Big update. Correct the behavior for true_lb and true_ub computation

when the size of the data is zero. Now they are not updated, which leave
us with the correct memory layout in all situations (so far). Update all
the comments to reflect exactly the supported behavior of the DDT engine.

This commit was SVN r14202.

ompi/datatype/dt_add.c

@@ -167,26 +167,20 @@ int32_t ompi_ddt_add( ompi_datatype_t* pdtBase, const ompi_datatype_t* pdtAdd,
         ub = LMAX( pdtBase->ub, ub );
     }
     /* While the true_lb and true_ub have to be ordered to have the true_lb lower
-     * than the true_ub, the ub and lb does not have to be ordered. They should be
+     * than the true_ub, the ub and lb do not have to be ordered. They should be
      * as the user define them.
      */
     pdtBase->lb = lb;
     pdtBase->ub = ub;
 
-    if( 0 == pdtBase->nbElems ) old_true_ub = disp;
-    else                        old_true_ub = pdtBase->true_ub;
-    pdtBase->true_lb = LMIN( true_lb, pdtBase->true_lb );
-    pdtBase->true_ub = LMAX( true_ub, pdtBase->true_ub );
-
     /* compute the new memory alignement */
     pdtBase->align = IMAX( pdtBase->align, pdtAdd->align );
-    pdtBase->size += count * pdtAdd->size;
 
     /* Now that we have the new ub and the alignment we should update the ub to match
-     * the new alignement. We have to add an epsilon that is the least nonnegative increment
-     * needed to roung the extent to the next multiple of the alignment. This rule
-     * apply only if there is user specified upper bound as stated in the MPI
-     * standard MPI 1.2 page 71.
+     * the new alignement. We have to add an epsilon that is the least nonnegative
+     * increment needed to roung the extent to the next multiple of the alignment.
+     * This rule apply only if there is user specified upper bound as stated in the
+     * MPI standard MPI 1.2 page 71.
      */
     if( !(pdtBase->flags & DT_FLAG_USER_UB) ) {
         epsilon = (pdtBase->ub - pdtBase->lb) % pdtBase->align;
@@ -196,17 +190,29 @@ int32_t ompi_ddt_add( ompi_datatype_t* pdtBase, const ompi_datatype_t* pdtAdd,
     }
 
     /*
-     * the count == 0 is LEGAL only for MPI_UB and MPI_LB. I accept it just as a nice way to set
-     * the soft UB for a data (without using a real UB marker). This approach can be used to
-     * create the subarray and darray datatype. However from the MPI level this function
-     * should never be called directly with a count set to 0.
-     * Adding a data-type with a size zero is legal but does not have to go through all the
-     * stuff below.
+     * the count == 0 is LEGAL only for MPI_UB and MPI_LB. Therefore we support it
+     * here in the upper part of this function. As an extension, the count set to
+     * zero can be used to reset the alignment of the data, but not for changing
+     * the true_lb and true_ub.
      */
     if( (0 == count) || (0 == pdtAdd->size) ) {
         return OMPI_SUCCESS;
     }
 
+    /* Now, once we know everything is fine and there are some bytes in
+     * the data-type we can update the size, true_lb and true_ub.
+     */
+    pdtBase->size += count * pdtAdd->size;
+    if( 0 == pdtBase->nbElems ) old_true_ub = disp;
+    else                        old_true_ub = pdtBase->true_ub;
+    if( 0 != pdtBase->size ) {
+        pdtBase->true_lb = LMIN( true_lb, pdtBase->true_lb );
+        pdtBase->true_ub = LMAX( true_ub, pdtBase->true_ub );
+    } else {
+        pdtBase->true_lb = true_lb;
+        pdtBase->true_ub = true_ub;
+    }
+
     pdtBase->bdt_used |= pdtAdd->bdt_used;
     newLength = pdtBase->desc.used + place_needed;
     if( newLength > pdtBase->desc.length ) {