Re: [PATCH 1/2] Mid engine setup [SU]ABDL

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

Oluwatamilore Adebayo <oluwatamilore.adebayo@arm.com> writes:
> From: oluade01 <oluwatamilore.adebayo@arm.com>
>
> This updates vect_recog_abd_pattern to recognize the widening
> variant of absolute difference (ABDL, ABDL2).
>
> gcc/ChangeLog:
>
> 	* internal-fn.cc (widening_fn_p, decomposes_to_hilo_fn_p):
> 	Add IFN_VEC_WIDEN_ABD to the switch statement.
> 	* internal-fn.def (VEC_WIDEN_ABD): New internal hilo optab.
> 	* optabs.def (vec_widen_sabd_optab,
> 	vec_widen_sabd_hi_optab, vec_widen_sabd_lo_optab,
> 	vec_widen_sabd_odd_even, vec_widen_sabd_even_optab,
> 	vec_widen_uabd_optab,
> 	vec_widen_uabd_hi_optab, vec_widen_uabd_lo_optab,
> 	vec_widen_uabd_odd_even, vec_widen_uabd_even_optab):
> 	New optabs.
> 	* tree-vect-patterns.cc (vect_recog_abd_pattern): Update to
> 	to build a VEC_WIDEN_ABD call if the input precision is smaller
> 	than the precision of the output.
> 	(vect_recog_widen_abd_pattern): Should an ABD expression be
> 	found preceeding an extension, replace the two with a
> 	VEC_WIDEN_ABD.
> ---
>  gcc/doc/md.texi           |  11 ++
>  gcc/internal-fn.def       |   5 +
>  gcc/optabs.def            |  10 ++
>  gcc/tree-vect-patterns.cc | 205 +++++++++++++++++++++++++++++---------
>  4 files changed, 183 insertions(+), 48 deletions(-)
>
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index e11b10d2fca11016232921bc85e47975f700e6c6..2ae6182b925d0cf8950dc830d083cf93baf2eaa1 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -5617,6 +5617,17 @@ signed/unsigned elements of size S@.  Subtract the high/low elements of 2 from
>  1 and widen the resulting elements. Put the N/2 results of size 2*S in the
>  output vector (operand 0).
>  
> +@cindex @code{vec_widen_sabdl_hi_@var{m}} instruction pattern
> +@cindex @code{vec_widen_sabdl_lo_@var{m}} instruction pattern
> +@cindex @code{vec_widen_uabdl_hi_@var{m}} instruction pattern
> +@cindex @code{vec_widen_uabdl_lo_@var{m}} instruction pattern
> +@item @samp{vec_widen_uabdl_hi_@var{m}}, @samp{vec_widen_uabdl_lo_@var{m}}
> +@itemx @samp{vec_widen_sabdl_hi_@var{m}}, @samp{vec_widen_sabdl_lo_@var{m}}

The optabs don't have the trailing “l” (long).  (Which is a good thing!)

The list should include the even/odd patterns as well.

> +Signed/Unsigned widening absolute difference long.  Operands 1 and 2 are

Similarly no “long” here.

> +vectors with N signed/unsigned elements of size S@.  Find the absolute
> +difference between 1 and 2 and widen the resulting elements.  Put the N/2

Maybe “operands 1 and 2”, or just “them”.

