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 sbitmap_free (load_index
);
1315 if (bitmap_bit_p (load_index
, lidx
))
1317 sbitmap_free (load_index
);
1320 bitmap_set_bit (load_index
, lidx
);
1322 for (i
= 0; i
< group_size
; i
++)
1323 if (!bitmap_bit_p (load_index
, i
))
1325 sbitmap_free (load_index
);
1328 sbitmap_free (load_index
);
1330 /* This permutation is valid for reduction. Since the order of the
1331 statements in the nodes is not important unless they are memory
1332 accesses, we can rearrange the statements in all the nodes
1333 according to the order of the loads. */
1334 vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn
), group_size
,
1335 node
->load_permutation
);
1337 /* We are done, no actual permutations need to be generated. */
1338 unsigned int unrolling_factor
= SLP_INSTANCE_UNROLLING_FACTOR (slp_instn
);
1339 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1341 gimple
*first_stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1342 first_stmt
= GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt
));
1343 /* But we have to keep those permutations that are required because
1344 of handling of gaps. */
1345 if (unrolling_factor
== 1
1346 || (group_size
== GROUP_SIZE (vinfo_for_stmt (first_stmt
))
1347 && GROUP_GAP (vinfo_for_stmt (first_stmt
)) == 0))
1348 SLP_TREE_LOAD_PERMUTATION (node
).release ();
1350 for (j
= 0; j
< SLP_TREE_LOAD_PERMUTATION (node
).length (); ++j
)
1351 SLP_TREE_LOAD_PERMUTATION (node
)[j
] = j
;
1357 /* Check if the required load permutations in the SLP instance
1358 SLP_INSTN are supported. */
1361 vect_supported_load_permutation_p (slp_instance slp_instn
)
1363 unsigned int group_size
= SLP_INSTANCE_GROUP_SIZE (slp_instn
);
1364 unsigned int i
, j
, k
, next
;
1366 gimple
*stmt
, *load
, *next_load
;
1368 if (dump_enabled_p ())
1370 dump_printf_loc (MSG_NOTE
, vect_location
, "Load permutation ");
1371 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1372 if (node
->load_permutation
.exists ())
1373 FOR_EACH_VEC_ELT (node
->load_permutation
, j
, next
)
1374 dump_printf (MSG_NOTE
, "%d ", next
);
1376 for (k
= 0; k
< group_size
; ++k
)
1377 dump_printf (MSG_NOTE
, "%d ", k
);
1378 dump_printf (MSG_NOTE
, "\n");
1381 /* In case of reduction every load permutation is allowed, since the order
1382 of the reduction statements is not important (as opposed to the case of
1383 grouped stores). The only condition we need to check is that all the
1384 load nodes are of the same size and have the same permutation (and then
1385 rearrange all the nodes of the SLP instance according to this
1388 /* Check that all the load nodes are of the same size. */
1389 /* ??? Can't we assert this? */
1390 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1391 if (SLP_TREE_SCALAR_STMTS (node
).length () != (unsigned) group_size
)
1394 node
= SLP_INSTANCE_TREE (slp_instn
);
1395 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1397 /* Reduction (there are no data-refs in the root).
1398 In reduction chain the order of the loads is not important. */
1399 if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
))
1400 && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
1401 vect_attempt_slp_rearrange_stmts (slp_instn
);
1403 /* In basic block vectorization we allow any subchain of an interleaving
1405 FORNOW: not supported in loop SLP because of realignment compications. */
1406 if (STMT_VINFO_BB_VINFO (vinfo_for_stmt (stmt
)))
1408 /* Check whether the loads in an instance form a subchain and thus
1409 no permutation is necessary. */
1410 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1412 if (!SLP_TREE_LOAD_PERMUTATION (node
).exists ())
1414 bool subchain_p
= true;
1416 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), j
, load
)
1419 && (next_load
!= load
1420 || GROUP_GAP (vinfo_for_stmt (load
)) != 1))
1425 next_load
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (load
));
1428 SLP_TREE_LOAD_PERMUTATION (node
).release ();
1431 /* Verify the permutation can be generated. */
1433 if (!vect_transform_slp_perm_load (node
, tem
, NULL
,
1434 1, slp_instn
, true))
1436 dump_printf_loc (MSG_MISSED_OPTIMIZATION
,
1438 "unsupported load permutation\n");
1446 /* For loop vectorization verify we can generate the permutation. */
1447 FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (slp_instn
), i
, node
)
1448 if (node
->load_permutation
.exists ()
1449 && !vect_transform_slp_perm_load
1451 SLP_INSTANCE_UNROLLING_FACTOR (slp_instn
), slp_instn
, true))
1458 /* Find the last store in SLP INSTANCE. */
1461 vect_find_last_scalar_stmt_in_slp (slp_tree node
)
1463 gimple
*last
= NULL
, *stmt
;
1465 for (int i
= 0; SLP_TREE_SCALAR_STMTS (node
).iterate (i
, &stmt
); i
++)
1467 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
1468 if (is_pattern_stmt_p (stmt_vinfo
))
1469 last
= get_later_stmt (STMT_VINFO_RELATED_STMT (stmt_vinfo
), last
);
1471 last
= get_later_stmt (stmt
, last
);
1477 /* Compute the cost for the SLP node NODE in the SLP instance INSTANCE. */
1480 vect_analyze_slp_cost_1 (slp_instance instance
, slp_tree node
,
1481 stmt_vector_for_cost
*prologue_cost_vec
,
1482 stmt_vector_for_cost
*body_cost_vec
,
1483 unsigned ncopies_for_cost
)
1488 stmt_vec_info stmt_info
;
1491 /* Recurse down the SLP tree. */
1492 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1493 if (SLP_TREE_DEF_TYPE (child
) == vect_internal_def
)
1494 vect_analyze_slp_cost_1 (instance
, child
, prologue_cost_vec
,
1495 body_cost_vec
, ncopies_for_cost
);
1497 /* Look at the first scalar stmt to determine the cost. */
1498 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1499 stmt_info
= vinfo_for_stmt (stmt
);
1500 if (STMT_VINFO_GROUPED_ACCESS (stmt_info
))
1502 if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info
)))
1503 vect_model_store_cost (stmt_info
, ncopies_for_cost
, false,
1504 vect_uninitialized_def
,
1505 node
, prologue_cost_vec
, body_cost_vec
);
1508 gcc_checking_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info
)));
1509 if (SLP_TREE_LOAD_PERMUTATION (node
).exists ())
1511 /* If the load is permuted then the alignment is determined by
1512 the first group element not by the first scalar stmt DR. */
1513 stmt
= GROUP_FIRST_ELEMENT (stmt_info
);
1514 stmt_info
= vinfo_for_stmt (stmt
);
1515 /* Record the cost for the permutation. */
1516 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vec_perm
,
1517 stmt_info
, 0, vect_body
);
1518 /* And adjust the number of loads performed. */
1520 = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
));
1522 = (GROUP_SIZE (stmt_info
) - GROUP_GAP (stmt_info
)
1523 + nunits
- 1) / nunits
;
1524 ncopies_for_cost
*= SLP_INSTANCE_UNROLLING_FACTOR (instance
);
1526 /* Record the cost for the vector loads. */
1527 vect_model_load_cost (stmt_info
, ncopies_for_cost
, false,
1528 node
, prologue_cost_vec
, body_cost_vec
);
1534 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vector_stmt
,
1535 stmt_info
, 0, vect_body
);
1536 if (SLP_TREE_TWO_OPERATORS (node
))
1538 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vector_stmt
,
1539 stmt_info
, 0, vect_body
);
1540 record_stmt_cost (body_cost_vec
, ncopies_for_cost
, vec_perm
,
1541 stmt_info
, 0, vect_body
);
1545 /* Push SLP node def-type to stmts. */
1546 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1547 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
1548 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
1549 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = SLP_TREE_DEF_TYPE (child
);
1551 /* Scan operands and account for prologue cost of constants/externals.
1552 ??? This over-estimates cost for multiple uses and should be
1554 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1555 lhs
= gimple_get_lhs (stmt
);
1556 for (i
= 0; i
< gimple_num_ops (stmt
); ++i
)
1558 tree op
= gimple_op (stmt
, i
);
1560 enum vect_def_type dt
;
1561 if (!op
|| op
== lhs
)
1563 if (vect_is_simple_use (op
, stmt_info
->vinfo
, &def_stmt
, &dt
))
1565 /* Without looking at the actual initializer a vector of
1566 constants can be implemented as load from the constant pool.
