[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Help making a decision on re-basing Slackware ARM from armv4 to armv5



On 14 December 2011 21:30, Stuart Winter <m-lists@biscuit.org.uk> wrote:
> Slackware, like Debian is a general purpose OS aiming to meet most users'
> needs on the common denominator hardware -- which on ARM is armv4.

Some instructions were added in v5E: count-leading-zeroes, 16-bit
multipliers, saturating arithmetic, a preload instruction, double word
loads and stores and a faster instruction to switch between ARM and
Thumb modes in function returns.
See Column 3 of
http://infocenter.arm.com/help/topic/com.arm.doc.qrc0001l/QRC0001_UAL.pdf
where is says "5E". It's unlikely that a compiler would be able to
make much use of any of these: CLZ is used in assembly-language
division routines, saturating arithmetic is unusable when compiling,
faster switching between ARM and Thumb is useless. The compiler may be
able to make use of the preloads to reduce the memory bottleneck and
of the double word load/stores.

> Will there be much of a performance difference for regular server-type
> workloads (e.g. file serving etc) as distinct from playing media and
> desktop usage with web browsers etc?

Probably not, since that kind of application will be mainly memory-,
network- and disk-bound.

> Also, will an OS compiled for armv5 perform better on armv7 hardware?

Probably the same size gain as you get on v5 hardware.

> My understanding is that only certain workloads -- e.g. playing video -- would
> benefit from compiling for armv5.

These will benefit more from the SIMD instructions, for which you can recompile
the specific libraries for your actual target processor.  A few
applications and libraries
will benefit from mandating a floating point processor such as VFP -
here the speedups
are a factor of three to four, and switching the hardfloat ABI gives
another 30% or so I gather.
However, that would also eliminate any board with no FPU.

> Thanks in advance for any ideas -- just rough instinctive feelings are all
> I'm after (unless somone does have any actual figures of running Debian
> armel userland compiled for armv4 vs armel compiled for armv5).

Of course, that is the real answer...  you could compare a Debian and
a Ubuntu userland, which are build from almost identical sourves using
the same mechanisms and compiler. I'd be very interested to know your
measured results.

My feeling is that the speed gain would be in the 5-10% range at best.
For a general-purpose operating system you need to evaluate whether
that's worth the effort and exclusion of low-end machines.

   M


Reply to: