Statistical Selection of Compiler Options

1
Statistical Selectionof Compiler Options

• R.P.J. Pinkers
• P.M.W. Knijnenburg
• M. Haneda
• H.A.G. Wijshoff

2
Motivation

• Modern compilers contain dozens of options.
• Options can positively or negatively interfere.
• Optimal setting of those options depends on
  application as well as target architecture.
• Standard Ox settings produce sub-optimal
  results.
• We propose almost automatic iterative procedure
  to select options for a given application and
  architecture based on statistical analysis

3
Fragment Optimization Space gcc
4
Fragment Optimization Space ijpeg
5
Orthogonal Arrays

• If there are N options or factors, the full
  optimization space contains 2N combinations.
• This space is called a full factorial design.
• A fractional factorial design is a subset of the
  full factorial design.
• An Orthogonal Array (OA) or Taguchi design is a
  well-known approach to fractional factorial
  designs.
• An OA allows us to determine the effect of a
  factor in the presence of other factors using a
  reduced space.

6
Definition of Orthogonal Arrays

• OA is N x k matrix of zeroes and ones.
• The columns are interpreted as options.
• Each row defines a compiler setting.
• An OA has the property that two arbitrary columns
  contain the patterns
• equally often.
• Each option Oi is turned on and off equally
  often.
• If Oi is turned on, then each other option Oj is
  turned on and off equally often.

00 01 10 11
7
Orthogonal Array (Example)
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 1 0 1 0
1 0 1 0 1 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 1 1 0 0 1
1 0 0 1 1 0 0 1 1 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1
1 0 1 1 0 1 0 0 1 0 1 1 0 1 0 0 1 1 1 1 0 0 0 0 1
1 1 1 0 0 1 1 0 1 0 0 1 0 1 1 0 1 0 0 1 0 0 0 0 0
0 0 1 1 1 1 1 1 1 1 1 0 1 0 1 0 1 1 0 1 0 1 0 1 0
0 1 1 0 0 1 1 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 1 0 0
1 1 0 0 1 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 1 0 1 1 0
1 0 1 0 1 0 0 1 0 1 0 1 1 1 1 0 0 1 1 0 0 0 0 1 1
1 1 0 1 0 0 1 1 0 0 1 0 1 1 0
8
Main Effects of Options

• An OA allows us to calculate the main effect of
  options.
• The main effect of an option Oi wrt an OA A is
  defined as
• where s denotes a row in A and T(s) denotes the
  execution time of the program when compiles with
  setting s.

9
Relative Effects

• Execution times are given in cycles or seconds.
• Effects of options for different benchmarks
  cannot be compared easily.
• We define relative effect of an option Oi as

10
Improvement of Options

• An option can improve or degrade performance.
• The main effect is always positive and does not
  distinguish between improvement or degradation.
• We define the improvement of an option Oi wrt an
  OA A as

11
Iterative Algorithm

• Repeat
• Compile application with each row from A as
  compiler setting and execute optimized
  application.
• Compute effect of each option.
• If effect is larger than threshold of 10
• if option has positive improvement, switch option
  on
• else switch option off.
• Construct new OA A by dropping columns
  corresponding to selected options.
• until all options are set

12
Experimental Setup

• SimpleScalar simulator with GCC 2.6.3.
• Contains 19 options divided in 14 factors
• 6 SPEC95 benchmarks.

O1 unroll-loops, strength-reduce, rerun-cse-after-loop
O2 fast-math
O3 schedule-insns, schedule-insns2
O4 inline, inline-functions
O5 cse-follow-jumps, cse-skip-blocks
O6 float-store
O7 Defer-pop
O8 Force-mem
O9 Force-addr
O10 Omit-frame-pointer
O11 Thread-jumps
O12 Expensive-optimizations
O13 Caller-saves
O14 peephole
13
Iteration 1 for gcc
14
Iteration 2 for gcc
15
Iteration 3 for gcc
16
Comparison with Standard Settings
17
Conclusions

• We have proposed a method based on statistical
  analysis of the effect of options using
  Orthogonal Arrays.
• Method uses (almost) no knowledge of compiler.
• Can be implemented as a simple driver on top of
  any compiler.
• Significant improvement can be obtained by
  automatically tuning compiler options over
  standard optimization settings.