On Wed, 3 Jun 2015, Richard Biener wrote:
>
> This allows all permutations we can generate (according to the target).
>
> Bootstrap and regtest pending on x86_64-unknown-linux-gnu.
So this turned up other issues thus the following is what I have
committed after bootstrapping and testing on x86_64-unknown-linux-gnu.
Richard.
2015-06-08 Richard Biener <rguent...@suse.de>
* tree-vect-stmts.c (vectorizable_load): Compute the pointer
adjustment for gaps at the end of a SLP load group properly.
* tree-vect-slp.c (vect_supported_load_permutation_p): Allow
all permutations we can generate.
(vect_transform_slp_perm_load): Use the correct group-size.
* gcc.dg/vect/slp-perm-10.c: New testcase.
* gcc.dg/vect/slp-23.c: Adjust.
* gcc.dg/torture/pr53366-2.c: Also verify cross-iteration
vector pointer update.
Index: gcc/tree-vect-stmts.c
===================================================================
*** gcc/tree-vect-stmts.c (revision 224077)
--- gcc/tree-vect-stmts.c (working copy)
*************** vectorizable_load (gimple stmt, gimple_s
*** 5807,5813 ****
gimple ptr_incr = NULL;
int nunits = TYPE_VECTOR_SUBPARTS (vectype);
int ncopies;
! int i, j, group_size = -1, group_gap;
tree msq = NULL_TREE, lsq;
tree offset = NULL_TREE;
tree byte_offset = NULL_TREE;
--- 5807,5813 ----
gimple ptr_incr = NULL;
int nunits = TYPE_VECTOR_SUBPARTS (vectype);
int ncopies;
! int i, j, group_size = -1, group_gap_adj;
tree msq = NULL_TREE, lsq;
tree offset = NULL_TREE;
tree byte_offset = NULL_TREE;
*************** vectorizable_load (gimple stmt, gimple_s
*** 6396,6421 ****
}
first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt));
/* VEC_NUM is the number of vect stmts to be created for this group. */
if (slp)
{
grouped_load = false;
vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
! group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt));
}
else
! {
! vec_num = group_size;
! group_gap = 0;
! }
}
else
{
first_stmt = stmt;
first_dr = dr;
group_size = vec_num = 1;
! group_gap = 0;
}
alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
--- 6396,6419 ----
}
first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt));
+ group_gap_adj = 0;
/* VEC_NUM is the number of vect stmts to be created for this group. */
if (slp)
{
grouped_load = false;
vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
! group_gap_adj = vf * group_size - nunits * vec_num;
}
else
! vec_num = group_size;
}
else
{
first_stmt = stmt;
first_dr = dr;
group_size = vec_num = 1;
! group_gap_adj = 0;
}
alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
*************** vectorizable_load (gimple stmt, gimple_s
*** 6831,6842 ****
if (slp && !slp_perm)
SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt);
}
! /* Bump the vector pointer to account for a gap. */
! if (slp && group_gap != 0)
{
! tree bump = size_binop (MULT_EXPR,
! TYPE_SIZE_UNIT (elem_type),
! size_int (group_gap));
dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi,
stmt, bump);
}
--- 6829,6843 ----
if (slp && !slp_perm)
SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt);
}
! /* Bump the vector pointer to account for a gap or for excess
! elements loaded for a permuted SLP load. */
! if (group_gap_adj != 0)
{
! bool ovf;
! tree bump
! = wide_int_to_tree (sizetype,
! wi::smul (TYPE_SIZE_UNIT (elem_type),
! group_gap_adj, &ovf));
dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi,
stmt, bump);
}
Index: gcc/tree-vect-slp.c
===================================================================
*** gcc/tree-vect-slp.c (revision 224077)
--- gcc/tree-vect-slp.c (working copy)
*************** vect_supported_load_permutation_p (slp_i
*** 1502,1548 ****
return true;
}
! /* FORNOW: the only supported permutation is 0..01..1.. of length equal to
! GROUP_SIZE and where each sequence of same drs is of GROUP_SIZE length as
! well (unless it's reduction). */
! if (SLP_INSTANCE_LOADS (slp_instn).length () != group_size)
! return false;
! FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
! if (!node->load_permutation.exists ())
! return false;
!
