Re: [RESEND PATCH v1] x86/build: add -fno-builtin flag to prevent shadowing

[Nathan Chancellor]

On Sun, May 08, 2022 at 09:37:14PM +0900, Vincent MAILHOL wrote:
> Hi Arnd,
> 
> +CC: Kees Cook
> 
> On Sun. 8 May 2022 at 19:27, Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
> > On Sun, May 8, 2022 at 12:09 PM Vincent Mailhol
> > <mailhol.vincent@xxxxxxxxxx> wrote:
> > >
> > > Aside of the __builtin_foo() ones, x86 does not directly rely on any
> > > builtin functions.
> > >
> > > However, such builtin functions are not explicitly deactivated,
> > > creating some collisions, concrete example being ffs() from bitops.h,
> > > c.f.:
> > >
> > > | ./arch/x86/include/asm/bitops.h:283:28: warning: declaration of 'ffs' shadows a built-in function [-Wshadow]
> > > |   283 | static __always_inline int ffs(int x)
> > >
> > > This patch adds -fno-builtin to KBUILD_CFLAGS for the x86
> > > architectures in order to prevent shadowing of builtin functions.
> > >
> > > Signed-off-by: Vincent Mailhol <mailhol.vincent@xxxxxxxxxx>
> > > ---
> > > FYI, I tested this patch on a "make allyesconfig" for both x86_32 and
> > > x86_64.
> > >
> > > This is a resend. Only difference is that I dropped the RFC flag and
> > > added Arnd in CC because he did a similar patch to fix ffs shadow
> > > warnings in the past:
> > >
> > > https://lore.kernel.org/all/20201026160006.3704027-1-arnd@xxxxxxxxxx/
> >
> > I think this is a correct change, but unfortunately it exposes a clang bug
> >  with -mregparm=3. Nick should be able to provide more details, I think
> > he has a plan.
> 
> Interesting. I admittedly did not do extensive tests on clang
> but I would have expected the Linux kernel bot to have warned me
> on my previous patch.
> 
> I did research on mregparm and clang. I found this thread:
> https://lore.kernel.org/r/20220208225350.1331628-9-keescook@xxxxxxxxxxxx
> 
> and the associated LLVM issue:
> https://github.com/llvm/llvm-project/issues/53645
> 
> Those threads mention that some clang builtins become unusable
> when combining -mregparm=3 and -m32. But I could not find a
> bug reference about -mregparm=3 and -fno-builtin combination.
> 
> Could you just double confirm that you indeed saw the issue with
> -fno-builtin? If that the case, I am really curious to get the
> details :)

-ffreestanding implies -fno-builtin; removing -ffreestanding from
arch/x86/Makefile for 32-bit x86 caused the problem so I don't think
that your patch would cause any issue but I could be missing something.

However, doesn't -fno-builtin remove the ability for GCC and clang to
perform certain libcall optimizations? I seem to recall this coming up
in previous threads but I am having a hard time finding the exact
language that I was looking for. This thread seems to be the most recent
one that I can remember:

https://lore.kernel.org/CAHk-=whn91ar+EbcBXQb9UXad00Q5WjU-TCB6UBzVba682a4ew@xxxxxxxxxxxxxx/

Cheers,
Nathan