On Saturday 17 June 2006 14:54, Erik Hofman wrote:
> Mathias Fröhlich wrote:
> > If you feel that they should be used, can you care for them and make them
> > alias safe?
>
> Not at this time, I've got other things on my prioritylist.
Ok, the ones in fastmath.hxx are easy to fix, I have already prepared a
checkin for that.
The ones in the fastmath.cxx look really ugly to fix.
If we would have no different compiler than gcc I would have suggested to add
a special compile flag to fastmath.cxx to avoid gcc optimizing too much away.
But we also have to support other compilers ...
So I have to note that many of these functions are expected to be misscompiled
by modern compilers that care for alias sets like current gcc's do.
Anyway, you still try to sell to me that (in case of fast_fabs for example) it
is faster to take a floating point value probably already being loaded in the
fpu push it into memory reload it into the integer unit do some bit fiddling
push it back into memory and reload it into the floating point unit to do
further computations with that is faster than just take the fpu register and
execute a fabs operation on that?
Seriously, I *can* think of fast_atan's polynomials being faster than a full
blown correct functioin.
But those simple fabs, sign or neg function implemented in the header cannot
even be faster than the cpu builtin assembler instructions usually available
on current cpu's. Also, in presence of sse instructions forcing that fist
legacy i386 fpu instruction in that assembler snipet will not make it faster
(unloading from sse unit, loading to legacy fpu, in contrast to do the same
directly in the sse unit ...).
Greetings
Mathias
--
Mathias Fröhlich, email: [EMAIL PROTECTED]
_______________________________________________
Flightgear-devel mailing list
Flightgear-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/flightgear-devel
