Hi Steve
On Tue, Mar 20, 2018 at 02:10:52PM -0700, Steve Langasek wrote:
> Considering debian/rules also hardcodes -O2 for the package itself, it
> doesn't seem inappropriate to also build the test shim with the same
> options?
> Yes, this is working around a toolchain issue. Feel free to reassign this
> bug to the toolchain, if you prefer.
I digged deeper and the arguments do not matter. The whole problem is
not a toolchain issue, it's actually triggered by a small different in
the used prototypes.
> FWIW I was able to identify a symbol reference that drops out of the .so
> when built with -O3 but is present when built with -O2; this may be the
> cause of the crash:
> -0000000000000000 DF *UND* 0000000000000000 GLIBC_2.17 0x60
> __stack_chk_fail
> And perhaps that points to an explanation of the toolchain bug.
In my own tests this different does not show up. Also it's a undefined
symbol, so it is for a call into something else.
> It's annoyingly difficult to debug this crash, as gdb doesn't appear to let
> me DTRT with 'set env LD_PRELOAD' and instead manages to crash itself before
> handing control over to the process under test.
Yes, it is. And adding debugging output makes it disappear.
Anyway, the cause is:
| static int (*orig_open)(const char *pathname, int flags, int mode);
This does not match the public prototype for the open function.
Regards,
Bastian
--
Insufficient facts always invite danger.
-- Spock, "Space Seed", stardate 3141.9
