1 /* Statement Analysis and Transformation for Vectorization
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 and Ira Rosen <irar@il.ibm.com>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
30 #include "basic-block.h"
31 #include "tree-pretty-print.h"
32 #include "gimple-pretty-print.h"
33 #include "tree-flow.h"
34 #include "tree-dump.h"
36 #include "cfglayout.h"
40 #include "diagnostic-core.h"
42 #include "tree-vectorizer.h"
43 #include "langhooks.h"
46 /* Utility functions used by vect_mark_stmts_to_be_vectorized. */
48 /* Function vect_mark_relevant.
50 Mark STMT as "relevant for vectorization" and add it to WORKLIST. */
53 vect_mark_relevant (VEC(gimple,heap) **worklist, gimple stmt,
54 enum vect_relevant relevant, bool live_p)
56 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
57 enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info);
58 bool save_live_p = STMT_VINFO_LIVE_P (stmt_info);
60 if (vect_print_dump_info (REPORT_DETAILS))
61 fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p);
63 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
67 /* This is the last stmt in a sequence that was detected as a
68 pattern that can potentially be vectorized. Don't mark the stmt
69 as relevant/live because it's not going to be vectorized.
70 Instead mark the pattern-stmt that replaces it. */
72 pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
74 if (vect_print_dump_info (REPORT_DETAILS))
75 fprintf (vect_dump, "last stmt in pattern. don't mark relevant/live.");
76 stmt_info = vinfo_for_stmt (pattern_stmt);
77 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt);
78 save_relevant = STMT_VINFO_RELEVANT (stmt_info);
79 save_live_p = STMT_VINFO_LIVE_P (stmt_info);
83 STMT_VINFO_LIVE_P (stmt_info) |= live_p;
84 if (relevant > STMT_VINFO_RELEVANT (stmt_info))
85 STMT_VINFO_RELEVANT (stmt_info) = relevant;
87 if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant
88 && STMT_VINFO_LIVE_P (stmt_info) == save_live_p)
90 if (vect_print_dump_info (REPORT_DETAILS))
91 fprintf (vect_dump, "already marked relevant/live.");
95 VEC_safe_push (gimple, heap, *worklist, stmt);
99 /* Function vect_stmt_relevant_p.
101 Return true if STMT in loop that is represented by LOOP_VINFO is
102 "relevant for vectorization".
104 A stmt is considered "relevant for vectorization" if:
105 - it has uses outside the loop.
106 - it has vdefs (it alters memory).
107 - control stmts in the loop (except for the exit condition).
109 CHECKME: what other side effects would the vectorizer allow? */
112 vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo,
113 enum vect_relevant *relevant, bool *live_p)
115 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
117 imm_use_iterator imm_iter;
121 *relevant = vect_unused_in_scope;
124 /* cond stmt other than loop exit cond. */
125 if (is_ctrl_stmt (stmt)
126 && STMT_VINFO_TYPE (vinfo_for_stmt (stmt))
127 != loop_exit_ctrl_vec_info_type)
128 *relevant = vect_used_in_scope;
130 /* changing memory. */
131 if (gimple_code (stmt) != GIMPLE_PHI)
132 if (gimple_vdef (stmt))
134 if (vect_print_dump_info (REPORT_DETAILS))
135 fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs.");
136 *relevant = vect_used_in_scope;
139 /* uses outside the loop. */
140 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF)
142 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p))
144 basic_block bb = gimple_bb (USE_STMT (use_p));
145 if (!flow_bb_inside_loop_p (loop, bb))
147 if (vect_print_dump_info (REPORT_DETAILS))
148 fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop.");
150 if (is_gimple_debug (USE_STMT (use_p)))
153 /* We expect all such uses to be in the loop exit phis
154 (because of loop closed form) */
155 gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI);
156 gcc_assert (bb == single_exit (loop)->dest);
163 return (*live_p || *relevant);
167 /* Function exist_non_indexing_operands_for_use_p
169 USE is one of the uses attached to STMT. Check if USE is
170 used in STMT for anything other than indexing an array. */
173 exist_non_indexing_operands_for_use_p (tree use, gimple stmt)
176 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
178 /* USE corresponds to some operand in STMT. If there is no data
179 reference in STMT, then any operand that corresponds to USE
180 is not indexing an array. */
181 if (!STMT_VINFO_DATA_REF (stmt_info))
184 /* STMT has a data_ref. FORNOW this means that its of one of
188 (This should have been verified in analyze_data_refs).
190 'var' in the second case corresponds to a def, not a use,
191 so USE cannot correspond to any operands that are not used
194 Therefore, all we need to check is if STMT falls into the
195 first case, and whether var corresponds to USE. */
197 if (!gimple_assign_copy_p (stmt))
199 if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
201 operand = gimple_assign_rhs1 (stmt);
202 if (TREE_CODE (operand) != SSA_NAME)
213 Function process_use.
216 - a USE in STMT in a loop represented by LOOP_VINFO
217 - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt
218 that defined USE. This is done by calling mark_relevant and passing it
219 the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant).
222 Generally, LIVE_P and RELEVANT are used to define the liveness and
223 relevance info of the DEF_STMT of this USE:
224 STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p
225 STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant
227 - case 1: If USE is used only for address computations (e.g. array indexing),
228 which does not need to be directly vectorized, then the liveness/relevance
229 of the respective DEF_STMT is left unchanged.
230 - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we
231 skip DEF_STMT cause it had already been processed.
232 - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will
233 be modified accordingly.
235 Return true if everything is as expected. Return false otherwise. */
238 process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p,
239 enum vect_relevant relevant, VEC(gimple,heap) **worklist)
241 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
242 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
243 stmt_vec_info dstmt_vinfo;
244 basic_block bb, def_bb;
247 enum vect_def_type dt;
249 /* case 1: we are only interested in uses that need to be vectorized. Uses
250 that are used for address computation are not considered relevant. */
251 if (!exist_non_indexing_operands_for_use_p (use, stmt))
254 if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt))
256 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
257 fprintf (vect_dump, "not vectorized: unsupported use in stmt.");
261 if (!def_stmt || gimple_nop_p (def_stmt))
264 def_bb = gimple_bb (def_stmt);
265 if (!flow_bb_inside_loop_p (loop, def_bb))
267 if (vect_print_dump_info (REPORT_DETAILS))
268 fprintf (vect_dump, "def_stmt is out of loop.");
272 /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT).
273 DEF_STMT must have already been processed, because this should be the
274 only way that STMT, which is a reduction-phi, was put in the worklist,
275 as there should be no other uses for DEF_STMT in the loop. So we just
276 check that everything is as expected, and we are done. */
277 dstmt_vinfo = vinfo_for_stmt (def_stmt);
278 bb = gimple_bb (stmt);
279 if (gimple_code (stmt) == GIMPLE_PHI
280 && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def
281 && gimple_code (def_stmt) != GIMPLE_PHI
282 && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def
283 && bb->loop_father == def_bb->loop_father)
285 if (vect_print_dump_info (REPORT_DETAILS))
286 fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest.");
287 if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo))
288 dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo));
289 gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction);
290 gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo)
291 || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope);
295 /* case 3a: outer-loop stmt defining an inner-loop stmt:
296 outer-loop-header-bb:
302 if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father))
304 if (vect_print_dump_info (REPORT_DETAILS))
305 fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt.");
309 case vect_unused_in_scope:
310 relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ?
311 vect_used_in_scope : vect_unused_in_scope;
314 case vect_used_in_outer_by_reduction:
315 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def);
316 relevant = vect_used_by_reduction;
319 case vect_used_in_outer:
320 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def);
321 relevant = vect_used_in_scope;
324 case vect_used_in_scope:
332 /* case 3b: inner-loop stmt defining an outer-loop stmt:
333 outer-loop-header-bb:
337 outer-loop-tail-bb (or outer-loop-exit-bb in double reduction):
339 else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father))
341 if (vect_print_dump_info (REPORT_DETAILS))
342 fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt.");
346 case vect_unused_in_scope:
347 relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def
348 || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ?
349 vect_used_in_outer_by_reduction : vect_unused_in_scope;
352 case vect_used_by_reduction:
353 relevant = vect_used_in_outer_by_reduction;
356 case vect_used_in_scope:
357 relevant = vect_used_in_outer;
365 vect_mark_relevant (worklist, def_stmt, relevant, live_p);
370 /* Function vect_mark_stmts_to_be_vectorized.
372 Not all stmts in the loop need to be vectorized. For example:
381 Stmt 1 and 3 do not need to be vectorized, because loop control and
382 addressing of vectorized data-refs are handled differently.
384 This pass detects such stmts. */
387 vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo)
389 VEC(gimple,heap) *worklist;
390 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
391 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
392 unsigned int nbbs = loop->num_nodes;
393 gimple_stmt_iterator si;
396 stmt_vec_info stmt_vinfo;
400 enum vect_relevant relevant, tmp_relevant;
401 enum vect_def_type def_type;
403 if (vect_print_dump_info (REPORT_DETAILS))
404 fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ===");
406 worklist = VEC_alloc (gimple, heap, 64);
408 /* 1. Init worklist. */
409 for (i = 0; i < nbbs; i++)
412 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
415 if (vect_print_dump_info (REPORT_DETAILS))
417 fprintf (vect_dump, "init: phi relevant? ");
418 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
421 if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p))
422 vect_mark_relevant (&worklist, phi, relevant, live_p);
424 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
426 stmt = gsi_stmt (si);
427 if (vect_print_dump_info (REPORT_DETAILS))
429 fprintf (vect_dump, "init: stmt relevant? ");
430 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
433 if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p))
434 vect_mark_relevant (&worklist, stmt, relevant, live_p);
438 /* 2. Process_worklist */
439 while (VEC_length (gimple, worklist) > 0)
444 stmt = VEC_pop (gimple, worklist);
445 if (vect_print_dump_info (REPORT_DETAILS))
447 fprintf (vect_dump, "worklist: examine stmt: ");
448 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
451 /* Examine the USEs of STMT. For each USE, mark the stmt that defines it
452 (DEF_STMT) as relevant/irrelevant and live/dead according to the
453 liveness and relevance properties of STMT. */
454 stmt_vinfo = vinfo_for_stmt (stmt);
455 relevant = STMT_VINFO_RELEVANT (stmt_vinfo);
456 live_p = STMT_VINFO_LIVE_P (stmt_vinfo);
458 /* Generally, the liveness and relevance properties of STMT are
459 propagated as is to the DEF_STMTs of its USEs:
460 live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO)
461 relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO)
463 One exception is when STMT has been identified as defining a reduction
464 variable; in this case we set the liveness/relevance as follows:
466 relevant = vect_used_by_reduction
467 This is because we distinguish between two kinds of relevant stmts -
468 those that are used by a reduction computation, and those that are
469 (also) used by a regular computation. This allows us later on to
470 identify stmts that are used solely by a reduction, and therefore the
471 order of the results that they produce does not have to be kept. */
473 def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo);
474 tmp_relevant = relevant;
477 case vect_reduction_def:
478 switch (tmp_relevant)
480 case vect_unused_in_scope:
481 relevant = vect_used_by_reduction;
484 case vect_used_by_reduction:
485 if (gimple_code (stmt) == GIMPLE_PHI)
490 if (vect_print_dump_info (REPORT_DETAILS))
491 fprintf (vect_dump, "unsupported use of reduction.");
493 VEC_free (gimple, heap, worklist);
500 case vect_nested_cycle:
501 if (tmp_relevant != vect_unused_in_scope
502 && tmp_relevant != vect_used_in_outer_by_reduction
503 && tmp_relevant != vect_used_in_outer)
505 if (vect_print_dump_info (REPORT_DETAILS))
506 fprintf (vect_dump, "unsupported use of nested cycle.");
508 VEC_free (gimple, heap, worklist);
515 case vect_double_reduction_def:
516 if (tmp_relevant != vect_unused_in_scope
517 && tmp_relevant != vect_used_by_reduction)
519 if (vect_print_dump_info (REPORT_DETAILS))
520 fprintf (vect_dump, "unsupported use of double reduction.");
522 VEC_free (gimple, heap, worklist);
533 FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE)
535 tree op = USE_FROM_PTR (use_p);
536 if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist))
538 VEC_free (gimple, heap, worklist);
542 } /* while worklist */
544 VEC_free (gimple, heap, worklist);
549 /* Get cost by calling cost target builtin. */
552 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost)
554 tree dummy_type = NULL;
557 return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
562 /* Get cost for STMT. */
565 cost_for_stmt (gimple stmt)
567 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
569 switch (STMT_VINFO_TYPE (stmt_info))
571 case load_vec_info_type:
572 return vect_get_stmt_cost (scalar_load);
573 case store_vec_info_type:
574 return vect_get_stmt_cost (scalar_store);
575 case op_vec_info_type:
576 case condition_vec_info_type:
577 case assignment_vec_info_type:
578 case reduc_vec_info_type:
579 case induc_vec_info_type:
580 case type_promotion_vec_info_type:
581 case type_demotion_vec_info_type:
582 case type_conversion_vec_info_type:
583 case call_vec_info_type:
584 return vect_get_stmt_cost (scalar_stmt);
585 case undef_vec_info_type:
591 /* Function vect_model_simple_cost.
593 Models cost for simple operations, i.e. those that only emit ncopies of a
594 single op. Right now, this does not account for multiple insns that could
595 be generated for the single vector op. We will handle that shortly. */
598 vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies,
599 enum vect_def_type *dt, slp_tree slp_node)
602 int inside_cost = 0, outside_cost = 0;
604 /* The SLP costs were already calculated during SLP tree build. */
605 if (PURE_SLP_STMT (stmt_info))
608 inside_cost = ncopies * vect_get_stmt_cost (vector_stmt);
610 /* FORNOW: Assuming maximum 2 args per stmts. */
611 for (i = 0; i < 2; i++)
613 if (dt[i] == vect_constant_def || dt[i] == vect_external_def)
614 outside_cost += vect_get_stmt_cost (vector_stmt);
617 if (vect_print_dump_info (REPORT_COST))
618 fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, "
619 "outside_cost = %d .", inside_cost, outside_cost);
621 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
622 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
623 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
627 /* Function vect_cost_strided_group_size
629 For strided load or store, return the group_size only if it is the first
630 load or store of a group, else return 1. This ensures that group size is
631 only returned once per group. */
634 vect_cost_strided_group_size (stmt_vec_info stmt_info)
636 gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info);
638 if (first_stmt == STMT_VINFO_STMT (stmt_info))
639 return DR_GROUP_SIZE (stmt_info);
645 /* Function vect_model_store_cost
647 Models cost for stores. In the case of strided accesses, one access
648 has the overhead of the strided access attributed to it. */
651 vect_model_store_cost (stmt_vec_info stmt_info, int ncopies,
652 enum vect_def_type dt, slp_tree slp_node)
655 unsigned int inside_cost = 0, outside_cost = 0;
656 struct data_reference *first_dr;
659 /* The SLP costs were already calculated during SLP tree build. */
660 if (PURE_SLP_STMT (stmt_info))
663 if (dt == vect_constant_def || dt == vect_external_def)
664 outside_cost = vect_get_stmt_cost (scalar_to_vec);
666 /* Strided access? */
667 if (DR_GROUP_FIRST_DR (stmt_info))
671 first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0);
676 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
677 group_size = vect_cost_strided_group_size (stmt_info);
680 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
682 /* Not a strided access. */
686 first_dr = STMT_VINFO_DATA_REF (stmt_info);
689 /* Is this an access in a group of stores, which provide strided access?
690 If so, add in the cost of the permutes. */
693 /* Uses a high and low interleave operation for each needed permute. */
694 inside_cost = ncopies * exact_log2(group_size) * group_size
695 * vect_get_stmt_cost (vector_stmt);
697 if (vect_print_dump_info (REPORT_COST))
698 fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .",
703 /* Costs of the stores. */
704 vect_get_store_cost (first_dr, ncopies, &inside_cost);
706 if (vect_print_dump_info (REPORT_COST))
707 fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, "
708 "outside_cost = %d .", inside_cost, outside_cost);
710 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
711 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
712 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
716 /* Calculate cost of DR's memory access. */
718 vect_get_store_cost (struct data_reference *dr, int ncopies,
719 unsigned int *inside_cost)
721 int alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
723 switch (alignment_support_scheme)
727 *inside_cost += ncopies * vect_get_stmt_cost (vector_store);
729 if (vect_print_dump_info (REPORT_COST))
730 fprintf (vect_dump, "vect_model_store_cost: aligned.");
735 case dr_unaligned_supported:
737 gimple stmt = DR_STMT (dr);
738 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
739 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
741 /* Here, we assign an additional cost for the unaligned store. */
742 *inside_cost += ncopies
743 * targetm.vectorize.builtin_vectorization_cost (unaligned_store,
744 vectype, DR_MISALIGNMENT (dr));
746 if (vect_print_dump_info (REPORT_COST))
747 fprintf (vect_dump, "vect_model_store_cost: unaligned supported by "
759 /* Function vect_model_load_cost
761 Models cost for loads. In the case of strided accesses, the last access
762 has the overhead of the strided access attributed to it. Since unaligned
763 accesses are supported for loads, we also account for the costs of the
764 access scheme chosen. */
767 vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node)
772 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
773 unsigned int inside_cost = 0, outside_cost = 0;
775 /* The SLP costs were already calculated during SLP tree build. */
776 if (PURE_SLP_STMT (stmt_info))
779 /* Strided accesses? */
780 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
781 if (first_stmt && !slp_node)
783 group_size = vect_cost_strided_group_size (stmt_info);
784 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
786 /* Not a strided access. */
793 /* Is this an access in a group of loads providing strided access?
