Improving Dataflow Analysis with Path Profiles

1
Improving Data-flow Analysis with Path Profiles

• Glenn Ammons James R. Larus
• University of Wisconsin-Madison
• 1998
• Presented by Jessica Friis

2
Problem and Approach

• Some paths in a CFG are not run
• A small number of paths often make up a large
  portion of the runtime
• By duplicating some frequently used paths, we can
  improve analysis on them

3
Outline

• Identify hot paths, frequently run paths in a
  CFG by doing a training run on the program
• Duplicate hot paths to form a new CFG, the Hot
  Paths Graph, or HPG
• The extra paths allow more precise analysis.
  Constant propagation is looked at here
• Reduce the graph to preserve only the valuable
  solutions for the rest of compilation.
• Implementation and Results

4
Example CFG

• Hot paths taken during sample executions
• A, B, C, E, F, H, I, X
• A, B, D, E, F, H
• B, D, E, G, H
• B, D, E, F, H, I, X
• Use Ball-Larus path profiles
• Chose paths so that they cover 97 of execution
  time

5
Finite Automaton for the path profile

• The retrieval tree is used as a simple
  representation of the automaton.
• CFG edges label the transitions
• At 13, 15, or 16, a B takes you back to 0
• Anything else not labeled goes to an error state
• These extra states/edges are left out for
  readability

6
Trace CFG and FA to get HPG

• Uses Holley and Rosen's data flow tracing
  algorithm
• The HPG will have nodes made up of a tuple v,q
  where v is a vertex from the original CFG and q
  is a state in the automaton
• A worklist algorithm starts with r,q? where r
  is the starting node in the CFG and q? is the
  starting node in the automaton. It follows the
  edges in the CFG and automaton to create a new HPG

7
New HPG

• The diagonally striped nodes (A0 and B0)
  represent the beginnings of the forward paths
• The shaded nodes represent the error states in
  the automaton (paths that are not 'hot.')
• Haven't lost any information from orginal CFG

8
New Knowledge with HPG

• At H14, ab6
• At H12, H15, ab5
• At H13, ab4
• At H14H15, i 1
• At I17, n is 1
• These are new constant results

9
Eliminate unneeded vertices

• Heuristic algorithm for identifying the most
  valuable duplicated vertices.
• Others can be reduced/combined to form rHPG for
  further compilation.

10
Implementation

• Implemented as two new paths in SUIF compiler, PP
  and PW
• Compile into intermediate form
• PP pass instrumented intermediate form for path
  profiling
• Running the result gives us our path profile
• Next the intermediate code and the path profile
  is run through PW
• PW generates the HPG, discovers new constants,
  and generates the rHPG
• The output is compiled into an executable

11
Speedup of SPEC95 benchmarks

• The benchmarks with the most new constants found
  sped up, the other slowed
• The increase in program size causes the slowdown

12
Program Costs

• Cost of duplication in size of CFG ? HPG
• Go increased 184
• All others increasaed an average of 32
• Cost of duplication in size of CFG ? rHPG
• Go increased 77
• All others increased less than 10
• Analysis time
• Go took 6 times longer
• All others took an average of 61 longer

13
Contributions

• Shows improvements in the precision of data flow
  analysis through guided duplication
• Describes how to reduce hot path graphs
• Preserves path profiling information through
  transformation from CFG to HPG to rHPG
• Applies to constant propagation to show
  performance increase

14
Issues

• The cost of analysis is significant
• Only should be used just before release
• The increased graph size slows down the running
  time
• Specific reasons are not known, but larger
  program size is a possibility
• They only tested with a single optimization
• Adding in other optimizations may give more
  speedup gains

15
Questions?