On Sat, Oct 30, 2004 at 10:57:40AM +1000, Finn Thain wrote:
> I wonder if there is a performance benefit to either solution. My
> suspicion is there is not, assuming the emulation code is equal. (Though
> the linux kernel FPU emulation does offer a choice of precision/speed
> tradeoff.)
If the emulation code is exactly the same quality, the soft float will
beat out the trapped instructions every time. An unimplimented instruction
trap is more expensive than a function call.
> There are two bugs? As Brad pointed out, apple fixed a bug that broke
> A-trap exceptions without significant overhead. (You aren't referring to
> that bug I gather). I guess it is also possible that Motorola gave Apple
> information about the bug(s) that we don't have, that permitted a simple
> solution.
There are a lot more than two bugs in these chips. :)
The big one that prevents FPU emulation on older 68LC040 chips affects
several vectors, including the A and F line. Since Apple was able to
fix it for A, we could fix it for F if we figure out how they did it.
> Yes. The linux 2.6.8.1 kernel will not boot my LC040 at all unless FPU
> emulation is configured in. Could I fix it simply building the kernel with
> a softfloat-enabled gcc?
I think this is just because the kernel thinks you should always have
either the emulation or a real FPU, so it does something stupid. The
kernel didn't do that before the emulation code was added.
Brad Boyer
[EMAIL PROTECTED]
