Re: warning: objtool: fn1 uses BP as a scratch register

[Nick Desaulniers]

On Fri, Jul 19, 2019 at 11:44 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>
> On Fri, Jul 19, 2019 at 8:31 PM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> >
> > On Fri, Jul 19, 2019 at 11:23:16AM -0700, Nick Desaulniers wrote:
> > > On Fri, Jul 19, 2019 at 11:10 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > > >
> > > > A lot of objtool fixes showed up in linux-next, so I looked at some
> > > > remaining ones.
> > > > This one comes a lot up in some configurations
> > > >
> > > > https://godbolt.org/z/ZZLVD-
> > > >
> > > > struct ov7670_win_size {
> > > >   int width;
> > > >   int height;
> > > > };
> > > > struct ov7670_devtype {
> > > >   struct ov7670_win_size *win_sizes;
> > > >   unsigned n_win_sizes;
> > > > };
> > > > struct ov7670_info {
> > > >   int min_width;
> > > >   int min_height;
> > > >   struct ov7670_devtype devtype;
> > > > } a;
> > > > int b;
> > > > int fn1() {
> > > >   struct ov7670_info c = a;
> > > >   int i = 0;
> > > >   for (; i < c.devtype.n_win_sizes; i++) {
> > > >     struct ov7670_win_size d = c.devtype.win_sizes[i];
> > > >     if (c.min_width && d.width < d.height < c.min_height)
> > > >       if (b)
> > > >         return 0;
> > > >   }
> > > >   return 2;
> > > > }
> > > >
> > > > $ clang-8 -O2 -fno-omit-frame-pointer -fno-strict-overflow -c ov7670.i
> > > > $ objtool check  --no-unreachable --uaccess ov7670.o
> > > > ov7670.o: warning: objtool: fn1 uses BP as a scratch register
> > >
> > > Thanks for the report and reduced test case.  From the godbolt link, I
> > > don't see %rbp, %ebp, %bp, or %bpl being referenced (other that %rbp
> > > in the typical epilogue).  Am I missing something? Is objtool maybe
> > > not reporting the precise function at fault?
> >
> > I haven't looked, but it could very well be an objtool bug (surprise).
>
> Actually the reproducer may be wrong. I reduced the test case using
> 9.0.0-svn363902-1~exp1+0~20190620001509.2315~1.gbp76e756,
> and this contains a link
>
>          testl %ebp, %ebp
>
> I get the same thing with clang-8, but godbolt.org shows it only
> with clang-8 (see https://godbolt.org/z/g1lZO0) , not with trunk.

(Sorry for sending a blank email just now)
+ Craig and Simon, in case they recall this being a recent fix in
LLVM's x86 backend.
Sounds like this is fixing in upstream LLVM.  Guessing this could
result in improper unwinding w/ clang-8, but that's kind of irrelevant
for x86 as there's no asm goto (though we don't need CONFIG_JUMP_LABEL
in LTS branches :P)
-- 
Thanks,
~Nick Desaulniers