Re: [RFC] [v2] Extend fold_vec_perm to handle VLA vectors

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

Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
> On Tue, 25 Jul 2023 at 18:25, Richard Sandiford
> <richard.sandiford@arm.com> wrote:
>>
>> Hi,
>>
>> Thanks for the rework and sorry for the slow review.
> Hi Richard,
> Thanks for the suggestions!  Please find my responses inline below.
>>
>> Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
>> > Hi Richard,
>> > This is reworking of patch to extend fold_vec_perm to handle VLA vectors.
>> > The attached patch unifies handling of VLS and VLA vector_csts, while
>> > using fallback code
>> > for ctors.
>> >
>> > For VLS vector, the patch ignores underlying encoding, and
>> > uses npatterns = nelts, and nelts_per_pattern = 1.
>> >
>> > For VLA patterns, if sel has a stepped sequence, then it
>> > only chooses elements from a particular pattern of a particular
>> > input vector.
>> >
>> > To make things simpler, the patch imposes following constraints:
>> > (a) op0_npatterns, op1_npatterns and sel_npatterns are powers of 2.
>> > (b) The step size for a stepped sequence is a power of 2, and
>> >       multiple of npatterns of chosen input vector.
>> > (c) Runtime vector length of sel is a multiple of sel_npatterns.
>> >      So, we don't handle sel.length = 2 + 2x and npatterns = 4.
>> >
>> > Eg:
>> > op0, op1: npatterns = 2, nelts_per_pattern = 3
>> > op0_len = op1_len = 16 + 16x.
>> > sel = { 0, 0, 2, 0, 4, 0, ... }
>> > npatterns = 2, nelts_per_pattern = 3.
>> >
>> > For pattern {0, 2, 4, ...}
>> > Let,
>> > a1 = 2
>> > S = step size = 2
>> >
>> > Let Esel denote number of elements per pattern in sel at runtime.
>> > Esel = (16 + 16x) / npatterns_sel
>> >         = (16 + 16x) / 2
>> >         = (8 + 8x)
>> >
>> > So, last element of pattern:
>> > ae = a1 + (Esel - 2) * S
>> >      = 2 + (8 + 8x - 2) * 2
>> >      = 14 + 16x
>> >
>> > a1 /trunc arg0_len = 2 / (16 + 16x) = 0
>> > ae /trunc arg0_len = (14 + 16x) / (16 + 16x) = 0
>> > Since both are equal with quotient = 0, we select elements from op0.
>> >
>> > Since step size (S) is a multiple of npatterns(op0), we select
>> > all elements from same pattern of op0.
>> >
>> > res_npatterns = max (op0_npatterns, max (op1_npatterns, sel_npatterns))
>> >                        = max (2, max (2, 2)
>> >                        = 2
>> >
>> > res_nelts_per_pattern = max (op0_nelts_per_pattern,
>> >                                                 max (op1_nelts_per_pattern,
>> >                                                          sel_nelts_per_pattern))
>> >                                     = max (3, max (3, 3))
>> >                                     = 3
>> >
>> > So res has encoding with npatterns = 2, nelts_per_pattern = 3.
>> > res: { op0[0], op0[0], op0[2], op0[0], op0[4], op0[0], ... }
>> >
>> > Unfortunately, this results in an issue for poly_int_cst index:
>> > For example,
>> > op0, op1: npatterns = 1, nelts_per_pattern = 3
>> > op0_len = op1_len = 4 + 4x
>> >
>> > sel: { 4 + 4x, 5 + 4x, 6 + 4x, ... } // should choose op1
>> >
>> > In this case,
>> > a1 = 5 + 4x
>> > S = (6 + 4x) - (5 + 4x) = 1
>> > Esel = 4 + 4x
>> >
>> > ae = a1 + (esel - 2) * S
>> >      = (5 + 4x) + (4 + 4x - 2) * 1
>> >      = 7 + 8x
>> >
>> > IIUC, 7 + 8x will always be index for last element of op1 ?
>> > if x = 0, len = 4, 7 + 8x = 7
>> > if x = 1, len = 8, 7 + 8x = 15, etc.
>> > So the stepped sequence will always choose elements
>> > from op1 regardless of vector length for above case ?
>> >
>> > However,
>> > ae /trunc op0_len
>> > = (7 + 8x) / (4 + 4x)
>> > which is not defined because 7/4 != 8/4
>> > and we return NULL_TREE, but I suppose the expected result would be:
>> > res: { op1[0], op1[1], op1[2], ... } ?
>> >
>> > The patch passes bootstrap+test on aarch64-linux-gnu with and without sve,
>> > and on x86_64-unknown-linux-gnu.
>> > I would be grateful for suggestions on how to proceed.
>> >
>> > Thanks,
>> > Prathamesh
>> >
>> > diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc
>> > index a02ede79fed..8028b3e8e9a 100644
>> > --- a/gcc/fold-const.cc
>> > +++ b/gcc/fold-const.cc
>> > @@ -85,6 +85,10 @@ along with GCC; see the file COPYING3.  If not see