794 If so, add in the cost of the permutes. */
797 /* Uses an even and odd extract operations for each needed permute. */
798 inside_cost = ncopies * exact_log2(group_size) * group_size
799 * vect_get_stmt_cost (vector_stmt);
801 if (vect_print_dump_info (REPORT_COST))
802 fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .",
806 /* The loads themselves. */
807 vect_get_load_cost (first_dr, ncopies,
808 ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node),
809 &inside_cost, &outside_cost);
811 if (vect_print_dump_info (REPORT_COST))
812 fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, "
813 "outside_cost = %d .", inside_cost, outside_cost);
815 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
816 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
817 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
821 /* Calculate cost of DR's memory access. */
823 vect_get_load_cost (struct data_reference *dr, int ncopies,
824 bool add_realign_cost, unsigned int *inside_cost,
825 unsigned int *outside_cost)
827 int alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
829 switch (alignment_support_scheme)
833 *inside_cost += ncopies * vect_get_stmt_cost (vector_load);
835 if (vect_print_dump_info (REPORT_COST))
836 fprintf (vect_dump, "vect_model_load_cost: aligned.");
840 case dr_unaligned_supported:
842 gimple stmt = DR_STMT (dr);
843 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
844 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
846 /* Here, we assign an additional cost for the unaligned load. */
847 *inside_cost += ncopies
848 * targetm.vectorize.builtin_vectorization_cost (unaligned_load,
849 vectype, DR_MISALIGNMENT (dr));
850 if (vect_print_dump_info (REPORT_COST))
851 fprintf (vect_dump, "vect_model_load_cost: unaligned supported by "
856 case dr_explicit_realign:
858 *inside_cost += ncopies * (2 * vect_get_stmt_cost (vector_load)
859 + vect_get_stmt_cost (vector_stmt));
861 /* FIXME: If the misalignment remains fixed across the iterations of
862 the containing loop, the following cost should be added to the
864 if (targetm.vectorize.builtin_mask_for_load)
865 *inside_cost += vect_get_stmt_cost (vector_stmt);
869 case dr_explicit_realign_optimized:
871 if (vect_print_dump_info (REPORT_COST))
872 fprintf (vect_dump, "vect_model_load_cost: unaligned software "
875 /* Unaligned software pipeline has a load of an address, an initial
876 load, and possibly a mask operation to "prime" the loop. However,
877 if this is an access in a group of loads, which provide strided
878 access, then the above cost should only be considered for one
879 access in the group. Inside the loop, there is a load op
880 and a realignment op. */
882 if (add_realign_cost)
884 *outside_cost = 2 * vect_get_stmt_cost (vector_stmt);
885 if (targetm.vectorize.builtin_mask_for_load)
886 *outside_cost += vect_get_stmt_cost (vector_stmt);
889 *inside_cost += ncopies * (vect_get_stmt_cost (vector_load)
890 + vect_get_stmt_cost (vector_stmt));
900 /* Function vect_init_vector.
902 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
903 the vector elements of VECTOR_VAR. Place the initialization at BSI if it
904 is not NULL. Otherwise, place the initialization at the loop preheader.
905 Return the DEF of INIT_STMT.
906 It will be used in the vectorization of STMT. */
909 vect_init_vector (gimple stmt, tree vector_var, tree vector_type,
910 gimple_stmt_iterator *gsi)
912 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
920 new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_");
921 add_referenced_var (new_var);
922 init_stmt = gimple_build_assign (new_var, vector_var);
923 new_temp = make_ssa_name (new_var, init_stmt);
924 gimple_assign_set_lhs (init_stmt, new_temp);
927 vect_finish_stmt_generation (stmt, init_stmt, gsi);
930 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
934 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
936 if (nested_in_vect_loop_p (loop, stmt))
939 pe = loop_preheader_edge (loop);
940 new_bb = gsi_insert_on_edge_immediate (pe, init_stmt);
941 gcc_assert (!new_bb);
945 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo);
947 gimple_stmt_iterator gsi_bb_start;
949 gcc_assert (bb_vinfo);
950 bb = BB_VINFO_BB (bb_vinfo);
951 gsi_bb_start = gsi_after_labels (bb);
952 gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT);
956 if (vect_print_dump_info (REPORT_DETAILS))
958 fprintf (vect_dump, "created new init_stmt: ");
959 print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM);
962 vec_oprnd = gimple_assign_lhs (init_stmt);
967 /* Function vect_get_vec_def_for_operand.
969 OP is an operand in STMT. This function returns a (vector) def that will be
970 used in the vectorized stmt for STMT.
972 In the case that OP is an SSA_NAME which is defined in the loop, then
973 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
975 In case OP is an invariant or constant, a new stmt that creates a vector def
976 needs to be introduced. */
979 vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def)
984 stmt_vec_info def_stmt_info = NULL;
985 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
987 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
991 enum vect_def_type dt;
995 if (vect_print_dump_info (REPORT_DETAILS))
997 fprintf (vect_dump, "vect_get_vec_def_for_operand: ");
998 print_generic_expr (vect_dump, op, TDF_SLIM);
1001 is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def,
1003 gcc_assert (is_simple_use);
1004 if (vect_print_dump_info (REPORT_DETAILS))
1008 fprintf (vect_dump, "def = ");
1009 print_generic_expr (vect_dump, def, TDF_SLIM);
1013 fprintf (vect_dump, " def_stmt = ");
1014 print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM);
1020 /* Case 1: operand is a constant. */
1021 case vect_constant_def:
1023 vector_type = get_vectype_for_scalar_type (TREE_TYPE (op));
1024 gcc_assert (vector_type);
1025 nunits = TYPE_VECTOR_SUBPARTS (vector_type);
1030 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1031 if (vect_print_dump_info (REPORT_DETAILS))
1032 fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits);
1034 vec_cst = build_vector_from_val (vector_type, op);
1035 return vect_init_vector (stmt, vec_cst, vector_type, NULL);
1038 /* Case 2: operand is defined outside the loop - loop invariant. */
1039 case vect_external_def:
1041 vector_type = get_vectype_for_scalar_type (TREE_TYPE (def));
1042 gcc_assert (vector_type);
1043 nunits = TYPE_VECTOR_SUBPARTS (vector_type);
1048 /* Create 'vec_inv = {inv,inv,..,inv}' */
1049 if (vect_print_dump_info (REPORT_DETAILS))
1050 fprintf (vect_dump, "Create vector_inv.");
1052 vec_inv = build_vector_from_val (vector_type, def);
1053 return vect_init_vector (stmt, vec_inv, vector_type, NULL);
1056 /* Case 3: operand is defined inside the loop. */
1057 case vect_internal_def:
1060 *scalar_def = NULL/* FIXME tuples: def_stmt*/;
1062 /* Get the def from the vectorized stmt. */
1063 def_stmt_info = vinfo_for_stmt (def_stmt);
1064 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1065 gcc_assert (vec_stmt);
1066 if (gimple_code (vec_stmt) == GIMPLE_PHI)
1067 vec_oprnd = PHI_RESULT (vec_stmt);
1068 else if (is_gimple_call (vec_stmt))
1069 vec_oprnd = gimple_call_lhs (vec_stmt);
1071 vec_oprnd = gimple_assign_lhs (vec_stmt);
1075 /* Case 4: operand is defined by a loop header phi - reduction */
1076 case vect_reduction_def:
1077 case vect_double_reduction_def:
1078 case vect_nested_cycle:
1082 gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
1083 loop = (gimple_bb (def_stmt))->loop_father;
1085 /* Get the def before the loop */
1086 op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop));
1087 return get_initial_def_for_reduction (stmt, op, scalar_def);
1090 /* Case 5: operand is defined by loop-header phi - induction. */
1091 case vect_induction_def:
1093 gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
1095 /* Get the def from the vectorized stmt. */
1096 def_stmt_info = vinfo_for_stmt (def_stmt);
1097 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1098 gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI);
1099 vec_oprnd = PHI_RESULT (vec_stmt);
1109 /* Function vect_get_vec_def_for_stmt_copy
1111 Return a vector-def for an operand. This function is used when the
1112 vectorized stmt to be created (by the caller to this function) is a "copy"
1113 created in case the vectorized result cannot fit in one vector, and several
1114 copies of the vector-stmt are required. In this case the vector-def is
1115 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1116 of the stmt that defines VEC_OPRND.
1117 DT is the type of the vector def VEC_OPRND.
1120 In case the vectorization factor (VF) is bigger than the number
1121 of elements that can fit in a vectype (nunits), we have to generate
1122 more than one vector stmt to vectorize the scalar stmt. This situation
1123 arises when there are multiple data-types operated upon in the loop; the
1124 smallest data-type determines the VF, and as a result, when vectorizing
1125 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1126 vector stmt (each computing a vector of 'nunits' results, and together
1127 computing 'VF' results in each iteration). This function is called when
1128 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1129 which VF=16 and nunits=4, so the number of copies required is 4):
1131 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1133 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1134 VS1.1: vx.1 = memref1 VS1.2
1135 VS1.2: vx.2 = memref2 VS1.3
1136 VS1.3: vx.3 = memref3
1138 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1139 VSnew.1: vz1 = vx.1 + ... VSnew.2
1140 VSnew.2: vz2 = vx.2 + ... VSnew.3
1141 VSnew.3: vz3 = vx.3 + ...
1143 The vectorization of S1 is explained in vectorizable_load.
1144 The vectorization of S2:
1145 To create the first vector-stmt out of the 4 copies - VSnew.0 -
1146 the function 'vect_get_vec_def_for_operand' is called to
1147 get the relevant vector-def for each operand of S2. For operand x it
1148 returns the vector-def 'vx.0'.
1150 To create the remaining copies of the vector-stmt (VSnew.j), this
1151 function is called to get the relevant vector-def for each operand. It is
1152 obtained from the respective VS1.j stmt, which is recorded in the
1153 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
1155 For example, to obtain the vector-def 'vx.1' in order to create the
1156 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
1157 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
1158 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
1159 and return its def ('vx.1').
1160 Overall, to create the above sequence this function will be called 3 times:
1161 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
1162 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
1163 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
1166 vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd)
1168 gimple vec_stmt_for_operand;
1169 stmt_vec_info def_stmt_info;
1171 /* Do nothing; can reuse same def. */
1172 if (dt == vect_external_def || dt == vect_constant_def )
1175 vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd);
1176 def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand);
1177 gcc_assert (def_stmt_info);
1178 vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info);
1179 gcc_assert (vec_stmt_for_operand);
1180 vec_oprnd = gimple_get_lhs (vec_stmt_for_operand);
1181 if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI)
1182 vec_oprnd = PHI_RESULT (vec_stmt_for_operand);
1184 vec_oprnd = gimple_get_lhs (vec_stmt_for_operand);
1189 /* Get vectorized definitions for the operands to create a copy of an original
1190 stmt. See vect_get_vec_def_for_stmt_copy () for details. */
1193 vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt,
1194 VEC(tree,heap) **vec_oprnds0,
1195 VEC(tree,heap) **vec_oprnds1)
1197 tree vec_oprnd = VEC_pop (tree, *vec_oprnds0);
1199 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd);
1200 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
1202 if (vec_oprnds1 && *vec_oprnds1)
1204 vec_oprnd = VEC_pop (tree, *vec_oprnds1);
1205 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd);
1206 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
1211 /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not
1215 vect_get_vec_defs (tree op0, tree op1, gimple stmt,
1216 VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1,
1220 vect_get_slp_defs (op0, op1, slp_node, vec_oprnds0, vec_oprnds1, -1);
1225 *vec_oprnds0 = VEC_alloc (tree, heap, 1);
1226 vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL);
1227 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
1231 *vec_oprnds1 = VEC_alloc (tree, heap, 1);
1232 vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL);
1233 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
1239 /* Function vect_finish_stmt_generation.
1241 Insert a new stmt. */
1244 vect_finish_stmt_generation (gimple stmt, gimple vec_stmt,
1245 gimple_stmt_iterator *gsi)
1247 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1248 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1249 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
1251 gcc_assert (gimple_code (stmt) != GIMPLE_LABEL);
1253 gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT);
1255 set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo,
1258 if (vect_print_dump_info (REPORT_DETAILS))
1260 fprintf (vect_dump, "add new stmt: ");
1261 print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM);
1264 gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi)));
1267 /* Checks if CALL can be vectorized in type VECTYPE. Returns
1268 a function declaration if the target has a vectorized version
1269 of the function, or NULL_TREE if the function cannot be vectorized. */
1272 vectorizable_function (gimple call, tree vectype_out, tree vectype_in)
1274 tree fndecl = gimple_call_fndecl (call);
1276 /* We only handle functions that do not read or clobber memory -- i.e.
1277 const or novops ones. */
1278 if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS)))
1282 || TREE_CODE (fndecl) != FUNCTION_DECL
1283 || !DECL_BUILT_IN (fndecl))
1286 return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out,
1290 /* Function vectorizable_call.