1567 ??? We need to pass down stmt_info for a vector type
1568 even if it points to the wrong stmt. */
1569 if (dt
== vect_constant_def
)
1570 record_stmt_cost (prologue_cost_vec
, 1, vector_load
,
1571 stmt_info
, 0, vect_prologue
);
1572 else if (dt
== vect_external_def
)
1573 record_stmt_cost (prologue_cost_vec
, 1, vec_construct
,
1574 stmt_info
, 0, vect_prologue
);
1578 /* Restore stmt def-types. */
1579 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
1580 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
1581 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
1582 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = vect_internal_def
;
1585 /* Compute the cost for the SLP instance INSTANCE. */
1588 vect_analyze_slp_cost (slp_instance instance
, void *data
)
1590 stmt_vector_for_cost body_cost_vec
, prologue_cost_vec
;
1591 unsigned ncopies_for_cost
;
1592 stmt_info_for_cost
*si
;
1595 if (dump_enabled_p ())
1596 dump_printf_loc (MSG_NOTE
, vect_location
,
1597 "=== vect_analyze_slp_cost ===\n");
1599 /* Calculate the number of vector stmts to create based on the unrolling
1600 factor (number of vectors is 1 if NUNITS >= GROUP_SIZE, and is
1601 GROUP_SIZE / NUNITS otherwise. */
1602 unsigned group_size
= SLP_INSTANCE_GROUP_SIZE (instance
);
1603 slp_tree node
= SLP_INSTANCE_TREE (instance
);
1604 stmt_vec_info stmt_info
= vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (node
)[0]);
1605 /* Adjust the group_size by the vectorization factor which is always one
1606 for basic-block vectorization. */
1607 if (STMT_VINFO_LOOP_VINFO (stmt_info
))
1608 group_size
*= LOOP_VINFO_VECT_FACTOR (STMT_VINFO_LOOP_VINFO (stmt_info
));
1609 unsigned nunits
= TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
));
1610 /* For reductions look at a reduction operand in case the reduction
1611 operation is widening like DOT_PROD or SAD. */
1612 if (!STMT_VINFO_GROUPED_ACCESS (stmt_info
))
1614 gimple
*stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
1615 switch (gimple_assign_rhs_code (stmt
))
1619 nunits
= TYPE_VECTOR_SUBPARTS (get_vectype_for_scalar_type
1620 (TREE_TYPE (gimple_assign_rhs1 (stmt
))));
1625 ncopies_for_cost
= least_common_multiple (nunits
, group_size
) / nunits
;
1627 prologue_cost_vec
.create (10);
1628 body_cost_vec
.create (10);
1629 vect_analyze_slp_cost_1 (instance
, SLP_INSTANCE_TREE (instance
),
1630 &prologue_cost_vec
, &body_cost_vec
,
1633 /* Record the prologue costs, which were delayed until we were
1634 sure that SLP was successful. */
1635 FOR_EACH_VEC_ELT (prologue_cost_vec
, i
, si
)
1637 struct _stmt_vec_info
*stmt_info
1638 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
1639 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
1640 si
->misalign
, vect_prologue
);
1643 /* Record the instance's instructions in the target cost model. */
1644 FOR_EACH_VEC_ELT (body_cost_vec
, i
, si
)
1646 struct _stmt_vec_info
*stmt_info
1647 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
1648 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
1649 si
->misalign
, vect_body
);
1652 prologue_cost_vec
.release ();
1653 body_cost_vec
.release ();
1656 /* Splits a group of stores, currently beginning at FIRST_STMT, into two groups:
1657 one (still beginning at FIRST_STMT) of size GROUP1_SIZE (also containing
1658 the first GROUP1_SIZE stmts, since stores are consecutive), the second
1659 containing the remainder.
1660 Return the first stmt in the second group. */
1663 vect_split_slp_store_group (gimple
*first_stmt
, unsigned group1_size
)
1665 stmt_vec_info first_vinfo
= vinfo_for_stmt (first_stmt
);
1666 gcc_assert (GROUP_FIRST_ELEMENT (first_vinfo
) == first_stmt
);
1667 gcc_assert (group1_size
> 0);
1668 int group2_size
= GROUP_SIZE (first_vinfo
) - group1_size
;
1669 gcc_assert (group2_size
> 0);
1670 GROUP_SIZE (first_vinfo
) = group1_size
;
1672 gimple
*stmt
= first_stmt
;
1673 for (unsigned i
= group1_size
; i
> 1; i
--)
1675 stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
));
1676 gcc_assert (GROUP_GAP (vinfo_for_stmt (stmt
)) == 1);
1678 /* STMT is now the last element of the first group. */
1679 gimple
*group2
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
));
1680 GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
)) = 0;
1682 GROUP_SIZE (vinfo_for_stmt (group2
)) = group2_size
;
1683 for (stmt
= group2
; stmt
; stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt
)))
1685 GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)) = group2
;
1686 gcc_assert (GROUP_GAP (vinfo_for_stmt (stmt
)) == 1);
1689 /* For the second group, the GROUP_GAP is that before the original group,
1690 plus skipping over the first vector. */
1691 GROUP_GAP (vinfo_for_stmt (group2
)) =
1692 GROUP_GAP (first_vinfo
) + group1_size
;
1694 /* GROUP_GAP of the first group now has to skip over the second group too. */
1695 GROUP_GAP (first_vinfo
) += group2_size
;
1697 if (dump_enabled_p ())
1698 dump_printf_loc (MSG_NOTE
, vect_location
, "Split group into %d and %d\n",
1699 group1_size
, group2_size
);
1704 /* Analyze an SLP instance starting from a group of grouped stores. Call
1705 vect_build_slp_tree to build a tree of packed stmts if possible.
1706 Return FALSE if it's impossible to SLP any stmt in the loop. */
1709 vect_analyze_slp_instance (vec_info
*vinfo
,
1710 gimple
*stmt
, unsigned max_tree_size
)
1712 slp_instance new_instance
;
1714 unsigned int group_size
= GROUP_SIZE (vinfo_for_stmt (stmt
));
1715 unsigned int unrolling_factor
= 1, nunits
;
1716 tree vectype
, scalar_type
= NULL_TREE
;
1719 unsigned int max_nunits
= 0;
1720 vec
<slp_tree
> loads
;
1721 struct data_reference
*dr
= STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
));
1722 vec
<gimple
*> scalar_stmts
;
1724 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
1728 scalar_type
= TREE_TYPE (DR_REF (dr
));
1729 vectype
= get_vectype_for_scalar_type (scalar_type
);
1733 gcc_assert (is_a
<loop_vec_info
> (vinfo
));
1734 vectype
= STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt
));
1737 group_size
= GROUP_SIZE (vinfo_for_stmt (stmt
));
1741 gcc_assert (is_a
<loop_vec_info
> (vinfo
));
1742 vectype
= STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt
));
1743 group_size
= as_a
<loop_vec_info
> (vinfo
)->reductions
.length ();
1748 if (dump_enabled_p ())
1750 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1751 "Build SLP failed: unsupported data-type ");
1752 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, scalar_type
);
1753 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
1758 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1760 /* Create a node (a root of the SLP tree) for the packed grouped stores. */
1761 scalar_stmts
.create (group_size
);
1763 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
1765 /* Collect the stores and store them in SLP_TREE_SCALAR_STMTS. */
1768 if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (next
))
1769 && STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next
)))
1770 scalar_stmts
.safe_push (
1771 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next
)));
1773 scalar_stmts
.safe_push (next
);
1774 next
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (next
));
1776 /* Mark the first element of the reduction chain as reduction to properly
1777 transform the node. In the reduction analysis phase only the last
1778 element of the chain is marked as reduction. */
1779 if (!STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt
)))
1780 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = vect_reduction_def
;
1784 /* Collect reduction statements. */
1785 vec
<gimple
*> reductions
= as_a
<loop_vec_info
> (vinfo
)->reductions
;
1786 for (i
= 0; reductions
.iterate (i
, &next
); i
++)
1787 scalar_stmts
.safe_push (next
);
1790 loads
.create (group_size
);
1792 /* Build the tree for the SLP instance. */
1793 bool *matches
= XALLOCAVEC (bool, group_size
);
1794 unsigned npermutes
= 0;
1795 node
= vect_build_slp_tree (vinfo
, scalar_stmts
, group_size
,
1796 &max_nunits
, &loads
, matches
, &npermutes
,
1797 NULL
, max_tree_size
);
1800 /* Calculate the unrolling factor based on the smallest type. */
1802 = least_common_multiple (max_nunits
, group_size
) / group_size
;
1804 if (unrolling_factor
!= 1
1805 && is_a
<bb_vec_info
> (vinfo
))
1808 if (max_nunits
> group_size
)
1810 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1811 "Build SLP failed: store group "
1812 "size not a multiple of the vector size "
1813 "in basic block SLP\n");
1814 vect_free_slp_tree (node
);
1818 /* Fatal mismatch. */
1819 matches
[group_size
/max_nunits
* max_nunits
] = false;
1820 vect_free_slp_tree (node
);
1825 /* Create a new SLP instance. */
1826 new_instance
= XNEW (struct _slp_instance
);
1827 SLP_INSTANCE_TREE (new_instance
) = node
;
1828 SLP_INSTANCE_GROUP_SIZE (new_instance
) = group_size
;
1829 SLP_INSTANCE_UNROLLING_FACTOR (new_instance
) = unrolling_factor
;
1830 SLP_INSTANCE_LOADS (new_instance
) = loads