>> >  #include "vec-perm-indices.h"
>> >  #include "asan.h"
>> >  #include "gimple-range.h"
>> > +#include <algorithm>
>> > +#include "tree-pretty-print.h"
>> > +#include "gimple-pretty-print.h"
>> > +#include "print-tree.h"
>> >
>> >  /* Nonzero if we are folding constants inside an initializer or a C++
>> >     manifestly-constant-evaluated context; zero otherwise.
>> > @@ -10493,15 +10497,9 @@ fold_mult_zconjz (location_t loc, tree type, tree expr)
>> >  static bool
>> >  vec_cst_ctor_to_array (tree arg, unsigned int nelts, tree *elts)
>> >  {
>> > -  unsigned HOST_WIDE_INT i, nunits;
>> > +  unsigned HOST_WIDE_INT i;
>> >
>> > -  if (TREE_CODE (arg) == VECTOR_CST
>> > -      && VECTOR_CST_NELTS (arg).is_constant (&nunits))
>> > -    {
>> > -      for (i = 0; i < nunits; ++i)
>> > -     elts[i] = VECTOR_CST_ELT (arg, i);
>> > -    }
>> > -  else if (TREE_CODE (arg) == CONSTRUCTOR)
>> > +  if (TREE_CODE (arg) == CONSTRUCTOR)
>> >      {
>> >        constructor_elt *elt;
>> >
>> > @@ -10519,6 +10517,230 @@ vec_cst_ctor_to_array (tree arg, unsigned int nelts, tree *elts)
>> >    return true;
>> >  }
>> >
>> > +/* Return a vector with (NPATTERNS, NELTS_PER_PATTERN) encoding.  */
>> > +
>> > +static tree
>> > +vector_cst_reshape (tree vec, unsigned npatterns, unsigned nelts_per_pattern)
>> > +{
>> > +  gcc_assert (pow2p_hwi (npatterns));
>> > +
>> > +  if (VECTOR_CST_NPATTERNS (vec) == npatterns
>> > +      && VECTOR_CST_NELTS_PER_PATTERN (vec) == nelts_per_pattern)
>> > +    return vec;
>> > +
>> > +  tree v = make_vector (exact_log2 (npatterns), nelts_per_pattern);
>> > +  TREE_TYPE (v) = TREE_TYPE (vec);
>> > +
>> > +  unsigned nelts = npatterns * nelts_per_pattern;
>> > +  for (unsigned i = 0; i < nelts; i++)
>> > +    VECTOR_CST_ENCODED_ELT(v, i) = vector_cst_elt (vec, i);
>> > +  return v;
>> > +}
>> > +
>> > +/* Helper routine for fold_vec_perm_vla to check if ARG is a suitable
>> > +   operand for VLA vec_perm folding. If arg is VLS, then set
>> > +   NPATTERNS = nelts and NELTS_PER_PATTERN = 1.  */
>> > +
>> > +static tree
>> > +valid_operand_for_fold_vec_perm_cst_p (tree arg)
>> > +{
>> > +  if (TREE_CODE (arg) != VECTOR_CST)
>> > +    return NULL_TREE;
>> > +
>> > +  unsigned HOST_WIDE_INT nelts;
>> > +  unsigned npatterns, nelts_per_pattern;
>> > +  if (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg)).is_constant (&nelts))
>> > +    {
>> > +      npatterns = nelts;
>> > +      nelts_per_pattern = 1;
>> > +    }
>> > +  else
>> > +    {
>> > +      npatterns = VECTOR_CST_NPATTERNS (arg);
>> > +      nelts_per_pattern = VECTOR_CST_NELTS_PER_PATTERN (arg);
>> > +    }
>> > +
>> > +  if (!pow2p_hwi (npatterns))
>> > +    return NULL_TREE;
>> > +
>> > +  return vector_cst_reshape (arg, npatterns, nelts_per_pattern);
>> > +}
>>
>> I don't think we should reshape the vectors for VLS, since it would
>> create more nodes for GC to clean up later.  Also, the "compact" encoding
>> is canonical even for VLS, so the reshaping would effectively create
>> noncanonical constants (even if only temporarily).
> Removed in the attached patch.
>>
>> Instead, I think we should change the later:
>>
>> > +  if (!valid_mask_for_fold_vec_perm_cst_p (arg0, arg1, sel, sel_npatterns,
>> > +                                        sel_nelts_per_pattern, reason, verbose))
>> > +    return NULL_TREE;
>>
>> so that it comes after the computation of res_npatterns and
>> res_nelts_per_pattern.  Then, if valid_mask_for_fold_vec_perm_cst_p
>> returns false, and if the result type has a constant number of elements,
>> we should:
>>
>> * set res_npatterns to that number of elements
>> * set res_nelts_per_pattern to 1
>> * continue instead of returning null
> Assuming we don't enforce only VLA or only VLS for input vectors and sel,
> won't that be still an issue if res (and sel) is VLS, and input
> vectors are VLA ?
> For eg:
> arg0, arg1 are type VNx4SI with npatterns = 2, nelts_per_pattern = 3, step = 2
> sel is V4SI constant with encoding { 0, 2, 4, ... }
> and res_type is V4SI.
> In this case, when it comes to index 4, the vector selection becomes ambiguous,
> since it can be arg1 for len = 4 + 4x, and arg0 for lengths > 4 + 4x ?

