Re: RFC: Add -mshared option to x86 ELF assembler

[Andy Lutomirski <luto at amacapital dot net>]

On Wed, May 13, 2015 at 5:59 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Wed, May 13, 2015 at 4:50 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>> On Tue, May 12, 2015 at 5:14 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>> On Fri, May 8, 2015 at 1:16 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>>> On Fri, May 8, 2015 at 5:09 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>> On Thu, May 7, 2015 at 8:22 PM, Andy Lutomirski <luto@amacapital.net> wrote:
>>>>>> On Thu, May 7, 2015 at 9:21 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>>>> On Thu, May 7, 2015 at 4:52 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>>>>>>> On 07.05.15 at 08:02, <luto@amacapital.net> wrote:
>>>>>>>>> AFAICT gas will produce relocations for jumps to global labels in the
>>>>>>>>> same file.  This doesn't seem directly harmful to me, except that, on
>>>>>>>>> x86, it forces five-byte jumps instead of two-byte jumps.
>>>>>>>>>
>>>>>>>>> This seems especially unfortunate, since even hidden and protected
>>>>>>>>> symbols have this problem.
>>>>>>>>>
>>>>>>>>> Given that many users don't want interposition support (especially the
>>>>>>>>> kernel and anyone using .hidden or .protected), it would be nice to
>>>>>>>>> have a command-line option to turn this off and probably also to turn
>>>>>>>>> it off by default for hidden and protected symbols.  Can gas do this?
>>>>>>>>
>>>>>>>> I've been running with the below changes (taken off of a bigger set
>>>>>>>> of changes, so the line numbers may look a little odd) for the last
>>>>>>>> couple of years. I never tried to submit this change because so far
>>>>>>>> I couldn't find the time to check whether this would have any
>>>>>>>> unwanted side effects on cases I don't normally use.
>>>>>>>>
>>>>>>>
>>>>>>> This is the patch I checked in.
>>>>>>>
>>>>>>> Thanks.
>>>>>>>
>>>>>>> --
>>>>>>> H.J.
>>>>>>> ---
>>>>>>> Branches to global non-weak symbols defined in the same segment with
>>>>>>> non-default visibility can be optimized the same way as branches to
>>>>>>> local symbols.
>>>>>>
>>>>>> Would it make sense to also add a command line option along the lines
>>>>>> of gcc's -fno-semantic-interposition or some way to override the
>>>>>> default visibility?  AFAICS this patch helps but only if asm code gets
>>>>>> liberally sprinkled with .hidden or .protected directives.
>>>>>>
>>>>>
>>>>> This is what I checked in.  With
>>>>>
>>>>> diff --git a/arch/x86/Makefile b/arch/x86/Makefile
>>>>> index 2fda005..186e6f7 100644
>>>>> --- a/arch/x86/Makefile
>>>>> +++ b/arch/x86/Makefile
>>>>> @@ -107,6 +107,10 @@ else
>>>>>          KBUILD_CFLAGS += $(call cc-option,-maccumulate-outgoing-args)
>>>>>  endif
>>>>>
>>>>> +NO_SHARED_CFLAGS = $(call as-option,-Wa$(comma)-mno-shared)
>>>>> +KBUILD_CFLAGS += $(NO_SHARED_CFLAGS)
>>>>> +KBUILD_AFLAGS += $(NO_SHARED_CFLAGS)
>>>>> +
>>>>>  # Make sure compiler does not have buggy stack-protector support.
>>>>>  ifdef CONFIG_CC_STACKPROTECTOR
>>>>>    cc_has_sp := $(srctree)/scripts/gcc-x86_$(BITS)-has-stack-protector.sh
>>>>>
>>>>> On kernel master branch, I got
>>>>>
>>>>>    text   data    bss    dec    hex filename
>>>>> 10934167 2275232 1609728 14819127 e21f37 vmlinux.old
>>>>> 10934119 2275232 1609728 14819079 e21f07 vmlinux
>>>>>
>>>>> It saves 48 bytes.
>>>>
>>>> This is before I fixed:
>>>>
>>>> /* This is global to keep gas from relaxing the jumps */
>>>> ENTRY(early_idt_handler)
>>>>         cld
>>>>
>>>> in arch/x86/kernel/head_64.S.  With -mno-shared, we must
>>>> make early_idt_handler weak to keep gas from relaxing the jumps.
>>>>
>>>
>>> Here is a patch to change the assembler default to optimize out
>>> relocations to defined non-weak global branch targets with default
>>> visibility.  It will generate slightly smaller object files.  But Linux
>>> kernel will be broken unless early_idt_handler is marked weak.
>>> I am little uncomfortable with -mshare and I don't like -mno-shared
>>> very much either.  I may just simply remove -mno-shared.
>>>
>>
>> I reverted the -mno-shared change.
>>
>
> Here is a patch to add -mshared, which is off by default.  On Linux kernel
> with this change:
>
> diff --git a/arch/x86/kernel/head_64.S b/arch/x86/kernel/head_64.S
> index a468c0a..9a10e05 100644
> --- a/arch/x86/kernel/head_64.S
> +++ b/arch/x86/kernel/head_64.S
> @@ -339,8 +339,8 @@ early_idt_handlers:
>   i = i + 1
>   .endr
>
> -/* This is global to keep gas from relaxing the jumps */
> -ENTRY(early_idt_handler)
> +/* This is weak to keep gas from relaxing the jumps */
> +WEAK(early_idt_handler)
>   cld
>
>   cmpl $2,(%rsp) # X86_TRAP_NMI
> --
> 2.1.0
>
> I got
>
> [hjl@gnu-tools-1 kernel.org]$ readelf -r old/vmlinux.o | head -5
>
> Relocation section '.rela.text' at offset 0xafea2f0 contains 205717 entries:
>   Offset          Info           Type           Sym. Value    Sym. Name + Addend
> 000000000001  1253100000002 R_X86_64_PC32     0000000000001e70 __fentry__ - 4
> 000000000009  1c8c00000002 R_X86_64_PC32     0000000000000000 .data + 51bc
> [hjl@gnu-tools-1 kernel.org]$ readelf -r new/vmlinux.o | head -5
>
> Relocation section '.rela.text' at offset 0xafea280 contains 205711 entries:
>   Offset          Info           Type           Sym. Value    Sym. Name + Addend
> 000000000001  1253100000002 R_X86_64_PC32     0000000000001e70 __fentry__ - 4
> 000000000009  1c8c00000002 R_X86_64_PC32     0000000000000000 .data + 51bc
> [hjl@gnu-tools-1 kernel.org]$
>
> It removes 6 relocations.  On gcc master branch,
>
> [hjl@gnu-tools-1 gcc-misc]$ size build-x86_64-linux*/gcc/cc1
>    text   data    bss    dec    hex filename
> 21529621  62256 1348312 22940189 15e0a1d build-x86_64-linux.branch/gcc/cc1
> 21529749  62256 1348312 22940317 15e0a9d build-x86_64-linux/gcc/cc1
> [hjl@gnu-tools-1 gcc-misc]$ size build-x86_64-linux*/gcc/cc1plus
>    text   data    bss    dec    hex filename
> 23713509  62400 1372760 25148669 17fbcfd build-x86_64-linux.branch/gcc/cc1plus
> 23713669  62400 1372760 25148829 17fbd9d build-x86_64-linux/gcc/cc1plus
> [hjl@gnu-tools-1 gcc-misc]$
>
> It is more effective.  I will run more tests.

This seems like a sensible idea, but I can imagine it breaking some
weird use cases (like that one Linux thing).  Is that okay?

--Andy