Access pattern optimizations

sam2y@koa.cs.Virginia.EDU (Steven A. Moyer)
Fri, 26 Jul 91 15:21:35 GMT

          From comp.compilers

Related articles
Access pattern optimizations sam2y@koa.cs.Virginia.EDU (1991-07-26)
Re: Access pattern optimizations moss@cs.umass.edu (1991-07-27)
| List of all articles for this month |

Newsgroups: comp.compilers
From: sam2y@koa.cs.Virginia.EDU (Steven A. Moyer)
Keywords: optimize
Organization: Dept. of Computer Science, University of Virginia
Date: Fri, 26 Jul 91 15:21:35 GMT

Does anyone know of any work on optimizations/transformations of
access patterns for scalar processors? In particular, given a
loop:
                        do 1 k = 1, n
                1 x(k) = y(k+1) + y(k)


a straight-forward translation would generate the accesses:


load y(2), load y(1), store x(1)
load y(3), load y(2), store x(2)
etc.


with 2 loads and 1 store per iteration. It would be preferable to
generate the access pattern:


load y(1)
load y(2), store x(1)
load y(3), store x(2)
etc.


which re-uses a previously loaded value so that only 1 load and 1 store
is required per iteration. Furthermore, the later access pattern loads
y values and stores x values as vectors with stride 1; given an
interleaved memory system, if the loop which generates this access pattern
were unrolled and accesses to each vector grouped then memory bandwidth
would be improved considerably.


It is easy to imagine more complex loops which would also benefit
from this type of transformation.


Maybe this could be handled by some form of global register coloring
scheme, I'm not sure.


If you know of any published work in this area I would greatly
appreciate references. Please send responses to: sam2y@virginia.edu


Thanks.
  --Steve


Steve Moyer
Dept. of Computer Science
University of Virginia


sam2y@virginia.edu
[This is pretty similar to software pipelining, a well-known optimization
used in unrolled loops. -John]


--


Post a followup to this message

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