Code Specialization

1
Code Specialization

• Michael E. LocastoSeptember 19, 2002

2
What is Code Specialization?

• General algorithm --gt domain-specific
  implementation
• eliminate (hold invariant) some inputs
• a compiler optimization

3
Related Terms Concepts

• Compiler Optimizations
• Invariants
• partial evaluation
• promotion and accumulation
• memoization

• JIT
• code obfuscation
• transformational programming
• finite differencing
• loop unrolling

4
Why Is It Useful?

• Performance boost 13-25, and sometimes more
  (60)
• Dont need to save the world (just your
  continent)
• Respond to change in environment (JIT)
• useful in embedded env (highly specialized
  hardware)

5
Why Is It Useful ? (con)

• MAC operation in DSP Harvard 56000
• three approaches to providing performance
  guarantees in algorithm design randomize,
  amortize, or optimize.
• manual optimizations (better programmers)
• automatic optimizations (save your fingers)

6
Some Manual Optimizations

• Algebraic and logical expression manipulation
• eliminate recursive method calls ( method call
  overhead in general)
• reduce excessive object creation
• write in ASM

7
The Idea of Invariants

• int foo( int a, int b) if( agtb ) //many
  instructions else if( ab ) //many
  instructions else return a

• int foo( int a, int b ) return a//or
  evenint foo( int a ) return a//or evena

8
Partial Evaluation

• Non-deterministic, may fail to terminate
• lazy evaluation
• Program p Data d new source (p_d)
• Basic idea your data has mostly been evaluated
  so the program does not waste time getting,
  scanning, and verifying large set of input

9
Another Example

• See sheet
• (http//www.cs.columbia.edu/locasto/projects/code
  spec/codespecialization.summary.html )

10
Specific Applications

• Database querying
• hardware simulation
• spreadsheet computations
• Configuration files (seldom-changing parameters)

• Support both modularity and performance
• crypto (many repeated operations with predictable
  data)
• R replacing complex function with multiplication
  by 2,3

11
Security and Code Specialization

• Partial Evaluation is used to create the
  intermediate permissions for java.security.Policy
• in some way make code simpler (proof of security
  may be easier)
• 2.29 Gb/s Rijndael processor (for AES) -
  designers used code specialization to turn the C
  ref imp to hardware

12
Rijndael Code Specialization

• Look up table (Galois field)
• large mem required, embedded device limits
• analysis revealed simple 2 or 3 mul

13
Embedded Devices

• Much work being done here (as well as for the
  JVM/JIT)
• save power, space, heat, cpu cycles
• these devices have highly specialized
  architecture, making it natural to want to
  automatically specialize an algorithm. Also,
  their data is of a predictable and uniform type.

14
Resources

• http//www.diku.dk/research-groups/topps/activitie
  s/PartialEvaluation.html
• http//www.cs.columbia.edu/locasto/projects/codes
  pec/codespecialization.survey.html
• http//www.dina.dk/sestoft/pebook/pebook.html

15
Questions /Comments/

• Specific problems?
• Applications
• how to apply to crypto
• optimizations