Re: [PATCH] asm-generic: Force inlining of get_order() to work around gcc10 poor decision

From: Joel Stanley <joel@jms.id.au>
To: Christophe Leroy <christophe.leroy@csgroup.eu>,
	Segher Boessenkool <segher@kernel.crashing.org>
Cc: Arnd Bergmann <arnd@arndb.de>,
	Masahiro Yamada <yamada.masahiro@socionext.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	linux-arch@vger.kernel.org,
	linuxppc-dev <linuxppc-dev@lists.ozlabs.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] asm-generic: Force inlining of get_order() to work around gcc10 poor decision
Date: Mon, 19 Oct 2020 04:55:55 +0000	[thread overview]
Message-ID: <CACPK8XfgK0Bj3sLjKCi80jS6vK34FN5BTkU8FvBGcMR=RQn4Xw@mail.gmail.com> (raw)
In-Reply-To: <96c6172d619c51acc5c1c4884b80785c59af4102.1602949927.git.christophe.leroy@csgroup.eu>

On Sat, 17 Oct 2020 at 15:55, Christophe Leroy
<christophe.leroy@csgroup.eu> wrote:
>
> When building mpc885_ads_defconfig with gcc 10.1,
> the function get_order() appears 50 times in vmlinux:
>
> [linux]# ppc-linux-objdump -x vmlinux | grep get_order | wc -l
> 50
>
> [linux]# size vmlinux
>    text    data     bss     dec     hex filename
> 3842620  675624  135160 4653404  47015c vmlinux
>
> In the old days, marking a function 'static inline' was forcing
> GCC to inline, but since commit ac7c3e4ff401 ("compiler: enable
> CONFIG_OPTIMIZE_INLINING forcibly") GCC may decide to not inline
> a function.
>
> It looks like GCC 10 is taking poor decisions on this.
>
> get_order() compiles into the following tiny function,
> occupying 20 bytes of text.
>
> 0000007c <get_order>:
>   7c:   38 63 ff ff     addi    r3,r3,-1
>   80:   54 63 a3 3e     rlwinm  r3,r3,20,12,31
>   84:   7c 63 00 34     cntlzw  r3,r3
>   88:   20 63 00 20     subfic  r3,r3,32
>   8c:   4e 80 00 20     blr
>
> By forcing get_order() to be __always_inline, the size of text is
> reduced by 1940 bytes, that is almost twice the space occupied by
> 50 times get_order()
>
> [linux-powerpc]# size vmlinux
>    text    data     bss     dec     hex filename
> 3840680  675588  135176 4651444  46f9b4 vmlinux

I see similar results with GCC 10.2 building for arm32. There are 143
instances of get_order with aspeed_g5_defconfig.

Before:
 9071838 2630138  186468 11888444         b5673c vmlinux
After:
 9069886 2630126  186468 11886480         b55f90 vmlinux

1952 bytes smaller with your patch applied. Did you raise this with
anyone from GCC?

Reviewed-by: Joel Stanley <joel@jms.id.au>

> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
> ---
>  include/asm-generic/getorder.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/include/asm-generic/getorder.h b/include/asm-generic/getorder.h
> index e9f20b813a69..f2979e3a96b6 100644
> --- a/include/asm-generic/getorder.h
> +++ b/include/asm-generic/getorder.h
> @@ -26,7 +26,7 @@
>   *
>   * The result is undefined if the size is 0.
>   */
> -static inline __attribute_const__ int get_order(unsigned long size)
> +static __always_inline __attribute_const__ int get_order(unsigned long size)
>  {
>         if (__builtin_constant_p(size)) {
>                 if (!size)
> --
> 2.25.0
>