1 /* SLP - Basic Block Vectorization
2 Copyright (C) 2007-2016 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
4 and Ira Rosen <irar@il.ibm.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
30 #include "tree-pass.h"
32 #include "optabs-tree.h"
33 #include "insn-config.h"
34 #include "recog.h" /* FIXME: for insn_data */
36 #include "fold-const.h"
37 #include "stor-layout.h"
38 #include "gimple-iterator.h"
40 #include "tree-vectorizer.h"
41 #include "langhooks.h"
42 #include "gimple-walk.h"
46 /* Recursively free the memory allocated for the SLP tree rooted at NODE. */
49 vect_free_slp_tree (slp_tree node
)
54 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
55 vect_free_slp_tree (child
);
58 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
59 /* After transform some stmts are removed and thus their vinfo is gone. */
60 if (vinfo_for_stmt (stmt
))
62 gcc_assert (STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt
)) > 0);
63 STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt
))--;
66 SLP_TREE_CHILDREN (node
).release ();
67 SLP_TREE_SCALAR_STMTS (node
).release ();
68 SLP_TREE_VEC_STMTS (node
).release ();
69 SLP_TREE_LOAD_PERMUTATION (node
).release ();
75 /* Free the memory allocated for the SLP instance. */
78 vect_free_slp_instance (slp_instance instance
)
80 vect_free_slp_tree (SLP_INSTANCE_TREE (instance
));
81 SLP_INSTANCE_LOADS (instance
).release ();
86 /* Create an SLP node for SCALAR_STMTS. */
89 vect_create_new_slp_node (vec
<gimple
*> scalar_stmts
)
92 gimple
*stmt
= scalar_stmts
[0];
95 if (is_gimple_call (stmt
))
96 nops
= gimple_call_num_args (stmt
);
97 else if (is_gimple_assign (stmt
))
99 nops
= gimple_num_ops (stmt
) - 1;
100 if (gimple_assign_rhs_code (stmt
) == COND_EXPR
)
106 node
= XNEW (struct _slp_tree
);
107 SLP_TREE_SCALAR_STMTS (node
) = scalar_stmts
;
108 SLP_TREE_VEC_STMTS (node
).create (0);
109 SLP_TREE_CHILDREN (node
).create (nops
);
110 SLP_TREE_LOAD_PERMUTATION (node
) = vNULL
;
111 SLP_TREE_TWO_OPERATORS (node
) = false;
112 SLP_TREE_DEF_TYPE (node
) = vect_internal_def
;
115 FOR_EACH_VEC_ELT (scalar_stmts
, i
, stmt
)
116 STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt
))++;
122 /* This structure is used in creation of an SLP tree. Each instance
123 corresponds to the same operand in a group of scalar stmts in an SLP
125 typedef struct _slp_oprnd_info
127 /* Def-stmts for the operands. */
128 vec
<gimple
*> def_stmts
;
129 /* Information about the first statement, its vector def-type, type, the
130 operand itself in case it's constant, and an indication if it's a pattern
132 enum vect_def_type first_dt
;
139 /* Allocate operands info for NOPS operands, and GROUP_SIZE def-stmts for each
141 static vec
<slp_oprnd_info
>
142 vect_create_oprnd_info (int nops
, int group_size
)
145 slp_oprnd_info oprnd_info
;
146 vec
<slp_oprnd_info
> oprnds_info
;
148 oprnds_info
.create (nops
);
149 for (i
= 0; i
< nops
; i
++)
151 oprnd_info
= XNEW (struct _slp_oprnd_info
);
152 oprnd_info
->def_stmts
.create (group_size
);
153 oprnd_info
->first_dt
= vect_uninitialized_def
;
154 oprnd_info
->first_op_type
= NULL_TREE
;
155 oprnd_info
->first_pattern
= false;
156 oprnd_info
->second_pattern
= false;
157 oprnds_info
.quick_push (oprnd_info
);
164 /* Free operands info. */
167 vect_free_oprnd_info (vec
<slp_oprnd_info
> &oprnds_info
)
170 slp_oprnd_info oprnd_info
;
172 FOR_EACH_VEC_ELT (oprnds_info
, i
, oprnd_info
)
174 oprnd_info
->def_stmts
.release ();
175 XDELETE (oprnd_info
);
178 oprnds_info
.release ();
182 /* Find the place of the data-ref in STMT in the interleaving chain that starts
183 from FIRST_STMT. Return -1 if the data-ref is not a part of the chain. */
186 vect_get_place_in_interleaving_chain (gimple
*stmt
, gimple
*first_stmt
)
188 gimple
*next_stmt
= first_stmt
;
191 if (first_stmt
!= GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
196 if (next_stmt
== stmt
)
198 next_stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt
));
200 result
+= GROUP_GAP (vinfo_for_stmt (next_stmt
));
208 /* Get the defs for the rhs of STMT (collect them in OPRNDS_INFO), check that
209 they are of a valid type and that they match the defs of the first stmt of
210 the SLP group (stored in OPRNDS_INFO). If there was a fatal error
211 return -1, if the error could be corrected by swapping operands of the
212 operation return 1, if everything is ok return 0. */
215 vect_get_and_check_slp_defs (vec_info
*vinfo
,
216 gimple
*stmt
, unsigned stmt_num
,
217 vec
<slp_oprnd_info
> *oprnds_info
)
220 unsigned int i
, number_of_oprnds
;
222 enum vect_def_type dt
= vect_uninitialized_def
;
223 bool pattern
= false;
224 slp_oprnd_info oprnd_info
;
225 int first_op_idx
= 1;
226 bool commutative
= false;
227 bool first_op_cond
= false;
228 bool first
= stmt_num
== 0;
229 bool second
= stmt_num
== 1;
231 if (is_gimple_call (stmt
))
233 number_of_oprnds
= gimple_call_num_args (stmt
);
236 else if (is_gimple_assign (stmt
))
238 enum tree_code code
= gimple_assign_rhs_code (stmt
);
239 number_of_oprnds
= gimple_num_ops (stmt
) - 1;
240 if (gimple_assign_rhs_code (stmt
) == COND_EXPR
241 && COMPARISON_CLASS_P (gimple_assign_rhs1 (stmt
)))
243 first_op_cond
= true;
248 commutative
= commutative_tree_code (code
);
253 bool swapped
= false;
254 for (i
= 0; i
< number_of_oprnds
; i
++)
259 if (i
== 0 || i
== 1)
260 oprnd
= TREE_OPERAND (gimple_op (stmt
, first_op_idx
),
263 oprnd
= gimple_op (stmt
, first_op_idx
+ i
- 1);
266 oprnd
= gimple_op (stmt
, first_op_idx
+ (swapped
? !i
: i
));
268 oprnd_info
= (*oprnds_info
)[i
];
270 if (!vect_is_simple_use (oprnd
, vinfo
, &def_stmt
, &dt
))
272 if (dump_enabled_p ())
274 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
275 "Build SLP failed: can't analyze def for ");
276 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, oprnd
);
277 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
283 /* Check if DEF_STMT is a part of a pattern in LOOP and get the def stmt
284 from the pattern. Check that all the stmts of the node are in the
286 if (def_stmt
&& gimple_bb (def_stmt
)
287 && vect_stmt_in_region_p (vinfo
, def_stmt
)
288 && vinfo_for_stmt (def_stmt
)
289 && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (def_stmt
))
290 && !STMT_VINFO_RELEVANT (vinfo_for_stmt (def_stmt
))
291 && !STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt
)))
294 if (!first
&& !oprnd_info
->first_pattern
295 /* Allow different pattern state for the defs of the
296 first stmt in reduction chains. */
297 && (oprnd_info
->first_dt
!= vect_reduction_def
298 || (!second
&& !oprnd_info
->second_pattern
)))
308 if (dump_enabled_p ())
310 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
311 "Build SLP failed: some of the stmts"
312 " are in a pattern, and others are not ");
313 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, oprnd
);
314 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
320 def_stmt
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt
));
321 dt
= STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
));
323 if (dt
== vect_unknown_def_type
)
325 if (dump_enabled_p ())
326 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
327 "Unsupported pattern.\n");
331 switch (gimple_code (def_stmt
))
338 if (dump_enabled_p ())
339 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
340 "unsupported defining stmt:\n");
346 oprnd_info
->second_pattern
= pattern
;
350 oprnd_info
->first_dt
= dt
;
351 oprnd_info
->first_pattern
= pattern
;
352 oprnd_info
->first_op_type
= TREE_TYPE (oprnd
);
356 /* Not first stmt of the group, check that the def-stmt/s match
357 the def-stmt/s of the first stmt. Allow different definition
358 types for reduction chains: the first stmt must be a
359 vect_reduction_def (a phi node), and the rest
360 vect_internal_def. */
361 if (((oprnd_info
->first_dt
!= dt
362 && !(oprnd_info
->first_dt
== vect_reduction_def
363 && dt
== vect_internal_def
)
364 && !((oprnd_info
->first_dt
== vect_external_def
365 || oprnd_info
->first_dt
== vect_constant_def
)
366 && (dt
== vect_external_def
367 || dt
== vect_constant_def
)))
368 || !types_compatible_p (oprnd_info
->first_op_type
,
371 /* Try swapping operands if we got a mismatch. */
380 if (dump_enabled_p ())
381 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
382 "Build SLP failed: different types\n");
388 /* Check the types of the definitions. */
391 case vect_constant_def
:
392 case vect_external_def
:
393 case vect_reduction_def
:
396 case vect_internal_def
:
397 oprnd_info
->def_stmts
.quick_push (def_stmt
);
401 /* FORNOW: Not supported. */
402 if (dump_enabled_p ())
404 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
405 "Build SLP failed: illegal type of def ");
406 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, oprnd
);
407 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
417 /* If there are already uses of this stmt in a SLP instance then
418 we've committed to the operand order and can't swap it. */
419 if (STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmt
)) != 0)
421 if (dump_enabled_p ())
423 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
424 "Build SLP failed: cannot swap operands of "
426 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
433 tree cond
= gimple_assign_rhs1 (stmt
);
434 swap_ssa_operands (stmt
, &TREE_OPERAND (cond
, 0),
435 &TREE_OPERAND (cond
, 1));
436 TREE_SET_CODE (cond
, swap_tree_comparison (TREE_CODE (cond
)));
439 swap_ssa_operands (stmt
, gimple_assign_rhs1_ptr (stmt
),
440 gimple_assign_rhs2_ptr (stmt
));
441 if (dump_enabled_p ())
443 dump_printf_loc (MSG_NOTE
, vect_location
,
444 "swapped operands to match def types in ");
445 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
453 /* Verify if the scalar stmts STMTS are isomorphic, require data
454 permutation or are of unsupported types of operation. Return
455 true if they are, otherwise return false and indicate in *MATCHES
456 which stmts are not isomorphic to the first one. If MATCHES[0]
457 is false then this indicates the comparison could not be
458 carried out or the stmts will never be vectorized by SLP. */
461 vect_build_slp_tree_1 (vec_info
*vinfo
,
462 vec
<gimple
*> stmts
, unsigned int group_size
,
463 unsigned nops
, unsigned int *max_nunits
,
464 bool *matches
, bool *two_operators
)
467 gimple
*first_stmt
= stmts
[0], *stmt
= stmts
[0];
468 enum tree_code first_stmt_code
= ERROR_MARK
;
469 enum tree_code alt_stmt_code
= ERROR_MARK
;
470 enum tree_code rhs_code
= ERROR_MARK
;
471 enum tree_code first_cond_code
= ERROR_MARK
;
473 bool need_same_oprnds
= false;
474 tree vectype
= NULL_TREE
, scalar_type
, first_op1
= NULL_TREE
;
477 machine_mode optab_op2_mode
;
478 machine_mode vec_mode
;
480 gimple
*first_load
= NULL
, *prev_first_load
= NULL
;
482 /* For every stmt in NODE find its def stmt/s. */
483 FOR_EACH_VEC_ELT (stmts
, i
, stmt
)
487 if (dump_enabled_p ())
489 dump_printf_loc (MSG_NOTE
, vect_location
, "Build SLP for ");
490 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
493 /* Fail to vectorize statements marked as unvectorizable. */
494 if (!STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt
)))
496 if (dump_enabled_p ())
498 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
499 "Build SLP failed: unvectorizable statement ");
500 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
501 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
503 /* Fatal mismatch. */
508 lhs
= gimple_get_lhs (stmt
);
509 if (lhs
== NULL_TREE
)
511 if (dump_enabled_p ())
513 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
514 "Build SLP failed: not GIMPLE_ASSIGN nor "
516 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
517 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
519 /* Fatal mismatch. */
524 scalar_type
= vect_get_smallest_scalar_type (stmt
, &dummy
, &dummy
);
525 vectype
= get_vectype_for_scalar_type (scalar_type
);
528 if (dump_enabled_p ())
530 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
531 "Build SLP failed: unsupported data-type ");
532 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
534 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
536 /* Fatal mismatch. */
541 /* If populating the vector type requires unrolling then fail
542 before adjusting *max_nunits for basic-block vectorization. */
543 if (is_a
<bb_vec_info
> (vinfo
)
544 && TYPE_VECTOR_SUBPARTS (vectype
) > group_size
)
546 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
547 "Build SLP failed: unrolling required "
548 "in basic block SLP\n");
549 /* Fatal mismatch. */
554 /* In case of multiple types we need to detect the smallest type. */
555 if (*max_nunits
< TYPE_VECTOR_SUBPARTS (vectype
))
556 *max_nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
558 if (gcall
*call_stmt
= dyn_cast
<gcall
*> (stmt
))
560 rhs_code
= CALL_EXPR
;
561 if (gimple_call_internal_p (call_stmt
)
562 || gimple_call_tail_p (call_stmt
)
563 || gimple_call_noreturn_p (call_stmt
)
564 || !gimple_call_nothrow_p (call_stmt
)
565 || gimple_call_chain (call_stmt
))
567 if (dump_enabled_p ())
569 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
570 "Build SLP failed: unsupported call type ");
571 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
573 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
575 /* Fatal mismatch. */
581 rhs_code
= gimple_assign_rhs_code (stmt
);
583 /* Check the operation. */
586 first_stmt_code
= rhs_code
;
588 /* Shift arguments should be equal in all the packed stmts for a
589 vector shift with scalar shift operand. */
590 if (rhs_code
== LSHIFT_EXPR
|| rhs_code
== RSHIFT_EXPR
591 || rhs_code
== LROTATE_EXPR
592 || rhs_code
== RROTATE_EXPR
)
594 vec_mode
= TYPE_MODE (vectype
);
596 /* First see if we have a vector/vector shift. */
597 optab
= optab_for_tree_code (rhs_code
, vectype
,
601 || optab_handler (optab
, vec_mode
) == CODE_FOR_nothing
)
603 /* No vector/vector shift, try for a vector/scalar shift. */
604 optab
= optab_for_tree_code (rhs_code
, vectype
,
609 if (dump_enabled_p ())
610 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
611 "Build SLP failed: no optab.\n");
612 /* Fatal mismatch. */
616 icode
= (int) optab_handler (optab
, vec_mode
);
617 if (icode
== CODE_FOR_nothing
)
619 if (dump_enabled_p ())
620 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
622 "op not supported by target.\n");
623 /* Fatal mismatch. */
627 optab_op2_mode
= insn_data
[icode
].operand
[2].mode
;
628 if (!VECTOR_MODE_P (optab_op2_mode
))
630 need_same_oprnds
= true;
631 first_op1
= gimple_assign_rhs2 (stmt