;
1832 /* Compute the load permutation. */
1834 bool loads_permuted
= false;
1835 FOR_EACH_VEC_ELT (loads
, i
, load_node
)
1837 vec
<unsigned> load_permutation
;
1839 gimple
*load
, *first_stmt
;
1840 bool this_load_permuted
= false;
1841 load_permutation
.create (group_size
);
1842 first_stmt
= GROUP_FIRST_ELEMENT
1843 (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node
)[0]));
1844 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node
), j
, load
)
1846 int load_place
= vect_get_place_in_interleaving_chain
1848 gcc_assert (load_place
!= -1);
1849 if (load_place
!= j
)
1850 this_load_permuted
= true;
1851 load_permutation
.safe_push (load_place
);
1853 if (!this_load_permuted
1854 /* The load requires permutation when unrolling exposes
1855 a gap either because the group is larger than the SLP
1856 group-size or because there is a gap between the groups. */
1857 && (unrolling_factor
== 1
1858 || (group_size
== GROUP_SIZE (vinfo_for_stmt (first_stmt
))
1859 && GROUP_GAP (vinfo_for_stmt (first_stmt
)) == 0)))
1861 load_permutation
.release ();
1864 SLP_TREE_LOAD_PERMUTATION (load_node
) = load_permutation
;
1865 loads_permuted
= true;
1870 if (!vect_supported_load_permutation_p (new_instance
))
1872 if (dump_enabled_p ())
1874 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1875 "Build SLP failed: unsupported load "
1877 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
,
1879 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
1881 vect_free_slp_instance (new_instance
);
1886 /* If the loads and stores can be handled with load/store-lan
1887 instructions do not generate this SLP instance. */
1888 if (is_a
<loop_vec_info
> (vinfo
)
1890 && dr
&& vect_store_lanes_supported (vectype
, group_size
))
1893 FOR_EACH_VEC_ELT (loads
, i
, load_node
)
1895 gimple
*first_stmt
= GROUP_FIRST_ELEMENT
1896 (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node
)[0]));
1897 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (first_stmt
);
1898 /* Use SLP for strided accesses (or if we
1899 can't load-lanes). */
1900 if (STMT_VINFO_STRIDED_P (stmt_vinfo
)
1901 || ! vect_load_lanes_supported
1902 (STMT_VINFO_VECTYPE (stmt_vinfo
),
1903 GROUP_SIZE (stmt_vinfo
)))
1906 if (i
== loads
.length ())
1908 if (dump_enabled_p ())
1909 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1910 "Built SLP cancelled: can use "
1911 "load/store-lanes\n");
1912 vect_free_slp_instance (new_instance
);
1917 vinfo
->slp_instances
.safe_push (new_instance
);
1919 if (dump_enabled_p ())
1921 dump_printf_loc (MSG_NOTE
, vect_location
,
1922 "Final SLP tree for instance:\n");
1923 vect_print_slp_tree (MSG_NOTE
, vect_location
, node
);
1931 /* Failed to SLP. */
1932 /* Free the allocated memory. */
1933 scalar_stmts
.release ();
1937 /* For basic block SLP, try to break the group up into multiples of the
1939 if (is_a
<bb_vec_info
> (vinfo
)
1940 && GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
))
1941 && STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (stmt
)))
1943 /* We consider breaking the group only on VF boundaries from the existing
1945 for (i
= 0; i
< group_size
; i
++)
1946 if (!matches
[i
]) break;
1948 if (i
>= nunits
&& i
< group_size
)
1950 /* Split into two groups at the first vector boundary before i. */
1951 gcc_assert ((nunits
& (nunits
- 1)) == 0);
1952 unsigned group1_size
= i
& ~(nunits
- 1);
1954 gimple
*rest
= vect_split_slp_store_group (stmt
, group1_size
);
1955 bool res
= vect_analyze_slp_instance (vinfo
, stmt
, max_tree_size
);
1956 /* If the first non-match was in the middle of a vector,
1957 skip the rest of that vector. */
1958 if (group1_size
< i
)
1960 i
= group1_size
+ nunits
;
1962 rest
= vect_split_slp_store_group (rest
, nunits
);
1965 res
|= vect_analyze_slp_instance (vinfo
, rest
, max_tree_size
);
1968 /* Even though the first vector did not all match, we might be able to SLP
1969 (some) of the remainder. FORNOW ignore this possibility. */
1976 /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP
1977 trees of packed scalar stmts if SLP is possible. */
1980 vect_analyze_slp (vec_info
*vinfo
, unsigned max_tree_size
)
1983 gimple
*first_element
;
1986 if (dump_enabled_p ())
1987 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vect_analyze_slp ===\n");
1989 /* Find SLP sequences starting from groups of grouped stores. */
1990 FOR_EACH_VEC_ELT (vinfo
->grouped_stores
, i
, first_element
)
1991 if (vect_analyze_slp_instance (vinfo
, first_element
, max_tree_size
))
1994 if (loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (vinfo
))
1996 if (loop_vinfo
->reduction_chains
.length () > 0)
1998 /* Find SLP sequences starting from reduction chains. */
1999 FOR_EACH_VEC_ELT (loop_vinfo
->reduction_chains
, i
, first_element
)
2000 if (vect_analyze_slp_instance (vinfo
, first_element
,
2006 /* Don't try to vectorize SLP reductions if reduction chain was
2011 /* Find SLP sequences starting from groups of reductions. */
2012 if (loop_vinfo
->reductions
.length () > 1
2013 && vect_analyze_slp_instance (vinfo
, loop_vinfo
->reductions
[0],
2022 /* For each possible SLP instance decide whether to SLP it and calculate overall
2023 unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at
2024 least one instance. */
2027 vect_make_slp_decision (loop_vec_info loop_vinfo
)
2029 unsigned int i
, unrolling_factor
= 1;
2030 vec
<slp_instance
> slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
2031 slp_instance instance
;
2032 int decided_to_slp
= 0;
2034 if (dump_enabled_p ())
2035 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vect_make_slp_decision ==="
2038 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
2040 /* FORNOW: SLP if you can. */
2041 if (unrolling_factor
< SLP_INSTANCE_UNROLLING_FACTOR (instance
))
2042 unrolling_factor
= SLP_INSTANCE_UNROLLING_FACTOR (instance
);
2044 /* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we
2045 call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and
2046 loop-based vectorization. Such stmts will be marked as HYBRID. */
2047 vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance
), pure_slp
, -1);
2051 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
) = unrolling_factor
;
2053 if (decided_to_slp
&& dump_enabled_p ())
2054 dump_printf_loc (MSG_NOTE
, vect_location
,
2055 "Decided to SLP %d instances. Unrolling factor %d\n",
2056 decided_to_slp
, unrolling_factor
);
2058 return (decided_to_slp
> 0);
2062 /* Find stmts that must be both vectorized and SLPed (since they feed stmts that
2063 can't be SLPed) in the tree rooted at NODE. Mark such stmts as HYBRID. */
2066 vect_detect_hybrid_slp_stmts (slp_tree node
, unsigned i
, slp_vect_type stype
)
2068 gimple
*stmt
= SLP_TREE_SCALAR_STMTS (node
)[i
];
2069 imm_use_iterator imm_iter
;
2071 stmt_vec_info use_vinfo
, stmt_vinfo
= vinfo_for_stmt (stmt
);
2073 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
2074 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2077 /* Propagate hybrid down the SLP tree. */
2078 if (stype
== hybrid
)
2080 else if (HYBRID_SLP_STMT (stmt_vinfo
))
2084 /* Check if a pure SLP stmt has uses in non-SLP stmts. */
2085 gcc_checking_assert (PURE_SLP_STMT (stmt_vinfo
));
2086 /* If we get a pattern stmt here we have to use the LHS of the
2087 original stmt for immediate uses. */
2088 if (! STMT_VINFO_IN_PATTERN_P (stmt_vinfo
)
2089 && STMT_VINFO_RELATED_STMT (stmt_vinfo
))
2090 stmt
= STMT_VINFO_RELATED_STMT (stmt_vinfo
);
2091 if (TREE_CODE (gimple_op (stmt
, 0)) == SSA_NAME
)
2092 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, gimple_op (stmt
, 0))
2094 if (!flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
)))
2096 use_vinfo
= vinfo_for_stmt (use_stmt
);
2097 if (STMT_VINFO_IN_PATTERN_P (use_vinfo
)
2098 && STMT_VINFO_RELATED_STMT (use_vinfo
))
2099 use_vinfo
= vinfo_for_stmt (STMT_VINFO_RELATED_STMT (use_vinfo
));
2100 if (!STMT_SLP_TYPE (use_vinfo
)
2101 && (STMT_VINFO_RELEVANT (use_vinfo
)
2102 || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (use_vinfo
)))
2103 && !(gimple_code (use_stmt
) == GIMPLE_PHI
2104 && STMT_VINFO_DEF_TYPE (use_vinfo
) == vect_reduction_def
))
2106 if (dump_enabled_p ())
2108 dump_printf_loc (MSG_NOTE
, vect_location
, "use of SLP "
2109 "def in non-SLP stmt: ");
2110 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, use_stmt
, 0);
2118 && !HYBRID_SLP_STMT (stmt_vinfo
))
2120 if (dump_enabled_p ())
2122 dump_printf_loc (MSG_NOTE
, vect_location
, "marking hybrid: ");
2123 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
2125 STMT_SLP_TYPE (stmt_vinfo
) = hybrid
;
2128 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), j
, child
)
2129 if (SLP_TREE_DEF_TYPE (child
) != vect_external_def
)
2130 vect_detect_hybrid_slp_stmts (child
, i
, stype
);
2133 /* Helpers for vect_detect_hybrid_slp walking pattern stmt uses. */
2136 vect_detect_hybrid_slp_1 (tree
*tp
, int *, void *data
)
2138 walk_stmt_info
*wi
= (walk_stmt_info
*)data
;
2139 struct loop
*loopp
= (struct loop
