On Wednesday 02 July 2003 05:36, Peter Chubb wrote:
> Bernardo> If there are architectures where gcc doesn't implement
> Bernardo> divisions correctly, this issue should be solved in gcc, not
> Bernardo> by adding a silly macro to the kernel.
> The issue is that on 32-bit platforms, 64bit divided by 32 bit is
> handed off to a subroutine _udivdi3 which isn't linked into the
> kernel, and which in any case does a full 64 bit by 64-bit division
> (which is slow).
I see. It's ashaming that the gcc people didn't care special casing
the quite common 64/32 case in the x86 machine description or at least
> Using do_div() allows one to generate near-optimal code for a 64by32
> bit division/remainder on platforms (e.g., IA32) which have problems,
> and generating something sane for other platforms (e.g., IA64).
I agree. I'd prefer to see it fixed in gcc, but until then...
> Platforms that never expect to deal with a 64-bit number just redefine
> the macro in terms of long. Which means that printing out long longs
> doesn't work properly on those architectures.
A function which changes its semantics depending on the platform is
definitely a ugly hack.
A cleaner way to address this problem would be using platform-specific
typedefs to reduce the size of specific objects to 32bits on smaller systems.
In mm/vmscan.c:shrink_slab() you'll find this:
long long delta;
delta = scanned * shrinker->seeks;
delta *= (*shrinker->shrinker)(0, gfp_mask);
do_div(delta, pages + 1);
shrinker->nr += delta;
This is _BAD_ because delta will be 128bits on some 64bit systems
and 64bit, but truncated to 32bit in do_div(), on some other systems.
Using at least uint64_t would solve the first problem, but anyone
looking at this code wouldn't realize the result could get truncated
by some do_div() implementations.
What can we do about that? This time I don't have a clean solution to
-- // Bernardo Innocenti - Develer S.r.l., R&D dept. \X/ http://www.develer.com/
Please don't send Word attachments - http://www.gnu.org/philosophy/no-word-attachments.html
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to email@example.com More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Mon Jul 07 2003 - 22:00:15 EST