[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Re: Rep:Re: Re: [f-cpu] New suggestion about call convention


>> "store [sp],r31" for example, if we call a sub-function with other
>> stores, we will crash the saved value of r31. So my question is : what must
>> I do to change the address register without discarding its pointer status ?
>> (you know, Whygee, you told me if I do "add 8,sp,sp", sp would no longer be
>> considered as a pointer).
>You have madd/msub for this purpose (an imediate version could be usefull too).

In the interim, addi/subi can still be used,
as i have evaluated the ("ideal") delay to be around 1 or 2
cycles of penalty for the "load" instruction (LSU already
contains data, no TLB miss ...)

But the decoder will refuse to "issue" the instruction
as long as the TLB is not checked and ok, so there is quite
a stall. Fortunately, the TLB is connected to the Xbar
and the ASU output so if the Load/store is located
right behind (1 or 2 instructions) the add, the TLB
can check the ASU output. Then the penalty will depend
on the TLB latency, then by the LSU and the cache etc...

(i hope this will work)


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