The ensemble team wants to be able to take advantage of the full range of representation for 64-bit numeric data types.

Paul Morris has already done some exploration on a branch, see note below.

Assigning to Mike for now since he seems to have thought about this the most.

From Paul (see also attached diff files) :

I have been able to successfully modify Europa to use the 'long' type for Time on the openeuropa-new 64-bit machine. It is probably too late to use this for the G3 delivery because I would recommend prolonged use before delivering this to acquire confidence. However, I would like to record here the changes needed, and the issues that needed to be addressed, for future use.

The core Europa files that needed to be modified are the following:

M TemporalNetwork/base/DistanceGraph.hh M TemporalNetwork/TemporalNetworkDefs.hh M Utils/CommonDefs.hh M Utils/base/Utils.cc M ConstraintEngine/component/IntervalIntDomain.hh M ConstraintEngine/component/IntervalIntDomain.cc

The most extensive changes are to the IntervalIntDomain files, which were modified to use an IntInt datatype instead of int. I have attached diff files for those. IntInt is then typedefed to long in CommonDefs.

Note that because the internal representation of values in ALL Domains uses the double type, IntervalIntDomains and hence Time must still be restricted to at most 52 or 53 bits (the significand part of double). I have restricted it to 50 bits in CommonDefs just to be on the safe side. This is still large enough to represent millseconds.

A further issue that arises is that PLUS_INFINITY and MINUS_INFINITY are used as generalized infinities through the Europa core code, including uses as ints, unsigned ints, and floats, so I couldn't just reset those to 50 bit sizes. Instead I separated out PLUS_INFINITE_TIME from PLUS_INFINITY as separate values, and used the former for Time and IntInt uses, and kept the latter for generalized uses.

One caveat is that durations (but not start/end times) are still restricted to [1, PLUS_INFINITY] because there are several places in the NddlResource.cc and NddlToken.cc files, where IntervalIntDomain(1, PLUS_INFINITY) is called for durations, that I didn't want to change. Another caveat is that PLUS_INFINITY needs to be chosen so that

   PLUS_INFINITY = (int)(float)PLUS_INFINITY

because the +inff float value in the NDDL Resource definitions gets translated to PLUS_INFINITY and is then checked to be <= PLUS_INFINITY, which could fail if there is roundoff error in the (int)(float) cast. I was able to make PLUS_INFINITY be INT_MAX/2 + 1, which casts ok.

These considerations allowed the following definitions to work in the CommonDefs.hh file.

typedef long IntInt;

DECLARE_GLOBAL_CONST(IntInt, g_maxInt);

DECLARE_GLOBAL_CONST(IntInt, g_infiniteTime);

DECLARE_GLOBAL_CONST(IntInt, g_noTime);

define PLUS_INFINITE_TIME (1125899906842624) // 250 (double mantissa)

define MINUS_INFINITE_TIME (-1125899906842624)

// NDDL use of float +inf relies on // PLUS_INFINITY = (int)(float)PLUS_INFINITY

define PLUS_INFINITY ( INT_MAX/2 + 1 )

define MINUS_INFINITY ( -INT_MAX/2 - 1 )

(These are the effective changes for INFINITE_TIME, but should be done in a cleaner way in terms of g_maxInt and g_infiniteTime.)

Now the 'long' definitions of Time in previous versions of the TemporalNetwork could be restored and work with a 64-bit long.

I also had to replace PLUS/MINUS_INFINITY in the DynamicEuropa files by PLUS/MINUS_INFINITE_TIME where appropriate (i.e. for Time uses).

Paul

Committed in r5923. I fully expect there to be bugs and revisions required.