Lecture 4: Program Tuning and Optimization

Background Material

• Reading assembly code (Chapter 3 of Bryant and O'Hallaron)

Reading

• Bryant and O'Hallaron (Chpt. 5)
• Bacon, Graham, and Sharp, Compiler Transformations for High-Performance Computing, ACM Computing Surveys, Vol. 26, No. 4, December 1994.

Topics

• Optimization blockers
  • Pointers and aliasing (see twiddle.c)
  • Functions and side-effects (see f.c)
• Machine Independent Optimizations
  • Compiler optimizations: code movement, strength reduction, common subexpression elimination, constant folding, register allocation.
  • Abstract vector class example.
    • Moving function calls outside of loops.
    • Removing access functions allows further optimization, but exposes implementation.
    • Accumulate into local variable to remove loads and allow accumulation into register.
• Machine Dependent Optimizations
  • Machine model with support for instruction level parallelism (separate Instruction control and Execution units, data and instruction cache, multiple functional units, register renaming and the retirement unit, out of order execution, branch prediction and speculative computation)
  • Latency and issue time for functional units (pipelining).
  • Translating instructions into operations. Simple renaming convention and pictorial diagram to show instruction overlap and dependencies.
  • Instruction scheduling
  • Loop unrolling.
  • Loop splitting and introducing parallelism.
  • Limits to parallelism - register spilling.
  • Branch prediction and misprediction penalties.

Tasks