[D-runtime] Why is clock_gettime() commented out?

[Jonathan M Davis]

On Friday, November 19, 2010 10:40:29 Sean Kelly wrote:
> On Nov 18, 2010, at 9:38 PM, Jonathan M Davis wrote:
> > On Thursday 18 November 2010 21:24:16 Sean Kelly wrote:
> >> On Nov 18, 2010, at 6:57 PM, Jonathan M Davis wrote:
> >>> Is there a particular reason why clock_gettime() and various related
> >>> posix C time functions are all commented out in core.sys.posix.time and
> >>> core.sys.posix.signals? They're necessary to get either a realtime
> >>> clock or a monotonic clock on Posix systems. It is true that that
> >>> necessetates linking with -lrt, but as far as I know, all Posix
> >>> systems should have librt, and the flag could be added to dmd.conf.
> >>> 
> >>> The stopwatch code in std.datetime is going to be potentially
> >>> inaccurate without clock_gettime(), and the normal time code won't be
> >>> quite as precise. Ideally, clock_gettime() would be available. Why are
> >>> all of those time functions commented out? Were they causing problems
> >>> somehow?
> >> 
> >> I don't recall.  I wrote all of the linux header blocks while the others
> >> were submitted.  My best guess is that they weren't defined back when I
> >> created the headers.  I see no problem with uncommenting them so long as
> >> the functions exist though.  I'll verify and then make the change for
> >> linux and OSX.
> > 
> > They definitely should exist. The caveat is that you have to link with
> > librt for them to work (why they didn't just stick them in glibc proper,
> > I have no idea - especially when anything that cares about precision
> > time at all should be using them).
> 
> clockid_t and related functions don't exist on OSX, so if you want to use
> them in std.datetime you'll have to fail gracefully to something less
> accurate.

That's annoying. Well, it uses gettimeofday() if clock_gettime() isn't 
available, which isn't as good, but as is close as you can get as far as I know, 
but because druntime and phobos are pre-compiled as libraries, it has to be 
based on the platform (you couldn't, for instance, use clock_gettime() on some 
Linux machines and gettimeofday() on others), but making MacOSX work differently 
from Linux works just fine. I do have to wonder though if MacOSX has its own 
functions for getting a monotonic and/or realtime clock. If so, we may want to 
use those on MacOSX. It's particularly undesirable to be using a non-monotonic 
clock for timing code such as std.datetime.StopWatch. But assuming that my 
proposed core.time and std.datetime are accepted, it will be an issue in both 
core.time and std.datetime, because while the code for getting the time for 
SysTime and the like is in std.datetime, because TickDuration has to go in 
core.time, the code which grabs the time from the monotonic clock is in 
core.time.

- Jonathan M Davis