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 18:59 schrieb 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.
>
> Hmm, but I arrived here with that not being the case … (the chosen Lay-out not zero)

By chosen layout for the partition, I meant:

      unsigned int from_layout_i = m_partitions[partition_i].layout;

That should be 0 for these external vector nodes due to the
start_choosing_layout code that you mentioned in the PR.  to_layout_i
(the layout that the caller wants for a use of the node) can be anything,
so yeah, to_layout_i != 0 has to go through the "else" arm and generate
a VEC_PERM_EXPR.  But because from_layout_i == 0 for these nodes,
the "else" arm should handle all cases correctly, not just the
to_layout_i != 0 case.

Thanks,
Richard