);
635 else if (rhs_code
== WIDEN_LSHIFT_EXPR
)
637 need_same_oprnds
= true;
638 first_op1
= gimple_assign_rhs2 (stmt
);
643 if (first_stmt_code
!= rhs_code
644 && alt_stmt_code
== ERROR_MARK
)
645 alt_stmt_code
= rhs_code
;
646 if (first_stmt_code
!= rhs_code
647 && (first_stmt_code
!= IMAGPART_EXPR
648 || rhs_code
!= REALPART_EXPR
)
649 && (first_stmt_code
!= REALPART_EXPR
650 || rhs_code
!= IMAGPART_EXPR
)
651 /* Handle mismatches in plus/minus by computing both
652 and merging the results. */
653 && !((first_stmt_code
== PLUS_EXPR
654 || first_stmt_code
== MINUS_EXPR
)
655 && (alt_stmt_code
== PLUS_EXPR
656 || alt_stmt_code
== MINUS_EXPR
)
657 && rhs_code
== alt_stmt_code
)
658 && !(STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt
))
659 && (first_stmt_code
== ARRAY_REF
660 || first_stmt_code
== BIT_FIELD_REF
661 || first_stmt_code
== INDIRECT_REF
662 || first_stmt_code
== COMPONENT_REF
663 || first_stmt_code
== MEM_REF
)))
665 if (dump_enabled_p ())
667 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
668 "Build SLP failed: different operation "
670 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
671 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
673 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
681 && !operand_equal_p (first_op1
, gimple_assign_rhs2 (stmt
), 0))
683 if (dump_enabled_p ())
685 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
686 "Build SLP failed: different shift "
688 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
689 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
695 if (rhs_code
== CALL_EXPR
)
697 gimple
*first_stmt
= stmts
[0];
698 if (gimple_call_num_args (stmt
) != nops
699 || !operand_equal_p (gimple_call_fn (first_stmt
),
700 gimple_call_fn (stmt
), 0)
701 || gimple_call_fntype (first_stmt
)
702 != gimple_call_fntype (stmt
))
704 if (dump_enabled_p ())
706 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
707 "Build SLP failed: different calls in ");
708 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
710 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
718 /* Grouped store or load. */
719 if (STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt
)))
721 if (REFERENCE_CLASS_P (lhs
))
729 first_load
= GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
));
732 /* Check that there are no loads from different interleaving
733 chains in the same node. */
734 if (prev_first_load
!= first_load
)
736 if (dump_enabled_p ())
738 dump_printf_loc (MSG_MISSED_OPTIMIZATION
,
740 "Build SLP failed: different "
741 "interleaving chains in one node ");
742 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
744 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
751 prev_first_load
= first_load
;
753 } /* Grouped access. */
756 if (TREE_CODE_CLASS (rhs_code
) == tcc_reference
)
758 /* Not grouped load. */
759 if (dump_enabled_p ())
761 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
762 "Build SLP failed: not grouped load ");
763 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
764 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
767 /* FORNOW: Not grouped loads are not supported. */
768 /* Fatal mismatch. */
773 /* Not memory operation. */
774 if (TREE_CODE_CLASS (rhs_code
) != tcc_binary
775 && TREE_CODE_CLASS (rhs_code
) != tcc_unary
776 && TREE_CODE_CLASS (rhs_code
) != tcc_expression
777 && TREE_CODE_CLASS (rhs_code
) != tcc_comparison
778 && rhs_code
!= CALL_EXPR
)
780 if (dump_enabled_p ())
782 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
783 "Build SLP failed: operation");
784 dump_printf (MSG_MISSED_OPTIMIZATION
, " unsupported ");
785 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
786 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
788 /* Fatal mismatch. */
793 if (rhs_code
== COND_EXPR
)
795 tree cond_expr
= gimple_assign_rhs1 (stmt
);
798 first_cond_code
= TREE_CODE (cond_expr
);
799 else if (first_cond_code
!= TREE_CODE (cond_expr
))
801 if (dump_enabled_p ())
803 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
804 "Build SLP failed: different"
806 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
808 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
819 for (i
= 0; i
< group_size
; ++i
)
823 /* If we allowed a two-operation SLP node verify the target can cope
824 with the permute we are going to use. */
825 if (alt_stmt_code
!= ERROR_MARK
826 && TREE_CODE_CLASS (alt_stmt_code
) != tcc_reference
)
829 = XALLOCAVEC (unsigned char, TYPE_VECTOR_SUBPARTS (vectype
));
830 for (i
= 0; i
< TYPE_VECTOR_SUBPARTS (vectype
); ++i
)
833 if (gimple_assign_rhs_code (stmts
[i
% group_size
]) == alt_stmt_code
)
834 sel
[i
] += TYPE_VECTOR_SUBPARTS (vectype
);
836 if (!can_vec_perm_p (TYPE_MODE (vectype
), false, sel
))
838 for (i
= 0; i
< group_size
; ++i
)
839 if (gimple_assign_rhs_code (stmts
[i
]) == alt_stmt_code
)
842 if (dump_enabled_p ())
844 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
845 "Build SLP failed: different operation "
847 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
849 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
851 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
857 *two_operators
= true;
863 /* Recursively build an SLP tree starting from NODE.
864 Fail (and return a value not equal to zero) if def-stmts are not
865 isomorphic, require data permutation or are of unsupported types of
866 operation. Otherwise, return 0.
867 The value returned is the depth in the SLP tree where a mismatch
871 vect_build_slp_tree (vec_info
*vinfo
,
872 vec
<gimple
*> stmts
, unsigned int group_size
,
873 unsigned int *max_nunits
,
874 vec
<slp_tree
> *loads
,
875 bool *matches
, unsigned *npermutes
, unsigned *tree_size
,
876 unsigned max_tree_size
)
878 unsigned nops
, i
, this_tree_size
= 0, this_max_nunits
= *max_nunits
;
885 if (is_gimple_call (stmt
))
886 nops
= gimple_call_num_args (stmt
);
887 else if (is_gimple_assign (stmt
))
889 nops
= gimple_num_ops (stmt
) - 1;
890 if (gimple_assign_rhs_code (stmt
) == COND_EXPR
)
896 bool two_operators
= false;
897 if (!vect_build_slp_tree_1 (vinfo
,
898 stmts
, group_size
, nops
,
899 &this_max_nunits
, matches
, &two_operators
))
902 /* If the SLP node is a load, terminate the recursion. */
903 if (STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt
))
904 && DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
))))
906 *max_nunits
= this_max_nunits
;
907 node
= vect_create_new_slp_node (stmts
);
908 loads
->safe_push (node
);
912 /* Get at the operands, verifying they are compatible. */
913 vec
<slp_oprnd_info
> oprnds_info
= vect_create_oprnd_info (nops
, group_size
);
914 slp_oprnd_info oprnd_info
;
915 FOR_EACH_VEC_ELT (stmts
, i
, stmt
)
917 switch (vect_get_and_check_slp_defs (vinfo
, stmt
, i
, &oprnds_info
))
923 vect_free_oprnd_info (oprnds_info
);
930 for (i
= 0; i
< group_size
; ++i
)
933 vect_free_oprnd_info (oprnds_info
);
937 auto_vec
<slp_tree
, 4> children
;
938 auto_vec
<slp_tree
> this_loads
;
942 /* Create SLP_TREE nodes for the definition node/s. */
943 FOR_EACH_VEC_ELT (oprnds_info
, i
, oprnd_info
)
946 unsigned old_nloads
= this_loads
.length ();
947 unsigned old_tree_size
= this_tree_size
;
950 if (oprnd_info
->first_dt
!= vect_internal_def
)
953 if (++this_tree_size
> max_tree_size
)
955 FOR_EACH_VEC_ELT (children
, j
, child
)
956 vect_free_slp_tree (child
);
957 vect_free_oprnd_info (oprnds_info
);
961 if ((child
= vect_build_slp_tree (vinfo
, oprnd_info
->def_stmts
,
962 group_size
, &this_max_nunits
,
963 &this_loads
, matches
, npermutes
,
965 max_tree_size
)) != NULL
)
967 /* If we have all children of child built up from scalars then just
968 throw that away and build it up this node from scalars. */
969 if (!SLP_TREE_CHILDREN (child
).is_empty ()
970 /* ??? Rejecting patterns this way doesn't work. We'd have to
971 do extra work to cancel the pattern so the uses see the
973 && !is_pattern_stmt_p
974 (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child
)[0])))
978 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child
), j
, grandchild
)
979 if (SLP_TREE_DEF_TYPE (grandchild
) == vect_internal_def
)
984 this_loads
.truncate (old_nloads
);
985 this_tree_size
= old_tree_size
;
986 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child
), j
, grandchild
)
987 vect_free_slp_tree (grandchild
);
988 SLP_TREE_CHILDREN (child
).truncate (0);
990 dump_printf_loc (MSG_NOTE
, vect_location
,
991 "Building parent vector operands from "
992 "scalars instead\n");
993 oprnd_info
->def_stmts
= vNULL
;
994 SLP_TREE_DEF_TYPE (child
) = vect_external_def
;
995 children
.safe_push (child
);
1000 oprnd_info
->def_stmts
= vNULL
;
1001 children
.safe_push (child
);
1005 /* If the SLP build failed fatally and we analyze a basic-block
1006 simply treat nodes we fail to build as externally defined
1007 (and thus build vectors from the scalar defs).
1008 The cost model will reject outright expensive cases.
1009 ??? This doesn't treat cases where permutation ultimatively
1010 fails (or we don't try permutation below). Ideally we'd
1011 even compute a permutation that will end up with the maximum
1013 if (is_a
<bb_vec_info
> (vinfo
)
1015 /* ??? Rejecting patterns this way doesn't work. We'd have to
1016 do extra work to cancel the pattern so the uses see the
1018 && !is_pattern_stmt_p (vinfo_for_stmt (stmt
)))
1020 dump_printf_loc (MSG_NOTE
, vect_location
,
1021 "Building vector operands from scalars\n");
1022 child
= vect_create_new_slp_node (oprnd_info
->def_stmts
);
1023 SLP_TREE_DEF_TYPE (child
) = vect_external_def
;
1024 children
.safe_push (child
);
1025 oprnd_info
->def_stmts
= vNULL
;
1029 /* If the SLP build for operand zero failed and operand zero
1030 and one can be commutated try that for the scalar stmts
1031 that failed the match. */
1033 /* A first scalar stmt mismatch signals a fatal mismatch. */
1035 /* ??? For COND_EXPRs we can swap the comparison operands
1036 as well as the arms under some constraints. */
1038 && oprnds_info
[1]->first_dt
== vect_internal_def
1039 && is_gimple_assign (stmt
)
1040 && commutative_tree_code (gimple_assign_rhs_code (stmt
))
1042 /* Do so only if the number of not successful permutes was nor more
1043 than a cut-ff as re-trying the recursive match on
1044 possibly each level of the tree would expose exponential
1048 /* Verify if we can safely swap or if we committed to a specific
1049 operand order already. */
1050 for (j
= 0; j
< group_size
; ++j
)
1052 && STMT_VINFO_NUM_SLP_USES (vinfo_for_stmt (stmts
[j
])) != 0)
1054 if (dump_enabled_p ())
1056 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1057 "Build SLP failed: cannot swap operands "
1059 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
1065 /* Swap mismatched definition stmts. */
1066 dump_printf_loc (MSG_NOTE
, vect_location
,
1067 "Re-trying with swapped operands of stmts ");
1068 for (j
= 0; j
< group_size
; ++j
)
1071 std::swap (oprnds_info
[0]->def_stmts
[j
],
1072 oprnds_info
[1]->def_stmts
[j
]);
1073 dump_printf (MSG_NOTE
, "%d ", j
);
1075 dump_printf (MSG_NOTE
, "\n");
1076 /* And try again with scratch 'matches' ... */
1077 bool *tem
= XALLOCAVEC (bool, group_size
);
1078 if ((child
= vect_build_slp_tree (vinfo
, oprnd_info
->def_stmts
,
1079 group_size
, &this_max_nunits
,
1080 &this_loads
, tem
, npermutes
,
1082 max_tree_size
)) != NULL
)
1084 /* ... so if successful we can apply the operand swapping
1085 to the GIMPLE IL. This is necessary because for example
1086 vect_get_slp_defs uses operand indexes and thus expects
1087 canonical operand order. This is also necessary even
1088 if we end up building the operand from scalars as
1089 we'll continue to process swapped operand two. */
1090 for (j
= 0; j
< group_size
; ++j
)
1092 gimple
*stmt
= stmts
[j
];
1093 gimple_set_plf (stmt
, GF_PLF_1
, false);
1095 for (j
= 0; j
< group_size
; ++j
)
1097 gimple
*stmt
= stmts
[j
];
1100 /* Avoid swapping operands twice. */
1101 if (gimple_plf (stmt
, GF_PLF_1
))
1103 swap_ssa_operands (stmt
, gimple_assign_rhs1_ptr (stmt
),
1104 gimple_assign_rhs2_ptr (stmt
));
1105 gimple_set_plf (stmt
, GF_PLF_1
, true);
1108 /* Verify we swap all duplicates or none. */
1110 for (j
= 0; j
< group_size
; ++j
)
1112 gimple
*stmt
= stmts
[j
];
1113 gcc_assert (gimple_plf (stmt
, GF_PLF_1
) == ! matches
[j
]);
1116 /* If we have all children of child built up from scalars then
1117 just throw that away and build it up this node from scalars. */
1118 if (!SLP_TREE_CHILDREN (child
).is_empty ()
1119 /* ??? Rejecting patterns this way doesn't work. We'd have
1120 to do extra work to cancel the pattern so the uses see the
1122 && !is_pattern_stmt_p
1123 (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child
)[0])))
1126 slp_tree grandchild
;
1128 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child
), j
, grandchild
)
1129 if (SLP_TREE_DEF_TYPE (grandchild
) == vect_internal_def
)
1134 this_loads
.truncate (old_nloads
);
1135 this_tree_size
= old_tree_size
;
1136 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (child
), j
, grandchild
)
1137 vect_free_slp_tree (grandchild
);
1138 SLP_TREE_CHILDREN (child
).truncate (0);
1140 dump_printf_loc (MSG_NOTE
, vect_location
,
1141 "Building parent vector operands from "
1142 "scalars instead\n");
1143 oprnd_info
->def_stmts
= vNULL
;
1144 SLP_TREE_DEF_TYPE (child
) = vect_external_def
;
1145 children
.safe_push (child
);
1150 oprnd_info
->def_stmts
= vNULL
;
1151 children
.safe_push (child
);
1159 gcc_assert (child
== NULL
);
1160 FOR_EACH_VEC_ELT (children
, j
, child
)
1161 vect_free_slp_tree (child
);
1162 vect_free_oprnd_info (oprnds_info
);
1166 vect_free_oprnd_info (oprnds_info
);
1169 *tree_size
+= this_tree_size
;
1170 *max_nunits
= this_max_nunits
;
1171 loads
->safe_splice (this_loads
);
1173 node
= vect_create_new_slp_node (stmts
);
1174 SLP_TREE_TWO_OPERATORS (node
) = two_operators
;
1175 SLP_TREE_CHILDREN (node
).splice (children
);
1179 /* Dump a slp tree NODE using flags specified in DUMP_KIND. */
1182 vect_print_slp_tree (int dump_kind
, location_t loc
, slp_tree node
)
1188 dump_printf_loc (dump_kind
, loc
, "node%s\n",
1189 SLP_TREE_DEF_TYPE (node
) != vect_internal_def
1190 ? " (external)" : "");
1191 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
1193 dump_printf_loc (dump_kind
, loc
, "\tstmt %d ", i
);
1194 dump_gimple_stmt (dump_kind
, TDF_SLIM
, stmt
, 0);
1196 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1197 vect_print_slp_tree (dump_kind
, loc
, child
);
1201 /* Mark the tree rooted at NODE with MARK (PURE_SLP or HYBRID).
