Optimize LIKE with turbo-boyer-more algoritm

Docs/manual.texi

@@ -26728,6 +26728,12 @@ In the first statement, the @code{LIKE} value begins with a wildcard
 character.  In the second statement, the @code{LIKE} value is not a
 constant.
 
+MySQL 4.0 does another optimization on @code{LIKE}.  If you are using
+@code{... LIKE "%string%"} and @code{string} is longer than 3 characters
+then MySQL will use the turbo-boyer-more algorithm to once initialize
+the pattern for the string and then use this pattern to quickly search
+after the given string.
+
 @findex IS NULL, and indexes
 @cindex indexes, and @code{IS NULL}
 Searching using @code{column_name IS NULL} will use indexes if column_name
@@ -49310,6 +49316,8 @@ Our TODO section contains what we plan to have in 4.0. @xref{TODO MySQL 4.0}.
 
 @itemize @bullet
 @item
+Use turbo-boyer-more to speed up @code{LIKE "%keyword%"} searches.
+@item
 Fixed bug in @code{DROP DATABASE} with symlink.
 @item
 Fixed crash in @code{REPAIR ... USE_FRM}.

mysql-test/r/func_like.result

@@ -15,4 +15,15 @@ test
 select * from t1 where a like "te_t";
 a
 test
+select * from t1 where a like "%a%";
+a
+a
+abc
+abcd
+select * from t1 where a like "%abcd%";
+a
+abcd
+select * from t1 where a like "%abc\d%";
+a
+abcd
 drop table t1;

mysql-test/t/func_like.test

@@ -9,4 +9,12 @@ select * from t1 where a like "abc%";
 select * from t1 where a like "ABC%"; 
 select * from t1 where a like "test%"; 
 select * from t1 where a like "te_t"; 
+
+#
+# The following will test the boyer-more code
+#
+select * from t1 where a like "%a%";
+select * from t1 where a like "%abcd%";
+select * from t1 where a like "%abc\d%";
+
 drop table t1;

sql/item_cmpfunc.cc

@@ -1228,23 +1228,23 @@ void Item_func_like::fix_length_and_dec()
   //  cmp_type=STRING_RESULT;			// For quick select
 }
 
-
 longlong Item_func_like::val_int()
 {
-  String *res,*res2;
+  String* res = args[0]->val_str(&tmp_value1);
-  res=args[0]->val_str(&tmp_value1);
   if (args[0]->null_value)
   {
     null_value=1;
     return 0;
   }
-  res2=args[1]->val_str(&tmp_value2);
+  String* res2 = args[1]->val_str(&tmp_value2);
   if (args[1]->null_value)
   {
     null_value=1;
     return 0;
   }
   null_value=0;
+  if (canDoTurboBM)
+    return turboBM_matches(res->ptr(), res->length()) ? 1 : 0;
   if (binary)
     return wild_compare(*res,*res2,escape) ? 0 : 1;
   else
@@ -1268,6 +1268,51 @@ Item_func::optimize_type Item_func_like::select_optimize() const
   return OPTIMIZE_NONE;
 }
 
+bool Item_func_like::fix_fields(THD *thd,struct st_table_list *tlist)
+{
+  if (Item_bool_func2::fix_fields(thd, tlist))
+    return 1;
+
+  /*
+    TODO--we could do it for non-const, but we'd have to
+    recompute the tables for each row--probably not worth it.
+  */
+  if (args[1]->const_item() && !(specialflag & SPECIAL_NO_NEW_FUNC))
+  {
+    String* res2 = args[1]->val_str(&tmp_value2);
+    const size_t len   = res2->length();
+    const char*  first = res2->ptr();
+    const char*  last  = first + len - 1;
+    /*
+      len must be > 2 ('%pattern%')
+      heuristic: only do TurboBM for pattern_len > 2
+    */
+
+    if (len > MIN_TURBOBM_PATTERN_LEN + 2 &&
+	*first == wild_many &&
+	*last  == wild_many)
+    {
+      const char* tmp = first + 1;
+      for ( ; *tmp != wild_many && *tmp != wild_one && *tmp != escape; tmp++) ;
+      canDoTurboBM = tmp == last;
+    }
+
+    if (canDoTurboBM)
+    {
+      pattern     = first + 1;
+      pattern_len = len - 2;
+      DBUG_PRINT("TurboBM", ("Initializing pattern: '%s'...", first));
+      int* suff = (int*)thd->alloc(sizeof(int[pattern_len + 1]));
+      bmGs      = (int*)thd->alloc(sizeof(int[pattern_len + 1]));
+      bmBc      = (int*)thd->alloc(sizeof(int[alphabet_size]));
+      turboBM_compute_good_suffix_shifts(suff);
+      turboBM_compute_bad_character_shifts();
+      DBUG_PRINT("turboBM",("done"));
+    }
+  }
+  return 0;
+}
+
 #ifdef USE_REGEX
 
 bool
@@ -1307,7 +1352,6 @@ Item_func_regex::fix_fields(THD *thd,TABLE_LIST *tables)
   return 0;
 }
 
-
 longlong Item_func_regex::val_int()
 {
   char buff[MAX_FIELD_WIDTH];
@@ -1364,3 +1408,215 @@ Item_func_regex::~Item_func_regex()
 }
 
 #endif /* USE_REGEX */
+
+
+#ifdef LIKE_CMP_TOUPPER
+#define likeconv(A) (uchar) toupper(A)
+#else
+#define likeconv(A) (uchar) my_sort_order[(uchar) (A)]
+#endif
+
+
+/**********************************************************************
+  turboBM_compute_suffixes()
+  Precomputation dependent only on pattern_len.
+**********************************************************************/
+
+void Item_func_like::turboBM_compute_suffixes(int* suff)
+{
+  const int   plm1 = pattern_len - 1;
+  int            f = 0;
+  int            g = plm1;
+  int* const splm1 = suff + plm1;
+
+  *splm1 = pattern_len;
+
+  if (binary)
+  {
+    int i;
+    for (i = pattern_len - 2; i >= 0; i--)
+    {
+      int tmp = *(splm1 + i - f);
+      if (g < i && tmp < i - g)
+	suff[i] = tmp;
+      else
+      {
+	if (i < g)
+	  g = i; // g = min(i, g)
+	f = i;
+	while (g >= 0 && pattern[g] == pattern[g + plm1 - f])
+	  g--;
+	suff[i] = f - g;
+      }
+    }
+  }
+  else
+  {
+    int i;
+    for (i = pattern_len - 2; 0 <= i; --i)
+    {
+      int tmp = *(splm1 + i - f);
+      if (g < i && tmp < i - g)
+	suff[i] = tmp;
+      else
+      {
+	if (i < g)
+	  g = i; // g = min(i, g)
+	f = i;
+	while (g >= 0 && likeconv(pattern[g]) == likeconv(pattern[g + plm1 - f]))
+	  g--;
+	suff[i] = f - g;
+      }
+    }
+  }
+}
+
+
+/**********************************************************************
+   turboBM_compute_good_suffix_shifts()
+   Precomputation dependent only on pattern_len.
+**********************************************************************/
+
+void Item_func_like::turboBM_compute_good_suffix_shifts(int* suff)
+{
+  turboBM_compute_suffixes(suff);
+
+  int* end = bmGs + pattern_len;
+  int* k;
+  for (k = bmGs; k < end; k++)
+    *k = pattern_len;
+
+  int tmp;
+  int i;
+  int j          = 0;
+  const int plm1 = pattern_len - 1;
+  for (i = plm1; i > -1; i--)
+  {
+    if (suff[i] == i + 1)
+    {
+      for (tmp = plm1 - i; j < tmp; j++)
+      {
+	int* tmp2 = bmGs + j;
+	if (*tmp2 == pattern_len)
+	  *tmp2 = tmp;
+      }
+    }
+  }
+
+  int* tmp2;
+  for (tmp = plm1 - i; j < tmp; j++)
+  {
+    tmp2 = bmGs + j;
+    if (*tmp2 == pattern_len)
+      *tmp2 = tmp;
+  }
+
+  tmp2 = bmGs + plm1;
+  for (i = 0; i <= pattern_len - 2; i++)
+    *(tmp2 - suff[i]) = plm1 - i;
+}
+
+
+/**********************************************************************
+   turboBM_compute_bad_character_shifts()
+   Precomputation dependent on pattern_len.
+**********************************************************************/
+
+void Item_func_like::turboBM_compute_bad_character_shifts()
+{
+  int*   i;
+  int* end = bmBc + alphabet_size;
+  for (i = bmBc; i < end; i++)
+    *i = pattern_len;
+
+  int j;
+  const int plm1 = pattern_len - 1;
+  if (binary)
+    for (j = 0; j < plm1; j++)
+      bmBc[pattern[j]] = plm1 - j;
+  else
+    for (j = 0; j < plm1; j++)
+      bmBc[likeconv(pattern[j])] = plm1 - j;
+}
+
+
+/**********************************************************************
+  turboBM_matches()
+  Search for pattern in text, returns true/false for match/no match
+**********************************************************************/
+
+bool Item_func_like::turboBM_matches(const char* text, int text_len) const
+{
+  register int bcShift;
+  register int turboShift;
+  int shift = pattern_len;
+  int j     = 0;
+  int u     = 0;
+
+  const int plm1  = pattern_len - 1;
+  const int tlmpl =    text_len - pattern_len;
+
+  /* Searching */
+  if (binary)
+  {
+    while (j <= tlmpl)
+    {
+      register int i = plm1;
+      while (i >= 0 && pattern[i] == text[i + j])
+      {
+	i--;
+	if (i == plm1 - shift)
+	  i -= u;
+      }
+      if (i < 0)
+	return true;
+
+      register const int v = plm1 - i;
+      turboShift = u - v;
+      bcShift    = bmBc[text[i + j]] - plm1 + i;
+      shift      = max(turboShift, bcShift);
+      shift      = max(shift, bmGs[i]);
+      if (shift == bmGs[i])
+	u = min(pattern_len - shift, v);
+      else
+      {
+	if (turboShift < bcShift)
+	  shift = max(shift, u + 1);
+	u = 0;
+      }
+      j += shift;
+    }
+    return false;
+  }
+  else
+  {
+    while (j <= tlmpl)
+    {
+      register int i = plm1;
+      while (i >= 0 && likeconv(pattern[i]) == likeconv(text[i + j]))
+      {
+	i--;
+	if (i == plm1 - shift)
+	  i -= u;
+      }
+      if (i < 0)
+	return true;
+
+      register const int v = plm1 - i;
+      turboShift = u - v;
+      bcShift    = bmBc[likeconv(text[i + j])] - plm1 + i;
+      shift      = max(turboShift, bcShift);
+      shift      = max(shift, bmGs[i]);
+      if (shift == bmGs[i])
+	u = min(pattern_len - shift, v);
+      else
+      {
+	if (turboShift < bcShift)
+	  shift = max(shift, u + 1);
+	u = 0;
+      }
+      j += shift;
+    }
+    return false;
+  }
+}

sql/item_cmpfunc.h

@@ -478,15 +478,40 @@ public:
 class Item_func_like :public Item_bool_func2
 {
   char escape;
+
+  // Turbo Boyer-Moore data
+  bool        canDoTurboBM;	// pattern is '%abcd%' case
+  const char* pattern;
+  int         pattern_len;
+
+  // TurboBM buffers, *this is owner
+  int* bmGs; //   good suffix shift table, size is pattern_len + 1
+  int* bmBc; // bad character shift table, size is alphabet_size
+
+  void turboBM_compute_suffixes(int* suff);
+  void turboBM_compute_good_suffix_shifts(int* suff);
+  void turboBM_compute_bad_character_shifts();
+  bool turboBM_matches(const char* text, int text_len) const;
+  enum { alphabet_size = 256 };
+
 public:
-  Item_func_like(Item *a,Item *b, char* escape_arg) :Item_bool_func2(a,b),escape(*escape_arg)
+  Item_func_like::Item_func_like(Item *a,Item *b, char* escape_arg) :
+    Item_bool_func2(a,b),
+    escape(*escape_arg),
+    canDoTurboBM(false),
+    pattern(0),
+    pattern_len(0),
+    bmGs(0),
+    bmBc(0)
   {}
+
   longlong val_int();
   enum Functype functype() const { return LIKE_FUNC; }
   optimize_type select_optimize() const;
   cond_result eq_cmp_result() const { return COND_TRUE; }
   const char *func_name() const { return "like"; }
   void fix_length_and_dec();
+  bool fix_fields(THD *thd,struct st_table_list *tlist);
 };
 
 #ifdef USE_REGEX

sql/unireg.h

@@ -122,6 +122,13 @@ bfill((A)->null_flags,(A)->null_bytes,255);\
 #define TE_INFO_LENGTH 3
 #define MTYP_NOEMPTY_BIT 128
 
+/*
+ *  Minimum length pattern before Turbo Boyer-Moore is used
+ *  for SELECT "text" LIKE "%pattern%", excluding the two
+ *  wildcards in class Item_func_like.
+ */
+#define MIN_TURBOBM_PATTERN_LEN 3
+
 	/* Include prototypes for unireg */
 
 #include "mysqld_error.h"