tree: compact the node's parent and extra fields into one

This shrinks the tree node to just 3*sizeof(void *).

Running perf_tree yields (times in seconds):

           |     32-bit    |     64-bit
           | non-c | comp  | non-c | comp
-----------+-------+-------+-------+-------
insertions | 0.529 | 0.631 | 0.525 | 0.649
finds      | 5.655 | 6.515 | 5.523 | 6.362
deletions  | 0.334 | 0.310 | 0.260 | 0.328

Because the compacting slows down the tree operations 15-20%, the
non-compacted layout is still supported by rebuilding the library without
JEFFPC_TREE_COMPACT defined.

Signed-off-by: Josef 'Jeff' Sipek <[email protected]>

Author: jeffpc

cmake/config.cmake

@@ -93,6 +93,8 @@ endif()
 set(JEFFPC_LOCK_DEP_COUNT 16)
 set(JEFFPC_LOCK_STACK_DEPTH 32)
 
+set(JEFFPC_TREE_COMPACT 1)
+
 include("${CMAKE_DIR}/config-errno.cmake")
 
 configure_file("${CMAKE_CURRENT_SOURCE_DIR}/include/jeffpc/config.h.in"

include/jeffpc/config.h.in

@@ -48,4 +48,6 @@
 #cmakedefine JEFFPC_LOCK_DEP_COUNT ${JEFFPC_LOCK_DEP_COUNT}
 #cmakedefine JEFFPC_LOCK_STACK_DEPTH ${JEFFPC_LOCK_STACK_DEPTH}
 
+#cmakedefine JEFFPC_TREE_COMPACT	${JEFFPC_TREE_COMPACT}
+
 #endif

include/jeffpc/tree_private.h

@@ -26,11 +26,23 @@
 #include <stdbool.h>
 
 #include <jeffpc/int.h>
+#include <jeffpc/config.h>
 
 struct tree_node {
 	struct tree_node *children[2];
+#ifdef JEFFPC_TREE_COMPACT
+	/*
+	 * The following field encodes both the parent pointer and the 2-bit
+	 * extra field.
+	 *
+	 * The extra field uses the 2 least significant bits.  The remainder
+	 * of the value is the parent pointer.
+	 */
+	uintptr_t _parent_and_extra;
+#else
 	struct tree_node *_parent;
 	unsigned int _extra:2;
+#endif
 };
 
 struct tree_tree {

tree_impl.h

@@ -43,22 +43,40 @@ static inline struct tree_node *obj2node(struct tree_tree *tree, void *obj)
 
 static inline struct tree_node *get_parent(struct tree_node *node)
 {
+#ifdef JEFFPC_TREE_COMPACT
+	return (void *) (node->_parent_and_extra & ~0x3ul);
+#else
 	return node->_parent;
+#endif
 }
 
 static inline void set_parent(struct tree_node *node, struct tree_node *parent)
 {
+#ifdef JEFFPC_TREE_COMPACT
+	node->_parent_and_extra = (node->_parent_and_extra & 0x3ul) |
+				  (((uintptr_t) parent) & ~0x3ul);
+#else
 	node->_parent = parent;
+#endif
 }
 
 static inline unsigned int get_extra(struct tree_node *node)
 {
+#ifdef JEFFPC_TREE_COMPACT
+	return node->_parent_and_extra & 0x3;
+#else
 	return node->_extra;
+#endif
 }
 
 static inline void set_extra(struct tree_node *node, unsigned int extra)
 {
+#ifdef JEFFPC_TREE_COMPACT
+	node->_parent_and_extra = (node->_parent_and_extra & ~0x3ul) |
+				  (extra & 0x3);
+#else
 	node->_extra = extra;
+#endif
 }
 
 static inline enum tree_dir which_dir(struct tree_node *parent,