For most ARM architecture variants, the only atomic operation is swap. The semantics of the SWP instruction are what gcc calls __sync_lock_test_and_set (a rather odd name since the set is unconditional). Would it be possible to add a __sync_lock_test_and_set builtin for ARM that generates a SWP instruction? It would be useful to also provide a predefined macro that can be tested to determine whether the builtin is implemented. Typically, code that wants to do architecture-independent atomic operations will want to do something like: #if have atomic add etc ... atomic operations using those builtins #elsif have atomic swap ... atomic operations using swap #else ... atomic operations using pthreads, or #error #end I understand that __GCC_HAVE_SYNC_COMPARE_AND_SWAP_n can be tested for the first #if above (on sufficiently new compilers), on the basis that an architecture that provides compare-and-swap can use it to synthesize all of the other builtin atomic operations. However, swap is not sufficient to implement the other builtins without some additional constraints, e.g. a sentinel value, and assumptions about the threading model. For discussions please see: http://thread.gmane.org/gmane.comp.gcc.help/21113 Thanks, Phil.
SWP is deprecated in the ARM architecture. It would be a really bad idea to get gcc to generate that by default as it will break compatibility.
Hi Richard, This is obviously less of an issue than it was in 2007. I think there are enough pre-ARMv6 systems still deployed that they cannot be ignored, though. For example, I believe that the Android native development kit assumes that the hardware is >= ARMv5. Currently, portable code need to do something like: #if v6 or newer ...use gcc builtins, which generate ldx/stx... #else ...use asm statements that generate swp... #endif It would be a bit less clunky if gcc builtins could be used in both cases. Personally, I use asm sufficiently rarely that it takes me a while to work out the syntax each time (both the actual assembler syntax and the gcc asm statement syntax). The builtins are much easier to use. Regards, Phil.