Joseph Kowalski wrote:
> Joerg Schilling wrote:
>> On amd64, all 64 bit binaries typically (with a few exceptions) run ~
>> 30% faster than their 32 bit couterparts.
>>
>> Shouldn't we take take care of this difference? Is there no way to
>> treat amd64
>> different from sparcv9?
>>
>
> Ah, the topic everyone seems to avoid talking about.
>
> There is a small set of "correction" issues on SPARC. There are
> *some* applications which are faster as 32-bit. There are *some*
> applications which are faster as 64-bits. All kind of worth a huge yawn.
>
> There is a significant performance issue on amd64 [1]. Its all about
> the registers. Unfortunately, since we still attempt to produce one
> "IA" distribution *and* support 32-bit IA processors, we are stuck at
> the lowest common denominator.
>
> Every Linix'ish distro I know of has both 64 and 32 bit variants. Sun
> has not shown a significant interest in producing a 64-bit, IA
> distribution. OpenSolaris certainly could.
>
> That said, its not this case, right?
I don't know if its this case or not. I derailed it, and case materials
need to be supplied. The idea that we could "fix" some of the 32-bit
performance problems, or offer an optional 64-bit only distro, certainly
has a certain amount of appeal.
There's a whole other issue, which is that for a lot of interfaces,
thunking between 64-bit and 32-bit structures when crossing the
userland/kernel boundary is somewhat expensive. But, due to
compatibility guarantees, I don't think we can eliminate the code in the
kernel to support it-- but we can reduce the amount that it gets
*used*. (There may be some private ioctls which could be entirely
cleaned up/made 64-bit only, though. But only for drivers that don't
require 32-bit userland support -- I'm not sure there are very many of
those.)
-- Garrett
>
> - jek3
>
> [1] Replace "amd64" with your choice of: i386, i486, i586, i686,
> x86, ia86, ia64, x64, amd64, IA64,.... :-)