1292 Check if STMT performs a function call that can be vectorized.
1293 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1294 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1295 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1298 vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt)
1303 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
1304 stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info;
1305 tree vectype_out, vectype_in;
1308 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1309 tree fndecl, new_temp, def, rhs_type;
1311 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1312 gimple new_stmt = NULL;
1314 VEC(tree, heap) *vargs = NULL;
1315 enum { NARROW, NONE, WIDEN } modifier;
1318 /* FORNOW: unsupported in basic block SLP. */
1319 gcc_assert (loop_vinfo);
1321 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1324 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1327 /* FORNOW: SLP not supported. */
1328 if (STMT_SLP_TYPE (stmt_info))
1331 /* Is STMT a vectorizable call? */
1332 if (!is_gimple_call (stmt))
1335 if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
1338 if (stmt_could_throw_p (stmt))
1341 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1343 /* Process function arguments. */
1344 rhs_type = NULL_TREE;
1345 vectype_in = NULL_TREE;
1346 nargs = gimple_call_num_args (stmt);
1348 /* Bail out if the function has more than three arguments, we do not have
1349 interesting builtin functions to vectorize with more than two arguments
1350 except for fma. No arguments is also not good. */
1351 if (nargs == 0 || nargs > 3)
1354 for (i = 0; i < nargs; i++)
1358 op = gimple_call_arg (stmt, i);
1360 /* We can only handle calls with arguments of the same type. */
1362 && !types_compatible_p (rhs_type, TREE_TYPE (op)))
1364 if (vect_print_dump_info (REPORT_DETAILS))
1365 fprintf (vect_dump, "argument types differ.");
1369 rhs_type = TREE_TYPE (op);
1371 if (!vect_is_simple_use_1 (op, loop_vinfo, NULL,
1372 &def_stmt, &def, &dt[i], &opvectype))
1374 if (vect_print_dump_info (REPORT_DETAILS))
1375 fprintf (vect_dump, "use not simple.");
1380 vectype_in = opvectype;
1382 && opvectype != vectype_in)
1384 if (vect_print_dump_info (REPORT_DETAILS))
1385 fprintf (vect_dump, "argument vector types differ.");
1389 /* If all arguments are external or constant defs use a vector type with
1390 the same size as the output vector type. */
1392 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1394 gcc_assert (vectype_in);
1397 if (vect_print_dump_info (REPORT_DETAILS))
1399 fprintf (vect_dump, "no vectype for scalar type ");
1400 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1407 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1408 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1409 if (nunits_in == nunits_out / 2)
1411 else if (nunits_out == nunits_in)
1413 else if (nunits_out == nunits_in / 2)
1418 /* For now, we only vectorize functions if a target specific builtin
1419 is available. TODO -- in some cases, it might be profitable to
1420 insert the calls for pieces of the vector, in order to be able
1421 to vectorize other operations in the loop. */
1422 fndecl = vectorizable_function (stmt, vectype_out, vectype_in);
1423 if (fndecl == NULL_TREE)
1425 if (vect_print_dump_info (REPORT_DETAILS))
1426 fprintf (vect_dump, "function is not vectorizable.");
1431 gcc_assert (!gimple_vuse (stmt));
1433 if (modifier == NARROW)
1434 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1436 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1438 /* Sanity check: make sure that at least one copy of the vectorized stmt
1439 needs to be generated. */
1440 gcc_assert (ncopies >= 1);
1442 if (!vec_stmt) /* transformation not required. */
1444 STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
1445 if (vect_print_dump_info (REPORT_DETAILS))
1446 fprintf (vect_dump, "=== vectorizable_call ===");
1447 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
1453 if (vect_print_dump_info (REPORT_DETAILS))
1454 fprintf (vect_dump, "transform operation.");
1457 scalar_dest = gimple_call_lhs (stmt);
1458 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1460 prev_stmt_info = NULL;
1464 for (j = 0; j < ncopies; ++j)
1466 /* Build argument list for the vectorized call. */
1468 vargs = VEC_alloc (tree, heap, nargs);
1470 VEC_truncate (tree, vargs, 0);
1472 for (i = 0; i < nargs; i++)
1474 op = gimple_call_arg (stmt, i);
1477 = vect_get_vec_def_for_operand (op, stmt, NULL);
1480 vec_oprnd0 = gimple_call_arg (new_stmt, i);
1482 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1485 VEC_quick_push (tree, vargs, vec_oprnd0);
1488 new_stmt = gimple_build_call_vec (fndecl, vargs);
1489 new_temp = make_ssa_name (vec_dest, new_stmt);
1490 gimple_call_set_lhs (new_stmt, new_temp);
1492 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1493 mark_symbols_for_renaming (new_stmt);
1496 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1498 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1500 prev_stmt_info = vinfo_for_stmt (new_stmt);
1506 for (j = 0; j < ncopies; ++j)
1508 /* Build argument list for the vectorized call. */
1510 vargs = VEC_alloc (tree, heap, nargs * 2);
1512 VEC_truncate (tree, vargs, 0);
1514 for (i = 0; i < nargs; i++)
1516 op = gimple_call_arg (stmt, i);
1520 = vect_get_vec_def_for_operand (op, stmt, NULL);
1522 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1526 vec_oprnd1 = gimple_call_arg (new_stmt, 2*i);
1528 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1);
1530 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1533 VEC_quick_push (tree, vargs, vec_oprnd0);
1534 VEC_quick_push (tree, vargs, vec_oprnd1);
1537 new_stmt = gimple_build_call_vec (fndecl, vargs);
1538 new_temp = make_ssa_name (vec_dest, new_stmt);
1539 gimple_call_set_lhs (new_stmt, new_temp);
1541 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1542 mark_symbols_for_renaming (new_stmt);
1545 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1547 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1549 prev_stmt_info = vinfo_for_stmt (new_stmt);
1552 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1557 /* No current target implements this case. */
1561 VEC_free (tree, heap, vargs);
1563 /* Update the exception handling table with the vector stmt if necessary. */
1564 if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt))
1565 gimple_purge_dead_eh_edges (gimple_bb (stmt));
1567 /* The call in STMT might prevent it from being removed in dce.
1568 We however cannot remove it here, due to the way the ssa name
1569 it defines is mapped to the new definition. So just replace
1570 rhs of the statement with something harmless. */
1572 type = TREE_TYPE (scalar_dest);
1573 new_stmt = gimple_build_assign (gimple_call_lhs (stmt),
1574 fold_convert (type, integer_zero_node));
1575 set_vinfo_for_stmt (new_stmt, stmt_info);
1576 set_vinfo_for_stmt (stmt, NULL);
1577 STMT_VINFO_STMT (stmt_info) = new_stmt;
1578 gsi_replace (gsi, new_stmt, false);
1579 SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt;
1585 /* Function vect_gen_widened_results_half
1587 Create a vector stmt whose code, type, number of arguments, and result
1588 variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are
1589 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
1590 In the case that CODE is a CALL_EXPR, this means that a call to DECL
1591 needs to be created (DECL is a function-decl of a target-builtin).
1592 STMT is the original scalar stmt that we are vectorizing. */
1595 vect_gen_widened_results_half (enum tree_code code,
1597 tree vec_oprnd0, tree vec_oprnd1, int op_type,
1598 tree vec_dest, gimple_stmt_iterator *gsi,
1604 /* Generate half of the widened result: */
1605 if (code == CALL_EXPR)
1607 /* Target specific support */
1608 if (op_type == binary_op)
1609 new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1);
1611 new_stmt = gimple_build_call (decl, 1, vec_oprnd0);
1612 new_temp = make_ssa_name (vec_dest, new_stmt);
1613 gimple_call_set_lhs (new_stmt, new_temp);
1617 /* Generic support */
1618 gcc_assert (op_type == TREE_CODE_LENGTH (code));
1619 if (op_type != binary_op)
1621 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0,
1623 new_temp = make_ssa_name (vec_dest, new_stmt);
1624 gimple_assign_set_lhs (new_stmt, new_temp);
1626 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1632 /* Check if STMT performs a conversion operation, that can be vectorized.
1633 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1634 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1635 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1638 vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi,
1639 gimple *vec_stmt, slp_tree slp_node)
1644 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
1645 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1646 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1647 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
1648 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
1652 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1653 gimple new_stmt = NULL;
1654 stmt_vec_info prev_stmt_info;
1657 tree vectype_out, vectype_in;
1661 enum { NARROW, NONE, WIDEN } modifier;
1663 VEC(tree,heap) *vec_oprnds0 = NULL;
1665 VEC(tree,heap) *dummy = NULL;
1668 /* Is STMT a vectorizable conversion? */
1670 /* FORNOW: unsupported in basic block SLP. */
1671 gcc_assert (loop_vinfo);
1673 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1676 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1679 if (!is_gimple_assign (stmt))
1682 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
1685 code = gimple_assign_rhs_code (stmt);
1686 if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
1689 /* Check types of lhs and rhs. */
1690 scalar_dest = gimple_assign_lhs (stmt);
1691 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1693 op0 = gimple_assign_rhs1 (stmt);
1694 rhs_type = TREE_TYPE (op0);
1695 /* Check the operands of the operation. */
1696 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
1697 &def_stmt, &def, &dt[0], &vectype_in))
1699 if (vect_print_dump_info (REPORT_DETAILS))
1700 fprintf (vect_dump, "use not simple.");
1703 /* If op0 is an external or constant defs use a vector type of
1704 the same size as the output vector type. */
1706 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1708 gcc_assert (vectype_in);
1711 if (vect_print_dump_info (REPORT_DETAILS))
1713 fprintf (vect_dump, "no vectype for scalar type ");
1714 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1721 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1722 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1723 if (nunits_in == nunits_out / 2)
1725 else if (nunits_out == nunits_in)
1727 else if (nunits_out == nunits_in / 2)
1732 if (modifier == NARROW)
1733 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1735 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1737 /* Multiple types in SLP are handled by creating the appropriate number of
1738 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
1743 /* Sanity check: make sure that at least one copy of the vectorized stmt
1744 needs to be generated. */
1745 gcc_assert (ncopies >= 1);
1747 /* Supportable by target? */
1748 if ((modifier == NONE
1749 && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in))
1750 || (modifier == WIDEN
1751 && !supportable_widening_operation (code, stmt,
1752 vectype_out, vectype_in,
1755 &dummy_int, &dummy))
1756 || (modifier == NARROW
1757 && !supportable_narrowing_operation (code, vectype_out, vectype_in,
1758 &code1, &dummy_int, &dummy)))
1760 if (vect_print_dump_info (REPORT_DETAILS))
1761 fprintf (vect_dump, "conversion not supported by target.");
1765 if (modifier != NONE)
1767 /* FORNOW: SLP not supported. */
1768 if (STMT_SLP_TYPE (stmt_info))
1772 if (!vec_stmt) /* transformation not required. */
1774 STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type;
1779 if (vect_print_dump_info (REPORT_DETAILS))
1780 fprintf (vect_dump, "transform conversion.");
1783 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1785 if (modifier == NONE && !slp_node)
1786 vec_oprnds0 = VEC_alloc (tree, heap, 1);
1788 prev_stmt_info = NULL;
1792 for (j = 0; j < ncopies; j++)
1795 vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node);
1797 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL);
1800 targetm.vectorize.builtin_conversion (code,
1801 vectype_out, vectype_in);
1802 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
1804 /* Arguments are ready. create the new vector stmt. */
1805 new_stmt = gimple_build_call (builtin_decl, 1, vop0);
1806 new_temp = make_ssa_name (vec_dest, new_stmt);
1807 gimple_call_set_lhs (new_stmt, new_temp);
1808 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1810 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
1814 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1816 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1817 prev_stmt_info = vinfo_for_stmt (new_stmt);
1822 /* In case the vectorization factor (VF) is bigger than the number
1823 of elements that we can fit in a vectype (nunits), we have to
1824 generate more than one vector stmt - i.e - we need to "unroll"
1825 the vector stmt by a factor VF/nunits. */
1826 for (j = 0; j < ncopies; j++)
1829 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1831 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1833 /* Generate first half of the widened result: */
1835 = vect_gen_widened_results_half (code1, decl1,
1836 vec_oprnd0, vec_oprnd1,
1837 unary_op, vec_dest, gsi, stmt);
1839 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1841 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1842 prev_stmt_info = vinfo_for_stmt (new_stmt);
1844 /* Generate second half of the widened result: */
1846 = vect_gen_widened_results_half (code2, decl2,
1847 vec_oprnd0, vec_oprnd1,
1848 unary_op, vec_dest, gsi, stmt);
1849 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1850 prev_stmt_info = vinfo_for_stmt (new_stmt);
1855 /* In case the vectorization factor (VF) is bigger than the number
1856 of elements that we can fit in a vectype (nunits), we have to
1857 generate more than one vector stmt - i.e - we need to "unroll"
1858 the vector stmt by a factor VF/nunits. */
1859 for (j = 0; j < ncopies; j++)
1864 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1865 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1869 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
1870 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1873 /* Arguments are ready. Create the new vector stmt. */
1874 new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0,
1876 new_temp = make_ssa_name (vec_dest, new_stmt);
1877 gimple_assign_set_lhs (new_stmt, new_temp);
1878 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1881 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1883 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1885 prev_stmt_info = vinfo_for_stmt (new_stmt);
1888 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1892 VEC_free (tree, heap, vec_oprnds0);
1898 /* Function vectorizable_assignment.
1900 Check if STMT performs an assignment (copy) that can be vectorized.
1901 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1902 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1903 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1906 vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi,
1907 gimple *vec_stmt, slp_tree slp_node)
1912 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1913 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
1914 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1918 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1919 unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype);
1922 VEC(tree,heap) *vec_oprnds = NULL;
1924 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
1925 gimple new_stmt = NULL;
1926 stmt_vec_info prev_stmt_info = NULL;
1927 enum tree_code code;
1930 /* Multiple types in SLP are handled by creating the appropriate number of
1931 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
1936 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
1938 gcc_assert (ncopies >= 1);
1940 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
1943 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1946 /* Is vectorizable assignment? */
1947 if (!is_gimple_assign (stmt))
1950 scalar_dest = gimple_assign_lhs (stmt);
1951 if (TREE_CODE (scalar_dest) != SSA_NAME)
1954 code = gimple_assign_rhs_code (stmt);
1955 if (gimple_assign_single_p (stmt)
1956 || code == PAREN_EXPR
1957 || CONVERT_EXPR_CODE_P (code))
1958 op = gimple_assign_rhs1 (stmt);
1962 if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo,
1963 &def_stmt, &def, &dt[0], &vectype_in))
1965 if (vect_print_dump_info (REPORT_DETAILS))
1966 fprintf (vect_dump, "use not simple.");
1970 /* We can handle NOP_EXPR conversions that do not change the number
1971 of elements or the vector size. */
1972 if (CONVERT_EXPR_CODE_P (code)
1974 || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits
1975 || (GET_MODE_SIZE (TYPE_MODE (vectype))
1976 != GET_MODE_SIZE (TYPE_MODE (vectype_in)))))
1979 if (!vec_stmt) /* transformation not required. */
1981 STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
1982 if (vect_print_dump_info (REPORT_DETAILS))
1983 fprintf (vect_dump, "=== vectorizable_assignment ===");
1984 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
1989 if (vect_print_dump_info (REPORT_DETAILS))
1990 fprintf (vect_dump, "transform assignment.");
1993 vec_dest = vect_create_destination_var (scalar_dest, vectype);
1996 for (j = 0; j < ncopies; j++)
2000 vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node);
2002 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL);
2004 /* Arguments are ready. create the new vector stmt. */
2005 FOR_EACH_VEC_ELT (tree, vec_oprnds, i, vop)
2007 if (CONVERT_EXPR_CODE_P (code))
2008 vop = build1 (VIEW_CONVERT_EXPR, vectype, vop);
2009 new_stmt = gimple_build_assign (vec_dest, vop);
2010 new_temp = make_ssa_name (vec_dest, new_stmt);
2011 gimple_assign_set_lhs (new_stmt, new_temp);
2012 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2014 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2021 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2023 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2025 prev_stmt_info = vinfo_for_stmt (new_stmt);
2028 VEC_free (tree, heap, vec_oprnds);
2033 /* Function vectorizable_shift.
2035 Check if STMT performs a shift operation that can be vectorized.
2036 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2037 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2038 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2041 vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi,
2042 gimple *vec_stmt, slp_tree slp_node)
2046 tree op0, op1 = NULL;
2047 tree vec_oprnd1 = NULL_TREE;
2048 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2050 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2051 enum tree_code code;
2052 enum machine_mode vec_mode;
2056 enum machine_mode optab_op2_mode;
2059 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2060 gimple new_stmt = NULL;
2061 stmt_vec_info prev_stmt_info;
2067 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2070 bool scalar_shift_arg = false;
2071 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
2074 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
2077 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2080 /* Is STMT a vectorizable binary/unary operation? */
2081 if (!is_gimple_assign (stmt))
2084 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2087 code = gimple_assign_rhs_code (stmt);
2089 if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR
2090 || code == RROTATE_EXPR))
2093 scalar_dest = gimple_assign_lhs (stmt);
2094 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2096 op0 = gimple_assign_rhs1 (stmt);
2097 if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo,
2098 &def_stmt, &def, &dt[0], &vectype))
2100 if (vect_print_dump_info (REPORT_DETAILS))
2101 fprintf (vect_dump, "use not simple.");
2104 /* If op0 is an external or constant def use a vector type with
2105 the same size as the output vector type. */
2107 vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2109 gcc_assert (vectype);
2112 if (vect_print_dump_info (REPORT_DETAILS))
2114 fprintf (vect_dump, "no vectype for scalar type ");
2115 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2121 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2122 nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
2123 if (nunits_out != nunits_in)
2126 op1 = gimple_assign_rhs2 (stmt);
2127 if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[1]))
2129 if (vect_print_dump_info (REPORT_DETAILS))
2130 fprintf (vect_dump, "use not simple.");
2135 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2139 /* Multiple types in SLP are handled by creating the appropriate number of
2140 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2145 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2147 gcc_assert (ncopies >= 1);
2149 /* Determine whether the shift amount is a vector, or scalar. If the
2150 shift/rotate amount is a vector, use the vector/vector shift optabs. */
2152 /* Vector shifted by vector. */
2153 if (dt[1] == vect_internal_def)
2155 optab = optab_for_tree_code (code, vectype, optab_vector);
2156 if (vect_print_dump_info (REPORT_DETAILS))
2157 fprintf (vect_dump, "vector/vector shift/rotate found.");
2159 /* See if the machine has a vector shifted by scalar insn and if not
2160 then see if it has a vector shifted by vector insn. */
2161 else if (dt[1] == vect_constant_def || dt[1] == vect_external_def)
2163 optab = optab_for_tree_code (code, vectype, optab_scalar);
2165 && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing)
2167 scalar_shift_arg = true;
2168 if (vect_print_dump_info (REPORT_DETAILS))
2169 fprintf (vect_dump, "vector/scalar shift/rotate found.");
2173 optab = optab_for_tree_code (code, vectype, optab_vector);
2175 && (optab_handler (optab, TYPE_MODE (vectype))
2176 != CODE_FOR_nothing))
2178 if (vect_print_dump_info (REPORT_DETAILS))
2179 fprintf (vect_dump, "vector/vector shift/rotate found.");
2181 /* Unlike the other binary operators, shifts/rotates have
2182 the rhs being int, instead of the same type as the lhs,
2183 so make sure the scalar is the right type if we are
2184 dealing with vectors of short/char. */
2185 if (dt[1] == vect_constant_def)
2186 op1 = fold_convert (TREE_TYPE (vectype), op1);
2192 if (vect_print_dump_info (REPORT_DETAILS))
2193 fprintf (vect_dump, "operand mode requires invariant argument.");
2197 /* Supportable by target? */
2200 if (vect_print_dump_info (REPORT_DETAILS))
2201 fprintf (vect_dump, "no optab.");
2204 vec_mode = TYPE_MODE (vectype);
2205 icode = (int) optab_handler (optab, vec_mode);
2206 if (icode == CODE_FOR_nothing)
2208 if (vect_print_dump_info (REPORT_DETAILS))
2209 fprintf (vect_dump, "op not supported by target.");
2210 /* Check only during analysis. */
2211 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2212 || (vf < vect_min_worthwhile_factor (code)
2215 if (vect_print_dump_info (REPORT_DETAILS))
2216 fprintf (vect_dump, "proceeding using word mode.");
2219 /* Worthwhile without SIMD support? Check only during analysis. */
2220 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2221 && vf < vect_min_worthwhile_factor (code)
2224 if (vect_print_dump_info (REPORT_DETAILS))
2225 fprintf (vect_dump, "not worthwhile without SIMD support.");
2229 if (!vec_stmt) /* transformation not required. */
2231 STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type;
2232 if (vect_print_dump_info (REPORT_DETAILS))
2233 fprintf (vect_dump, "=== vectorizable_shift ===");
2234 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2240 if (vect_print_dump_info (REPORT_DETAILS))
2241 fprintf (vect_dump, "transform binary/unary operation.");
2244 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2246 /* Allocate VECs for vector operands. In case of SLP, vector operands are
2247 created in the previous stages of the recursion, so no allocation is
2248 needed, except for the case of shift with scalar shift argument. In that
2249 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
2250 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
2251 In case of loop-based vectorization we allocate VECs of size 1. We
2252 allocate VEC_OPRNDS1 only in case of binary operation. */
2255 vec_oprnds0 = VEC_alloc (tree, heap, 1);
2256 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2258 else if (scalar_shift_arg)
2259 vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size);
2261 prev_stmt_info = NULL;
2262 for (j = 0; j < ncopies; j++)
2267 if (scalar_shift_arg)
2269 /* Vector shl and shr insn patterns can be defined with scalar
2270 operand 2 (shift operand). In this case, use constant or loop
2271 invariant op1 directly, without extending it to vector mode
2273 optab_op2_mode = insn_data[icode].operand[2].mode;
2274 if (!VECTOR_MODE_P (optab_op2_mode))
2276 if (vect_print_dump_info (REPORT_DETAILS))
2277 fprintf (vect_dump, "operand 1 using scalar mode.");
2279 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2282 /* Store vec_oprnd1 for every vector stmt to be created
2283 for SLP_NODE. We check during the analysis that all
2284 the shift arguments are the same.
