As far as I can see, this breaks macOS < 10.12

(In reply to Ozkan Sezer from comment #1)
> As far as I can see, this breaks macOS < 10.12

Fwiw, mach_absolute_time() is apparently deprecated as of 11.0, so we'll probably have to wrap this in preprocessor check in any case.  :/

--ryan.

Ah older still supported. I can try detection at config time in both autoconf/Cmake fashions.

Created attachment 4668 [details]
patch 2

Created attachment 4671 [details]
patch 3

The second patch didn't apply, I regenerated it by guess-work:
Please confirm that it matches your intention.

Yes sorry I m not comfortable with mercurial really much better with git :-) but the diff looks good indeed, except maybe those two lines removal

`
-#else
-#define SDL_MONOTONIC_CLOCK CLOCK_MONOTONIC
`

Created attachment 4672 [details]
patch 4

(In reply to David Carlier from comment #6)
> but the diff looks good indeed, except maybe those two lines removal

Attached v4 patch.

+1

Is there any benefit for using the newer API?
If it's performance, can you attach your test and the timing results you got?

Not that much (no more init but that s once) but more reliability, less error prone and easier to use. It s usually advised to move on to this newer api when possible (ideally I would have prefered removing the ancient one but depends on support wanted).

Patch added, thanks!
https://hg.libsdl.org/SDL/rev/8cad7d7ec50c

Created attachment 4697 [details]
better check for clock_gettime_nsec_np()

(In reply to Ozkan Sezer from comment #12)
> Created attachment 4697 [details]
> better check for clock_gettime_nsec_np()

The attached patch makes sure that the build is targeting macOS >= 10.12
for clock_gettime_nsec_np() availability.  It also adds missing HAVE_xxx
to SDL_config_macosx.h.  Please review.

Notes:
I do NOT know how to check IOS version requirements. time.h from 10.12
SDK says:
__IOS_AVAILABLE(10.0) __TVOS_AVAILABLE(10.0) __WATCHOS_AVAILABLE(3.0)

If these are already the minimum requirements from SDL, then I can add
the availability unconditionally:  Tell me what to do for it and I will
complete the patch and push to hg.

> I do NOT know how to check IOS version requirements

I do not know if that answers your question but the attribute available can

as :

if (@available(iOS 10, *)) // It is variadic to be able to check tvOs and so on

(In reply to David Carlier from comment #14)
> I do not know if that answers your question but the attribute available can
> 
> as :
> 
> if (@available(iOS 10, *)) // It is variadic to be able to check tvOs and so on

I believe that's for detection at run time.  We need something for compile-time.

ah right I think you have constants such as :
 - __IPHONE_OS_VERSION_MIN_REQUIRED
- more broadly __OSX_AVAILABLE_STARTING(__MAC_10_2,__IPHONE_10_0)

from the top of my head.

otherwise there are these in case :
- __WATCHOS_AVAILABLE(<version>) and __TVOS_AVAILABLE(<version>)

Created attachment 4700 [details]
better check for clock_gettime_nsec_np()

-- patch rebased to current hg

Created attachment 4701 [details]
better check for clock_gettime_nsec_np() - v6

Updated patch, v6:

- Unconditionally defines HAVE_CLOCK_GETTIME_NSEC_NP for iOS / TVOS
  assuming iOS >= 10.0.
  This is similar to SDL_JOYSTICK_MFI being defined for iOS or TVOS
  which depend on GameController framework which require iOS >=10.0
  at the least.

Comments?  OK to apply?

Seems ok to me

Latest patch looks good to me.

--ryan.

(In reply to David Carlier from comment #20)
> Seems ok to me

(In reply to Ryan C. Gordon from comment #21)
> Latest patch looks good to me.
> 
> --ryan.

OK, applied the latest patch as https://hg.libsdl.org/SDL/rev/d01593bd58d9

Re-closing as fixed.

SDL_GetTicks() was broken, so I backed this out:
https://hg.libsdl.org/SDL/rev/88c79c6a8457

Ah surprising but understood... Sorry for this.

It may make sense to use clock_gettime_nsec_np() for the performance timer and leave the original code for SDL_GetTicks().

Thoughts?

Could work out indeed.

Is there any advantage in performance or resolution?

Performance not really, resolution most likely and more consistent across architectures.

As far as I know there's no benefit to using clock_gettime_nsec_np versus the current code aside from mach_absolute_time being marked as deprecated. I suspect Apple only marked it as deprecated in the first place because other code (not SDL) was using it wrong on macOS in a way that will produce correct results on x64 but not Apple Silicon CPUs.