Re: [dpdk-dev] Build is broken in dpdk-next-net

[Neil Horman <nhorman@tuxdriver.com>]

On Sat, Mar 31, 2018 at 06:21:41PM +0200, Gaëtan Rivet wrote:
> On Sat, Mar 31, 2018 at 11:27:55AM -0400, Neil Horman wrote:
> > On Sat, Mar 31, 2018 at 05:09:47PM +0200, Gaëtan Rivet wrote:
> > > On Sat, Mar 31, 2018 at 09:33:43AM -0400, Neil Horman wrote:
> > > > On Fri, Mar 30, 2018 at 10:47:09PM +0800, Tonghao Zhang wrote:
> > > > > I rebuild it on ubuntu 17.10 and cash it. I use the 'RTE_SET_USED' to fix it.
> > > > > 
> > > > > 
> > > > > diff --git a/lib/librte_vhost/fd_man.c b/lib/librte_vhost/fd_man.c
> > > > > index 771675718..f11803191 100644
> > > > > --- a/lib/librte_vhost/fd_man.c
> > > > > +++ b/lib/librte_vhost/fd_man.c
> > > > > @@ -279,7 +279,8 @@ fdset_pipe_read_cb(int readfd, void *dat __rte_unused,
> > > > >                    int *remove __rte_unused)
> > > > >  {
> > > > >         char charbuf[16];
> > > > > -       read(readfd, charbuf, sizeof(charbuf));
> > > > > +       int r = read(readfd, charbuf, sizeof(charbuf));
> > > > > +       RTE_SET_USED(r);
> > > > >  }
> > > > > 
> > > > >  void
> > > > > @@ -319,5 +320,6 @@ fdset_pipe_init(struct fdset *fdset)
> > > > >  void
> > > > >  fdset_pipe_notify(struct fdset *fdset)
> > > > >  {
> > > > > -       write(fdset->u.writefd, "1", 1);
> > > > > +       int r = write(fdset->u.writefd, "1", 1);
> > > > > +       RTE_SET_USED(r);
> > > > >  }
> > > > > 
> > > > 
> > > > A better option might be to use _Pragma
> > > > 
> > > > Something like this perhaps
> > > > 
> > > > #define ALLOW_UNUSED(x) \
> > > > _Pragma(push) \
> > > > _Pragma(diagnostic ignored "-Wunused-result") \
> > > > #x;\
> > > > _Pragma(pop)
> > > > 
> > > > This is of course untested, so it probably needs some tweaking, but this method
> > > > avoids the need to declare an additional stack variable, which i don't think can
> > > > be eliminated due to the cast.  I believe that this method should also work
> > > > accross compilers (the gcc and clang compilers support this, and i think the
> > > > intel compiler should as well)
> > > > 
> > > > Neil
> > > > 
> > > 
> > > It would be nice to avoid the definition of a useless variable.
> > > An alternative could be
> > > 
> > >    if (read() < 0) {
> > >        /* Failure here is acceptable for such and such reason. */
> > >    }
> > > 
> > > to ensure all-around compatibility, and the definition or another macro.
> > > Just a suggestion.
> > > 
> > That would be a good alternative, but I think its effectiveness is dependent on
> > when the compiler does with the return value check. Without any code inside the
> > conditional, the compiler may optimize the check out, meaning the warning will
> > still be asserted.  If it doesn't optimize the check out, then you have a
> > useless compare and jump instruction left in the code path.
> > 
> > Best
> > Neil
> > 
> 
> I tested quickly, I see no difference with the three methods:

gcc seems to be sufficiently smart to optimize out the conditional, clang not so
much:

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>

__attribute__((warn_unused_result))
int wur(void)
{
	printf("CALLING WUR!\n");
        return read(0, NULL, 0);
}

#define UNUSED_RESULT(x) if (x) {}

int main(void)
{
	UNUSED_RESULT(wur());
        return 0;
}

[nhorman@neilslaptop ~]$ gcc -g -Wunused-result -Werror ./test.c
[nhorman@neilslaptop ~]$ objdump -d -S a.out > ./results
[nhorman@neilslaptop ~]$ cat results
... 
000000000040054b <main>:

#define UNUSED_RESULT(x) if (x) {}

int main(void)
{
  40054b:       55                      push   %rbp
  40054c:       48 89 e5                mov    %rsp,%rbp
        UNUSED_RESULT(wur());
  40054f:       e8 d3 ff ff ff          callq  400527 <wur>
        return 0;
  400554:       b8 00 00 00 00          mov    $0x0,%eax
}
  400559:       5d                      pop    %rbp
  40055a:       c3                      retq
  40055b:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)


[nhorman@neilslaptop ~]$ clang -g -Wunused-result -Werror ./test.c
[nhorman@neilslaptop ~]$ objdump -d -S a.out > ./results
[nhorman@neilslaptop ~]$ cat results 
...
0000000000400570 <main>:
}

#define UNUSED_RESULT(x) if (x) {}

int main(void)
{
  400570:       55                      push   %rbp
  400571:       48 89 e5                mov    %rsp,%rbp
  400574:       48 83 ec 10             sub    $0x10,%rsp
  400578:       c7 45 fc 00 00 00 00    movl   $0x0,-0x4(%rbp)
        UNUSED_RESULT(wur());
  40057f:       e8 ac ff ff ff          callq  400530 <wur>
  400584:       83 f8 00                cmp    $0x0,%eax
  400587:       0f 84 05 00 00 00       je     400592 <main+0x22>
  40058d:       e9 00 00 00 00          jmpq   400592 <main+0x22>
  400592:       31 c0                   xor    %eax,%eax
        return 0;
  400594:       48 83 c4 10             add    $0x10,%rsp
  400598:       5d                      pop    %rbp
  400599:       c3                      retq
  40059a:       66 0f 1f 44 00 00       nopw   0x0(%rax,%rax,1)


There is an additional compare and two jump statements there.  I'm sure
eventually most compilers will figure out how to eliminate this, and it might
even do so now with the right optimization flags, but I think its best to just
organize the source such that no conditional branching is implied.  Assuming the
intel compiler supports it (which I think it should, can someone with access to
it confirm), the _Pragma utility is probably the most clear way to do that.

Regards
Neil