On 06/23/2017 05:39 AM, Richard Biener wrote: > > SLP induction vectorization runs into the issue that it remembers > pointers to PHI nodes in the SLP tree during analysis. But those > may get invalidated by loop copying (for prologue/epilogue peeling > or versioning) as the low-level CFG helper copy_bbs works in the > way of copying individual BBs plus their outgoing edges but with > old destinations and at the end re-directing the edges to the > desired location. In SSA this triggers the whole machinery of > making room for new PHI nodes -- that is undesirable because it > causes re-allocation of PHI nodes in the set of source blocks. > > After much pondering I arrived at the following (least ugly) solution > to this "problem" (well, I define it as a problem, it's at least > an inefficiency and a workaround in the vectorizer would be way > uglier). Namely simply do not trigger the SSA machinery for > blocks with BB_DUPLICATED (I skimmed all other users and they seem > fine). > > In the process I also implemented some poisoning of the old PHI node > when we reallocate (well, free) PHI nodes. But that triggers some > other issues, one fixed by the tree-ssa-phionlycoprop.c hunk below. > So I'm not submitting it as part of this fix. > > Bootstrapped (with the poisoning sofar, plain patch still running) > on x86_64-unknown-linux-gnu, testing in progress. > > Comments welcome, testing won't finish before I leave for the > weekend. I fully support poisoning the old PHI nodes -- I tracked down a similar problem just a few months back that probably would have been obvious if we had poisoned the old nodes (79621 which is now a missed optimization bug).
I wouldn't be surprised if there's others lurking and given the general trend of using block duplication to enable various optimizations, catching this stuff early would definitely be good. Jeff
