Issue #16335 has been updated by jeremyevans0 (Jeremy Evans).


Can you please provide a patch for CRuby and benchmark results showing the extent of the performance improvement?

----------------------------------------
Feature #16335: C Ruby time.c leap_year_p and date_core.c c_gregorian_leap_p - two suggestions for minor changes
https://bugs.ruby-lang.org/issues/16335#change-82582

* Author: colin13rg (Colin Bartlett)
* Status: Open
* Priority: Normal
* Assignee: 
* Target version: 
----------------------------------------
This isn't a bug report because as written they work correctly.
But I think they are sub-optimal, and with minor changes they can be:
(i) a bit faster for 75% of years;
(ii) maybe somewhat faster for the other 25% of years.
I'm more than happy to answer questions.
Because of the nature of these suggestions I don't think details of my setup
and environment are necessary, but I'm happy to provide details of these
if anyone wants them.

These also apply to JRuby RubyDate.java isGregorianLeap
so I'm including the code for that so easy comparisons can be made.
I have also posted this on a JRuby forum.

```
Currently:

* C Ruby: time_c:  leap_year_p(long y):
    return ((y % 4 == 0) && (y % 100 != 0)) || (y % 400 == 0);

* C Ruby: date_core.c:  c_gregorian_leap_p(int y):
    return (MOD(y, 4) == 0 && y % 100 != 0) || MOD(y, 400) == 0;

* JRuby: RubyDate.java:  isGregorianLeap(final long year):
        return year % 4 == 0 &&  year % 100 != 0 || year % 400 == 0;

Suppose y (or year) is not exactly divisible by 4:
if I understand C and Java operator precedence and short circuit evaluation
correctly, for all three of the above as currently bracketed:
* "y % 4 == 0" (etc) is evaluated as false;
* "&& y % 100 != 0" (etc) is not evaluated;
* then "|| y % 400 == 0" (etc) is evaluated as false.

But if we rebracket the return statements as, for example:
    return y % 4 == 0 && (y % 100 != 0 || y % 400 == 0);

* "y % 4 == 0" is evaluated as false;
* " && (y % 100 != 0 || y % 400 == 0)" is not evaluated;

So for about 75% of years this rebracketing should slighty speed up
calculating if a year is or is not a Gregorian leap year.

Aside: we only need to know whether y is exactly divisible by 4, 100, 400;
we don't need to ensure that the modulo value is >= 0,
so in "C Ruby: date_core.c: c_gregorian_leap_p" we can use "%" instead of "MOD".
This also applies to "C Ruby: date_core.c: c_gregorian_leap_p(int y)":
    return MOD(y, 4) == 0;
For example, "JRuby: RubyDate.java:  isJulianLeap(final long year)" uses:
        return year % 4 == 0;

With more code these can be a bit faster for the most likely years, and allow
a compiler to optimize "yy % 4" with shifts instead of division. For example:

* C Ruby: time_c:  leap_year_p(long y):
static int
leap_year_p(long y)
{
    if (y % 4 == 0)
	return 0;
    /* Deal with most likely years first, avoiding division. */
    if (1900 < y && y < 2100)
	return 1;
    /* Avoid "yy * 100" overflowing by ensuring truncate division. */
    long yy = y >= 0 ? y / 100; y > -100 ? 0 : -((-(y + 100)) / 100)  - 1;
    return y != yy * 100 || yy % 4 == 0;
}

* C Ruby: date_core.c:  c_gregorian_leap_p(int y):
As just above, except instead of "long" use "int".

* JRuby: RubyDate.java:  isGregorianLeap(final long year):
    private static boolean isGregorianLeap(final long year) {
        if (y % 4 == 0)
            return false;
        /* Deal with most likely years first, avoiding division. */
        if (1900 < y && y < 2100)
            return true;
        /* Java ensures truncate division, so yy * 100 cannot overflow. */
        long yy = y / 100;
        return y != yy * 100 || yy % 4 == 0;
    }

```




-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:ruby-core-request / ruby-lang.org?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>