Re: ARM port(s) BoF at DebConf
On Fri, Jul 20, 2012 at 10:08:03PM +0200, Martin Guy wrote:
>On 19 July 2012 19:35, Steve McIntyre <steve@einval.com> wrote:
>> armel
>> =====
>>
>> First released with Lenny. Soft-float EABI, Software floating point
>> assumed by default. v4t which also runs smaller-size thumb instruction
>> set. Targeting old hardware like openmoko. Discussed (again!) moving
>> forwards from v4. Declared that v5 is no faster than v4t, but there
>> are doubts elsewhere in the community. Later discussion suggests
>> moving to v5te would be worth it. Some good benchmarks would help -
>> volunteers welcome!
>
>Actually, supporting less machines is a move backward, not forward.
>The speed advantage for standard apps on v5+ machines is less than 1%,
>i.e. negligable.
>
>Of course, I have a vested interest in continued armv4t support, since
>my company has an armv4t board on the market that ships with Debian as
>its standard distribution. It would also impact Technologic Systems,
>Bluewater Systems and other small companies for similar reasons.
Right. I've asked before who's still using v4t machines, and the
common response is "just openmoko". Thanks for mentioning others.
>Who is it that keeps bringing this up? I can see that ARM Ltd would
>want this, as it would eliminate Linux distro support for devices from
>which they no longer see any royalties., but I don't see any advantage
>for anyone else except chronic speed freaks who would kill other
>people's boards off to get a half of a percent faster for themselves.
We've been pestered several times by toolchain developers and
upstreams for various other projects that generate code for ARM
(e.g. JITs in browsers). It seems that Debian is about the only place
where anybody still cares about v4t any more, so we'll be the only
people seeing bugs related to broken (or even missing) v4 support
now.
>If somebody has a critical need to multiple two shorts with result as
>a long in a single instruction (which is what the E in 5TE brings),
>surely they can compile their own armel packages changing the cpu
>type, rather than making Debian do that and breaking other people's
>systems needlessly?
>
>Isn't Debian supposed to be the "Universal Operating System", where
>"Universal" includes running on as many different computers as
>possible?
Even Debian moves on, dropping support for older platforms from time
to time. We don't support actual i386 hardware any more, for example.
>And the speed freaks can always build their own v5t and v5te
>repositories and use those to install from, leaving everybody happy.
I've asked people for benchmarks to show the improvements (or lack of
them) that might come from a move to v5te. If there isn't a compelling
case for moving, we won't!
--
Steve McIntyre, Cambridge, UK. steve@einval.com
Welcome my son, welcome to the machine.
Reply to: