Thanks. That's good to know. At the moment a fix to NSAllocateObject will solve my problem, but one may never know, I could end up with structures requiring 16 bytes alignments in any objects. (btw, this is coming from intrinsic accessing SSE2 registers).
Thanks for looking into this!
-- Laurent
On Jul 6, 2013, at 3:58 AM, Richard Frith-Macdonald < address@hidden> wrote: On 5 Jul 2013, at 14:41, Laurent Michel <address@hidden> wrote:
Ok, so that's a bug then. On x64, one should add 8 bytes of padding before (or after) the counter in front of any NSObject to preserve a clean alignment.
NSAllocateObject() already contained code to do that, but was getting it wrong using the alignment of double. With gcc it can use __BIGGEST_ALIGNMENT__ to get the right calculation, but that's not available with clang. So I changed the code in svn trunk to always use 16 if we are compiling with clang.
If anyone knows a clang equiivalent of __BIGGEST_ALIGNMENT__, we could/should change to use it, since using 16 may waste space on some architectures and may not be enough on others (though I know of none).
PS. this is for NSAllocateObject() ... if you call class_createInstance() then the runtime will probably not provide the required padding :-(
However, if you are using the old runtime, the compatibility layer in gnustep-base now implements class_createInstance() by calling NSDAllocateObject(), so in this case it should be safe to use that runtime function.
|