RE: [PATCH 5/19]middle-end: Enable bit-field vectorization to work correctly when we're vectoring inside conds

[Richard Biener <rguenther@suse.de>]

On Mon, 10 Jul 2023, Tamar Christina wrote:

> > > -  *type_out = STMT_VINFO_VECTYPE (stmt_info);
> > > +  if (cond_cst)
> > > +    {
> > > +      append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
> > > +      pattern_stmt
> > > +	= gimple_build_cond (gimple_cond_code (cond_stmt),
> > > +			     gimple_get_lhs (pattern_stmt),
> > > +			     fold_convert (ret_type, cond_cst),
> > > +			     gimple_cond_true_label (cond_stmt),
> > > +			     gimple_cond_false_label (cond_stmt));
> > > +      *type_out = STMT_VINFO_VECTYPE (stmt_info);
> > 
> > is there any vectype set for a gcond?
> 
> No, because gconds can't be codegen'd yet, atm we must replace the original
> gcond when generating code.
> 
> However looking at the diff this code, don't think the else is needed here.
> Testing an updated patch.
> 
> > 
> > I must say the flow of the function is a bit convoluted now.  Is it possible to
> > factor out a helper so we can fully separate the gassign vs. gcond handling in
> > this function?
> 
> I am not sure, the only place the changes are are at the start (e.g. how we determine bf_stmt)
> and how we determine ret_type, and when determining shift_first for the single use case.
> 
> Now I can't move the ret_type anywhere as I need to decompose bf_stmt first.  And the shift_first
> can be simplified by moving it up into the part that determined bf_stmt, but then we walk the
> immediate uses even on cases where we early exit.  Which seems inefficient.
> 
> Then there's the final clause which just generates an additional gcond if the original statement was
> a gcond. But not sure that'll help, since it's just something done *in addition* to the normal assign.
> 
> So there doesn't seem to be enough, or big enough divergence to justify a split.   I have however made
> an attempt at cleaning it up a bit, is this one better?

Yeah, it is.
 
> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> 
> Ok for master?

OK.

Richard.