1202 If MARK is HYBRID, it refers to a specific stmt in NODE (the stmt at index
1203 J). Otherwise, MARK is PURE_SLP and J is -1, which indicates that all the
1204 stmts in NODE are to be marked. */
1207 vect_mark_slp_stmts (slp_tree node
, enum slp_vect_type mark
, int j
)
1213 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
1216 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
1217 if (j
< 0 || i
== j
)
1218 STMT_SLP_TYPE (vinfo_for_stmt (stmt
)) = mark
;
1220 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1221 vect_mark_slp_stmts (child
, mark
, j
);
1225 /* Mark the statements of the tree rooted at NODE as relevant (vect_used). */
1228 vect_mark_slp_stmts_relevant (slp_tree node
)
1232 stmt_vec_info stmt_info
;
1235 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
1238 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
1240 stmt_info
= vinfo_for_stmt (stmt
);
1241 gcc_assert (!STMT_VINFO_RELEVANT (stmt_info
)
1242 || STMT_VINFO_RELEVANT (stmt_info
) == vect_used_in_scope
);
1243 STMT_VINFO_RELEVANT (stmt_info
) = vect_used_in_scope
;
1246 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1247 vect_mark_slp_stmts_relevant (child
);
1251 /* Rearrange the statements of NODE according to PERMUTATION. */
1254 vect_slp_rearrange_stmts (slp_tree node
, unsigned int group_size
,
1255 vec
<unsigned> permutation
)
1258 vec
<gimple
*> tmp_stmts
;
1262 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1263 vect_slp_rearrange_stmts (child
, group_size
, permutation
);
1265 gcc_assert (group_size
== SLP_TREE_SCALAR_STMTS (node
).length ());
1266 tmp_stmts
.create (group_size
);
1267 tmp_stmts
.quick_grow_cleared (group_size
);
1269 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
1270 tmp_stmts
[permutation
[i
]] = stmt
;
1272 SLP_TREE_SCALAR_STMTS (node
).release ();
1273 SLP_TREE_SCALAR_STMTS (node
) = tmp_stmts
;
1277 /* Attempt to reorder stmts in a reduction chain so that we don't
1278 require any load permutation. Return true if that was possible,
1279 otherwise return false. */
1282 vect_attempt_slp_rearrange_stmts (slp_instance slp_instn
)
1284 unsigned int group_size
= SLP_INSTANCE_GROUP_SIZE (slp_instn
);
1288 slp_tree node
, load
;
1290 /* Compare all the permutation sequences to the first one. We know
1291 that at least one load is permuted. */
1292 node
= SLP_INSTANCE_LOADS (slp_instn
)[0];
1293 if (!node
->load_permutation
.exists ())
1295 for (i
= 1; SLP_INSTANCE_LOADS (slp_instn
).iterate (i
, &load
); ++i
)
1297 if (!load
->load_permutation
.exists ())
1299 FOR_EACH_VEC_ELT (load
->load_permutation
, j
, lidx
)
1300 if (lidx
!= node
->load_permutation
[j
])
1304 /* Check that the loads in the first sequence are different and there
1305 are no gaps between them. */
1306 load_index
= sbitmap_alloc (group_size
);
1307 bitmap_clear (load_index
);
1308 FOR_EACH_VEC_ELT (node
->load_permutation
, i
, lidx
)
1310 if (lidx
>= group_size
)
1312 if (bitmap_bit_p (load_index
, lidx
))
1314 sbitmap_free (load_index
);
1317 bitmap_set_bit (load_index
, lidx
);
1319 for (i
= 0; i
< group_size
; i
++)
1320 if (!bitmap_bit_p (load_index
, i
))
1322 sbitmap_free (load_index
);
1325 sbitmap_free (load_index
);
1327 /* This permutation is valid for reduction. Since the order of the
1328 statements in the nodes is not important unless they are memory
1329 accesses, we can rearrange the statements in all the nodes
1330 according to the order of the loads. */
1331 vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn
), group_size
,
1332 node
->load_permutation
);
1334 /* We are done, no actual permutations need to be generated. */
1335 unsigned int unrolling_factor
= SLP_INSTANCE_UNROLLING_FACTOR (slp_instn
);
1336 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1338 gimple
*first_stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1339 first_stmt
= GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt
));
1340 /* But we have to keep those permutations that are required because
1341 of handling of gaps. */
1342 if (unrolling_factor
== 1
1343 || (group_size
== GROUP_SIZE (vinfo_for_stmt (first_stmt
))
1344 && GROUP_GAP (vinfo_for_stmt (first_stmt
)) == 0))
1345 SLP_TREE_LOAD_PERMUTATION (node
).release ();
1351 /* Check if the required load permutations in the SLP instance
1352 SLP_INSTN are supported. */
1355 vect_supported_load_permutation_p (slp_instance slp_instn
)
1357 unsigned int group_size
= SLP_INSTANCE_GROUP_SIZE (slp_instn
);
1358 unsigned int i
, j
, k
, next
;
1360 gimple
*stmt
, *load
, *next_load
;
1362 if (dump_enabled_p ())
1364 dump_printf_loc (MSG_NOTE
, vect_location
, "Load permutation ");
1365 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1366 if (node
->load_permutation
.exists ())
1367 FOR_EACH_VEC_ELT (node
->load_permutation
, j
, next
)
1368 dump_printf (MSG_NOTE
, "%d ", next
);
1370 for (k
= 0; k
< group_size
; ++k
)
1371 dump_printf (MSG_NOTE
, "%d ", k
);
1372 dump_printf (MSG_NOTE
, "\n");
1375 /* In case of reduction every load permutation is allowed, since the order
1376 of the reduction statements is not important (as opposed to the case of
1377 grouped stores). The only condition we need to check is that all the
1378 load nodes are of the same size and have the same permutation (and then
1379 rearrange all the nodes of the SLP instance according to this
1382 /* Check that all the load nodes are of the same size. */
1383 /* ??? Can't we assert this? */
1384 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1385 if (SLP_TREE_SCALAR_STMTS (node
).length () != (unsigned) group_size
)
1388 node
= SLP_INSTANCE_TREE (slp_instn
);
1389 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1391 /* Reduction (there are no data-refs in the root).
1392 In reduction chain the order of the loads is not important. */
1393 if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
))
1394 && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
1395 vect_attempt_slp_rearrange_stmts (slp_instn
);
1397 /* In basic block vectorization we allow any subchain of an interleaving
1399 FORNOW: not supported in loop SLP because of realignment compications. */
1400 if (STMT_VINFO_BB_VINFO (vinfo_for_stmt (stmt
)))
1402 /* Check whether the loads in an instance form a subchain and thus
1403 no permutation is necessary. */
1404 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1406 if (!SLP_TREE_LOAD_PERMUTATION (node
).exists ())
1408 bool subchain_p
= true;
1410 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), j
, load
)
1413 && (next_load
!= load
1414 || GROUP_GAP (vinfo_for_stmt (load
)) != 1))
1419 next_load
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (load
));
1422 SLP_TREE_LOAD_PERMUTATION (node
).release ();
1425 /* Verify the permutation can be generated. */
1427 if (!vect_transform_slp_perm_load (node
, tem
, NULL
,
1428 1, slp_instn
, true))
1430 dump_printf_loc (MSG_MISSED_OPTIMIZATION
,
1432 "unsupported load permutation\n");
1440 /* For loop vectorization verify we can generate the permutation. */
1441 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1442 if (node
->load_permutation
.exists ()
1443 && !vect_transform_slp_perm_load
1445 SLP_INSTANCE_UNROLLING_FACTOR (slp_instn
), slp_instn
, true))
1452 /* Find the last store in SLP INSTANCE. */
1455 vect_find_last_scalar_stmt_in_slp (slp_tree node
)
1457 gimple
*last
= NULL
, *stmt
;
1459 for (int i
= 0; SLP_TREE_SCALAR_STMTS (node
).iterate (i
, &stmt
); i
++)
1461 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
1462 if (is_pattern_stmt_p (stmt_vinfo
))
1463 last
= get_later_stmt (STMT_VINFO_RELATED_STMT (stmt_vinfo
), last
);
1465 last
= get_later_stmt (stmt
, last
);
1471 /* Compute the cost for the SLP node NODE in the SLP instance INSTANCE. */
1474 vect_analyze_slp_cost_1 (slp_instance instance
, slp_tree node
,
1475 stmt_vector_for_cost
*prologue_cost_vec
,
1476 stmt_vector_for_cost
*body_cost_vec
,
1477 unsigned ncopies_for_cost
)
1482 stmt_vec_info stmt_info
;
1485 /* Recurse down the SLP tree. */
1486 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1487 if (SLP_TREE_DEF_TYPE (child
) == vect_internal_def
)
1488 vect_analyze_slp_cost_1 (instance
, child
, prologue_cost_vec
,
1489 body_cost_vec
, ncopies_for_cost
);
1491 /* Look at the first scalar stmt to determine the cost. */
1492 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1493 stmt_info
= vinfo_for_stmt (stmt
);
1494 if (STMT_VINFO_GROUPED_ACCESS (stmt_info
))
1496 if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info
)))
1497 vect_model_store_cost (stmt_info
, ncopies_for_cost
, false,
1498 vect_uninitialized_def
,
1499 node
, prologue_cost_vec
, body_cost_vec
);
1502 gcc_checking_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info
)));
1503 if (SLP_TREE_LOAD_PERMUTATION (node
).exists ())
1505 /* If the load is permuted then the alignment is determined by
1506 the first group element not by the first scalar stmt DR. */
1507 stmt
= GROUP_FIRST_ELEMENT (stmt_info
);
1508 stmt_info
= vinfo_for_stmt (stmt
);
1509 /* Record the cost for the permutation. */
1510 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vec_perm
,
1511 stmt_info
, 0, vect_body
);
1512 /* And adjust the number of loads performed. */
1514 = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
));
1516 = (GROUP_SIZE (stmt_info
) - GROUP_GAP (stmt_info
)
1517 + nunits
- 1) / nunits
;
1518 ncopies_for_cost
*= SLP_INSTANCE_UNROLLING_FACTOR (instance
);
1520 /* Record the cost for the vector loads. */
1521 vect_model_load_cost (stmt_info
, ncopies_for_cost
, false,
1522 node
, prologue_cost_vec
, body_cost_vec
);
1528 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vector_stmt
,
1529 stmt_info
, 0, vect_body
);
1530 if (SLP_TREE_TWO_OPERATORS (node
))
1532 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vector_stmt
,
1533 stmt_info
, 0, vect_body
);
1534 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vec_perm
,
1535 stmt_info
, 0, vect_body
);
1539 /* Push SLP node def-type to stmts. */
1540 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1541 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
1542 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
1543 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = SLP_TREE_DEF_TYPE (child
);
1545 /* Scan operands and account for prologue cost of constants/externals.
1546 ??? This over-estimates cost for multiple uses and should be
1548 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1549 lhs
= gimple_get_lhs (stmt
);
1550 for (i
= 0; i
< gimple_num_ops (stmt
); ++i
)
1552 tree op
= gimple_op (stmt
, i
);
1554 enum vect_def_type dt
;
1555 if (!op
|| op
== lhs
)
1557 if (vect_is_simple_use (op
, stmt_info
->vinfo
, &def_stmt
, &dt
))
1559 /* Without looking at the actual initializer a vector of
1560 constants can be implemented as load from the constant pool.
