Optimizing General Compiler Optimization

1
Optimizing General Compiler Optimization

• M. Haneda,
• P.M.W. Knijnenburg,
• and H.A.G. Wijshoff

2
Problem Optimizing optimizations

• A compiler usually has many optimization settings
  (e.g. peephole, delayed-branch, etc)
• gcc 3.3 has 54 optimization options
• gcc 4 has over 100 possible settings
• Very little is known about how these options
  affect each other
• Compiler writers typically include switches that
  bundle together many optimization options
• gcc O1, -O2, -O3

3
but can we do better?

• It is possible to perform better than these
  predefined optimization settings, but doing so
  requires extensive knowledge of the code as well
  as the available optimization options
• How do we define one set of options that would
  work well with a large variety of programs?

4
Motivation for this paper

• Since there are too many optimization settings,
  an exhaustive search would cost too much
• gcc 3 250 different combinations!
• We want to define a systematic method to find the
  optimal settings, with a reduced search space
• Ideally, we would like to do this with minimal
  knowledge of what the options actually will do

5
Big Idea

• We want to find the biggest subsets of compiler
  options that positively interact with each other
• Once we obtain these subsets, we will try to
  combine them together, under the condition that
  they do not negatively affect each other
• We will select our ultimate optimal compiler
  setting from the result of these set combinations

6
Full vs. Fractional Factorial Design

• Full Factorial Design explores the entire search
  space, with every possible combination
• Given k options, this will take O(2k) time
• Fractional Factorial Design explores a reduced
  search space, that is representative of the full
  search space
• This can be done using orthogonal arrays

7
Orthogonal Arrays

• An Orthogonal Array is a matrix of 0s and 1s.
• The rows represent the experiments to be
  performed.
• The columns represent the factors that the
  experiment tries to analyze
• Any option is equally likely to be turned on/off.
• Given a particular experiment with a particular
  option turned on, all the other options are still
  equally likely to be turned on/off

8
Orthogonal Arrays
9
Algorithm Step 1

• Finding maximum subsets of positively interacting
  options
• Step 1.1 Find a set of options that give the
  best overall improvement
• For any single optimization setting i, compute
  the average speedup for all the settings in the
  search space in which i is turned on
• Select M of the highest average improvement
  settings

10
Step 1.1 Selecting the initial set
11
Algorithm Step 1(cont.)

• Step 1.2 Iteratively add new options to the
  already obtained sets, to get a maximum set of
  positively reinforcing optimizations
• Ex If using options A and B together produces a
  more optimal setting than just using A, then add
  B
• If using A, B and C together produces a more
  optimal setting than A, B, then add C to A, B

12
Algorithm Step 2

• Take the sets that we already have and try to
  combine them together, assuming that they do not
  negatively influence each other.
• This is done to maximize the number of settings
  turned on for each set
• Example
• If A, B, C and D, E do not counteract each
  other, then we can combine them into A, B, C, D,
  E
• Otherwise, leave them separate

13
Algorithm Step 3

• Take the resulting sets from step 2, and select
  the one with the best overall improvement.
• The result would be the ideal combination of
  optimization settings, according to this
  methodology.

14
Final Result
15
Comparing with existing compiler switches
16
Comparing results

• The compiler setting obtained by this methodology
  outpeforms O1, -O2, and O3 on almost all the
  SPECint95 benchmarks
• -O3 performs better on li (39.2 vs. 38.4)
• The new setting delivers the best performance for
  perl (18.4 vs. 10.5)

17
Conclusion

• The paper introduced a systematic way of
  combining compiler optimization settings
• Used a reduced search space, constructed as an
  orthogonal array
• Can be done with no knowledge of actual options
• Can be done independently of architecture
• Can be applied to a wide variety of applications

18
Future work

• Using the same methodology to find a good
  optimization setting for a particular domain of
  applications
• Applying the methodology to newer versions of the
  gcc compiler, such as gcc 4.0.1