*)wi
->info
;
2144 if (TREE_CODE (*tp
) == SSA_NAME
2145 && !SSA_NAME_IS_DEFAULT_DEF (*tp
))
2147 gimple
*def_stmt
= SSA_NAME_DEF_STMT (*tp
);
2148 if (flow_bb_inside_loop_p (loopp
, gimple_bb (def_stmt
))
2149 && PURE_SLP_STMT (vinfo_for_stmt (def_stmt
)))
2151 if (dump_enabled_p ())
2153 dump_printf_loc (MSG_NOTE
, vect_location
, "marking hybrid: ");
2154 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, def_stmt
, 0);
2156 STMT_SLP_TYPE (vinfo_for_stmt (def_stmt
)) = hybrid
;
2164 vect_detect_hybrid_slp_2 (gimple_stmt_iterator
*gsi
, bool *handled
,
2167 /* If the stmt is in a SLP instance then this isn't a reason
2168 to mark use definitions in other SLP instances as hybrid. */
2169 if (STMT_SLP_TYPE (vinfo_for_stmt (gsi_stmt (*gsi
))) != loop_vect
)
2174 /* Find stmts that must be both vectorized and SLPed. */
2177 vect_detect_hybrid_slp (loop_vec_info loop_vinfo
)
2180 vec
<slp_instance
> slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
2181 slp_instance instance
;
2183 if (dump_enabled_p ())
2184 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vect_detect_hybrid_slp ==="
2187 /* First walk all pattern stmt in the loop and mark defs of uses as
2188 hybrid because immediate uses in them are not recorded. */
2189 for (i
= 0; i
< LOOP_VINFO_LOOP (loop_vinfo
)->num_nodes
; ++i
)
2191 basic_block bb
= LOOP_VINFO_BBS (loop_vinfo
)[i
];
2192 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
2195 gimple
*stmt
= gsi_stmt (gsi
);
2196 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2197 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
2200 memset (&wi
, 0, sizeof (wi
));
2201 wi
.info
= LOOP_VINFO_LOOP (loop_vinfo
);
2202 gimple_stmt_iterator gsi2
2203 = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info
));
2204 walk_gimple_stmt (&gsi2
, vect_detect_hybrid_slp_2
,
2205 vect_detect_hybrid_slp_1
, &wi
);
2206 walk_gimple_seq (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
),
2207 vect_detect_hybrid_slp_2
,
2208 vect_detect_hybrid_slp_1
, &wi
);
2213 /* Then walk the SLP instance trees marking stmts with uses in
2214 non-SLP stmts as hybrid, also propagating hybrid down the
2215 SLP tree, collecting the above info on-the-fly. */
2216 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
2218 for (unsigned i
= 0; i
< SLP_INSTANCE_GROUP_SIZE (instance
); ++i
)
2219 vect_detect_hybrid_slp_stmts (SLP_INSTANCE_TREE (instance
),
2225 /* Create and initialize a new bb_vec_info struct for BB, as well as
2226 stmt_vec_info structs for all the stmts in it. */
2229 new_bb_vec_info (gimple_stmt_iterator region_begin
,
2230 gimple_stmt_iterator region_end
)
2232 basic_block bb
= gsi_bb (region_begin
);
2233 bb_vec_info res
= NULL
;
2234 gimple_stmt_iterator gsi
;
2236 res
= (bb_vec_info
) xcalloc (1, sizeof (struct _bb_vec_info
));
2237 res
->kind
= vec_info::bb
;
2238 BB_VINFO_BB (res
) = bb
;
2239 res
->region_begin
= region_begin
;
2240 res
->region_end
= region_end
;
2242 for (gsi
= region_begin
; gsi_stmt (gsi
) != gsi_stmt (region_end
);
2245 gimple
*stmt
= gsi_stmt (gsi
);
2246 gimple_set_uid (stmt
, 0);
2247 set_vinfo_for_stmt (stmt
, new_stmt_vec_info (stmt
, res
));
2250 BB_VINFO_GROUPED_STORES (res
).create (10);
2251 BB_VINFO_SLP_INSTANCES (res
).create (2);
2252 BB_VINFO_TARGET_COST_DATA (res
) = init_cost (NULL
);
2259 /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the
2260 stmts in the basic block. */
2263 destroy_bb_vec_info (bb_vec_info bb_vinfo
)
2265 slp_instance instance
;
2271 vect_destroy_datarefs (bb_vinfo
);
2272 free_dependence_relations (BB_VINFO_DDRS (bb_vinfo
));
2273 BB_VINFO_GROUPED_STORES (bb_vinfo
).release ();
2274 FOR_EACH_VEC_ELT (BB_VINFO_SLP_INSTANCES (bb_vinfo
), i
, instance
)
2275 vect_free_slp_instance (instance
);
2276 BB_VINFO_SLP_INSTANCES (bb_vinfo
).release ();
2277 destroy_cost_data (BB_VINFO_TARGET_COST_DATA (bb_vinfo
));
2279 for (gimple_stmt_iterator si
= bb_vinfo
->region_begin
;
2280 gsi_stmt (si
) != gsi_stmt (bb_vinfo
->region_end
); gsi_next (&si
))
2282 gimple
*stmt
= gsi_stmt (si
);
2283 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2286 /* Free stmt_vec_info. */
2287 free_stmt_vec_info (stmt
);
2289 /* Reset region marker. */
2290 gimple_set_uid (stmt
, -1);
2293 BB_VINFO_BB (bb_vinfo
)->aux
= NULL
;
2298 /* Analyze statements contained in SLP tree node after recursively analyzing
2299 the subtree. Return TRUE if the operations are supported. */
2302 vect_slp_analyze_node_operations (slp_tree node
)
2309 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
2312 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
2313 if (!vect_slp_analyze_node_operations (child
))
2317 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
2319 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2320 gcc_assert (stmt_info
);
2321 gcc_assert (STMT_SLP_TYPE (stmt_info
) != loop_vect
);
2323 /* Push SLP node def-type to stmt operands. */
2324 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), j
, child
)
2325 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
2326 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child
)[i
]))
2327 = SLP_TREE_DEF_TYPE (child
);
2328 res
= vect_analyze_stmt (stmt
, &dummy
, node
);
2329 /* Restore def-types. */
2330 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), j
, child
)
2331 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
2332 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (child
)[i
]))
2333 = vect_internal_def
;
2342 /* Analyze statements in SLP instances of the basic block. Return TRUE if the
2343 operations are supported. */
2346 vect_slp_analyze_operations (vec
<slp_instance
> slp_instances
, void *data
)
2348 slp_instance instance
;
2351 if (dump_enabled_p ())
2352 dump_printf_loc (MSG_NOTE
, vect_location
,
2353 "=== vect_slp_analyze_operations ===\n");
2355 for (i
= 0; slp_instances
.iterate (i
, &instance
); )
2357 if (!vect_slp_analyze_node_operations (SLP_INSTANCE_TREE (instance
)))
2359 dump_printf_loc (MSG_NOTE
, vect_location
,
2360 "removing SLP instance operations starting from: ");
2361 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
2362 SLP_TREE_SCALAR_STMTS
2363 (SLP_INSTANCE_TREE (instance
))[0], 0);
2364 vect_free_slp_instance (instance
);
2365 slp_instances
.ordered_remove (i
);
2369 /* Compute the costs of the SLP instance. */
2370 vect_analyze_slp_cost (instance
, data
);
2375 if (!slp_instances
.length ())
2382 /* Compute the scalar cost of the SLP node NODE and its children
2383 and return it. Do not account defs that are marked in LIFE and
2384 update LIFE according to uses of NODE. */
2387 vect_bb_slp_scalar_cost (basic_block bb
,
2388 slp_tree node
, vec
<bool, va_heap
> *life
)
2390 unsigned scalar_cost
= 0;
2395 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
2398 ssa_op_iter op_iter
;
2399 def_operand_p def_p
;
2400 stmt_vec_info stmt_info
;
2405 /* If there is a non-vectorized use of the defs then the scalar
2406 stmt is kept live in which case we do not account it or any
2407 required defs in the SLP children in the scalar cost. This
2408 way we make the vectorization more costly when compared to
2410 FOR_EACH_SSA_DEF_OPERAND (def_p
, stmt
, op_iter
, SSA_OP_DEF
)
2412 imm_use_iterator use_iter
;
2414 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, DEF_FROM_PTR (def_p
))
2415 if (!is_gimple_debug (use_stmt
)
2416 && (! vect_stmt_in_region_p (vinfo_for_stmt (stmt
)->vinfo
,
2418 || ! PURE_SLP_STMT (vinfo_for_stmt (use_stmt
))))
2421 BREAK_FROM_IMM_USE_STMT (use_iter
);
2427 /* Count scalar stmts only once. */
2428 if (gimple_visited_p (stmt
))
2430 gimple_set_visited (stmt
, true);
2432 stmt_info
= vinfo_for_stmt (stmt
);
2433 if (STMT_VINFO_DATA_REF (stmt_info
))
2435 if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info
)))
2436 stmt_cost
= vect_get_stmt_cost (scalar_load
);
2438 stmt_cost
= vect_get_stmt_cost (scalar_store
);
2441 stmt_cost
= vect_get_stmt_cost (scalar_stmt
);
2443 scalar_cost
+= stmt_cost
;
2446 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
2447 if (SLP_TREE_DEF_TYPE (child
) == vect_internal_def
)
2448 scalar_cost
+= vect_bb_slp_scalar_cost (bb
, child
, life
);
2453 /* Check if vectorization of the basic block is profitable. */
2456 vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo
)
2458 vec
<slp_instance
> slp_instances
= BB_VINFO_SLP_INSTANCES (bb_vinfo
);
2459 slp_instance instance
;
2461 unsigned int vec_inside_cost
= 0, vec_outside_cost
= 0, scalar_cost
= 0;
2462 unsigned int vec_prologue_cost
= 0, vec_epilogue_cost
= 0;
2464 /* Calculate scalar cost. */
2465 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
2467 auto_vec
<bool, 20> life
;
2468 life
.safe_grow_cleared (SLP_INSTANCE_GROUP_SIZE (instance
));
2469 scalar_cost
+= vect_bb_slp_scalar_cost (BB_VINFO_BB (bb_vinfo
),
2470 SLP_INSTANCE_TREE (instance
),
2474 /* Unset visited flag. */
2475 for (gimple_stmt_iterator gsi
= bb_vinfo
->region_begin
;
2476 gsi_stmt (gsi
) != gsi_stmt (bb_vinfo
->region_end
