[1/2] Minor refactoring of tree-vect-patterns.c

Hello,

in working on a fix for PR 52633, I noticed that tree-vect-patterns.c now
contains a number of copies of rather similar code (of which my patch
would have added another copy), so it seems to make sense to do a bit of
refactoring first.

This patch introduced a new helper function "vect_same_loop_or_bb_p"
to decide whether a new statement is in the same loop or basic-block
(depending on whether we're currently doing loop-based or basic-block-based
vectorization) as an existing statement.  The helper is then used everywhere
where this test is currently open-coded.

As a side-effect, the patch actually contains two bug fixes:

- The condition in some places
     (!loop && gimple_bb (def_stmt) != BB_VINFO_BB (bb_vinfo)
            && gimple_code (def_stmt) != GIMPLE_PHI)

  doesn't seem to make sense.  It allows PHI statements from random
  other basic blocks -- but those will have an uninitialized
  vinfo_for_stmt, so if that test ever hits, we're going to crash.

  The check was introduced in this patch:

    http://gcc.gnu.org/ml/gcc-patches/2012-02/msg00191.html

  which likewise introduces a similar check in vect_get_and_check_slp_defs:

      (!loop && gimple_bb (def_stmt) == BB_VINFO_BB (bb_vinfo)
             && gimple_code (def_stmt) != GIMPLE_PHI))

  This makes sense: basic-block vectorization operates only on statements
  in the current basic block, but *not* on the incoming PHIs.

  It seems that the condition simply was incorrectly negated when moving
  it over to the tree-vect-pattern.c uses.


- In vect_recog_widen_shift_pattern, the test for same loop / basic-block
  is still missing.  In my recent patch to PR 52870 I added the check to
  vect_recog_widen_mult_pattern ... it is still missing in ...shift_pattern.


Tested on i386-linux-gnu and arm-linux-gnueabi with no regressions.

OK for mainline?

Bye,
Ulrich


ChangeLog:

	* tree-vect-patterns.c (vect_same_loop_or_bb_p): New function.
	(vect_handle_widen_op_by_const): Use it instead of open-coding test.
	(vect_recog_widen_mult_pattern): Likewise.
	(vect_operation_fits_smaller_type): Likewise.
	(vect_recog_over_widening_pattern): Likewise.
	(vect_recog_widen_shift_pattern): Add to vect_same_loop_or_bb_p test.

On Mon, Apr 30, 2012 at 6:19 PM, Ulrich Weigand <uweigand@de.ibm.com> wrote:
> Hello,
>
> in working on a fix for PR 52633, I noticed that tree-vect-patterns.c now
> contains a number of copies of rather similar code (of which my patch
> would have added another copy), so it seems to make sense to do a bit of
> refactoring first.
>
> This patch introduced a new helper function "vect_same_loop_or_bb_p"
> to decide whether a new statement is in the same loop or basic-block
> (depending on whether we're currently doing loop-based or basic-block-based
> vectorization) as an existing statement.  The helper is then used everywhere
> where this test is currently open-coded.
>
> As a side-effect, the patch actually contains two bug fixes:
>
> - The condition in some places
>     (!loop && gimple_bb (def_stmt) != BB_VINFO_BB (bb_vinfo)
>            && gimple_code (def_stmt) != GIMPLE_PHI)
>
>  doesn't seem to make sense.  It allows PHI statements from random
>  other basic blocks -- but those will have an uninitialized
>  vinfo_for_stmt, so if that test ever hits, we're going to crash.
>
>  The check was introduced in this patch:
>
>    http://gcc.gnu.org/ml/gcc-patches/2012-02/msg00191.html
>
>  which likewise introduces a similar check in vect_get_and_check_slp_defs:
>
>      (!loop && gimple_bb (def_stmt) == BB_VINFO_BB (bb_vinfo)
>             && gimple_code (def_stmt) != GIMPLE_PHI))
>
>  This makes sense: basic-block vectorization operates only on statements
>  in the current basic block, but *not* on the incoming PHIs.
>
>  It seems that the condition simply was incorrectly negated when moving
>  it over to the tree-vect-pattern.c uses.
>
>
> - In vect_recog_widen_shift_pattern, the test for same loop / basic-block
>  is still missing.  In my recent patch to PR 52870 I added the check to
>  vect_recog_widen_mult_pattern ... it is still missing in ...shift_pattern.
>
>
> Tested on i386-linux-gnu and arm-linux-gnueabi with no regressions.
>
> OK for mainline?

