Re: [PATCH] asan unit tests from llvm lit-test incremental changes

[Konstantin Serebryany <konstantin.s.serebryany@gmail.com>]

On Thu, Dec 13, 2012 at 1:30 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Wed, Dec 12, 2012 at 10:16:49PM +0100, Dodji Seketeli wrote:
>> Independently of this review, I think it's be interesting to hear
>> Kostya's voice on:
>>
>> Jakub Jelinek <jakub@redhat.com> writes:
>>
>> > 2) In large-func-test-1.C, I had to stop matching the backtrace after
>> > _Znw[jm], because libasan is using the fast but inaccurate backtrace,
>> > and while the tests can be easily tweaked to compile with
>> > -fno-omit-frame-pointer, we definitely can't rely on libstdc++.so to be
>> > built with that option.  Most likely it isn't.

The tests should be built with -fno-omit-frame-pointer and we don't
need that for libstdc++.so.
This is how it works in LLVM.
asan's interceptors are written in such a way that they don't care
if libstdc++.so or the asan run-time have frame pointers.

>>> I repeat that I think
>> > that at least for Linux libasan should use the _Unwind* based backtrace
>> > at least for the fatal functions (__asan_report* etc.),

We are discussing it from time to time.
Sometimes, if e.g. an error happens inside a qsort callback,
the fp-based unwinder fails to unwind through libc, while _Unwind would work.

I was opposed to this sometime ago because _Unwind often produced
buggy stack traces on Ubuntu Lucid (the version we cared about).
I also vaguely remember some problems with _Unwind* depending on
malloc (or maybe that's something else?)
Now we mostly care about Ubuntu Precise and we need to test whether
_Unwind produces good enough results there before switching.

unwinding on malloc/free should keep using the fp-based unwinder, at
least by default.

We'll be tracking the issue in
https://code.google.com/p/address-sanitizer/issues/detail?id=137

>>  and perhaps for
>> > these malloc wrappers like ::operator new, ::operator new[] and their
>> > const std::nothrow_t& variants libasan could intercept them, call
>> > malloc and if that returns NULL, call the original corresponding function
>> > so that it deals with exceptions, new handler etc.

Hmm.. Why's that?
Calling libc's malloc or libstdc++'s operator new in asan run-time is
really a bad idea.
asan's allocator should never return 0 anyway, it should simply crash.
I don't think we want to support new handler at all.

>
> Yeah, I'd appreciate that too.





>
>> and on:
>>
>> > 3) deep-thread-stack-1.C fails for me right now with some libasan assertion,
>> > Kostya, can you please look at that?
>> >   AsanThread *t = asanThreadRegistry().GetCurrent();
>> >   CHECK(t);
>> > where it failed on the CHECK, because t was NULL.  I've skipped the test for
>> > now.
>>
>> [...]
>
> This one is for the testcase solved right now already by the -lasan -lpthread
> linking instead of just -lpthread (and driver adding -lasan afterwards).
> We'll need to think about how to tweak the driver to add -lasan early on the
> command line, before user passed -l* options.
>>
>> > --- gcc/testsuite/g++.dg/asan/deep-tail-call-1.C.jj 2012-12-04 20:24:10.000000000 +0100
>> > +++ gcc/testsuite/g++.dg/asan/deep-tail-call-1.C    2012-12-05 11:01:48.600443834 +0100
>> > @@ -1,21 +1,22 @@
>> > -// { dg-do run }
>> > +// { dg-do run }
>> >  // { dg-options "-fno-omit-frame-pointer -fno-optimize-sibling-calls" }
>> >  // { dg-additional-options "-mno-omit-leaf-frame-pointer" { target { i?86-*-* x86_64-*-* } } }
>> > -// { dg-shouldfail "asan" }
>> > +// { dg-shouldfail "asan" }
>> >
>> >  int global[10];
>> >  void __attribute__((noinline)) call4(int i) { global[i+10]++; }
>> >  void __attribute__((noinline)) call3(int i) { call4(i); }
>> >  void __attribute__((noinline)) call2(int i) { call3(i); }
>> >  void __attribute__((noinline)) call1(int i) { call2(i); }
>> > -int main(int argc, char **argv) {
>> > -  call1(argc);
>> > +volatile int one = 1;
>>
>> Just curious, why do we need this variable to be volatile, especially
>> since the test is compiled without optimization?
>
> asan.exp tests are torture tests, they iterate over several -O* options,
> unless explicitly dg-skip-if skipped.  It could be non-volatile with
> asm volatile ("" : : : "memory");
> or asm volatile ("" : "+m" (one)); or similar too, sure.
> I just don't want to rely on argc being one, and the compiler shouldn't know
> that one is 1 in the test.
>
>> [...]
>>
>> The patch looks OK to me in any case.
>
> Thanks.
>
>         Jakub


--kcc