1561 ??? We need to pass down stmt_info for a vector type
1562 even if it points to the wrong stmt. */
1563 if (dt
== vect_constant_def
)
1564 record_stmt_cost (prologue_cost_vec
, 1, vector_load
,
1565 stmt_info
, 0, vect_prologue
);
1566 else if (dt
== vect_external_def
)
1567 record_stmt_cost (prologue_cost_vec
, 1, vec_construct
,
1568 stmt_info
, 0, vect_prologue
);
1572 /* Restore stmt def-types. */
1573 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1574 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
1575 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
1576 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = vect_internal_def
;
1579 /* Compute the cost for the SLP instance INSTANCE. */
1582 vect_analyze_slp_cost (slp_instance instance
, void *data
)
1584 stmt_vector_for_cost body_cost_vec
, prologue_cost_vec
;
1585 unsigned ncopies_for_cost
;
1586 stmt_info_for_cost
*si
;
1589 if (dump_enabled_p ())
1590 dump_printf_loc (MSG_NOTE
, vect_location
,
1591 "=== vect_analyze_slp_cost ===\n");
1593 /* Calculate the number of vector stmts to create based on the unrolling
1594 factor (number of vectors is 1 if NUNITS >= GROUP_SIZE, and is
1595 GROUP_SIZE / NUNITS otherwise. */
1596 unsigned group_size
= SLP_INSTANCE_GROUP_SIZE (instance
);
1597 slp_tree node
= SLP_INSTANCE_TREE (instance
);
1598 stmt_vec_info stmt_info
= vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (node
)[0]);
1599 /* Adjust the group_size by the vectorization factor which is always one
1600 for basic-block vectorization. */
1601 if (STMT_VINFO_LOOP_VINFO (stmt_info
))
1602 group_size
*= LOOP_VINFO_VECT_FACTOR (STMT_VINFO_LOOP_VINFO (stmt_info
));
1603 unsigned nunits
= TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
));
1604 /* For reductions look at a reduction operand in case the reduction
1605 operation is widening like DOT_PROD or SAD. */
1606 if (!STMT_VINFO_GROUPED_ACCESS (stmt_info
))
1608 gimple
*stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1609 switch (gimple_assign_rhs_code (stmt
))
1613 nunits
= TYPE_VECTOR_SUBPARTS (get_vectype_for_scalar_type
1614 (TREE_TYPE (gimple_assign_rhs1 (stmt
))));
1619 ncopies_for_cost
= least_common_multiple (nunits
, group_size
) / nunits
;
1621 prologue_cost_vec
.create (10);
1622 body_cost_vec
.create (10);
1623 vect_analyze_slp_cost_1 (instance
, SLP_INSTANCE_TREE (instance
),
1624 &prologue_cost_vec
, &body_cost_vec
,
1627 /* Record the prologue costs, which were delayed until we were
1628 sure that SLP was successful. */
1629 FOR_EACH_VEC_ELT (prologue_cost_vec
, i
, si
)
1631 struct _stmt_vec_info
*stmt_info
1632 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
1633 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
1634 si
->misalign
, vect_prologue
);
1637 /* Record the instance's instructions in the target cost model. */
1638 FOR_EACH_VEC_ELT (body_cost_vec
, i
, si
)
1640 struct _stmt_vec_info
*stmt_info
1641 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
1642 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
1643 si
->misalign
, vect_body
);
1646 prologue_cost_vec
.release ();
1647 body_cost_vec
.release ();
1650 /* Splits a group of stores, currently beginning at FIRST_STMT, into two groups:
1651 one (still beginning at FIRST_STMT) of size GROUP1_SIZE (also containing
1652 the first GROUP1_SIZE stmts, since stores are consecutive), the second
1653 containing the remainder.
1654 Return the first stmt in the second group. */
1657 vect_split_slp_store_group (gimple
*first_stmt
, unsigned group1_size
)
1659 stmt_vec_info first_vinfo
= vinfo_for_stmt (first_stmt
);
1660 gcc_assert (GROUP_FIRST_ELEMENT (first_vinfo
) == first_stmt
);
1661 gcc_assert (group1_size
> 0);
1662 int group2_size
= GROUP_SIZE (first_vinfo
) - group1_size
;
1663 gcc_assert (group2_size
> 0);
1664 GROUP_SIZE (first_vinfo
) = group1_size
;
1666 gimple
*stmt
= first_stmt
;
1667 for (unsigned i
= group1_size
; i
> 1; i
--)
1669 stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
));
1670 gcc_assert (GROUP_GAP (vinfo_for_stmt (stmt
)) == 1);
1672 /* STMT is now the last element of the first group. */
1673 gimple
*group2
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
));
1674 GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
)) = 0;
1676 GROUP_SIZE (vinfo_for_stmt (group2
)) = group2_size
;
1677 for (stmt
= group2
; stmt
; stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
)))
1679 GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)) = group2
;
1680 gcc_assert (GROUP_GAP (vinfo_for_stmt (stmt
)) == 1);
1683 /* For the second group, the GROUP_GAP is that before the original group,
1684 plus skipping over the first vector. */
1685 GROUP_GAP (vinfo_for_stmt (group2
)) =
1686 GROUP_GAP (first_vinfo
) + group1_size
;
1688 /* GROUP_GAP of the first group now has to skip over the second group too. */
1689 GROUP_GAP (first_vinfo
) += group2_size
;
1691 if (dump_enabled_p ())
1692 dump_printf_loc (MSG_NOTE
, vect_location
, "Split group into %d and %d\n",
1693 group1_size
, group2_size
);
1698 /* Analyze an SLP instance starting from a group of grouped stores. Call
1699 vect_build_slp_tree to build a tree of packed stmts if possible.
1700 Return FALSE if it's impossible to SLP any stmt in the loop. */
1703 vect_analyze_slp_instance (vec_info
*vinfo
,
1704 gimple
*stmt
, unsigned max_tree_size
)
1706 slp_instance new_instance
;
1708 unsigned int group_size
= GROUP_SIZE (vinfo_for_stmt (stmt
));
1709 unsigned int unrolling_factor
= 1, nunits
;
1710 tree vectype
, scalar_type
= NULL_TREE
;
1713 unsigned int max_nunits
= 0;
1714 vec
<slp_tree
> loads
;
1715 struct data_reference
*dr
= STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
));
1716 vec
<gimple
*> scalar_stmts
;
1718 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
1722 scalar_type
= TREE_TYPE (DR_REF (dr
));
1723 vectype
= get_vectype_for_scalar_type (scalar_type
);
1727 gcc_assert (is_a
<loop_vec_info
> (vinfo
));
1728 vectype
= STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt
));
1731 group_size
= GROUP_SIZE (vinfo_for_stmt (stmt
));
1735 gcc_assert (is_a
<loop_vec_info
> (vinfo
));
1736 vectype
= STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt
));
1737 group_size
= as_a
<loop_vec_info
> (vinfo
)->reductions
.length ();
1742 if (dump_enabled_p ())
1744 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1745 "Build SLP failed: unsupported data-type ");
1746 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, scalar_type
);
1747 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
1752 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1754 /* Calculate the unrolling factor. */
1755 unrolling_factor
= least_common_multiple (nunits
, group_size
) / group_size
;
1756 if (unrolling_factor
!= 1 && is_a
<bb_vec_info
> (vinfo
))
1758 if (dump_enabled_p ())
1759 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1760 "Build SLP failed: unrolling required in basic"
1766 /* Create a node (a root of the SLP tree) for the packed grouped stores. */
1767 scalar_stmts
.create (group_size
);
1769 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
1771 /* Collect the stores and store them in SLP_TREE_SCALAR_STMTS. */
1774 if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (next
))
1775 && STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next
)))
1776 scalar_stmts
.safe_push (
1777 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next
)));
1779 scalar_stmts
.safe_push (next
);
1780 next
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (next
));
1782 /* Mark the first element of the reduction chain as reduction to properly
1783 transform the node. In the reduction analysis phase only the last
1784 element of the chain is marked as reduction. */
1785 if (!STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt
)))
1786 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = vect_reduction_def
;
1790 /* Collect reduction statements. */
1791 vec
<gimple
*> reductions
= as_a
<loop_vec_info
> (vinfo
)->reductions
;
1792 for (i
= 0; reductions
.iterate (i
, &next
); i
++)
1793 scalar_stmts
.safe_push (next
);
1796 loads
.create (group_size
);
1798 /* Build the tree for the SLP instance. */
1799 bool *matches
= XALLOCAVEC (bool, group_size
);
1800 unsigned npermutes
= 0;
1801 if ((node
= vect_build_slp_tree (vinfo
, scalar_stmts
, group_size
,
1802 &max_nunits
, &loads
, matches
, &npermutes
,
1803 NULL
, max_tree_size
)) != NULL
)
1805 /* Calculate the unrolling factor based on the smallest type. */
1806 if (max_nunits
> nunits
)
1807 unrolling_factor
= least_common_multiple (max_nunits
, group_size
)
1810 if (unrolling_factor
!= 1 && is_a
<bb_vec_info
> (vinfo
))
1812 if (dump_enabled_p ())
1813 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1814 "Build SLP failed: unrolling required in basic"
1816 vect_free_slp_tree (node
);
1821 /* Create a new SLP instance. */
1822 new_instance
= XNEW (struct _slp_instance
);
1823 SLP_INSTANCE_TREE (new_instance
) = node
;
1824 SLP_INSTANCE_GROUP_SIZE (new_instance
) = group_size
;
1825 SLP_INSTANCE_UNROLLING_FACTOR (new_instance
) = unrolling_factor
;
1826 SLP_INSTANCE_LOADS (new_instance
) = loads
;
1828 /* Compute the load permutation. */
1830 bool loads_permuted
= false;
1831 FOR_EACH_VEC_ELT (loads
, i
, load_node
)
1833 vec
<unsigned> load_permutation
;
1835 gimple
*load
, *first_stmt
;
1836 bool this_load_permuted
= false;
1837 load_permutation
.create (group_size
);
1838 first_stmt
= GROUP_FIRST_ELEMENT
1839 (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node
)[0]));
1840 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node
), j
, load
)
1843 = vect_get_place_in_interleaving_chain (load
, first_stmt
);
1844 gcc_assert (load_place
!= -1);
1845 if (load_place
!= j
)
1846 this_load_permuted
= true;
1847 load_permutation
.safe_push (load_place
);
1849 if (!this_load_permuted
1850 /* The load requires permutation when unrolling exposes
1851 a gap either because the group is larger than the SLP
1852 group-size or because there is a gap between the groups. */
1853 && (unrolling_factor
== 1
1854 || (group_size
== GROUP_SIZE (vinfo_for_stmt (first_stmt
))
1855 && GROUP_GAP (vinfo_for_stmt (first_stmt
)) == 0)))
1857 load_permutation
.release ();
1860 SLP_TREE_LOAD_PERMUTATION (load_node
) = load_permutation
;
1861 loads_permuted
= true;
1866 if (!vect_supported_load_permutation_p (new_instance
))
1868 if (dump_enabled_p ())
1870 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1871 "Build SLP failed: unsupported load "
1873 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
1874 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
1876 vect_free_slp_instance (new_instance
);
1881 /* If the loads and stores can be handled with load/store-lane
1882 instructions do not generate this SLP instance. */
1883 if (is_a
<loop_vec_info
> (vinfo
)
1885 && dr
&& vect_store_lanes_supported (vectype
, group_size
))
1888 FOR_EACH_VEC_ELT (loads
, i
, load_node
)
1890 gimple
*first_stmt
= GROUP_FIRST_ELEMENT
1891 (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node
)[0]));
1892 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (first_stmt
);
1893 /* Use SLP for strided accesses (or if we can't load-lanes). */
1894 if (STMT_VINFO_STRIDED_P (stmt_vinfo
)
1895 || ! vect_load_lanes_supported
1896 (STMT_VINFO_VECTYPE (stmt_vinfo
),
1897 GROUP_SIZE (stmt_vinfo
)))
1900 if (i
== loads
.length ())
1902 if (dump_enabled_p ())
1903 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1904 "Built SLP cancelled: can use "
1905 "load/store-lanes\n");
1906 vect_free_slp_instance (new_instance
);
1911 vinfo
->slp_instances
.safe_push (new_instance
);
1913 if (dump_enabled_p ())
1915 dump_printf_loc (MSG_NOTE
, vect_location
,
1916 "Final SLP tree for instance:\n");
1917 vect_print_slp_tree (MSG_NOTE
, vect_location
, node
);
1923 /* Failed to SLP. */
1924 /* Free the allocated memory. */
1925 scalar_stmts
.release ();
1928 /* For basic block SLP, try to break the group up into multiples of the
1930 if (is_a
<bb_vec_info
> (vinfo
)
1931 && GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
))
1932 && STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt
)))
1934 /* We consider breaking the group only on VF boundaries from the existing
1936 for (i
= 0; i
< group_size
; i
++)
1937 if (!matches
[i
]) break;
1939 if (i
>= nunits
&& i
< group_size
)
1941 /* Split into two groups at the first vector boundary before i. */
1942 gcc_assert ((nunits
& (nunits
- 1)) == 0);
1943 unsigned group1_size
= i
& ~(nunits
- 1);
1945 gimple
*rest
= vect_split_slp_store_group (stmt
, group1_size
);
1946 bool res
= vect_analyze_slp_instance (vinfo
, stmt
, max_tree_size
);
1947 /* If the first non-match was in the middle of a vector,
1948 skip the rest of that vector. */
1949 if (group1_size
< i
)
1951 i
= group1_size
+ nunits
;
1953 rest
= vect_split_slp_store_group (rest
, nunits
);
1956 res
