Re: [PATCH] Optimize nested permutation to single VEC_PERM_EXPR [PR54346]

["Xu, Liwei" <liwei.xu@intel.com>]

Hi Richard 

Sure, I’ll also check the VLA case once I’m back from vacation, thank you for your help.

Best Regards
Liwei

> On 30 Sep 2022, at 5:17 pm, Richard Sandiford <richard.sandiford@arm.com> wrote:
> 
> Richard Biener <richard.guenther@gmail.com> writes:
>>> On Mon, Sep 26, 2022 at 8:58 AM Liwei Xu <liwei.xu@intel.com> wrote:
>>> 
>>>        This patch implemented the optimization in PR 54346, which Merges
>>> 
>>>        c = VEC_PERM_EXPR <a, b, VCST0>;
>>>        d = VEC_PERM_EXPR <c, c, VCST1>;
>>>                to
>>>        d = VEC_PERM_EXPR <a, b, NEW_VCST>;
>>> 
>>>        Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
>>>        tree-ssa/forwprop-19.c fail to pass but I'm not sure whether it
>>>        is ok to removed it.
>> 
>> Looks good, but leave Richard a chance to ask for VLA vector support which
>> might be trivial to do.
> 
> Sorry for the slow reply.  It might be tricky to handle the general case,
> so I'd suggest going with this for now and dealing with VLA as a follow-up.
> (Probably after Prathamesh's changes to fold_vec_perm_expr.)
> 
> Thanks,
> Richard
> 
>> Btw, doesn't this handle the VEC_PERM + VEC_PERM case in
>> tree-ssa-forwprop.cc:simplify_permutation as well?  Note _that_ does
>> seem to handle VLA vectors.
>> 
>> Thanks,
>> Richard.
>> 
>>> gcc/ChangeLog:
>>> 
>>>        PR target/54346
>>>        * match.pd: Merge the index of VCST then generates the new vec_perm.
>>> 
>>> gcc/testsuite/ChangeLog:
>>> 
>>>        PR target/54346
>>>        * gcc.dg/pr54346.c: New test.
>>> 
>>> Co-authored-by: liuhongt <hongtao.liu@intel.com>
>>> ---
>>> gcc/match.pd                   | 41 ++++++++++++++++++++++++++++++++++
>>> gcc/testsuite/gcc.dg/pr54346.c | 13 +++++++++++
>>> 2 files changed, 54 insertions(+)
>>> create mode 100755 gcc/testsuite/gcc.dg/pr54346.c
>>> 
>>> diff --git a/gcc/match.pd b/gcc/match.pd
>>> index 345bcb701a5..9219b0a10e1 100644
>>> --- a/gcc/match.pd
>>> +++ b/gcc/match.pd
>>> @@ -8086,6 +8086,47 @@ and,
>>>   (minus (mult (vec_perm @1 @1 @3) @2) @4)))
>>> 
>>> 
>>> +/* (PR54346) Merge
>>> +   c = VEC_PERM_EXPR <a, b, VCST0>;
>>> +   d = VEC_PERM_EXPR <c, c, VCST1>;
>>> +   to
>>> +   d = VEC_PERM_EXPR <a, b, NEW_VCST>; */
>>> +
>>> +(simplify
>>> + (vec_perm (vec_perm@0 @1 @2 VECTOR_CST@3) @0 VECTOR_CST@4)
>>> + (with
>>> +  {
>>> +    if(!TYPE_VECTOR_SUBPARTS (type).is_constant())
>>> +      return NULL_TREE;
>>> +
>>> +    tree op0;
>>> +    machine_mode result_mode = TYPE_MODE (type);
>>> +    machine_mode op_mode = TYPE_MODE (TREE_TYPE (@1));
>>> +    int nelts = TYPE_VECTOR_SUBPARTS (type).to_constant();
>>> +    vec_perm_builder builder0;
>>> +    vec_perm_builder builder1;
>>> +    vec_perm_builder builder2 (nelts, nelts, 1);
>>> +
>>> +    if (!tree_to_vec_perm_builder (&builder0, @3)
>>> +    || !tree_to_vec_perm_builder (&builder1, @4))
>>> +      return NULL_TREE;
>>> +
>>> +    vec_perm_indices sel0 (builder0, 2, nelts);
>>> +    vec_perm_indices sel1 (builder1, 1, nelts);
>>> +
>>> +    for (int i = 0; i < nelts; i++)
>>> +      builder2.quick_push (sel0[sel1[i].to_constant()]);
>>> +
>>> +    vec_perm_indices sel2 (builder2, 2, nelts);
>>> +
>>> +    if (!can_vec_perm_const_p (result_mode, op_mode, sel2, false))
>>> +      return NULL_TREE;
>>> +
>>> +    op0 = vec_perm_indices_to_tree (TREE_TYPE (@4), sel2);
>>> +  }
>>> +  (vec_perm @1 @2 { op0; })))
>>> +
>>> +
>>> /* Match count trailing zeroes for simplify_count_trailing_zeroes in fwprop.
>>>    The canonical form is array[((x & -x) * C) >> SHIFT] where C is a magic
>>>    constant which when multiplied by a power of 2 contains a unique value
>>> diff --git a/gcc/testsuite/gcc.dg/pr54346.c b/gcc/testsuite/gcc.dg/pr54346.c
>>> new file mode 100755
>>> index 00000000000..d87dc3a79a5
>>> --- /dev/null
>>> +++ b/gcc/testsuite/gcc.dg/pr54346.c
>>> @@ -0,0 +1,13 @@
>>> +/* { dg-do compile } */
>>> +/* { dg-options "-O -fdump-tree-dse1" } */
>>> +
>>> +typedef int veci __attribute__ ((vector_size (4 * sizeof (int))));
>>> +
>>> +void fun (veci a, veci b, veci *i)
>>> +{
>>> +  veci c = __builtin_shuffle (a, b, __extension__ (veci) {1, 4, 2, 7});
>>> +  *i = __builtin_shuffle (c, __extension__ (veci) { 7, 2, 1, 5 });
>>> +}
>>> +
>>> +/* { dg-final { scan-tree-dump "VEC_PERM_EXPR.*{ 3, 6, 0, 0 }" "dse1" } } */
>>> +/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 1 "dse1" } } */
>>> \ No newline at end of file
>>> --
>>> 2.18.2
>>>