Re: [PATCH] um: clang: Strip out -mno-global-merge from USER_CFLAGS

[Nathan Chancellor <nathan@kernel.org>]

On Thu, Mar 03, 2022 at 10:04:28AM -0800, Kees Cook wrote:
> On Thu, Mar 03, 2022 at 10:30:47AM -0700, Nathan Chancellor wrote:
> > Hi David,
> > 
> > On Thu, Mar 03, 2022 at 05:06:42PM +0800, David Gow wrote:
> > > The things built with USER_CFLAGS don't seem to recognise it as a
> > > compiler option, and print a warning:
> > > clang: warning: argument unused during compilation: '-mno-global-merge' [-Wunused-command-line-argument]
> > > 
> > > Fixes: 744814d2fa ("um: Allow builds with Clang")
> > > Signed-off-by: David Gow <davidgow@google.com>
> > > ---
> > > 
> > > This warning shows up after merging:
> > > https://lore.kernel.org/lkml/20220227184517.504931-6-keescook@chromium.org/
> > > 
> > > I'm not 100% sure why this is necessary, but it does seem to work. All
> > > the attempts to get rid of -mno-global-merge entirely have been met with
> > > skepticism, but I'm guessing that it's not a problem for just the UML
> > > "user" files, as they shouldn't(?) interact too much with modules.
> > 
> > Thank you for the patch! I think it is correct, as this flag only works
> > for AArch64 and ARM, as it is only used in Clang::AddAArch64TargetArgs()
> > and Clang::AddARMTargetArgs() in clang/lib/Driver/ToolChains/Clang.cpp,
> > which are obviously never called with UML. I am not sure why we do not
> > see warning during regular kernel builds, maybe something about how UML
> > objects are compiled exposes this?
> 
> Yeah, I got really turned around when I tried to find this last week.
> 
> > 
> > Regardless, I would definitely like to clean up this instance of the
> > warning because I would like to make this warning a hard error so that
> > we do not get cryptic cc-option failures:
> > 
> > https://github.com/ClangBuiltLinux/linux/issues/1587
> > 
> > Reviewed-by: Nathan Chancellor <nathan@kernel.org>
> > 
> > One small comment below.
> > 
> > >  arch/um/Makefile | 4 ++++
> > >  1 file changed, 4 insertions(+)
> > > 
> > > diff --git a/arch/um/Makefile b/arch/um/Makefile
> > > index f2fe63bfd819..320b09cd513c 100644
> > > --- a/arch/um/Makefile
> > > +++ b/arch/um/Makefile
> > > @@ -75,6 +75,10 @@ USER_CFLAGS = $(patsubst $(KERNEL_DEFINES),,$(patsubst -I%,,$(KBUILD_CFLAGS))) \
> > >  		-D_FILE_OFFSET_BITS=64 -idirafter $(srctree)/include \
> > >  		-idirafter $(objtree)/include -D__KERNEL__ -D__UM_HOST__
> > >  
> > > +ifdef CONFIG_CC_IS_CLANG
> > 
> > Is this ifdef needed?
> > 
> > > +USER_CFLAGS := $(patsubst -mno-global-merge,,$(USER_CFLAGS))
> > > +endif
> 
> How does -mno-global-merge get KBUILD_CFLAGS in the first place? If it's
> arm/arm64 only, shouldn't that get relocated to those architectures?
> 
> *time travel* found it:
> 
> 61163efae020 ("kbuild: LLVMLinux: Add Kbuild support for building kernel with Clang")
> 
> So I think this may have been universally true long ago, and now only
> arm/arm64 need it?

I don't think so. -mno-global-merge was introduced in [1], which only
added the option handling in Clang::AddARMTargetArgs(). In fact, when
-mglobal-merge was added, there was a test added that checked that
-mglobal-merge/-mno-global-merge had no impact with an x86_64 triple [2].

Perhaps at the time that 61163efae020 was written, they were testing
against either ARCH=arm or ARCH=arm64, which would have shown the
problem that they mentioned? There is no context in the commit message
so *shrug*.

Regardless, I think your patch is equally viable, perhaps even better,
as I am sure this is not the only area where this warning might show up.

[1]: https://github.com/llvm/llvm-project/commit/ba3df1d3ca8569ff8f6e289594c53926e71e519a
[2]: https://github.com/llvm/llvm-project/commit/256a869d312b37ed7b5d131329050fcdadb11148

Cheers,
Nathan

> diff --git a/Makefile b/Makefile
> index 2f6fbba88a0e..fbc42c3c389e 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -852,10 +852,6 @@ ifdef CONFIG_CC_IS_CLANG
>  KBUILD_CPPFLAGS += -Qunused-arguments
>  # The kernel builds with '-std=gnu89' so use of GNU extensions is acceptable.
>  KBUILD_CFLAGS += -Wno-gnu
> -# CLANG uses a _MergedGlobals as optimization, but this breaks modpost, as the
> -# source of a reference will be _MergedGlobals and not on of the whitelisted names.
> -# See modpost pattern 2
> -KBUILD_CFLAGS += -mno-global-merge
>  else
>  
>  # gcc inanely warns about local variables called 'main'
> diff --git a/arch/arm/Makefile b/arch/arm/Makefile
> index a2391b8de5a5..dcab28c44c26 100644
> --- a/arch/arm/Makefile
> +++ b/arch/arm/Makefile
> @@ -48,6 +48,13 @@ CHECKFLAGS	+= -D__ARMEL__
>  KBUILD_LDFLAGS	+= -EL
>  endif
>  
> +ifdef CONFIG_CC_IS_CLANG
> +# CLANG uses a _MergedGlobals as optimization, but this breaks modpost, as the
> +# source of a reference will be _MergedGlobals and not on of the whitelisted names.
> +# See modpost pattern 2
> +KBUILD_CFLAGS	+= -mno-global-merge
> +endif
> +
>  #
>  # The Scalar Replacement of Aggregates (SRA) optimization pass in GCC 4.9 and
>  # later may result in code being generated that handles signed short and signed
> diff --git a/arch/arm64/Makefile b/arch/arm64/Makefile
> index 2f1de88651e6..2d20fb2ea664 100644
> --- a/arch/arm64/Makefile
> +++ b/arch/arm64/Makefile
> @@ -36,6 +36,13 @@ ifeq ($(CONFIG_BROKEN_GAS_INST),y)
>  $(warning Detected assembler with broken .inst; disassembly will be unreliable)
>  endif
>  
> +ifdef CONFIG_CC_IS_CLANG
> +# CLANG uses a _MergedGlobals as optimization, but this breaks modpost, as the
> +# source of a reference will be _MergedGlobals and not on of the whitelisted names.
> +# See modpost pattern 2
> +KBUILD_CFLAGS	+= -mno-global-merge
> +endif
> +
>  KBUILD_CFLAGS	+= -mgeneral-regs-only	\
>  		   $(compat_vdso) $(cc_has_k_constraint)
>  KBUILD_CFLAGS	+= $(call cc-disable-warning, psabi)
> 
> > > +
> > >  #This will adjust *FLAGS accordingly to the platform.
> > >  include $(srctree)/$(ARCH_DIR)/Makefile-os-$(OS)
> > >  
> > > -- 
> > > 2.35.1.616.g0bdcbb4464-goog
> > > 
> 
> -- 
> Kees Cook