Re: [PATCH 1/2] Missed opportunity to use [SU]ABD

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

Oluwatamilore Adebayo <oluwatamilore.adebayo@arm.com> writes:
> From: oluade01 <oluwatamilore.adebayo@arm.com>
>
> This adds a recognition pattern for the non-widening
> absolute difference (ABD).
>
> gcc/ChangeLog:
>
> 	* doc/md.texi (sabd, uabd): Document them.
> 	* internal-fn.def (ABD): Use new optab.
> 	* optabs.def (sabd_optab, uabd_optab): New optabs,
> 	* tree-vect-patterns.cc (vect_recog_absolute_difference):
> 	Recognize the following idiom abs (a - b).
> 	(vect_recog_sad_pattern): Refactor to use
> 	vect_recog_absolute_difference.
> 	(vect_recog_abd_pattern): Use patterns found by
> 	vect_recog_absolute_difference to build a new ABD
> 	internal call.
> ---
>  gcc/doc/md.texi           |  10 ++
>  gcc/internal-fn.def       |   3 +
>  gcc/optabs.def            |   2 +
>  gcc/tree-vect-patterns.cc | 245 +++++++++++++++++++++++++++++++++-----
>  4 files changed, 230 insertions(+), 30 deletions(-)
>
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index 6a435eb44610960513e9739ac9ac1e8a27182c10..e11b10d2fca11016232921bc85e47975f700e6c6 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -5787,6 +5787,16 @@ Other shift and rotate instructions, analogous to the
>  Vector shift and rotate instructions that take vectors as operand 2
>  instead of a scalar type.
>  
> +@cindex @code{uabd@var{m}} instruction pattern
> +@cindex @code{sabd@var{m}} instruction pattern
> +@item @samp{uabd@var{m}}, @samp{sabd@var{m}}
> +Signed and unsigned absolute difference instructions.  These
> +instructions find the difference between operands 1 and 2
> +then return the absolute value.  A C code equivalent would be:
> +@smallexample
> +op0 = op1 > op2 ? op1 - op2 : op2 - op1;
> +@end smallexample
> +
>  @cindex @code{avg@var{m}3_floor} instruction pattern
>  @cindex @code{uavg@var{m}3_floor} instruction pattern
>  @item @samp{avg@var{m}3_floor}
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index 3ac9d82aace322bd8ef108596e5583daa18c76e3..116965f4830cec8f60642ff011a86b6562e2c509 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -191,6 +191,9 @@ DEF_INTERNAL_OPTAB_FN (FMS, ECF_CONST, fms, ternary)
>  DEF_INTERNAL_OPTAB_FN (FNMA, ECF_CONST, fnma, ternary)
>  DEF_INTERNAL_OPTAB_FN (FNMS, ECF_CONST, fnms, ternary)
>  
> +DEF_INTERNAL_SIGNED_OPTAB_FN (ABD, ECF_CONST | ECF_NOTHROW, first,
> +			      sabd, uabd, binary)
> +
>  DEF_INTERNAL_SIGNED_OPTAB_FN (AVG_FLOOR, ECF_CONST | ECF_NOTHROW, first,
>  			      savg_floor, uavg_floor, binary)
>  DEF_INTERNAL_SIGNED_OPTAB_FN (AVG_CEIL, ECF_CONST | ECF_NOTHROW, first,
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 6c064ff4993620067d38742a0bfe0a3efb511069..35b835a6ac56d72417dac8ddfd77a8a7e2475e65 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -359,6 +359,8 @@ OPTAB_D (mask_fold_left_plus_optab, "mask_fold_left_plus_$a")
>  OPTAB_D (extract_last_optab, "extract_last_$a")
>  OPTAB_D (fold_extract_last_optab, "fold_extract_last_$a")
>  
> +OPTAB_D (uabd_optab, "uabd$a3")
> +OPTAB_D (sabd_optab, "sabd$a3")
>  OPTAB_D (savg_floor_optab, "avg$a3_floor")
>  OPTAB_D (uavg_floor_optab, "uavg$a3_floor")
>  OPTAB_D (savg_ceil_optab, "avg$a3_ceil")
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index dc102c919352a0328cf86eabceb3a38c41a7e4fd..4b63febc33e90b3caa854404a241afb2f09d755e 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -782,6 +782,86 @@ vect_split_statement (vec_info *vinfo, stmt_vec_info stmt2_info, tree new_rhs,
>      }
>  }
>  
> +/* Look for the following pattern
> +	X = x[i]
> +	Y = y[i]
> +	DIFF = X - Y
> +	DAD = ABS_EXPR<DIFF>
> +
> +   ABS_STMT should point to a statement of code ABS_EXPR or ABSU_EXPR.
> +   HALF_TYPE and UNPROM will be set should the statement be found to
> +   be a widened operation.
> +   DIFF_STMT will be set to the MINUS_EXPR
> +   statement that precedes the ABS_STMT unless vect_widened_op_tree
> +   succeeds.
> + */
> +static bool
> +vect_recog_absolute_difference (vec_info *vinfo, gassign *abs_stmt,
> +				tree *half_type,
> +				vect_unpromoted_value unprom[2],
> +				gassign **diff_stmt)
> +{
> +  if (!abs_stmt)
> +    return false;
> +
> +  /* FORNOW.  Can continue analyzing the def-use chain when this stmt in a phi
> +     inside the loop (in case we are analyzing an outer-loop).  */
> +  enum tree_code code = gimple_assign_rhs_code (abs_stmt);
> +  if (code != ABS_EXPR && code != ABSU_EXPR)
> +    return false;
> +
> +  tree abs_oprnd = gimple_assign_rhs1 (abs_stmt);
> +  tree abs_type = TREE_TYPE (abs_oprnd);
> +  if (!abs_oprnd)
> +    return false;
> +  if (!ANY_INTEGRAL_TYPE_P (abs_type)
> +      || TYPE_OVERFLOW_WRAPS (abs_type)
> +      || TYPE_UNSIGNED (abs_type))
> +    return false;
> +
> +  /* Peel off conversions from the ABS input.  This can involve sign
> +     changes (e.g. from an unsigned subtraction to a signed ABS input)
> +     or signed promotion, but it can't include unsigned promotion.
> +     (Note that ABS of an unsigned promotion should have been folded
> +     away before now anyway.)  */
> +  vect_unpromoted_value unprom_diff;
> +  abs_oprnd = vect_look_through_possible_promotion (vinfo, abs_oprnd,
> +						    &unprom_diff);
> +  if (!abs_oprnd)
> +    return false;
> +  if (TYPE_PRECISION (unprom_diff.type) != TYPE_PRECISION (abs_type)
> +      && TYPE_UNSIGNED (unprom_diff.type))
> +    return false;
> +
> +  /* We then detect if the operand of abs_expr is defined by a minus_expr.  */
> +  stmt_vec_info diff_stmt_vinfo = vect_get_internal_def (vinfo, abs_oprnd);
> +  if (!diff_stmt_vinfo)
> +    return false;
> +
> +  /* FORNOW.  Can continue analyzing the def-use chain when this stmt in a phi
> +     inside the loop (in case we are analyzing an outer-loop).  */
> +  if (vect_widened_op_tree (vinfo, diff_stmt_vinfo,
> +			    MINUS_EXPR, IFN_VEC_WIDEN_MINUS,
> +			    false, 2, unprom, half_type))
> +    return true;
> +
> +  /* Failed to find a widen operation so we check for a regular MINUS_EXPR.  */
> +  gassign *diff = dyn_cast <gassign *> (STMT_VINFO_STMT (diff_stmt_vinfo));
> +  if (!diff || gimple_assign_rhs_code (diff) != MINUS_EXPR)
> +    return false;
> +
> +  if (diff_stmt)
> +    *diff_stmt = diff;
> +
> +  if (TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (abs_oprnd)))
> +    {
> +      *half_type = NULL_TREE;
> +      return true;
> +    }

