() > sym.dataoffset()) {
+    if (sym.dataoffset() >= 0 && (int)m_heapsize > sym.dataoffset()) {
         // lives on the heap
-        return &m_heap[sym.dataoffset()];
+        return m_heap.get() + sym.dataoffset();
     }
 
     // doesn't live on the heap

src/liboslexec/oslexec_pvt.h

@@ -1741,6 +1741,13 @@ class OSLEXECPUBLIC ShadingContext {
         record_error(ErrorHandler::EH_MESSAGE, Strutil::sprintf (fmt, args...));
     }
 
+    void reserve_heap(size_t size) {
+        if (size > m_heapsize) {
+            m_heap.reset((char *)OIIO::aligned_malloc(size, OIIO_CACHE_LINE_SIZE));
+            m_heapsize = size;
+        }
+    }
+
 private:
 
     void free_dict_resources ();
@@ -1750,7 +1757,9 @@ class OSLEXECPUBLIC ShadingContext {
     PerThreadInfo *m_threadinfo;        ///< Ptr to our thread's info
     mutable TextureSystem::Perthread *m_texture_thread_info; ///< Ptr to texture thread info
     ShaderGroup *m_group;               ///< Ptr to shader group
-    std::vector<char> m_heap;           ///< Heap memory
+    // Heap memory
+    std::unique_ptr<char, decltype(&OIIO::aligned_free)> m_heap { nullptr, &OIIO::aligned_free };
+    size_t m_heapsize = 0;
     typedef std::unordered_map<ustring, std::unique_ptr<regex>, ustringHash> RegexMap;
     RegexMap m_regex_map;               ///< Compiled regex's
     MessageList m_messages;             ///< Message blackboard