The following reverts the previous bogus fixes and instead fixes
strided store support to properly get at SLP operands. It also
fixes a typo and exchanges i == i for i == 0.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Richard.
2015-06-16 Richard Biener <rguent...@suse.de>
PR tree-optimization/66251
* tree-vect-stmts.c (vectorizable_store): Fix gathering of vectorized
stmts for SLP strided stores.
* gfortran.fortran-torture/compile/pr66251-2.f90: New testcase.
Revert
2015-05-22 Richard Biener <rguent...@suse.de>
PR tree-optimization/66251
* tree-vect-stmts.c (vectorizable_conversion): Properly
set STMT_VINFO_VEC_STMT even for the SLP case.
2015-05-26 Michael Matz <m...@suse.de>
PR middle-end/66251
* tree-vect-stmts.c (vect_create_vectorized_demotion_stmts): Always set
STMT_VINFO_VEC_STMT, also with SLP.
Index: gcc/testsuite/gfortran.fortran-torture/compile/pr66251-2.f90
===================================================================
--- gcc/testsuite/gfortran.fortran-torture/compile/pr66251-2.f90
(revision 0)
+++ gcc/testsuite/gfortran.fortran-torture/compile/pr66251-2.f90
(working copy)
@@ -0,0 +1,23 @@
+subroutine mv(m,nc,irp,ja,val,x,ldx,y,ldy,acc)
+ use iso_fortran_env
+ implicit none
+
+ integer, parameter :: ipk_ = int32
+ integer, parameter :: spk_ = real32
+ complex(spk_), parameter :: czero=(0.0_spk_,0.0_spk_)
+
+ integer(ipk_), intent(in) :: m,ldx,ldy,nc,irp(*),ja(*)
+ complex(spk_), intent(in) :: x(ldx,*),val(*)
+ complex(spk_), intent(inout) :: y(ldy,*)
+ complex(spk_), intent(inout) :: acc(*)
+ integer(ipk_) :: i,j,k, ir, jc
+
+ do i=1,m
+ acc(1:nc) = czero
+ do j=irp(i), irp(i+1)-1
+ acc(1:nc) = acc(1:nc) + val(j) * x(ja(j),1:nc)
+ enddo
+ y(i,1:nc) = -acc(1:nc)
+ end do
+
+end subroutine mv
Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c (revision 224514)
+++ gcc/tree-vect-stmts.c (working copy)
@@ -3370,13 +3370,15 @@ vect_create_vectorized_demotion_stmts (v
(or in STMT_VINFO_RELATED_STMT chain). */
if (slp_node)
SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt);
-
- if (!*prev_stmt_info)
- STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
else
- STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt;
+ {
+ if (!*prev_stmt_info)
+ STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
+ else
+ STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt;
- *prev_stmt_info = vinfo_for_stmt (new_stmt);
+ *prev_stmt_info = vinfo_for_stmt (new_stmt);
+ }
}
}
@@ -3955,12 +3957,14 @@ vectorizable_conversion (gimple stmt, gi
if (slp_node)
SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt);
-
- if (!prev_stmt_info)
- STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
else
- STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
- prev_stmt_info = vinfo_for_stmt (new_stmt);
+ {
+ if (!prev_stmt_info)
+ STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
+ else
+ STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
+ prev_stmt_info = vinfo_for_stmt (new_stmt);
+ }
}
}
@@ -5327,9 +5331,23 @@ vectorizable_store (gimple stmt, gimple_
/* We've set op and dt above, from gimple_assign_rhs1(stmt),
and first_stmt == stmt. */
if (j == 0)
- vec_oprnd = vect_get_vec_def_for_operand (op, first_stmt, NULL);
+ {
+ if (slp)
+ {
+ vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL,
+ slp_node, -1);
+ vec_oprnd = vec_oprnds[0];
+ }
+ else
+ vec_oprnd = vect_get_vec_def_for_operand (op, first_stmt, NULL);
+ }
else
- vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd);
+ {
+ if (slp)
+ vec_oprnd = vec_oprnds[j];
+ else
+ vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd);
+ }
for (i = 0; i < nstores; i++)
{
@@ -5359,7 +5377,7 @@ vectorizable_store (gimple stmt, gimple_
vect_finish_stmt_generation (stmt, incr, gsi);
running_off = newoff;
- if (j == 0 && i == i)
+ if (j == 0 && i == 0)
STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = assign;
else
STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign;