); gsi_next (&gsi
))
2477 gimple_set_visited (gsi_stmt (gsi
), false);
2479 /* Complete the target-specific cost calculation. */
2480 finish_cost (BB_VINFO_TARGET_COST_DATA (bb_vinfo
), &vec_prologue_cost
,
2481 &vec_inside_cost
, &vec_epilogue_cost
);
2483 vec_outside_cost
= vec_prologue_cost
+ vec_epilogue_cost
;
2485 if (dump_enabled_p ())
2487 dump_printf_loc (MSG_NOTE
, vect_location
, "Cost model analysis: \n");
2488 dump_printf (MSG_NOTE
, " Vector inside of basic block cost: %d\n",
2490 dump_printf (MSG_NOTE
, " Vector prologue cost: %d\n", vec_prologue_cost
);
2491 dump_printf (MSG_NOTE
, " Vector epilogue cost: %d\n", vec_epilogue_cost
);
2492 dump_printf (MSG_NOTE
, " Scalar cost of basic block: %d\n", scalar_cost
);
2495 /* Vectorization is profitable if its cost is more than the cost of scalar
2496 version. Note that we err on the vector side for equal cost because
2497 the cost estimate is otherwise quite pessimistic (constant uses are
2498 free on the scalar side but cost a load on the vector side for
2500 if (vec_outside_cost
+ vec_inside_cost
> scalar_cost
)
2506 /* Check if the basic block can be vectorized. Returns a bb_vec_info
2507 if so and sets fatal to true if failure is independent of
2508 current_vector_size. */
2511 vect_slp_analyze_bb_1 (gimple_stmt_iterator region_begin
,
2512 gimple_stmt_iterator region_end
,
2513 vec
<data_reference_p
> datarefs
, int n_stmts
,
2516 bb_vec_info bb_vinfo
;
2517 slp_instance instance
;
2521 /* The first group of checks is independent of the vector size. */
2524 if (n_stmts
> PARAM_VALUE (PARAM_SLP_MAX_INSNS_IN_BB
))
2526 if (dump_enabled_p ())
2527 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2528 "not vectorized: too many instructions in "
2530 free_data_refs (datarefs
);
2534 bb_vinfo
= new_bb_vec_info (region_begin
, region_end
);
2538 BB_VINFO_DATAREFS (bb_vinfo
) = datarefs
;
2540 /* Analyze the data references. */
2542 if (!vect_analyze_data_refs (bb_vinfo
, &min_vf
))
2544 if (dump_enabled_p ())
2545 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2546 "not vectorized: unhandled data-ref in basic "
2549 destroy_bb_vec_info (bb_vinfo
);
2553 if (BB_VINFO_DATAREFS (bb_vinfo
).length () < 2)
2555 if (dump_enabled_p ())
2556 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2557 "not vectorized: not enough data-refs in "
2560 destroy_bb_vec_info (bb_vinfo
);
2564 if (!vect_analyze_data_ref_accesses (bb_vinfo
))
2566 if (dump_enabled_p ())
2567 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2568 "not vectorized: unhandled data access in "
2571 destroy_bb_vec_info (bb_vinfo
);
2575 /* If there are no grouped stores in the region there is no need
2576 to continue with pattern recog as vect_analyze_slp will fail
2578 if (bb_vinfo
->grouped_stores
.is_empty ())
2580 if (dump_enabled_p ())
2581 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2582 "not vectorized: no grouped stores in "
2585 destroy_bb_vec_info (bb_vinfo
);
2589 /* While the rest of the analysis below depends on it in some way. */
2592 vect_pattern_recog (bb_vinfo
);
2594 /* Check the SLP opportunities in the basic block, analyze and build SLP
2596 if (!vect_analyze_slp (bb_vinfo
, n_stmts
))
2598 if (dump_enabled_p ())
2600 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2601 "Failed to SLP the basic block.\n");
2602 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2603 "not vectorized: failed to find SLP opportunities "
2604 "in basic block.\n");
2607 destroy_bb_vec_info (bb_vinfo
);
2611 /* Analyze and verify the alignment of data references and the
2612 dependence in the SLP instances. */
2613 for (i
= 0; BB_VINFO_SLP_INSTANCES (bb_vinfo
).iterate (i
, &instance
); )
2615 if (! vect_slp_analyze_and_verify_instance_alignment (instance
)
2616 || ! vect_slp_analyze_instance_dependence (instance
))
2618 dump_printf_loc (MSG_NOTE
, vect_location
,
2619 "removing SLP instance operations starting from: ");
2620 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
2621 SLP_TREE_SCALAR_STMTS
2622 (SLP_INSTANCE_TREE (instance
))[0], 0);
2623 vect_free_slp_instance (instance
);
2624 BB_VINFO_SLP_INSTANCES (bb_vinfo
).ordered_remove (i
);
2628 /* Mark all the statements that we want to vectorize as pure SLP and
2630 vect_mark_slp_stmts (SLP_INSTANCE_TREE (instance
), pure_slp
, -1);
2631 vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance
));
2635 if (! BB_VINFO_SLP_INSTANCES (bb_vinfo
).length ())
2637 destroy_bb_vec_info (bb_vinfo
);
2641 if (!vect_slp_analyze_operations (BB_VINFO_SLP_INSTANCES (bb_vinfo
),
2642 BB_VINFO_TARGET_COST_DATA (bb_vinfo
)))
2644 if (dump_enabled_p ())
2645 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2646 "not vectorized: bad operation in basic block.\n");
2648 destroy_bb_vec_info (bb_vinfo
);
2652 /* Cost model: check if the vectorization is worthwhile. */
2653 if (!unlimited_cost_model (NULL
)
2654 && !vect_bb_vectorization_profitable_p (bb_vinfo
))
2656 if (dump_enabled_p ())
2657 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2658 "not vectorized: vectorization is not "
2661 destroy_bb_vec_info (bb_vinfo
);
2665 if (dump_enabled_p ())
2666 dump_printf_loc (MSG_NOTE
, vect_location
,
2667 "Basic block will be vectorized using SLP\n");
2673 /* Main entry for the BB vectorizer. Analyze and transform BB, returns
2674 true if anything in the basic-block was vectorized. */
2677 vect_slp_bb (basic_block bb
)
2679 bb_vec_info bb_vinfo
;
2680 gimple_stmt_iterator gsi
;
2681 unsigned int vector_sizes
;
2682 bool any_vectorized
= false;
2684 if (dump_enabled_p ())
2685 dump_printf_loc (MSG_NOTE
, vect_location
, "===vect_slp_analyze_bb===\n");
2687 /* Autodetect first vector size we try. */
2688 current_vector_size
= 0;
2689 vector_sizes
= targetm
.vectorize
.autovectorize_vector_sizes ();
2691 gsi
= gsi_start_bb (bb
);
2695 if (gsi_end_p (gsi
))
2698 gimple_stmt_iterator region_begin
= gsi
;
2699 vec
<data_reference_p
> datarefs
= vNULL
;
2702 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
2704 gimple
*stmt
= gsi_stmt (gsi
);
2705 if (is_gimple_debug (stmt
))
2709 if (gimple_location (stmt
) != UNKNOWN_LOCATION
)
2710 vect_location
= gimple_location (stmt
);
2712 if (!find_data_references_in_stmt (NULL
, stmt
, &datarefs
))
2716 /* Skip leading unhandled stmts. */
2717 if (gsi_stmt (region_begin
) == gsi_stmt (gsi
))
2723 gimple_stmt_iterator region_end
= gsi
;
2725 bool vectorized
= false;
2727 bb_vinfo
= vect_slp_analyze_bb_1 (region_begin
, region_end
,
2728 datarefs
, insns
, fatal
);
2730 && dbg_cnt (vect_slp
))
2732 if (dump_enabled_p ())
2733 dump_printf_loc (MSG_NOTE
, vect_location
, "SLPing BB part\n");
2735 vect_schedule_slp (bb_vinfo
);
2737 if (dump_enabled_p ())
2738 dump_printf_loc (MSG_NOTE
, vect_location
,
2739 "basic block part vectorized\n");
2741 destroy_bb_vec_info (bb_vinfo
);
2746 destroy_bb_vec_info (bb_vinfo
);
2748 any_vectorized
|= vectorized
;
2750 vector_sizes
&= ~current_vector_size
;
2752 || vector_sizes
== 0
2753 || current_vector_size
== 0
2754 /* If vect_slp_analyze_bb_1 signaled that analysis for all
2755 vector sizes will fail do not bother iterating. */
2758 if (gsi_end_p (region_end
))
2761 /* Skip the unhandled stmt. */
2764 /* And reset vector sizes. */
2765 current_vector_size
= 0;
2766 vector_sizes
= targetm
.vectorize
.autovectorize_vector_sizes ();
2770 /* Try the next biggest vector size. */
2771 current_vector_size
= 1 << floor_log2 (vector_sizes
);
2772 if (dump_enabled_p ())
2773 dump_printf_loc (MSG_NOTE
, vect_location
,
2774 "***** Re-trying analysis with "
2775 "vector size %d\n", current_vector_size
);
2782 return any_vectorized
;
2786 /* Return 1 if vector type of boolean constant which is OPNUM
2787 operand in statement STMT is a boolean vector. */
2790 vect_mask_constant_operand_p (gimple
*stmt
, int opnum
)
2792 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
2793 enum tree_code code
= gimple_expr_code (stmt
);
2796 enum vect_def_type dt
;
2798 /* For comparison and COND_EXPR type is chosen depending
2799 on the other comparison operand. */
2800 if (TREE_CODE_CLASS (code
) == tcc_comparison
)
2803 op
= gimple_assign_rhs1 (stmt
);
2805 op
= gimple_assign_rhs2 (stmt
);
2807 if (!vect_is_simple_use (op
, stmt_vinfo
->vinfo
, &def_stmt
,
2811 return !vectype
|| VECTOR_BOOLEAN_TYPE_P (vectype
);
2814 if (code
== COND_EXPR
)
2816 tree cond
= gimple_assign_rhs1 (stmt
);
2818 if (TREE_CODE (cond
) == SSA_NAME
)
2822 op
= TREE_OPERAND (cond
, 1);
2824 op
= TREE_OPERAND (cond
, 0);
2826 if (!vect_is_simple_use (op
, stmt_vinfo
->vinfo
, &def_stmt
,
2830 return !vectype
|| VECTOR_BOOLEAN_TYPE_P (vectype
);
2833 return VECTOR_BOOLEAN_TYPE_P (STMT_VINFO_VECTYPE (stmt_vinfo
));
2837 /* For constant and loop invariant defs of SLP_NODE this function returns
2838 (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts.
2839 OP_NUM determines if we gather defs for operand 0 or operand 1 of the RHS of
2840 scalar stmts. NUMBER_OF_VECTORS is the number of vector defs to create.
