Re: heap corruption in du

[Mikulas Patocka]

On Mon, 10 Oct 2005, Paul Eggert wrote:

> Mikulas Patocka <address@hidden> writes:
>
> > I mean if there's some probability that these experimental versions
> > will break system, possibly even breaking make install and preventing
> > reinstallation of old versions?
>
> I doubt that would happen.  But if you're worried about it, save the
> old versions before installing the new ones.

Just a bug report --- I got a failed build on intel C compiler 9.0. The 
problem is in autoconf. autoconf creates conftest.c files like this:

char pstat_getdynamic();
char (*f)()=pstat_getdynamic;
int main()
{
return f != pstat_getdynamic;
}

Intel compiler's interprocedural optimizations (-ipo) optimizes this out 
even though the symbol is unknown. Linking succeeds and ./a.out executable 
returning 0 is generated.

Coreutils than attempt to use the function and fail.

With this optimization, the compiler doesn't generate regular *.o files, 
but generates *.o files in its own intercode --- and when you attempt to 
link these files, the compiler takes intercode of all modules, optimizes 
it and runs linker on it. This whole-project optimization allows more 
advanced optimizations, such as finding that "f" is constant through the 
program, replacing "f != pstat_getdynamic" with zero and finally removing 
"f" because it is unused. Linker than sees no reference to 
pstat_getdynamic and generates proper ./a.out.

If you change the conftest file to somehow use the call to the function 
being tested, for example:

char pstat_getdynamic();
char (*f)()=pstat_getdynamic;
int main(int argc, char *argv[])
{
if (argc == 1) return f != pstat_getdynamic;
return f();
}

, the compiler will properly fail on unknown function.

Mikulas