Ok.

Thanks,
Richard.

> Bye,
> Ulrich
>
>
> ChangeLog:
>
>        * tree-vect-patterns.c (vect_same_loop_or_bb_p): New function.
>        (vect_handle_widen_op_by_const): Use it instead of open-coding test.
>        (vect_recog_widen_mult_pattern): Likewise.
>        (vect_operation_fits_smaller_type): Likewise.
>        (vect_recog_over_widening_pattern): Likewise.
>        (vect_recog_widen_shift_pattern): Add to vect_same_loop_or_bb_p test.
>
>
> Index: gcc-head/gcc/tree-vect-patterns.c
> ===================================================================
> --- gcc-head.orig/gcc/tree-vect-patterns.c      2012-04-26 15:53:48.000000000 +0200
> +++ gcc-head/gcc/tree-vect-patterns.c   2012-04-26 19:46:12.000000000 +0200
> @@ -84,6 +84,41 @@ new_pattern_def_seq (stmt_vec_info stmt_
>   append_pattern_def_seq (stmt_info, stmt);
>  }
>
> +/* Check whether STMT2 is in the same loop or basic block as STMT1.
> +   Which of the two applies depends on whether we're currently doing
> +   loop-based or basic-block-based vectorization, as determined by
> +   the vinfo_for_stmt for STMT1 (which must be defined).
> +
> +   If this returns true, vinfo_for_stmt for STMT2 is guaranteed
> +   to be defined as well.  */
> +
> +static bool
> +vect_same_loop_or_bb_p (gimple stmt1, gimple stmt2)
> +{
> +  stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt1);
> +  loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
> +  bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo);
> +
> +  if (!gimple_bb (stmt2))
> +    return false;
> +
> +  if (loop_vinfo)
> +    {
> +      struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
> +      if (!flow_bb_inside_loop_p (loop, gimple_bb (stmt2)))
> +       return false;
> +    }
> +  else
> +    {
> +      if (gimple_bb (stmt2) != BB_VINFO_BB (bb_vinfo)
> +         || gimple_code (stmt2) == GIMPLE_PHI)
> +       return false;
> +    }
> +
> +  gcc_assert (vinfo_for_stmt (stmt2));
> +  return true;
> +}
> +
>  /* Check whether NAME, an ssa-name used in USE_STMT,
>    is a result of a type promotion or demotion, such that:
>      DEF_STMT: NAME = NOP (name0)
> @@ -400,16 +435,6 @@ vect_handle_widen_op_by_const (gimple st
>  {
>   tree new_type, new_oprnd, tmp;
>   gimple new_stmt;
> -  loop_vec_info loop_vinfo;
> -  struct loop *loop = NULL;
> -  bb_vec_info bb_vinfo;
> -  stmt_vec_info stmt_vinfo;
> -
> -  stmt_vinfo = vinfo_for_stmt (stmt);
> -  loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
> -  bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo);
> -  if (loop_vinfo)
> -    loop = LOOP_VINFO_LOOP (loop_vinfo);
>
>   if (code != MULT_EXPR && code != LSHIFT_EXPR)
>     return false;
> @@ -425,12 +450,10 @@ vect_handle_widen_op_by_const (gimple st
>       return true;
>     }
>
> -  if (TYPE_PRECISION (type) < (TYPE_PRECISION (*half_type) * 4)
> -      || !gimple_bb (def_stmt)
> -      || (loop && !flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)))
> -      || (!loop && gimple_bb (def_stmt) != BB_VINFO_BB (bb_vinfo)
> -         && gimple_code (def_stmt) != GIMPLE_PHI)
> -      || !vinfo_for_stmt (def_stmt))
> +  if (TYPE_PRECISION (type) < (TYPE_PRECISION (*half_type) * 4))
> +    return false;
> +
> +  if (!vect_same_loop_or_bb_p (stmt, def_stmt))
>     return false;
>
>   /* TYPE is 4 times bigger than HALF_TYPE, try widening operation for
> @@ -564,16 +587,6 @@ vect_recog_widen_mult_pattern (VEC (gimp
>   VEC (tree, heap) *dummy_vec;
>   bool op1_ok;
>   bool promotion;
> -  loop_vec_info loop_vinfo;
> -  struct loop *loop = NULL;
