WRaP-IT

Objectives & Compilation Methods

Compilation methods

Classical compilation

Heavily using static cost models, classical compilation build a binary from a source file, using various intermediate form during the compilation process.

• Optimisation choices based on static cost models
• Static cost models not reliable for more and more complex and dynamic architectures

Iterative compilation

Because building detailed static cost models for modern dynamic architectures is no more possible, iterative compilation rely on the following mechanism :

1. Using static cost models, choose which transformations to apply
2. Choose some parameters for the tranformations
3. Build an intermedite binary with those parameters
4. Test the binary performances, and feedback those informations to the compiler
5. Considering those performance informations, update the parameters and go to 3.
6. When the performance is good enough, build the final binary.

• Allows to search only for optimal transformation parameters

Expert-driven iterative compilation

Within expert-driven iterative compilation framework, the feedback information described above are sent to the expert system.

Then the expert system can choose which transformation to apply from a pool of transformations all described the same way. Those choices are made based of some empiric results from "by hand" program optimisation (A decision tree based on the feedback information selects the relevant transformations)

Evolutinary algorithms may also help in the search of the optimal parameters or transformations.

• Allow to search not only for transformation parameters, but also for transformations to apply.
• Requires a unified model for loop transformations.