> +results of size 2*S in the output vector (operand 0).
> +
>  @cindex @code{vec_addsub@var{m}3} instruction pattern
>  @item @samp{vec_addsub@var{m}3}
>  Alternating subtract, add with even lanes doing subtract and odd
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index 116965f4830cec8f60642ff011a86b6562e2c509..d67274d68b49943a88c531e903fd03b42343ab97 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -352,6 +352,11 @@ DEF_INTERNAL_WIDENING_OPTAB_FN (VEC_WIDEN_MINUS,
>  				first,
>  				vec_widen_ssub, vec_widen_usub,
>  				binary)
> +DEF_INTERNAL_WIDENING_OPTAB_FN (VEC_WIDEN_ABD,
> +				ECF_CONST | ECF_NOTHROW,
> +				first,
> +				vec_widen_sabd, vec_widen_uabd,
> +				binary)
>  DEF_INTERNAL_OPTAB_FN (VEC_FMADDSUB, ECF_CONST, vec_fmaddsub, ternary)
>  DEF_INTERNAL_OPTAB_FN (VEC_FMSUBADD, ECF_CONST, vec_fmsubadd, ternary)
>  
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 35b835a6ac56d72417dac8ddfd77a8a7e2475e65..68dfa1550f791a2fe833012157601ecfa68f1e09 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -418,6 +418,11 @@ OPTAB_D (vec_widen_sadd_hi_optab, "vec_widen_sadd_hi_$a")
>  OPTAB_D (vec_widen_sadd_lo_optab, "vec_widen_sadd_lo_$a")
>  OPTAB_D (vec_widen_sadd_odd_optab, "vec_widen_sadd_odd_$a")
>  OPTAB_D (vec_widen_sadd_even_optab, "vec_widen_sadd_even_$a")
> +OPTAB_D (vec_widen_sabd_optab, "vec_widen_sabd_$a")
> +OPTAB_D (vec_widen_sabd_hi_optab, "vec_widen_sabd_hi_$a")
> +OPTAB_D (vec_widen_sabd_lo_optab, "vec_widen_sabd_lo_$a")
> +OPTAB_D (vec_widen_sabd_odd_optab, "vec_widen_sabd_odd_$a")
> +OPTAB_D (vec_widen_sabd_even_optab, "vec_widen_sabd_even_$a")
>  OPTAB_D (vec_widen_sshiftl_hi_optab, "vec_widen_sshiftl_hi_$a")
>  OPTAB_D (vec_widen_sshiftl_lo_optab, "vec_widen_sshiftl_lo_$a")
>  OPTAB_D (vec_widen_umult_even_optab, "vec_widen_umult_even_$a")
> @@ -436,6 +441,11 @@ OPTAB_D (vec_widen_uadd_hi_optab, "vec_widen_uadd_hi_$a")
>  OPTAB_D (vec_widen_uadd_lo_optab, "vec_widen_uadd_lo_$a")
>  OPTAB_D (vec_widen_uadd_odd_optab, "vec_widen_uadd_odd_$a")
>  OPTAB_D (vec_widen_uadd_even_optab, "vec_widen_uadd_even_$a")
> +OPTAB_D (vec_widen_uabd_optab, "vec_widen_uabd_$a")
> +OPTAB_D (vec_widen_uabd_hi_optab, "vec_widen_uabd_hi_$a")
> +OPTAB_D (vec_widen_uabd_lo_optab, "vec_widen_uabd_lo_$a")
> +OPTAB_D (vec_widen_uabd_odd_optab, "vec_widen_uabd_odd_$a")
> +OPTAB_D (vec_widen_uabd_even_optab, "vec_widen_uabd_even_$a")
>  OPTAB_D (vec_addsub_optab, "vec_addsub$a3")
>  OPTAB_D (vec_fmaddsub_optab, "vec_fmaddsub$a4")
>  OPTAB_D (vec_fmsubadd_optab, "vec_fmsubadd$a4")
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index e2392113bff4065c909aefc760b4c48978b73a5a..281d7bc2e9945ee415be051f5ec1cce19251fbbf 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -1404,15 +1404,28 @@ vect_recog_sad_pattern (vec_info *vinfo,
>        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)
> +	  || gimple_call_num_args (abd_stmt) != 2)
>  	return NULL;
>  
>        tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
>        tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
>  
> -      if (!vect_look_through_possible_promotion (vinfo, abd_oprnd0, &unprom[0])
> -	  || !vect_look_through_possible_promotion (vinfo, abd_oprnd1,
> -						    &unprom[1]))
> +      if (gimple_call_internal_fn (abd_stmt) == IFN_ABD)
> +	{
> +	  if (!vect_look_through_possible_promotion (vinfo, abd_oprnd0,
> +						     &unprom[0])
> +	      || !vect_look_through_possible_promotion (vinfo, abd_oprnd1,
> +							&unprom[1]))
> +	    return NULL;
> +	}
> +      else if (gimple_call_internal_fn (abd_stmt) == IFN_VEC_WIDEN_ABD)
> +	{
> +	  unprom[0].op = abd_oprnd0;
> +	  unprom[0].type = TREE_TYPE (abd_oprnd0);
> +	  unprom[1].op = abd_oprnd1;
> +	  unprom[1].type = TREE_TYPE (abd_oprnd1);
> +	}
> +      else
>  	return NULL;
>  
>        half_type = unprom[0].type;
> @@ -1442,16 +1455,19 @@ vect_recog_sad_pattern (vec_info *vinfo,
>  
>  /* Function vect_recog_abd_pattern
>  
> -   Try to find the following ABsolute Difference (ABD) pattern:
> +   Try to find the following ABsolute Difference (ABD) or
> +   widening ABD (WIDEN_ABD) pattern:
>  
> -     VTYPE x, y, out;
> -     type diff;
> -   loop i in range:
> -     S1 diff = x[i] - y[i]
> -     S2 out[i] = ABS_EXPR <diff>;
> +   TYPE1 x;
> +   TYPE2 y;
> +   TYPE3 x_cast = (TYPE3) x;		  // widening or no-op
> +   TYPE3 y_cast = (TYPE3) y;		  // widening or no-op
> +   TYPE3 diff = x_cast - y_cast;
> +   TYPE4 diff_cast = (TYPE4) diff;	  // widening or no-op
> +   TYPE5 abs = ABS(U)_EXPR <diff_cast>;
>  
> -   where 'type' is a integer and 'VTYPE' is a vector of integers
> -   the same size as 'type'
> +   WIDEN_ABD exists to optimize the case where WTYPE is at least
> +   twice as wide as VTYPE.