|= vect_analyze_slp_instance (vinfo
, rest
, max_tree_size
);
1959 /* Even though the first vector did not all match, we might be able to SLP
1960 (some) of the remainder. FORNOW ignore this possibility. */
1967 /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP
1968 trees of packed scalar stmts if SLP is possible. */
1971 vect_analyze_slp (vec_info
*vinfo
, unsigned max_tree_size
)
1974 gimple
*first_element
;
1977 if (dump_enabled_p ())
1978 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vect_analyze_slp ===\n");
1980 /* Find SLP sequences starting from groups of grouped stores. */
1981 FOR_EACH_VEC_ELT (vinfo
->grouped_stores
, i
, first_element
)
1982 if (vect_analyze_slp_instance (vinfo
, first_element
, max_tree_size
))
1985 if (loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (vinfo
))
1987 if (loop_vinfo
->reduction_chains
.length () > 0)
1989 /* Find SLP sequences starting from reduction chains. */
1990 FOR_EACH_VEC_ELT (loop_vinfo
->reduction_chains
, i
, first_element
)
1991 if (vect_analyze_slp_instance (vinfo
, first_element
,
1997 /* Don't try to vectorize SLP reductions if reduction chain was
2002 /* Find SLP sequences starting from groups of reductions. */
2003 if (loop_vinfo
->reductions
.length () > 1
2004 && vect_analyze_slp_instance (vinfo
, loop_vinfo
->reductions
[0],
2013 /* For each possible SLP instance decide whether to SLP it and calculate overall
2014 unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at
2015 least one instance. */
2018 vect_make_slp_decision (loop_vec_info loop_vinfo
)
2020 unsigned int i
, unrolling_factor
= 1;
2021 vec
<slp_instance
> slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
2022 slp_instance instance
;
2023 int decided_to_slp
= 0;
2025 if (dump_enabled_p ())
2026 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vect_make_slp_decision ==="
2029 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
2031 /* FORNOW: SLP if you can. */
2032 if (unrolling_factor
< SLP_INSTANCE_UNROLLING_FACTOR (instance
))
2033 unrolling_factor
= SLP_INSTANCE_UNROLLING_FACTOR (instance
);
2035 /* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we
2036 call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and
2037 loop-based vectorization. Such stmts will be marked as HYBRID. */
2038 vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance
), pure_slp
, -1);
2042 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
) = unrolling_factor
;
2044 if (decided_to_slp
&& dump_enabled_p ())
2045 dump_printf_loc (MSG_NOTE
, vect_location
,
2046 "Decided to SLP %d instances. Unrolling factor %d\n",
2047 decided_to_slp
, unrolling_factor
);
2049 return (decided_to_slp
> 0);
2053 /* Find stmts that must be both vectorized and SLPed (since they feed stmts that
2054 can't be SLPed) in the tree rooted at NODE. Mark such stmts as HYBRID. */
2057 vect_detect_hybrid_slp_stmts (slp_tree node
, unsigned i
, slp_vect_type stype
)
2059 gimple
*stmt
= SLP_TREE_SCALAR_STMTS (node
)[i
];
2060 imm_use_iterator imm_iter
;
2062 stmt_vec_info use_vinfo
, stmt_vinfo
= vinfo_for_stmt (stmt
);
2064 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
2065 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2068 /* Propagate hybrid down the SLP tree. */
2069 if (stype
== hybrid
)
2071 else if (HYBRID_SLP_STMT (stmt_vinfo
))
2075 /* Check if a pure SLP stmt has uses in non-SLP stmts. */
2076 gcc_checking_assert (PURE_SLP_STMT (stmt_vinfo
));
2077 /* If we get a pattern stmt here we have to use the LHS of the
2078 original stmt for immediate uses. */
2079 if (! STMT_VINFO_IN_PATTERN_P (stmt_vinfo
)
2080 && STMT_VINFO_RELATED_STMT (stmt_vinfo
))
2081 stmt
= STMT_VINFO_RELATED_STMT (stmt_vinfo
);
2082 if (TREE_CODE (gimple_op (stmt
, 0)) == SSA_NAME
)
2083 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, gimple_op (stmt
, 0))
2085 if (!flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
)))
2087 use_vinfo
= vinfo_for_stmt (use_stmt
);
2088 if (STMT_VINFO_IN_PATTERN_P (use_vinfo
)
2089 && STMT_VINFO_RELATED_STMT (use_vinfo
))
2090 use_vinfo
= vinfo_for_stmt (STMT_VINFO_RELATED_STMT (use_vinfo
));
2091 if (!STMT_SLP_TYPE (use_vinfo
)
2092 && (STMT_VINFO_RELEVANT (use_vinfo
)
2093 || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (use_vinfo
)))
2094 && !(gimple_code (use_stmt
) == GIMPLE_PHI
2095 && STMT_VINFO_DEF_TYPE (use_vinfo
) == vect_reduction_def
))
2097 if (dump_enabled_p ())
2099 dump_printf_loc (MSG_NOTE
, vect_location
, "use of SLP "
2100 "def in non-SLP stmt: ");
2101 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, use_stmt
, 0);
2109 && !HYBRID_SLP_STMT (stmt_vinfo
))
2111 if (dump_enabled_p ())
2113 dump_printf_loc (MSG_NOTE
, vect_location
, "marking hybrid: ");
2114 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
2116 STMT_SLP_TYPE (stmt_vinfo
) = hybrid
;
2119 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), j
, child
)
2120 if (SLP_TREE_DEF_TYPE (child
) != vect_external_def
)
2121 vect_detect_hybrid_slp_stmts (child
, i
, stype
);
2124 /* Helpers for vect_detect_hybrid_slp walking pattern stmt uses. */
2127 vect_detect_hybrid_slp_1 (tree
*tp
, int *, void *data
)
2129 walk_stmt_info
*wi
= (walk_stmt_info
*)data
;
2130 struct loop
*loopp
= (struct loop
*)wi
->info
;
2135 if (TREE_CODE (*tp
) == SSA_NAME
2136 && !SSA_NAME_IS_DEFAULT_DEF (*tp
))
2138 gimple
*def_stmt
= SSA_NAME_DEF_STMT (*tp
);
2139 if (flow_bb_inside_loop_p (loopp
, gimple_bb (def_stmt
))
2140 && PURE_SLP_STMT (vinfo_for_stmt (def_stmt
)))
2142 if (dump_enabled_p ())
2144 dump_printf_loc (MSG_NOTE
, vect_location
, "marking hybrid: ");
2145 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, def_stmt
, 0);
2147 STMT_SLP_TYPE (vinfo_for_stmt (def_stmt
)) = hybrid
;
2155 vect_detect_hybrid_slp_2 (gimple_stmt_iterator
*gsi
, bool *handled
,
2158 /* If the stmt is in a SLP instance then this isn't a reason
2159 to mark use definitions in other SLP instances as hybrid. */
2160 if (STMT_SLP_TYPE (vinfo_for_stmt (gsi_stmt (*gsi
))) != loop_vect
)
2165 /* Find stmts that must be both vectorized and SLPed. */
2168 vect_detect_hybrid_slp (loop_vec_info loop_vinfo
)
2171 vec
<slp_instance
> slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
2172 slp_instance instance
;
2174 if (dump_enabled_p ())
2175 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vect_detect_hybrid_slp ==="
2178 /* First walk all pattern stmt in the loop and mark defs of uses as
2179 hybrid because immediate uses in them are not recorded. */
2180 for (i
= 0; i
< LOOP_VINFO_LOOP (loop_vinfo
)->num_nodes
; ++i
)
2182 basic_block bb
= LOOP_VINFO_BBS (loop_vinfo
)[i
];
2183 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
2186 gimple
*stmt
= gsi_stmt (gsi
);
2187 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2188 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
2191 memset (&wi
, 0, sizeof (wi
));
2192 wi
.info
= LOOP_VINFO_LOOP (loop_vinfo
);
2193 gimple_stmt_iterator gsi2
2194 = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info
));
2195 walk_gimple_stmt (&gsi2
, vect_detect_hybrid_slp_2
,
2196 vect_detect_hybrid_slp_1
, &wi
);
2197 walk_gimple_seq (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
),
2198 vect_detect_hybrid_slp_2
,
2199 vect_detect_hybrid_slp_1
, &wi
);
2204 /* Then walk the SLP instance trees marking stmts with uses in
2205 non-SLP stmts as hybrid, also propagating hybrid down the
2206 SLP tree, collecting the above info on-the-fly. */
2207 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
2209 for (unsigned i
= 0; i
< SLP_INSTANCE_GROUP_SIZE (instance
); ++i
)
2210 vect_detect_hybrid_slp_stmts (SLP_INSTANCE_TREE (instance
),
2216 /* Create and initialize a new bb_vec_info struct for BB, as well as
2217 stmt_vec_info structs for all the stmts in it. */
2220 new_bb_vec_info (gimple_stmt_iterator region_begin
,
2221 gimple_stmt_iterator region_end
)
2223 basic_block bb
= gsi_bb (region_begin
);
2224 bb_vec_info res
= NULL
;
2225 gimple_stmt_iterator gsi
;
2227 res
= (bb_vec_info
) xcalloc (1, sizeof (struct _bb_vec_info
));
2228 res
->kind
= vec_info::bb
;
2229 BB_VINFO_BB (res
) = bb
;
2230 res
->region_begin
= region_begin
;
2231 res
->region_end
= region_end
;
2233 for (gsi
= region_begin
; gsi_stmt (gsi
) != gsi_stmt (region_end
);
2236 gimple
*stmt
= gsi_stmt (gsi
);
2237 gimple_set_uid (stmt
, 0);
2238 set_vinfo_for_stmt (stmt
, new_stmt_vec_info (stmt
, res
));
2241 BB_VINFO_GROUPED_STORES (res
).create (10);
2242 BB_VINFO_SLP_INSTANCES (res
).create (2);
2243 BB_VINFO_TARGET_COST_DATA (res
) = init_cost (NULL
);
2250 /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the
2251 stmts in the basic block. */
2254 destroy_bb_vec_info (bb_vec_info bb_vinfo
)
2256 slp_instance instance
;
2262 vect_destroy_datarefs (bb_vinfo
);
2263 free_dependence_relations (BB_VINFO_DDRS (bb_vinfo
));
2264 BB_VINFO_GROUPED_STORES (bb_vinfo
).release ();
2265 FOR_EACH_VEC_ELT (BB_VINFO_SLP_INSTANCES (bb_vinfo
), i
, instance
)
2266 vect_free_slp_instance (instance
);
2267 BB_VINFO_SLP_INSTANCES (bb_vinfo
).release ();
2268 destroy_cost_data (BB_VINFO_TARGET_COST_DATA (bb_vinfo
));
2270 for (gimple_stmt_iterator si
= bb_vinfo
->region_begin
;
2271 gsi_stmt (si
) != gsi_stmt (bb_vinfo
->region_end
); gsi_next (&si
))
2273 gimple
*stmt
= gsi_stmt (si
);
2274 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2277 /* Free stmt_vec_info. */
2278 free_stmt_vec_info (stmt
);
2280 /* Reset region marker. */
2281 gimple_set_uid (stmt
, -1);
2284 BB_VINFO_BB (bb_vinfo
)->aux
= NULL
;
2289 /* Analyze statements contained in SLP tree node after recursively analyzing
2290 the subtree. Return TRUE if the operations are supported. */
2293 vect_slp_analyze_node_operations (slp_tree node
)
2300 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
2303 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
2304 if (!vect_slp_analyze_node_operations (child
))
2308 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
2310 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2311 gcc_assert (stmt_info
);
2312 gcc_assert (STMT_SLP_TYPE (stmt_info
) != loop_vect
);
2314 /* Push SLP node def-type to stmt operands. */
2315 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), j
, child
)
2316 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
2317 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child
)[i
]))
2318 = SLP_TREE_DEF_TYPE (child
);
2319 res
= vect_analyze_stmt (stmt
, &dummy
, node
);
2320 /* Restore def-types. */
2321 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), j
, child
)
2322 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
2323 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child
)[i
]))
2324 = vect_internal_def
;
2333 /* Analyze statements in SLP instances of the basic block. Return TRUE if the
2334 operations are supported. */
2337 vect_slp_analyze_operations (vec
<slp_instance
> slp_instances
, void *data
)
2339 slp_instance instance
;
2342 if (dump_enabled_p ())
2343 dump_printf_loc (MSG_NOTE
, vect_location
,
2344 "=== vect_slp_analyze_operations ===\n");
2346 for (i
= 0; slp_instances
.iterate (i
, &instance
); )
2348 if (!vect_slp_analyze_node_operations (SLP_INSTANCE_TREE (instance
)))
2350 dump_printf_loc (MSG_NOTE
, vect_location
,
2351 "removing SLP instance operations starting from: ");
2352 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
2353 SLP_TREE_SCALAR_STMTS
2354 (SLP_INSTANCE_TREE (instance
))[0], 0);
2355 vect_free_slp_instance (instance
);
2356 slp_instances
.ordered_remove (i
);
2360 /* Compute the costs of the SLP instance. */
2361 vect_analyze_slp_cost (instance
, data
);
2366 if (!slp_instances
.length ())
2373 /* Compute the scalar cost of the SLP node NODE and its children
2374 and return it. Do not account defs that are marked in LIFE and
2375 update LIFE according to uses of NODE. */
2378 vect_bb_slp_scalar_cost (basic_block bb
,
2379 slp_tree node
, vec
<bool, va_heap
> *life
)
2381 unsigned scalar_cost
= 0;
2386 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
2389 ssa_op_iter op_iter
;
2390 def_operand_p def_p
;
2391 stmt_vec_info stmt_info
;
2396 /* If there is a non-vectorized use of the defs then the scalar
2397 stmt is kept live in which case we do not account it or any
2398 required defs in the SLP children in the scalar cost. This
2399 way we make the vectorization more costly when compared to
2401 FOR_EACH_SSA_DEF_OPERAND (def_p
, stmt
, op_iter
, SSA_OP_DEF
)
2403 imm_use_iterator use_iter
;
2405 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, DEF_FROM_PTR (def_p
))
2406 if (!is_gimple_debug (use_stmt
)
2407 && (! vect_stmt_in_region_p (vinfo_for_stmt (stmt
)->vinfo
,
2409 || ! PURE_SLP_STMT (vinfo_for_stmt (use_stmt
))))
2412 BREAK_FROM_IMM_USE_STMT (use_iter
);
2418 /* Count scalar stmts only once. */
2419 if (gimple_visited_p (stmt
))
2421 gimple_set_visited (stmt
, true);
2423 stmt_info
= vinfo_for_stmt (stmt
);