Ah, right.  So the condition is whether the result type and the data
input type have a constant number of elements, rather than just the result.

> In the attached patch if sel is not a suitable mask, and input vectors
> have constant length, then it sets:
> res_npatterns = nelts of input vector
> res_nelts_per_pattern = 1
> Does that look OK ?
> This part of the code has few tests in test_mixed().

res_npatterns should be nelts of the result vector, not the input.

>> The loop that follows will then do the correct thing for each element.
>>
>> The check for a power of 2 would then go in
>> valid_mask_for_fold_vec_perm_cst_p rather than
>> valid_operand_for_fold_vec_perm_cst_p.
>>
>> With that change, I think:
>>
>> > +
>> > +/* Helper routine for fold_vec_perm_cst to check if SEL is a suitable
>> > +   mask for VLA vec_perm folding. Set SEL_NPATTERNS and SEL_NELTS_PER_PATTERN
>> > +   similarly.  */
>> > +
>> > +static bool
>> > +valid_mask_for_fold_vec_perm_cst_p (tree arg0, tree arg1,
>> > +                                 const vec_perm_indices &sel,
>> > +                                 unsigned& sel_npatterns,
>> > +                                 unsigned& sel_nelts_per_pattern,
>> > +                                 char *reason = NULL,
>> > +                                 bool verbose = false)
>> > +{
>> > +  unsigned HOST_WIDE_INT nelts;
>> > +  if (sel.length ().is_constant (&nelts))
>> > +    {
>> > +      sel_npatterns = nelts;
>> > +      sel_nelts_per_pattern = 1;
>> > +    }
>> > +  else
>> > +    {
>> > +      sel_npatterns = sel.encoding ().npatterns ();
>> > +      sel_nelts_per_pattern = sel.encoding ().nelts_per_pattern ();
>> > +    }
>>
>> ...we should use the "else" code unconditionally.
> Done.
>>
>> The function comment should describe "reason" and "verbose".  It looks
>> like these are debug parameters, is that right?
> Yes, verbose is just for printf debugging, and reason is used in unit-tests when
> result is NULL_TREE, to verify that it's NULL for the intended reason,
> and not due to something else.

