------- Additional Comments From law at redhat dot com 2004-12-28 08:22 ------- A few more notes.
I would suggest starting the testcase in comment #8. While the simpler testcase shows the underlying issue, it's simplified so far that the "problem" could be fixed in various ways which totally avoid the copy, but which do not help the larger testcases. If we go back to the testcase in comment #8 and analyze the code closely we will see that a copy is inevitable. If we copy propagate into the PHI, then the copy gets inserted into its own block. If we do not copy propagate into the PHI, then the copy sits just before the loop exit test. Having the copy before the loop exit test rather than splitting the loop backedge is better for a number of reasons. So the issue at hand is not an extraneous copy, but instead the *placement* of that copy. We queue the copy to be inserted on the backedge of the loop, which is clearly a correct insertion point. However, if we use life analysis we can determine that we can place the copy before the loop exit test. The out-of-ssa code computes object lifetimes, so, in theory we have the information we need. What we'd need to do is enhance the copy placement code to be a little smarter when we want to insert a copy on a critical edge. Given a copy dest = src that we want to place on a critical edge E, we'd want to check if dest is live on any of the other edges in e->src->succs -- if dest is not live on any of those edges, then we can place the copy at the end of e->src. Or something along those lines. I have no idea how hard that would be to implement within the current out-of-ssa framework we have in place. It's worth nothing that copy placement on loop backedges for out-of-ssa has been an area that has been problematical before. I ran into a number of regressions in the RTL SSA code with poor placements (mostly in ways that confused the old old rtl loop optimizers). So it's not a _huge_ surprise that we're seeing some issues in the same areas with the tree out-of-ssa code. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19038
