On Thu, Oct 18, 2012 at 4:48 PM, David Blaikie <dblai...@gmail.com> wrote: > On Thu, Oct 18, 2012 at 1:45 PM, Douglas Gregor <dgre...@apple.com> wrote: >> >> On Oct 18, 2012, at 9:57 AM, David Blaikie <dblai...@gmail.com> wrote: >> >>> Author: dblaikie >>> Date: Thu Oct 18 11:57:32 2012 >>> New Revision: 166188 >>> >>> URL: http://llvm.org/viewvc/llvm-project?rev=166188&view=rev >>> Log: >>> PR14021: Copy lookup results to ensure safe iteration. >>> >>> Within the body of the loop the underlying map may be modified via >>> >>> Sema::AddOverloadCandidate >>> -> Sema::CompareReferenceRelationship >>> -> Sema::RequireCompleteType >>> >>> to avoid the use of invalid iterators the sequence is copied first. >> >> Did you audit other uses of LookupConstructors to ensure that this is the >> only ticking time bomb in this area? > > [+Robert Muth] > > I've not performed any such audit, no. With the hack debug check > inserted we could just run the whole test suite to see if anything > pops up.
I did not perform any such audit either. I feel those are only of limited use and the time would be better spend addressing the root of the issue. If there is interest I could try add some diagnostic code to the DenseMap which would only be enabled in Debug mode. We kicked around some ideas over lunch and here was one suggestion that could work: add an atomic timestamp to each DenseMap. increment the time step whenever iterators are invalidated (e.g. insertions), Each iterator also holds a copy of the timestamp form when it was created so we can compare it against the parent container timestamp when\ever the iterator is used. I have not thought this through all this much, so it may not work in the end. Any feedback would be welcome. _______________________________________________ cfe-commits mailing list cfe-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits