摘要
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.
原文 | English |
---|---|
頁(從 - 到) | 1-12 |
頁數 | 12 |
期刊 | International Journal of Embedded Systems |
卷 | 17 |
發行號 | 5 |
DOIs | |
出版狀態 | Published - 2024 |