Abstract
Traditionally, function invocation is represented as the (static) call graph or the (dynamic) execution tree in compilers. We define the new call-site tree, in which two different executions of a call-site (say α that is located within a function f) are represented by the same node if and only if the calling chains from main to f in the two different executions of α are identical. Function inlining is a very profitable optimisation that replaces a call-site with the body of the called function. Intuitively, it is preferable to inline the call-sites that are executed most often. Call-sites are suitable for function inlining because they allow to adjust the execution counts of (new and existing) call-sites without re-profiling after a call-site is inlined. We also propose analysis algorithms of the call-site tree and implement an inliner in LLVM. The experimental results on SPEC INT 2006 are reported.
Original language | English |
---|---|
Pages (from-to) | 1-12 |
Number of pages | 12 |
Journal | International Journal of Embedded Systems |
Volume | 17 |
Issue number | 5 |
DOIs | |
State | Published - 2024 |
Keywords
- call graph
- call-site trees
- compiler
- execution tree
- inlining
- LLVM
- optimisation
- programming language
- transformation