How does the register allocator supports modulo scheduled code?

How does the register allocator supports modulo scheduled code?

[ritu sabharwal]

Hello,
    We have done the modulo scheduling for the loops and added that to the 
gcc instruction scheduler. We have replaced the loop block with our 
scheduled code and it is working fine. But now when this is passed to the 
gcc register allocator, we are getting errors in reload pass. Will the gcc 
register allocator can not handle this modulo scheduled code?
Also we need to this modulo scheduling for IA64, so we need to do the 
predication for the modulo loop. Is there a way through which we can 
generate the predicate for these loop instructions i.e. how can we change 
the rtl format so that after generating the predicate it will keep that 
predicate in the rtl format and does the remainging gcc passes will support 
this predicated instructions?

Best,
Ritu.

_________________________________________________________________
MSN 8 with e-mail virus protection service: 3 months FREE*. 
http://join.msn.com/?page=features/virus&xAPID=42&PS=47575&PI=7324&DI=7474&SU= 
http://www.hotmail.msn.com/cgi-bin/getmsg&HL=1216hotmailtaglines_eliminateviruses_3mf

Re: How does the register allocator supports modulo scheduled code?

[Jan Hubicka]

> 
> 
> Hello,
>    We have done the modulo scheduling for the loops and added that to the 
> gcc instruction scheduler. We have replaced the loop block with our 
> scheduled code and it is working fine. But now when this is passed to the 
> gcc register allocator, we are getting errors in reload pass. Will the gcc 
> register allocator can not handle this modulo scheduled code?

What kind of errors do you get exactly?  In case the insturctions you
produced match the constraints it should work just OK.  In case they
don't reload won't be able to fix up the hard registers.  One possible
trick is to do the same as new regalloc does - use pseudos one for each
hard register and manage register allocator to allocate these special
pseudos in the hard registers you want.  This is somewhat hackish
tought..

> Also we need to this modulo scheduling for IA64, so we need to do the 
> predication for the modulo loop. Is there a way through which we can 
> generate the predicate for these loop instructions i.e. how can we change 
> the rtl format so that after generating the predicate it will keep that 
> predicate in the rtl format and does the remainging gcc passes will support 
> this predicated instructions?

There is support for predicated instructions in RTL chain (see if
conversion pass implementation) but currently only few passes are
updated to deal with these.  Updating others is very desirable but may
be dificult so I think it should be done soeparately once you have
working modulo scheduling without this feature.

Honza
> 
> Best,
> Ritu.
> 
> _________________________________________________________________
> MSN 8 with e-mail virus protection service: 3 months FREE*. 
> http://join.msn.com/?page=features/virus&xAPID=42&PS=47575&PI=7324&DI=7474&SU= 
> http://www.hotmail.msn.com/cgi-bin/getmsg&HL=1216hotmailtaglines_eliminateviruses_3mf

Re: How does the register allocator supports modulo scheduled code?

[Andrew Haley]

Jan Hubicka writes:
 > > 
 > >    We have done the modulo scheduling for the loops and added that to the 
 > > gcc instruction scheduler. We have replaced the loop block with our 
 > > scheduled code and it is working fine. But now when this is passed to the 
 > > gcc register allocator, we are getting errors in reload pass. Will the gcc 
 > > register allocator can not handle this modulo scheduled code?
 > 
 > What kind of errors do you get exactly?  In case the insturctions you
 > produced match the constraints it should work just OK.  In case they
 > don't reload won't be able to fix up the hard registers.

Surely this can't work because reload will mess up the carefully
arranged schedule.

Andrew.

Re: How does the register allocator supports modulo scheduled code?

[Jan Hubicka]

> Jan Hubicka writes:
>  > > 
>  > >    We have done the modulo scheduling for the loops and added that to the 
>  > > gcc instruction scheduler. We have replaced the loop block with our 
>  > > scheduled code and it is working fine. But now when this is passed to the 
>  > > gcc register allocator, we are getting errors in reload pass. Will the gcc 
>  > > register allocator can not handle this modulo scheduled code?
>  > 
>  > What kind of errors do you get exactly?  In case the insturctions you
>  > produced match the constraints it should work just OK.  In case they
>  > don't reload won't be able to fix up the hard registers.
> 
> Surely this can't work because reload will mess up the carefully
> arranged schedule.
In case you do the register allocation job, reload will do nothing, so
this can work.

Honza
> 
> Andrew.

Re: How does the register allocator supports modulo scheduled code?

[Daniel Berlin]

On Mon, 23 Dec 2002, ritu sabharwal wrote:

>
>
> Hello,
>     We have done the modulo scheduling for the loops and added that to the
> gcc instruction scheduler.
> We have replaced the loop block with our
> scheduled code and it is working fine. But now when this is passed to the
> gcc register allocator, we are getting errors in reload pass. Will the gcc
> register allocator can not handle this modulo scheduled code?
Errr, we can't diagnose problems that only occur with your modulo
scheduling pass, without having your modulo scheduling pass.
People need to see and be able to work with the code, to determine if it's
reload's fault, your code's fault, or some combination of the two.
I would suggest you post your modulo schedulnig patch (even if it's rough)
if you want people to be able to help you.

--Dan