[M3commit] CVS Update: cm3

[Tony Hosking]

It is sometimes more efficient, depending on hardware.

I forgot to mention that to use these intrinsics you need to tell the  
backend in your cm3.cfg to take advantage of suitable instructions  
where available.  On x86 I use "-march=i686".

On 17 Feb 2009, at 16:06, Jay wrote:

>
> Is CASP trivially constructed from CAS?
>
> CASP(value, old, new) == (CAS(value, old, new) == old) ?
>
> P for predicate? ie: returns boolean?
>
>
> Or a) am I just wrong?
>    b) sometimes it is more efficient implemented more directly?
>
>
> For example, I know "InterlockedIncrement" can be implemented as an  
> infinite retry loop around compare_and_swap, but that doesn't seem  
> most efficient, e.g:
>
>
> long InterlockedIncrement(volatile long* p):
>  while (1)
>    long old = *p;
>    if (InterlockedCompareExchange(p, old, old + 1) == old)
>      return (old + 1);
>
>
>
> I assume we don't care about 386 support, nor maybe even 486.
> Somewhere along the line way back "lock cmpxchg8b" is not supported.
>
>
> - Jay
>
>
> ----------------------------------------
>> Date: Tue, 17 Feb 2009 05:45:48 +0000
>> To: m3commit at elegosoft.com
>> From: hosking at elego.de
>> Subject: [M3commit] CVS Update: cm3
>>
>> CVSROOT: /usr/cvs
>> Changes by: hosking at birch. 09/02/17 05:45:48
>>
>> Modified files:
>> cm3/m3-sys/m3front/src/builtinOps/: BuiltinOps.m3 m3makefile
>> cm3/m3-sys/m3front/src/misc/: CG.i3 CG.m3 Token.m3
>> cm3/m3-sys/m3front/src/values/: Formal.i3 Formal.m3
>> Added files:
>> cm3/m3-sys/m3front/src/builtinOps/: Cas.i3 Cas.m3 CasP.i3
>> CasP.m3
>>
>> Log message:
>> Experimental support for compare_and_swap (CAS) and compare_and_set  
>> (CASP) builtin operations.
>> CAS is defined as follows:
>>
>> CAS(v, o, n): BaseType(v)
>>
>> where v is an integer variable (INTEGER/LONGINT), and o and n are  
>> assignable to v,
>> atomically sets v to value n, so long as the previous value of v is  
>> o, returning the previous
>> value of v in either case.
>>
>> Similarly, CASP is defined as:
>>
>> CASP(v, o, n): BOOLEAN
>>
>> atomically sets v to n, so long as the previous value of v is o,  
>> returning TRUE if the
>> comparison is successful and n was assigned to v.
>>
>> I may yet generalize these to permit v to have any ordinal type.
>> In unsafe modules I may also permit v to be a subtype of ADDRESS.
>>