Very Computer

I'm wondering if anyone can provide me a quick tip on how to do dead function
elimination optimizations with gcc-2.95.2, without resorting to
-finline-functions?

The compiler can know that a function is unused if it is static, and there are
no references to it in the same file. Using -finline-functions (with the
default -fno-keep-inline-functions) will eliminate such functions as a side
effect of attempting to inline them, provided their complexity is below the
inline-limit. However, I don't necessarily want to force all functions below
the inline-limit to be inlined; I just want static functions that don't get
called to be eliminated.

Here's why: I'm attempting to cross-compile the Linux kernel for an embedded
system which has no need for the /proc filesystem. Hence, CONFIG_PROC_FS is
undefined. The kernel code contains files with constructs like:

blah.c:
    static proc_blah_blah()
    {
        /* function dealing with /proc stuff */
    }

    #ifdef CONFIG_PROC_FS
    struct with reference to proc_blah_blah
    #endif

With CONFIG_PROC_FS undefined, I'd like to optimise away the function
proc_blah_blah automatically, without having to spray more
#ifdef CONFIG_PROC_FS's around all such functions. The compiler already has
enough information to eliminate them, but how do I get it to do so without
also asking it to try to inline _everything_?

Any suggestions?

Thanks,
Graham

Quote:> Hi there,

> I'm wondering if anyone can provide me a quick tip on how to do dead
function
> elimination optimizations with gcc-2.95.2, without resorting to
> -finline-functions?

> The compiler can know that a function is unused if it is static, and there
are
> no references to it in the same file. Using -finline-functions (with the
> default -fno-keep-inline-functions) will eliminate such functions as a
side
> effect of attempting to inline them, provided their complexity is below
the
> inline-limit. However, I don't necessarily want to force all functions
below
> the inline-limit to be inlined; I just want static functions that don't
get
> called to be eliminated.

> Here's why: I'm attempting to cross-compile the Linux kernel for an
embedded
> system which has no need for the /proc filesystem. Hence, CONFIG_PROC_FS
is
> undefined. The kernel code contains files with constructs like:

> blah.c:
>     static proc_blah_blah()
>     {
> /* function dealing with /proc stuff */
>     }

>     #ifdef CONFIG_PROC_FS
>     struct with reference to proc_blah_blah
>     #endif

> With CONFIG_PROC_FS undefined, I'd like to optimise away the function
> proc_blah_blah automatically, without having to spray more
> #ifdef CONFIG_PROC_FS's around all such functions. The compiler already
has
> enough information to eliminate them, but how do I get it to do so without
> also asking it to try to inline _everything_?

> Any suggestions?

> Thanks,
> Graham

Any other suggestions?

Thanks,
Graham

  That's all.

  Hope that helps,
  Etienne.

Quote:>  Compile with GCC option -ffunction-sections, which put every functions
> in their own section, rewrite the linker description file to put all
> sections ".text.<functionname>" in the ".text" segment, and more important
> "ask to binutil" / "send them a patch" to accept the flag --gc-section
> for the i386 target and process it as it is documented (ld manual).

Regards,
Graham

Quote:> Compile with GCC option -ffunction-sections, which put every functions
> in their own section, rewrite the linker description file to put all
> sections ".text.<functionname>" in the ".text" segment, and more important
> "ask to binutil" / "send them a patch" to accept the flag --gc-section
> for the i386 target and process it as it is documented (ld manual).

    *(.text.*)

However, it won't link without --gc-section in LINKFLAGS, and with --gc-section
in LINKFLAGS, it generates a tiny vmlinux file which nm won't even recognise.

Any more suggestions?

Thanks,
Graham