An earlier version had:

+  if (diff_oprnds && diff
+      && gimple_assign_rhs_code (diff) == MINUS_EXPR)

which IMO correctly made this path conditional on whether the caller
was prepared to accept a MINUS_EXPR on unwidened inputs.  The code
above instead returns true for such a MINUS_EXPR even if diff_stmt
is null.

So I think this should be:

  /* Failed to find a widen operation so we check for a regular MINUS_EXPR.  */
  gassign *diff = dyn_cast <gassign *> (STMT_VINFO_STMT (diff_stmt_vinfo));
  if (diff_stmt
      && diff
      && gimple_assign_rhs_code (diff) == MINUS_EXPR
      && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (abs_oprnd)))
    {
      *diff_stmt = diff;
      *half_type = NULL_TREE;
      return true;
    }

With that change, there would be no need for a separate half_type
check...

> +
> +  return false;
> +}
> +
>  /* Convert UNPROM to TYPE and return the result, adding new statements
>     to STMT_INFO's pattern definition statements if no better way is
>     available.  VECTYPE is the vector form of TYPE.
> @@ -1320,41 +1400,31 @@ vect_recog_sad_pattern (vec_info *vinfo,
>    /* FORNOW.  Can continue analyzing the def-use chain when this stmt in a phi
>       inside the loop (in case we are analyzing an outer-loop).  */
>    gassign *abs_stmt = dyn_cast <gassign *> (abs_stmt_vinfo->stmt);
> -  if (!abs_stmt
> -      || (gimple_assign_rhs_code (abs_stmt) != ABS_EXPR
> -	  && gimple_assign_rhs_code (abs_stmt) != ABSU_EXPR))
> -    return NULL;
> +  vect_unpromoted_value unprom[2];
>  
> -  tree abs_oprnd = gimple_assign_rhs1 (abs_stmt);
> -  tree abs_type = TREE_TYPE (abs_oprnd);
> -  if (TYPE_UNSIGNED (abs_type))
> -    return NULL;
> +  if (!abs_stmt)
> +    {
> +      gcall *abd_stmt = dyn_cast <gcall *> (abs_stmt_vinfo->stmt);
> +      if (!abd_stmt
> +	  || !gimple_call_internal_p (abd_stmt)
> +	  || gimple_call_internal_fn (abd_stmt) != IFN_ABD)
> +	return NULL;
>  
> -  /* Peel off conversions from the ABS input.  This can involve sign
> -     changes (e.g. from an unsigned subtraction to a signed ABS input)
> -     or signed promotion, but it can't include unsigned promotion.
> -     (Note that ABS of an unsigned promotion should have been folded
> -     away before now anyway.)  */
> -  vect_unpromoted_value unprom_diff;
> -  abs_oprnd = vect_look_through_possible_promotion (vinfo, abs_oprnd,
> -						    &unprom_diff);
> -  if (!abs_oprnd)
> -    return NULL;
> -  if (TYPE_PRECISION (unprom_diff.type) != TYPE_PRECISION (abs_type)
> -      && TYPE_UNSIGNED (unprom_diff.type))
> -    return NULL;
> +      tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
> +      tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
>  
> -  /* We then detect if the operand of abs_expr is defined by a minus_expr.  */
> -  stmt_vec_info diff_stmt_vinfo = vect_get_internal_def (vinfo, abs_oprnd);
> -  if (!diff_stmt_vinfo)
> +      if (!vect_look_through_possible_promotion (vinfo, abd_oprnd0, &unprom[0])
> +	  || !vect_look_through_possible_promotion (vinfo, abd_oprnd1,
> +						    &unprom[1]))
> +	return NULL;
> +
> +      half_type = unprom[0].type;
> +    }
> +  else if (!vect_recog_absolute_difference (vinfo, abs_stmt, &half_type,
> +					    unprom, NULL))
>      return NULL;
>  
> -  /* FORNOW.  Can continue analyzing the def-use chain when this stmt in a phi
> -     inside the loop (in case we are analyzing an outer-loop).  */
> -  vect_unpromoted_value unprom[2];
> -  if (!vect_widened_op_tree (vinfo, diff_stmt_vinfo, MINUS_EXPR,
> -			     IFN_VEC_WIDEN_MINUS,
> -			     false, 2, unprom, &half_type))
> +  if (!half_type)
>      return NULL;

...here.

>  
>    vect_pattern_detected ("vect_recog_sad_pattern", last_stmt);
> @@ -1376,6 +1446,120 @@ vect_recog_sad_pattern (vec_info *vinfo,
>    return pattern_stmt;
>  }
>  
> +/* Function vect_recog_abd_pattern
> +
> +   Try to find the following ABsolute Difference (ABD) pattern:
> +
> +     VTYPE x, y, out;
> +     type diff;
> +   loop i in range:
> +     S1 diff = x[i] - y[i]
> +     S2 out[i] = ABS_EXPR <diff>;
> +
> +   where 'type' is a integer and 'VTYPE' is a vector of integers
> +   the same size as 'type'
> +
> +   Input:
> +
> +   * STMT_VINFO: The stmt from which the pattern search begins
> +
> +   Output:
> +
> +   * TYPE_out: The type of the output of this pattern
> +
> +   * Return value: A new stmt that will be used to replace the sequence of
> +     stmts that constitute the pattern; either SABD or UABD:
> +	SABD_EXPR<x, y, out>
> +	UABD_EXPR<x, y, out>
> + */
> +
> +static gimple *
> +vect_recog_abd_pattern (vec_info *vinfo,
> +			stmt_vec_info stmt_vinfo, tree *type_out)
> +{
> +  /* Look for the following patterns
> +	X = x[i]
> +	Y = y[i]
> +	DIFF = X - Y
> +	DAD = ABS_EXPR<DIFF>
> +	out[i] = DAD
> +
> +     In which
> +      - X, Y, DIFF, DAD all have the same type
> +      - x, y, out are all vectors of the same type
> +  */
> +
> +  gassign *last_stmt = dyn_cast <gassign *> (STMT_VINFO_STMT (stmt_vinfo));
> +  if (!last_stmt)
> +    return NULL;
> +
> +  tree out_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
> +
> +  vect_unpromoted_value unprom[2];
> +  gassign *diff_stmt;
> +  tree half_type;
> +  if (!vect_recog_absolute_difference (vinfo, last_stmt, &half_type,
> +				       unprom, &diff_stmt))
> +    return NULL;
> +
> +  tree abd_type = out_type, vectype;
> +  tree abd_oprnds[2];
> +  bool extend = false;
> +  if (half_type)
> +    {
> +      vectype = get_vectype_for_scalar_type (vinfo, half_type);
> +      if (!vectype)
> +	return NULL;
> +
> +      abd_type = half_type;
> +      vect_convert_inputs (vinfo, stmt_vinfo, 2, abd_oprnds,
> +			   half_type, unprom, vectype);
> +
> +      extend = TYPE_PRECISION (abd_type) < TYPE_PRECISION (out_type);
> +    }
> +  else
> +    {
> +      unprom[0].op = gimple_assign_rhs1 (diff_stmt);
> +      unprom[1].op = gimple_assign_rhs2 (diff_stmt);
> +      tree signed_out = signed_type_for (out_type);
> +      vectype = get_vectype_for_scalar_type (vinfo, signed_out);
> +      if (!vectype)
> +	return NULL;
> +
> +      vect_convert_inputs (vinfo, stmt_vinfo, 2, abd_oprnds,
> +			   signed_out, unprom, vectype);
> +    }

We should postpone doing the vect_convert_inputs until...

> +
> +  vect_pattern_detected ("vect_recog_abd_pattern", last_stmt);
> +
> +  if (!vectype
> +      || !direct_internal_fn_supported_p (IFN_ABD, vectype,
> +					  OPTIMIZE_FOR_SPEED))
> +    return NULL;

...here.  The following should work for both cases:

      vect_convert_inputs (vinfo, stmt_vinfo, 2, abd_oprnds,
			   TREE_TYPE (vectype), unprom, vectype);

LGTM otherwise, thanks.

Richard

> +
> +  *type_out = get_vectype_for_scalar_type (vinfo, out_type);
> +
> +  tree abd_result = vect_recog_temp_ssa_var (abd_type, NULL);
> +  gcall *abd_stmt = gimple_build_call_internal (IFN_ABD, 2,
> +						abd_oprnds[0], abd_oprnds[1]);
> +  gimple_call_set_lhs (abd_stmt, abd_result);
> +  gimple_set_location (abd_stmt, gimple_location (last_stmt));
> +
> +  if (!extend)
> +    return abd_stmt;
> +
> +  gimple *stmt = abd_stmt;
> +  if (!TYPE_UNSIGNED (abd_type))
> +    {
> +      tree unsign = unsigned_type_for (abd_type);
> +      tree unsign_vectype = get_vectype_for_scalar_type (vinfo, unsign);
> +      stmt = vect_convert_output (vinfo, stmt_vinfo, unsign, stmt,
> +				  unsign_vectype);
> +    }
> +
> +  return vect_convert_output (vinfo, stmt_vinfo, out_type, stmt, vectype);
> +}
> +
>  /* Recognize an operation that performs ORIG_CODE on widened inputs,
>     so that it can be treated as though it had the form:
>  
> @@ -6471,6 +6655,7 @@ struct vect_recog_func
>  static vect_recog_func vect_vect_recog_func_ptrs[] = {
>    { vect_recog_bitfield_ref_pattern, "bitfield_ref" },
>    { vect_recog_bit_insert_pattern, "bit_insert" },
> +  { vect_recog_abd_pattern, "abd" },
>    { vect_recog_over_widening_pattern, "over_widening" },
>    /* Must come after over_widening, which narrows the shift as much as
>       possible beforehand.  */