st.c: use power-of-two sizes to avoid slow modulo ops

* st.c (hash_pos): use bitwise AND to avoid slow modulo op
  (new_size): power-of-two sizes for hash_pos change
  (st_numhash): adjust for common keys due to lack of prime modulo
  [Feature ruby#9425]
* hash.c (rb_any_hash): right shift for symbols
* benchmark/bm_hash_aref_miss.rb: added to show improvement
* benchmark/bm_hash_aref_sym_long.rb: ditto
* benchmark/bm_hash_aref_str.rb: ditto
* benchmark/bm_hash_aref_sym.rb: ditto
* benchmark/bm_hash_ident_num.rb: added to prevent regression
* benchmark/bm_hash_ident_obj.rb: ditto
* benchmark/bm_hash_ident_str.rb: ditto
* benchmark/bm_hash_ident_sym.rb: ditto

git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@45384 b2dd03c8-39d4-4d8f-98ff-823fe69b080e

Author: normal

ChangeLog

@@ -1,3 +1,19 @@
+Sun Mar 23 08:12:27 2014  Eric Wong  <[email protected]>
+
+	* st.c (hash_pos): use bitwise AND to avoid slow modulo op
+	  (new_size): power-of-two sizes for hash_pos change
+	  (st_numhash): adjust for common keys due to lack of prime modulo
+	  [Feature #9425]
+	* hash.c (rb_any_hash): right shift for symbols
+	* benchmark/bm_hash_aref_miss.rb: added to show improvement
+	* benchmark/bm_hash_aref_sym_long.rb: ditto
+	* benchmark/bm_hash_aref_str.rb: ditto
+	* benchmark/bm_hash_aref_sym.rb: ditto
+	* benchmark/bm_hash_ident_num.rb: added to prevent regression
+	* benchmark/bm_hash_ident_obj.rb: ditto
+	* benchmark/bm_hash_ident_str.rb: ditto
+	* benchmark/bm_hash_ident_sym.rb: ditto
+
 Sat Mar 22 22:56:45 2014  NARUSE, Yui  <[email protected]>
 
 	* addr2line.c (fill_lines): compare the file names of object in which

NEWS

@@ -84,3 +84,9 @@ with all sufficient information, see the ChangeLog file.
 
 * struct RBignum is hidden.  [Feature #6083]
   Use rb_integer_pack and rb_integer_unpack instead.
+
+* st hash table uses power-of-two sizes for speed [Feature #9425].
+  Lookups are 10-25% faster if using appropriate hash functions.
+  However, weaknesses in hash distribution can no longer be masked
+  by prime number-sized tables, so extensions may need to tweak
+  hash functions to ensure good distribution.

benchmark/bm_hash_aref_miss.rb

@@ -0,0 +1,5 @@
+h = {}
+strs = ('a'..'z').to_a.map!(&:freeze)
+strs.each { |s| h[s] = s }
+strs = ('A'..'Z').to_a
+200_000.times { strs.each { |s| h[s] } }

benchmark/bm_hash_aref_str.rb

@@ -0,0 +1,4 @@
+h = {}
+strs = ('a'..'z').to_a.map!(&:freeze)
+strs.each { |s| h[s] = s }
+200_000.times { strs.each { |s| h[s] } }

benchmark/bm_hash_aref_sym.rb

@@ -0,0 +1,4 @@
+h = {}
+syms = ('a'..'z').to_a.map(&:to_sym)
+syms.each { |s| h[s] = s }
+200_000.times { syms.each { |s| h[s] } }

benchmark/bm_hash_aref_sym_long.rb

@@ -0,0 +1,8 @@
+h = {}
+syms = %w[puts warn syswrite write stat bacon lettuce tomato
+some symbols in this array may already be interned  others should not be
+hash browns make good breakfast but not cooked using prime numbers
+shift for division entries delete_if keys exist?
+].map!(&:to_sym)
+syms.each { |s| h[s] = s }
+200_000.times { syms.each { |s| h[s] } }

benchmark/bm_hash_ident_num.rb

@@ -0,0 +1,4 @@
+h = {}.compare_by_identity
+nums = (1..26).to_a
+nums.each { |n| h[n] = n }
+200_000.times { nums.each { |n| h[n] } }

benchmark/bm_hash_ident_obj.rb

@@ -0,0 +1,4 @@
+h = {}.compare_by_identity
+objs = 26.times.map { Object.new }
+objs.each { |o| h[o] = o }
+200_000.times { objs.each { |o| h[o] } }

benchmark/bm_hash_ident_str.rb

@@ -0,0 +1,4 @@
+h = {}.compare_by_identity
+strs = ('a'..'z').to_a
+strs.each { |s| h[s] = s }
+200_000.times { strs.each { |s| h[s] } }

benchmark/bm_hash_ident_sym.rb

@@ -0,0 +1,4 @@
+h = {}.compare_by_identity
+syms = ('a'..'z').to_a.map(&:to_sym)
+syms.each { |s| h[s] = s }
+200_000.times { syms.each { |s| h[s] } }

