Re: IRA patch: move clobbers downwards

[Jeff Law <law at redhat dot com>]

On 07/13/10 09:29, Vladimir Makarov wrote:
> Bernd Schmidt wrote:
> > This is an older patch I'm resubmitting.  The scheduler can move reg
> > clobbers away from the insns for which they were generated, generating a
> > window during which IRA thinks the register is live when it isn't.  This
> > can cause unnecessary conflicts.
> I am not sure that this is the right place to fix this problem.  
> Common sense is saying me that it should be solved in the first place 
> (insn scheduling).
>
> Moreover I am not sure that this problem is not solved already with 
> -fsched-pressure.  If it is not it could be easily solved by using the 
> infrastructure for -fsched-pressure.
You know, this sounds an awful lot like a discussion I had with David E. 
a while back.

ISTM the CLOBBER should be completely ignored by scheduling.  The 
clobber serves no purpose other than to tell the dataflow analyzers that 
a double-word value is going to be completely replaced.

Ideally, we'd just ignore the clobbers in the scheduler; however, that 
might be a PITA to implement.  But ISTM we ought to be able to shrink 
the clobber's lifetime in a single pass over the insns after the 
scheduler has run.

Or we could build a real pressure reduction pass using much of the 
scheduler's infrastructure.

> > The real problem here is that IRA does a backwards scan; just as in
> > peep2, it would probably be better to use a forwards scan using REG_DEAD
> > notes.  However, I'm told it's expensive to create those.
> Originally, it was done on the forward scan in IRA.  As I know for 
> many people the final goal was to remove REG_DEAD notes.  I was not in 
> a hurry to implement the backward scan because the major problem of 
> removing REG_DEAD is the reload pass which uses it.  But Richard 
> Sandiford made a patch to use a backward pass.  We invested a lot of 
> efforts to solve all (non-trivial) problems of the backward pass 
> patch.  So I think it is quite unreasonable to switch back to the 
> forward pass.
Interestingly enough, reload's use of REG_DEAD notes caused me some 
major grief recently.    I'll spare everyone the details.

Jeff