There have been lately a discussion on the GCC mailing list and other forums about the efficiency of SMS (the current implementation of software pipelining in GCC). One of the issues that currently limit SMS is the lack of loop carried memory dependency information. The long term solution for this is to propagate alias information from tree level down to RTL level and then through the RTL passes. After we have the alias info we perform a data reference analysis inside loops in RTL level and get the dependence information. This infrastructure is important since other optimizations can gain from it. However, there is another (short-term) way that we can follow to improve the loop carried memory dependence information in SMS. The main Idea in this approach (which was raised by Sebastian Bob several months ago) is to propagate the loop carried memory dependancy in a form of a maximum dependancy distance of a given loop from the trees (where we have this information available) down to the RTL and propagate it thought the RTL passes. This solution can buy us a good estimation of what alias information would give us (when available) in SMS. However, it requires the loop information to live across the RTL passes (currently we build the loop information at the beginning of the pass that needs it and destroy that information at the end of the pass). A third approach - which is the same as the previous one with only one difference - is to use a NOTE inside one of the basic blocks of the loop (latch for example) and put the information regarding the maximum memory dependence distance in that NOTE; since notes are naturally propagated through the RTL passes there is no need for special handling to propagate this information. I wanted to share these thoughts with others, because I see increasing interest in improving SMS, and see if anyone expect problems in any of these approaches. Mostafa.
