Parallel Architectures and Compilation Techniques PACT 2001

Foreword This year's International Conference on Parallel Architectures and Compilation Techniques (PACT 2001) includes for the first time a Work-in-Progress session. We received many excellent short abstracts – some of which were wild ideas. From these we chose seven abstracts for presentation, and we have included them in this booklet. We selected them for their novelty and the brevity with which they were expressed. Thanks to Kevin Skadron, TCCA Newsletter Editor, we are also pleased to announce that all of the accepted and presented abstracts will appear in the upcoming TCCA newsletter. This booklet therefore is only an intermediate presentation of the work for the conference attendees, and it is intended to form a better basis for a discussion of the presented work during the WiP session. We hope that the session itself will be lively and fun for both the presenters and the attendees and will serve as a forum for intensive, perhaps even controversial discussions. Keeping with the spirit of this session, there will be some interesting prizes for the best abstracts selected by a panel of judges. conference steering committee for giving us the opportunity to institute this new tradition. We would also like to direct special thanks to Eduard Ayguade, Josep-Lluis Larriba, and Marta Jiminez for their excellent support regarding the web presentation, the local organisation, and the printing of these booklets. Last, but certainly not least, we also thank all those who took the time to submit abstracts to this special session. We hope that this WiP session will be a success and that it will become a long and productive tradition for future PACT conferences. Introduction This abstract describes work to minimize re-execution of control independent instructions. This technique differs from prior work in its emphasis on compiler scheduling in order to minimize changes to the hardware of an out-of-order processor. Work so far has focused on single-sided branch hammocks. A branch hammock [1] is an instruction sequence corresponding to an 'if' language construct. It is called double-sided when it corresponds to an 'if-then-else' construct and single-sided when it corresponds to an 'if-then' construct alone (i.e., without the 'else' part). When the 'then' and the 'else' contexts contain only one basic block each, the branch hammock is called simple. In typical speculative processors, when a conditional branch corresponding to such a branch hammock is mispredicted, all the instructions fetched after the branch are squashed. …