Re: ld.lld: error: undefined symbol: __udivdi3

[Nathan Chancellor <nathan@kernel.org>]

On Wed, Dec 07, 2022 at 09:32:41PM +0300, Mikhail Zhilkin wrote:
> Hi Nathan,
> 
> On 12/6/2022 8:03 PM, Nathan Chancellor wrote:
> 
> > Hi Mikhail,
> >
> > First and foremost, apologies for the delay in my response!
> >
> > On Thu, Dec 01, 2022 at 09:39:22PM +0300, Mikhail Zhilkin wrote:
> >> Hi Miquèl, Nathan,
> >>
> >> On 12/1/2022 7:13 PM, Miquel Raynal wrote:
> >>> Hi Nathan,
> >>>
> >>> nathan@kernel.org wrote on Thu, 1 Dec 2022 09:00:57 -0700:
> >>>
> >>>> On Thu, Dec 01, 2022 at 09:16:10AM +0100, Miquel Raynal wrote:
> >>>>> Hi Mikhail, Hiroshi,
> >>>>>
> >>>>> lkp@intel.com wrote on Thu, 1 Dec 2022 01:46:36 +0800:
> >>>>>
> >>>>>> tree:   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> >>>>>> head:   01f856ae6d0ca5ad0505b79bf2d22d7ca439b2a1
> >>>>>> commit: 9b78ef0c7997052e9eaa0f7a4513d546fa17358c mtd: parsers: add support for Sercomm partitions
> >>>>>> date:   6 months ago
> >>>>>> config: mips-randconfig-r033-20221128
> >>>>>> compiler: clang version 16.0.0 (https://github.com/llvm/llvm-project 6e4cea55f0d1104408b26ac574566a0e4de48036)
> >>>>>> reproduce (this is a W=1 build):
> >>>>>>         wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
> >>>>>>         chmod +x ~/bin/make.cross
> >>>>>>         # install mips cross compiling tool for clang build
> >>>>>>         # apt-get install binutils-mipsel-linux-gnu
> >>>>>>         # https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=9b78ef0c7997052e9eaa0f7a4513d546fa17358c
> >>>>>>         git remote add linus https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
> >>>>>>         git fetch --no-tags linus master
> >>>>>>         git checkout 9b78ef0c7997052e9eaa0f7a4513d546fa17358c
> >>>>>>         # save the config file
> >>>>>>         mkdir build_dir && cp config build_dir/.config
> >>>>>>         COMPILER_INSTALL_PATH=$HOME/0day COMPILER=clang make.cross W=1 O=build_dir ARCH=mips SHELL=/bin/bash
> >>>>>>
> >>>>>> If you fix the issue, kindly add following tag where applicable
> >>>>>> | Reported-by: kernel test robot <lkp@intel.com>
> >>>>>>
> >>>>>> All errors (new ones prefixed by >>):
> >>>>>>
> >>>>>>>> ld.lld: error: undefined symbol: __udivdi3
> >>>>>>    >>> referenced by scpart.c
> >>>>>>    >>>               mtd/parsers/scpart.o:(scpart_parse) in archive drivers/built-in.a
> >>>>>>
> >>>>> Can you please try to reproduce this with a mips toolchain as advised
> >>>>> and send a fix?
> >>>> For what it's worth, this is likely our bug:
> >>>>
> >>>> https://github.com/ClangBuiltLinux/linux/issues/1635
> >>>>
> >>>> It sounds like there are some merged LLVM patches that should help but I
> >>>> guess not based on this report...
> >>> Ah, yeah, thanks a lot for the share.
> >>>
> >>> Cheers, Miquèl
> >> I could reproduce the issue and found that reference to __udivdi3
> >> appears in this function:
> >>
> >> static uint32_t scpart_desc_is_valid(struct sc_part_desc *pdesc)
> >> {
> >>         return ((pdesc->part_id != 0xffffffffUL) &&
> >>                 (pdesc->part_offs != 0xffffffffUL) &&
> >>                 (pdesc->part_bytes != 0xffffffffUL));
> >> }
> >>
> >> There is no LLVM compiler error if:
> >>
> >> static uint32_t scpart_desc_is_valid(struct sc_part_desc *pdesc)
> >> {
> >>         return (0);
> >> }
> >>
> >> Currently I have no ideas where does __udivdi3 come from, because
> >> there's no division here. Should I try to rewrite somehow the function
> >> to avoid the error?
> >> I would greatly appreciate any ideas.
> > I do not think it is that function that is causing the problem. Making
> > it return zero just causes the code generation to change enough to avoid
> > this issue.
> >
> > Based on the previous analysis and reduction [1], LLVM is transforming
> > the for loop in scpart_scan_partmap() into division; you can verify this
> > by commenting it out some way then rebuilding and running 'readelf -s'
> > to look for __udivdi3.
> >
> > Arnd suggested working around this problem by making offs a 32-bit type,
> > as it is already bounded by ->erasesize, which appears to be a 32-bit
> > value from what I can tell. Making a change like so appears to resolve
> > the issue for me but I am not sure if that has any other implications.
> >
> > diff --git a/drivers/mtd/parsers/scpart.c b/drivers/mtd/parsers/scpart.c
> > index 02601bb33de4..6e5e11c37078 100644
> > --- a/drivers/mtd/parsers/scpart.c
> > +++ b/drivers/mtd/parsers/scpart.c
> > @@ -50,7 +50,7 @@ static int scpart_scan_partmap(struct mtd_info *master, loff_t partmap_offs,
> >  	int cnt = 0;
> >  	int res = 0;
> >  	int res2;
> > -	loff_t offs;
> > +	uint32_t offs;
> >  	size_t retlen;
> >  	struct sc_part_desc *pdesc = NULL;
> >  	struct sc_part_desc *tmpdesc;
> >
> > [1]: https://github.com/ClangBuiltLinux/linux/issues/1635
> 
> Many thanks for your time and very clear explanation!
> 
> I've made compile and run tests. It seems that have the working
> workaround and I'm going to send the patch. After analyzing the code, I
> don't expect any other implications.
> 
> Who should I add to Suggested-by of the patch: You or maybe Arndt?

I think Arnd should be the suggested-by on that patch, you should be
able to get his info from either MAINTAINERS or git log. Thanks for
taking a look and sending a patch!

Cheers,
Nathan