Analysis of various grafting techniques for VLIW processors

Affiliated to

Location

Timing

Description

We found that an additional dedicated grafting phase in GCC improves the performance of code for TriMedia processor significantly.
Grafting is the code-duplication of basic blocks with multiple incoming control edges to eliminate jumps. It is also known as tail-duplication.
This is a powerful optimization step for VLIW processors where jump-overhead is very high. The optimization is usually a trade-off between code-size and performance, although on the TriMedia in some cases it is a win-win.
We would like to investigate a few strategies/heuristics for TriMedia
grafting and report on those. The grafting framework implementation is in place. The project consists of creating and implementing grafting strategies
and test them on our TriMedia bench-suites. The desired end-result is an implementation in GCC of the heuristics and a technical report on insights
gained.