This detects 1:1 permutations and avoids asking the target if it can
create those as well as generating VEC_PERM_EXPR in the vectorized code.
Bootstrap / regtest in progress on x86_64-unknown-linux-gnu.
Richard.
2015-11-27 Richard Biener <rguent...@suse.de>
PR tree-optimization/68553
* tree-vect-slp.c (vect_create_mask_and_perm): Skip VEC_PERM_EXPR
generation for 1:1 permutations.
(vect_transform_slp_perm_load): Detect 1:1 permutations.
Index: gcc/tree-vect-slp.c
===================================================================
*** gcc/tree-vect-slp.c (revision 230993)
--- gcc/tree-vect-slp.c (working copy)
*************** vect_create_mask_and_perm (gimple *stmt,
*** 3224,3235 ****
first_vec = dr_chain[first_vec_indx];
second_vec = dr_chain[second_vec_indx];
! /* Generate the permute statement. */
! perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
! first_vec, second_vec, mask);
! data_ref = make_ssa_name (perm_dest, perm_stmt);
! gimple_set_lhs (perm_stmt, data_ref);
! vect_finish_stmt_generation (stmt, perm_stmt, gsi);
/* Store the vector statement in NODE. */
SLP_TREE_VEC_STMTS (node)[stride_out * i + vect_stmts_counter]
--- 3253,3270 ----
first_vec = dr_chain[first_vec_indx];
second_vec = dr_chain[second_vec_indx];
! /* Generate the permute statement if necessary. */
! if (mask)
! {
! perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
! first_vec, second_vec, mask);
! data_ref = make_ssa_name (perm_dest, perm_stmt);
! gimple_set_lhs (perm_stmt, data_ref);
! vect_finish_stmt_generation (stmt, perm_stmt, gsi);
! }
! else
! /* If mask was NULL_TREE generate the requested identity transform. */
! perm_stmt = SSA_NAME_DEF_STMT (first_vec);
/* Store the vector statement in NODE. */
SLP_TREE_VEC_STMTS (node)[stride_out * i + vect_stmts_counter]
*************** vect_transform_slp_perm_load (slp_tree n
*** 3315,3320 ****
--- 3350,3356 ----
int index = 0;
int first_vec_index = -1;
int second_vec_index = -1;
+ bool noop_p = true;
for (int j = 0; j < unroll_factor; j++)
{
*************** vect_transform_slp_perm_load (slp_tree n
*** 3351,3361 ****
gcc_assert (mask_element >= 0
&& mask_element < 2 * nunits);
mask[index++] = mask_element;
if (index == nunits)
{
! if (!can_vec_perm_p (mode, false, mask))
{
if (dump_enabled_p ())
{
--- 3387,3400 ----
gcc_assert (mask_element >= 0
&& mask_element < 2 * nunits);
+ if (mask_element != index)
+ noop_p = false;
mask[index++] = mask_element;
if (index == nunits)
{
! if (! noop_p
! && ! can_vec_perm_p (mode, false, mask))
{
if (dump_enabled_p ())
{
*************** vect_transform_slp_perm_load (slp_tree n
*** 3371,3381 ****
if (!analyze_only)
{
! tree mask_vec, *mask_elts;
! mask_elts = XALLOCAVEC (tree, nunits);
! for (int l = 0; l < nunits; ++l)
! mask_elts[l] = build_int_cst (mask_element_type, mask[l]);
! mask_vec = build_vector (mask_type, mask_elts);
if (second_vec_index == -1)
second_vec_index = first_vec_index;
--- 3410,3425 ----
if (!analyze_only)
{
! tree mask_vec = NULL_TREE;
!
! if (! noop_p)
! {
! tree *mask_elts = XALLOCAVEC (tree, nunits);
! for (int l = 0; l < nunits; ++l)
! mask_elts[l] = build_int_cst (mask_element_type,
! mask[l]);
! mask_vec = build_vector (mask_type, mask_elts);
! }
if (second_vec_index == -1)
second_vec_index = first_vec_index;
*************** vect_transform_slp_perm_load (slp_tree n
*** 3388,3393 ****
--- 3432,3438 ----
index = 0;
first_vec_index = -1;
second_vec_index = -1;
+ noop_p = true;
}
}
}
