Re: [PATCH v2] ARM: Add Cortex-A15 optimized NEON and VFP memcpy routines, with IFUNC.

[Will Newton <will.newton@linaro.org>]

On 18 April 2013 18:58, Siddhesh Poyarekar <siddhesh.poyarekar@gmail.com> wrote:

Hi Siddesh,

> On 18 April 2013 13:55, Siddhesh Poyarekar <siddhesh.poyarekar@gmail.com> wrote:
>> On 18 April 2013 13:24, Will Newton <will.newton@linaro.org> wrote:
>>> Thanks for the offer, I don't have any armv6 hardware to test with.
>>> The easiest way to test is probably using the cortex-strings package
>>> (as the patch I posted is for armv7).
>>>
>>> https://launchpad.net/cortex-strings
>>>
>>> "try-this -t memcpy" will benchmark the code I submitted, "try-glibc
>>> -t memcpy" will benchmark the current glibc code.
>>
>> OK, I'll try to do this tonight (i.e. ~8 hours from now) since the
>> board is at home and I'm not.  I assume you want 'current glibc code'
>> to be master?  The distribution version is 2.17 IIRC or at worst 2.16.
>
> `try-this` crashes with a SIGILL:
>
> Program received signal SIGILL, Illegal instruction.
> memset () at src/linaro-a9/memset.S:64
> 64              cbz     r2, 10f         @ Exit if 0 length
> (gdb) bt
>
> I don't know much arm assembly, but digging around a bit I found that
> cbz is implemented on armv6T2 or later and hence not available on my
> board.  If you're interested in making this work on armv6 then I can
> give this a shot - I've been meaning to start playing with arm
> assembly anyway.

Thanks for trying this.

All the routines in cortex-strings are intended for armv7 (i.e.
Cortex-A cores) and it looks like the memset code is being run as a
side effect of the test. It might be quite simple to work around or it
could involve lots of work! I guess you could always drop in a known
armv6 compatible memset e.g. the glibc one.

--
Will Newton
Toolchain Working Group, Linaro