Re: [MiNT] MiNTlib timing related functions behavior/performance

[Alan Hourihane <alanh@fairlite.co.uk>]

On 04/17/13 15:59, Eero Tamminen wrote:
> Hi,
>
> I was profiling Quake (II) and Doom code compiled with Vincent's toolchain
> i.e. with new MiNTlib, on (emulated) Falcon and found something you might
> be interested...
>
> By far most of the time in Quake startup, and in rendering a frame
> in Doom, is spent by localtime.o from MiNTlib, between these symbols
> in the binary:
> ---
> d0fd0 T localtime.o
> d30fc T _tzsetwall
> ---
> tzsetwall is symbol from localtime, but for some reason there are no
> symbols before that in the linked in localtime.o content, although
> one can see them with "mn":
> ---------------------
> $ ar x libc.a
> $ nm localtime.o | sort -n
> ...
>           U _strncpy
> 000025e6 T _tzset
> 000027ae T _mktime
> 000027c6 T _timelocal
> 000027dc T _timegm
> 000027fc T _timeoff
> 0000274a T _gmtime_r
> 0000275e T _offtime
> 0000278a T _ctime_r
> 0000281e T _time2posix
> 00002572 T _tzsetwall
> ...
> ---------------------
>
> And most of the time in those two programs is spent in addresses
> between these symbols, "localtime.o" and "_tzsetwall"...
>
>
> When I started to look what happens in MiNTlib timing related functions,
> I found something I consider a bit horrible...
>
>
> Quake code uses gettimeofday() and (u)sleep. Because of memory
> requirements etc, it's not possible to run it under MiNT.
>
>
> Sleep does some funny stuff with sigalarm() on MiNT and calls clock()
> on TOS.  Usleep() uses Fselect() on MiNT and calls _clock() on TOS.
> Usage of clock() is fine performance-wise.
>
> However, gettimeofday():
> * uses Tgettimeofday() on MiNT and Tgettime() & Tgetdate() on TOS.
> * Then it does mktime() which does timezone stuff, and if non-null
>    value is given as timezone arg (like Quake does), it will also
>    call timezone stuff directly.
>
> Currently it's assumed that this is the culprit for bad performance,
> it will hopefully be soon verified (by removal of this call & re-testing).

mktime() is a very bad performer - known fact. One of the packages
I was building decided mktime() was too slow and deemed it unusable
because it took longer than 60 seconds (when compiled with m68000,
m68020-60 is quicker, but still slow.)

>  From my very quick reading of MiNTlib localtime.c, it seems that most
> of timezone parsing stuff is done always (.e. twice in current Quake
> case).  Is this true?
>
>
> nanosleep() (not used in Quake) is even worse:
> - it uses gettimeofday() (with NULL timezone parameter),
> - then it does select(), which on TOS returns error I think, and
> - as result it will call gettimeofday() again.
>
> nanosleep() can, and is used, for shorter sleeps than sleep(), so
> it taking a lot more CPU than sleep(), especially on TOS, is IMHO
> not really acceptable.

Feel free to optimize ! It would be good to have much better performing
functions on the mintlib.

Alan.