On Thu 2025-07-17 11:48:33, Richard Biener wrote:
> On Thu, 17 Jul 2025, Filip Kastl wrote:
>
> > On Thu 2025-07-17 10:00:01, Richard Biener wrote:
> > > On Thu, 17 Jul 2025, Filip Kastl wrote:
> > >
> > > > Hi,
> > > >
> > > > This patch cuts out the solver from tree-ssa-structalias.cc. Soon, I'd
> > > > like to
> > > > also cut out the part that generates constraints and perhaps also other
> > > > parts.
> > > >
> > > > My big goal is to implement Steensgaard-style points-to analysis and
> > > > use it
> > > > interprocedurally. Unlike the IPA-PTA we currently have, Steensgaard
> > > > should be
> > > > fast enough to be enabled by default. Once this matures a bit, my hope
> > > > is that
> > > > this could enable GCC to optimize in many interesting cases where it
> > > > didn't
> > > > have enough alias information before. But that is a very long-term
> > > > vision.
> > > >
> > > > Cutting tree-ssa-structalias.cc into smaller files is preparatory work.
> > > > Even ignoring the upcoming changes, cutting a file of this size into
> > > > smaller
> > > > files seems like a good idea to me. I believe it will improve
> > > > readability.
> > > > It will also be clearer what I'm modifying in each of the upcoming
> > > > IPA-PTA
> > > > patches. For example, I'm not planning to modify the existing solver
> > > > at all.
> > > > The splitting will also make it easier to eventually plug in the new
> > > > Steensgaard solver.
> > > >
> > > > I've tried to make as few changes as seemed reasonable. I go into more
> > > > detail
> > > > about how I did the splitting in the commit message.
> > > >
> > > > I've named the new files *-andersen.{cc,h} because the current solver
> > > > would be
> > > > refered to as Andersen-style in points-to analysis literature. When
> > > > I'll be
> > > > adding the Steensgard-style solver, I'm planning to call those files
> > > > *-steens.{cc,h}.
> > > >
> > > > Btw, I copied the license comment at the top of tree-ssa-structalias.cc
> > > > into
> > > > all the files I created. This means that it says "Copyright (C)
> > > > 2005-2025" in
> > > > files which were just created. Is that ok?
> > > >
> > > > Bootstrapped and regtested on x86_64. Ok to push?
> > >
> > > I'm OK with splitting up the file, but can we get rid of
> > > 'tree' and 'structalias'? I'd say ssa-pta or for the solver,
> > > which has nothing to do with SSA, use pta-andersen.{cc,h} instead?
> > > At least for the new file, don't rename the old one.
> >
> > Yes, I'm all for that. I thought about doing that in a later patch but I
> > guess
> > it will be nicer to do that now.
> >
> > > That extends to the choice of the namespace name as well (I'd use 'pta').
> >
> > Ok.
> >
> > > The constraint building is mostly gimple based (it doesn't use SSA),
> > > so gimple-pta-constraints.{cc,h} might be appropriate and the pass
> > > itself would be ssa-pta.{cc,h} since it puts the solutions on SSA
> > > vars only.
> >
> > You say we should keep tree-ssa-structalias.cc and that the pass (the alias,
> > ealias and ipa-pta passes, right?) should be in ssa-pta.{cc,h}. So you
> > suggest
> > splitting out the pass classes into ssa-pta.{cc,h}? I'm confused about
> > which
> > parts you think should then be in tree-ssa-structalias.cc and which in
> > ssa-pta.cc.
>
> For now those pieces would remain in tree-ssa-structalias.cc. After all
> reorg is complete we'd move that to ssa-pta.{cc,h}. So I suggest to
> defer this part until it gets clear how this all plays out.
Ah, ok. I see.
> > Also, why not also rename tree-ssa-structalias.cc to be consistent? Is it a
> > no-no to rename old files?
>
> See above.
>
> > If we ignored the fact that the constraint building doesn't really interact
> > with SSA (but AFAIU, it benefits from SSA) and that the passes don't really
> > modify gimple statements, we could have:
> >
> > pta-andersen.{cc,h}
> > pta-steens.{cc,h}
> > gimple-ssa-pta-constraints.{cc,h}
> > gimple-ssa-pta.{cc,h}
> >
> > That seems nicer to me. A justification for calling those files
> > gimple-ssa-*
> > could be "to convey that this analysis happens on GIMPLE SSA IR". But you
> > are
> > the boss, of course.
>
> Sure, that works for me and follows existing practice.
>
> > Btw, I also thought about renaming the passes. ipa-pta is fine. But ealias
> > and alias could be called epta and pta. That would convey what they are
> > actually doing better, I think. It would also be more consistent with file
> > names.
>
> Yeah, but that's a separate change affecting dump file names and
> -fdisable-tree-XYZ only.
I'll do that as a separate patch, then.
I'll write a second version of this patch where I rename
tree-ssa-structalias-andersen.{cc,h} to pta-andersen.{cc,h} and the structalias
namespace to pta (and update the commit message accordingly). Can I consider
that pre-approved?
Thanks,
Filip Kastl
