Re: Enable SSA at -O0

[Andrew MacLeod <amacleod at redhat dot com>]

Jan Hubicka wrote:
> Well, I don't know either...  I didn't run into debug info quality
> problems during development and I tried to use this as build compiler.
>   
thats a good sign anyway!

>   
> > right now -O0 is the only saving grace we have, and as much as I want to 
> > see the SSA path at -O0, I'd also hate for us to end up with numerous 
> > cases where we no longer have some basic info.
> >     
>
> I've disabled TERing across statements, just within one statement or
> putting statements with no debug info into statement with debug info.
> Also I re-coalesce all user variables into single variable.  So in
> theory we should do no worse then we did before tree-SSA as trees fed
> into expander are now only closer to pre-tree-SSA form.
>   

right. And as long as we don't do any SSA optimizations which move code 
around, we should never have to worry about overlapping live ranges 
which play havoc with debug output. in theory, it should be OK, but the 
specter of compiler temps without debug info finding their way in where 
they shouldn't be lurks at the window...  we certainly have to be 
careful what ssa passes are run.

> > > Bootstrapped/regtested i686-linux.  I am testing x86_64-linux and 
> > > PPC-linux but both
> > > passed with earlier version of the patch.
> > >
> > >  
> > >       
> > I presume this is being targeted for 4.4?
> >     
>
> I guess it depends on overall consensus.  I started with this patch as
> simple hack as Richard needed it for the lowering passes.  It turned out
> to need a lot of fixing here and there and enabling unit-at-a-time by
> default, so it ended up being more work. If for nothing, this patch was
> useful to fix a lot of latent or semi-latent problems ;)) and I think
> the results are not discougrating.  
> Seeing it from the other POV, this patch would enforce us fixing debug
> info quality issues in SSA that should improve debug quality when
> optimizing.
>   
Yeah. I'm currently working on a mechanism to provide a quantitative 
comparison of the 'user quality' of debug info issued between different 
runs of the compiler. It should also identify what isn't correct and 
where.  At least that is the plan :-)  This would be another useful 
thing to run through it, -O0 with SSA and -O0 without SSA and see how 
similar it is. And also as a test of the comparator mechanism itself.

We might then be able to detect that we have a problem we need to look 
at :-)

Andrew