2285 TODO: Allow different constants for different vector
2286 stmts generated for an SLP instance. */
2287 for (k = 0; k < slp_node->vec_stmts_size - 1; k++)
2288 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2293 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
2294 (a special case for certain kind of vector shifts); otherwise,
2295 operand 1 should be of a vector type (the usual case). */
2297 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
2300 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
2304 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
2306 /* Arguments are ready. Create the new vector stmt. */
2307 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
2309 vop1 = VEC_index (tree, vec_oprnds1, i);
2310 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2311 new_temp = make_ssa_name (vec_dest, new_stmt);
2312 gimple_assign_set_lhs (new_stmt, new_temp);
2313 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2315 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2322 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2324 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2325 prev_stmt_info = vinfo_for_stmt (new_stmt);
2328 VEC_free (tree, heap, vec_oprnds0);
2329 VEC_free (tree, heap, vec_oprnds1);
2335 /* Function vectorizable_operation.
2337 Check if STMT performs a binary or unary operation that can be vectorized.
2338 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2339 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2340 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2343 vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi,
2344 gimple *vec_stmt, slp_tree slp_node)
2348 tree op0, op1 = NULL;
2349 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2351 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2352 enum tree_code code;
2353 enum machine_mode vec_mode;
2360 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2361 gimple new_stmt = NULL;
2362 stmt_vec_info prev_stmt_info;
2368 VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2370 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
2373 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
2376 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2379 /* Is STMT a vectorizable binary/unary operation? */
2380 if (!is_gimple_assign (stmt))
2383 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2386 code = gimple_assign_rhs_code (stmt);
2388 /* For pointer addition, we should use the normal plus for
2389 the vector addition. */
2390 if (code == POINTER_PLUS_EXPR)
2393 /* Support only unary or binary operations. */
2394 op_type = TREE_CODE_LENGTH (code);
2395 if (op_type != unary_op && op_type != binary_op)
2397 if (vect_print_dump_info (REPORT_DETAILS))
2398 fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type);
2402 scalar_dest = gimple_assign_lhs (stmt);
2403 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2405 op0 = gimple_assign_rhs1 (stmt);
2406 if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo,
2407 &def_stmt, &def, &dt[0], &vectype))
2409 if (vect_print_dump_info (REPORT_DETAILS))
2410 fprintf (vect_dump, "use not simple.");
2413 /* If op0 is an external or constant def use a vector type with
2414 the same size as the output vector type. */
2416 vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2418 gcc_assert (vectype);
2421 if (vect_print_dump_info (REPORT_DETAILS))
2423 fprintf (vect_dump, "no vectype for scalar type ");
2424 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2430 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2431 nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
2432 if (nunits_out != nunits_in)
2435 if (op_type == binary_op)
2437 op1 = gimple_assign_rhs2 (stmt);
2438 if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def,
2441 if (vect_print_dump_info (REPORT_DETAILS))
2442 fprintf (vect_dump, "use not simple.");
2448 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2452 /* Multiple types in SLP are handled by creating the appropriate number of
2453 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2458 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2460 gcc_assert (ncopies >= 1);
2462 /* Shifts are handled in vectorizable_shift (). */
2463 if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR
2464 || code == RROTATE_EXPR)
2467 optab = optab_for_tree_code (code, vectype, optab_default);
2469 /* Supportable by target? */
2472 if (vect_print_dump_info (REPORT_DETAILS))
2473 fprintf (vect_dump, "no optab.");
2476 vec_mode = TYPE_MODE (vectype);
2477 icode = (int) optab_handler (optab, vec_mode);
2478 if (icode == CODE_FOR_nothing)
2480 if (vect_print_dump_info (REPORT_DETAILS))
2481 fprintf (vect_dump, "op not supported by target.");
2482 /* Check only during analysis. */
2483 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2484 || (vf < vect_min_worthwhile_factor (code)
2487 if (vect_print_dump_info (REPORT_DETAILS))
2488 fprintf (vect_dump, "proceeding using word mode.");
2491 /* Worthwhile without SIMD support? Check only during analysis. */
2492 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2493 && vf < vect_min_worthwhile_factor (code)
2496 if (vect_print_dump_info (REPORT_DETAILS))
2497 fprintf (vect_dump, "not worthwhile without SIMD support.");
2501 if (!vec_stmt) /* transformation not required. */
2503 STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
2504 if (vect_print_dump_info (REPORT_DETAILS))
2505 fprintf (vect_dump, "=== vectorizable_operation ===");
2506 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2512 if (vect_print_dump_info (REPORT_DETAILS))
2513 fprintf (vect_dump, "transform binary/unary operation.");
2516 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2518 /* Allocate VECs for vector operands. In case of SLP, vector operands are
2519 created in the previous stages of the recursion, so no allocation is
2520 needed, except for the case of shift with scalar shift argument. In that
2521 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
2522 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
2523 In case of loop-based vectorization we allocate VECs of size 1. We
2524 allocate VEC_OPRNDS1 only in case of binary operation. */
2527 vec_oprnds0 = VEC_alloc (tree, heap, 1);
2528 if (op_type == binary_op)
2529 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2532 /* In case the vectorization factor (VF) is bigger than the number
2533 of elements that we can fit in a vectype (nunits), we have to generate
2534 more than one vector stmt - i.e - we need to "unroll" the
2535 vector stmt by a factor VF/nunits. In doing so, we record a pointer
2536 from one copy of the vector stmt to the next, in the field
2537 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
2538 stages to find the correct vector defs to be used when vectorizing
2539 stmts that use the defs of the current stmt. The example below
2540 illustrates the vectorization process when VF=16 and nunits=4 (i.e.,
2541 we need to create 4 vectorized stmts):
2543 before vectorization:
2544 RELATED_STMT VEC_STMT
2548 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
2550 RELATED_STMT VEC_STMT
2551 VS1_0: vx0 = memref0 VS1_1 -
2552 VS1_1: vx1 = memref1 VS1_2 -
2553 VS1_2: vx2 = memref2 VS1_3 -
2554 VS1_3: vx3 = memref3 - -
2555 S1: x = load - VS1_0
2558 step2: vectorize stmt S2 (done here):
2559 To vectorize stmt S2 we first need to find the relevant vector
2560 def for the first operand 'x'. This is, as usual, obtained from
2561 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
2562 that defines 'x' (S1). This way we find the stmt VS1_0, and the
2563 relevant vector def 'vx0'. Having found 'vx0' we can generate
2564 the vector stmt VS2_0, and as usual, record it in the
2565 STMT_VINFO_VEC_STMT of stmt S2.
2566 When creating the second copy (VS2_1), we obtain the relevant vector
2567 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
2568 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
2569 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
2570 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
2571 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
2572 chain of stmts and pointers:
2573 RELATED_STMT VEC_STMT
2574 VS1_0: vx0 = memref0 VS1_1 -
2575 VS1_1: vx1 = memref1 VS1_2 -
2576 VS1_2: vx2 = memref2 VS1_3 -
2577 VS1_3: vx3 = memref3 - -
2578 S1: x = load - VS1_0
2579 VS2_0: vz0 = vx0 + v1 VS2_1 -
2580 VS2_1: vz1 = vx1 + v1 VS2_2 -
2581 VS2_2: vz2 = vx2 + v1 VS2_3 -
2582 VS2_3: vz3 = vx3 + v1 - -
2583 S2: z = x + 1 - VS2_0 */
2585 prev_stmt_info = NULL;
2586 for (j = 0; j < ncopies; j++)
2591 if (op_type == binary_op)
2592 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
2595 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
2599 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
2601 /* Arguments are ready. Create the new vector stmt. */
2602 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
2604 vop1 = ((op_type == binary_op)
2605 ? VEC_index (tree, vec_oprnds1, i) : NULL);
2606 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2607 new_temp = make_ssa_name (vec_dest, new_stmt);
2608 gimple_assign_set_lhs (new_stmt, new_temp);
2609 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2611 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2618 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2620 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2621 prev_stmt_info = vinfo_for_stmt (new_stmt);
2624 VEC_free (tree, heap, vec_oprnds0);
2626 VEC_free (tree, heap, vec_oprnds1);
2632 /* Get vectorized definitions for loop-based vectorization. For the first
2633 operand we call vect_get_vec_def_for_operand() (with OPRND containing
2634 scalar operand), and for the rest we get a copy with
2635 vect_get_vec_def_for_stmt_copy() using the previous vector definition
2636 (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details.
2637 The vectors are collected into VEC_OPRNDS. */
2640 vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt,
2641 VEC (tree, heap) **vec_oprnds, int multi_step_cvt)
2645 /* Get first vector operand. */
2646 /* All the vector operands except the very first one (that is scalar oprnd)
2648 if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE)
2649 vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL);
2651 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd);
2653 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2655 /* Get second vector operand. */
2656 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd);
2657 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2661 /* For conversion in multiple steps, continue to get operands
2664 vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1);
2668 /* Create vectorized demotion statements for vector operands from VEC_OPRNDS.
2669 For multi-step conversions store the resulting vectors and call the function
2673 vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds,
2674 int multi_step_cvt, gimple stmt,
2675 VEC (tree, heap) *vec_dsts,
2676 gimple_stmt_iterator *gsi,
2677 slp_tree slp_node, enum tree_code code,
2678 stmt_vec_info *prev_stmt_info)
2681 tree vop0, vop1, new_tmp, vec_dest;
2683 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2685 vec_dest = VEC_pop (tree, vec_dsts);
2687 for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2)
2689 /* Create demotion operation. */
2690 vop0 = VEC_index (tree, *vec_oprnds, i);
2691 vop1 = VEC_index (tree, *vec_oprnds, i + 1);
2692 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2693 new_tmp = make_ssa_name (vec_dest, new_stmt);
2694 gimple_assign_set_lhs (new_stmt, new_tmp);
2695 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2698 /* Store the resulting vector for next recursive call. */
2699 VEC_replace (tree, *vec_oprnds, i/2, new_tmp);
2702 /* This is the last step of the conversion sequence. Store the
2703 vectors in SLP_NODE or in vector info of the scalar statement
2704 (or in STMT_VINFO_RELATED_STMT chain). */
2706 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2709 if (!*prev_stmt_info)
2710 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
2712 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt;
2714 *prev_stmt_info = vinfo_for_stmt (new_stmt);
2719 /* For multi-step demotion operations we first generate demotion operations
2720 from the source type to the intermediate types, and then combine the
2721 results (stored in VEC_OPRNDS) in demotion operation to the destination
2725 /* At each level of recursion we have have of the operands we had at the
2727 VEC_truncate (tree, *vec_oprnds, (i+1)/2);
2728 vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1,
2729 stmt, vec_dsts, gsi, slp_node,
2730 code, prev_stmt_info);
2735 /* Function vectorizable_type_demotion
2737 Check if STMT performs a binary or unary operation that involves
2738 type demotion, and if it can be vectorized.
2739 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2740 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2741 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2744 vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi,
2745 gimple *vec_stmt, slp_tree slp_node)
2750 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2751 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2752 enum tree_code code, code1 = ERROR_MARK;
2755 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2756 stmt_vec_info prev_stmt_info;
2763 int multi_step_cvt = 0;
2764 VEC (tree, heap) *vec_oprnds0 = NULL;
2765 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
2766 tree last_oprnd, intermediate_type;
2768 /* FORNOW: not supported by basic block SLP vectorization. */
2769 gcc_assert (loop_vinfo);
2771 if (!STMT_VINFO_RELEVANT_P (stmt_info))
2774 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2777 /* Is STMT a vectorizable type-demotion operation? */
2778 if (!is_gimple_assign (stmt))
2781 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2784 code = gimple_assign_rhs_code (stmt);
2785 if (!CONVERT_EXPR_CODE_P (code))
2788 scalar_dest = gimple_assign_lhs (stmt);
2789 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2791 /* Check the operands of the operation. */
2792 op0 = gimple_assign_rhs1 (stmt);
2793 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
2794 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
2795 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
2796 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
2797 && CONVERT_EXPR_CODE_P (code))))
2799 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
2800 &def_stmt, &def, &dt[0], &vectype_in))
2802 if (vect_print_dump_info (REPORT_DETAILS))
2803 fprintf (vect_dump, "use not simple.");
2806 /* If op0 is an external def use a vector type with the
2807 same size as the output vector type if possible. */
2809 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2811 gcc_assert (vectype_in);
2814 if (vect_print_dump_info (REPORT_DETAILS))
2816 fprintf (vect_dump, "no vectype for scalar type ");
2817 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2823 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
2824 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2825 if (nunits_in >= nunits_out)
2828 /* Multiple types in SLP are handled by creating the appropriate number of
2829 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2834 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
2835 gcc_assert (ncopies >= 1);
2837 /* Supportable by target? */
2838 if (!supportable_narrowing_operation (code, vectype_out, vectype_in,
2839 &code1, &multi_step_cvt, &interm_types))
2842 if (!vec_stmt) /* transformation not required. */
2844 STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type;
2845 if (vect_print_dump_info (REPORT_DETAILS))
2846 fprintf (vect_dump, "=== vectorizable_demotion ===");
2847 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2852 if (vect_print_dump_info (REPORT_DETAILS))
2853 fprintf (vect_dump, "transform type demotion operation. ncopies = %d.",
2856 /* In case of multi-step demotion, we first generate demotion operations to
2857 the intermediate types, and then from that types to the final one.