hash.c

@@ -132,6 +132,7 @@ rb_any_hash(VALUE a)
 
     if (SPECIAL_CONST_P(a)) {
 	if (a == Qundef) return 0;
+	if (SYMBOL_P(a)) a >>= (RUBY_SPECIAL_SHIFT + 3); /* 3=ID_SCOPE_SHIFT */
 	hnum = rb_objid_hash((st_index_t)a);
     }
     else if (BUILTIN_TYPE(a) == T_STRING) {

st.c

@@ -33,8 +33,7 @@ typedef struct st_packed_entry {
 #define STATIC_ASSERT(name, expr) typedef int static_assert_##name##_check[(expr) ? 1 : -1];
 
 #define ST_DEFAULT_MAX_DENSITY 5
-#define ST_DEFAULT_INIT_TABLE_SIZE 11
-#define ST_DEFAULT_SECOND_TABLE_SIZE 19
+#define ST_DEFAULT_INIT_TABLE_SIZE 16
 #define ST_DEFAULT_PACKED_TABLE_SIZE 18
 #define PACKED_UNIT (int)(sizeof(st_packed_entry) / sizeof(st_table_entry*))
 #define MAX_PACKED_HASH (int)(ST_DEFAULT_PACKED_TABLE_SIZE * sizeof(st_table_entry*) / sizeof(st_packed_entry))
@@ -85,7 +84,7 @@ static void rehash(st_table *);
 #define EQUAL(table,x,y) ((x)==(y) || (*(table)->type->compare)((x),(y)) == 0)
 
 #define do_hash(key,table) (st_index_t)(*(table)->type->hash)((key))
-#define hash_pos(h,n) ((h) % (n))
+#define hash_pos(h,n) ((h) & (n - 1))
 #define do_hash_bin(key,table) hash_pos(do_hash((key), (table)), (table)->num_bins)
 
 /* preparation for possible allocation improvements */
@@ -140,69 +139,18 @@ remove_safe_packed_entry(st_table *table, st_index_t i, st_data_t never)
     PHASH_SET(table, i, 0);
 }
 
-/*
- * MINSIZE is the minimum size of a dictionary.
- */
-
-#define MINSIZE 8
-
-/*
-Table of prime numbers 2^n+a, 2<=n<=30.
-*/
-static const unsigned int primes[] = {
-	ST_DEFAULT_INIT_TABLE_SIZE,
-	ST_DEFAULT_SECOND_TABLE_SIZE,
-	32 + 5,
-	64 + 3,
-	128 + 3,
-	256 + 27,
-	512 + 9,
-	1024 + 9,
-	2048 + 5,
-	4096 + 3,
-	8192 + 27,
-	16384 + 43,
-	32768 + 3,
-	65536 + 45,
-	131072 + 29,
-	262144 + 3,
-	524288 + 21,
-	1048576 + 7,
-	2097152 + 17,
-	4194304 + 15,
-	8388608 + 9,
-	16777216 + 43,
-	33554432 + 35,
-	67108864 + 15,
-	134217728 + 29,
-	268435456 + 3,
-	536870912 + 11,
-	1073741824 + 85,
-	0
-};
-
 static st_index_t
 new_size(st_index_t size)
 {
-    int i;
+    st_index_t i;
 
-#if 0
     for (i=3; i<31; i++) {
-	if ((1<<i) > size) return 1<<i;
-    }
-    return -1;
-#else
-    st_index_t newsize;
-
-    for (i = 0, newsize = MINSIZE; i < numberof(primes); i++, newsize <<= 1) {
-	if (newsize > size) return primes[i];
+	if ((st_index_t)(1<<i) > size) return 1<<i;
     }
-    /* Ran out of primes */
 #ifndef NOT_RUBY
     rb_raise(rb_eRuntimeError, "st_table too big");
 #endif
     return -1;			/* should raise exception */
-#endif
 }
 
 #ifdef HASH_LOG
@@ -1685,5 +1633,17 @@ st_numcmp(st_data_t x, st_data_t y)
 st_index_t
 st_numhash(st_data_t n)
 {
-    return (st_index_t)n;
+    /*
+     * This hash function is lightly-tuned for Ruby.  Further tuning
+     * should be possible.  Notes:
+     *
+     * - (n >> 3) alone is great for heap objects and OK for fixnum,
+     *   however symbols perform poorly.
+     * - (n >> (RUBY_SPECIAL_SHIFT+3)) was added to make symbols hash well,
+     *   n.b.: +3 to remove ID scope, +1 worked well initially, too
+     * - (n << 3) was finally added to avoid losing bits for fixnums
+     * - avoid expensive modulo instructions, it is currently only
+     *   shifts and bitmask operations.
+     */
+    return (st_index_t)((n>>(RUBY_SPECIAL_SHIFT+3)|(n<<3)) ^ (n>>3));
 }