On Mon, May 11, 2020 at 12:34 PM Brian Gerst <brge...@gmail.com> wrote:
>
> On Mon, May 11, 2020 at 2:46 PM Nick Desaulniers
> <ndesaulni...@google.com> wrote:
> >
> > On Mon, May 11, 2020 at 11:09 AM Brian Gerst <brge...@gmail.com> wrote:
> > > This looks like the same issue that we just discussed for bitops.h.
> > > Add the "b" operand size modifier to force it to use the 8-bit
> > > register names (and probably also needs the "w" modifier in the 16-bit
> > > case).
> >
> > While it does feel familiar, it is slightly different.
> > https://godbolt.org/z/Rme4Zg
> > That case was both compilers validating the inline asm, yet generating
> > assembly that the assembler would choke on. This case is validation
> > in the front end failing.
>
> > long long ret;
> > switch (sizeof(ret)) {
> > case 1:
> > asm ("movb $5, %0" : "=q" (ret));
> > break;
> > case 8:;
> > }
>
> So if the issue here is that the output variable type is long long,
> what code is using a 64-bit percpu variable on a 32-bit kernel? Can
> you give a specific file that fails to build with Clang? If Clang is
> choking on it it may be silently miscompiling on GCC.
I'm not sure that's the case. Applying this patch, undoing the hunk
in percpu_from_op() we get tons of errors. Looking at one:
kernel/events/core.c:8679:8: error: invalid output size for constraint '=q'
./include/linux/percpu-defs.h:446:2: note: expanded from macro '__this_cpu_read'
raw_cpu_read(pcp); \
^
...
There's nothing wrong with this line, it's reading a percpu u64 into a
local u64. The error comes from validating the inline asm in the dead
branch.
--
Thanks,
~Nick Desaulniers