2858 We create vector destinations for the intermediate type (TYPES) received
2859 from supportable_narrowing_operation, and store them in the correct order
2860 for future use in vect_create_vectorized_demotion_stmts(). */
2862 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
2864 vec_dsts = VEC_alloc (tree, heap, 1);
2866 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
2867 VEC_quick_push (tree, vec_dsts, vec_dest);
2871 for (i = VEC_length (tree, interm_types) - 1;
2872 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
2874 vec_dest = vect_create_destination_var (scalar_dest,
2876 VEC_quick_push (tree, vec_dsts, vec_dest);
2880 /* In case the vectorization factor (VF) is bigger than the number
2881 of elements that we can fit in a vectype (nunits), we have to generate
2882 more than one vector stmt - i.e - we need to "unroll" the
2883 vector stmt by a factor VF/nunits. */
2885 prev_stmt_info = NULL;
2886 for (j = 0; j < ncopies; j++)
2890 vect_get_slp_defs (op0, NULL_TREE, slp_node, &vec_oprnds0, NULL, -1);
2893 VEC_free (tree, heap, vec_oprnds0);
2894 vec_oprnds0 = VEC_alloc (tree, heap,
2895 (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2));
2896 vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0,
2897 vect_pow2 (multi_step_cvt) - 1);
2900 /* Arguments are ready. Create the new vector stmts. */
2901 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
2902 vect_create_vectorized_demotion_stmts (&vec_oprnds0,
2903 multi_step_cvt, stmt, tmp_vec_dsts,
2904 gsi, slp_node, code1,
2908 VEC_free (tree, heap, vec_oprnds0);
2909 VEC_free (tree, heap, vec_dsts);
2910 VEC_free (tree, heap, tmp_vec_dsts);
2911 VEC_free (tree, heap, interm_types);
2913 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
2918 /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0
2919 and VEC_OPRNDS1 (for binary operations). For multi-step conversions store
2920 the resulting vectors and call the function recursively. */
2923 vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0,
2924 VEC (tree, heap) **vec_oprnds1,
2925 int multi_step_cvt, gimple stmt,
2926 VEC (tree, heap) *vec_dsts,
2927 gimple_stmt_iterator *gsi,
2928 slp_tree slp_node, enum tree_code code1,
2929 enum tree_code code2, tree decl1,
2930 tree decl2, int op_type,
2931 stmt_vec_info *prev_stmt_info)
2934 tree vop0, vop1, new_tmp1, new_tmp2, vec_dest;
2935 gimple new_stmt1, new_stmt2;
2936 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2937 VEC (tree, heap) *vec_tmp;
2939 vec_dest = VEC_pop (tree, vec_dsts);
2940 vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2);
2942 FOR_EACH_VEC_ELT (tree, *vec_oprnds0, i, vop0)
2944 if (op_type == binary_op)
2945 vop1 = VEC_index (tree, *vec_oprnds1, i);
2949 /* Generate the two halves of promotion operation. */
2950 new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1,
2951 op_type, vec_dest, gsi, stmt);
2952 new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1,
2953 op_type, vec_dest, gsi, stmt);
2954 if (is_gimple_call (new_stmt1))
2956 new_tmp1 = gimple_call_lhs (new_stmt1);
2957 new_tmp2 = gimple_call_lhs (new_stmt2);
2961 new_tmp1 = gimple_assign_lhs (new_stmt1);
2962 new_tmp2 = gimple_assign_lhs (new_stmt2);
2967 /* Store the results for the recursive call. */
2968 VEC_quick_push (tree, vec_tmp, new_tmp1);
2969 VEC_quick_push (tree, vec_tmp, new_tmp2);
2973 /* Last step of promotion sequience - store the results. */
2976 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1);
2977 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2);
2981 if (!*prev_stmt_info)
2982 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1;
2984 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1;
2986 *prev_stmt_info = vinfo_for_stmt (new_stmt1);
2987 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2;
2988 *prev_stmt_info = vinfo_for_stmt (new_stmt2);
2995 /* For multi-step promotion operation we first generate we call the
2996 function recurcively for every stage. We start from the input type,
2997 create promotion operations to the intermediate types, and then
2998 create promotions to the output type. */
2999 *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp);
3000 vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1,
3001 multi_step_cvt - 1, stmt,
3002 vec_dsts, gsi, slp_node, code1,
3003 code2, decl2, decl2, op_type,
3007 VEC_free (tree, heap, vec_tmp);
3011 /* Function vectorizable_type_promotion
3013 Check if STMT performs a binary or unary operation that involves
3014 type promotion, and if it can be vectorized.
3015 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3016 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3017 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3020 vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi,
3021 gimple *vec_stmt, slp_tree slp_node)
3025 tree op0, op1 = NULL;
3026 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
3027 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3028 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3029 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
3030 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
3034 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3035 stmt_vec_info prev_stmt_info;
3042 tree intermediate_type = NULL_TREE;
3043 int multi_step_cvt = 0;
3044 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
3045 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
3047 /* FORNOW: not supported by basic block SLP vectorization. */
3048 gcc_assert (loop_vinfo);
3050 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3053 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3056 /* Is STMT a vectorizable type-promotion operation? */
3057 if (!is_gimple_assign (stmt))
3060 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
3063 code = gimple_assign_rhs_code (stmt);
3064 if (!CONVERT_EXPR_CODE_P (code)
3065 && code != WIDEN_MULT_EXPR)
3068 scalar_dest = gimple_assign_lhs (stmt);
3069 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
3071 /* Check the operands of the operation. */
3072 op0 = gimple_assign_rhs1 (stmt);
3073 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
3074 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
3075 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
3076 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
3077 && CONVERT_EXPR_CODE_P (code))))
3079 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
3080 &def_stmt, &def, &dt[0], &vectype_in))
3082 if (vect_print_dump_info (REPORT_DETAILS))
3083 fprintf (vect_dump, "use not simple.");
3086 /* If op0 is an external or constant def use a vector type with
3087 the same size as the output vector type. */
3089 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
3091 gcc_assert (vectype_in);
3094 if (vect_print_dump_info (REPORT_DETAILS))
3096 fprintf (vect_dump, "no vectype for scalar type ");
3097 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
3103 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
3104 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3105 if (nunits_in <= nunits_out)
3108 /* Multiple types in SLP are handled by creating the appropriate number of
3109 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3114 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3116 gcc_assert (ncopies >= 1);
3118 op_type = TREE_CODE_LENGTH (code);
3119 if (op_type == binary_op)
3121 op1 = gimple_assign_rhs2 (stmt);
3122 if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1]))
3124 if (vect_print_dump_info (REPORT_DETAILS))
3125 fprintf (vect_dump, "use not simple.");
3130 /* Supportable by target? */
3131 if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in,
3132 &decl1, &decl2, &code1, &code2,
3133 &multi_step_cvt, &interm_types))
3136 /* Binary widening operation can only be supported directly by the
3138 gcc_assert (!(multi_step_cvt && op_type == binary_op));
3140 if (!vec_stmt) /* transformation not required. */
3142 STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type;
3143 if (vect_print_dump_info (REPORT_DETAILS))
3144 fprintf (vect_dump, "=== vectorizable_promotion ===");
3145 vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL);
3151 if (vect_print_dump_info (REPORT_DETAILS))
3152 fprintf (vect_dump, "transform type promotion operation. ncopies = %d.",
3156 /* In case of multi-step promotion, we first generate promotion operations
3157 to the intermediate types, and then from that types to the final one.
3158 We store vector destination in VEC_DSTS in the correct order for
3159 recursive creation of promotion operations in
3160 vect_create_vectorized_promotion_stmts(). Vector destinations are created
3161 according to TYPES recieved from supportable_widening_operation(). */
3163 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
3165 vec_dsts = VEC_alloc (tree, heap, 1);
3167 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
3168 VEC_quick_push (tree, vec_dsts, vec_dest);
3172 for (i = VEC_length (tree, interm_types) - 1;
3173 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
3175 vec_dest = vect_create_destination_var (scalar_dest,
3177 VEC_quick_push (tree, vec_dsts, vec_dest);
3183 vec_oprnds0 = VEC_alloc (tree, heap,
3184 (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1));
3185 if (op_type == binary_op)
3186 vec_oprnds1 = VEC_alloc (tree, heap, 1);
3189 /* In case the vectorization factor (VF) is bigger than the number
3190 of elements that we can fit in a vectype (nunits), we have to generate
3191 more than one vector stmt - i.e - we need to "unroll" the
3192 vector stmt by a factor VF/nunits. */
3194 prev_stmt_info = NULL;
3195 for (j = 0; j < ncopies; j++)
3201 vect_get_slp_defs (op0, op1, slp_node, &vec_oprnds0,
3205 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
3206 VEC_quick_push (tree, vec_oprnds0, vec_oprnd0);
3207 if (op_type == binary_op)
3209 vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
3210 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
3216 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3217 VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0);
3218 if (op_type == binary_op)
3220 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
3221 VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1);
3225 /* Arguments are ready. Create the new vector stmts. */
3226 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
3227 vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1,
3228 multi_step_cvt, stmt,
3230 gsi, slp_node, code1, code2,
3231 decl1, decl2, op_type,
3235 VEC_free (tree, heap, vec_dsts);
3236 VEC_free (tree, heap, tmp_vec_dsts);
3237 VEC_free (tree, heap, interm_types);
3238 VEC_free (tree, heap, vec_oprnds0);
3239 VEC_free (tree, heap, vec_oprnds1);
3241 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3246 /* Function vectorizable_store.
3248 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
3250 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3251 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3252 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3255 vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3261 tree vec_oprnd = NULL_TREE;
3262 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3263 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL;
3264 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3265 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3266 struct loop *loop = NULL;
3267 enum machine_mode vec_mode;
3269 enum dr_alignment_support alignment_support_scheme;
3272 enum vect_def_type dt;
3273 stmt_vec_info prev_stmt_info = NULL;
3274 tree dataref_ptr = NULL_TREE;
3275 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3278 gimple next_stmt, first_stmt = NULL;
3279 bool strided_store = false;
3280 unsigned int group_size, i;
3281 VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL;
3283 VEC(tree,heap) *vec_oprnds = NULL;
3284 bool slp = (slp_node != NULL);
3285 unsigned int vec_num;
3286 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3289 loop = LOOP_VINFO_LOOP (loop_vinfo);
3291 /* Multiple types in SLP are handled by creating the appropriate number of
3292 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3297 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3299 gcc_assert (ncopies >= 1);
3301 /* FORNOW. This restriction should be relaxed. */
3302 if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3304 if (vect_print_dump_info (REPORT_DETAILS))
3305 fprintf (vect_dump, "multiple types in nested loop.");
3309 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3312 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3315 /* Is vectorizable store? */
3317 if (!is_gimple_assign (stmt))
3320 scalar_dest = gimple_assign_lhs (stmt);
3321 if (TREE_CODE (scalar_dest) != ARRAY_REF
3322 && TREE_CODE (scalar_dest) != INDIRECT_REF
3323 && TREE_CODE (scalar_dest) != COMPONENT_REF
3324 && TREE_CODE (scalar_dest) != IMAGPART_EXPR
3325 && TREE_CODE (scalar_dest) != REALPART_EXPR
3326 && TREE_CODE (scalar_dest) != MEM_REF)
3329 gcc_assert (gimple_assign_single_p (stmt));
3330 op = gimple_assign_rhs1 (stmt);
3331 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt))
3333 if (vect_print_dump_info (REPORT_DETAILS))
3334 fprintf (vect_dump, "use not simple.");
3338 /* The scalar rhs type needs to be trivially convertible to the vector
3339 component type. This should always be the case. */
3340 if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op)))
3342 if (vect_print_dump_info (REPORT_DETAILS))
3343 fprintf (vect_dump, "??? operands of different types");
3347 vec_mode = TYPE_MODE (vectype);
3348 /* FORNOW. In some cases can vectorize even if data-type not supported
3349 (e.g. - array initialization with 0). */
3350 if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing)
3353 if (!STMT_VINFO_DATA_REF (stmt_info))
3356 if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0)
3358 if (vect_print_dump_info (REPORT_DETAILS))
3359 fprintf (vect_dump, "negative step for store.");
3363 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3365 strided_store = true;
3366 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3367 if (!vect_strided_store_supported (vectype)
3368 && !PURE_SLP_STMT (stmt_info) && !slp)
3371 if (first_stmt == stmt)
3373 /* STMT is the leader of the group. Check the operands of all the
3374 stmts of the group. */
3375 next_stmt = DR_GROUP_NEXT_DR (stmt_info);
3378 gcc_assert (gimple_assign_single_p (next_stmt));
3379 op = gimple_assign_rhs1 (next_stmt);
3380 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt,
3383 if (vect_print_dump_info (REPORT_DETAILS))
3384 fprintf (vect_dump, "use not simple.");
3387 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3392 if (!vec_stmt) /* transformation not required. */
3394 STMT_VINFO_TYPE (stmt_info) = store_vec_info_type;
3395 vect_model_store_cost (stmt_info, ncopies, dt, NULL);
3403 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3404 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3406 DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++;
3409 gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt));
3411 /* We vectorize all the stmts of the interleaving group when we
3412 reach the last stmt in the group. */
3413 if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))
3414 < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt))
3423 strided_store = false;
3424 /* VEC_NUM is the number of vect stmts to be created for this
3426 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3427 first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0);
3428 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3431 /* VEC_NUM is the number of vect stmts to be created for this
3433 vec_num = group_size;
3439 group_size = vec_num = 1;
3442 if (vect_print_dump_info (REPORT_DETAILS))
3443 fprintf (vect_dump, "transform store. ncopies = %d",ncopies);
3445 dr_chain = VEC_alloc (tree, heap, group_size);
3446 oprnds = VEC_alloc (tree, heap, group_size);
3448 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3449 gcc_assert (alignment_support_scheme);
3451 /* In case the vectorization factor (VF) is bigger than the number
3452 of elements that we can fit in a vectype (nunits), we have to generate
3453 more than one vector stmt - i.e - we need to "unroll" the
3454 vector stmt by a factor VF/nunits. For more details see documentation in
3455 vect_get_vec_def_for_copy_stmt. */
3457 /* In case of interleaving (non-unit strided access):
3464 We create vectorized stores starting from base address (the access of the
3465 first stmt in the chain (S2 in the above example), when the last store stmt
3466 of the chain (S4) is reached:
3469 VS2: &base + vec_size*1 = vx0
3470 VS3: &base + vec_size*2 = vx1
3471 VS4: &base + vec_size*3 = vx3
3473 Then permutation statements are generated:
3475 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
3476 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
3479 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3480 (the order of the data-refs in the output of vect_permute_store_chain
3481 corresponds to the order of scalar stmts in the interleaving chain - see
3482 the documentation of vect_permute_store_chain()).
3484 In case of both multiple types and interleaving, above vector stores and
3485 permutation stmts are created for every copy. The result vector stmts are
3486 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
3487 STMT_VINFO_RELATED_STMT for the next copies.
3490 prev_stmt_info = NULL;
3491 for (j = 0; j < ncopies; j++)
3500 /* Get vectorized arguments for SLP_NODE. */
3501 vect_get_slp_defs (NULL_TREE, NULL_TREE, slp_node, &vec_oprnds,
3504 vec_oprnd = VEC_index (tree, vec_oprnds, 0);
3508 /* For interleaved stores we collect vectorized defs for all the
3509 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
3510 used as an input to vect_permute_store_chain(), and OPRNDS as
3511 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
3513 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3514 OPRNDS are of size 1. */
3515 next_stmt = first_stmt;
3516 for (i = 0; i < group_size; i++)
3518 /* Since gaps are not supported for interleaved stores,
3519 GROUP_SIZE is the exact number of stmts in the chain.
3520 Therefore, NEXT_STMT can't be NULL_TREE. In case that
3521 there is no interleaving, GROUP_SIZE is 1, and only one
3522 iteration of the loop will be executed. */
3523 gcc_assert (next_stmt
3524 && gimple_assign_single_p (next_stmt));
3525 op = gimple_assign_rhs1 (next_stmt);
3527 vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt,
3529 VEC_quick_push(tree, dr_chain, vec_oprnd);
3530 VEC_quick_push(tree, oprnds, vec_oprnd);
3531 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3535 /* We should have catched mismatched types earlier. */
3536 gcc_assert (useless_type_conversion_p (vectype,
3537 TREE_TYPE (vec_oprnd)));
3538 dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE,
3539 &dummy, &ptr_incr, false,
3541 gcc_assert (bb_vinfo || !inv_p);
3545 /* For interleaved stores we created vectorized defs for all the
3546 defs stored in OPRNDS in the previous iteration (previous copy).