2841 REDUC_INDEX is the index of the reduction operand in the statements, unless
2845 vect_get_constant_vectors (tree op
, slp_tree slp_node
,
2846 vec
<tree
> *vec_oprnds
,
2847 unsigned int op_num
, unsigned int number_of_vectors
,
2850 vec
<gimple
*> stmts
= SLP_TREE_SCALAR_STMTS (slp_node
);
2851 gimple
*stmt
= stmts
[0];
2852 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
2856 unsigned j
, number_of_places_left_in_vector
;
2859 int group_size
= stmts
.length ();
2860 unsigned int vec_num
, i
;
2861 unsigned number_of_copies
= 1;
2863 voprnds
.create (number_of_vectors
);
2864 bool constant_p
, is_store
;
2865 tree neutral_op
= NULL
;
2866 enum tree_code code
= gimple_expr_code (stmt
);
2869 gimple_seq ctor_seq
= NULL
;
2871 /* Check if vector type is a boolean vector. */
2872 if (TREE_CODE (TREE_TYPE (op
)) == BOOLEAN_TYPE
2873 && vect_mask_constant_operand_p (stmt
, op_num
))
2875 = build_same_sized_truth_vector_type (STMT_VINFO_VECTYPE (stmt_vinfo
));
2877 vector_type
= get_vectype_for_scalar_type (TREE_TYPE (op
));
2878 nunits
= TYPE_VECTOR_SUBPARTS (vector_type
);
2880 if (STMT_VINFO_DEF_TYPE (stmt_vinfo
) == vect_reduction_def
2881 && reduc_index
!= -1)
2883 op_num
= reduc_index
;
2884 op
= gimple_op (stmt
, op_num
+ 1);
2885 /* For additional copies (see the explanation of NUMBER_OF_COPIES below)
2886 we need either neutral operands or the original operands. See
2887 get_initial_def_for_reduction() for details. */
2890 case WIDEN_SUM_EXPR
:
2897 if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op
)))
2898 neutral_op
= build_real (TREE_TYPE (op
), dconst0
);
2900 neutral_op
= build_int_cst (TREE_TYPE (op
), 0);
2905 if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (op
)))
2906 neutral_op
= build_real (TREE_TYPE (op
), dconst1
);
2908 neutral_op
= build_int_cst (TREE_TYPE (op
), 1);
2913 neutral_op
= build_int_cst (TREE_TYPE (op
), -1);
2916 /* For MIN/MAX we don't have an easy neutral operand but
2917 the initial values can be used fine here. Only for
2918 a reduction chain we have to force a neutral element. */
2921 if (!GROUP_FIRST_ELEMENT (stmt_vinfo
))
2925 def_stmt
= SSA_NAME_DEF_STMT (op
);
2926 loop
= (gimple_bb (stmt
))->loop_father
;
2927 neutral_op
= PHI_ARG_DEF_FROM_EDGE (def_stmt
,
2928 loop_preheader_edge (loop
));
2933 gcc_assert (!GROUP_FIRST_ELEMENT (stmt_vinfo
));
2938 if (STMT_VINFO_DATA_REF (stmt_vinfo
))
2941 op
= gimple_assign_rhs1 (stmt
);
2948 if (CONSTANT_CLASS_P (op
))
2953 /* NUMBER_OF_COPIES is the number of times we need to use the same values in
2954 created vectors. It is greater than 1 if unrolling is performed.
2956 For example, we have two scalar operands, s1 and s2 (e.g., group of
2957 strided accesses of size two), while NUNITS is four (i.e., four scalars
2958 of this type can be packed in a vector). The output vector will contain
2959 two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
2962 If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
2963 containing the operands.
2965 For example, NUNITS is four as before, and the group size is 8
2966 (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
2967 {s5, s6, s7, s8}. */
2969 number_of_copies
= nunits
* number_of_vectors
/ group_size
;
2971 number_of_places_left_in_vector
= nunits
;
2972 elts
= XALLOCAVEC (tree
, nunits
);
2973 bool place_after_defs
= false;
2974 for (j
= 0; j
< number_of_copies
; j
++)
2976 for (i
= group_size
- 1; stmts
.iterate (i
, &stmt
); i
--)
2979 op
= gimple_assign_rhs1 (stmt
);
2986 tree cond
= gimple_assign_rhs1 (stmt
);
2987 if (TREE_CODE (cond
) == SSA_NAME
)
2988 op
= gimple_op (stmt
, op_num
+ 1);
2989 else if (op_num
== 0 || op_num
== 1)
2990 op
= TREE_OPERAND (cond
, op_num
);
2994 op
= gimple_assign_rhs2 (stmt
);
2996 op
= gimple_assign_rhs3 (stmt
);
3002 op
= gimple_call_arg (stmt
, op_num
);
3009 op
= gimple_op (stmt
, op_num
+ 1);
3010 /* Unlike the other binary operators, shifts/rotates have
3011 the shift count being int, instead of the same type as
3012 the lhs, so make sure the scalar is the right type if
3013 we are dealing with vectors of
3014 long long/long/short/char. */
3015 if (op_num
== 1 && TREE_CODE (op
) == INTEGER_CST
)
3016 op
= fold_convert (TREE_TYPE (vector_type
), op
);
3020 op
= gimple_op (stmt
, op_num
+ 1);
3025 if (reduc_index
!= -1)
3027 loop
= (gimple_bb (stmt
))->loop_father
;
3028 def_stmt
= SSA_NAME_DEF_STMT (op
);
3032 /* Get the def before the loop. In reduction chain we have only
3033 one initial value. */
3034 if ((j
!= (number_of_copies
- 1)
3035 || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
))
3040 op
= PHI_ARG_DEF_FROM_EDGE (def_stmt
,
3041 loop_preheader_edge (loop
));
3044 /* Create 'vect_ = {op0,op1,...,opn}'. */
3045 number_of_places_left_in_vector
--;
3047 if (!types_compatible_p (TREE_TYPE (vector_type
), TREE_TYPE (op
)))
3049 if (CONSTANT_CLASS_P (op
))
3051 if (VECTOR_BOOLEAN_TYPE_P (vector_type
))
3053 /* Can't use VIEW_CONVERT_EXPR for booleans because
3054 of possibly different sizes of scalar value and
3056 if (integer_zerop (op
))
3057 op
= build_int_cst (TREE_TYPE (vector_type
), 0);
3058 else if (integer_onep (op
))
3059 op
= build_int_cst (TREE_TYPE (vector_type
), 1);
3064 op
= fold_unary (VIEW_CONVERT_EXPR
,
3065 TREE_TYPE (vector_type
), op
);
3066 gcc_assert (op
&& CONSTANT_CLASS_P (op
));
3070 tree new_temp
= make_ssa_name (TREE_TYPE (vector_type
));
3072 if (VECTOR_BOOLEAN_TYPE_P (vector_type
))
3074 gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (op
)));
3075 init_stmt
= gimple_build_assign (new_temp
, NOP_EXPR
, op
);
3079 op
= build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (vector_type
),
3082 = gimple_build_assign (new_temp
, VIEW_CONVERT_EXPR
,
3085 gimple_seq_add_stmt (&ctor_seq
, init_stmt
);
3089 elts
[number_of_places_left_in_vector
] = op
;
3090 if (!CONSTANT_CLASS_P (op
))
3092 if (TREE_CODE (orig_op
) == SSA_NAME
3093 && !SSA_NAME_IS_DEFAULT_DEF (orig_op
)
3094 && STMT_VINFO_BB_VINFO (stmt_vinfo
)
3095 && (STMT_VINFO_BB_VINFO (stmt_vinfo
)->bb
3096 == gimple_bb (SSA_NAME_DEF_STMT (orig_op
))))
3097 place_after_defs
= true;
3099 if (number_of_places_left_in_vector
== 0)
3101 number_of_places_left_in_vector
= nunits
;
3104 vec_cst
= build_vector (vector_type
, elts
);
3107 vec
<constructor_elt
, va_gc
> *v
;
3109 vec_alloc (v
, nunits
);
3110 for (k
= 0; k
< nunits
; ++k
)
3111 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, elts
[k
]);
3112 vec_cst
= build_constructor (vector_type
, v
);
3115 gimple_stmt_iterator gsi
;
3116 if (place_after_defs
)
3119 (vect_find_last_scalar_stmt_in_slp (slp_node
));
3120 init
= vect_init_vector (stmt
, vec_cst
, vector_type
, &gsi
);
3123 init
= vect_init_vector (stmt
, vec_cst
, vector_type
, NULL
);
3124 if (ctor_seq
!= NULL
)
3126 gsi
= gsi_for_stmt (SSA_NAME_DEF_STMT (init
));
3127 gsi_insert_seq_before_without_update (&gsi
, ctor_seq
,
3131 voprnds
.quick_push (init
);
3132 place_after_defs
= false;
3137 /* Since the vectors are created in the reverse order, we should invert
3139 vec_num
= voprnds
.length ();
3140 for (j
= vec_num
; j
!= 0; j
--)
3142 vop
= voprnds
[j
- 1];
3143 vec_oprnds
->quick_push (vop
);
3148 /* In case that VF is greater than the unrolling factor needed for the SLP
3149 group of stmts, NUMBER_OF_VECTORS to be created is greater than
3150 NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
3151 to replicate the vectors. */
3152 while (number_of_vectors
> vec_oprnds
->length ())
3154 tree neutral_vec
= NULL
;
3159 neutral_vec
= build_vector_from_val (vector_type
, neutral_op
);
3161 vec_oprnds
->quick_push (neutral_vec
);
3165 for (i
= 0; vec_oprnds
->iterate (i
, &vop
) && i
< vec_num
; i
++)
3166 vec_oprnds
->quick_push (vop
);
3172 /* Get vectorized definitions from SLP_NODE that contains corresponding
3173 vectorized def-stmts. */
3176 vect_get_slp_vect_defs (slp_tree slp_node
, vec
<tree
> *vec_oprnds
)
3179 gimple
*vec_def_stmt
;
3182 gcc_assert (SLP_TREE_VEC_STMTS (slp_node
).exists ());
3184 FOR_EACH_VEC_ELT (SLP_TREE_VEC_STMTS (slp_node
), i
, vec_def_stmt
)
3186 gcc_assert (vec_def_stmt
);
3187 vec_oprnd
= gimple_get_lhs (vec_def_stmt
);
3188 vec_oprnds
->quick_push (vec_oprnd
);
3193 /* Get vectorized definitions for SLP_NODE.
3194 If the scalar definitions are loop invariants or constants, collect them and
3195 call vect_get_constant_vectors() to create vector stmts.
3196 Otherwise, the def-stmts must be already vectorized and the vectorized stmts
3197 must be stored in the corresponding child of SLP_NODE, and we call
3198 vect_get_slp_vect_defs () to retrieve them. */
3201 vect_get_slp_defs (vec
<tree
> ops
, slp_tree slp_node
,
3202 vec
<vec
<tree
> > *vec_oprnds
, int reduc_index
)
3205 int number_of_vects
= 0, i
;
3206 unsigned int child_index
= 0;
3207 HOST_WIDE_INT lhs_size_unit
, rhs_size_unit
;
3208 slp_tree child
= NULL
;
3211 bool vectorized_defs
;
3213 first_stmt
= SLP_TREE_SCALAR_STMTS (slp_node
)[0];
3214 FOR_EACH_VEC_ELT (ops
, i
, oprnd
)
3216 /* For each operand we check if it has vectorized definitions in a child
3217 node or we need to create them (for invariants and constants). We
3218 check if the LHS of the first stmt of the next child matches OPRND.
