Re: Autoconf 2.69 AC_PROG_CXX accepts plain C compilers

[Eric Blake]

On 02/05/2014 01:04 PM, Peter De Wachter wrote:
> I was testing a C++ project of mine with older versions of GCC, and
> mistakingly used the following command:
> $ ./configure CXX=gcc-4.7
> 
> Autoconf didn't spot any problem:
>> checking whether the C++ compiler works... yes
> But of course gcc-4.7 doesn't know how to link a C++ program, so my
> build failed.
> 
> The attached patch fixes this problem by changing the C++ null program
> into:
> 
> int main() { (void) new int; return 0; }
> 
> Any C++ compiler, even an ancient one, should be able to pass that test,
> but C compilers will fail. In GCC's case, the program will compile but
> the link will fail because libc doesn't contain an 'operator new'.

Actually, in gcc's case, your proposed program is now a syntax error:

$ gcc -o foo foo.c
foo.c: In function ‘main’:
foo.c:1:21: error: ‘new’ undeclared (first use in this function)
 int main() { (void) new int; return 0; }
                     ^
foo.c:1:21: note: each undeclared identifier is reported only once for
each function it appears in
foo.c:1:25: error: expected ‘;’ before ‘int’
 int main() { (void) new int; return 0; }

But that's equally useful for weeding out invalid C++ compilers :)

I'm wondering if we should avoid the memory leak by using 'delete new
int' instead of '(void) new int', just so we are less likely to trip up
on a compiler warning causing a false negative.  Or maybe even some
other construct that doesn't involve memory allocation but is truly a
no-op C++ program that fails to compile under C (such as
AC_LANG_PROGRAM([class foo]) by exploiting 'class' rather than 'new').
Any opinions?  Otherwise, the idea for your patch looks good to me.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature