On Fri, 14 Mar 2008 20:55:20 +0100, "Andi Kleen" <andi@firstfloor.org>
said:
quoted text
> 
> The CMOV define should probably be dependent on what CPU the kernel
> is tuned for. It was originally written for when x86-64 was only
> K8 which has fast CMOV, but e.g. on P4 CMOV is actually deprecated
> over jumps.

Hi Andi,

I guess you are right. But there is quite a big number of different
types of P4. Let's see what the current situation is... defconfigs
(of current x86#testing+this patch/current linus) with
CONFIG_GENERIC_CPU=n:

Athlon: 4764 / 4667 occurences of cmovxx
Pentium-IV: 4079 / 3982 occurences of cmovxx
Pentium-M: 3939 / 3841 occurences of cmovxx
Core-2: 4335 / 4330 occurences of cmovxx

So it adds a few percent extra cmovxx's. The last one is fishy...
But I'm too hungry and sleepy to go hunt that one down.

quoted text
> > Both define fls64(), but i386 uses a generic one and x86_64 defines
> > one all by itself. The generic one is currently not suitable for
> > use by 64-bit archs... that can change.
> 
> It is very unlikely a generic one will ever be able to compete
> with a single instruction.

Generic is maybe not the right term: asm-generic/bitops/fls64.h has:

static inline int fls64(__u64 x)
{
        __u32 h = x >> 32;
        if (h)
                return fls(h) + 32;
        return fls(x);
}

I just wanted to move the 64-bit version to that header, with some
ifdefs to select the right one.

quoted text
> > x86_64 defines ARCH_HAS_FAST_MULTIPLIER, i386 not. This affects a
> > choice of generated code in the (generic) hweight function. It would
> > be nice if that could move to some other file.
> 
> It depends on the CPU, but it can be probably safely set on pretty much
> all modern x86 cores.

In fact I just found out that it only had an effect for 64 bit
machines. Still, setting it unconditionally feels wrong.

quoted text
> > x86_64 has a mysterious inline function set_bit_string, which is
> > only used by pci-calgary_64.c and pci-gart_64.c. Not sure what to
> > do with it.
> 
> It's generic and could really live in linux/bitops.h

It could. But it is a trivial (slow?) implementation. Probably fine
for the uses in pci-calgary_64.c and pci-gart_64.c (small ranges?),
but I would worry about people using it, thinking it was a near-
optimal implementation.

Greetings,
    Alexander
-- 
  Alexander van Heukelum
  heukelum@fastmail.fm

-- 
http://www.fastmail.fm - IMAP accessible web-mail

--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/