Re: [PATCH] tree-optimization/110742 - fix latent issue with permuting existing vectors

[Richard Sandiford <richard.sandiford@arm.com>]

Richard Biener <rguenther@suse.de> writes:
>> Am 20.07.2023 um 16:09 schrieb Richard Sandiford <richard.sandiford@arm.com>:
>> 
>> Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
>>> When we materialize a layout we push edge permutes to constant/external
>>> defs without checking we can actually do so.  For externals defined
>>> by vector stmts rather than scalar components we can't.
>>> 
>>> Bootstrapped and tested on x86_64-unknown-linux-gnu.
>>> 
>>> OK?
>>> 
>>> Thanks,
>>> Richard.
>>> 
>>>    PR tree-optimization/110742
>>>    * tree-vect-slp.cc (vect_optimize_slp_pass::get_result_with_layout):
>>>    Do not materialize an edge permutation in an external node with
>>>    vector defs.
>>>    (vect_slp_analyze_node_operations_1): Guard purely internal
>>>    nodes better.
>>> 
>>>    * g++.dg/torture/pr110742.C: New testcase.
>>> ---
>>> gcc/testsuite/g++.dg/torture/pr110742.C | 47 +++++++++++++++++++++++++
>>> gcc/tree-vect-slp.cc                    |  8 +++--
>>> 2 files changed, 53 insertions(+), 2 deletions(-)
>>> create mode 100644 gcc/testsuite/g++.dg/torture/pr110742.C
>>> 
>>> diff --git a/gcc/testsuite/g++.dg/torture/pr110742.C b/gcc/testsuite/g++.dg/torture/pr110742.C
>>> new file mode 100644
>>> index 00000000000..d41ac0479d2
>>> --- /dev/null
>>> +++ b/gcc/testsuite/g++.dg/torture/pr110742.C
>>> @@ -0,0 +1,47 @@
>>> +// { dg-do compile }
>>> +
>>> +struct HARD_REG_SET {
>>> +  HARD_REG_SET operator~() const {
>>> +    HARD_REG_SET res;
>>> +    for (unsigned int i = 0; i < (sizeof(elts) / sizeof((elts)[0])); ++i)
>>> +      res.elts[i] = ~elts[i];
>>> +    return res;
>>> +  }
>>> +  HARD_REG_SET operator&(const HARD_REG_SET &other) const {
>>> +    HARD_REG_SET res;
>>> +    for (unsigned int i = 0; i < (sizeof(elts) / sizeof((elts)[0])); ++i)
>>> +      res.elts[i] = elts[i] & other.elts[i];
>>> +    return res;
>>> +  }
>>> +  unsigned long elts[4];
>>> +};
>>> +typedef const HARD_REG_SET &const_hard_reg_set;
>>> +inline bool hard_reg_set_subset_p(const_hard_reg_set x, const_hard_reg_set y) {
>>> +  unsigned long bad = 0;
>>> +  for (unsigned int i = 0; i < (sizeof(x.elts) / sizeof((x.elts)[0])); ++i)
>>> +    bad |= (x.elts[i] & ~y.elts[i]);
>>> +  return bad == 0;
>>> +}
>>> +inline bool hard_reg_set_empty_p(const_hard_reg_set x) {
>>> +  unsigned long bad = 0;
>>> +  for (unsigned int i = 0; i < (sizeof(x.elts) / sizeof((x.elts)[0])); ++i)
>>> +    bad |= x.elts[i];
>>> +  return bad == 0;
>>> +}
>>> +extern HARD_REG_SET rr[2];
>>> +extern int t[2];
>>> +extern HARD_REG_SET nn;
>>> +static HARD_REG_SET mm;
>>> +void setup_reg_class_relations(void) {
>>> +  HARD_REG_SET intersection_set, union_set, temp_set2;
>>> +  for (int cl2 = 0; cl2 < 2; cl2++) {
>>> +    temp_set2 = rr[cl2] & ~nn;
>>> +    if (hard_reg_set_empty_p(mm) && hard_reg_set_empty_p(temp_set2)) {
>>> +      mm = rr[0] & nn;
>>> +      if (hard_reg_set_subset_p(mm, intersection_set))
>>> +        if (!hard_reg_set_subset_p(mm, temp_set2) ||
>>> +            hard_reg_set_subset_p(rr[0], rr[t[cl2]]))
>>> +          t[cl2] = 0;
>>> +    }
>>> +  }
>>> +}
>>> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
>>> index 693621ca990..1d79c77e8ce 100644
>>> --- a/gcc/tree-vect-slp.cc
>>> +++ b/gcc/tree-vect-slp.cc
>>> @@ -5198,7 +5198,10 @@ vect_optimize_slp_pass::get_result_with_layout (slp_tree node,
>>>     return result;
>>> 
>>>   if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
>>> -      || SLP_TREE_DEF_TYPE (node) == vect_external_def)
>>> +      || (SLP_TREE_DEF_TYPE (node) == vect_external_def
>>> +      && (to_layout_i == 0
>>> +          /* We can't permute vector defs.  */
>>> +          || SLP_TREE_VEC_DEFS (node).is_empty ())))
>> 
>> Guess it's personal preference, but IMO it's easier to follow without the
>> to_layout_i condition, so that it ties directly to the create_partitions
>> test.
>
> I don’t understand- in the code guarding this we seem to expect to_layout_i == 0 and that’s the case we can handle as noop.  I didn’t understand why the function doesn’t always just do nothing in this case though, so I must have missed something.

OK, so I guess that disproves that my way is easier to understand :)

I think logically, the code is doing the equivalent of:

  int partition_i = m_vertices[node->vertex].partition;
  if (partition < 0)
    {
      /* If the vector is uniform or unchanged, there's nothing to do.  */
      ...      
    }
  else
    {
      ... Return node if to_layout_i matches this partition's chosen layout...
    }

And I guess I should have written it that way.

So when there is no partition, we have a constant or external def
built from individual scalars.  We can use the node as-is if the
caller wants an unpermuted node or if all elements are equal
(so that the permutation doesn't matter).  Otherwise we need
to permute the scalars.

When there is a partition, we can use the node as-is if the caller
wants the layout that was chosen for that partition.  Otherwise we
need a new VEC_PERM_EXPR node.

In the particular case of external defs built from vectors, we're
guaranteed that the node's chosen layout is 0 (i.e. the original layout),
and so both ways work.  But in principle this case fits the "else" arm
better than the "then" arm, because we're dealing with a node that is in
a partition, and that is not built from scalars.

Thanks,
Richard