AVL: Remove obsolete branching optimizations

Modern Clang and GCC can successfully implement simple conditions
without branching with math and flag operations.  Use of arrays for
translation no longer helps as much as it was 14+ years ago.

Disassemble of the code generated by Clang 13.0.0 on FreeBSD 13.1,
Clang 14.0.4 on FreeBSD 14 and GCC 10.2.1 on Debian 11 with this
change still shows no branching instructions.

Profiling of CPU-bound scan stage of sorted scrub shows reproducible
reduction of time spent inside avl_find() from 6.52% to 4.58%.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes #13540
This commit is contained in:
Alexander Motin 2022-06-09 18:27:36 -04:00 committed by GitHub
parent cdeb98a116
commit fc5200aa9b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -108,21 +108,6 @@
#include <sys/cmn_err.h>
#include <sys/mod.h>
/*
* Small arrays to translate between balance (or diff) values and child indices.
*
* Code that deals with binary tree data structures will randomly use
* left and right children when examining a tree. C "if()" statements
* which evaluate randomly suffer from very poor hardware branch prediction.
* In this code we avoid some of the branch mispredictions by using the
* following translation arrays. They replace random branches with an
* additional memory reference. Since the translation arrays are both very
* small the data should remain efficiently in cache.
*/
static const int avl_child2balance[] = {-1, 1};
static const int avl_balance2child[] = {0, 0, 1};
/*
* Walk from one node to the previous valued node (ie. an infix walk
* towards the left). At any given node we do one of 2 things:
@ -278,8 +263,7 @@ avl_find(avl_tree_t *tree, const void *value, avl_index_t *where)
#endif
return (AVL_NODE2DATA(node, off));
}
child = avl_balance2child[1 + diff];
child = (diff > 0);
}
if (where != NULL)
@ -527,7 +511,7 @@ avl_insert(avl_tree_t *tree, void *new_data, avl_index_t where)
* Compute the new balance
*/
old_balance = AVL_XBALANCE(node);
new_balance = old_balance + avl_child2balance[which_child];
new_balance = old_balance + (which_child ? 1 : -1);
/*
* If we introduced equal balance, then we are done immediately
@ -693,7 +677,7 @@ avl_remove(avl_tree_t *tree, void *data)
* choose node to swap from whichever side is taller
*/
old_balance = AVL_XBALANCE(delete);
left = avl_balance2child[old_balance + 1];
left = (old_balance > 0);
right = 1 - left;
/*
@ -777,7 +761,7 @@ avl_remove(avl_tree_t *tree, void *data)
*/
node = parent;
old_balance = AVL_XBALANCE(node);
new_balance = old_balance - avl_child2balance[which_child];
new_balance = old_balance - (which_child ? 1 : -1);
parent = AVL_XPARENT(node);
which_child = AVL_XCHILD(node);