3547 DR_CHAIN is then used as an input to vect_permute_store_chain(),
3548 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
3550 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3551 OPRNDS are of size 1. */
3552 for (i = 0; i < group_size; i++)
3554 op = VEC_index (tree, oprnds, i);
3555 vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def,
3557 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op);
3558 VEC_replace(tree, dr_chain, i, vec_oprnd);
3559 VEC_replace(tree, oprnds, i, vec_oprnd);
3562 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
3567 result_chain = VEC_alloc (tree, heap, group_size);
3569 if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi,
3574 next_stmt = first_stmt;
3575 for (i = 0; i < vec_num; i++)
3577 struct ptr_info_def *pi;
3580 /* Bump the vector pointer. */
3581 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
3585 vec_oprnd = VEC_index (tree, vec_oprnds, i);
3586 else if (strided_store)
3587 /* For strided stores vectorized defs are interleaved in
3588 vect_permute_store_chain(). */
3589 vec_oprnd = VEC_index (tree, result_chain, i);
3591 data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr,
3592 build_int_cst (reference_alias_ptr_type
3593 (DR_REF (first_dr)), 0));
3594 pi = get_ptr_info (dataref_ptr);
3595 pi->align = TYPE_ALIGN_UNIT (vectype);
3596 if (aligned_access_p (first_dr))
3598 else if (DR_MISALIGNMENT (first_dr) == -1)
3600 TREE_TYPE (data_ref)
3601 = build_aligned_type (TREE_TYPE (data_ref),
3602 TYPE_ALIGN (TREE_TYPE (vectype)));
3603 pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype));
3608 TREE_TYPE (data_ref)
3609 = build_aligned_type (TREE_TYPE (data_ref),
3610 TYPE_ALIGN (TREE_TYPE (vectype)));
3611 pi->misalign = DR_MISALIGNMENT (first_dr);
3614 /* Arguments are ready. Create the new vector stmt. */
3615 new_stmt = gimple_build_assign (data_ref, vec_oprnd);
3616 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3617 mark_symbols_for_renaming (new_stmt);
3623 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3625 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3627 prev_stmt_info = vinfo_for_stmt (new_stmt);
3628 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3634 VEC_free (tree, heap, dr_chain);
3635 VEC_free (tree, heap, oprnds);
3637 VEC_free (tree, heap, result_chain);
3639 VEC_free (tree, heap, vec_oprnds);
3644 /* Given a vector type VECTYPE returns a builtin DECL to be used
3645 for vector permutation and stores a mask into *MASK that implements
3646 reversal of the vector elements. If that is impossible to do
3647 returns NULL (and *MASK is unchanged). */
3650 perm_mask_for_reverse (tree vectype, tree *mask)
3653 tree mask_element_type, mask_type;
3654 tree mask_vec = NULL;
3657 if (!targetm.vectorize.builtin_vec_perm)
3660 builtin_decl = targetm.vectorize.builtin_vec_perm (vectype,
3661 &mask_element_type);
3662 if (!builtin_decl || !mask_element_type)
3665 mask_type = get_vectype_for_scalar_type (mask_element_type);
3666 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3668 || TYPE_VECTOR_SUBPARTS (vectype) != TYPE_VECTOR_SUBPARTS (mask_type))
3671 for (i = 0; i < nunits; i++)
3672 mask_vec = tree_cons (NULL, build_int_cst (mask_element_type, i), mask_vec);
3673 mask_vec = build_vector (mask_type, mask_vec);
3675 if (!targetm.vectorize.builtin_vec_perm_ok (vectype, mask_vec))
3679 return builtin_decl;
3682 /* Given a vector variable X, that was generated for the scalar LHS of
3683 STMT, generate instructions to reverse the vector elements of X,
3684 insert them a *GSI and return the permuted vector variable. */
3687 reverse_vec_elements (tree x, gimple stmt, gimple_stmt_iterator *gsi)
3689 tree vectype = TREE_TYPE (x);
3690 tree mask_vec, builtin_decl;
3691 tree perm_dest, data_ref;
3694 builtin_decl = perm_mask_for_reverse (vectype, &mask_vec);
3696 perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype);
3698 /* Generate the permute statement. */
3699 perm_stmt = gimple_build_call (builtin_decl, 3, x, x, mask_vec);
3700 data_ref = make_ssa_name (perm_dest, perm_stmt);
3701 gimple_call_set_lhs (perm_stmt, data_ref);
3702 vect_finish_stmt_generation (stmt, perm_stmt, gsi);
3707 /* vectorizable_load.
3709 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
3711 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3712 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3713 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3716 vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3717 slp_tree slp_node, slp_instance slp_node_instance)
3720 tree vec_dest = NULL;
3721 tree data_ref = NULL;
3722 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3723 stmt_vec_info prev_stmt_info;
3724 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3725 struct loop *loop = NULL;
3726 struct loop *containing_loop = (gimple_bb (stmt))->loop_father;
3727 bool nested_in_vect_loop = false;
3728 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
3729 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3731 enum machine_mode mode;
3732 gimple new_stmt = NULL;
3734 enum dr_alignment_support alignment_support_scheme;
3735 tree dataref_ptr = NULL_TREE;
3737 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3739 int i, j, group_size;
3740 tree msq = NULL_TREE, lsq;
3741 tree offset = NULL_TREE;
3742 tree realignment_token = NULL_TREE;
3744 VEC(tree,heap) *dr_chain = NULL;
3745 bool strided_load = false;
3750 bool compute_in_loop = false;
3751 struct loop *at_loop;
3753 bool slp = (slp_node != NULL);
3754 bool slp_perm = false;
3755 enum tree_code code;
3756 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3761 loop = LOOP_VINFO_LOOP (loop_vinfo);
3762 nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
3763 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
3768 /* Multiple types in SLP are handled by creating the appropriate number of
3769 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3774 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3776 gcc_assert (ncopies >= 1);
3778 /* FORNOW. This restriction should be relaxed. */
3779 if (nested_in_vect_loop && ncopies > 1)
3781 if (vect_print_dump_info (REPORT_DETAILS))
3782 fprintf (vect_dump, "multiple types in nested loop.");
3786 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3789 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3792 /* Is vectorizable load? */
3793 if (!is_gimple_assign (stmt))
3796 scalar_dest = gimple_assign_lhs (stmt);
3797 if (TREE_CODE (scalar_dest) != SSA_NAME)
3800 code = gimple_assign_rhs_code (stmt);
3801 if (code != ARRAY_REF
3802 && code != INDIRECT_REF
3803 && code != COMPONENT_REF
3804 && code != IMAGPART_EXPR
3805 && code != REALPART_EXPR
3809 if (!STMT_VINFO_DATA_REF (stmt_info))
3812 negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0;
3813 if (negative && ncopies > 1)
3815 if (vect_print_dump_info (REPORT_DETAILS))
3816 fprintf (vect_dump, "multiple types with negative step.");
3820 scalar_type = TREE_TYPE (DR_REF (dr));
3821 mode = TYPE_MODE (vectype);
3823 /* FORNOW. In some cases can vectorize even if data-type not supported
3824 (e.g. - data copies). */
3825 if (optab_handler (mov_optab, mode) == CODE_FOR_nothing)
3827 if (vect_print_dump_info (REPORT_DETAILS))
3828 fprintf (vect_dump, "Aligned load, but unsupported type.");
3832 /* The vector component type needs to be trivially convertible to the
3833 scalar lhs. This should always be the case. */
3834 if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype)))
3836 if (vect_print_dump_info (REPORT_DETAILS))
3837 fprintf (vect_dump, "??? operands of different types");
3841 /* Check if the load is a part of an interleaving chain. */
3842 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3844 strided_load = true;
3846 gcc_assert (! nested_in_vect_loop);
3848 /* Check if interleaving is supported. */
3849 if (!vect_strided_load_supported (vectype)
3850 && !PURE_SLP_STMT (stmt_info) && !slp)
3856 gcc_assert (!strided_load);
3857 alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
3858 if (alignment_support_scheme != dr_aligned
3859 && alignment_support_scheme != dr_unaligned_supported)
3861 if (vect_print_dump_info (REPORT_DETAILS))
3862 fprintf (vect_dump, "negative step but alignment required.");
3865 if (!perm_mask_for_reverse (vectype, NULL))
3867 if (vect_print_dump_info (REPORT_DETAILS))
3868 fprintf (vect_dump, "negative step and reversing not supported.");
3873 if (!vec_stmt) /* transformation not required. */
3875 STMT_VINFO_TYPE (stmt_info) = load_vec_info_type;
3876 vect_model_load_cost (stmt_info, ncopies, NULL);
3880 if (vect_print_dump_info (REPORT_DETAILS))
3881 fprintf (vect_dump, "transform load.");
3887 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3888 /* Check if the chain of loads is already vectorized. */
3889 if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)))
3891 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3894 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3895 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3897 /* VEC_NUM is the number of vect stmts to be created for this group. */
3900 strided_load = false;
3901 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3902 if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance))
3906 vec_num = group_size;
3908 dr_chain = VEC_alloc (tree, heap, vec_num);
3914 group_size = vec_num = 1;
3917 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3918 gcc_assert (alignment_support_scheme);
3920 /* In case the vectorization factor (VF) is bigger than the number
3921 of elements that we can fit in a vectype (nunits), we have to generate
3922 more than one vector stmt - i.e - we need to "unroll" the
3923 vector stmt by a factor VF/nunits. In doing so, we record a pointer
3924 from one copy of the vector stmt to the next, in the field
3925 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
3926 stages to find the correct vector defs to be used when vectorizing
3927 stmts that use the defs of the current stmt. The example below
3928 illustrates the vectorization process when VF=16 and nunits=4 (i.e., we
3929 need to create 4 vectorized stmts):
3931 before vectorization:
3932 RELATED_STMT VEC_STMT
3936 step 1: vectorize stmt S1:
3937 We first create the vector stmt VS1_0, and, as usual, record a
3938 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
3939 Next, we create the vector stmt VS1_1, and record a pointer to
3940 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
3941 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
3943 RELATED_STMT VEC_STMT
3944 VS1_0: vx0 = memref0 VS1_1 -
3945 VS1_1: vx1 = memref1 VS1_2 -
3946 VS1_2: vx2 = memref2 VS1_3 -
3947 VS1_3: vx3 = memref3 - -
3948 S1: x = load - VS1_0
3951 See in documentation in vect_get_vec_def_for_stmt_copy for how the
3952 information we recorded in RELATED_STMT field is used to vectorize
3955 /* In case of interleaving (non-unit strided access):
3962 Vectorized loads are created in the order of memory accesses
3963 starting from the access of the first stmt of the chain:
3966 VS2: vx1 = &base + vec_size*1
3967 VS3: vx3 = &base + vec_size*2
3968 VS4: vx4 = &base + vec_size*3
3970 Then permutation statements are generated:
3972 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
3973 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
3976 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3977 (the order of the data-refs in the output of vect_permute_load_chain
3978 corresponds to the order of scalar stmts in the interleaving chain - see
3979 the documentation of vect_permute_load_chain()).
3980 The generation of permutation stmts and recording them in
3981 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
3983 In case of both multiple types and interleaving, the vector loads and
3984 permutation stmts above are created for every copy. The result vector
3985 stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the
3986 corresponding STMT_VINFO_RELATED_STMT for the next copies. */
3988 /* If the data reference is aligned (dr_aligned) or potentially unaligned
3989 on a target that supports unaligned accesses (dr_unaligned_supported)
3990 we generate the following code:
3994 p = p + indx * vectype_size;
3999 Otherwise, the data reference is potentially unaligned on a target that
4000 does not support unaligned accesses (dr_explicit_realign_optimized) -
4001 then generate the following code, in which the data in each iteration is
4002 obtained by two vector loads, one from the previous iteration, and one
4003 from the current iteration:
4005 msq_init = *(floor(p1))
4006 p2 = initial_addr + VS - 1;
4007 realignment_token = call target_builtin;
4010 p2 = p2 + indx * vectype_size
4012 vec_dest = realign_load (msq, lsq, realignment_token)
4017 /* If the misalignment remains the same throughout the execution of the
4018 loop, we can create the init_addr and permutation mask at the loop
4019 preheader. Otherwise, it needs to be created inside the loop.
4020 This can only occur when vectorizing memory accesses in the inner-loop
4021 nested within an outer-loop that is being vectorized. */
4023 if (loop && nested_in_vect_loop_p (loop, stmt)
4024 && (TREE_INT_CST_LOW (DR_STEP (dr))
4025 % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0))
4027 gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized);
4028 compute_in_loop = true;
4031 if ((alignment_support_scheme == dr_explicit_realign_optimized
4032 || alignment_support_scheme == dr_explicit_realign)
4033 && !compute_in_loop)
4035 msq = vect_setup_realignment (first_stmt, gsi, &realignment_token,
4036 alignment_support_scheme, NULL_TREE,
4038 if (alignment_support_scheme == dr_explicit_realign_optimized)
4040 phi = SSA_NAME_DEF_STMT (msq);
4041 offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
4048 offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1);
4050 prev_stmt_info = NULL;
4051 for (j = 0; j < ncopies; j++)
4053 /* 1. Create the vector pointer update chain. */
4055 dataref_ptr = vect_create_data_ref_ptr (first_stmt,
4057 &dummy, &ptr_incr, false,
4061 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
4063 for (i = 0; i < vec_num; i++)
4066 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
4069 /* 2. Create the vector-load in the loop. */
4070 switch (alignment_support_scheme)
4073 case dr_unaligned_supported:
4075 struct ptr_info_def *pi;
4077 = build2 (MEM_REF, vectype, dataref_ptr,
4078 build_int_cst (reference_alias_ptr_type
4079 (DR_REF (first_dr)), 0));
4080 pi = get_ptr_info (dataref_ptr);
4081 pi->align = TYPE_ALIGN_UNIT (vectype);
4082 if (alignment_support_scheme == dr_aligned)
4084 gcc_assert (aligned_access_p (first_dr));
4087 else if (DR_MISALIGNMENT (first_dr) == -1)
4089 TREE_TYPE (data_ref)
4090 = build_aligned_type (TREE_TYPE (data_ref),
4091 TYPE_ALIGN (TREE_TYPE (vectype)));
4092 pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype));
4097 TREE_TYPE (data_ref)
4098 = build_aligned_type (TREE_TYPE (data_ref),
4099 TYPE_ALIGN (TREE_TYPE (vectype)));
4100 pi->misalign = DR_MISALIGNMENT (first_dr);
4104 case dr_explicit_realign:
4107 tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
4109 if (compute_in_loop)
4110 msq = vect_setup_realignment (first_stmt, gsi,
4112 dr_explicit_realign,
4115 new_stmt = gimple_build_assign_with_ops
4116 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
4118 (TREE_TYPE (dataref_ptr),
4119 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
4120 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
4121 gimple_assign_set_lhs (new_stmt, ptr);
4122 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4124 = build2 (MEM_REF, vectype, ptr,
4125 build_int_cst (reference_alias_ptr_type
4126 (DR_REF (first_dr)), 0));
4127 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4128 new_stmt = gimple_build_assign (vec_dest, data_ref);
4129 new_temp = make_ssa_name (vec_dest, new_stmt);
4130 gimple_assign_set_lhs (new_stmt, new_temp);
4131 gimple_set_vdef (new_stmt, gimple_vdef (stmt));
4132 gimple_set_vuse (new_stmt, gimple_vuse (stmt));
4133 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4136 bump = size_binop (MULT_EXPR, vs_minus_1,
4137 TYPE_SIZE_UNIT (scalar_type));
4138 ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump);
4139 new_stmt = gimple_build_assign_with_ops
4140 (BIT_AND_EXPR, NULL_TREE, ptr,
4143 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
4144 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
4145 gimple_assign_set_lhs (new_stmt, ptr);
4146 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4148 = build2 (MEM_REF, vectype, ptr,
4149 build_int_cst (reference_alias_ptr_type
4150 (DR_REF (first_dr)), 0));
4153 case dr_explicit_realign_optimized:
4154 new_stmt = gimple_build_assign_with_ops
4155 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
4157 (TREE_TYPE (dataref_ptr),
4158 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
4159 new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
4160 gimple_assign_set_lhs (new_stmt, new_temp);
4161 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4163 = build2 (MEM_REF, vectype, new_temp,
4164 build_int_cst (reference_alias_ptr_type
4165 (DR_REF (first_dr)), 0));
4170 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4171 new_stmt = gimple_build_assign (vec_dest, data_ref);
4172 new_temp = make_ssa_name (vec_dest, new_stmt);
4173 gimple_assign_set_lhs (new_stmt, new_temp);
4174 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4175 mark_symbols_for_renaming (new_stmt);
4177 /* 3. Handle explicit realignment if necessary/supported. Create in
4178 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
4179 if (alignment_support_scheme == dr_explicit_realign_optimized
4180 || alignment_support_scheme == dr_explicit_realign)
4184 lsq = gimple_assign_lhs (new_stmt);
4185 if (!realignment_token)
4186 realignment_token = dataref_ptr;
4187 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4188 tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq,
4190 new_stmt = gimple_build_assign (vec_dest, tmp);
4191 new_temp = make_ssa_name (vec_dest, new_stmt);
4192 gimple_assign_set_lhs (new_stmt, new_temp);
4193 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4195 if (alignment_support_scheme == dr_explicit_realign_optimized)
4198 if (i == vec_num - 1 && j == ncopies - 1)
4199 add_phi_arg (phi, lsq, loop_latch_edge (containing_loop),
4205 /* 4. Handle invariant-load. */
4206 if (inv_p && !bb_vinfo)
4208 gcc_assert (!strided_load);
4209 gcc_assert (nested_in_vect_loop_p (loop, stmt));
4214 tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type);
4216 /* CHECKME: bitpos depends on endianess? */
4217 bitpos = bitsize_zero_node;
4218 vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp,
4221 vect_create_destination_var (scalar_dest, NULL_TREE);
4222 new_stmt = gimple_build_assign (vec_dest, vec_inv);
4223 new_temp = make_ssa_name (vec_dest, new_stmt);
4224 gimple_assign_set_lhs (new_stmt, new_temp);
4225 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4227 for (k = nunits - 1; k >= 0; --k)
4228 t = tree_cons (NULL_TREE, new_temp, t);
4229 /* FIXME: use build_constructor directly. */
4230 vec_inv = build_constructor_from_list (vectype, t);
4231 new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi);
4232 new_stmt = SSA_NAME_DEF_STMT (new_temp);
4235 gcc_unreachable (); /* FORNOW. */
4240 new_temp = reverse_vec_elements (new_temp, stmt, gsi);
4241 new_stmt = SSA_NAME_DEF_STMT (new_temp);
4244 /* Collect vector loads and later create their permutation in
4245 vect_transform_strided_load (). */
4246 if (strided_load || slp_perm)
4247 VEC_quick_push (tree, dr_chain, new_temp);
4249 /* Store vector loads in the corresponding SLP_NODE. */
4250 if (slp && !slp_perm)
4251 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
4254 if (slp && !slp_perm)
4259 if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf,
4260 slp_node_instance, false))
4262 VEC_free (tree, heap, dr_chain);
4270 if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi))
4273 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
4274 VEC_free (tree, heap, dr_chain);
4275 dr_chain = VEC_alloc (tree, heap, group_size);
4280 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4282 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4283 prev_stmt_info = vinfo_for_stmt (new_stmt);
4289 VEC_free (tree, heap, dr_chain);
4294 /* Function vect_is_simple_cond.
