Course Description:

Compilation is ordinarily a one-shot deal: you compile your program, and then execute it. What then, is adaptive compilation, and why is it interesting?

Compilation is adaptive if the normal compilation strategy (compile, then execute) is changed, based on information obtained outside of compile-time. Much more information about the program may be available after compile-time, and so using this information may lead to much greater performance.

Adaptive compilation can be:

profile-driven:
the compilation strategy is to compile, execute while collecting profile information, and then to optimize during the recompile, based on the collected profile.

template-based:
the compiler generates optimized code with hooks to be filled in later (such as the values of yet-to-be determined runtime constants).
dynamic-feedback-based:
the compiler generates multiple variants optimized for different conditions, and then monitors and changes at run-time the selection of the variant for best performance.
dynamic:
run-time optimization and code generation. This can be a blend of interpretation and compilation.

We will read papers that cover all these aspects of adaptive compilation, particularly as they relate to performance issues. Another theme we will examine in the course is the need for adaptation of the compiler structure itself. Adaptive compilation may require anything from simple, small changes, as in profile-based compilation, to total reorganization which makes the compiler just one component in a suite of highly interactive tools.