• redsea <redsea 163.com> Jan 07 2009
• "Jarrett Billingsley" <jarrett.billingsley gmail.com> Jan 07 2009
• redsea <redsea 163.com> Jan 07 2009
• "Robert Jacques" <sandford jhu.edu> Jan 07 2009
• Sean Kelly <sean invisibleduck.org> Jan 07 2009
• redsea <redsea 163.com> Jan 07 2009

redsea <redsea 163.com> writes:

When I implement some high performance program in linux 32bit, 32bit atomic if
often not sufficient, for example, If I want to hold a index & a timestamp to a
atomic variable, 32bit is not enough,  in 64bit environment, I can do that.

"Jarrett Billingsley" <jarrett.billingsley gmail.com> writes:

On Wed, Jan 7, 2009 at 10:47 AM, redsea <redsea 163.com> wrote:
 When I implement some high performance program in linux 32bit, 32bit atomic if
often not sufficient, for example, If I want to hold a index & a timestamp to a
atomic variable, 32bit is not enough,  in 64bit environment, I can do that.


..Congratulations?

redsea <redsea 163.com> writes:

Jarrett Billingsley Wrote:

 On Wed, Jan 7, 2009 at 10:47 AM, redsea <redsea 163.com> wrote:
 When I implement some high performance program in linux 32bit, 32bit atomic if
often not sufficient, for example, If I want to hold a index & a timestamp to a
atomic variable, 32bit is not enough,  in 64bit environment, I can do that.


 ..Congratulations?


You mean scope & RAII ?
Yes, it works, very well,  thanks Walter.

I test some dstress fail test case too, according to 

http://www.incasoftware.de/~kamm/ldc/tests/dmd-1.038-dstress1601/fail.html

scope_06_M  is also fixed, but some case with goto  still failed. 

scope_06_M in run/s/scope_06_M.d      ok now
scope_13_A in run/s/scope_13_A.d       still no ok, but Walter said it is not a
bug.
scope_14_A in run/s/scope_14_A.d       
scope_14_B in run/s/scope_14_B.d
scope_14_C in run/s/scope_14_C.d
scope_14_E in run/s/scope_14_E.d
scope_17_A in run/s/scope_17_A.d      failed
scope_17_C in run/s/scope_17_C.d       failed
scope_17_D in run/s/scope_17_D.d

other ones can not pass compile (so I wouldn't be afraid of them).

And 17_A, 17_C  is using goto combine scope/RAII, so, if we avoid the use of
goto, we could be safe.

"Robert Jacques" <sandford jhu.edu> writes:

On Wed, 07 Jan 2009 10:47:03 -0500, redsea <redsea 163.com> wrote:

 When I implement some high performance program in linux 32bit, 32bit  
 atomic if often not sufficient, for example, If I want to hold a index &  
 a timestamp to a atomic variable, 32bit is not enough,  in 64bit  
 environment, I can do that.


and Swap operation is availible, which allows you to use the index + tag  
method to advoid the ABA problem. However, a decent number of 64-bit  
systems don't support this operation leaving the vast majority of  
lock-free algorithms high and dry. (While there are high order bits you  
could hide a tag in, kernel developers are already using them for OS  
stuff). Of course, if you're referring to indexing into an array, you can  
use two shorts, etc. on 32-bit machines for much the same effect.

Sean Kelly <sean invisibleduck.org> writes:

== Quote from redsea (redsea 163.com)'s article
 When I implement some high performance program in linux 32bit, 32bit atomic if
often not sufficient,


environment, I can do that.

The IA-32 instruction set contains a 64-bit CAS operation in addition to
the 32-bit CAS.  You don't even need a 64-bit machine to do it!  The
atomics module in Tango supports 64-bit storeIf operations on 32-bit
Intel platforms.


Sean

redsea <redsea 163.com> writes:

Sean Kelly Wrote:
 The IA-32 instruction set contains a 64-bit CAS operation in addition to
 the 32-bit CAS.  You don't even need a 64-bit machine to do it!  The
 atomics module in Tango supports 64-bit storeIf operations on 32-bit
 Intel platforms.


Thanks  Sean Kelly and Robert Jacques for teaching me.

I read following in linux kernel code, so I think x86 32 can not support 64 CAS.
I'm going to read tango atomic code  :)

#ifdef CONFIG_X86_32
# include "atomic_32.h"
#else
# include "atomic_64.h"
#endif