4297 LOOP - the loop that is being vectorized.
4298 COND - Condition that is checked for simple use.
4300 Returns whether a COND can be vectorized. Checks whether
4301 condition operands are supportable using vec_is_simple_use. */
4304 vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo)
4308 enum vect_def_type dt;
4310 if (!COMPARISON_CLASS_P (cond))
4313 lhs = TREE_OPERAND (cond, 0);
4314 rhs = TREE_OPERAND (cond, 1);
4316 if (TREE_CODE (lhs) == SSA_NAME)
4318 gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs);
4319 if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def,
4323 else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST
4324 && TREE_CODE (lhs) != FIXED_CST)
4327 if (TREE_CODE (rhs) == SSA_NAME)
4329 gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs);
4330 if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def,
4334 else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST
4335 && TREE_CODE (rhs) != FIXED_CST)
4341 /* vectorizable_condition.
4343 Check if STMT is conditional modify expression that can be vectorized.
4344 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4345 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
4348 When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable
4349 to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in
4350 else caluse if it is 2).
4352 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4355 vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi,
4356 gimple *vec_stmt, tree reduc_def, int reduc_index)
4358 tree scalar_dest = NULL_TREE;
4359 tree vec_dest = NULL_TREE;
4360 tree op = NULL_TREE;
4361 tree cond_expr, then_clause, else_clause;
4362 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4363 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4364 tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE;
4365 tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE;
4366 tree vec_compare, vec_cond_expr;
4368 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4369 enum machine_mode vec_mode;
4371 enum vect_def_type dt, dts[4];
4372 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
4373 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
4374 enum tree_code code;
4375 stmt_vec_info prev_stmt_info = NULL;
4378 /* FORNOW: unsupported in basic block SLP. */
4379 gcc_assert (loop_vinfo);
4381 gcc_assert (ncopies >= 1);
4382 if (reduc_index && ncopies > 1)
4383 return false; /* FORNOW */
4385 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4388 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
4389 && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle
4393 /* FORNOW: SLP not supported. */
4394 if (STMT_SLP_TYPE (stmt_info))
4397 /* FORNOW: not yet supported. */
4398 if (STMT_VINFO_LIVE_P (stmt_info))
4400 if (vect_print_dump_info (REPORT_DETAILS))
4401 fprintf (vect_dump, "value used after loop.");
4405 /* Is vectorizable conditional operation? */
4406 if (!is_gimple_assign (stmt))
4409 code = gimple_assign_rhs_code (stmt);
4411 if (code != COND_EXPR)
4414 gcc_assert (gimple_assign_single_p (stmt));
4415 op = gimple_assign_rhs1 (stmt);
4416 cond_expr = TREE_OPERAND (op, 0);
4417 then_clause = TREE_OPERAND (op, 1);
4418 else_clause = TREE_OPERAND (op, 2);
4420 if (!vect_is_simple_cond (cond_expr, loop_vinfo))
4423 /* We do not handle two different vector types for the condition
4425 if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)),
4426 TREE_TYPE (vectype)))
4429 if (TREE_CODE (then_clause) == SSA_NAME)
4431 gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause);
4432 if (!vect_is_simple_use (then_clause, loop_vinfo, NULL,
4433 &then_def_stmt, &def, &dt))
4436 else if (TREE_CODE (then_clause) != INTEGER_CST
4437 && TREE_CODE (then_clause) != REAL_CST
4438 && TREE_CODE (then_clause) != FIXED_CST)
4441 if (TREE_CODE (else_clause) == SSA_NAME)
4443 gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause);
4444 if (!vect_is_simple_use (else_clause, loop_vinfo, NULL,
4445 &else_def_stmt, &def, &dt))
4448 else if (TREE_CODE (else_clause) != INTEGER_CST
4449 && TREE_CODE (else_clause) != REAL_CST
4450 && TREE_CODE (else_clause) != FIXED_CST)
4454 vec_mode = TYPE_MODE (vectype);
4458 STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type;
4459 return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode);
4465 scalar_dest = gimple_assign_lhs (stmt);
4466 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4468 /* Handle cond expr. */
4469 for (j = 0; j < ncopies; j++)
4476 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0),
4478 vect_is_simple_use (TREE_OPERAND (cond_expr, 0), loop_vinfo,
4479 NULL, >emp, &def, &dts[0]);
4481 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1),
4483 vect_is_simple_use (TREE_OPERAND (cond_expr, 1), loop_vinfo,
4484 NULL, >emp, &def, &dts[1]);
4485 if (reduc_index == 1)
4486 vec_then_clause = reduc_def;
4489 vec_then_clause = vect_get_vec_def_for_operand (then_clause,
4491 vect_is_simple_use (then_clause, loop_vinfo,
4492 NULL, >emp, &def, &dts[2]);
4494 if (reduc_index == 2)
4495 vec_else_clause = reduc_def;
4498 vec_else_clause = vect_get_vec_def_for_operand (else_clause,
4500 vect_is_simple_use (else_clause, loop_vinfo,
4501 NULL, >emp, &def, &dts[3]);
4506 vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], vec_cond_lhs);
4507 vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], vec_cond_rhs);
4508 vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2],
4510 vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3],
4514 /* Arguments are ready. Create the new vector stmt. */
4515 vec_compare = build2 (TREE_CODE (cond_expr), vectype,
4516 vec_cond_lhs, vec_cond_rhs);
4517 vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
4518 vec_compare, vec_then_clause, vec_else_clause);
4520 new_stmt = gimple_build_assign (vec_dest, vec_cond_expr);
4521 new_temp = make_ssa_name (vec_dest, new_stmt);
4522 gimple_assign_set_lhs (new_stmt, new_temp);
4523 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4525 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4527 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4529 prev_stmt_info = vinfo_for_stmt (new_stmt);
4536 /* Make sure the statement is vectorizable. */
4539 vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node)
4541 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4542 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
4543 enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info);
4545 tree scalar_type, vectype;
4547 if (vect_print_dump_info (REPORT_DETAILS))
4549 fprintf (vect_dump, "==> examining statement: ");
4550 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4553 if (gimple_has_volatile_ops (stmt))
4555 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4556 fprintf (vect_dump, "not vectorized: stmt has volatile operands");
4561 /* Skip stmts that do not need to be vectorized. In loops this is expected
4563 - the COND_EXPR which is the loop exit condition
4564 - any LABEL_EXPRs in the loop
4565 - computations that are used only for array indexing or loop control.
4566 In basic blocks we only analyze statements that are a part of some SLP
4567 instance, therefore, all the statements are relevant. */
4569 if (!STMT_VINFO_RELEVANT_P (stmt_info)
4570 && !STMT_VINFO_LIVE_P (stmt_info))
4572 if (vect_print_dump_info (REPORT_DETAILS))
4573 fprintf (vect_dump, "irrelevant.");
4578 switch (STMT_VINFO_DEF_TYPE (stmt_info))
4580 case vect_internal_def:
4583 case vect_reduction_def:
4584 case vect_nested_cycle:
4585 gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer
4586 || relevance == vect_used_in_outer_by_reduction
4587 || relevance == vect_unused_in_scope));
4590 case vect_induction_def:
4591 case vect_constant_def:
4592 case vect_external_def:
4593 case vect_unknown_def_type:
4600 gcc_assert (PURE_SLP_STMT (stmt_info));
4602 scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
4603 if (vect_print_dump_info (REPORT_DETAILS))
4605 fprintf (vect_dump, "get vectype for scalar type: ");
4606 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4609 vectype = get_vectype_for_scalar_type (scalar_type);
4612 if (vect_print_dump_info (REPORT_DETAILS))
4614 fprintf (vect_dump, "not SLPed: unsupported data-type ");
4615 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4620 if (vect_print_dump_info (REPORT_DETAILS))
4622 fprintf (vect_dump, "vectype: ");
4623 print_generic_expr (vect_dump, vectype, TDF_SLIM);
4626 STMT_VINFO_VECTYPE (stmt_info) = vectype;
4629 if (STMT_VINFO_RELEVANT_P (stmt_info))
4631 gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt))));
4632 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
4633 *need_to_vectorize = true;
4638 && (STMT_VINFO_RELEVANT_P (stmt_info)
4639 || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def))
4640 ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL)
4641 || vectorizable_type_demotion (stmt, NULL, NULL, NULL)
4642 || vectorizable_conversion (stmt, NULL, NULL, NULL)
4643 || vectorizable_shift (stmt, NULL, NULL, NULL)
4644 || vectorizable_operation (stmt, NULL, NULL, NULL)
4645 || vectorizable_assignment (stmt, NULL, NULL, NULL)
4646 || vectorizable_load (stmt, NULL, NULL, NULL, NULL)
4647 || vectorizable_call (stmt, NULL, NULL)
4648 || vectorizable_store (stmt, NULL, NULL, NULL)
4649 || vectorizable_reduction (stmt, NULL, NULL, NULL)
4650 || vectorizable_condition (stmt, NULL, NULL, NULL, 0));
4654 ok = (vectorizable_shift (stmt, NULL, NULL, node)
4655 || vectorizable_operation (stmt, NULL, NULL, node)
4656 || vectorizable_assignment (stmt, NULL, NULL, node)
4657 || vectorizable_load (stmt, NULL, NULL, node, NULL)
4658 || vectorizable_store (stmt, NULL, NULL, node));
4663 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4665 fprintf (vect_dump, "not vectorized: relevant stmt not ");
4666 fprintf (vect_dump, "supported: ");
4667 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4676 /* Stmts that are (also) "live" (i.e. - that are used out of the loop)
4677 need extra handling, except for vectorizable reductions. */
4678 if (STMT_VINFO_LIVE_P (stmt_info)
4679 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4680 ok = vectorizable_live_operation (stmt, NULL, NULL);
4684 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4686 fprintf (vect_dump, "not vectorized: live stmt not ");
4687 fprintf (vect_dump, "supported: ");
4688 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4694 if (!PURE_SLP_STMT (stmt_info))
4696 /* Groups of strided accesses whose size is not a power of 2 are not
4697 vectorizable yet using loop-vectorization. Therefore, if this stmt
4698 feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and
4699 loop-based vectorized), the loop cannot be vectorized. */
4700 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
4701 && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt (
4702 DR_GROUP_FIRST_DR (stmt_info)))) == -1)
4704 if (vect_print_dump_info (REPORT_DETAILS))
4706 fprintf (vect_dump, "not vectorized: the size of group "
4707 "of strided accesses is not a power of 2");
4708 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4719 /* Function vect_transform_stmt.
4721 Create a vectorized stmt to replace STMT, and insert it at BSI. */
4724 vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi,
4725 bool *strided_store, slp_tree slp_node,
4726 slp_instance slp_node_instance)
4728 bool is_store = false;
4729 gimple vec_stmt = NULL;
4730 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4731 gimple orig_stmt_in_pattern, orig_scalar_stmt = stmt;
4734 switch (STMT_VINFO_TYPE (stmt_info))
4736 case type_demotion_vec_info_type:
4737 done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node);
4741 case type_promotion_vec_info_type:
4742 done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node);
4746 case type_conversion_vec_info_type:
4747 done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node);
4751 case induc_vec_info_type:
4752 gcc_assert (!slp_node);
4753 done = vectorizable_induction (stmt, gsi, &vec_stmt);
4757 case shift_vec_info_type:
4758 done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node);
4762 case op_vec_info_type:
4763 done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node);
4767 case assignment_vec_info_type:
4768 done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node);
4772 case load_vec_info_type:
4773 done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node,
4778 case store_vec_info_type:
4779 done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node);
4781 if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node)
4783 /* In case of interleaving, the whole chain is vectorized when the
4784 last store in the chain is reached. Store stmts before the last
4785 one are skipped, and there vec_stmt_info shouldn't be freed
4787 *strided_store = true;
4788 if (STMT_VINFO_VEC_STMT (stmt_info))
4795 case condition_vec_info_type:
4796 gcc_assert (!slp_node);
4797 done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0);
4801 case call_vec_info_type:
4802 gcc_assert (!slp_node);
4803 done = vectorizable_call (stmt, gsi, &vec_stmt);
4804 stmt = gsi_stmt (*gsi);
4807 case reduc_vec_info_type:
4808 done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node);
4813 if (!STMT_VINFO_LIVE_P (stmt_info))
4815 if (vect_print_dump_info (REPORT_DETAILS))
4816 fprintf (vect_dump, "stmt not supported.");
4821 /* Handle inner-loop stmts whose DEF is used in the loop-nest that
4822 is being vectorized, but outside the immediately enclosing loop. */
4824 && STMT_VINFO_LOOP_VINFO (stmt_info)
4825 && nested_in_vect_loop_p (LOOP_VINFO_LOOP (
4826 STMT_VINFO_LOOP_VINFO (stmt_info)), stmt)
4827 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type
4828 && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer
4829 || STMT_VINFO_RELEVANT (stmt_info) ==
4830 vect_used_in_outer_by_reduction))
4832 struct loop *innerloop = LOOP_VINFO_LOOP (
4833 STMT_VINFO_LOOP_VINFO (stmt_info))->inner;
4834 imm_use_iterator imm_iter;
4835 use_operand_p use_p;
4839 if (vect_print_dump_info (REPORT_DETAILS))
4840 fprintf (vect_dump, "Record the vdef for outer-loop vectorization.");
4842 /* Find the relevant loop-exit phi-node, and reord the vec_stmt there
4843 (to be used when vectorizing outer-loop stmts that use the DEF of
4845 if (gimple_code (stmt) == GIMPLE_PHI)
4846 scalar_dest = PHI_RESULT (stmt);
4848 scalar_dest = gimple_assign_lhs (stmt);
4850 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4852 if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p))))
4854 exit_phi = USE_STMT (use_p);
4855 STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt;
4860 /* Handle stmts whose DEF is used outside the loop-nest that is
4861 being vectorized. */
4862 if (STMT_VINFO_LIVE_P (stmt_info)
4863 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4865 done = vectorizable_live_operation (stmt, gsi, &vec_stmt);
4871 STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
4872 orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info);
4873 if (orig_stmt_in_pattern)
4875 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern);
4876 /* STMT was inserted by the vectorizer to replace a computation idiom.
4877 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
4878 computed this idiom. We need to record a pointer to VEC_STMT in
4879 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
4880 documentation of vect_pattern_recog. */
4881 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
4883 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo)
4884 == orig_scalar_stmt);
4885 STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
4894 /* Remove a group of stores (for SLP or interleaving), free their
4898 vect_remove_stores (gimple first_stmt)
4900 gimple next = first_stmt;
4902 gimple_stmt_iterator next_si;
4906 /* Free the attached stmt_vec_info and remove the stmt. */
4907 next_si = gsi_for_stmt (next);
4908 gsi_remove (&next_si, true);
4909 tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next));
4910 free_stmt_vec_info (next);
4916 /* Function new_stmt_vec_info.
