Commit c14ecc75 authored by Alexander Barkov's avatar Alexander Barkov

MDEV-23568 Improve performance of my_{time|date|datetime}_to_str()

This patch improves functions my_xxx_to_str() using the idea
introduced in this change in MySQL-8.0:

  commit 8d10f2fff6bbdea7f436b868ebb5fd811defc68a
  Author: Knut Anders Hatlen <knut.hatlen@oracle.com>
  Date:   Thu Oct 10 13:55:07 2019 +0200
    Bug#30472888: IMPROVE THE PERFORMANCE OF INTEGER HANDLING IN THE TEXT PROTOCOL

The new way prints 2 digits at a time and demonstrates a very impressing query time reduce:
10% to 38%, depending on the exact data type and the number of fractional digits:

  SELECT BENCHMARK(10*1000*1000,CONCAT(TIME'10:20:30'));
  SELECT BENCHMARK(10*1000*1000,CONCAT(TIME'10:20:30.123456'));
  SELECT BENCHMARK(10*1000*1000,CONCAT(DATE'2001-01-01'));
  SELECT BENCHMARK(10*1000*1000,CONCAT(TIMESTAMP'2001-01-01 10:20:30'));
  SELECT BENCHMARK(10*1000*1000,CONCAT(TIMESTAMP'2001-01-01 10:20:30.123456'));

