[MPlayer-dev-eng] Re: Vorbis decoding performance

[Rich Felker]

On Fri, Aug 18, 2006 at 02:42:13AM +0200, Michael Niedermayer wrote:
> > > hmm floating point emulation is slow sure but it shouldnt be that slow ...
> > 
> > Why not? At least Nx slower should be expected where every (1-cycle)
> > opcode becomes a N-cycle invalid opcode fault and context switch to
> > kernelspace handler. I would expect N is somewhere between 200 and
> > 5000.
> 
> well, for some reason i thought that the compiler should not generate invalid
> opcodes to begin with but rather inline / call the emu code ...
> and keep in mind that some cpu architectures have no fp-opcodes at all, so
> gcc should be able to do that, but then again ive not looked at what gcc
> really does on architectures without fpu or what it doess when the cpu
> type specified has no fpu but the architectures has ...

it depends on the arch and the compiler options. -msoft-float or
similar is needed to do this. even with soft float, the default on
some archs is still to use the (nonexistent, exception-generating) fpu
registers to return floating point return values; otherwise the abi
would be different with/without soft float and objects compiled with
different options could not be linked.

rich