Re: Theory on loop transformations

assmann@karlsruhe.gmd.de (Uwe Assmann) writes:
|> I remember that ... Whenever a transformation is performed this amounts to
|> a matrix operation over the loop.


I find the matrix approach easy to use, and have successfully used it for
improving data locality on parallel machines with memory hierarchies such
as the BBN Butterfly and KSR1. All you need to do is to construct one
matrix that represents the transformation.


My question to folks who have implemented loop transformations:


if you have done it in the traditional way, i.e. a sequence of loop
transformations as opposed to the matrix approach, what is your experience
about coming up with the right sequence of transformations to apply?


We can start with a small set of transformations such as loop interchange,
loop skewing, loop distribution/jamming etc.


Thanks.
--
Wei Li
Department of Computer Science
Cornell University
Ithaca, NY 14853
--

Return to the comp.compilers page.
Search the comp.compilers archives again.