On Mon, 7 Jun 1999, Philip Blundell wrote:
> >in the FPEmulator module in RISC OS. If I'm right about that, presumably
> >you'd need to install the fpem module in linux as usual, but the
> >instructions which *are* in the silicon wouldn't generate traps and would
> >be executed by the FPU?
>
> It's not quite that simple. Yes, there are some instructions that the FPA
> will never claim and the emulator will always get hold of, and those will just
> work. The more tricky case is that under some circumstances I think the FPA
> can claim and start to execute an instruction but then decide that it's too
> complicated and punt to software to finish things up. This avoids the need to
> dedicate silicon to handle all the corner cases of IEEE arithmetic. I don't
> have the FPA datasheet to hand right now but I suspect operations involving
> denormals probably fall into this category. The FPA support code needs to
> know how to extract the information about bounced instructions from the status
> word and finish things up. NWFPE certainly can't do this; I have no idea
> whether the Linux version of the Acron emulator can.
That's fair enough as far as it goes, but would I be completely wrong in
thinking that when the FPU gives up on an instruction, it will look to
software exactly as if the instruction was never even started and the FPEM
will do it's job normally on it?
--
James Craig <[EMAIL PROTECTED]>
<[EMAIL PROTECTED]>
unsubscribe: body of `unsubscribe linux-arm' to [EMAIL PROTECTED]
