Re: objtool warning "uses BP as a scratch register" with clang-9

[Nick Desaulniers]

On Tue, Aug 27, 2019 at 12:47 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>
> On Tue, Aug 27, 2019 at 9:23 PM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > On Tue, Aug 27, 2019 at 09:00:52PM +0200, Arnd Bergmann wrote:
> > > On Tue, Aug 27, 2019 at 5:00 PM Ilie Halip <ilie.halip@xxxxxxxxx> wrote:
> > > >
> > > > > > $ clang-9 -c  crc32.i  -O2   ; objtool check  crc32.o
> > > > > > crc32.o: warning: objtool: fn1 uses BP as a scratch register
> > > >
> > > > Yes, I see it too. https://godbolt.org/z/N56HW1
> > > >
> > > > > Do you still see this warning with -fno-omit-frame-pointer (assuming
> > > > > clang has that option)?
> > > >
> > > > Using this makes the warning go away. Running objtool with --no-fp
> > > > also gets rid of it.
> > >
> > > I still see the warning after adding back the -fno-omit-frame-pointer
> > > in my reduced test case:
> > >
> > > $ clang-9 -c  crc32.i -Werror -Wno-address-of-packed-member -Wall
> > > -Wno-pointer-sign -Wno-unused-value -Wno-constant-logical-operand -O2
> > > -Wno-unused -fno-omit-frame-pointer
> > > $ objtool check  crc32.o
> > > crc32.o: warning: objtool: fn1 uses BP as a scratch register
> >
> > This warning most likely means that clang is clobbering RBP in leaf
> > functions.  With -fno-omit-frame-pointer, leaf functions don't need to
> > set up the frame pointer, but they at least need to leave RBP untouched,
> > so that an interrupts/exceptions can unwind through the function.
>
> Yes, that clearly matches what I see in the output where it does
>
>    0: 55                    push   %rbp
> ...
>   73: 0f b6 ef              movzbl %bh,%ebp
>   76: 8b 1c 99              mov    (%rcx,%rbx,4),%ebx
>   79: 33 1c aa              xor    (%rdx,%rbp,4),%ebx
> ...
>   95: 5d                    pop    %rbp
>   96: c3                    retq
>
> I just did another simple test: an x86-64 defconfig build with
> UNWINDER_FRAME_POINTER shows the exact symptom as
> my randconfig, so it sounds like any configuration with frame
> pointers would, and there is nothing else to it (this also makes
> sense given that it happens with a relatively simple test case
> outside of the kernel).
>
>        Arnd

Thanks for the description of the issue and the reduced test case.  It
almost reminds me of
https://github.com/ClangBuiltLinux/linux/issues/612.

I've filed https://bugs.llvm.org/show_bug.cgi?id=43128, anything I
should add to the bug report?

-- 
Thanks,
~Nick Desaulniers