>>>>> "Fred" == Fred Lindberg <[EMAIL PROTECTED]> writes:
Fred> Monty:
Fred> Here is fmt_ulong() from Dan Bernstein's library:
Fred> #define FMT_ULONG 40 /* enough space to hold 2^128 - 1 in decimal, plus
Fred> \0 */
Fred> unsigned int fmt_ulong(s,u) register char *s; register unsigned long u;
Fred> {
Fred> register unsigned int len; register unsigned long q;
Fred> len = 1; q = u;
Fred> while (q > 9) { ++len; q /= 10; }
Fred> if (s) {
Fred> s += len;
Fred> do { *--s = '0' + (u % 10); u /= 10; } while(u); /* handles u == 0
Fred> */
Fred> }
Fred> return len;
Fred> }
Fred> It makes the number as short as possible, but by fixing len and
Fred> skipping, it's easy to convert to a function that fills YYYY-MM-DD
Fred> HH:MM:DD or whatever is needed and circumvent sprintf() entirely. You
Fred> are now using a probably quite general routine if sprintf() to fill the
Fred> multiple %02d fields, each of which has an overhead in the same order
Fred> of magnitude as the entire conversion.
Hi!
I thought about this a little and came up with the following:
-------
String *Field_datetime::val_str(String *val_buffer,String *val_ptr)
{
val_buffer->alloc(field_length);
val_buffer->length(field_length);
ulonglong tmp;
ulong part1,part2;
char *pos;
int part3;
longlongget(tmp,ptr);
/*
Avoid problem with slow longlong aritmetic and sprintf
*/
part1=(ulong) (tmp/LL(1000000));
part2=(ulong) (tmp - (ulonglong) part1*LL(1000000));
pos=(char*) val_buffer->ptr()+19;
*pos--=0;
*pos--='0'+part2%10; part2/=10;
*pos--='0'+part2%10; part3= (int) (part2 / 10);
*pos--=':';
*pos--='0'+part3%10; part3/=10;
*pos--='0'+part3%10; part3/=10;
*pos--=':';
*pos--='0'+part3%10; part3/=10;
*pos--='0'+part3;
*pos--=' ';
*pos--='0'+part1%10; part1/=10;
*pos--='0'+part1%10; part1/=10;
*pos--='-';
*pos--='0'+part1%10; part1/=10;
*pos--='0'+part1%10; part3= (int) (part1/10);
*pos--='-';
*pos--='0'+part3%10; part3/=10;
*pos--='0'+part3%10; part3/=10;
*pos--='0'+part3%10; part3/=10;
*pos='0'+part3;
return val_buffer;
}
--------
This is about 5 times faster than the code in MySQL 3.22.15
(This doesn't of course solve the general itoa() problem with
longlong for which we still need a solution, but makes datetime
handling much better)
Regards,
Monty
-----------------------------------------------------------
Send a mail to [EMAIL PROTECTED] with
unsubscribe mysql [EMAIL PROTECTED]
in the body of the message to unsubscribe from this list.
