Bug#35296563 performance regression since 8.0.22+ of simple query doing table scan with LIKE '%....%';

The benchmarks from
    Bug #32831830 REGEXP TAKES MORE THAN 10X TIME WHEN COMPARED TO LIKE
have been backported to mysql-8.0.21-release

Timings from
./bin/item_func_regexp-t --gtest_filter="Microbenchmarks.BM_LikeNoMatch"
on my desktop, Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz

mysql-8.0.21-release      22 ns/iter
trunk, original           32 ns/iter
trunk, with current patch 22 ns/iter

The patch optimizes eval_string_arg() for the common case: do string
operations with arguments belonging to the same character set.

Conversion verification and error checking is now delegatedd to
eval_string_arg_noinline()

Change-Id: Iad1fe8675fa2aee1a6a1840629d988ee0dfd37dd

Author: Tor Didriksen

include/my_sys.h

@@ -894,7 +894,6 @@ extern bool resolve_collation(const char *cl_name,
                               const CHARSET_INFO *default_cl,
                               const CHARSET_INFO **cl);
 extern char *get_charsets_dir(char *buf);
-extern bool my_charset_same(const CHARSET_INFO *cs1, const CHARSET_INFO *cs2);
 extern bool init_compiled_charsets(myf flags);
 extern size_t escape_string_for_mysql(const CHARSET_INFO *charset_info,
                                       char *to, size_t to_length,

include/mysql/strings/m_ctype.h

@@ -30,9 +30,11 @@
 
 #include <sys/types.h>
 
+#include <cassert>
 #include <cstddef>
 #include <cstdint>
 #include <cstdlib>
+#include <cstring>
 #include <deque>
 
 #include "mysql/attribute.h"
@@ -513,6 +515,12 @@ inline bool my_charset_is_ascii_based(const CHARSET_INFO *cs) {
   return (cs->state & MY_CS_NONASCII) == 0;
 }
 
+inline bool my_charset_same(const CHARSET_INFO *cs1, const CHARSET_INFO *cs2) {
+  assert(0 != strcmp(cs1->csname, "utf8"));
+  assert(0 != strcmp(cs2->csname, "utf8"));
+  return ((cs1 == cs2) || !strcmp(cs1->csname, cs2->csname));
+}
+
 MYSQL_STRINGS_EXPORT unsigned my_charset_repertoire(const CHARSET_INFO *cs);
 
 MYSQL_STRINGS_EXPORT unsigned my_strxfrm_flag_normalize(unsigned flags);

include/sql_string.h

@@ -41,6 +41,7 @@
 #include "memory_debugging.h"
 #include "my_alloc.h"
 #include "my_compiler.h"
+#include "my_sys.h"
 
 #include "my_inttypes.h"
 #include "mysql/components/services/bits/psi_bits.h"
@@ -607,6 +608,35 @@ class String {
   char *dup(MEM_ROOT *root) const;
 };
 
+/**
+  Checks that the source string can be just copied to the destination string
+  without conversion.
+
+  @param arg_length     Length of string to copy.
+  @param from_cs        Character set to copy from
+  @param to_cs          Character set to copy to
+  @param *offset	Returns number of unaligned characters.
+
+  @returns true if conversion is required, false otherwise.
+
+  @note
+  to_cs may be nullptr for "no conversion" if the system variable
+  character_set_results is NULL.
+*/
+
+inline bool String::needs_conversion(size_t arg_length,
+                                     const CHARSET_INFO *from_cs,
+                                     const CHARSET_INFO *to_cs,
+                                     size_t *offset) {
+  *offset = 0;
+  if (to_cs == nullptr || (to_cs == &my_charset_bin) || from_cs == to_cs ||
+      my_charset_same(from_cs, to_cs) ||
+      ((from_cs == &my_charset_bin) &&
+       (0 == (*offset = (arg_length % to_cs->mbminlen)))))
+    return false;
+  return true;
+}
+
 static inline void swap(String &a, String &b) noexcept { a.swap(b); }
 
 static inline std::string to_string(const String &str) {

mysys/charset.cc

@@ -75,12 +75,6 @@
 
 extern CHARSET_INFO my_charset_cp932_japanese_ci;
 
