On 12/25/2014 11:53 AM, Maxim Uvarov wrote: > On 12/25/2014 12:34 PM, Taras Kondratiuk wrote: >> On 12/24/2014 12:21 PM, Maxim Uvarov wrote: >>> On 12/23/2014 07:15 PM, Taras Kondratiuk wrote: >>>> Linux-generic shouldn't have arch specific parts, because it is >>>> *generic*. Instead it can fall back to strong __sync functions if >>>> __atomic are not available. >>> Taras I think that even linux-generic should have it's own requirements. >>> It might be compiler version. >>> Dependence on libssl. Kernel version and etc. >>> >>> We do not say that linux-generic should work on any linux, even 20 years >>> old. We just define needed requirements >>> to compile it. The same is with atomics. Instead of creating 20 branches >>> for all gcc and not gcc version we just >>> stick with more fresh gcc and say that it's minimal requirement to run >>> linux-generic. >> I don't think it worth to make such a tight restriction to use only gcc >> 4.7+. Having fallback to __sync_* function will relax it to something >> like gcc 4.1+. Sure performance will be worse, but it should be ok for >> linux-generic. >> > Why it is ok? If possible we need to make linux-generic version fast. > Support of old gcc is not required. If it's needed for somebody he can > do this in his own platform.
Because performance is not the main target for linux-generic. I'm a bit puzzled with your arguments so want to clarify your point. I see three options here: 1. __atomic_* only, so we don't support older compilers. 2. __atomic_* with fallback to __sync_*. 3. __atomic_* and inline assembler for some architectures. Currently we have #3 in the code. As far as I understand you are for #1. Right? I'd support #2. -- Taras Kondratiuk _______________________________________________ lng-odp mailing list [email protected] http://lists.linaro.org/mailman/listinfo/lng-odp
