PEPM `97 Talk Outline

Susan suggested that the talk serve to interest the audience in reading the gory details in the paper. To that end, we should probably talk about what we do and why, rather than how.

This is an outline that follows the high-level structure of the paper. Although it doesn't give any details about how we do anything, it may already be too detailed in saying what we do.

1. Advantages of Run-Time Specialization
   • More automatic
   • More widely applicable
   • Code explosion
   • Termination
2. Applications
   • Same applications for which PE has been successful
   • Operating Systems
   • Interpreters
   • Simulators
   • Graphics applications
3. Our Approach
   • Declarative
   • Annotations map to run-time specialization primitives
   • Dynamic compilation performed automatically
4. Advantages of Our DC System
   • General-purpose, imperative programming language (C)
   • Declarative, but has flexibility
   • Control
5. Annotations
   • Easy to use by human programmers
   • Target for fully automatic system
   • make_static signals dynamic-to-static promotion and sets policies
   • Specialization on static variables defines dynamic region
   • Function annotations extend dynamic region interprocedurally
6. Introduce Interpreter Example
7. Functionality Needed
   • Specialization for (potentially) changing variables
   • Multi-way loop unrolling
   • Conditional specialization
   • Caching and reuse of dynamically compiled code
   • Mechanism(s) to manage cost of specialization
8. Features
   • Support for both polyvariant specialization and polyvariant division
   • Intra- and interprocedural specialization, with caller and callee separately compilable
   • Program-point-specific specialization over arbitrary regions of code
   • Automatic caching, reuse, and reclamation of dynamically generated code
   • Automatic interleaving of specialization and dynamic execution
   • Run-time optimizations, including constant propagation, constant folding, strenth reduction, conditional branch folding and dead-code elimination, loop unrolling and merge splitting, and procedure-call specialization
9. System Architecture (Picture)
10. Run-Time Specializer
    • Basis is standard polyvariant specializer
    • Consults cache policies of static variables
    • Initially invoked from ordinary, static code
    • Can emit callback and defer specialization
11. Analysis
    • Intraprocedural alias analysis
    • Live and used variables analyses
    • Pointwise polyvariant BTA over CFG
    • Identify lazy branch successors
    • Output
      • Static variables with cache policies
      • Specialization points
      • Deferral points
12. Status
13. Conclusions and Future Work
    • Interprocedural BTA and specializer
    • Run-time inlining
    • Gain experience on real applications
    • Beyond run-time specialization

Last updated May 8, 1997.
Brian Grant (grant@cs.washington.edu)