2424 if (STMT_VINFO_DATA_REF (stmt_info
))
2426 if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info
)))
2427 stmt_cost
= vect_get_stmt_cost (scalar_load
);
2429 stmt_cost
= vect_get_stmt_cost (scalar_store
);
2432 stmt_cost
= vect_get_stmt_cost (scalar_stmt
);
2434 scalar_cost
+= stmt_cost
;
2437 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
2438 if (SLP_TREE_DEF_TYPE (child
) == vect_internal_def
)
2439 scalar_cost
+= vect_bb_slp_scalar_cost (bb
, child
, life
);
2444 /* Check if vectorization of the basic block is profitable. */
2447 vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo
)
2449 vec
<slp_instance
> slp_instances
= BB_VINFO_SLP_INSTANCES (bb_vinfo
);
2450 slp_instance instance
;
2452 unsigned int vec_inside_cost
= 0, vec_outside_cost
= 0, scalar_cost
= 0;
2453 unsigned int vec_prologue_cost
= 0, vec_epilogue_cost
= 0;
2455 /* Calculate scalar cost. */
2456 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
2458 auto_vec
<bool, 20> life
;
2459 life
.safe_grow_cleared (SLP_INSTANCE_GROUP_SIZE (instance
));
2460 scalar_cost
+= vect_bb_slp_scalar_cost (BB_VINFO_BB (bb_vinfo
),
2461 SLP_INSTANCE_TREE (instance
),
2465 /* Unset visited flag. */
2466 for (gimple_stmt_iterator gsi
= bb_vinfo
->region_begin
;
2467 gsi_stmt (gsi
) != gsi_stmt (bb_vinfo
->region_end
); gsi_next (&gsi
))
2468 gimple_set_visited (gsi_stmt (gsi
), false);
2470 /* Complete the target-specific cost calculation. */
2471 finish_cost (BB_VINFO_TARGET_COST_DATA (bb_vinfo
), &vec_prologue_cost
,
2472 &vec_inside_cost
, &vec_epilogue_cost
);
2474 vec_outside_cost
= vec_prologue_cost
+ vec_epilogue_cost
;
2476 if (dump_enabled_p ())
2478 dump_printf_loc (MSG_NOTE
, vect_location
, "Cost model analysis: \n");
2479 dump_printf (MSG_NOTE
, " Vector inside of basic block cost: %d\n",
2481 dump_printf (MSG_NOTE
, " Vector prologue cost: %d\n", vec_prologue_cost
);
2482 dump_printf (MSG_NOTE
, " Vector epilogue cost: %d\n", vec_epilogue_cost
);
2483 dump_printf (MSG_NOTE
, " Scalar cost of basic block: %d\n", scalar_cost
);
2486 /* Vectorization is profitable if its cost is more than the cost of scalar
2487 version. Note that we err on the vector side for equal cost because
2488 the cost estimate is otherwise quite pessimistic (constant uses are
2489 free on the scalar side but cost a load on the vector side for
2491 if (vec_outside_cost
+ vec_inside_cost
> scalar_cost
)
2497 /* Check if the basic block can be vectorized. Returns a bb_vec_info
2498 if so and sets fatal to true if failure is independent of
2499 current_vector_size. */
2502 vect_slp_analyze_bb_1 (gimple_stmt_iterator region_begin
,
2503 gimple_stmt_iterator region_end
,
2504 vec
<data_reference_p
> datarefs
, int n_stmts
,
2507 bb_vec_info bb_vinfo
;
2508 slp_instance instance
;
2512 /* The first group of checks is independent of the vector size. */
2515 if (n_stmts
> PARAM_VALUE (PARAM_SLP_MAX_INSNS_IN_BB
))
2517 if (dump_enabled_p ())
2518 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2519 "not vectorized: too many instructions in "
2521 free_data_refs (datarefs
);
2525 bb_vinfo
= new_bb_vec_info (region_begin
, region_end
);
2529 BB_VINFO_DATAREFS (bb_vinfo
) = datarefs
;
2531 /* Analyze the data references. */
2533 if (!vect_analyze_data_refs (bb_vinfo
, &min_vf
))
2535 if (dump_enabled_p ())
2536 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2537 "not vectorized: unhandled data-ref in basic "
2540 destroy_bb_vec_info (bb_vinfo
);
2544 if (BB_VINFO_DATAREFS (bb_vinfo
).length () < 2)
2546 if (dump_enabled_p ())
2547 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2548 "not vectorized: not enough data-refs in "
2551 destroy_bb_vec_info (bb_vinfo
);
2555 if (!vect_analyze_data_ref_accesses (bb_vinfo
))
2557 if (dump_enabled_p ())
2558 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2559 "not vectorized: unhandled data access in "
2562 destroy_bb_vec_info (bb_vinfo
);
2566 /* If there are no grouped stores in the region there is no need
2567 to continue with pattern recog as vect_analyze_slp will fail
2569 if (bb_vinfo
->grouped_stores
.is_empty ())
2571 if (dump_enabled_p ())
2572 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2573 "not vectorized: no grouped stores in "
2576 destroy_bb_vec_info (bb_vinfo
);
2580 /* While the rest of the analysis below depends on it in some way. */
2583 vect_pattern_recog (bb_vinfo
);
2585 /* Check the SLP opportunities in the basic block, analyze and build SLP
2587 if (!vect_analyze_slp (bb_vinfo
, n_stmts
))
2589 if (dump_enabled_p ())
2591 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2592 "Failed to SLP the basic block.\n");
2593 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2594 "not vectorized: failed to find SLP opportunities "
2595 "in basic block.\n");
2598 destroy_bb_vec_info (bb_vinfo
);
2602 /* Analyze and verify the alignment of data references and the
2603 dependence in the SLP instances. */
2604 for (i
= 0; BB_VINFO_SLP_INSTANCES (bb_vinfo
).iterate (i
, &instance
); )
2606 if (! vect_slp_analyze_and_verify_instance_alignment (instance
)
2607 || ! vect_slp_analyze_instance_dependence (instance
))
2609 dump_printf_loc (MSG_NOTE
, vect_location
,
2610 "removing SLP instance operations starting from: ");
2611 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
2612 SLP_TREE_SCALAR_STMTS
2613 (SLP_INSTANCE_TREE (instance
))[0], 0);
2614 vect_free_slp_instance (instance
);
2615 BB_VINFO_SLP_INSTANCES (bb_vinfo
).ordered_remove (i
);
2619 /* Mark all the statements that we want to vectorize as pure SLP and
2621 vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance
), pure_slp
, -1);
2622 vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance
));
2626 if (! BB_VINFO_SLP_INSTANCES (bb_vinfo
).length ())
2628 destroy_bb_vec_info (bb_vinfo
);
2632 if (!vect_slp_analyze_operations (BB_VINFO_SLP_INSTANCES (bb_vinfo
),
2633 BB_VINFO_TARGET_COST_DATA (bb_vinfo
)))
2635 if (dump_enabled_p ())
2636 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2637 "not vectorized: bad operation in basic block.\n");
2639 destroy_bb_vec_info (bb_vinfo
);
2643 /* Cost model: check if the vectorization is worthwhile. */
2644 if (!unlimited_cost_model (NULL
)
2645 && !vect_bb_vectorization_profitable_p (bb_vinfo
))
2647 if (dump_enabled_p ())
2648 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2649 "not vectorized: vectorization is not "
2652 destroy_bb_vec_info (bb_vinfo
);
2656 if (dump_enabled_p ())
2657 dump_printf_loc (MSG_NOTE
, vect_location
,
2658 "Basic block will be vectorized using SLP\n");
2664 /* Main entry for the BB vectorizer. Analyze and transform BB, returns
2665 true if anything in the basic-block was vectorized. */
2668 vect_slp_bb (basic_block bb
)
2670 bb_vec_info bb_vinfo
;
2671 gimple_stmt_iterator gsi
;
2672 unsigned int vector_sizes
;
2673 bool any_vectorized
= false;
2675 if (dump_enabled_p ())
2676 dump_printf_loc (MSG_NOTE
, vect_location
, "===vect_slp_analyze_bb===\n");
2678 /* Autodetect first vector size we try. */
2679 current_vector_size
= 0;
2680 vector_sizes
= targetm
.vectorize
.autovectorize_vector_sizes ();
2682 gsi
= gsi_start_bb (bb
);
2686 if (gsi_end_p (gsi
))
2689 gimple_stmt_iterator region_begin
= gsi
;
2690 vec
<data_reference_p
> datarefs
= vNULL
;
2693 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
2695 gimple
*stmt
= gsi_stmt (gsi
);
2696 if (is_gimple_debug (stmt
))
2700 if (gimple_location (stmt
) != UNKNOWN_LOCATION
)
2701 vect_location
= gimple_location (stmt
);
2703 if (!find_data_references_in_stmt (NULL
, stmt
, &datarefs
))
2707 /* Skip leading unhandled stmts. */
2708 if (gsi_stmt (region_begin
) == gsi_stmt (gsi
))
2714 gimple_stmt_iterator region_end
= gsi
;
2716 bool vectorized
= false;
2718 bb_vinfo
= vect_slp_analyze_bb_1 (region_begin
, region_end
,
2719 datarefs
, insns
, fatal
);
2721 && dbg_cnt (vect_slp
))
2723 if (dump_enabled_p ())
2724 dump_printf_loc (MSG_NOTE
, vect_location
, "SLPing BB part\n");
2726 vect_schedule_slp (bb_vinfo
);
2728 if (dump_enabled_p ())
2729 dump_printf_loc (MSG_NOTE
, vect_location
,
2730 "basic block part vectorized\n");
2732 destroy_bb_vec_info (bb_vinfo
);
2737 destroy_bb_vec_info (bb_vinfo
);
2739 any_vectorized
|= vectorized
;
2741 vector_sizes
&= ~current_vector_size
;
2743 || vector_sizes
== 0
2744 || current_vector_size
== 0
2745 /* If vect_slp_analyze_bb_1 signaled that analysis for all
2746 vector sizes will fail do not bother iterating. */
2749 if (gsi_end_p (region_end
))
2752 /* Skip the unhandled stmt. */
2755 /* And reset vector sizes. */
2756 current_vector_size
= 0;
2757 vector_sizes
= targetm
.vectorize
.autovectorize_vector_sizes ();
2761 /* Try the next biggest vector size. */
2762 current_vector_size
= 1 << floor_log2 (vector_sizes
);
2763 if (dump_enabled_p ())
2764 dump_printf_loc (MSG_NOTE
, vect_location
,
2765 "***** Re-trying analysis with "
2766 "vector size %d\n", current_vector_size
);
2773 return any_vectorized
;
2777 /* Return 1 if vector type of boolean constant which is OPNUM
2778 operand in statement STMT is a boolean vector. */
2781 vect_mask_constant_operand_p (gimple
*stmt
, int opnum
)
2783 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
2784 enum tree_code code
= gimple_expr_code (stmt
);
2787 enum vect_def_type dt
;
2789 /* For comparison and COND_EXPR type is chosen depending
2790 on the other comparison operand. */
2791 if (TREE_CODE_CLASS (code
) == tcc_comparison
)
2794 op
= gimple_assign_rhs1 (stmt
);
2796 op
= gimple_assign_rhs2 (stmt
);
2798 if (!vect_is_simple_use (op
, stmt_vinfo
->vinfo
, &def_stmt
,
2802 return !vectype
|| VECTOR_BOOLEAN_TYPE_P (vectype
);
2805 if (code
== COND_EXPR
)
2807 tree cond
= gimple_assign_rhs1 (stmt
);
2809 if (TREE_CODE (cond
) == SSA_NAME
)
2813 op
= TREE_OPERAND (cond
, 1);
2815 op
= TREE_OPERAND (cond
, 0);
2817 if (!vect_is_simple_use (op
, stmt_vinfo
->vinfo
, &def_stmt
,
2821 return !vectype
|| VECTOR_BOOLEAN_TYPE_P (vectype
);
2824 return VECTOR_BOOLEAN_TYPE_P (STMT_VINFO_VECTYPE (stmt_vinfo
));
2828 /* For constant and loop invariant defs of SLP_NODE this function returns
2829 (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts.
2830 OP_NUM determines if we gather defs for operand 0 or operand 1 of the RHS of
2831 scalar stmts. NUMBER_OF_VECTORS is the number of vector defs to create.
2832 REDUC_INDEX is the index of the reduction operand in the statements, unless
2836 vect_get_constant_vectors (tree op
, slp_tree slp_node
,
2837 vec
<tree
> *vec_oprnds
,
2838 unsigned int op_num
, unsigned int number_of_vectors
,
2841 vec
<gimple
*> stmts
= SLP_TREE_SCALAR_STMTS (slp_node
);
2842 gimple
*stmt
= stmts
[0];
2843 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
2847 unsigned j
, number_of_places_left_in_vector
;
2850 int group_size
= stmts
.length ();
2851 unsigned int vec_num
, i
;
2852 unsigned number_of_copies
= 1;
2854 voprnds
.create (number_of_vectors
);
2855 bool constant_p
, is_store
;
2856 tree neutral_op
= NULL
;
2857 enum tree_code code
= gimple_expr_code (stmt
);
2860 gimple_seq ctor_seq
= NULL
;
2862 /* Check if vector type is a boolean vector. */
2863 if (TREE_CODE (TREE_TYPE (op
)) == BOOLEAN_TYPE
2864 && vect_mask_constant_operand_p (stmt
, op_num
))
2866 = build_same_sized_truth_vector_type (STMT_VINFO_VECTYPE (stmt_vinfo
));
2868 vector_type
= get_vectype_for_scalar_type (TREE_TYPE (op
));
2869 nunits
= TYPE_VECTOR_SUBPARTS (vector_type
);
2871 if (STMT_VINFO_DEF_TYPE (stmt_vinfo
) == vect_reduction_def
2872 && reduc_index
!= -1)
2874 op_num
= reduc_index
;
2875 op
= gimple_op (stmt
, op_num
+ 1);
2876 /* For additional copies (see the explanation of NUMBER_OF_COPIES below)
2877 we need either neutral operands or the original operands. See
2878 get_initial_def_for_reduction() for details. */
2881 case WIDEN_SUM_EXPR
:
2888 if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op
)))
2889 neutral_op
= build_real (TREE_TYPE (op
), dconst0
);
2891 neutral_op
= build_int_cst (TREE_TYPE (op
), 0);
2896 if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op
)))
2897 neutral_op
= build_real (TREE_TYPE (op
), dconst1
);
2899 neutral_op
= build_int_cst (TREE_TYPE (op
), 1);
2904 neutral_op
= build_int_cst (TREE_TYPE (op
), -1);
2907 /* For MIN/MAX we don't have an easy neutral operand but
2908 the initial values can be used fine here. Only for
2909 a reduction chain we have to force a neutral element. */
2912 if (!GROUP_FIRST_ELEMENT (stmt_vinfo
))
2916 def_stmt
= SSA_NAME_DEF_STMT (op
);
2917 loop
= (gimple_bb (stmt
))->loop_father
;
2918 neutral_op
= PHI_ARG_DEF_FROM_EDGE (def_stmt
,
2919 loop_preheader_edge (loop
));
2924 gcc_assert (!GROUP_FIRST_ELEMENT (stmt_vinfo
));
2929 if (STMT_VINFO_DATA_REF (stmt_vinfo
))
2932 op
= gimple_assign_rhs1 (stmt
);
2939 if (CONSTANT_CLASS_P (op
))
2944 /* NUMBER_OF_COPIES is the number of times we need to use the same values in
2945 created vectors. It is greater than 1 if unrolling is performed.
