Re: Machine Dependent Reorganization Pass

[Jim Wilson <wilson at specifixinc dot com>]

kernel_learner wrote:
> I have made some changes including moving the call to
> MACHINE_DEPENDENT_REORG to after the call that does
> the delay slot scheduling.

I believe this has been suggested before, but I don't think we ever came 
to agreement that this would work for all targets, and we never agreed 
to add a second reorg hook.  This is an open issue.

> independent.  I am not sure if the the instruction
> dependency information which might have been
> previously computed is VALID at the point where the
> MACHINE_DEPENDENT_REORG pass is called.

This info is only used in combine and in the scheduler.  There is not 
much in between sched2 and machine_reorg, so the info is likely still 
accurate.  If you move machine_reorg after delay slot filling it is less 
likely to be accurate.

> I need to figure out how I can compute or use
> dependecy information between instructions in this
> final list. Anyone have any suggestions?

You could try calling the dependency routines again.  There are no 
target ports that do this though, so this is not guaranteed to be safe. 
 See the sched_analyze function in sched-deps.c.  You will need 
init_deps before calling that, and maybe some other stuff.  It might 
just be simpler to run a scheduling pass.  The IA-64 machine_reorg does 
this for other reasons.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com