[gecode-users] kernel issues

[Filip Konvička]

Hi,

so Lubos started working on the FloatVar implementation and he's 
discovered something that looks like a bug in Gecode. Initially, we just 
copied "integer.vis" to "float.vis" and renamed the identifiers so we 
got "_FLOAT_" instead of "_INT_" everywhere. (Later on, we'll probably 
at least remove the "DOM" modification event and propagation condition.)

So in short:

[General]
Name:		Float
Ifdef:		GECODE_HAS_FLOAT_VARS
Namespace:	Gecode::Float

and the rest goes exactly the same as with IntVar.

Then we tried posting a propagator on a FloatVar, but unexpectedly 
nothing happened. Lubos tracked this down to the following code fragment 
  in kernel/var-type.icc:

case ME_FLOAT_BND:
      {
        static const Gecode::ModEvent me_c = (
          ((ME_FLOAT_NONE ^ ME_FLOAT_BND ) <<  0) |
          ((ME_FLOAT_VAL  ^ ME_FLOAT_VAL ) <<  4) |
          ((ME_FLOAT_BND  ^ ME_FLOAT_BND ) <<  8) |
          ((ME_FLOAT_DOM  ^ ME_FLOAT_BND ) << 12)
        );
        Gecode::ModEvent me_o = (med & med_mask) >> med_fst;
        Gecode::ModEvent me_n = (me_c >> (me_o << 2)) & med_mask;
        if (me_n == 0)
          return false;
        med ^= me_n << med_fst;
        break;
      }

This is similar to what is generated for IntVars (ME_INT_BND), but it 
seems that the condition (me_n == 0) always evaluates to true for 
FloatVars, and so nothing ever happens.

Lubos fixed this by changing the appropriate line to

  Gecode::ModEvent me_n = (me_c >> (me_o << 2)) & (med_mask >> med_fst);

This seems logical to me. Do you know what script to fix in order to get 
it generated this way?

Thanks,
Filip