tree-vect-generic: Fix up ICE with SSA_NAME_OCCURS_IN_ABNORMAL_PHI [PR109392]
Checks
Commit Message
Hi!
tree_vect_extract uses gimple-fold/match.pd to attempt to simplify
the BIT_FIELD_REF immediately.
Unfortunately, maybe_push_res_to_seq has:
/* Play safe and do not allow abnormals to be mentioned in
newly created statements. */
for (unsigned int i = 0; i < num_ops; ++i)
if (TREE_CODE (ops[i]) == SSA_NAME
&& SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ops[i]))
return NULL_TREE;
if (num_ops > 0 && COMPARISON_CLASS_P (ops[0]))
for (unsigned int i = 0; i < 2; ++i)
if (TREE_CODE (TREE_OPERAND (ops[0], i)) == SSA_NAME
&& SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND (ops[0], i)))
return NULL_TREE;
and so can fail in presence of SSA_NAME_OCCURS_IN_ABNORMAL_PHI SSA_NAMEs,
where we then trigger the assert that maybe_push_res_to_seq doesn't return
NULL. The above is perfectly reasonable when trying to actually simplify
something that has been already created and let us just punt, but in the
tree_vect_extract case we have to build something always.
The following patch fixes it by building the BIT_FIELD_REF manually in
that case if the build+simplification failed.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2023-04-05 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/109392
* tree-vect-generic.cc (tree_vec_extract): If maybe_push_res_to_seq
fails, build BIT_FIELD_REF insn without trying to simplify it.
* gcc.dg/pr109392.c: New test.
Jakub
Comments
> Am 05.04.2023 um 10:58 schrieb Jakub Jelinek via Gcc-patches <gcc-patches@gcc.gnu.org>:
>
> Hi!
>
> tree_vect_extract uses gimple-fold/match.pd to attempt to simplify
> the BIT_FIELD_REF immediately.
> Unfortunately, maybe_push_res_to_seq has:
> /* Play safe and do not allow abnormals to be mentioned in
> newly created statements. */
> for (unsigned int i = 0; i < num_ops; ++i)
> if (TREE_CODE (ops[i]) == SSA_NAME
> && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ops[i]))
> return NULL_TREE;
>
> if (num_ops > 0 && COMPARISON_CLASS_P (ops[0]))
> for (unsigned int i = 0; i < 2; ++i)
> if (TREE_CODE (TREE_OPERAND (ops[0], i)) == SSA_NAME
> && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND (ops[0], i)))
> return NULL_TREE;
> and so can fail in presence of SSA_NAME_OCCURS_IN_ABNORMAL_PHI SSA_NAMEs,
> where we then trigger the assert that maybe_push_res_to_seq doesn't return
> NULL. The above is perfectly reasonable when trying to actually simplify
> something that has been already created and let us just punt, but in the
> tree_vect_extract case we have to build something always.
>
> The following patch fixes it by building the BIT_FIELD_REF manually in
> that case if the build+simplification failed.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Ok, but maybe there’s a gimple_build overload that meanwhile implements the desired semantics? It would probably need to specify the valueization hook to follow SSA edges beyond the sequence we’re currently building.
Richard
> 2023-04-05 Jakub Jelinek <jakub@redhat.com>
>
> PR tree-optimization/109392
> * tree-vect-generic.cc (tree_vec_extract): If maybe_push_res_to_seq
> fails, build BIT_FIELD_REF insn without trying to simplify it.
>
> * gcc.dg/pr109392.c: New test.
>
> --- gcc/tree-vect-generic.cc.jj 2023-03-23 10:02:18.997935620 +0100
> +++ gcc/tree-vect-generic.cc 2023-04-04 14:28:32.977729134 +0200
> @@ -174,7 +174,16 @@ tree_vec_extract (gimple_stmt_iterator *
> opr.resimplify (NULL, follow_all_ssa_edges);
> gimple_seq stmts = NULL;
> tree res = maybe_push_res_to_seq (&opr, &stmts);
> - gcc_assert (res);
> + if (!res)
> + {
> + /* This can happen if SSA_NAME_OCCURS_IN_ABNORMAL_PHI are
> + used. Build BIT_FIELD_REF manually otherwise. */
> + t = build3 (BIT_FIELD_REF, type, t, bitsize, bitpos);
> + res = make_ssa_name (type);
> + gimple *g = gimple_build_assign (res, t);
> + gsi_insert_before (gsi, g, GSI_SAME_STMT);
> + return res;
> + }
> gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> return res;
> }
> --- gcc/testsuite/gcc.dg/pr109392.c.jj 2023-04-04 14:36:03.096109943 +0200
> +++ gcc/testsuite/gcc.dg/pr109392.c 2023-04-04 14:35:39.784452751 +0200
> @@ -0,0 +1,15 @@
> +/* PR tree-optimization/109392 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -Wno-psabi" } */
> +
> +typedef short __attribute__ ((__vector_size__ (64))) V;
> +V v, w;
> +void bar (void) __attribute__((returns_twice));
> +
> +V
> +foo (V a, V b)
> +{
> + bar ();
> + b &= v < b;
> + return (V) { foo (b, w)[3], (V) {}[3] };
> +}
>
> Jakub
>
On Wed, Apr 05, 2023 at 02:11:08PM +0200, Richard Biener wrote:
> Ok, but maybe there’s a gimple_build overload that meanwhile implements
> the desired semantics? It would probably need to specify the valueization
> hook to follow SSA edges beyond the sequence we’re currently building.
Jeff has apparently committed the patch in the meantime.
For gimple_build, did you mean to use it just for this fallback case,
or instead of the initial resimplification as well?