! load_index = sbitmap_alloc (group_size);
! bitmap_clear (load_index);
! FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
! {
! unsigned int lidx = node->load_permutation[0];
! if (bitmap_bit_p (load_index, lidx))
! {
! sbitmap_free (load_index);
! return false;
! }
! bitmap_set_bit (load_index, lidx);
! FOR_EACH_VEC_ELT (node->load_permutation, j, k)
! if (k != lidx)
! {
! sbitmap_free (load_index);
! return false;
! }
! }
! for (i = 0; i < group_size; i++)
! if (!bitmap_bit_p (load_index, i))
! {
! sbitmap_free (load_index);
! return false;
! }
! sbitmap_free (load_index);
!
FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
if (node->load_permutation.exists ()
&& !vect_transform_slp_perm_load
(node, vNULL, NULL,
SLP_INSTANCE_UNROLLING_FACTOR (slp_instn), slp_instn, true))
return false;
return true;
}
--- 1517,1530 ----
return true;
}
! /* For loop vectorization verify we can generate the permutation. */
FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn), i, node)
if (node->load_permutation.exists ()
&& !vect_transform_slp_perm_load
(node, vNULL, NULL,
SLP_INSTANCE_UNROLLING_FACTOR (slp_instn), slp_instn, true))
return false;
+
return true;
}
*************** vect_transform_slp_perm_load (slp_tree n
*** 3287,3292 ****
--- 3269,3276 ----
if (!STMT_VINFO_GROUPED_ACCESS (stmt_info))
return false;
+ stmt_info = vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info));
+
/* Generate permutation masks for every NODE. Number of masks for each NODE
is equal to GROUP_SIZE.
E.g., we have a group of three nodes with three loads from the same
*************** vect_transform_slp_perm_load (slp_tree n
*** 3321,3327 ****
for (k = 0; k < group_size; k++)
{
i = SLP_TREE_LOAD_PERMUTATION (node)[k];
! first_mask_element = i + j * group_size;
if (!vect_get_mask_element (stmt, first_mask_element, 0,
nunits, only_one_vec, index,
mask, ¤t_mask_element,
--- 3305,3311 ----
for (k = 0; k < group_size; k++)
{
i = SLP_TREE_LOAD_PERMUTATION (node)[k];
! first_mask_element = i + j * STMT_VINFO_GROUP_SIZE (stmt_info);
if (!vect_get_mask_element (stmt, first_mask_element, 0,
nunits, only_one_vec, index,
mask, ¤t_mask_element,
Index: gcc/testsuite/gcc.dg/vect/slp-perm-10.c
===================================================================
*** gcc/testsuite/gcc.dg/vect/slp-perm-10.c (revision 0)
--- gcc/testsuite/gcc.dg/vect/slp-perm-10.c (working copy)
***************
*** 0 ****
--- 1,53 ----
+ /* { dg-require-effective-target vect_int } */
+
+ #include "tree-vect.h"
+
+ int a[256], b[256];
+
+ void __attribute__((noinline))
+ foo (void)
+ {
+ int i;
+ for (i = 0; i < 32; ++i)
+ {
+ b[i*8+0] = a[i*8+0];
+ b[i*8+1] = a[i*8+0];
+ b[i*8+2] = a[i*8+3];
+ b[i*8+3] = a[i*8+3];
+ b[i*8+4] = a[i*8+4];
+ b[i*8+5] = a[i*8+6];
+ b[i*8+6] = a[i*8+4];
+ b[i*8+7] = a[i*8+6];
+ }
+ }
+
+ int main ()
+ {
+ int i;
+
+ check_vect ();
+
+ for (i = 0; i < 256; ++i)
+ {
+ a[i] = i;
+ __asm__ volatile ("");
+ }
+
+ foo ();
+
+ for (i = 0; i < 32; ++i)
+ if (b[i*8+0] != i*8+0
+ || b[i*8+1] != i*8+0
+ || b[i*8+2] != i*8+3
+ || b[i*8+3] != i*8+3
+ || b[i*8+4] != i*8+4
+ || b[i*8+5] != i*8+6
