Re: [PATCH] tree-vect-generic: Fix up ICE with SSA_NAME_OCCURS_IN_ABNORMAL_PHI [PR109392]

[Richard Biener <rguenther@suse.de>]

> 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
>