[PATCH] Fix PR69882

[Richard Biener]

The following fixes an oversight in SLP reduction permutation 
optimization which will happily delete permutations required for
handling of gaps.

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

Richard.

2016-02-22  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/69882
	* tree-vect-slp.c (vect_attempt_slp_rearrange_stmts): Properly
	preserve permutations present because of gaps.
	(vect_supported_load_permutation_p): Always continue checking
	permutations after vect_attempt_slp_rearrange_stmts.

	* gfortran.dg/vect/pr69882.f90: New testcase.

Index: gcc/tree-vect-slp.c
===================================================================
*** gcc/tree-vect-slp.c	(revision 233598)
--- gcc/tree-vect-slp.c	(working copy)
*************** vect_attempt_slp_rearrange_stmts (slp_in
*** 1332,1339 ****
  			    node->load_permutation);
  
    /* We are done, no actual permutations need to be generated.  */
    FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
!     SLP_TREE_LOAD_PERMUTATION (node).release ();
    return true;
  }
  
--- 1332,1350 ----
  			    node->load_permutation);
  
    /* We are done, no actual permutations need to be generated.  */
+   unsigned int unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_instn);
    FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
!     {
!       gimple *first_stmt = SLP_TREE_SCALAR_STMTS (node)[0];
!       first_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt));
!       /* But we have to keep those permutations that are required because
!          of handling of gaps.  */
!       if (unrolling_factor == 1
! 	  || (group_size == GROUP_SIZE (vinfo_for_stmt (first_stmt))
! 	      && GROUP_GAP (vinfo_for_stmt (first_stmt)) == 0))
! 	SLP_TREE_LOAD_PERMUTATION (node).release ();
!     }
! 
    return true;
  }
  
*************** vect_supported_load_permutation_p (slp_i
*** 1381,1392 ****
       In reduction chain the order of the loads is not important.  */
    if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt))
        && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
!     {
!       if (vect_attempt_slp_rearrange_stmts (slp_instn))
! 	return true;
! 
!       /* Fallthru to general load permutation handling.  */
!     }
  
    /* In basic block vectorization we allow any subchain of an interleaving
       chain.
--- 1392,1398 ----
       In reduction chain the order of the loads is not important.  */
    if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt))
        && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
!     vect_attempt_slp_rearrange_stmts (slp_instn);
  
    /* In basic block vectorization we allow any subchain of an interleaving
       chain.
Index: gcc/testsuite/gfortran.dg/vect/pr69882.f90
===================================================================
*** gcc/testsuite/gfortran.dg/vect/pr69882.f90	(revision 0)
--- gcc/testsuite/gfortran.dg/vect/pr69882.f90	(working copy)
***************
*** 0 ****
--- 1,41 ----
+ ! { dg-additional-options "-Ofast" }
+ ! { dg-additional-options "-mavx" { target avx_runtime } }
+ 
+ subroutine foo(a, x)
+   implicit none
+ 
+   integer, parameter :: XX=4, YY=26
+   integer, intent(in) :: x
+   real *8, intent(in) :: a(XX,YY)
+   real *8 :: c(XX)
+ 
+   integer i, k
+ 
+   c = 0
+ 
+   do k=x,YY
+      do i=1,2
+         c(i) = max(c(i), a(i,k))
+      end do
+   end do
+ 
+   PRINT *, "c=", c
+ 
+   IF (c(1) .gt. 0.0) THEN
+      CALL ABORT
+   END IF
+ 
+   IF (c(2) .gt. 0.0) THEN
+      CALL ABORT
+   END IF
+ end subroutine foo
+ 
+ PROGRAM MAIN
+   real *8 a(4, 26)
+ 
+   a = 0
+   a(3,1) = 100.0
+   a(4,1) = 100.0
+   
+   CALL FOO(a, 1)
+ END PROGRAM