Re: Linux 3.1-rc9

[Ingo Molnar <mingo@elte.hu>]

* H. Peter Anvin <hpa@zytor.com> wrote:

> On 10/17/2011 11:49 AM, Ingo Molnar wrote:
> > Do we have some hard data on this, which we could put into comments 
> > in include/linux/ktime.h and such? Older versions of GCC used to do a 
> > bad job of long long handling on 32-bit systems - that might be a 
> > factor in the performance figures.
> > 
> > But i suspect you are right that the cost is still very much there 
> 
> 64/64 division is done bit by bit on most (all?) 32-bit architectures.
> 
> 64/32 division can be done in hardware on some architectures, e.g. x86.

it's 64/32 division - it's the /1000000000 /1000000 /1000 divisions 
in the large majority of cases, to convert between 
seconds/milliseconds/microseconds and scalar nanoseconds.

the kernel-internal ktime_t in the 32-bit optimized case is:

union ktime {
        s32     sec, nsec;
};

which is the same as timespec and arithmetically close to timeval, 
which many ABIs use. So conversion is easy in that case - but 
arithmetics gets a bit harder.

If we used a scalar 64-bit form for all kernel internal time 
representations:

	s64	nsecs;

then conversions back to timespec/timeval would involve dividing this 
64-bit value with 1000000000 or 1000000.

Is there no faster approximation for those than bit by bit?

In particular we could try something like:

	(high*2^32 + low)/1e9 ~==  ( high * (2^64/1e9) ) / 2^32

... which reduces it all to a 64-bit multiplication (or two 32-bit 
multiplications) with a known constant, at the cost of 1 nsec 
imprecision of the result - but that's an OK approximation in my 
opinion.

But it's late here and math is hard - lets go shopping ;-)

Thanks,

	Ingo