John

     Can you make the following changes to my workplan...

[...]

-- Modify prototype translator to exploit this data-base, unpeeling the
layers of the run-time operations to generate efficient low-level
code.  (4MM).
        ^^^

-- Integration of these optimisations with those investigated in thread (2)
   ^^^^^^^^^^^
(2MM).
 ^^^ 
[...]

...and add something like this...


Relationship of tasks 2 and 3
-----------------------------

     The prototype implementation will employ a rather naive basic
translation scheme, making calls to the highest layer of the run-time
libary, which directly implements HPF-style distributed data.  Task 2
is concerned with designing improved translation schemes.  Recognising
situations in the source code where better use can be made of more
primitive operations, lower layers of the run-time library will be
targetted directly.  These might be block data moves, for example.  We
envisage that the level of operations targetted here should still be
largely platform independent.

     Task 3 incorporates a class of less ambitious local optimisations,
which are, however, likely to be important in practise.  They can be
applied to the code produced by the basic translation scheme of the
prototype, or the more advanced translator produced by task 2.  These
relatively simple optimisations are captured in a separate database,
freeing the basic translator to concentrate on more global strategies.

     Both classes of optimisation are likely to be essential in
production HPF compilers.