Re: [cp-patches] GregorianCalendar mess

[Bryce McKinlay]

Jeroen Frijters wrote:

> Hi,
>
> Working on this almost made my head explode and I'm sure it's still
> broken, but I think it's little bit less broken than before.
>
> Two of my Mauve java.util.Date tests turned out to be broken. I had used
> Sun JDK 1.4.1 to calculate the reference dates, but it was buggy (fixed
> in JDK 1.5).
>  
>
> +    int day = (int) (time / (24 * 60 * 60 * 1000L));
> +    millisInDay = (int) (time % (24 * 60 * 60 * 1000L));
> +    if (millisInDay < 0)
> +      {
> +    millisInDay += (24 * 60 * 60 * 1000);
> +    day--;
> +      }
> +
> +    int[] f = new int[FIELD_COUNT];
> +    calculateDay(f, day, time - rawOffset >= gregorianCutover);
> +    year = f[YEAR];
> +    int month = f[MONTH];
> +    day = f[DAY_OF_MONTH];
> +    int weekday = f[DAY_OF_WEEK];
>      int dstOffset = isSet[DST_OFFSET]
>        ? fields[DST_OFFSET] : (zone.getOffset((year < 0) ? BC : AD,
>                           (year < 0) ? 1 - year : year,


We need to find a way to avoid the "new int[]" in calculateTime(), as 
this is a pretty substantial performance regression. Perhaps a separate 
pass to normalize the fields is needed, which can be called by both 
calculateDay() and computeTime()? I think this could also simplify much 
of the other code, since it can assume the fields are already properly 
rolled over after each set() call, etc? Thanks for working on this!

Regards

Bryce