Hi all, For a loop optimization at our backend (loop-pipelining) I would like to move all the loop iv increment and loop condition calculation to the head of a loop. This iv-acceleration should be visible at the final phase of the compilation before the ASM code generation.
Currently, I did it right after the RTL loop-optimization by creating a new basic block at the head of the loop, move all biv calculation, invariant instructions and branch to calculate the loop exit to the basic block. Then create an unconditional jump from the end of the loop to that block. The reason I did it here because the information about loop is available here. However, the new basic block ordering is unoptimal in the view of the later optimization passes especially the bb_reorder optimization. This pass clear all the modification I made for the iv-acceleration above. So my question is where in the optimization process should I do the iv-acceleration above so that the form of the CFG will not be modified? Should I do it after the bb_reordering pass? At this point can I still get the loop information such as biv and invariant instruction? Is there any difficulty if I do the iv-acceleration after the register allocation? Thanks very much for your reading. Any help is really appreciated. -- View this message in context: http://old.nabble.com/Moving-loop-induction-variable-calculation-to-the-head-of-loop.-tp26542157p26542157.html Sent from the gcc - Dev mailing list archive at Nabble.com.
