Re: [f-cpu] Conditionnal load and store, the return

[Cedric BAIL <cedric.bail@free.fr>]

> > I have reread the discussion about conditionnal load and store, and I
> > think that we forgot something : exception. In fact when we do a load
> > or a store and check the condition only on write.
 
> Spoken about the bit in a LSU0 entry telling us if we can write ? if it
> is like a write-right token, setting it to 1 allows the first one to reset
> it and have right to write. If an exception occurs at this place, it means
> that there is no matching LSU0 entry (am I wrong ?), so there is no right
> to write and condition fails. Meanwhile, the exception is executed. At
> exception exit, condition fails so we can reexecute the faulty
> conditional store instruction.

In fact we were speaking about a possible store[z/nz/m/l/nm/nl] and 
load[z/nz/m/l/nm/nl] instructions. And the problem was that currently the
test is checked on wright, that mean you do first load the data and then you
verify if the test is ok. The problem is that if you access to a not valid
address, a exeption must occur before you wait for the answer from the memory
system. Or the test is false, so no trap must occur and no request to memory 
must be send. (with store, we didn't have the problem, because you write only 
if the test is ok).

But in fact we still have a problem with the conditional store/load with
the LSU test in multi processor architecture.

> > The problem is what append if page
> > fault occur ? I think we must clarify that.
> > From my point of view I think that we must do the test before starting any
> > memory operation, so a trap only occur if the test is true. We must do the
> > same if we have a conditionnal prefetch. It's for me the only way to
> > correctly execute the 2 branch of a if simultaneously.

> > Finally what did we do with the cachemm instruction ? Did we remove or
> > change this strange instruction ?

For next release of the manual I mark this instruction a deprecated, so if no 
body cry it will be change or removed.

Cedric
*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/