procyberian/gcc
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

re PR tree-optimization/69882 (Excessive reduction statements generated by SLP)

Browse source 
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.

From-SVN: r233605
This commit is contained in:
Richard Biener 2016-02-22 14:53:17 +00:00 • committed by Richard Biener
parent 0b1587d4b3
commit c4e360f44f
4 changed files with 67 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
gcc/ChangeLog
View file

@ -1,3 +1,11 @@
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.
2016-02-22 Bin Cheng <bin.cheng@arm.com>
* tree-vect-loop.c (vect_estimate_min_profitable_iters): Dump

5
gcc/testsuite/ChangeLog
View file

@ -1,3 +1,8 @@
2016-02-22 Richard Biener <rguenther@suse.de>
PR tree-optimization/69882
* gfortran.dg/vect/pr69882.f90: New testcase.
2016-02-22 Jakub Jelinek <jakub@redhat.com>
PR target/69885

41
gcc/testsuite/gfortran.dg/vect/pr69882.f90 Normal file
View file

@ -0,0 +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

20
gcc/tree-vect-slp.c
View file

@ -1332,8 +1332,19 @@ vect_attempt_slp_rearrange_stmts (slp_instance slp_instn)
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)
SLP_TREE_LOAD_PERMUTATION (node).release ();
{
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;
}
@ -1381,12 +1392,7 @@ vect_supported_load_permutation_p (slp_instance slp_instn)
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. */
}
vect_attempt_slp_rearrange_stmts (slp_instn);
/* In basic block vectorization we allow any subchain of an interleaving
chain.