> Thanks,
> Tamar
> 
> gcc/ChangeLog:
> 
> 	* tree-vect-patterns.cc (vect_init_pattern_stmt): Copy STMT_VINFO_TYPE
> 	from original statement.
> 	(vect_recog_bitfield_ref_pattern): Support bitfields in gcond.
> 
> Co-Authored-By:  Andre Vieira <andre.simoesdiasvieira@arm.com>
> 
> --- inline copy of patch ---
> 
> diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> index 60bc9be6819af9bd28a81430869417965ba9d82d..b842f7d983405cd04f6760be7d91c1f55b30aac4 100644
> --- a/gcc/tree-vect-patterns.cc
> +++ b/gcc/tree-vect-patterns.cc
> @@ -128,6 +128,7 @@ vect_init_pattern_stmt (vec_info *vinfo, gimple *pattern_stmt,
>    STMT_VINFO_RELATED_STMT (pattern_stmt_info) = orig_stmt_info;
>    STMT_VINFO_DEF_TYPE (pattern_stmt_info)
>      = STMT_VINFO_DEF_TYPE (orig_stmt_info);
> +  STMT_VINFO_TYPE (pattern_stmt_info) = STMT_VINFO_TYPE (orig_stmt_info);
>    if (!STMT_VINFO_VECTYPE (pattern_stmt_info))
>      {
>        gcc_assert (!vectype
> @@ -2441,6 +2442,10 @@ vect_recog_widen_sum_pattern (vec_info *vinfo,
>     bf_value = BIT_FIELD_REF (container, bitsize, bitpos);
>     result = (type_out) bf_value;
>  
> +   or
> +
> +   if (BIT_FIELD_REF (container, bitsize, bitpos) `cmp` <constant>)
> +
>     where type_out is a non-bitfield type, that is to say, it's precision matches
>     2^(TYPE_SIZE(type_out) - (TYPE_UNSIGNED (type_out) ? 1 : 2)).
>  
> @@ -2450,6 +2455,10 @@ vect_recog_widen_sum_pattern (vec_info *vinfo,
>     here it starts with:
>     result = (type_out) bf_value;
>  
> +   or
> +
> +   if (BIT_FIELD_REF (container, bitsize, bitpos) `cmp` <constant>)
> +
>     Output:
>  
>     * TYPE_OUT: The vector type of the output of this pattern.
> @@ -2482,33 +2491,45 @@ vect_recog_widen_sum_pattern (vec_info *vinfo,
>  
>     The shifting is always optional depending on whether bitpos != 0.
>  
> +   When the original bitfield was inside a gcond then an new gcond is also
> +   generated with the newly `result` as the operand to the comparison.
> +
>  */
>  
>  static gimple *
>  vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
>  				 tree *type_out)
>  {
> -  gassign *first_stmt = dyn_cast <gassign *> (stmt_info->stmt);
> -
> -  if (!first_stmt)
> -    return NULL;
> -
> -  gassign *bf_stmt;
> -  if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (first_stmt))
> -      && TREE_CODE (gimple_assign_rhs1 (first_stmt)) == SSA_NAME)
> +  gimple *bf_stmt = NULL;
> +  tree lhs = NULL_TREE;
> +  tree ret_type = NULL_TREE;
> +  gimple *stmt = STMT_VINFO_STMT (stmt_info);
> +  if (gcond *cond_stmt = dyn_cast <gcond *> (stmt))
> +    {
> +      tree op = gimple_cond_lhs (cond_stmt);
> +      if (TREE_CODE (op) != SSA_NAME)
> +	return NULL;
> +      bf_stmt = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (op));
> +      if (TREE_CODE (gimple_cond_rhs (cond_stmt)) != INTEGER_CST)
> +	return NULL;
> +    }
> +  else if (is_gimple_assign (stmt)
> +	   && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
> +	   && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
>      {
> -      gimple *second_stmt
> -	= SSA_NAME_DEF_STMT (gimple_assign_rhs1 (first_stmt));
> +      gimple *second_stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
>        bf_stmt = dyn_cast <gassign *> (second_stmt);
> -      if (!bf_stmt
> -	  || gimple_assign_rhs_code (bf_stmt) != BIT_FIELD_REF)
> -	return NULL;
> +      lhs = gimple_assign_lhs (stmt);
> +      ret_type = TREE_TYPE (lhs);
>      }
> -  else
> +
> +  if (!bf_stmt
> +      || gimple_assign_rhs_code (bf_stmt) != BIT_FIELD_REF)
>      return NULL;
>  
>    tree bf_ref = gimple_assign_rhs1 (bf_stmt);
>    tree container = TREE_OPERAND (bf_ref, 0);
> +  ret_type = ret_type ? ret_type : TREE_TYPE (container);
>  
>    if (!bit_field_offset (bf_ref).is_constant ()
>        || !bit_field_size (bf_ref).is_constant ()
> @@ -2522,8 +2543,6 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
>  
>    gimple *use_stmt, *pattern_stmt;
>    use_operand_p use_p;
> -  tree ret = gimple_assign_lhs (first_stmt);
> -  tree ret_type = TREE_TYPE (ret);
>    bool shift_first = true;
>    tree container_type = TREE_TYPE (container);
>    tree vectype = get_vectype_for_scalar_type (vinfo, container_type);
> @@ -2560,7 +2579,7 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
>    /* If the only use of the result of this BIT_FIELD_REF + CONVERT is a
>       PLUS_EXPR then do the shift last as some targets can combine the shift and
>       add into a single instruction.  */
> -  if (single_imm_use (gimple_assign_lhs (first_stmt), &use_p, &use_stmt))
> +  if (lhs && single_imm_use (lhs, &use_p, &use_stmt))
>      {
>        if (gimple_code (use_stmt) == GIMPLE_ASSIGN
>  	  && gimple_assign_rhs_code (use_stmt) == PLUS_EXPR)
> @@ -2620,6 +2639,19 @@ vect_recog_bitfield_ref_pattern (vec_info *vinfo, stmt_vec_info stmt_info,
>  			       NOP_EXPR, result);
>      }
>  
> +  if (!lhs)
> +    {
> +      append_pattern_def_seq (vinfo, stmt_info, pattern_stmt, vectype);
> +      gcond *cond_stmt = dyn_cast <gcond *> (stmt_info->stmt);
> +      tree cond_cst = gimple_cond_rhs (cond_stmt);
> +      pattern_stmt
> +	= gimple_build_cond (gimple_cond_code (cond_stmt),
> +			     gimple_get_lhs (pattern_stmt),
> +			     fold_convert (ret_type, cond_cst),
> +			     gimple_cond_true_label (cond_stmt),
> +			     gimple_cond_false_label (cond_stmt));
> +    }
> +
>    *type_out = STMT_VINFO_VECTYPE (stmt_info);
>    vect_pattern_detected ("bitfield_ref pattern", stmt_info->stmt);
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)