Jakub,
Great result!
Ideally, I would like to limit the differences from upstream.
I'll put some of your changes upstream, for others I'd ask you to
consider other choices.
-#include "gtest/gtest.h"
+#include "dejagnu-gtest.h"
Maybe like this?
#if ASAN_USE_DEJAGNU_GTEST
#include "dejagnu-gtest.h"
#else
#include "gtest/gtest.h"
#endif
Can we have gcc_asan_test.C which will #include the unchanged (modulo
the comment header) asan_test.cc
and have dejagnu lines there?
Like this:
// { dg-do run { target { mmap && pthread } } }
...
#include asan_test.cc
+#elif defined(__SANITIZE_ADDRESS__)
+ bool asan = 1;
I'll put this upstream.
+#ifdef __SANITIZE_ADDRESS__
+ // Avoid this test during dejagnu testing, it is too expensive
+ if (getenv ("GCC_TEST_RUN_EXPENSIVE") == NULL)
+ return;
+#endif
I'd prefer to simply put this w/o ifdef.
-# error "please define ASAN_HAS_BLACKLIST"
+//# error "please define ASAN_HAS_BLACKLIST"
You can add -DASAN_HAS_BLACKLIST=0 to the command line.
If/when gcc gets blacklist support, we'll redefine it to 1
+#if __has_feature(address_sanitizer) || defined(__SANITIZE_ADDRESS__)
this is upstreamable
+#ifdef __GNUC__
+# define break_optimization(arg) __asm__ __volatile__ ("" : : "r"
(arg) : "memory")
+#endif
+
That's a nice piece of magic, let me use this too.
Cool!
--kcc
On Fri, Nov 30, 2012 at 12:46 AM, Jakub Jelinek <[email protected]> wrote:
> Hi!
>
> On Wed, Nov 28, 2012 at 03:13:14PM +0400, Konstantin Serebryany wrote:
>> That's a bit scary (and will be slower than with gtest).
>> But if we can limit the changes by replacing
>> asan/tests/asan_test_config.h (and maybe some minimal other changes)
>> that may work.
>
> So, here is a rough port of asan_test.cc (haven't added the few smaller
> other tests yet until this one is resolved).
>
> I'm attaching both gcc patch and diff between the current llvm files
> and the files added in the patch to make it clear what changes were done.
> Primarily it is replacing the gtest/gtest.h include with new dejagnu-gtest.h
> header that contains needed macros, adding some tcl stuff around it and
> a few dejagnu lines at the beginning.
>
> I've disabled the ManyThreads test, because it seems to be almost completely
> library bound and was very slow (order of minutes), many people run make
> check for gcc with very high -jN factors and having some test spawn 1000
> threads and in each allocate lots of memory is undersirable for normal
> testing.
>
> Attaching also make check-g++ RUNTESTFLAGS='--target_board=unix\{-m32,-m64\}
> asan.exp'
> result, there are some failures that need analysis (some of it might be
> because of the missing __asan_handle_no_return instrumentation from GCC,
> waiting there for review of the prerequisite patch, another thing is
> instrumentation of bitfields).
> But for -m32/-m64 together that is still:
>
> # of expected passes 2301
> # of unexpected failures 61
> # of unsupported tests 18
>
> so not that bad. Both -m32 and -m64 testing together took around 90 seconds
> (without the ManyThreads tests, with it (GCC_TEST_RUN_EXPENSIVE=1 in
> environment) more than 4 minutes).
>
> Jakub
