I'll add in the prefix before the final version for commit. The goal is to detect globals with dynamic initialization per the C++ standard, so that ASan knows which globals can only be accessed during . It's valid to access a statically initialized global inside of an initializer, even if that global resides in a different TU. Thus, LLVM/ASan need to be able to distinguish between dynamically/statically initialized globals. I'm not certain this is optimal, but there haven't been any unexpected false positives with it. We may be missing a few case, and the optimizer can certainly end up leaving us with false negatives, but those are much better than false positives.
On Thu, Aug 16, 2012 at 5:27 PM, Eric Christopher <[email protected]> wrote: > > > On Aug 16, 2012, at 1:05 AM, Kostya Serebryany <[email protected]> wrote: > >> +llvm-commits >> >> Reid, >> >> The LLVM and compiler-rt patches look good. >> Please fix the remaining small issues (see my code review comments) and >> commit. >> Hold on with the output tests for a bit since Alexey Samsonov is migrating >> them to cmake (please coordinate with him and commit as a separate patch). >> >> The stress test should contain X files, Y linker initialized globals and Z >> dynamically initialized globals. >> Such test only makes sense where all 3 numbers are large. >> I guess you can commit a single .sh script into compiler-rt/lib/asan/scripts > > The metadata should at least be prefixed with something like > llvm.asan.<whatever> instead of just the name. That way it's more > identifiable. > > What's the idea behind the metadata use anyhow? > > -eric _______________________________________________ cfe-commits mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