Sorry, my fault.  I was using the original type names in this
suggestion, rather than the TYPE1…TYPE5 ones.  Should be:

   WIDEN_ABD exists to optimize the case where TYPE4 is at least
   twice as wide as TYPE3.

>  
>     Input:
>  
> @@ -1459,30 +1475,18 @@ vect_recog_sad_pattern (vec_info *vinfo,
>  
>     Output:
>  
> -   * TYPE_out: The type of the output of this pattern
> +   * 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>
> +     stmts that constitute the pattern; either SABD, UABD, SABDL or UABDL:
> +	IFN_ABD<x, y, out>
> +	IFN_WIDEN_ABD<x, y, out>

Lingering use of “L” suffixes here.  Maybe:

     stmts that constitute the pattern, principally:
	out = IFN_ABD (x, y)
	out = IFN_WIDEN_ABD (x, y)
>   */
>  
>  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;
> @@ -1496,54 +1500,83 @@ vect_recog_abd_pattern (vec_info *vinfo,
>  				       unprom, &diff_stmt))
>      return NULL;
>  
> -  tree abd_type = out_type, vectype;
> -  tree abd_oprnds[2];
> -  bool extend = false;
> +  tree abd_in_type, abd_out_type;
> +
>    if (half_type)
>      {
> -      vectype = get_vectype_for_scalar_type (vinfo, half_type);
> -      abd_type = half_type;
> -      extend = TYPE_PRECISION (abd_type) < TYPE_PRECISION (out_type);
> +      abd_in_type = half_type;
> +      abd_out_type = abd_in_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);
> +      abd_in_type = signed_type_for (out_type);
> +      abd_out_type = abd_in_type;
>      }
>  
> -  vect_pattern_detected ("vect_recog_abd_pattern", last_stmt);
> +  tree vectype_in = get_vectype_for_scalar_type (vinfo, abd_in_type);
> +  if (!vectype_in)
> +    return NULL;
>  
> -  if (!vectype
> -      || !direct_internal_fn_supported_p (IFN_ABD, vectype,
> +  internal_fn ifn = IFN_ABD;
> +  tree vectype_out = vectype_in;
> +
> +  if (TYPE_PRECISION (out_type) >= TYPE_PRECISION (abd_in_type) * 2
> +      && TYPE_PRECISION (abd_out_type) != stmt_vinfo->min_output_precision)

Sorry for not noticing last time, but I think the second condition
would be more natural as:

      && stmt_vinfo->min_output_precision >= TYPE_PRECISION (abd_in_type) * 2)

(There's no distinction between abs_in_type and abs_out_type at this point,
so it seems clearer to use the same value in both conditions.)

