Re: [PATCH] ARM: Emit __gnu_mcount_nc when using Clang 10.0.0 or newer
From: Nick Desaulniers
Date: Thu Aug 29 2019 - 16:57:50 EST
On Thu, Aug 29, 2019 at 1:21 PM Stefan Agner <stefan@xxxxxxxx> wrote:
> On 2019-08-29 21:34, Nathan Chancellor wrote:
> > On Thu, Aug 29, 2019 at 10:55:28AM -0700, Nick Desaulniers wrote:
> >> On Wed, Aug 28, 2019 at 11:27 PM Nathan Chancellor
> >> <natechancellor@xxxxxxxxx> wrote:
> >> >
> >> > Currently, multi_v7_defconfig + CONFIG_FUNCTION_TRACER fails to build
> >> > with clang:
> >> >
> >> > arm-linux-gnueabi-ld: kernel/softirq.o: in function `_local_bh_enable':
> >> > softirq.c:(.text+0x504): undefined reference to `mcount'
> >> > arm-linux-gnueabi-ld: kernel/softirq.o: in function `__local_bh_enable_ip':
> >> > softirq.c:(.text+0x58c): undefined reference to `mcount'
> >> > arm-linux-gnueabi-ld: kernel/softirq.o: in function `do_softirq':
> >> > softirq.c:(.text+0x6c8): undefined reference to `mcount'
> >> > arm-linux-gnueabi-ld: kernel/softirq.o: in function `irq_enter':
> >> > softirq.c:(.text+0x75c): undefined reference to `mcount'
> >> > arm-linux-gnueabi-ld: kernel/softirq.o: in function `irq_exit':
> >> > softirq.c:(.text+0x840): undefined reference to `mcount'
> >> > arm-linux-gnueabi-ld: kernel/softirq.o:softirq.c:(.text+0xa50): more undefined references to `mcount' follow
> >> >
> >> > clang can emit a working mcount symbol, __gnu_mcount_nc, when
> >> > '-meabi gnu' is passed to it. Until r369147 in LLVM, this was
> >> > broken and caused the kernel not to boot because the calling
> >> > convention was not correct. Now that it is fixed, add this to
> >> > the command line when clang is 10.0.0 or newer so everything
> >> > works properly.
> >> >
> >> > Link: https://github.com/ClangBuiltLinux/linux/issues/35
> >> > Link: https://bugs.llvm.org/show_bug.cgi?id=33845
> >> > Link: https://github.com/llvm/llvm-project/commit/16fa8b09702378bacfa3d07081afe6b353b99e60
> >> > Signed-off-by: Nathan Chancellor <natechancellor@xxxxxxxxx>
> >> > ---
> >> > arch/arm/Makefile | 6 ++++++
> >> > 1 file changed, 6 insertions(+)
> >> >
> >> > diff --git a/arch/arm/Makefile b/arch/arm/Makefile
> >> > index c3624ca6c0bc..7b5a26a866fc 100644
> >> > --- a/arch/arm/Makefile
> >> > +++ b/arch/arm/Makefile
> >> > @@ -112,6 +112,12 @@ ifeq ($(CONFIG_ARM_UNWIND),y)
> >> > CFLAGS_ABI +=-funwind-tables
> >> > endif
> >> >
> >> > +ifeq ($(CONFIG_CC_IS_CLANG),y)
> >> > +ifeq ($(shell test $(CONFIG_CLANG_VERSION) -ge 100000; echo $$?),0)
> >> > +CFLAGS_ABI +=-meabi gnu
> >> > +endif
> >> > +endif
> >> > +
> >> Thanks for the patch! I think this is one of the final issues w/ 32b
> >> ARM configs when building w/ Clang.
> >> I'm not super enthused about the version check. The flag is indeed
> >> not recognized by GCC, but I think it would actually be more concise
> >> with $(cc-option) and no compiler or version check.
> >> Further, I think that the working __gnu_mcount_nc in Clang would
> >> better be represented as marking the arch/arm/KConfig option for
> >> CONFIG_FUNCTION_TRACER for dependent on a version of Clang greater
> >> than or equal to Clang 10, not conditionally adding this flag. (We
> >> should always add the flag when supported, IMO. __gnu_mcount_nc's
> >> calling convention being broken is orthogonal to the choice of
> >> __gnu_mcount_nc vs mcount, and it's the former's that should be
> >> checked, not the latter as in this patch.
> > I will test with or without CONFIG_AEABI like Matthias asked and I will
> > implement your Kconfig suggestion if it passes all of my tests. The
> > reason that I did it this way is because I didn't want a user to end up
> > with a non-booting kernel since -meabi gnu works with older versions of
> > clang at build time, the issue happens at boot time but the Kconfig
> > suggestion + cc-option should fix that.
> I agree with Nathan here, I'd rather prefer the build system to fail
> building rather than runtime error.
> If we decide we want to have it building despite it not building a
> functional kernel, we should at least add a #warning...
Just to be clear...I was suggesting a build failure, but for
__gnu_mcount_nc not having the correct calling convention in older
clang releases, which is orthogonal to passing `-meabi gnu`. This
patch uses the __gnu_mcount_nc calling convention problem to justify a
version check for a flag that while closely related, is not actually
the problem, IMO.