[RFC] md reorg plans for 4.6?

[RFC] md reorg plans for 4.6?

[Paolo Bonzini]

Hi all,

with the new plugin infrastructure, it makes sense to replace the 
one-catches-all md reorg pass with target-specific passes plugged into 
the pass manager.

If the md reorg is doing just complex peephole optimizations that cannot 
be achieved with targets, that's fine.

This has the advantage of being able to move passes where they really 
belong.  Many targets could benefit:

- Blackfin, IA64 and picochip is doing clever tricks to move splitting, 
sched2 and vartracking in the middle of its reorg

- MEP doesn't do the same tricks with vartracking, but also requires 
dataflow

- MIPS also could be split in several passes with nice dumps, 
eliminating crtl->dbr_scheduled_p which is there just to place 
delay-slot scheduling in the middle of md reorg

- i386... md reorg is sane and it is already doing plugin-style pass 
registration... but what about moving reg-stack.c to config/i386? :-)

- the split passes are a mess and some targets split again at the 
beginning of reorg -- having separate dumps would be nice (s390)

- there's also M68HC11 and SH... well, I doubt there will be volunteers.

- for SPU, there is an additional scheduling pass done as part of 
md_reorg; again, splitting dumps would be nice and in addition I wonder 
if doing sched2+this pass is really necessary if more pass placement 
freedom is given to the backend.

At the same time, it would be nice to free the CFG only for targets that 
absolutely abhor it (doing so in their own md reorg pass) rather than 
doing it for all targets.

Paolo

Re: [RFC] md reorg plans for 4.6?

[Ian Lance Taylor]

Paolo Bonzini <bonzini@gnu.org> writes:

> with the new plugin infrastructure, it makes sense to replace the
> one-catches-all md reorg pass with target-specific passes plugged into
> the pass manager.

Yes, please.

> - there's also M68HC11 and SH... well, I doubt there will be volunteers.

But note that, e.g., bt-load.c looks like a generic file but is
actually SH specific, so moving it into config/sh will be an
improvement for everybody.

Ian

Re: [RFC] md reorg plans for 4.6?

[Steven Bosscher]

On Saturday, October 31, 2009, Ian Lance Taylor <iant@google.com> wrote:
> Paolo Bonzini <bonzini@gnu.org> writes:
>> - there's also M68HC11 and SH... well, I doubt there will be volunteers.
>
> But note that, e.g., bt-load.c looks like a generic file but is
> actually SH specific, so moving it into config/sh will be an
> improvement for everybody.

Actually bt-load or something similar would be useful for at least one
other target, ia64.

Ciao,
Steven