[PATCH] Fix PR68892

[PATCH] Fix PR68892

[Richard Biener]

The following fixes PR68892, the BB vectorizer now happily creates
a load of dead vector loads (we had a similar bug with loop
single-element interleaving support in the past).  Fixed as a side-effect
of making the SLP load cost reflect reality.

Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.

Richard.

2015-12-14  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/68892
	* tree-vect-slp.c (vect_analyze_slp_cost_1): Properly compute
	cost for permuted loads.

	* gcc.dg/vect/bb-slp-pr68892.c: New testcase.

Index: gcc/tree-vect-slp.c
===================================================================
*** gcc/tree-vect-slp.c	(revision 231617)
--- gcc/tree-vect-slp.c	(working copy)
*************** vect_analyze_slp_cost_1 (slp_instance in
*** 1405,1414 ****
  {
    unsigned i;
    slp_tree child;
!   gimple *stmt, *s;
    stmt_vec_info stmt_info;
    tree lhs;
-   unsigned group_size = SLP_INSTANCE_GROUP_SIZE (instance);
  
    /* Recurse down the SLP tree.  */
    FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
--- 1405,1413 ----
  {
    unsigned i;
    slp_tree child;
!   gimple *stmt;
    stmt_vec_info stmt_info;
    tree lhs;
  
    /* Recurse down the SLP tree.  */
    FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
*************** vect_analyze_slp_cost_1 (slp_instance in
*** 1427,1470 ****
  			       node, prologue_cost_vec, body_cost_vec);
        else
  	{
- 	  int i;
  	  gcc_checking_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)));
- 	  /* If the load is permuted then the alignment is determined by
- 	     the first group element not by the first scalar stmt DR.  */
  	  if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
  	    {
  	      stmt = GROUP_FIRST_ELEMENT (stmt_info);
  	      stmt_info = vinfo_for_stmt (stmt);
  	    }
  	  vect_model_load_cost (stmt_info, ncopies_for_cost, false,
  				node, prologue_cost_vec, body_cost_vec);
- 	  /* If the load is permuted record the cost for the permutation.
- 	     ???  Loads from multiple chains are let through here only
- 	     for a single special case involving complex numbers where
- 	     in the end no permutation is necessary.  */
- 	  FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, s)
- 	    if ((STMT_VINFO_GROUP_FIRST_ELEMENT (vinfo_for_stmt (s))
- 		 == STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info))
- 		&& vect_get_place_in_interleaving_chain
- 		     (s, STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info)) != i)
- 	      {
- 		record_stmt_cost (body_cost_vec, group_size, vec_perm,
- 				  stmt_info, 0, vect_body);
- 		break;
- 	      }
  	}
      }
!   else
      {
        record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt,
  			stmt_info, 0, vect_body);
!       if (SLP_TREE_TWO_OPERATORS (node))
! 	{
! 	  record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt,
! 			    stmt_info, 0, vect_body);
! 	  record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm,
! 			    stmt_info, 0, vect_body);
! 	}
      }
  
    /* Scan operands and account for prologue cost of constants/externals.
--- 1426,1464 ----
  			       node, prologue_cost_vec, body_cost_vec);
        else
  	{
  	  gcc_checking_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)));
  	  if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
  	    {
+ 	      /* If the load is permuted then the alignment is determined by
+ 		 the first group element not by the first scalar stmt DR.  */
  	      stmt = GROUP_FIRST_ELEMENT (stmt_info);
  	      stmt_info = vinfo_for_stmt (stmt);
+ 	      /* Record the cost for the permutation.  */
+ 	      record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm,
+ 				stmt_info, 0, vect_body);
+ 	      /* And adjust the number of loads performed.  */
+ 	      unsigned nunits
+ 		= TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
+ 	      ncopies_for_cost
+ 	        = (GROUP_SIZE (stmt_info) - GROUP_GAP (stmt_info)
+ 		   + nunits - 1) / nunits;
+ 	      ncopies_for_cost *= SLP_INSTANCE_UNROLLING_FACTOR (instance);
  	    }
+ 	  /* Record the cost for the vector loads.  */
  	  vect_model_load_cost (stmt_info, ncopies_for_cost, false,
  				node, prologue_cost_vec, body_cost_vec);
  	}
+       return;
      }
! 
!   record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt,
! 		    stmt_info, 0, vect_body);
!   if (SLP_TREE_TWO_OPERATORS (node))
      {
        record_stmt_cost (body_cost_vec, ncopies_for_cost, vector_stmt,
  			stmt_info, 0, vect_body);
!       record_stmt_cost (body_cost_vec, ncopies_for_cost, vec_perm,
! 			stmt_info, 0, vect_body);
      }
  
    /* Scan operands and account for prologue cost of constants/externals.
Index: gcc/testsuite/gcc.dg/vect/bb-slp-pr68892.c
===================================================================
*** gcc/testsuite/gcc.dg/vect/bb-slp-pr68892.c	(revision 0)
--- gcc/testsuite/gcc.dg/vect/bb-slp-pr68892.c	(revision 0)
***************
*** 0 ****
--- 1,16 ----
+ /* { dg-do compile } */
+ /* { dg-additional-options "-fvect-cost-model=dynamic" } */
+ 
+ double a[128][128];
+ double b[128];
+ 
+ void foo(void)
+ {
+   b[0] = a[0][0];
+   b[1] = a[1][0];
+   b[2] = a[2][0];
+   b[3] = a[3][0];
+ }
+ 
+ /* { dg-final { scan-tree-dump "not profitable" "slp2" } } */
+ /* { dg-final { scan-tree-dump-times "Basic block will be vectorized" 0 "slp2" } } */

Re: [PATCH] Fix PR68892

[H.J. Lu]

On Mon, Dec 14, 2015 at 7:16 AM, Richard Biener <rguenther@suse.de> wrote:
>
> The following fixes PR68892, the BB vectorizer now happily creates
> a load of dead vector loads (we had a similar bug with loop
> single-element interleaving support in the past).  Fixed as a side-effect
> of making the SLP load cost reflect reality.
>
> Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
>

This caused:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69123


-- 
H.J.