[PATCH] riscv: Handle zicsr/zifencei issue between gcc and binutils

[Conor Dooley]

On Tue, Jul 25, 2023 at 07:57:54PM +0100, Conor Dooley wrote:
> On Tue, Jul 25, 2023 at 10:23:44AM -0700, Nathan Chancellor wrote:
> > On Wed, Jul 26, 2023 at 01:04:05AM +0800, Mingzheng Xing wrote:
> > > When compiling the kernel with the toolchain composed of GCC >= 12.1.0 and
> > > binutils < 2.38, default ISA spec used when building binutils and GCC, the
> > > following build failure will appear because the
> > > CONFIG_TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI is not turned on.
> > > (i.e, gcc-12.1.0 and binutils-2.36, or gcc-12.3.0 and binutils-2.37, use
> > > default ISA spec.)
> > > 
> > >   CC      arch/riscv/kernel/vdso/vgettimeofday.o
> > >   <<BUILDDIR>>/arch/riscv/include/asm/vdso/gettimeofday.h: Assembler messages:
> > >   <<BUILDDIR>>/arch/riscv/include/asm/vdso/gettimeofday.h:79: Error: unrecognized opcode `csrr a5,0xc01'
> > 
> > The gift that keeps on giving :/
> > 
> > > Binutils has updated the default ISA spec version, and the community has
> > > responded well to this[1][2][3], but it appears that this is not over yet.

Also, I just noticed this comment. I disagree with the wording "well",
and more like "with weeping and gnashing of teeth" ;) This stuff is a
huge pain in the ass, and mixing toolchains between LLVM & GNU stuff (or
using an older binutils with a newer GCC) really makes it a lot worse.
Thanks for submitting a fix for this so that Nathan or I didn't have to!

> > > 
> > > We also need to consider the situation of binutils < 2.38 but
> > > GCC >= 12.1.0, since the combination between different versions of GCC and
> > > binutils is not unique, which is to some extent flexible. GCC release
> > > 12.1.0 updated the default ISA spec version in GCC commit[4].
> > 
> > I suspect this combination is not too common because binutils 2.38 came
> > out before GCC 12.1.0 but as you note, it is obviously possible. What
> > toolchain has this combination in the wild, which would be helpful for
> > documentation purposes?
> 
> Yeah, that'd be great to know, at least the other niche stuff that we
> are working around had a clear use-case (testing LLVM in debian containers)
> whereas there's no clear user for this.
> That's doubly interesting, as this patch seems to break things for binutils
> < 2.35, and if we have to make a trade-off between those too, then it'd
> be good to be able to weigh up the options.
> Do we perhaps need the misa-spec workaround instead for this case?
> Haven't tested that though, trying to dig myself out of email backlog.

I don't think the misa-spec stuff is what we need actually. Instead, the
workaround/fix that this patch implements just needs to be constrained to
versions of GAS greater than 2.35.

Thanks,
Conor.

> 
> > > For more information, please refer to:
> > > 
> > > commit 6df2a016c0c8 ("riscv: fix build with binutils 2.38")
> > > commit e89c2e815e76 ("riscv: Handle zicsr/zifencei issues between clang and binutils")
> > > 
> > > [1]: https://groups.google.com/a/groups.riscv.org/g/sw-dev/c/aE1ZeHHCYf4
> > > [2]: https://lore.kernel.org/all/20230308220842.1231003-1-conor@kernel.org
> > > [3]: https://lore.kernel.org/all/20230223220546.52879-1-conor@kernel.org
> > > [4]: https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=98416dbb0a62579d4a7a4a76bab51b5b52fec2cd
> 
> btw, please make these regular Link: tags (with a [N] at EOL) and drop
> the space between them and the sign off. Also, this probably needs to be
> CC: stable at vger.kernel.org too.

> > > 
> > > Signed-off-by: Mingzheng Xing <xingmingzheng at iscas.ac.cn>
> > > ---
> > >  arch/riscv/Kconfig | 6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> > > index 4c07b9189c86..b49cea30f6cc 100644
> > > --- a/arch/riscv/Kconfig
> > > +++ b/arch/riscv/Kconfig
> > > @@ -570,11 +570,15 @@ config TOOLCHAIN_HAS_ZIHINTPAUSE
> > >  config TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI
> > >  	def_bool y
> > >  	# https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=aed44286efa8ae8717a77d94b51ac3614e2ca6dc
> > > -	depends on AS_IS_GNU && AS_VERSION >= 23800
> > > +	# https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=98416dbb0a62579d4a7a4a76bab51b5b52fec2cd
> > > +	depends on CC_IS_GCC && GCC_VERSION >= 120100 || \
> > > +		   AS_IS_GNU && AS_VERSION >= 23800
> > 
> > GCC_VERSION will be 0 for clang, so you don't need the CC_IS_GCC check.
> > With that change, this should be able to stay on one line:
> > 
> >     depends on GCC_VERSION >= 120100 || (AS_IS_GNU && AS_VERSION >= 23800)
> > 
> > >  	help
> > >  	  Newer binutils versions default to ISA spec version 20191213 which
> > >  	  moves some instructions from the I extension to the Zicsr and Zifencei
> > >  	  extensions.
> > > +	  Similarly, GCC release 12.1.0 has changed the default ISA spec version to
> > > +	  20191213, so the above situation requires this option to be enabled.
> > >  
> > >  config TOOLCHAIN_NEEDS_OLD_ISA_SPEC
> > >  	def_bool y
> > > -- 
> > > 2.34.1
> > > 
> 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 228 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-riscv/attachments/20230725/3784559b/attachment.sig>