Re: libsanitizer merge from upstream r196090

[Jeff Law <law at redhat dot com>]

On 12/03/13 22:28, Konstantin Serebryany wrote:
> I really think that we need to disable libsanitizer on old systems
> until someone volunteers to set up a proper testing process upstream.
> If no one volunteers -- no one really needs it.
The right way to do this is to do feature tests rather than just 
declaring something to be too old based on a version number.  Using 
version #s is a decent, but not great fall back (though it is often 
easier to writing the appropriate feature tests).

Do you use autoconf upstream?  If so, testing for things like the 
existence of particular header files is trivial and avoids nonsense like 
this entirely:

> >   #if !SANITIZER_ANDROID
> > +#include <linux/version.h>
> > +// <linux/perf_event.h> has been added in 2.6.32
> > +#if LINUX_VERSION_CODE >= 132640
> >   #include <linux/perf_event.h>
> >   #endif
> > +#endif
This kind of testing is so easy with autoconf that it's just silly.


> >   namespace __sanitizer {
> >     unsigned struct_statfs64_sz = sizeof(struct statfs64);
> > @@ -75,15 +79,18 @@ CHECK_SIZE_AND_OFFSET(io_event, res);
> >   CHECK_SIZE_AND_OFFSET(io_event, res2);
> >
> >   #if !SANITIZER_ANDROID
> > +#if LINUX_VERSION_CODE >= 132640
> >   COMPILER_CHECK(sizeof(struct __sanitizer_perf_event_attr) <=
> >                  sizeof(struct perf_event_attr));
> >   CHECK_SIZE_AND_OFFSET(perf_event_attr, type);
> >   CHECK_SIZE_AND_OFFSET(perf_event_attr, size);
> >   #endif
> > +#endif
Couldn't this be done with a test as well.  Given header files and the 
ability to run the compiler, it should be fairly easy to test for this, 
even in cross environments.

> >   COMPILER_CHECK(iocb_cmd_pread == IOCB_CMD_PREAD);
> >   COMPILER_CHECK(iocb_cmd_pwrite == IOCB_CMD_PWRITE);
> > -#if !SANITIZER_ANDROID
> > +#if !SANITIZER_ANDROID && LINUX_VERSION_CODE >= 132627
> > +// IOCB_CMD_PREADV/PWRITEV has been added in 2.6.19
> >   COMPILER_CHECK(iocb_cmd_preadv == IOCB_CMD_PREADV);
> >   COMPILER_CHECK(iocb_cmd_pwritev == IOCB_CMD_PWRITEV);
> >   #endif
Also trivial to do with autoconf.

[ ... ]

But more generally, I'd look real closely at anything including linux/ 
headers directly and see if it can be reasonably avoided.    The 
ultimate result will actually be easier maintenance upstream over the 
long term.

jeff