two dfa problems

[Richard Henderson]

On Sun, May 05, 2002 at 01:45:31PM -0400, Vladimir N. Makarov wrote:
> Richard, I'll look at this problem tomorrow.  But I think your estimation is
> not accurate.

Correct.  I later determined that the effect was caused by the
reservation problem that you diagnosed.  Indeed, your analysis

> [...] It means that we can issue infinite number of such insns.

suggests the reason for this failure.

> To prevent this we should put ev5_e1 and ev5_e0 in the same automaton.

I have done this in my current patch.

> If such solution were not acceptable, we could remove occurrence
> of ev5_fa from fdiv reservations nearby the end of the long reservations.

I have also done this in my current patch.  You're right -- modeling
the late reservation is silly.  Especially since the divide instruction
has a data-dependent latency.

> Because It is easy to miss such problem, I've made a patch which generates
> error (and reports the units) if the units should be in the same automaton.

Thanks for looking at this.


r~