Re: [cpuops cmpxchg double V1 1/4] Generic support for this_cpu_cmpxchg_double

["H. Peter Anvin" <hpa@zytor.com>]

On 12/23/2010 04:16 PM, Christoph Lameter wrote:
> On Wed, 22 Dec 2010, Tejun Heo wrote:
> 
>>> I'm a bit confused on this one.  The standard cmpxchg() takes a scalar
>>> and a pointer, and returns a scalar.  The equivalent for the "double"
>>> variety would be to return a compound object, basically:
>>>
>>> struct double_ulong {
>>> 	unsigned long v[2];
>>> };
>>>
>>> ... which can be returned in registers on both i386 and x86-64.
> 
> Really? How would that work? I tried with uint128 but could not get the
> compiler to do the right thing.
> 

There are two return registers; two machine registers can be returned in
registers.  [u]int128 is poorly implemented in a lot of gcc versions,
since it really hasn't been exercised.  However, two-word structures
should work.  I do not believe a two-word *array* works, though.

>>> It's a bit clumsy from a type perspective, but I'm not sure that that is
>>> a bad thing.  Doing too much type genericity has caused us problems in
>>> the past.
>>
>> Yeah, the above might be better too.  Is there any reason to use
>> cmpxchg_double on anything smaller?
> 
> Yes. You may want to use cmpxchg_double on 32 bit entities for backwards
> compatibilities sake or any other smaller unit size. But those could also
> be realized using this_cpu_cmpxchg_<double the size> by just aggregating
> the amount.
> 
> If we can indeed pass 128 bit entities (as claimed by hpa) via registers
> then the logical choice would be to do
> 
> 	this_cpu_cmpxchg_16(pcp, old, new)
> 
> instead of cmpxchg_double. All parameters would have to be bit.
> Then we can avoid the strange cmpxchg_double semantics and can completely
> avoid introducing those.

I'm not sure it works with passing in a structure.

	-hpa