On Tue, 10 Oct 2017 09:03:50 +0200, Geert Uytterhoeven wrote:
> Hi Jakub,
>
> On Tue, Oct 10, 2017 at 12:53 AM, Jakub Kicinski
> <jakub.kicinski@xxxxxxxxxxxxx> wrote:
> > On Mon, 9 Oct 2017 10:40:49 +0200, Geert Uytterhoeven wrote:
> >> On most architectures[*], gcc turns __builtin_ffsll() into a call to
> >> __ffsdi2(), which is not provided by any architecture, leading to
> >> failures like:
> >>
> >> rcar-gen3-cpg.c:(.text+0x289): undefined reference to `__ffsdi2'
> >>
> >> To fix this, use __ffs64() instead, which is available on all
> >> architectures.
> >>
> >> [*] Known exceptions are some 64-bit architectures like amd64, arm64,
> >> ia64, powerpc64, and tilegx.
> >>
> >> Reported-by: kbuild test robot <fengguang.wu@xxxxxxxxx>
> >> Fixes: 3e9b3112ec74f192 ("add basic register-field manipulation macros")
> >> Signed-off-by: Geert Uytterhoeven <geert+renesas@xxxxxxxxx>
> >> ---
> >> include/linux/bitfield.h | 3 ++-
> >> 1 file changed, 2 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/include/linux/bitfield.h b/include/linux/bitfield.h
> >> index 8b9d6fff002db113..0a827677372756fa 100644
> >> --- a/include/linux/bitfield.h
> >> +++ b/include/linux/bitfield.h
> >> @@ -15,6 +15,7 @@
> >> #ifndef _LINUX_BITFIELD_H
> >> #define _LINUX_BITFIELD_H
> >>
> >> +#include <linux/bitops.h>
> >> #include <linux/bug.h>
> >>
> >> /*
> >> @@ -46,7 +47,7 @@
> >> * reg |= FIELD_PREP(REG_FIELD_C, c);
> >> */
> >>
> >> -#define __bf_shf(x) (__builtin_ffsll(x) - 1)
> >> +#define __bf_shf(x) __ffs64(x)
> >
> > Hm. The build bot failure made me think. I think rcar-gen3-cpg.c may
> > be doing something wrong here, could you point me at the patch in
> > question? I don't see any FIELD_* there in Linus's tree.
>
> See series "[PATCH v3 0/6] clk: renesas: r8a779[56]: Add Z and Z2 clock
> support" (https://www.spinics.net/lists/arm-kernel/msg609499.html).
>
> > __bf_shf() is supposed to be used with constant masks only, therefore
> > the call must be optimized away completely.
> >
> >> #define __BF_FIELD_CHECK(_mask, _reg, _val, _pfx) \
> >> ({ \
>
> IC.
>
> Yes, it looks like __ffs64() can't be optimized away like __builtin_ffsll() :-(
>
> Apparently the patch series above uses __bf_shf() directly, to avoid the
> BUILD_BUG_ON_MSG(), which doesn't work when the call isn't optimized away.
> Sorry for not noticing that before...
>
> One way to fix that (non-)API abuse would be to get rid of __bf_shf(),
> and open code it as __builtin_ffsll(x) - 1 everywhere...
>
> What do you think?
I'm starting to feel like we/I should come up with a way to convert the
masks into something that can be passed to functions :( The "mask must
be constant" problem is coming up again and again. The only problem is
I'm not sure how to pack the mask+shift into a single argument :( Also
the mask can be 32 or 64 bit, it would be nice to not force the output
to always be 64bit...