> +    {
> +      tree mid_type
> +	= build_nonstandard_integer_type (TYPE_PRECISION (abd_in_type) * 2,
> +					  TYPE_UNSIGNED (abd_in_type));
> +      tree mid_vectype = get_vectype_for_scalar_type (vinfo, mid_type);
> +
> +      code_helper dummy_code;
> +      int dummy_int;
> +      auto_vec<tree> dummy_vec;
> +      if (mid_vectype
> +	  && supportable_widening_operation (vinfo, IFN_VEC_WIDEN_ABD,
> +					     stmt_vinfo, mid_vectype,
> +					     vectype_in,
> +					     &dummy_code, &dummy_code,
> +					     &dummy_int, &dummy_vec))
> +	{
> +	  ifn = IFN_VEC_WIDEN_ABD;
> +	  abd_out_type = mid_type;
> +	  vectype_out = mid_vectype;
> +	}
> +    }
> +
> +  if (ifn == IFN_ABD
> +      && !direct_internal_fn_supported_p (ifn, vectype_in,
>  					  OPTIMIZE_FOR_SPEED))
>      return NULL;
>  
> +  vect_pattern_detected ("vect_recog_abd_pattern", last_stmt);
> +
> +  tree abd_oprnds[2];
>    vect_convert_inputs (vinfo, stmt_vinfo, 2, abd_oprnds,
> -		       TREE_TYPE (vectype), unprom, vectype);
> +		       abd_in_type, unprom, vectype_in);
>  
>    *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,
> +  tree abd_result = vect_recog_temp_ssa_var (abd_out_type, NULL);
> +  gcall *abd_stmt = gimple_build_call_internal (ifn, 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))
> +  if (TYPE_PRECISION (abd_in_type) == TYPE_PRECISION (abd_out_type)
> +      && TYPE_PRECISION (abd_out_type) < TYPE_PRECISION (out_type)
> +      && !TYPE_UNSIGNED (abd_out_type))
>      {
> -      tree unsign = unsigned_type_for (abd_type);
> +      tree unsign = unsigned_type_for (abd_out_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);
> +  return vect_convert_output (vinfo, stmt_vinfo, out_type, stmt, vectype_out);
>  }
>  
>  /* Recognize an operation that performs ORIG_CODE on widened inputs,
> @@ -1703,6 +1736,81 @@ vect_recog_widen_minus_pattern (vec_info *vinfo, stmt_vec_info last_stmt_info,
>  				      &subtype);
>  }
>  
> +/* Try to detect abd on widened inputs, converting IFN_ABD
> +   to IFN_VEC_WIDEN_ABD.  */
> +static gimple *
> +vect_recog_widen_abd_pattern (vec_info *vinfo, stmt_vec_info stmt_vinfo,
> +			      tree *type_out)
> +{
> +  gassign *last_stmt = dyn_cast <gassign *> (STMT_VINFO_STMT (stmt_vinfo));
> +  if (!last_stmt || !gimple_assign_cast_p (last_stmt))

I think this should be:

  if (!last_stmt || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (last_stmt)))

gimple_assign_cast_p is more general, and allows conversions
between integral and non-integral types.

> +    return NULL;
> +
> +  tree last_rhs = gimple_assign_rhs1 (last_stmt);
> +
> +  tree in_type = TREE_TYPE (last_rhs);
> +  tree out_type = TREE_TYPE (gimple_assign_lhs (last_stmt));
> +  if (TYPE_PRECISION (in_type) * 2 != TYPE_PRECISION (out_type))
> +    return NULL;

I think this also needs to require TYPE_UNSIGNED (in_type):

  if (TYPE_PRECISION (in_type) * 2 != TYPE_PRECISION (out_type)
      || !TYPE_UNSIGNED (in_type))
    return NULL;

That is, the extension has to be a zero extension rather than
a sign extension.

For example:

  int32_t a, b, c;
  int64_t d;

  c = IFN_ABD (a, b);
  d = (int64_t) c;

sign-extends the ABD result to 64 bits, and so a == INT_MAX
&& b == INT_MIN gives:

  c = -1    (UINT_MAX converted to signed)
  d = -1

But IFN_WIDEN_ABD would give d == UINT_MAX instead.

