I'll attach a testcase on which we're pretty slow. It's one of the tests at <http://www.24fun.com/downloadcenter/benchjs/benchjs.html>. I profiled it, and got: Total hit count: 55549 Count %Total Function Name 41335 74.4 PRMJ_basetime (that's _in_, not under; total hits under PRMJ_basetime are 45354, of which 3859 are under PRMJ_ToExtendedTime). The problem is that the "compute the year" loop takes forever. Patch coming up that changes the profile on the same testcase to: Total hit count: 14032, of which total hits under PRMJ_basetime are 3862, almost all of them under PRMJ_ToExtendedTime.

Shaver, I did consider your suggestion that we compare to a known date like the beginning of 2000 or 2001 or something. Doing that made computing isleap kinda complicated, though, so I stuck with this. Assuming that LL_UDIVMOD is at all efficient (which it should be on any platform with HAS_LONG_LONG), it should be fine.

I ran the date tests at http://test.bclary.com/tests/mozilla.org/js/menu.html and this patch passes them all.

Comment on attachment 245402 [details] [diff] [review] Proposed change Looks right to me, though I admit I had to furrow my brow a bit to follow the math. I blame JSLL_*! r=shaver

Fixed.