[PATCH] Add Prefer_MAP_32BIT_EXEC for Silvermont

[H.J. Lu]

On Mon, Dec 14, 2015 at 7:15 AM, Florian Weimer <fweimer@redhat.com> wrote:
> On 12/12/2015 06:36 PM, H.J. Lu wrote:
>
>> Here is the updated patch to make it opt-in.  OK for master?
>
> I would still like to see performance numbers for the PIE and vDSO
> cases, as requested here:

This optimization has no performance impact on PIE nor
vDSO.

>   <https://sourceware.org/ml/libc-alpha/2015-12/>
>
> As far as I can tell, the application impact of this setting has not
> been investigated, either:
>
>   <https://sourceware.org/ml/libc-alpha/2015-12/msg00261.html>
>
> And it probably makes sense to guard this by a configure switch because
> for almost all users, it's just bloat.

Adding a configure switch defeats LD_ENABLE_PREFER_MAP_32BIT_EXEC.

old:

[hjl@gnu-6 build-x86_64-linux]$ size elf/dl-sysdep.os
   text   data    bss    dec    hex filename
   3976     20      4   4000    fa0 elf/dl-sysdep.os

and new:

[hjl@gnu-6 build-x86_64-linux]$ size elf/dl-sysdep.os
   text   data    bss    dec    hex filename
   4201     20      4   4225   1081 elf/dl-sysdep.os

It increases elf/dl-sysdep.os by 225 bytes.  Calling it "bloat" is
a stretch.

> I'm somewhat sympathetic to Zack's request to remove the special-casing
> on CPU settings.  It would certainly simplify application compatibility
> testing without access to the defective silicon.
>
> Florian
>



-- 
H.J.