2947 For example, we have two scalar operands, s1 and s2 (e.g., group of
2948 strided accesses of size two), while NUNITS is four (i.e., four scalars
2949 of this type can be packed in a vector). The output vector will contain
2950 two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
2953 If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
2954 containing the operands.
2956 For example, NUNITS is four as before, and the group size is 8
2957 (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
2958 {s5, s6, s7, s8}. */
2960 number_of_copies
= nunits
* number_of_vectors
/ group_size
;
2962 number_of_places_left_in_vector
= nunits
;
2963 elts
= XALLOCAVEC (tree
, nunits
);
2964 bool place_after_defs
= false;
2965 for (j
= 0; j
< number_of_copies
; j
++)
2967 for (i
= group_size
- 1; stmts
.iterate (i
, &stmt
); i
--)
2970 op
= gimple_assign_rhs1 (stmt
);
2977 tree cond
= gimple_assign_rhs1 (stmt
);
2978 if (TREE_CODE (cond
) == SSA_NAME
)
2979 op
= gimple_op (stmt
, op_num
+ 1);
2980 else if (op_num
== 0 || op_num
== 1)
2981 op
= TREE_OPERAND (cond
, op_num
);
2985 op
= gimple_assign_rhs2 (stmt
);
2987 op
= gimple_assign_rhs3 (stmt
);
2993 op
= gimple_call_arg (stmt
, op_num
);
3000 op
= gimple_op (stmt
, op_num
+ 1);
3001 /* Unlike the other binary operators, shifts/rotates have
3002 the shift count being int, instead of the same type as
3003 the lhs, so make sure the scalar is the right type if
3004 we are dealing with vectors of
3005 long long/long/short/char. */
3006 if (op_num
== 1 && TREE_CODE (op
) == INTEGER_CST
)
3007 op
= fold_convert (TREE_TYPE (vector_type
), op
);
3011 op
= gimple_op (stmt
, op_num
+ 1);
3016 if (reduc_index
!= -1)
3018 loop
= (gimple_bb (stmt
))->loop_father
;
3019 def_stmt
= SSA_NAME_DEF_STMT (op
);
3023 /* Get the def before the loop. In reduction chain we have only
3024 one initial value. */
3025 if ((j
!= (number_of_copies
- 1)
3026 || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
))
3031 op
= PHI_ARG_DEF_FROM_EDGE (def_stmt
,
3032 loop_preheader_edge (loop
));
3035 /* Create 'vect_ = {op0,op1,...,opn}'. */
3036 number_of_places_left_in_vector
--;
3038 if (!types_compatible_p (TREE_TYPE (vector_type
), TREE_TYPE (op
)))
3040 if (CONSTANT_CLASS_P (op
))
3042 if (VECTOR_BOOLEAN_TYPE_P (vector_type
))
3044 /* Can't use VIEW_CONVERT_EXPR for booleans because
3045 of possibly different sizes of scalar value and
3047 if (integer_zerop (op
))
3048 op
= build_int_cst (TREE_TYPE (vector_type
), 0);
3049 else if (integer_onep (op
))
3050 op
= build_int_cst (TREE_TYPE (vector_type
), 1);
3055 op
= fold_unary (VIEW_CONVERT_EXPR
,
3056 TREE_TYPE (vector_type
), op
);
3057 gcc_assert (op
&& CONSTANT_CLASS_P (op
));
3061 tree new_temp
= make_ssa_name (TREE_TYPE (vector_type
));
3063 if (VECTOR_BOOLEAN_TYPE_P (vector_type
))
3065 gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (op
)));
3066 init_stmt
= gimple_build_assign (new_temp
, NOP_EXPR
, op
);
3070 op
= build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (vector_type
),
3073 = gimple_build_assign (new_temp
, VIEW_CONVERT_EXPR
,
3076 gimple_seq_add_stmt (&ctor_seq
, init_stmt
);
3080 elts
[number_of_places_left_in_vector
] = op
;
3081 if (!CONSTANT_CLASS_P (op
))
3083 if (TREE_CODE (orig_op
) == SSA_NAME
3084 && !SSA_NAME_IS_DEFAULT_DEF (orig_op
)
3085 && STMT_VINFO_BB_VINFO (stmt_vinfo
)
3086 && (STMT_VINFO_BB_VINFO (stmt_vinfo
)->bb
3087 == gimple_bb (SSA_NAME_DEF_STMT (orig_op
))))
3088 place_after_defs
= true;
3090 if (number_of_places_left_in_vector
== 0)
3092 number_of_places_left_in_vector
= nunits
;
3095 vec_cst
= build_vector (vector_type
, elts
);
3098 vec
<constructor_elt
, va_gc
> *v
;
3100 vec_alloc (v
, nunits
);
3101 for (k
= 0; k
< nunits
; ++k
)
3102 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, elts
[k
]);
3103 vec_cst
= build_constructor (vector_type
, v
);
3106 gimple_stmt_iterator gsi
;
3107 if (place_after_defs
)
3110 (vect_find_last_scalar_stmt_in_slp (slp_node
));
3111 init
= vect_init_vector (stmt
, vec_cst
, vector_type
, &gsi
);
3114 init
= vect_init_vector (stmt
, vec_cst
, vector_type
, NULL
);
3115 if (ctor_seq
!= NULL
)
3117 gsi
= gsi_for_stmt (SSA_NAME_DEF_STMT (init
));
3118 gsi_insert_seq_before_without_update (&gsi
, ctor_seq
,
3122 voprnds
.quick_push (init
);
3123 place_after_defs
= false;
3128 /* Since the vectors are created in the reverse order, we should invert
3130 vec_num
= voprnds
.length ();
3131 for (j
= vec_num
; j
!= 0; j
--)
3133 vop
= voprnds
[j
- 1];
3134 vec_oprnds
->quick_push (vop
);
3139 /* In case that VF is greater than the unrolling factor needed for the SLP
3140 group of stmts, NUMBER_OF_VECTORS to be created is greater than
3141 NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
3142 to replicate the vectors. */
3143 while (number_of_vectors
> vec_oprnds
->length ())
3145 tree neutral_vec
= NULL
;
3150 neutral_vec
= build_vector_from_val (vector_type
, neutral_op
);
3152 vec_oprnds
->quick_push (neutral_vec
);
3156 for (i
= 0; vec_oprnds
->iterate (i
, &vop
) && i
< vec_num
; i
++)
3157 vec_oprnds
->quick_push (vop
);
3163 /* Get vectorized definitions from SLP_NODE that contains corresponding
3164 vectorized def-stmts. */
3167 vect_get_slp_vect_defs (slp_tree slp_node
, vec
<tree
> *vec_oprnds
)
3170 gimple
*vec_def_stmt
;
3173 gcc_assert (SLP_TREE_VEC_STMTS (slp_node
).exists ());
3175 FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (slp_node
), i
, vec_def_stmt
)
3177 gcc_assert (vec_def_stmt
);
3178 vec_oprnd
= gimple_get_lhs (vec_def_stmt
);
3179 vec_oprnds
->quick_push (vec_oprnd
);
3184 /* Get vectorized definitions for SLP_NODE.
3185 If the scalar definitions are loop invariants or constants, collect them and
3186 call vect_get_constant_vectors() to create vector stmts.
3187 Otherwise, the def-stmts must be already vectorized and the vectorized stmts
3188 must be stored in the corresponding child of SLP_NODE, and we call
3189 vect_get_slp_vect_defs () to retrieve them. */
3192 vect_get_slp_defs (vec
<tree
> ops
, slp_tree slp_node
,
3193 vec
<vec
<tree
> > *vec_oprnds
, int reduc_index
)
3196 int number_of_vects
= 0, i
;
3197 unsigned int child_index
= 0;
3198 HOST_WIDE_INT lhs_size_unit
, rhs_size_unit
;
3199 slp_tree child
= NULL
;
3202 bool vectorized_defs
;
3204 first_stmt
= SLP_TREE_SCALAR_STMTS (slp_node
)[0];
3205 FOR_EACH_VEC_ELT (ops
, i
, oprnd
)
3207 /* For each operand we check if it has vectorized definitions in a child
3208 node or we need to create them (for invariants and constants). We
3209 check if the LHS of the first stmt of the next child matches OPRND.
3210 If it does, we found the correct child. Otherwise, we call
3211 vect_get_constant_vectors (), and not advance CHILD_INDEX in order
3212 to check this child node for the next operand. */
3213 vectorized_defs
= false;
3214 if (SLP_TREE_CHILDREN (slp_node
).length () > child_index
)
3216 child
= SLP_TREE_CHILDREN (slp_node
)[child_index
];
3218 /* We have to check both pattern and original def, if available. */
3219 if (SLP_TREE_DEF_TYPE (child
) == vect_internal_def
)
3221 gimple
*first_def
= SLP_TREE_SCALAR_STMTS (child
)[0];
3223 = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (first_def
));
3225 if (operand_equal_p (oprnd
, gimple_get_lhs (first_def
), 0)
3227 && operand_equal_p (oprnd
, gimple_get_lhs (related
), 0)))
3229 /* The number of vector defs is determined by the number of
3230 vector statements in the node from which we get those
3232 number_of_vects
= SLP_TREE_NUMBER_OF_VEC_STMTS (child
);
3233 vectorized_defs
= true;
3241 if (!vectorized_defs
)
3245 number_of_vects
= SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node
);
3246 /* Number of vector stmts was calculated according to LHS in
3247 vect_schedule_slp_instance (), fix it by replacing LHS with
3248 RHS, if necessary. See vect_get_smallest_scalar_type () for
3250 vect_get_smallest_scalar_type (first_stmt
, &lhs_size_unit
,
3252 if (rhs_size_unit
!= lhs_size_unit
)
3254 number_of_vects
*= rhs_size_unit
;
3255 number_of_vects
/= lhs_size_unit
;
3260 /* Allocate memory for vectorized defs. */
3262 vec_defs
.create (number_of_vects
);
3264 /* For reduction defs we call vect_get_constant_vectors (), since we are
3265 looking for initial loop invariant values. */
3266 if (vectorized_defs
&& reduc_index
== -1)
3267 /* The defs are already vectorized. */
3268 vect_get_slp_vect_defs (child
, &vec_defs
);
3270 /* Build vectors from scalar defs. */
3271 vect_get_constant_vectors (oprnd
, slp_node
, &vec_defs
, i
,
3272 number_of_vects
, reduc_index
);
3274 vec_oprnds
->quick_push (vec_defs
);
3276 /* For reductions, we only need initial values. */
3277 if (reduc_index
!= -1)
3283 /* Create NCOPIES permutation statements using the mask MASK_BYTES (by
3284 building a vector of type MASK_TYPE from it) and two input vectors placed in
3285 DR_CHAIN at FIRST_VEC_INDX and SECOND_VEC_INDX for the first copy and
3286 shifting by STRIDE elements of DR_CHAIN for every copy.
3287 (STRIDE is the number of vectorized stmts for NODE divided by the number of
3289 VECT_STMTS_COUNTER specifies the index in the vectorized stmts of NODE, where
3290 the created stmts must be inserted. */
3293 vect_create_mask_and_perm (gimple
*stmt
,
3294 tree mask
, int first_vec_indx
, int second_vec_indx
,
3295 gimple_stmt_iterator
*gsi
, slp_tree node
,
3296 tree vectype
, vec
<tree
> dr_chain
,
3297 int ncopies
, int vect_stmts_counter
)
3300 gimple
*perm_stmt
= NULL
;
3301 int i
, stride_in
, stride_out
;
3302 tree first_vec
, second_vec
, data_ref
;
3304 stride_out
= SLP_TREE_NUMBER_OF_VEC_STMTS (node
) / ncopies
;
3305 stride_in
= dr_chain
.length () / ncopies
;
3307 /* Initialize the vect stmts of NODE to properly insert the generated
3309 for (i
= SLP_TREE_VEC_STMTS (node
).length ();
3310 i
< (int) SLP_TREE_NUMBER_OF_VEC_STMTS (node
); i
++)
3311 SLP_TREE_VEC_STMTS (node
).quick_push (NULL
);
3313 perm_dest
= vect_create_destination_var (gimple_assign_lhs (stmt
), vectype
);
3314 for (i
= 0; i
< ncopies
; i
++)
3316 first_vec
= dr_chain
[first_vec_indx
];
3317 second_vec
= dr_chain
[second_vec_indx
];
3319 /* Generate the permute statement if necessary. */
3322 perm_stmt
= gimple_build_assign (perm_dest
, VEC_PERM_EXPR
,
3323 first_vec
, second_vec
, mask
);
3324 data_ref
= make_ssa_name (perm_dest
, perm_stmt
);
3325 gimple_set_lhs (perm_stmt
, data_ref
);
3326 vect_finish_stmt_generation (stmt
, perm_stmt
, gsi
);
3329 /* If mask was NULL_TREE generate the requested identity transform. */
3330 perm_stmt
= SSA_NAME_DEF_STMT (first_vec
);
3332 /* Store the vector statement in NODE. */
3333 SLP_TREE_VEC_STMTS (node
)[stride_out
* i
+ vect_stmts_counter
]
3336 first_vec_indx
+= stride_in
;
3337 second_vec_indx
+= stride_in
;
3342 /* Generate vector permute statements from a list of loads in DR_CHAIN.
