Re: [PATCH] Partially revert "compiler: enable CONFIG_OPTIMIZE_INLINING forcibly"

[Masahiro Yamada]

On Mon, Sep 30, 2019 at 8:45 PM Will Deacon <will@xxxxxxxxxx> wrote:
>
> This reverts commit ac7c3e4ff401b304489a031938dbeaab585bfe0a for ARM and
> arm64.
>
> Building an arm64 kernel with CONFIG_OPTIMIZE_INLINING=y has been shown
> to violate fixed register allocations of local variables passed to
> inline assembly with GCC prior to version 9 which can lead to subtle
> failures at runtime:
>
>   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91111
>
> A very similar has been reported for 32-bit ARM as well:
>
>   https://lkml.kernel.org/r/f5c221f5749e5768c9f0d909175a14910d349456.camel@xxxxxxx


For reviewers:
The main discussion is here:

https://lore.kernel.org/patchwork/patch/1122097/



> Although GCC 9.1 appears to work for the specific case in the bugzilla
> above, the exact issue has not been root-caused so play safe and disable
> the option for now on these architectures.
>
> Cc: Nicolas Saenz Julienne <nsaenzjulienne@xxxxxxx>
> Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> Cc: Russell King <linux@xxxxxxxxxxxxxxx>
> Cc: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>,
> Cc: Arnd Bergmann <arnd@xxxxxxxx>
> Signed-off-by: Will Deacon <will@xxxxxxxxxx>
> ---
>  lib/Kconfig.debug | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
> index 93d97f9b0157..c37c72adaeff 100644
> --- a/lib/Kconfig.debug
> +++ b/lib/Kconfig.debug
> @@ -312,6 +312,7 @@ config HEADERS_CHECK
>
>  config OPTIMIZE_INLINING
>         def_bool y
> +       depends on !(ARM || ARM64) # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91111


This is a too big hammer.

For ARM, it is not a compiler bug, so I am trying to fix the kernel code.

For ARM64, even if it is a compiler bug, you can add __always_inline
to the functions in question.
(arch_atomic64_dec_if_positive in this case).

You do not need to force __always_inline globally.




>         help
>           This option determines if the kernel forces gcc to inline the functions
>           developers have marked 'inline'. Doing so takes away freedom from gcc to
> --
> 2.23.0.444.g18eeb5a265-goog
>




--
Best Regards
Masahiro Yamada