- User Since
- Apr 14 2017, 1:59 PM (83 w, 5 d)
Tue, Nov 13
@fjricci Now the test is actually failing, because the size of e.g. "__objc_classlist" is 16, and kMaxSegName is also 16. I'm changing < to <=, and adding +1 to the second strnlen arguments.
(unless there are any objections)
LGTM, but I would drop the word "convenient", and add an explicit \return section in doxygen.
Mon, Nov 12
In the future, for macOS-specific changes I think it would be better to wait for a sign-off from at least one maintainer who is an expert on Apple tools.
I suspect it broke http://green.lab.llvm.org/green/job/clang-stage2-Rthinlto/14090/console as well.
I don't quite understand the need for this patch.
If we are talking about a binary built from source, wouldn't it make more sense to build libcxx from source as well?
If libcxx is built from source in the same repo, clang-based tools do find it.
From a first glance, it's missing screenshots and an introduction section.
Also, remove diags::note_suggest_disabling_all_checkers.
Tue, Nov 6
Mon, Nov 5
@kcc OK makes sense.
Intuitively it still seems to make sense to me that vectorization and unrolling would make fuzzing worse, but you are right that a lot more data would be needed to make that decision.
@kcc the other one was shrink.test failing on x86_64h, because it was more eager to unroll.
@kcc My experience with this is limited, but so far I have seen at least two examples where it makes things worse.
Extra signal is not necessarily better, especially if it is noise, and hides the good signal.
I did some more digging: it seems the source of the problem is the same as for the loop unroller.
Before the transformation, on the example of interest (counters.test) there exists a single sanitizer counter responsible for the "interesting" branch,
and the test is doing the search for the input which would cause the counter to be incremented 4 times.
@fhahn Setting force-target-max-scalar-interleave=1 seems to work on this example.
Would it make more sense to read the attribute in LoopVectorize, or to try to set this option in this option in the driver?
@lebedev.ri yeah ASAN is making stack frame size larger.
Fri, Nov 2
@Szelethus you can see the command required to get modules build in the mentioned log. I think -DLLVM_ENABLE_MODULES=On might be enough.
Thu, Nov 1
oh, you mean specifically no optimizations here. you are right.
@kcc counters.test on ARM64?
@vitalybuka it depends on the parent revision, which is more complicated.
Minor nit: let's create a custom substitution for your diff command, like diff_plist.
Wed, Oct 31
Tue, Oct 30
Actually, I'm reverting this: the rule seems fairly ubiquitous.
I much prefer this version.
We've had the same problem with diffing plist output.
One thing we have learned is using a FileCheck was definitely a bad idea, as it leads to unreadable, unmaintainable, and very hard to update tests,
so either diff or your custom tool is way better.
Mon, Oct 29
I don't think a new PathGenerationScheme is needed, unless you plan changes to BugReporter.cpp.
Minor style notes + context missing.
Patch context is missing.
Thu, Oct 25
Or maybe it does not: we should still be able to catch cases where OSDynamicCast fails.
Actually, apologies again: I have rushed through this too much.
@NoQ has a good point: we need to preserve the distinction between the things analyzer "assumes" and the things analyzer "knows".
Minor nitpicks, otherwise makes sense!
Thanks a lot, this is almost ready!
I think it should be good to go after this round of nitpicks.
Wed, Oct 24
I'll try to take a look this week.
Good to go provided you will add an example.
If we want to be serious about this page, it really has to be auto-generated (like clang-tidy one), but I understand that this is a larger undertaking.