3219 If it does, we found the correct child. Otherwise, we call
3220 vect_get_constant_vectors (), and not advance CHILD_INDEX in order
3221 to check this child node for the next operand. */
3222 vectorized_defs
= false;
3223 if (SLP_TREE_CHILDREN (slp_node
).length () > child_index
)
3225 child
= SLP_TREE_CHILDREN (slp_node
)[child_index
];
3227 /* We have to check both pattern and original def, if available. */
3228 if (SLP_TREE_DEF_TYPE (child
) == vect_internal_def
)
3230 gimple
*first_def
= SLP_TREE_SCALAR_STMTS (child
)[0];
3232 = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (first_def
));
3234 if (operand_equal_p (oprnd
, gimple_get_lhs (first_def
), 0)
3236 && operand_equal_p (oprnd
, gimple_get_lhs (related
), 0)))
3238 /* The number of vector defs is determined by the number of
3239 vector statements in the node from which we get those
3241 number_of_vects
= SLP_TREE_NUMBER_OF_VEC_STMTS (child
);
3242 vectorized_defs
= true;
3250 if (!vectorized_defs
)
3254 number_of_vects
= SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node
);
3255 /* Number of vector stmts was calculated according to LHS in
3256 vect_schedule_slp_instance (), fix it by replacing LHS with
3257 RHS, if necessary. See vect_get_smallest_scalar_type () for
3259 vect_get_smallest_scalar_type (first_stmt
, &lhs_size_unit
,
3261 if (rhs_size_unit
!= lhs_size_unit
)
3263 number_of_vects
*= rhs_size_unit
;
3264 number_of_vects
/= lhs_size_unit
;
3269 /* Allocate memory for vectorized defs. */
3271 vec_defs
.create (number_of_vects
);
3273 /* For reduction defs we call vect_get_constant_vectors (), since we are
3274 looking for initial loop invariant values. */
3275 if (vectorized_defs
&& reduc_index
== -1)
3276 /* The defs are already vectorized. */
3277 vect_get_slp_vect_defs (child
, &vec_defs
);
3279 /* Build vectors from scalar defs. */
3280 vect_get_constant_vectors (oprnd
, slp_node
, &vec_defs
, i
,
3281 number_of_vects
, reduc_index
);
3283 vec_oprnds
->quick_push (vec_defs
);
3285 /* For reductions, we only need initial values. */
3286 if (reduc_index
!= -1)
3292 /* Create NCOPIES permutation statements using the mask MASK_BYTES (by
3293 building a vector of type MASK_TYPE from it) and two input vectors placed in
3294 DR_CHAIN at FIRST_VEC_INDX and SECOND_VEC_INDX for the first copy and
3295 shifting by STRIDE elements of DR_CHAIN for every copy.
3296 (STRIDE is the number of vectorized stmts for NODE divided by the number of
3298 VECT_STMTS_COUNTER specifies the index in the vectorized stmts of NODE, where
3299 the created stmts must be inserted. */
3302 vect_create_mask_and_perm (gimple
*stmt
,
3303 tree mask
, int first_vec_indx
, int second_vec_indx
,
3304 gimple_stmt_iterator
*gsi
, slp_tree node
,
3305 tree vectype
, vec
<tree
> dr_chain
,
3306 int ncopies
, int vect_stmts_counter
)
3309 gimple
*perm_stmt
= NULL
;
3310 int i
, stride_in
, stride_out
;
3311 tree first_vec
, second_vec
, data_ref
;
3313 stride_out
= SLP_TREE_NUMBER_OF_VEC_STMTS (node
) / ncopies
;
3314 stride_in
= dr_chain
.length () / ncopies
;
3316 /* Initialize the vect stmts of NODE to properly insert the generated
3318 for (i
= SLP_TREE_VEC_STMTS (node
).length ();
3319 i
< (int) SLP_TREE_NUMBER_OF_VEC_STMTS (node
); i
++)
3320 SLP_TREE_VEC_STMTS (node
).quick_push (NULL
);
3322 perm_dest
= vect_create_destination_var (gimple_assign_lhs (stmt
), vectype
);
3323 for (i
= 0; i
< ncopies
; i
++)
3325 first_vec
= dr_chain
[first_vec_indx
];
3326 second_vec
= dr_chain
[second_vec_indx
];
3328 /* Generate the permute statement if necessary. */
3331 perm_stmt
= gimple_build_assign (perm_dest
, VEC_PERM_EXPR
,
3332 first_vec
, second_vec
, mask
);
3333 data_ref
= make_ssa_name (perm_dest
, perm_stmt
);
3334 gimple_set_lhs (perm_stmt
, data_ref
);
3335 vect_finish_stmt_generation (stmt
, perm_stmt
, gsi
);
3338 /* If mask was NULL_TREE generate the requested identity transform. */
3339 perm_stmt
= SSA_NAME_DEF_STMT (first_vec
);
3341 /* Store the vector statement in NODE. */
3342 SLP_TREE_VEC_STMTS (node
)[stride_out
* i
+ vect_stmts_counter
]
3345 first_vec_indx
+= stride_in
;
3346 second_vec_indx
+= stride_in
;
3351 /* Generate vector permute statements from a list of loads in DR_CHAIN.
3352 If ANALYZE_ONLY is TRUE, only check that it is possible to create valid
3353 permute statements for the SLP node NODE of the SLP instance
3354 SLP_NODE_INSTANCE. */
3357 vect_transform_slp_perm_load (slp_tree node
, vec
<tree
> dr_chain
,
3358 gimple_stmt_iterator
*gsi
, int vf
,
3359 slp_instance slp_node_instance
, bool analyze_only
)
3361 gimple
*stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
3362 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3363 tree mask_element_type
= NULL_TREE
, mask_type
;
3364 int nunits
, vec_index
= 0;
3365 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
3366 int group_size
= SLP_INSTANCE_GROUP_SIZE (slp_node_instance
);
3367 int unroll_factor
, mask_element
, ncopies
;
3368 unsigned char *mask
;
3371 if (!STMT_VINFO_GROUPED_ACCESS (stmt_info
))
3374 stmt_info
= vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info
));
3376 mode
= TYPE_MODE (vectype
);
3378 /* The generic VEC_PERM_EXPR code always uses an integral type of the
3379 same size as the vector element being permuted. */
3380 mask_element_type
= lang_hooks
.types
.type_for_mode
3381 (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype
))), 1);
3382 mask_type
= get_vectype_for_scalar_type (mask_element_type
);
3383 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3384 mask
= XALLOCAVEC (unsigned char, nunits
);
3385 unroll_factor
= SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance
);
3387 /* Number of copies is determined by the final vectorization factor
3388 relatively to SLP_NODE_INSTANCE unrolling factor. */
3389 ncopies
= vf
/ SLP_INSTANCE_UNROLLING_FACTOR (slp_node_instance
);
3391 /* Generate permutation masks for every NODE. Number of masks for each NODE
3392 is equal to GROUP_SIZE.
3393 E.g., we have a group of three nodes with three loads from the same
3394 location in each node, and the vector size is 4. I.e., we have a
3395 a0b0c0a1b1c1... sequence and we need to create the following vectors:
3396 for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3
3397 for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3
3400 The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}.
3401 The last mask is illegal since we assume two operands for permute
3402 operation, and the mask element values can't be outside that range.
