[llvm-dev] [RFC] ASan: patches to support 32-byte shadow granularity (original) (raw)

Walter Lee via llvm-dev llvm-dev at lists.llvm.org
Wed Nov 1 08:36:06 PDT 2017

Thanks Kostya for the feedback.

On Tue, Oct 31, 2017 at 6:00 PM Kostya Serebryany <kcc at google.com> wrote:

- Stack errors: it seems sensible not to insert 32-byte sentinels between every object, but the result is that some stack overflow gets reported as unknown or use-after-scope. I have a patch that improves on the default behavior, but there remains cases where the error reports will not be as good.

Hmm. Not sure what's the problem here. It's totally fine to insert 32-byte redzone around stack objects. (in 32-byte granularity mode)

I was concerned about stack overhead, but I will go with your suggestion for now, and revisit when I have more data.

- Would it make sense to provide an internal compiler flag to set the shadow granularity, so that there we can at least run the instrumentation tests for 32-byte granularity in normal builds?

I'd prefer a proper flag, like -fsanitize-address-granularity=N (8,16,32)

Ok I'll do that.

- Is there a reasonable subset of tests I can port to 32-byte granularity to provide reasonable coverage, or should I aim to port all tests? Let's see what tests won't work out of the box and decide. We can mark all failing tests as UNSUPPORTED: 32-bit-granularity but ideally we shouldn't have to mark too many of those.

I will write up a classification of the failures. There are many tests that fail because they assume the shadow granularity; I will propose fixes for them.

Thanks,

Walter -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20171101/0e52ebdc/attachment.html>

More information about the llvm-dev mailing list