Thanks to all!

I see, I am glad there are good alternative libraries!
For my problem, I preferred to avoid having dependence on one more library for only 10 lines of code, so I wrote some code to do the conversion.  It works only for dates after 1970, and it is somewhat inelegant; it is here: http://wghstrfg.blogspot.com/2010/03/i-hate-daylight-savings-time.html

If you want to know why this bug drove me crazy for a couple of evenings, you can read this blog post.

Many thanks, and I am glad my email helped Dave.

Daylight savings time is a huge headache for anyone working on historically-timestamped data.

All the best,

Luca

On Sat, Mar 6, 2010 at 7:27 PM, Dave Benjamin <dave@ramenlabs.com> wrote:
Luca de Alfaro wrote:
I need to convert a time, expressed in yyyy/mm/dd hh:mm:ss form, into a floating point.
The conversion has to be done in GMT, but the real point is, the conversion must NOT change due to daylight savings time.

Ocaml seems to have only one conversion function, however: Unix.mktime, which takes a time, and makes a float in the local time zone.

No problem, I thought: I will simply add 3600 if the conversion result tells me that dst is active (and then convert for the difference between GMT and winter time).
NO! This does not work!  Look at the two conversions below.  The tmrec differs by one hour.
However, the two floating point numbers returned are identical, and tm_isdst is set to true in both cases!

This means that I have no way of using the standard libraries to convert a time expressed in yyyy/mm/ss hh:mm:ss to a float!

You are absolutely right, and I unfortunately did not notice this subtlety when I wrote the XmlRpcDateTime module that is part of XmlRpc-Light, so this means there is a bug in XmlRpcDateTime.to_unixfloat_utc on systems with time zones that observe daylight savings. I did not notice the problem because I live in Arizona, one of only two states in the US that do not observe daylight savings!

The culprit can be seen here in the C implementation of Unix.mktime:

CAMLprim value unix_mktime(value t)
{
 struct tm tm;
 time_t clock;
 value res;
 value tmval = Val_unit, clkval = Val_unit;

 Begin_roots2(tmval, clkval);
   tm.tm_sec = Int_val(Field(t, 0));
   tm.tm_min = Int_val(Field(t, 1));
   tm.tm_hour = Int_val(Field(t, 2));
   tm.tm_mday = Int_val(Field(t, 3));
   tm.tm_mon = Int_val(Field(t, 4));
   tm.tm_year = Int_val(Field(t, 5));
   tm.tm_wday = Int_val(Field(t, 6));
   tm.tm_yday = Int_val(Field(t, 7));
   tm.tm_isdst = -1; /* tm.tm_isdst = Bool_val(Field(t, 8)); */
   clock = mktime(&tm);
   if (clock == (time_t) -1) unix_error(ERANGE, "mktime", Nothing);
   tmval = alloc_tm(&tm);
   clkval = copy_double((double) clock);
   res = alloc_small(2, 0);
   Field(res, 0) = clkval;
   Field(res, 1) = tmval;
 End_roots ();
 return res;
}

The field tm.tm_isdst is not really a boolean from C's perspective. It can be one of *three* states: positive for DST, zero for non-DST, and negative to query the system timezone database for the value. It looks like at one point Unix.mktime was written to take the value you gave it, but this was commented out and the value was fixed to -1. This is why it uses the time zone's daylight savings correction regardless of what you pass in.

Would it be possible to have a new function in the standard library with the commented-out behavior instead? As it stands now I don't see any reasonable way to get a UTC float from a Unix.tm.

As far as XmlRpc-Light is concerned, I will probably rewrite this function in terms of Netdate, since Ocamlnet is already one of my dependencies. Apologies to anyone who is affected by this bug (hopefully, no one).

Thanks,
Dave