Re: [debuglocus] PHI source locations

[Andrew MacLeod <amacleod at redhat dot com>]

Andrew Pinski wrote:
> On Mon, Mar 23, 2009 at 6:43 AM, Andrew MacLeod <amacleod@redhat.com> wrote:
>   
> > This is the first patch on the debuglocus branch.
> >
> > It enhances PHI nodes to carry around a source location field which allows
> > out-of-ssa to properly replace the locus field when copies are issued.
> > Previously, we lost a *lot* of locus information here.  The cost is an extra
> > word (the source locus field) in each phi argument.
> >
> > debuglocus requires this change in order to retain the locus field
> > throughout all optimizations, but its probably something that should have
> > been considered at the very beginning.
> >     
>
> Does this fix PR 26475?  Should this go in 4.5 even if debuglocus is
> not merged as there are cases where we are missing line info
> otherwise?
>   
Its certainly something to consider. It independent of the debuglocus 
stuff, just a prerequisite.   As for that testcase, I took a quick look, 
and it misses a statement there. One of the initial tasks will be to 
find all the places where we lose source locus information, and fix 
them.  So this patch plus a few follow ons should resolve that PR.  I 
will make sure to maintain the set of changes for locus retention 
separate. 

The primary thing I would investigate before checking that into 4.5 as a 
bug fix is see if there is any impact on compile time and memory with 
the additional phi arg word.  With virtual uses changing so much, I 
imagine the extra memory used would be vastly overshadowed by the 
savings in virtual PHI consumption.

I'll look at resolving the bits that are lost in that testcase today.

Andrew