Integer programming approaches to simultaneous scheduling and register allocation on the C6X

1
Integer programming approaches to simultaneous
scheduling and register allocation on the C6X
Andrew Gilpin, Rebecca Hutchinson April 12, 2005
2
Project Outline

• Formulate scheduling as an ILP.
• Sanity check formulation using CPLEX to solve.
• Translate Pegasus format to ILP translate CPLEX
  solution back to Pegasus.
• Formulate and incorporate register allocation
  into the ILP and translation.

3
ILP Formulation(examples of basic constraints)
4
ILP Formulation(examples of dealing with the
crosspath)
5
Modeling implications in ILP
6
Implementation Progress

• Formulation complete, implementation halfway done
• Taking advantage of new technology
• Latest version of CPLEX allows warm-starts
• Allows us to use list-scheduling solution to
  seed the ILP, hopefully yielding faster solve
  times
• Hopefully we can extend this model to include
  register allocation as well

7
Proposed evaluation

• Compare schedule lengths of our algorithm to
  schedule sizes from Leupers algorithm
• Test data functions in test.c
• Also compare compile times
• We expect to be slower, but how much?
• Compare running times in simulator

8
References

• Wilken, Liu, and Heffernan. Optimal Instruction
  Scheulding Using Integer Programming. PLDI
  2000.
• Chang, Chen, and King. Using Integer Linear
  Programming for Instruction Scheduling and
  Register Allocation in Multi-issue Processors.
  Computers and Mathematics with Applications, Vol
  34(9)1-14.
• Goodwin and Wilken. Optimal and Near-optimal
  Global Register Allocation Using 0-1 Integer
  Programming. Software-Practice and Experience,
  Vol 26(8), 929-965, August 1996.
• Texas Instruments TMS320C62xx CPU and
  Instruction Set Reference Guide,
  www.ti.com/sc/c6x, 1998