+ || b[i*8+6] != i*8+4
+ || b[i*8+7] != i*8+6)
+ abort ();
+
+ return 0;
+ }
+
+ /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target
vect_perm } } } */
+ /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" {
target vect_perm } } } */
Index: gcc/testsuite/gcc.dg/vect/slp-23.c
===================================================================
*** gcc/testsuite/gcc.dg/vect/slp-23.c (revision 224077)
--- gcc/testsuite/gcc.dg/vect/slp-23.c (working copy)
*************** int main (void)
*** 108,112 ****
/* { dg-final { scan-tree-dump-times "vectorized 2 loops" 1 "vect" { target {
vect_strided8 && { ! { vect_no_align} } } } } } */
/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target {
! { vect_strided8 || vect_no_align } } } } } */
! /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" }
} */
--- 108,113 ----
/* { dg-final { scan-tree-dump-times "vectorized 2 loops" 1 "vect" { target {
vect_strided8 && { ! { vect_no_align} } } } } } */
/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target {
! { vect_strided8 || vect_no_align } } } } } */
! /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" {
target { ! vect_perm } } } } */
! /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 2 "vect" {
target vect_perm } } } */
Index: gcc/testsuite/gcc.dg/torture/pr53366-2.c
===================================================================
*** gcc/testsuite/gcc.dg/torture/pr53366-2.c (revision 224077)
--- gcc/testsuite/gcc.dg/torture/pr53366-2.c (working copy)
***************
*** 4,21 ****
extern void abort (void);
struct T { float r[3], i[3]; };
! struct U { struct T j[2]; };
void __attribute__ ((noinline))
foo (struct U *__restrict y, const float _Complex *__restrict x)
{
int i, j;
! for (j = 0; j < 2; ++j)
{
float a = __real__ x[j];
float b = __imag__ x[j];
! float c = __real__ x[j + 2];
! float d = __imag__ x[j + 2];
for (i = 0; i < 3; ++i)
{
y->j[j].r[i] = y->j[j].r[i] + a + c;
--- 4,21 ----
extern void abort (void);
struct T { float r[3], i[3]; };
! struct U { struct T j[4]; };
void __attribute__ ((noinline))
foo (struct U *__restrict y, const float _Complex *__restrict x)
{
int i, j;
! for (j = 0; j < 4; ++j)
{
float a = __real__ x[j];
float b = __imag__ x[j];
! float c = __real__ x[j + 4];
! float d = __imag__ x[j + 4];
for (i = 0; i < 3; ++i)
{
y->j[j].r[i] = y->j[j].r[i] + a + c;
*************** foo (struct U *__restrict y, const float
*** 24,43 ****
}
}
! _Complex float x[4];
struct U y;
int
main ()
{
int i, j;
! for (i = 0; i < 4; ++i)
! x[i] = i + 1.0iF * (2 * i);
foo (&y, x);
! for (j = 0; j < 2; ++j)
for (i = 0; i < 3; ++i)
! if (y.j[j].r[i] != __real__ (x[j] + x[j + 2])
! || y.j[j].i[i] != __imag__ (x[j] + x[j + 2]))
__builtin_abort ();
return 0;
}
--- 24,46 ----
}
}
! _Complex float x[8];
struct U y;
int
main ()
{
int i, j;
! for (i = 0; i < 8; ++i)
! {
! x[i] = i + 1.0iF * (2 * i);
! __asm__ volatile ("");
! }
foo (&y, x);
! for (j = 0; j < 4; ++j)
for (i = 0; i < 3; ++i)
! if (y.j[j].r[i] != __real__ (x[j] + x[j + 4])
! || y.j[j].i[i] != __imag__ (x[j] + x[j + 4]))
__builtin_abort ();
return 0;
}
