On Tue, 2 Feb 2010, Ralf Baechle wrote:
> > > > + if (cpu_has_64bits) {
> > > > + write_c0_entryhi(0xfffffffffffff000ULL);
> > >
> > > macro indicated that we need to avoid hazards here on R4000.
>
> A MTC0 instruction on an R4000 writes EntryHi on pipeline stage 7 but
> will read from the same register on stage 4 which leaves a window of
> 2 instructions, that is 2 NOP instructions needed.
A footnote says: "An MTC0 of a CPR must not be immediately followed by
MFC0 of the same CPR." -- that seems to imply a single intermediate
instruction is sufficient, but that's not stated explicitly and obviously
adding an extraneous instruction here, where performance does not matter,
cannot hurt.
> > > Perhaps adding:
> > >
> > > back_to_back_c0_hazard();
> > >
> > Compiler already added a nop, so I thought it wasn't necessary.
> > Doesn't hurt either, so I'll put it in.
>
> This probe is needed as per MIPSxx architecture spec and several CPUs will
> missbehave without it. The 74K which of course is 32-bit but it
> illustrates the issue might even issue these instructions out of order.
> back_to_back_c0_hazard will expand into a suitable sequence to handle
> the pipeline hazard. And we can't trust on the compiler doing the right
> thing here; as explained above we might need multiple nops and some CPUs
> will need other instructions to deal with the hazard, for example a number
> of SSNOPs or an EHB instruction.
I reckon there are MIPS64r2 ISA implementations out there already, so an
EHB is a necessity where appropriate.
Maciej
|