> -  bb_vec_info bb_vinfo;
> -  stmt_vec_info stmt_vinfo;
> -
> -  stmt_vinfo = vinfo_for_stmt (last_stmt);
> -  loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
> -  bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo);
> -  if (loop_vinfo)
> -    loop = LOOP_VINFO_LOOP (loop_vinfo);
>
>   if (!is_gimple_assign (last_stmt))
>     return NULL;
> @@ -645,9 +658,7 @@ vect_recog_widen_mult_pattern (VEC (gimp
>           || gimple_assign_rhs_code (use_stmt) != NOP_EXPR)
>         return NULL;
>
> -      if (!gimple_bb (use_stmt)
> -         || (loop && !flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
> -         || (!loop && gimple_bb (use_stmt) != BB_VINFO_BB (bb_vinfo)))
> +      if (!vect_same_loop_or_bb_p (last_stmt, use_stmt))
>        return NULL;
>
>       use_lhs = gimple_assign_lhs (use_stmt);
> @@ -952,14 +963,8 @@ vect_operation_fits_smaller_type (gimple
>   tree interm_type = NULL_TREE, half_type, tmp, new_oprnd, type;
>   gimple def_stmt, new_stmt;
>   bool first = false;
> -  loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (vinfo_for_stmt (stmt));
> -  bb_vec_info bb_info = STMT_VINFO_BB_VINFO (vinfo_for_stmt (stmt));
> -  struct loop *loop = NULL;
>   bool promotion;
>
> -  if (loop_info)
> -    loop = LOOP_VINFO_LOOP (loop_info);
> -
>   *op0 = NULL_TREE;
>   *op1 = NULL_TREE;
>   *new_def_stmt = NULL;
> @@ -991,13 +996,9 @@ vect_operation_fits_smaller_type (gimple
>     {
>       first = true;
>       if (!type_conversion_p (oprnd, stmt, false, &half_type, &def_stmt,
> -                               &promotion)
> -         || !promotion
> -          || !gimple_bb (def_stmt)
> -          || (loop && !flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)))
> -         || (!loop && gimple_bb (def_stmt) != BB_VINFO_BB (bb_info)
> -             && gimple_code (def_stmt) != GIMPLE_PHI)
> -          || !vinfo_for_stmt (def_stmt))
> +                             &promotion)
> +         || !promotion
> +         || !vect_same_loop_or_bb_p (stmt, def_stmt))
>         return false;
>     }
>
> @@ -1171,16 +1172,6 @@ vect_recog_over_widening_pattern (VEC (g
>   tree var = NULL_TREE, new_type = NULL_TREE, tmp, new_oprnd;
>   bool first;
>   tree type = NULL;
> -  loop_vec_info loop_vinfo;
> -  struct loop *loop = NULL;
> -  bb_vec_info bb_vinfo;
> -  stmt_vec_info stmt_vinfo;
> -
> -  stmt_vinfo = vinfo_for_stmt (stmt);
> -  loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
> -  bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo);
> -  if (loop_vinfo)
> -    loop = LOOP_VINFO_LOOP (loop_vinfo);
>
>   first = true;
>   while (1)
> @@ -1212,9 +1203,7 @@ vect_recog_over_widening_pattern (VEC (g
>         }
>
>       if (nuses != 1 || !is_gimple_assign (use_stmt)
> -          || !gimple_bb (use_stmt)
> -          || (loop && !flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
> -         || (!loop && gimple_bb (use_stmt) != BB_VINFO_BB (bb_vinfo)))
> +         || !vect_same_loop_or_bb_p (stmt, use_stmt))
>         return NULL;
>
>       /* Create pattern statement for STMT.  */
> @@ -1495,6 +1484,9 @@ vect_recog_widen_shift_pattern (VEC (gim
>              || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (use_stmt)))
>            return NULL;
>
> +         if (!vect_same_loop_or_bb_p (last_stmt, use_stmt))
> +           return NULL;
> +
>           use_lhs = gimple_assign_lhs (use_stmt);
>           use_type = TREE_TYPE (use_lhs);
>
> --
>  Dr. Ulrich Weigand
>  GNU Toolchain for Linux on System z and Cell BE
>  Ulrich.Weigand@de.ibm.com
>