3403 Hence, the last mask must be converted into {2,5,5,5}.
3404 For the first two permutations we need the first and the second input
3405 vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation
3406 we need the second and the third vectors: {b1,c1,a2,b2} and
3409 int vect_stmts_counter
= 0;
3411 int first_vec_index
= -1;
3412 int second_vec_index
= -1;
3415 for (int j
= 0; j
< unroll_factor
; j
++)
3417 for (int k
= 0; k
< group_size
; k
++)
3419 int i
= (SLP_TREE_LOAD_PERMUTATION (node
)[k
]
3420 + j
* STMT_VINFO_GROUP_SIZE (stmt_info
));
3421 vec_index
= i
/ nunits
;
3422 mask_element
= i
% nunits
;
3423 if (vec_index
== first_vec_index
3424 || first_vec_index
== -1)
3426 first_vec_index
= vec_index
;
3428 else if (vec_index
== second_vec_index
3429 || second_vec_index
== -1)
3431 second_vec_index
= vec_index
;
3432 mask_element
+= nunits
;
3436 if (dump_enabled_p ())
3438 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
3439 "permutation requires at "
3440 "least three vectors ");
3441 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
3443 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
3448 gcc_assert (mask_element
>= 0
3449 && mask_element
< 2 * nunits
);
3450 if (mask_element
!= index
)
3452 mask
[index
++] = mask_element
;
3454 if (index
== nunits
)
3457 && ! can_vec_perm_p (mode
, false, mask
))
3459 if (dump_enabled_p ())
3461 dump_printf_loc (MSG_MISSED_OPTIMIZATION
,
3463 "unsupported vect permute { ");
3464 for (i
= 0; i
< nunits
; ++i
)
3465 dump_printf (MSG_MISSED_OPTIMIZATION
, "%d ", mask
[i
]);
3466 dump_printf (MSG_MISSED_OPTIMIZATION
, "}\n");
3473 tree mask_vec
= NULL_TREE
;
3477 tree
*mask_elts
= XALLOCAVEC (tree
, nunits
);
3478 for (int l
= 0; l
< nunits
; ++l
)
3479 mask_elts
[l
] = build_int_cst (mask_element_type
,
3481 mask_vec
= build_vector (mask_type
, mask_elts
);
3484 if (second_vec_index
== -1)
3485 second_vec_index
= first_vec_index
;
3486 vect_create_mask_and_perm (stmt
, mask_vec
, first_vec_index
,
3488 gsi
, node
, vectype
, dr_chain
,
3489 ncopies
, vect_stmts_counter
++);
3493 first_vec_index
= -1;
3494 second_vec_index
= -1;
3505 /* Vectorize SLP instance tree in postorder. */
3508 vect_schedule_slp_instance (slp_tree node
, slp_instance instance
,
3509 unsigned int vectorization_factor
)
3512 bool grouped_store
, is_store
;
3513 gimple_stmt_iterator si
;
3514 stmt_vec_info stmt_info
;
3515 unsigned int vec_stmts_size
, nunits
, group_size
;
3520 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
3523 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3524 vect_schedule_slp_instance (child
, instance
, vectorization_factor
);
3526 /* Push SLP node def-type to stmts. */
3527 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3528 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
3529 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
3530 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = SLP_TREE_DEF_TYPE (child
);
3532 stmt
= SLP_TREE_SCALAR_STMTS (node
)[0];
3533 stmt_info
= vinfo_for_stmt (stmt
);
3535 /* VECTYPE is the type of the destination. */
3536 vectype
= STMT_VINFO_VECTYPE (stmt_info
);
3537 nunits
= (unsigned int) TYPE_VECTOR_SUBPARTS (vectype
);
3538 group_size
= SLP_INSTANCE_GROUP_SIZE (instance
);
3540 /* For each SLP instance calculate number of vector stmts to be created
3541 for the scalar stmts in each node of the SLP tree. Number of vector
3542 elements in one vector iteration is the number of scalar elements in
3543 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
3545 Unless this is a SLP reduction in which case the number of vector
3546 stmts is equal to the number of vector stmts of the children. */
3547 if (GROUP_FIRST_ELEMENT (stmt_info
)
3548 && !STMT_VINFO_GROUPED_ACCESS (stmt_info
))
3549 vec_stmts_size
= SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_CHILDREN (node
)[0]);
3551 vec_stmts_size
= (vectorization_factor
* group_size
) / nunits
;
3553 if (!SLP_TREE_VEC_STMTS (node
).exists ())
3555 SLP_TREE_VEC_STMTS (node
).create (vec_stmts_size
);
3556 SLP_TREE_NUMBER_OF_VEC_STMTS (node
) = vec_stmts_size
;
3559 if (dump_enabled_p ())
3561 dump_printf_loc (MSG_NOTE
,vect_location
,
3562 "------>vectorizing SLP node starting from: ");
3563 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
3564 dump_printf (MSG_NOTE
, "\n");
3567 /* Vectorized stmts go before the last scalar stmt which is where
3568 all uses are ready. */
3569 si
= gsi_for_stmt (vect_find_last_scalar_stmt_in_slp (node
));
3571 /* Mark the first element of the reduction chain as reduction to properly
3572 transform the node. In the analysis phase only the last element of the
3573 chain is marked as reduction. */
3574 if (GROUP_FIRST_ELEMENT (stmt_info
) && !STMT_VINFO_GROUPED_ACCESS (stmt_info
)
3575 && GROUP_FIRST_ELEMENT (stmt_info
) == stmt
)
3577 STMT_VINFO_DEF_TYPE (stmt_info
) = vect_reduction_def
;
3578 STMT_VINFO_TYPE (stmt_info
) = reduc_vec_info_type
;
3581 /* Handle two-operation SLP nodes by vectorizing the group with
3582 both operations and then performing a merge. */
3583 if (SLP_TREE_TWO_OPERATORS (node
))
3585 enum tree_code code0
= gimple_assign_rhs_code (stmt
);
3586 enum tree_code ocode
= ERROR_MARK
;
3588 unsigned char *mask
= XALLOCAVEC (unsigned char, group_size
);
3589 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, ostmt
)
3590 if (gimple_assign_rhs_code (ostmt
) != code0
)
3593 ocode
= gimple_assign_rhs_code (ostmt
);
3597 if (ocode
!= ERROR_MARK
)
3602 tree tmask
= NULL_TREE
;
3603 vect_transform_stmt (stmt
, &si
, &grouped_store
, node
, instance
);
3604 v0
= SLP_TREE_VEC_STMTS (node
).copy ();
3605 SLP_TREE_VEC_STMTS (node
).truncate (0);
3606 gimple_assign_set_rhs_code (stmt
, ocode
);
3607 vect_transform_stmt (stmt
, &si
, &grouped_store
, node
, instance
);
3608 gimple_assign_set_rhs_code (stmt
, code0
);
3609 v1
= SLP_TREE_VEC_STMTS (node
).copy ();
3610 SLP_TREE_VEC_STMTS (node
).truncate (0);
3611 tree meltype
= build_nonstandard_integer_type
3612 (GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype
))), 1);
3613 tree mvectype
= get_same_sized_vectype (meltype
, vectype
);
3615 for (j
= 0; j
< v0
.length (); ++j
)
3617 tree
*melts
= XALLOCAVEC (tree
, TYPE_VECTOR_SUBPARTS (vectype
));
3618 for (l
= 0; l
< TYPE_VECTOR_SUBPARTS (vectype
); ++l
)
3620 if (k
>= group_size
)
3622 melts
[l
] = build_int_cst
3623 (meltype
, mask
[k
++] * TYPE_VECTOR_SUBPARTS (vectype
) + l
);
3625 tmask
= build_vector (mvectype
, melts
);
3627 /* ??? Not all targets support a VEC_PERM_EXPR with a
3628 constant mask that would translate to a vec_merge RTX
3629 (with their vec_perm_const_ok). We can either not
3630 vectorize in that case or let veclower do its job.
3631 Unfortunately that isn't too great and at least for
3632 plus/minus we'd eventually like to match targets
3633 vector addsub instructions. */
3635 vstmt
= gimple_build_assign (make_ssa_name (vectype
),
3637 gimple_assign_lhs (v0
[j
]),
3638 gimple_assign_lhs (v1
[j
]), tmask
);
3639 vect_finish_stmt_generation (stmt
, vstmt
, &si
);
3640 SLP_TREE_VEC_STMTS (node
).quick_push (vstmt
);
3647 is_store
= vect_transform_stmt (stmt
, &si
, &grouped_store
, node
, instance
);
3649 /* Restore stmt def-types. */
3650 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3651 if (SLP_TREE_DEF_TYPE (child
) != vect_internal_def
)
3652 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (child
), j
, stmt
)
3653 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt
)) = vect_internal_def
;
3658 /* Replace scalar calls from SLP node NODE with setting of their lhs to zero.
3659 For loop vectorization this is done in vectorizable_call, but for SLP
3660 it needs to be deferred until end of vect_schedule_slp, because multiple
3661 SLP instances may refer to the same scalar stmt. */
3664 vect_remove_slp_scalar_calls (slp_tree node
)
3666 gimple
*stmt
, *new_stmt
;
3667 gimple_stmt_iterator gsi
;
3671 stmt_vec_info stmt_info
;
3673 if (SLP_TREE_DEF_TYPE (node
) != vect_internal_def
)
3676 FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node
), i
, child
)
3677 vect_remove_slp_scalar_calls (child
);
3679 FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node
), i
, stmt
)
3681 if (!is_gimple_call (stmt
) || gimple_bb (stmt
) == NULL
)
3683 stmt_info
= vinfo_for_stmt (stmt
);
3684 if (stmt_info
== NULL
3685 || is_pattern_stmt_p (stmt_info
)
3686 || !PURE_SLP_STMT (stmt_info
))
3688 lhs
= gimple_call_lhs (stmt
);
3689 new_stmt
= gimple_build_assign (lhs
, build_zero_cst (TREE_TYPE (lhs
)));
3690 set_vinfo_for_stmt (new_stmt
, stmt_info
);
3691 set_vinfo_for_stmt (stmt
, NULL
);
3692 STMT_VINFO_STMT (stmt_info
) = new_stmt
;
3693 gsi
= gsi_for_stmt (stmt
);
3694 gsi_replace (&gsi
, new_stmt
, false);
3695 SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt
)) = new_stmt
;
3699 /* Generate vector code for all SLP instances in the loop/basic block. */
3702 vect_schedule_slp (vec_info
*vinfo
)
3704 vec
<slp_instance
> slp_instances
;
3705 slp_instance instance
;
3707 bool is_store
= false;
3709 slp_instances
= vinfo
->slp_instances
;
3710 if (is_a
<loop_vec_info
> (vinfo
))
3711 vf
= as_a
<loop_vec_info
> (vinfo
)->vectorization_factor
;
3715 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
3717 /* Schedule the tree of INSTANCE. */
3718 is_store
= vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance
),
3720 if (dump_enabled_p ())
3721 dump_printf_loc (MSG_NOTE
, vect_location
,
3722 "vectorizing stmts using SLP.\n");
3725 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
3727 slp_tree root
= SLP_INSTANCE_TREE (instance
);
3730 gimple_stmt_iterator gsi
;
3732 /* Remove scalar call stmts. Do not do this for basic-block
3733 vectorization as not all uses may be vectorized.
3734 ??? Why should this be necessary? DCE should be able to
3735 remove the stmts itself.
3736 ??? For BB vectorization we can as well remove scalar
3737 stmts starting from the SLP tree root if they have no
3739 if (is_a
<loop_vec_info
> (vinfo
))
3740 vect_remove_slp_scalar_calls (root
);
3742 for (j
= 0; SLP_TREE_SCALAR_STMTS (root
).iterate (j
, &store
)
3743 && j
< SLP_INSTANCE_GROUP_SIZE (instance
); j
++)
3745 if (!STMT_VINFO_DATA_REF (vinfo_for_stmt (store
)))
3748 if (is_pattern_stmt_p (vinfo_for_stmt (store
)))
3749 store
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (store
));
3750 /* Free the attached stmt_vec_info and remove the stmt. */
3751 gsi
= gsi_for_stmt (store
);
3752 unlink_stmt_vdef (store
);
3753 gsi_remove (&gsi
, true);
3754 release_defs (store
);
3755 free_stmt_vec_info (store
);