Re: GUB on kainhofer: still cross/gcc

[Anthony W. Youngman]

In message <address@hidden>, Jan Nieuwenhuizen 
<address@hidden> writes
> Op donderdag 04-06-2009 om 10:46 uur [tijdzone +0100], schreef Anthony
> W. Youngman:
>
> > >I think that's a pretty usual setup (most people I know have a 32bit version
> > >of Linux installed on their laptop even though their CPU is actually 
> > >64bit).
>
> Sometimes it makes sense to "do what most people do", esp. if you do it
> as a deliberate choice :-)

Just as long as you're aware of the consequences ... "what most people 
do" is usually a pretty stupid thing to do. Following the herd is fine 
if you don't want to stand out, but if you want to make your mark it's 
not an option.
> > Note also, that running 32-bit on a 64-bit system can OFTEN be a
> > performance WIN, so you DON'T want to upgrade "just because you can".
>
> I call BS.  Ref please?

Are you saying that 64-bit code is *inherently* more efficient than 
32-bit on a 64-bit system? As I understand it, the CPU doesn't give a 
monkeys whether the code is 32 or 64 bit.

But. Compile *the* *same* program, in both 32 and 64 bit mode. The 
64-bit executable will be larger, maybe by quite a lot! This can easily 
hurt the program, as the program becomes too large to fit in cache, 
loops become too large to fit in the CPU cache, the OS scheduling 
mechanism penalises large programs over small, etc etc.

As I understand it, the rule is "if you don't need 64-bit mode, then 
your executable will be smaller in 32-bit mode, therefore it won't run 
slower but could well run faster precisely because it is smaller". If 
you *DO* need 64-bit mode, of course running in 32-bit mode will be far 
slower because the cost of emulating 64-bit in 32-bit is expensive.

I can't give an actual reference, I'm afraid, but - I think it was on 
the database mailing list I'm on - they were saying "benchmark your app" 
and I think they had several such benchmarks that said small systems ran 
better in 32-bit mode, even on a 64-bit processor.
> Greetings,
> Jan.

Cheers,
Wol
--
Anthony W. Youngman - address@hidden