> +
> +  stmt_vec_info abs_vinfo = vect_get_internal_def (vinfo, last_rhs);
> +  if (!abs_vinfo)
> +    return NULL;
> +
> +  stmt_vec_info abd_pattern_vinfo = STMT_VINFO_RELATED_STMT (abs_vinfo);
> +  if (!abd_pattern_vinfo)
> +    return NULL;
> +
> +  gimple *pattern_stmt = STMT_VINFO_STMT (abd_pattern_vinfo);
> +  if (gimple_assign_cast_p (pattern_stmt))
> +    {
> +       tree op = gimple_assign_rhs1 (pattern_stmt);
> +       vect_unpromoted_value unprom;
> +       op = vect_look_through_possible_promotion (vinfo, op, &unprom);
> +
> +      if (!op)
> +	return NULL;
> +
> +      abd_pattern_vinfo = vect_get_internal_def (vinfo, op);
> +      if (!abd_pattern_vinfo)
> +	return NULL;
> +
> +      pattern_stmt = STMT_VINFO_STMT (abd_pattern_vinfo);
> +    }

I think the code quoted above reduces to:

  vect_unpromoted_value unprom;
  tree op = vect_look_through_possible_promotion (vinfo, last_rhs, &unprom);
  if (!op || TYPE_PRECISION (TREE_TYPE (op)) != TYPE_PRECISION (in_type))
    return NULL;

  stmt_vec_info abd_pattern_vinfo = vect_get_internal_def (vinfo, op);
  if (!abd_pattern_vinfo)
    return NULL;
  abd_pattern_vinfo = vect_stmt_to_vectorize (abd_pattern_vinfo);

> +
> +  gcall *abd_stmt = dyn_cast <gcall *> (pattern_stmt);
> +  if (!abd_stmt || gimple_call_internal_fn (abd_stmt) != IFN_ABD)
> +    return NULL;
> +
> +  tree abd_oprnd0 = gimple_call_arg (abd_stmt, 0);
> +  tree abd_oprnd1 = gimple_call_arg (abd_stmt, 1);
> +  if (TYPE_PRECISION (TREE_TYPE (abd_oprnd0)) != TYPE_PRECISION (in_type))
> +    return NULL;

With the changes above, this check would not be necessary.

LGTM otherwise, thanks.

Richard

> +
> +  tree vectype_in = get_vectype_for_scalar_type (vinfo, in_type);
> +  tree vectype_out = get_vectype_for_scalar_type (vinfo, out_type);
> +
> +  code_helper dummy_code;
> +  int dummy_int;
> +  auto_vec<tree> dummy_vec;
> +  if (!supportable_widening_operation (vinfo, IFN_VEC_WIDEN_ABD, stmt_vinfo,
> +				       vectype_out, vectype_in,
> +				       &dummy_code, &dummy_code,
> +				       &dummy_int, &dummy_vec))
> +    return NULL;
> +
> +  vect_pattern_detected ("vect_recog_widen_abd_pattern", last_stmt);
> +
> +  *type_out = vectype_out;
> +
> +  tree widen_abd_result = vect_recog_temp_ssa_var (out_type, NULL);
> +  gcall *widen_abd_stmt = gimple_build_call_internal (IFN_VEC_WIDEN_ABD, 2,
> +						      abd_oprnd0, abd_oprnd1);
> +  gimple_call_set_lhs (widen_abd_stmt, widen_abd_result);
> +  gimple_set_location (widen_abd_stmt, gimple_location (last_stmt));
> +  return widen_abd_stmt;
> +}
> +
>  /* Function vect_recog_ctz_ffs_pattern
>  
>     Try to find the following pattern:
> @@ -6670,6 +6778,7 @@ static vect_recog_func vect_vect_recog_func_ptrs[] = {
>    { vect_recog_mask_conversion_pattern, "mask_conversion" },
>    { vect_recog_widen_plus_pattern, "widen_plus" },
>    { vect_recog_widen_minus_pattern, "widen_minus" },
> +  { vect_recog_widen_abd_pattern, "widen_abd" },
>    /* These must come after the double widening ones.  */
>  };