Improving Region Selection in Dynamic Optimization Systems

1
Improving Region Selection in Dynamic
Optimization Systems

• Author David Hiniker, Kim Hazelwood, Michael D.
  Smith
• Presenter Oh Hyeong-Seok
• Date 2006.11.22

2
Contents

• Introduction
• NET Trace and Shortcomings
• Last-Executed Iterating (LEI) Algorithm
• Trace Combination Algorithm
• Conclusion

3
Introduction

• Dynamic Optimization System

• Region Selection
• Whole Method
• Trace
• Orthogonal problem trace separation excessive
  code duplication

4
NET trace

• Select begin of trace
• Backward branch
• Exit from an existing trace
• Start select trace when execution count threshold
  is reached
• Trace continue
• Backward branch
• Branch taken to another trace
• Size limit

5
Three Shortcoming - Loops

• Cant span interprocedural cycle
• Cant extend trace to interprocedural backward
  path

NET
Ideal
A
A
E
A
B
F
B
B
C
D
to A
D
to C
E
D
E
call
to E
F
to C
F
return
E is lower address than D
6
Three Shortcomings Nested Loops

• Duplicate the beginning of an inner loop in a
  trace for an outer loop

NET
Ideal
A
C
A
B
B
C
to A
to C
B
to C
A
B
to C
to B
C
7
Three Shortcomings Unbiased branch

• Trace can contain only one of the branch target

NET
A
A
B
50
50
C
D
C
B
D
F
F
to E
D
to B
to E
95
5
F
E
8
Last-Executed Iteration

• Select cyclic trace
• Based on history buffer
• Form from backward branch or exit from code cache
• Trace end when cycle is completed or next
  instruction begin existing trace
• Select frequently executed trace
• Trace after the branch executes predefined number
  of time Tcyc

9
LEI Algorithm
10
Forming Trace
11
Trace Combination

• Simple extension of trace selection
• Lower trace selection threshold
• Observe the trace generated for the next several
  execution, and combine
• Trace combination select
• 1st step include only those blocks that occur in
  frequently executed traces
• 2nd step include executed paths that rejoin
  those blocks

12
Algorithm for Trace Combination
13
Storing a Trace

• Store each observed trace independently
• Compact representation

14
Constructing the CFG

• Construct a control-flow graph for the observed
  traces
• Incrementally adding each observed trace
• Mark each block which appear in traces Tmin times

15
Marking Paths that Rejoin

• Mark if any successor of a unmarked basic block
  is marked

16
Experiment

• Create a Framework for simulating
• Relies on the Pin
• Unbounded code cache
• SPECint2000 benchmark
• Definition
• Hit Rate percentage of executed program
  instruction from code cache
• Code Expansion number of program instructions
  that are copied into the code cache
• Region Transition jump between regions in the
  code caches

17
LEI Trace Result

• Code Expansion and Region Transitions
• 99 of execution occur natively from the code
  cache (difference between NET and LEI less than
  0.2 in most benchmark)

18
Region Combination Result

• Region Transitions

19
Conclusion

• Identified NET trace-selection algorithm
• Develop two new region-selection algorithm
• Better select regions of frequently executing
  code and improve locality of execution