Could we make it a "const char **" instead, and just store the string
pointers there?  It looks like all the reasons are fixed strings rather
than dynamic.

>>
>> > +
>> > +  if (!pow2p_hwi (sel_npatterns))
>> > +    {
>> > +      if (reason)
>> > +     strcpy (reason, "sel_npatterns is not power of 2");
>> > +      return false;
>> > +    }
>> > +
>> > +  /* We want to avoid cases where sel.length is not a multiple of npatterns.
>> > +     For eg: sel.length = 2 + 2x, and sel npatterns = 4.  */
>> > +  poly_uint64 esel;
>> > +  if (!multiple_p (sel.length (), sel_npatterns, &esel))
>> > +    {
>> > +      if (reason)
>> > +     strcpy (reason, "sel.length is not multiple of sel_npatterns");
>> > +      return false;
>> > +    }
>> > +
>> > +  if (sel_nelts_per_pattern < 3)
>> > +    return true;
>> > +
>> > +  for (unsigned pattern = 0; pattern < sel_npatterns; pattern++)
>> > +    {
>> > +      poly_uint64 a1 = sel[pattern + sel_npatterns];
>> > +      poly_uint64 a2 = sel[pattern + 2 * sel_npatterns];
>> > +
>> > +      poly_uint64 step = a2 - a1;
>> > +      if (!step.is_constant ())
>> > +     {
>> > +       if (reason)
>> > +         strcpy (reason, "step is not constant");
>> > +       return false;
>> > +     }
>> > +      int S = step.to_constant ();
>> > +      if (S == 0)
>> > +     continue;
>>
>> Might be simpler as:
>>
>>       HOST_WIDE_INT step;
>>       if (!(a2 - a1).is_constant (&step))
>>         {
>>           ...
>>         }
> This resulted in following error:
> ../../gcc/gcc/fold-const.cc:10563:34: error: no matching function for
> call to ‘poly_int<2, long unsigned int>::is_constant(long int*)’
> 10563 |       if (!(a2 - a1).is_constant (&S))
>            |  ~~~~~~~~~~~~~~~~~~~~~~^~~~

OK:

       HOST_WIDE_INT step;
       if (!poly_int64 (a2 - a1).is_constant (&step))

> Could you please elaborate on the following issue, because it still
> stands in this patch,
> which is in Case 2 of test_fold_vec_perm_cst::test_stepped.
>
> For example,
> op0, op1: npatterns = 1, nelts_per_pattern = 3
> op0_len = op1_len = 4 + 4x
>
> sel: { 4 + 4x, 5 + 4x, 6 + 4x, ... } // should choose op1
>
> In this case,
> a1 = 5 + 4x
> S = (6 + 4x) - (5 + 4x) = 1
> Esel = 4 + 4x
>
> ae = a1 + (esel - 2) * S
>      = (5 + 4x) + (4 + 4x - 2) * 1
>      = 7 + 8x
>
> IIUC, 7 + 8x will always be index for last element of op1 ?
> if x = 0, len = 4, 7 + 8x = 7
> if x = 1, len = 8, 7 + 8x = 15, etc.
> So the stepped sequence will always choose elements
> from op1 regardless of vector length for above case ?
>
> However,
> ae /trunc op0_len
> = (7 + 8x) / (4 + 4x)
> which is not defined because 7/4 != 8/4

Ah, that was me being lazy and not handling a boundary case that
seemed difficult to keep overflow-free.  I've now pushed a fix.

Thanks,
Richard