See MDEV for details on the benchmark results.
parent 482cf29e
......@@ -1454,23 +1454,148 @@ void set_zero_time(MYSQL_TIME *tm, enum enum_mysql_timestamp_type time_type)
/*
Helper function for datetime formatting.
Format number as string, left-padded with 0.
A formatting routine to print a 2 digit zero padded number.
It prints 2 digits at a time, which gives a performance improvement.
The idea is taken from "class TwoDigitWriter" in MySQL.
The old implementation printed one digit at a time, using the division
and the remainder operators, which appeared to be slow.
It's cheaper to have a cached array of 2-digit numbers
in their string representation.
Benchmark results showed a 10% to 23% time reduce for these queries:
SELECT BENCHMARK(10*1000*1000,CONCAT(TIME'10:20:30'));
SELECT BENCHMARK(10*1000*1000,CONCAT(DATE'2001-01-01'));
SELECT BENCHMARK(10*1000*1000,CONCAT(TIMESTAMP'2001-01-01 10:20:30'));
SELECT BENCHMARK(10*1000*1000,CONCAT(TIME'10:20:30.123456'));
SELECT BENCHMARK(10*1000*1000,CONCAT(TIMESTAMP'2001-01-01 10:20:30.123456'));
(depending on the exact data type and fractional precision).
The array has extra elements for uint8 values 100..255.
This is done for safety. If the caller passes a value
outside of the expected range 0..99, it will be printed as "XX".
*/
static const char two_digit_numbers[512+1]=
{
/* 0..99 */
"00010203040506070809"
"10111213141516171819"
"20212223242526272829"
"30313233343536373839"
"40414243444546474849"
"50515253545556575859"
"60616263646566676869"
"70717273747576777879"
"80818283848586878889"
"90919293949596979899"
/* 100..199 - safety */
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
/* 200..255 - safety */
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXXXXXXXXXX"
"XXXXXXXXXXXX"
};
The reason to use own formatting rather than sprintf() is performance - in a
datetime benchmark it helped to reduced the datetime formatting overhead
from ~30% down to ~4%.
static inline char* fmt_number2(uint8 val, char *out)
{
const char *src= two_digit_numbers + val * 2;
*out++= *src++;
*out++= *src++;
return out;
}
/*
We tried the same trick with an array of 16384 zerofill 4-digit numbers,
with 10000 elements with numbers 0000..9999, and a tail filled with "XXXX".
Benchmark results for a RelWithDebInfo build:
SELECT BENCHMARK(10*1000*1000,CONCAT(TIMESTAMP'2001-01-01 10:20:30.123456'));
- 0.379 sec (current)
- 0.369 sec (array)
SELECT BENCHMARK(10*1000*1000,CONCAT(DATE'2001-01-01'));
- 0.225 sec (current)
- 0.219 sec (array)
It demonstrated an additional 3% performance imrovement one these queries.
However, as the array size is too huge, we afraid that it will flush data
from the CPU memory cache, which under real load may affect negatively.
Let's keep using the fmt_number4() version with division and remainder
for now. This can be revised later. We could try some smaller array,
e.g. for YEARs in the range 1970..2098 (fitting into a 256 element array).
*/
/*
static inline char* fmt_number4(uint16 val, char *out)
{
const char *src= four_digit_numbers + (val & 0x3FFF) * 4;
memcpy(out, src, 4);
return out + 4;
}
*/
static char* fmt_number(uint val, char *out, uint digits)
/*
A formatting routine to print a 4 digit zero padded number.
*/
static inline char* fmt_number4(uint16 val, char *out)
{
uint i;
for(i= 0; i < digits; i++)
{
out[digits-i-1]= '0' + val%10;
val/=10;
out= fmt_number2((uint8) (val / 100), out);
out= fmt_number2((uint8) (val % 100), out);
return out;
}
/*
A formatting routine to print a 6 digit zero padded number.
*/
static inline char* fmt_number6(uint val, char *out)
{
out= fmt_number2((uint8) (val / 10000), out);
val%= 10000;
out= fmt_number2((uint8) (val / 100), out);
out= fmt_number2((uint8) (val % 100), out);
return out;
}
static char* fmt_usec(uint val, char *out, uint digits)
{
switch (digits)
{
case 1:
*out++= '0' + (val % 10);
return out;
case 2:
return fmt_number2((uint8) val, out);
case 3:
*out++= '0' + (val / 100) % 10;
return fmt_number2((uint8) (val % 100), out);
case 4:
return fmt_number4((uint16) val, out);
case 5:
*out++= '0' + (val / 10000) % 10;
return fmt_number4((uint16) (val % 10000), out);
case 6:
return fmt_number6(val, out);
}
return out + digits;
DBUG_ASSERT(0);
return out;
}
......@@ -1480,13 +1605,13 @@ static int my_mmssff_to_str(const MYSQL_TIME *ltime, char *to, uint fsp)
if (fsp == AUTO_SEC_PART_DIGITS)
fsp= ltime->second_part ? TIME_SECOND_PART_DIGITS : 0;
DBUG_ASSERT(fsp <= TIME_SECOND_PART_DIGITS);
pos= fmt_number(ltime->minute, pos, 2);
pos= fmt_number2((uint8) ltime->minute, pos);
*pos++= ':';
pos= fmt_number(ltime->second, pos, 2);
pos= fmt_number2((uint8) ltime->second, pos);
if (fsp)
{
*pos++= '.';
pos= fmt_number((uint)sec_part_shift(ltime->second_part, fsp), pos, fsp);
pos= fmt_usec((uint)sec_part_shift(ltime->second_part, fsp), pos, fsp);
}
return (int) (pos - to);
}
......@@ -1506,7 +1631,7 @@ int my_interval_DDhhmmssff_to_str(const MYSQL_TIME *ltime, char *to, uint fsp)
pos= longlong10_to_str((longlong) hour / 24, pos, 10);
*pos++= ' ';
}
pos= fmt_number(hour % 24, pos, 2);
pos= fmt_number2((uint8) (hour % 24), pos);
*pos++= ':';
pos+= my_mmssff_to_str(ltime, pos, fsp);
*pos= 0;
......@@ -1538,7 +1663,7 @@ int my_time_to_str(const MYSQL_TIME *l_time, char *to, uint digits)
/* Need more than 2 digits for hours in string representation. */
pos= longlong10_to_str((longlong)hour, pos, 10);
else
pos= fmt_number(hour, pos, 2);
pos= fmt_number2((uint8) hour, pos);
*pos++= ':';
pos+= my_mmssff_to_str(l_time, pos, digits);
......@@ -1550,11 +1675,11 @@ int my_time_to_str(const MYSQL_TIME *l_time, char *to, uint digits)
int my_date_to_str(const MYSQL_TIME *l_time, char *to)
{
char *pos=to;
pos= fmt_number(l_time->year, pos, 4);
pos= fmt_number4((uint16) l_time->year, pos);
*pos++='-';
pos= fmt_number(l_time->month, pos, 2);
pos= fmt_number2((uint8) l_time->month, pos);
*pos++='-';
pos= fmt_number(l_time->day, pos, 2);
pos= fmt_number2((uint8) l_time->day, pos);
*pos= 0;
return (int)(pos - to);
}
......@@ -1563,13 +1688,13 @@ int my_date_to_str(const MYSQL_TIME *l_time, char *to)
int my_datetime_to_str(const MYSQL_TIME *l_time, char *to, uint digits)
{
char *pos= to;
pos= fmt_number(l_time->year, pos, 4);
pos= fmt_number4((uint16) l_time->year, pos);
*pos++='-';
pos= fmt_number(l_time->month, pos, 2);
pos= fmt_number2((uint8) l_time->month, pos);
*pos++='-';
pos= fmt_number(l_time->day, pos, 2);
pos= fmt_number2((uint8) l_time->day, pos);
*pos++=' ';
pos= fmt_number(l_time->hour, pos, 2);
pos= fmt_number2((uint8) l_time->hour, pos);
*pos++= ':';
pos+= my_mmssff_to_str(l_time, pos, digits);
*pos= 0;
......@@ -1625,7 +1750,7 @@ int my_timeval_to_str(const struct timeval *tm, char *to, uint dec)
if (dec)
{
*pos++= '.';
pos= fmt_number((uint) sec_part_shift(tm->tv_usec, dec), pos, dec);
pos= fmt_usec((uint) sec_part_shift(tm->tv_usec, dec), pos, dec);
}
*pos= '\0';
return (int) (pos - to);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment