On Mon, 3 Nov 2025, Andrew Pinski wrote:
> On Mon, Nov 3, 2025 at 7:21 AM Richard Biener <[email protected]> wrote:
> >
> > The following fixes forwprop using FOR_EACH_IMM_USE_STMT to iterate
> > over stmts and then eventually removing the active stmt, releasing
> > its defs. This can cause debug stmt insertion with a RHS referencing
> > the SSA name we iterate over, adding to its immediate use list
> > but also adjusting all other debug stmts refering to the released
> > SSA name, updating those. And those can refer to the original
> > iterated over variable.
> >
> > In the end the destructive behavior of update_stmt is a problem
> > here, which unlinks all uses of a stmt and then links in the
> > newly computed ones instead of leaving those in place that are.
> >
> > * tree-ssa-forwprop.cc (forward_propagate_addr_expr):
> > Avoid removing stmts inside FOR_EACH_IMM_USE_STMT.
>
> LGTM.
> Note optimize_vector_load has the same issue. Maybe you did hit that
> but didn't go and didn't analyze all of the fallout of the checking.
I also think forward_propagate_addr_expr should be changed to work
in the use->def direction which would also fix this.
Hitting all fallout is quite difficult with the current checking
patch, so I fear we'll have to play whack-a-mole a lot. I'm
diverting to try inventing some better abstraction for most
of the FOR_EACH_IMM_USE_STMT uses.
Richard.
> Thanks,
> Andrew
>
>
> > ---
> > gcc/tree-ssa-forwprop.cc | 15 +++++++++------
> > 1 file changed, 9 insertions(+), 6 deletions(-)
> >
> > diff --git a/gcc/tree-ssa-forwprop.cc b/gcc/tree-ssa-forwprop.cc
> > index 9a993ab04de..4bac938cf54 100644
> > --- a/gcc/tree-ssa-forwprop.cc
> > +++ b/gcc/tree-ssa-forwprop.cc
> > @@ -930,6 +930,7 @@ forward_propagate_addr_expr (tree name, tree rhs, bool
> > parent_single_use_p)
> > bool all = true;
> > bool single_use_p = parent_single_use_p && has_single_use (name);
> >
> > + auto_vec<gimple *, 4> to_remove;
> > FOR_EACH_IMM_USE_STMT (use_stmt, iter, name)
> > {
> > bool result;
> > @@ -963,12 +964,14 @@ forward_propagate_addr_expr (tree name, tree rhs,
> > bool parent_single_use_p)
> > && TREE_CODE (gimple_assign_lhs (use_stmt)) == SSA_NAME
> > && TREE_CODE (use_rhs) == SSA_NAME
> > && has_zero_uses (gimple_assign_lhs (use_stmt)))
> > - {
> > - gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
> > - fwprop_invalidate_lattice (gimple_get_lhs (use_stmt));
> > - release_defs (use_stmt);
> > - gsi_remove (&gsi, true);
> > - }
> > + to_remove.safe_push (use_stmt);
> > + }
> > + for (gimple *use : to_remove)
> > + {
> > + gimple_stmt_iterator gsi = gsi_for_stmt (use);
> > + fwprop_invalidate_lattice (gimple_get_lhs (use));
> > + release_defs (use);
> > + gsi_remove (&gsi, true);
> > }
> >
> > return all && has_zero_uses (name);
> > --
> > 2.51.0
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
