Subsets, masked ops question 2

[bjoo]

One issue that arises with sets and subsets is the annoying difference between ordered and unordered sets. In principle the idea in the 'scalarsite' ordering was that in an ordered subset one can do a straight for loop from 'start' to 'end'. In the scalarvecsite (e.g. parscalarvec) this is not immediately so straightforward since
contiguity may not span inner/outer site boundaries (ie, the subset could be contiguous but end mid-way through the inner sites of an outer block.)

Further there are subsets which are a discontiguous collection of contiguous subsets, e.g. for t-slice sums. WIth time running slowest, typically there is a portion (one timeslice, one checkerboard) that is contiguous.

A potential solution may be to store the subset as a 'list of contiquous bits'. Each member of this list could b
for scalarsite arch: just start/end sites
for scalarvecsite arch: start/end outer blocks, and mask on the last outer block

The lattice traversal could then proceed through the generic loop (written in very pseudocode)

scalarsite:

         for i in list of contiguous bits:
              for site in startsite[i]..endsite[i], site++
                   forEach( ...)  on site 'i'

scalarvecsite:

         for i in list of contiguous bits:
               for block in startblocks[i]..endblocks[i], block++
                     if ( block == endblocks[i] ) 
                          mask=endMask[i]
                     else  
                          mask[:] = true
               forEach(   ) on block with mask.

NB: For 2 D vectorization etc, it may be that each block will need a mask,
as linear 'filling up' may not be guaranteed then.

This should help subsets with lots of contiguous bits, should hopefully stay efficient for
all contiguous subsets, and fragment naturally to truly noncontiguous subsets.

It would also need to generalize ForEach() to work with masks. How to do that?
ForEach with masks can also help with where() statements in scalarvec/parscalarvec?
Thoughts ?