[cfe-commits] [libcxx] r115633 - /libcxx/trunk/www/atomic_design.html

[Howard Hinnant]

On Oct 5, 2010, at 1:55 PM, John McCall wrote:

> On Oct 5, 2010, at 10:22 AM, Howard Hinnant wrote:
>> Author: hhinnant
>> Date: Tue Oct  5 12:22:28 2010
>> New Revision: 115633
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=115633&view=rev
>> Log:
>> A compiler writer's guide to <atomic>, minor update
>> 
>> Modified:
>>   libcxx/trunk/www/atomic_design.html
>> 
>> Modified: libcxx/trunk/www/atomic_design.html
>> URL: http://llvm.org/viewvc/llvm-project/libcxx/trunk/www/atomic_design.html?rev=115633&r1=115632&r2=115633&view=diff
>> ==============================================================================
>> --- libcxx/trunk/www/atomic_design.html (original)
>> +++ libcxx/trunk/www/atomic_design.html Tue Oct  5 12:22:28 2010
>> @@ -377,11 +377,17 @@
>> <p>
>> On some platforms, the compiler vendor can offer some or even all of the above
>> intrinsics at one or more weaker levels of memory synchronization.  This might
>> -lead for example to not issuing an <tt>mfense</tt> instruction on the x86.  If
>> -the compiler does not offer any given operation, at any given memory ordering
>> +lead for example to not issuing an <tt>mfense</tt> instruction on the x86.
>> +</p>
> 
> "mfence"

Some day I'll learn to spell. :-)

> 
> The design seems reasonable to me.
> 
> Have you considered how to support TR18037 address space-qualified pointers?  I assume <atomic> does a reinterpret_cast to void** to use the generic intrinsic, but this can silently do the wrong thing for AS-qualified pointers, which are not necessarily the same size as a generic pointer.  Fortunately, I don't think that happens for any Apple-supported platforms.

No I haven't considered such pointers at all.  I didn't design the top level API, only the implementation of it.  You are correct that there is a generic T* interface that feeds down to void*.  The top level API has a:

struct atomic_address;

which is void*-based.  And on top of that:

template <class T> struct atomic<T*> : atomic_address {..};

I suppose something along the lines of struct atomic_space_qualified_address could be added under an #ifdef, assuming someone was motivated to support such platforms.

-Howard