3343 If ANALYZE_ONLY is TRUE, only check that it is possible to create valid
3344 permute statements for the SLP node NODE of the SLP instance
3345 SLP_NODE_INSTANCE. */
3348 vect_transform_slp_perm_load (slp_tree node
, vec
<tree
> dr_chain
,
3349 gimple_stmt_iterator
*gsi
, int vf
,
3350 slp_instance slp_node_instance
, bool analyze_only
)
3352 gimple
*stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
3353 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3354 tree mask_element_type
= NULL_TREE
, mask_type
;
3355 int nunits
, vec_index
= 0;
3356 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
3357 int group_size
= SLP_INSTANCE_GROUP_SIZE (slp_node_instance
);
3358 int unroll_factor
, mask_element
, ncopies
;
3359 unsigned char *mask
;
3362 if (!STMT_VINFO_GROUPED_ACCESS (stmt_info
))
3365 stmt_info
= vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info
));
3367 mode
= TYPE_MODE (vectype
);
3369 /* The generic VEC_PERM_EXPR code always uses an integral type of the
3370 same size as the vector element being permuted. */
3371 mask_element_type
= lang_hooks
.types
.type_for_mode
3372 (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype
))), 1);
3373 mask_type
= get_vectype_for_scalar_type (mask_element_type
);
3374 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3375 mask
= XALLOCAVEC (unsigned char, nunits
);
3376 unroll_factor
= SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance
);
3378 /* Number of copies is determined by the final vectorization factor
3379 relatively to SLP_NODE_INSTANCE unrolling factor. */
3380 ncopies
= vf
/ SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance
);
3382 /* Generate permutation masks for every NODE. Number of masks for each NODE
3383 is equal to GROUP_SIZE.
3384 E.g., we have a group of three nodes with three loads from the same
3385 location in each node, and the vector size is 4. I.e., we have a
3386 a0b0c0a1b1c1... sequence and we need to create the following vectors:
3387 for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3
3388 for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3
3391 The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}.
3392 The last mask is illegal since we assume two operands for permute
3393 operation, and the mask element values can't be outside that range.
3394 Hence, the last mask must be converted into {2,5,5,5}.
3395 For the first two permutations we need the first and the second input
3396 vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation
3397 we need the second and the third vectors: {b1,c1,a2,b2} and
3400 int vect_stmts_counter
= 0;
3402 int first_vec_index
= -1;
3403 int second_vec_index
= -1;
3406 for (int j
= 0; j
< unroll_factor
; j
++)
3408 for (int k
= 0; k
< group_size
; k
++)
3410 int i
= (SLP_TREE_LOAD_PERMUTATION (node
)[k
]
3411 + j
* STMT_VINFO_GROUP_SIZE (stmt_info
));
3412 vec_index
= i
/ nunits
;
3413 mask_element
= i
% nunits
;
3414 if (vec_index
== first_vec_index
3415 || first_vec_index
== -1)
3417 first_vec_index
= vec_index
;
3419 else if (vec_index
== second_vec_index
3420 || second_vec_index
== -1)
3422 second_vec_index
= vec_index
;
3423 mask_element
+= nunits
;
3427 if (dump_enabled_p ())
3429 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
3430 "permutation requires at "
3431 "least three vectors ");
3432 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
3434 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
3439 gcc_assert (mask_element
>= 0
3440 && mask_element
< 2 * nunits
);
3441 if (mask_element
!= index
)
3443 mask
[index
++] = mask_element
;
3445 if (index
== nunits
)
3448 && ! can_vec_perm_p (mode
, false, mask
))
3450 if (dump_enabled_p ())
3452 dump_printf_loc (MSG_MISSED_OPTIMIZATION
,
3454 "unsupported vect permute { ");
3455 for (i
= 0; i
< nunits
; ++i
)
3456 dump_printf (MSG_MISSED_OPTIMIZATION
, "%d ", mask
[i
]);
3457 dump_printf (MSG_MISSED_OPTIMIZATION
, "}\n");
3464 tree mask_vec
= NULL_TREE
;
3468 tree
*mask_elts
= XALLOCAVEC (tree
, nunits
);
3469 for (int l
= 0; l
< nunits
; ++l
)
3470 mask_elts
[l
] = build_int_cst (mask_element_type
,
3472 mask_vec
= build_vector (mask_type
, mask_elts
);
3475 if (second_vec_index
== -1)
3476 second_vec_index
= first_vec_index
;
3477 vect_create_mask_and_perm (stmt
, mask_vec
, first_vec_index
,
3479 gsi
, node
, vectype
, dr_chain
,
3480 ncopies
, vect_stmts_counter
++);
3484 first_vec_index
= -1;
3485 second_vec_index
= -1;
3496 /* Vectorize SLP instance tree in postorder. */
3499 vect_schedule_slp_instance (slp_tree node
, slp_instance instance
,
3500 unsigned int vectorization_factor
)
3503 bool grouped_store
, is_store
;
3504 gimple_stmt_iterator si
;
3505 stmt_vec_info stmt_info
;
3506 unsigned int vec_stmts_size
, nunits
, group_size
;
3511 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
3514 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3515 vect_schedule_slp_instance (child
, instance
, vectorization_factor
);
3517 /* Push SLP node def-type to stmts. */
3518 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3519 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
3520 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
3521 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = SLP_TREE_DEF_TYPE (child
);
3523 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
3524 stmt_info
= vinfo_for_stmt (stmt
);
3526 /* VECTYPE is the type of the destination. */
3527 vectype
= STMT_VINFO_VECTYPE (stmt_info
);
3528 nunits
= (unsigned int) TYPE_VECTOR_SUBPARTS (vectype
);
3529 group_size
= SLP_INSTANCE_GROUP_SIZE (instance
);
3531 /* For each SLP instance calculate number of vector stmts to be created
3532 for the scalar stmts in each node of the SLP tree. Number of vector
3533 elements in one vector iteration is the number of scalar elements in
3534 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
3536 Unless this is a SLP reduction in which case the number of vector
3537 stmts is equal to the number of vector stmts of the children. */
3538 if (GROUP_FIRST_ELEMENT (stmt_info
)
3539 && !STMT_VINFO_GROUPED_ACCESS (stmt_info
))
3540 vec_stmts_size
= SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_CHILDREN (node
)[0]);
3542 vec_stmts_size
= (vectorization_factor
* group_size
) / nunits
;
3544 if (!SLP_TREE_VEC_STMTS (node
).exists ())
3546 SLP_TREE_VEC_STMTS (node
).create (vec_stmts_size
);
3547 SLP_TREE_NUMBER_OF_VEC_STMTS (node
) = vec_stmts_size
;
3550 if (dump_enabled_p ())
3552 dump_printf_loc (MSG_NOTE
,vect_location
,
3553 "------>vectorizing SLP node starting from: ");
3554 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
3555 dump_printf (MSG_NOTE
, "\n");
3558 /* Vectorized stmts go before the last scalar stmt which is where
3559 all uses are ready. */
3560 si
= gsi_for_stmt (vect_find_last_scalar_stmt_in_slp (node
));
3562 /* Mark the first element of the reduction chain as reduction to properly
3563 transform the node. In the analysis phase only the last element of the
3564 chain is marked as reduction. */
3565 if (GROUP_FIRST_ELEMENT (stmt_info
) && !STMT_VINFO_GROUPED_ACCESS (stmt_info
)
3566 && GROUP_FIRST_ELEMENT (stmt_info
) == stmt
)
3568 STMT_VINFO_DEF_TYPE (stmt_info
) = vect_reduction_def
;
3569 STMT_VINFO_TYPE (stmt_info
) = reduc_vec_info_type
;
3572 /* Handle two-operation SLP nodes by vectorizing the group with
3573 both operations and then performing a merge. */
3574 if (SLP_TREE_TWO_OPERATORS (node
))
3576 enum tree_code code0
= gimple_assign_rhs_code (stmt
);
3577 enum tree_code ocode
= ERROR_MARK
;
3579 unsigned char *mask
= XALLOCAVEC (unsigned char, group_size
);
3580 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, ostmt
)
3581 if (gimple_assign_rhs_code (ostmt
) != code0
)
3584 ocode
= gimple_assign_rhs_code (ostmt
);
3588 if (ocode
!= ERROR_MARK
)
3593 tree tmask
= NULL_TREE
;
3594 vect_transform_stmt (stmt
, &si
, &grouped_store
, node
, instance
);
3595 v0
= SLP_TREE_VEC_STMTS (node
).copy ();
3596 SLP_TREE_VEC_STMTS (node
).truncate (0);
3597 gimple_assign_set_rhs_code (stmt
, ocode
);
3598 vect_transform_stmt (stmt
, &si
, &grouped_store
, node
, instance
);
3599 gimple_assign_set_rhs_code (stmt
, code0
);
3600 v1
= SLP_TREE_VEC_STMTS (node
).copy ();
3601 SLP_TREE_VEC_STMTS (node
).truncate (0);
3602 tree meltype
= build_nonstandard_integer_type
3603 (GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype
))), 1);
3604 tree mvectype
= get_same_sized_vectype (meltype
, vectype
);
3606 for (j
= 0; j
< v0
.length (); ++j
)
3608 tree
*melts
= XALLOCAVEC (tree
, TYPE_VECTOR_SUBPARTS (vectype
));
3609 for (l
= 0; l
< TYPE_VECTOR_SUBPARTS (vectype
); ++l
)
3611 if (k
>= group_size
)
3613 melts
[l
] = build_int_cst
3614 (meltype
, mask
[k
++] * TYPE_VECTOR_SUBPARTS (vectype
) + l
);
3616 tmask
= build_vector (mvectype
, melts
);
3618 /* ??? Not all targets support a VEC_PERM_EXPR with a
3619 constant mask that would translate to a vec_merge RTX
3620 (with their vec_perm_const_ok). We can either not
3621 vectorize in that case or let veclower do its job.
3622 Unfortunately that isn't too great and at least for
3623 plus/minus we'd eventually like to match targets
3624 vector addsub instructions. */
3626 vstmt
= gimple_build_assign (make_ssa_name (vectype
),
3628 gimple_assign_lhs (v0
[j
]),
3629 gimple_assign_lhs (v1
[j
]), tmask
);
3630 vect_finish_stmt_generation (stmt
, vstmt
, &si
);
3631 SLP_TREE_VEC_STMTS (node
).quick_push (vstmt
);
3638 is_store
= vect_transform_stmt (stmt
, &si
, &grouped_store
, node
, instance
);
3640 /* Restore stmt def-types. */
3641 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3642 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
3643 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
3644 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = vect_internal_def
;
3649 /* Replace scalar calls from SLP node NODE with setting of their lhs to zero.
3650 For loop vectorization this is done in vectorizable_call, but for SLP
3651 it needs to be deferred until end of vect_schedule_slp, because multiple
3652 SLP instances may refer to the same scalar stmt. */
3655 vect_remove_slp_scalar_calls (slp_tree node
)
3657 gimple
*stmt
, *new_stmt
;
3658 gimple_stmt_iterator gsi
;
3662 stmt_vec_info stmt_info
;
3664 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
3667 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3668 vect_remove_slp_scalar_calls (child
);
3670 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
3672 if (!is_gimple_call (stmt
) || gimple_bb (stmt
) == NULL
)
3674 stmt_info
= vinfo_for_stmt (stmt
);
3675 if (stmt_info
== NULL
3676 || is_pattern_stmt_p (stmt_info
)
3677 || !PURE_SLP_STMT (stmt_info
))
3679 lhs
= gimple_call_lhs (stmt
);
3680 new_stmt
= gimple_build_assign (lhs
, build_zero_cst (TREE_TYPE (lhs
)));
3681 set_vinfo_for_stmt (new_stmt
, stmt_info
);
3682 set_vinfo_for_stmt (stmt
, NULL
);
3683 STMT_VINFO_STMT (stmt_info
) = new_stmt
;
3684 gsi
= gsi_for_stmt (stmt
);
3685 gsi_replace (&gsi
, new_stmt
, false);
3686 SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt
)) = new_stmt
;
3690 /* Generate vector code for all SLP instances in the loop/basic block. */
3693 vect_schedule_slp (vec_info
*vinfo
)
3695 vec
<slp_instance
> slp_instances
;
3696 slp_instance instance
;
3698 bool is_store
= false;
3700 slp_instances
= vinfo
->slp_instances
;
3701 if (is_a
<loop_vec_info
> (vinfo
))
3702 vf
= as_a
<loop_vec_info
> (vinfo
)->vectorization_factor
;
3706 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
3708 /* Schedule the tree of INSTANCE. */
3709 is_store
= vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance
),
3711 if (dump_enabled_p ())
3712 dump_printf_loc (MSG_NOTE
, vect_location
,
3713 "vectorizing stmts using SLP.\n");
3716 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
3718 slp_tree root
= SLP_INSTANCE_TREE (instance
);
3721 gimple_stmt_iterator gsi
;
3723 /* Remove scalar call stmts. Do not do this for basic-block
3724 vectorization as not all uses may be vectorized.
3725 ??? Why should this be necessary? DCE should be able to
3726 remove the stmts itself.
3727 ??? For BB vectorization we can as well remove scalar
3728 stmts starting from the SLP tree root if they have no
3730 if (is_a
<loop_vec_info
> (vinfo
))
3731 vect_remove_slp_scalar_calls (root
);
3733 for (j
= 0; SLP_TREE_SCALAR_STMTS (root
).iterate (j
, &store
)
3734 && j
< SLP_INSTANCE_GROUP_SIZE (instance
); j
++)
3736 if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (store
)))
3739 if (is_pattern_stmt_p (vinfo_for_stmt (store
)))
3740 store
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (store
));
3741 /* Free the attached stmt_vec_info and remove the stmt. */
3742 gsi
= gsi_for_stmt (store
);
3743 unlink_stmt_vdef (store
);
3744 gsi_remove (&gsi
, true);
3745 release_defs (store
);
3746 free_stmt_vec_info (store
);