Re: [PATCH] arm64: Delete the space separator in __emit_inst

[Nick Desaulniers]

On Sun, Apr 12, 2020 at 8:38 PM 'Fangrui Song' via Clang Built Linux
<clang-built-linux@xxxxxxxxxxxxxxxx> wrote:
>
> Many instances of __emit_inst(x) expand to a directive. In a few places
> it is used as a macro argument, e.g.
>
>   arch/arm64/include/asm/sysreg.h
>   #define __emit_inst(x)                       .inst (x)
>
>   arch/arm64/include/asm/sysreg.h
>   #define SET_PSTATE_PAN(x)            __emit_inst(0xd500401f | PSTATE_PAN | ((!!x) << PSTATE_Imm_shift))
>
>   arch/arm64/kvm/hyp/entry.S
>   ALTERNATIVE(nop, SET_PSTATE_PAN(1), ARM64_HAS_PAN, CONFIG_ARM64_PAN)
>
> Clang integrated assembler parses `.inst (x)` as two arguments passing
> to a macro. We delete the space separator so that `.inst(x)` will be
> parsed as one argument.
>
> Note, GNU as parsing `.inst (x)` as one argument is unintentional (for
> example the x86 backend will parse the construct as two arguments).
> See https://sourceware.org/bugzilla/show_bug.cgi?id=25750#c10

Thanks for the patch and for leading the discussion with the binutils
developers on this curious parsing case.

>
> Link: https://github.com/ClangBuiltLinux/linux/issues/939
> Cc: clang-built-linux@xxxxxxxxxxxxxxxx
> Signed-off-by: Fangrui Song <maskray@xxxxxxxxxx>

Shouldn't this have:
Suggested-by: Ilie Halip <ilie.halip@xxxxxxxxx>
Since Ilie sugguested this in
https://github.com/ClangBuiltLinux/linux/issues/939#issuecomment-601776123?

> ---
>  arch/arm64/include/asm/sysreg.h | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
> index ebc622432831..af21e2ec5e3e 100644
> --- a/arch/arm64/include/asm/sysreg.h
> +++ b/arch/arm64/include/asm/sysreg.h
> @@ -49,7 +49,9 @@
>  #ifndef CONFIG_BROKEN_GAS_INST
>
>  #ifdef __ASSEMBLY__
> -#define __emit_inst(x)                 .inst (x)
> +// The space separator is omitted so that __emit_inst(x) can be parsed as
> +// either a directive or a macro argument.
> +#define __emit_inst(x)                 .inst(x)
>  #else
>  #define __emit_inst(x)                 ".inst " __stringify((x)) "\n\t"

What happens if someone starts using `__emit_inst` from inline
assembly, and passes that subexpression to a macro?  There are no
users today in arch/arm64/, but I think it's better to change both.

With this applied on -next, and testing via:
$ ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- make LLVM=1 LLVM_IAS=1
-j71 defconfig
$ qemu-system-aarch64 -kernel arch/arm64/boot/Image.gz -machine virt
-cpu cortex-a57 -nographic --append "console=ttyAMA0" -m 2048 -initrd
<path to my buildroot.cpio>

I was able to build and boot, modulo:
https://github.com/ClangBuiltLinux/linux/issues/988 (new)
https://github.com/ClangBuiltLinux/linux/issues/716
https://github.com/ClangBuiltLinux/linux/issues/510 (more specific to
LLD than AS)
so we're pretty close to being able to assemble an arm64 defconfig with Clang.
-- 
Thanks,
~Nick Desaulniers