> > 2023-04-05 Jakub Jelinek <jakub@redhat.com>
> >
> > PR tree-optimization/109392
> > * tree-vect-generic.cc (tree_vec_extract): If maybe_push_res_to_seq
> > fails, build BIT_FIELD_REF insn without trying to simplify it.
> >
> > * gcc.dg/pr109392.c: New test.
> >
> > --- gcc/tree-vect-generic.cc.jj 2023-03-23 10:02:18.997935620 +0100
> > +++ gcc/tree-vect-generic.cc 2023-04-04 14:28:32.977729134 +0200
> > @@ -174,7 +174,16 @@ tree_vec_extract (gimple_stmt_iterator *
> > opr.resimplify (NULL, follow_all_ssa_edges);
> > gimple_seq stmts = NULL;
> > tree res = maybe_push_res_to_seq (&opr, &stmts);
> > - gcc_assert (res);
> > + if (!res)
> > + {
> > + /* This can happen if SSA_NAME_OCCURS_IN_ABNORMAL_PHI are
> > + used. Build BIT_FIELD_REF manually otherwise. */
> > + t = build3 (BIT_FIELD_REF, type, t, bitsize, bitpos);
> > + res = make_ssa_name (type);
> > + gimple *g = gimple_build_assign (res, t);
> > + gsi_insert_before (gsi, g, GSI_SAME_STMT);
> > + return res;
> > + }
> > gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> > return res;
> > }
Jakub
On Tue, 11 Apr 2023, Jakub Jelinek wrote:
> On Wed, Apr 05, 2023 at 02:11:08PM +0200, Richard Biener wrote:
> > Ok, but maybe there?s a gimple_build overload that meanwhile implements
> > the desired semantics? It would probably need to specify the valueization
> > hook to follow SSA edges beyond the sequence we?re currently building.
>
> Jeff has apparently committed the patch in the meantime.
> For gimple_build, did you mean to use it just for this fallback case,
> or instead of the initial resimplification as well?
I was originally thinking of the initial resimplification.
But yes, using gimple_build would have simplified it to
if (!res)
res = gimple_build (&stmts, BIT_FIELD_REF, type, t, bitsize, bitpos);
With the gsi overloads the whole function should now be able to
be turned into
return gimple_build (gsi, true, GSI_SAME_STMT, UNKNOWN_LOCATION,
BIT_FIELD_REF, type, t, bitsize, bitpos);
I think. Note that's not 100% equivalent to what we do now, so it's
probably better to defer to stage1. What it doesn't do is
/* We're using the resimplify API and maybe_push_res_to_seq to
simplify the BIT_FIELD_REF but restrict the simplification to
a single stmt while at the same time following SSA edges for
simplification with already emitted CTORs. */
which is achieved by passing NULL as the gimple_seq * argument to
.resimplify. The above replacement would instead do what a later
forwprop pass would do when folding. ISTR I wanted to do minimal
changes when rewriting this.
Richard.
> > > 2023-04-05 Jakub Jelinek <jakub@redhat.com>
> > >
> > > PR tree-optimization/109392
> > > * tree-vect-generic.cc (tree_vec_extract): If maybe_push_res_to_seq
> > > fails, build BIT_FIELD_REF insn without trying to simplify it.
> > >
> > > * gcc.dg/pr109392.c: New test.
> > >
> > > --- gcc/tree-vect-generic.cc.jj 2023-03-23 10:02:18.997935620 +0100
> > > +++ gcc/tree-vect-generic.cc 2023-04-04 14:28:32.977729134 +0200
> > > @@ -174,7 +174,16 @@ tree_vec_extract (gimple_stmt_iterator *
> > > opr.resimplify (NULL, follow_all_ssa_edges);
> > > gimple_seq stmts = NULL;
> > > tree res = maybe_push_res_to_seq (&opr, &stmts);
> > > - gcc_assert (res);
> > > + if (!res)
> > > + {
> > > + /* This can happen if SSA_NAME_OCCURS_IN_ABNORMAL_PHI are
> > > + used. Build BIT_FIELD_REF manually otherwise. */
> > > + t = build3 (BIT_FIELD_REF, type, t, bitsize, bitpos);
> > > + res = make_ssa_name (type);
> > > + gimple *g = gimple_build_assign (res, t);
> > > + gsi_insert_before (gsi, g, GSI_SAME_STMT);
> > > + return res;
> > > + }
> > > gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> > > return res;
> > > }
>
> Jakub
>
>
@@ -174,7 +174,16 @@ tree_vec_extract (gimple_stmt_iterator *
opr.resimplify (NULL, follow_all_ssa_edges);
gimple_seq stmts = NULL;
tree res = maybe_push_res_to_seq (&opr, &stmts);
- gcc_assert (res);
+ if (!res)
+ {
+ /* This can happen if SSA_NAME_OCCURS_IN_ABNORMAL_PHI are
+ used. Build BIT_FIELD_REF manually otherwise. */
+ t = build3 (BIT_FIELD_REF, type, t, bitsize, bitpos);
+ res = make_ssa_name (type);
+ gimple *g = gimple_build_assign (res, t);
+ gsi_insert_before (gsi, g, GSI_SAME_STMT);
+ return res;
+ }
gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
return res;
}
@@ -0,0 +1,15 @@
+/* PR tree-optimization/109392 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -Wno-psabi" } */
+
+typedef short __attribute__ ((__vector_size__ (64))) V;
+V v, w;
+void bar (void) __attribute__((returns_twice));
+
+V
+foo (V a, V b)
+{
+ bar ();
+ b &= v < b;
+ return (V) { foo (b, w)[3], (V) {}[3] };
+}