[PATCH 5/6] include: Add __rdtsc and __rdtscp intrinsics.

[Jacek Caban]

On 11.03.2021 21:37, Rémi Bernon wrote:
> On 3/11/21 9:23 PM, Jacek Caban wrote:
>> On 11.03.2021 11:47, Rémi Bernon wrote:
>>> On 3/10/21 10:20 PM, Rémi Bernon wrote:
>>>> On 3/10/21 9:44 PM, Jacek Caban wrote:
>>>>> On 10.03.2021 17:13, Rémi Bernon wrote:
>>>>>> Apparently clang >= 4.0.0 has __rdtsc (but not __rdtscp). Can we 
>>>>>> just assume clang >= 4.0.0 is used and something like that would 
>>>>>> be acceptable instead?
>>>>>
>>>>>
>>>>> There is __has_builtin(), which would be cleaner in this case, IMHO.
>>>>>
>>>>>
>>>>> Both compilers ship __rdtscp in ia32intrin.h header (gcc also 
>>>>> __rdtscp, which clang doesn't need). I was considering something 
>>>>> like this inside our intrin.h:
>>>>>
>>>>>
>>>>> #if defined(__i386__) || defined(__x86_64__)
>>>>>
>>>>> # include <x86intrin.h>
>>>>>
>>>>> #endif
>>>>>
>>>>>
>>>>> I believe that this should solve your problem. This rises some 
>>>>> compatibility concerns, but I think it may be fine. It's shipped 
>>>>> with compilers themselves for all platforms (not by mingw-w64).
>>>>>
>>>>>
>>>>> Jacek
>>>>>
>>>>
>>>> Ah yes sure, I didn't know these were available in intel intrinsic 
>>>> headers. I guess that would be perfect then.
>>>
>>> Except that it has some conflicts with msvcrt _(l)rot[r|l] functions 
>>> in stdlib.h. Can we remove the definitions from there?
>>
>>
>> #undefs, like in the attached patch, seem to be enough.
>>
>>
>> Jacek
>>
>
> IMHO if these are definitions, it would be more efficient to have 
> #ifndef instead, and use the compiler builtin intrinsic when it is 
> available.


Sure, that should work too.


> But then, I'm not sure we can reliably assume they are always 
> definitions (it seems to be the case on Clang and GCC, but there's no 
> guarantee it'll always be).


I kind of expected compatibility glitches in that area, that's why I 
wasn't sure if x86intrin.h is the way to go. It doesn't seem like a deal 
breaker to me yet.


Jacek