-bool my_charset_same(const CHARSET_INFO *cs1, const CHARSET_INFO *cs2) {
-  assert(0 != strcmp(cs1->csname, "utf8"));
-  assert(0 != strcmp(cs2->csname, "utf8"));
-  return ((cs1 == cs2) || !strcmp(cs1->csname, cs2->csname));
-}
-
 namespace {
 
 class Mysys_charset_loader : public MY_CHARSET_LOADER {

sql-common/sql_string.cc

@@ -249,39 +249,6 @@ bool String::copy(const char *str, size_t arg_length, const CHARSET_INFO *cs) {
   return false;
 }
 
-/*
-  Checks that the source string can be just copied to the destination string
-  without conversion.
-
-  SYNPOSIS
-
-  needs_conversion()
-  arg_length		Length of string to copy.
-  from_cs		Character set to copy from
-  to_cs			Character set to copy to
-  uint32 *offset	Returns number of unaligned characters.
-
-  RETURN
-   0  No conversion needed
-   1  Either character set conversion or adding leading  zeros
-      (e.g. for UCS-2) must be done
-
-  NOTE
-  to_cs may be NULL for "no conversion" if the system variable
-  character_set_results is NULL.
-*/
-
-bool String::needs_conversion(size_t arg_length, const CHARSET_INFO *from_cs,
-                              const CHARSET_INFO *to_cs, size_t *offset) {
-  *offset = 0;
-  if (!to_cs || (to_cs == &my_charset_bin) || (to_cs == from_cs) ||
-      my_charset_same(from_cs, to_cs) ||
-      ((from_cs == &my_charset_bin) &&
-       (!(*offset = (arg_length % to_cs->mbminlen)))))
-    return false;
-  return true;
-}
-
 /*
   Checks that the source string can just be copied to the destination string
   without conversion.

sql/item_func.cc

@@ -258,19 +258,17 @@ bool simplify_string_args(THD *thd, const DTCollation &c, Item **args,
   @returns string pointer if success, NULL if error or NULL value
 */
 
-String *eval_string_arg(const CHARSET_INFO *to_cs, Item *arg, String *buffer) {
-  StringBuffer<STRING_BUFFER_USUAL_SIZE> local_string(nullptr, 0, to_cs);
-
+String *eval_string_arg_noinline(const CHARSET_INFO *to_cs, Item *arg,
+                                 String *buffer) {
   size_t offset;
   const bool convert =
       String::needs_conversion(0, arg->collation.collation, to_cs, &offset);
-  String *res = arg->val_str(convert ? &local_string : buffer);
 
-  // Return immediately if argument is a NULL value, or there was an error
-  if (res == nullptr) {
-    return nullptr;
-  }
   if (convert) {
+    StringBuffer<STRING_BUFFER_USUAL_SIZE> local_string(nullptr, 0, to_cs);
+    String *res = arg->val_str(&local_string);
+    // Return immediately if argument is a NULL value, or there was an error
+    if (res == nullptr) return nullptr;
     /*
       String must be converted from source character set. It has been built
       in the "local_string" buffer and will be copied with conversion into the
@@ -285,6 +283,10 @@ String *eval_string_arg(const CHARSET_INFO *to_cs, Item *arg, String *buffer) {
     }
     return buffer;
   }
+  String *res = arg->val_str(buffer);
+  // Return immediately if argument is a NULL value, or there was an error
+  if (res == nullptr) return nullptr;
+
   // If source is a binary string, the string may have to be validated:
   if (to_cs != &my_charset_bin && arg->collation.collation == &my_charset_bin &&
       !res->is_valid_string(to_cs)) {

sql/item_func.h

@@ -88,7 +88,15 @@ void report_conversion_error(const CHARSET_INFO *to_cs, const char *from,
 bool simplify_string_args(THD *thd, const DTCollation &c, Item **items,
                           uint nitems);
 
-String *eval_string_arg(const CHARSET_INFO *to_cs, Item *arg, String *buffer);
+String *eval_string_arg_noinline(const CHARSET_INFO *to_cs, Item *arg,
+                                 String *buffer);
+
+inline String *eval_string_arg(const CHARSET_INFO *to_cs, Item *arg,
+                               String *buffer) {
+  if (my_charset_same(to_cs, arg->collation.collation))
+    return arg->val_str(buffer);
+  return eval_string_arg_noinline(to_cs, arg, buffer);
+}
 
 class Item_func : public Item_result_field {
  protected: