Added OPENVDB_ASSUME macros and applied them to Leaf ValueIters - fixes vectorization misses with clang

Signed-off-by: Nick Avramoussis <[email protected]>

Author: Idclip

openvdb/openvdb/Platform.h

@@ -92,6 +92,29 @@
 #define OPENVDB_UNLIKELY(x) (x)
 #endif
 
+/// Macros for assume builtins. Note that we currently don't simply apply these
+/// in place of asserts (when asserts are disabled) - they should be only be
+/// applied with an assert once profiled
+#ifdef __has_cpp_attribute
+  #if __has_cpp_attribute(assume) >= 202207L
+    #define OPENVDB_ASSUME(...) [[assume(__VA_ARGS__)]]
+  #endif
+#endif
+#ifndef OPENVDB_ASSUME
+  #if defined(__clang__)
+    #define OPENVDB_ASSUME(...)  __builtin_assume(__VA_ARGS__);
+  #elif defined(_MSC_VER)
+    #define OPENVDB_ASSUME(...) __assume(__VA_ARGS__);
+  #elif defined(__GNUC__)
+    #if __GNUC__ >= 13
+      #define OPENVDB_ASSUME(...) __attribute__((__assume__(__VA_ARGS__)))
+    #endif
+  #endif
+#endif
+#ifndef OPENVDB_ASSUME
+  #define OPENVDB_ASSUME(...)
+#endif
+
 /// Force inline function macros. These macros do not necessary guarantee that
 /// the decorated function will be inlined, but provide the strongest vendor
 /// annotations to that end.

openvdb/openvdb/tree/LeafNode.h

@@ -242,6 +242,8 @@ class LeafNode
                     "ValueIter::setItem cannot be called on const iterators");
             }
             else {
+                OPENVDB_ASSERT(pos < SIZE);
+                OPENVDB_ASSUME(pos < SIZE);
                 mData[pos] = value;
             }
         }
@@ -257,6 +259,8 @@ class LeafNode
                     "ValueIter::modifyItem cannot be called on const iterators");
             }
             else {
+                OPENVDB_ASSERT(n < SIZE);
+                OPENVDB_ASSUME(n < SIZE);
                 op(mData[n]);
                 this->parent().setValueOn(n);
             }

pendingchanges/leaf_value_iters.txt

@@ -1,6 +1,8 @@
 OpenVDB:
  - Improvements:
-    - Significantly improved the performance of all LeafNode ValueIterators,
-    especially when delay loading is enabled. Construction of a ValueIterator
-    on a leaf node now requests the leaf buffers ahead of iteration to avoid
-    expensive API calls.
+   - Significantly improved the performance of all LeafNode ValueIterators,
+     up to 5x on some platforms and up to 10x when delay loading is enabled.
+     Construction of a ValueIterator from a leaf node now requests the leaf
+     buffers ahead of iteration to avoid potentially expensive API calls.
+   - Added OPENVDB_ASSUME macros to mimic builtin assume and C++23 assume
+     attributes.