4918 Create and initialize a new stmt_vec_info struct for STMT. */
4921 new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo,
4922 bb_vec_info bb_vinfo)
4925 res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info));
4927 STMT_VINFO_TYPE (res) = undef_vec_info_type;
4928 STMT_VINFO_STMT (res) = stmt;
4929 STMT_VINFO_LOOP_VINFO (res) = loop_vinfo;
4930 STMT_VINFO_BB_VINFO (res) = bb_vinfo;
4931 STMT_VINFO_RELEVANT (res) = vect_unused_in_scope;
4932 STMT_VINFO_LIVE_P (res) = false;
4933 STMT_VINFO_VECTYPE (res) = NULL;
4934 STMT_VINFO_VEC_STMT (res) = NULL;
4935 STMT_VINFO_VECTORIZABLE (res) = true;
4936 STMT_VINFO_IN_PATTERN_P (res) = false;
4937 STMT_VINFO_RELATED_STMT (res) = NULL;
4938 STMT_VINFO_DATA_REF (res) = NULL;
4940 STMT_VINFO_DR_BASE_ADDRESS (res) = NULL;
4941 STMT_VINFO_DR_OFFSET (res) = NULL;
4942 STMT_VINFO_DR_INIT (res) = NULL;
4943 STMT_VINFO_DR_STEP (res) = NULL;
4944 STMT_VINFO_DR_ALIGNED_TO (res) = NULL;
4946 if (gimple_code (stmt) == GIMPLE_PHI
4947 && is_loop_header_bb_p (gimple_bb (stmt)))
4948 STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
4950 STMT_VINFO_DEF_TYPE (res) = vect_internal_def;
4952 STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5);
4953 STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0;
4954 STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0;
4955 STMT_SLP_TYPE (res) = loop_vect;
4956 DR_GROUP_FIRST_DR (res) = NULL;
4957 DR_GROUP_NEXT_DR (res) = NULL;
4958 DR_GROUP_SIZE (res) = 0;
4959 DR_GROUP_STORE_COUNT (res) = 0;
4960 DR_GROUP_GAP (res) = 0;
4961 DR_GROUP_SAME_DR_STMT (res) = NULL;
4962 DR_GROUP_READ_WRITE_DEPENDENCE (res) = false;
4968 /* Create a hash table for stmt_vec_info. */
4971 init_stmt_vec_info_vec (void)
4973 gcc_assert (!stmt_vec_info_vec);
4974 stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50);
4978 /* Free hash table for stmt_vec_info. */
4981 free_stmt_vec_info_vec (void)
4983 gcc_assert (stmt_vec_info_vec);
4984 VEC_free (vec_void_p, heap, stmt_vec_info_vec);
4988 /* Free stmt vectorization related info. */
4991 free_stmt_vec_info (gimple stmt)
4993 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4998 VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info));
4999 set_vinfo_for_stmt (stmt, NULL);
5004 /* Function get_vectype_for_scalar_type_and_size.
5006 Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported
5010 get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size)
5012 enum machine_mode inner_mode = TYPE_MODE (scalar_type);
5013 enum machine_mode simd_mode;
5014 unsigned int nbytes = GET_MODE_SIZE (inner_mode);
5021 /* We can't build a vector type of elements with alignment bigger than
5023 if (nbytes < TYPE_ALIGN_UNIT (scalar_type))
5026 /* If we'd build a vector type of elements whose mode precision doesn't
5027 match their types precision we'll get mismatched types on vector
5028 extracts via BIT_FIELD_REFs. This effectively means we disable
5029 vectorization of bool and/or enum types in some languages. */
5030 if (INTEGRAL_TYPE_P (scalar_type)
5031 && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type))
5034 if (GET_MODE_CLASS (inner_mode) != MODE_INT
5035 && GET_MODE_CLASS (inner_mode) != MODE_FLOAT)
5038 /* If no size was supplied use the mode the target prefers. Otherwise
5039 lookup a vector mode of the specified size. */
5041 simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode);
5043 simd_mode = mode_for_vector (inner_mode, size / nbytes);
5044 nunits = GET_MODE_SIZE (simd_mode) / nbytes;
5048 vectype = build_vector_type (scalar_type, nunits);
5049 if (vect_print_dump_info (REPORT_DETAILS))
5051 fprintf (vect_dump, "get vectype with %d units of type ", nunits);
5052 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
5058 if (vect_print_dump_info (REPORT_DETAILS))
5060 fprintf (vect_dump, "vectype: ");
5061 print_generic_expr (vect_dump, vectype, TDF_SLIM);
5064 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
5065 && !INTEGRAL_MODE_P (TYPE_MODE (vectype)))
5067 if (vect_print_dump_info (REPORT_DETAILS))
5068 fprintf (vect_dump, "mode not supported by target.");
5075 unsigned int current_vector_size;
5077 /* Function get_vectype_for_scalar_type.
5079 Returns the vector type corresponding to SCALAR_TYPE as supported
5083 get_vectype_for_scalar_type (tree scalar_type)
5086 vectype = get_vectype_for_scalar_type_and_size (scalar_type,
5087 current_vector_size);
5089 && current_vector_size == 0)
5090 current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype));
5094 /* Function get_same_sized_vectype
5096 Returns a vector type corresponding to SCALAR_TYPE of size
5097 VECTOR_TYPE if supported by the target. */
5100 get_same_sized_vectype (tree scalar_type, tree vector_type)
5102 return get_vectype_for_scalar_type_and_size
5103 (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type)));
5106 /* Function vect_is_simple_use.
5109 LOOP_VINFO - the vect info of the loop that is being vectorized.
5110 BB_VINFO - the vect info of the basic block that is being vectorized.
5111 OPERAND - operand of a stmt in the loop or bb.
5112 DEF - the defining stmt in case OPERAND is an SSA_NAME.
5114 Returns whether a stmt with OPERAND can be vectorized.
5115 For loops, supportable operands are constants, loop invariants, and operands
5116 that are defined by the current iteration of the loop. Unsupportable
5117 operands are those that are defined by a previous iteration of the loop (as
5118 is the case in reduction/induction computations).
5119 For basic blocks, supportable operands are constants and bb invariants.
5120 For now, operands defined outside the basic block are not supported. */
5123 vect_is_simple_use (tree operand, loop_vec_info loop_vinfo,
5124 bb_vec_info bb_vinfo, gimple *def_stmt,
5125 tree *def, enum vect_def_type *dt)
5128 stmt_vec_info stmt_vinfo;
5129 struct loop *loop = NULL;
5132 loop = LOOP_VINFO_LOOP (loop_vinfo);
5137 if (vect_print_dump_info (REPORT_DETAILS))
5139 fprintf (vect_dump, "vect_is_simple_use: operand ");
5140 print_generic_expr (vect_dump, operand, TDF_SLIM);
5143 if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST)
5145 *dt = vect_constant_def;
5149 if (is_gimple_min_invariant (operand))
5152 *dt = vect_external_def;
5156 if (TREE_CODE (operand) == PAREN_EXPR)
5158 if (vect_print_dump_info (REPORT_DETAILS))
5159 fprintf (vect_dump, "non-associatable copy.");
5160 operand = TREE_OPERAND (operand, 0);
5163 if (TREE_CODE (operand) != SSA_NAME)
5165 if (vect_print_dump_info (REPORT_DETAILS))
5166 fprintf (vect_dump, "not ssa-name.");
5170 *def_stmt = SSA_NAME_DEF_STMT (operand);
5171 if (*def_stmt == NULL)
5173 if (vect_print_dump_info (REPORT_DETAILS))
5174 fprintf (vect_dump, "no def_stmt.");
5178 if (vect_print_dump_info (REPORT_DETAILS))
5180 fprintf (vect_dump, "def_stmt: ");
5181 print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM);
5184 /* Empty stmt is expected only in case of a function argument.
5185 (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */
5186 if (gimple_nop_p (*def_stmt))
5189 *dt = vect_external_def;
5193 bb = gimple_bb (*def_stmt);
5195 if ((loop && !flow_bb_inside_loop_p (loop, bb))
5196 || (!loop && bb != BB_VINFO_BB (bb_vinfo))
5197 || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI))
5198 *dt = vect_external_def;
5201 stmt_vinfo = vinfo_for_stmt (*def_stmt);
5202 *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo);
5205 if (*dt == vect_unknown_def_type)
5207 if (vect_print_dump_info (REPORT_DETAILS))
5208 fprintf (vect_dump, "Unsupported pattern.");
5212 if (vect_print_dump_info (REPORT_DETAILS))
5213 fprintf (vect_dump, "type of def: %d.",*dt);
5215 switch (gimple_code (*def_stmt))
5218 *def = gimple_phi_result (*def_stmt);
5222 *def = gimple_assign_lhs (*def_stmt);
5226 *def = gimple_call_lhs (*def_stmt);
5231 if (vect_print_dump_info (REPORT_DETAILS))
5232 fprintf (vect_dump, "unsupported defining stmt: ");
5239 /* Function vect_is_simple_use_1.
5241 Same as vect_is_simple_use_1 but also determines the vector operand
5242 type of OPERAND and stores it to *VECTYPE. If the definition of
5243 OPERAND is vect_uninitialized_def, vect_constant_def or
5244 vect_external_def *VECTYPE will be set to NULL_TREE and the caller
5245 is responsible to compute the best suited vector type for the
5249 vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo,
5250 bb_vec_info bb_vinfo, gimple *def_stmt,
5251 tree *def, enum vect_def_type *dt, tree *vectype)
5253 if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt))
5256 /* Now get a vector type if the def is internal, otherwise supply
5257 NULL_TREE and leave it up to the caller to figure out a proper
5258 type for the use stmt. */
5259 if (*dt == vect_internal_def
5260 || *dt == vect_induction_def
5261 || *dt == vect_reduction_def
5262 || *dt == vect_double_reduction_def
5263 || *dt == vect_nested_cycle)
5265 stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt);
5266 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
5267 stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info));
5268 *vectype = STMT_VINFO_VECTYPE (stmt_info);
5269 gcc_assert (*vectype != NULL_TREE);
5271 else if (*dt == vect_uninitialized_def
5272 || *dt == vect_constant_def
5273 || *dt == vect_external_def)
5274 *vectype = NULL_TREE;
5282 /* Function supportable_widening_operation
5284 Check whether an operation represented by the code CODE is a
5285 widening operation that is supported by the target platform in
5286 vector form (i.e., when operating on arguments of type VECTYPE_IN
5287 producing a result of type VECTYPE_OUT).
5289 Widening operations we currently support are NOP (CONVERT), FLOAT
5290 and WIDEN_MULT. This function checks if these operations are supported
5291 by the target platform either directly (via vector tree-codes), or via
5295 - CODE1 and CODE2 are codes of vector operations to be used when
5296 vectorizing the operation, if available.
5297 - DECL1 and DECL2 are decls of target builtin functions to be used
5298 when vectorizing the operation, if available. In this case,
5299 CODE1 and CODE2 are CALL_EXPR.
5300 - MULTI_STEP_CVT determines the number of required intermediate steps in
5301 case of multi-step conversion (like char->short->int - in that case
5302 MULTI_STEP_CVT will be 1).
5303 - INTERM_TYPES contains the intermediate type required to perform the
5304 widening operation (short in the above example). */
5307 supportable_widening_operation (enum tree_code code, gimple stmt,
5308 tree vectype_out, tree vectype_in,
5309 tree *decl1, tree *decl2,
5310 enum tree_code *code1, enum tree_code *code2,
5311 int *multi_step_cvt,
5312 VEC (tree, heap) **interm_types)
5314 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5315 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info);
5316 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
5318 enum machine_mode vec_mode;
5319 enum insn_code icode1, icode2;
5320 optab optab1, optab2;
5321 tree vectype = vectype_in;
5322 tree wide_vectype = vectype_out;
5323 enum tree_code c1, c2;
5325 /* The result of a vectorized widening operation usually requires two vectors
5326 (because the widened results do not fit int one vector). The generated
5327 vector results would normally be expected to be generated in the same
5328 order as in the original scalar computation, i.e. if 8 results are
5329 generated in each vector iteration, they are to be organized as follows:
5330 vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8].
5332 However, in the special case that the result of the widening operation is
5333 used in a reduction computation only, the order doesn't matter (because
5334 when vectorizing a reduction we change the order of the computation).
5335 Some targets can take advantage of this and generate more efficient code.
5336 For example, targets like Altivec, that support widen_mult using a sequence
5337 of {mult_even,mult_odd} generate the following vectors:
5338 vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8].
5340 When vectorizing outer-loops, we execute the inner-loop sequentially
5341 (each vectorized inner-loop iteration contributes to VF outer-loop
5342 iterations in parallel). We therefore don't allow to change the order
5343 of the computation in the inner-loop during outer-loop vectorization. */
5345 if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction
5346 && !nested_in_vect_loop_p (vect_loop, stmt))
5352 && code == WIDEN_MULT_EXPR
5353 && targetm.vectorize.builtin_mul_widen_even
5354 && targetm.vectorize.builtin_mul_widen_even (vectype)
5355 && targetm.vectorize.builtin_mul_widen_odd
5356 && targetm.vectorize.builtin_mul_widen_odd (vectype))
5358 if (vect_print_dump_info (REPORT_DETAILS))
5359 fprintf (vect_dump, "Unordered widening operation detected.");
5361 *code1 = *code2 = CALL_EXPR;
5362 *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype);
5363 *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype);
5369 case WIDEN_MULT_EXPR:
5370 if (BYTES_BIG_ENDIAN)
5372 c1 = VEC_WIDEN_MULT_HI_EXPR;
5373 c2 = VEC_WIDEN_MULT_LO_EXPR;
5377 c2 = VEC_WIDEN_MULT_HI_EXPR;
5378 c1 = VEC_WIDEN_MULT_LO_EXPR;
5383 if (BYTES_BIG_ENDIAN)
5385 c1 = VEC_UNPACK_HI_EXPR;
5386 c2 = VEC_UNPACK_LO_EXPR;
5390 c2 = VEC_UNPACK_HI_EXPR;
5391 c1 = VEC_UNPACK_LO_EXPR;
5396 if (BYTES_BIG_ENDIAN)
5398 c1 = VEC_UNPACK_FLOAT_HI_EXPR;
5399 c2 = VEC_UNPACK_FLOAT_LO_EXPR;
5403 c2 = VEC_UNPACK_FLOAT_HI_EXPR;
5404 c1 = VEC_UNPACK_FLOAT_LO_EXPR;
5408 case FIX_TRUNC_EXPR:
5409 /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/
5410 VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for
5411 computing the operation. */
5418 if (code == FIX_TRUNC_EXPR)
5420 /* The signedness is determined from output operand. */
5421 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
5422 optab2 = optab_for_tree_code (c2, vectype_out, optab_default);
5426 optab1 = optab_for_tree_code (c1, vectype, optab_default);
5427 optab2 = optab_for_tree_code (c2, vectype, optab_default);
5430 if (!optab1 || !optab2)
5433 vec_mode = TYPE_MODE (vectype);
5434 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing
5435 || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing)
5438 /* Check if it's a multi-step conversion that can be done using intermediate
5440 if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype)
5441 || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype))
5444 tree prev_type = vectype, intermediate_type;
5445 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5446 optab optab3, optab4;
5448 if (!CONVERT_EXPR_CODE_P (code))
5454 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5455 intermediate steps in promotion sequence. We try
5456 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
5458 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5459 for (i = 0; i < 3; i++)
5461 intermediate_mode = insn_data[icode1].operand[0].mode;
5462 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5463 TYPE_UNSIGNED (prev_type));
5464 optab3 = optab_for_tree_code (c1, intermediate_type, optab_default);
5465 optab4 = optab_for_tree_code (c2, intermediate_type, optab_default);
5467 if (!optab3 || !optab4
5468 || ((icode1 = optab_handler (optab1, prev_mode))
5469 == CODE_FOR_nothing)
5470 || insn_data[icode1].operand[0].mode != intermediate_mode
5471 || ((icode2 = optab_handler (optab2, prev_mode))
5472 == CODE_FOR_nothing)
5473 || insn_data[icode2].operand[0].mode != intermediate_mode
5474 || ((icode1 = optab_handler (optab3, intermediate_mode))
5475 == CODE_FOR_nothing)
5476 || ((icode2 = optab_handler (optab4, intermediate_mode))
5477 == CODE_FOR_nothing))
5480 VEC_quick_push (tree, *interm_types, intermediate_type);
5481 (*multi_step_cvt)++;
5483 if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype)
5484 && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype))
5487 prev_type = intermediate_type;
5488 prev_mode = intermediate_mode;
5500 /* Function supportable_narrowing_operation
5502 Check whether an operation represented by the code CODE is a
5503 narrowing operation that is supported by the target platform in
5504 vector form (i.e., when operating on arguments of type VECTYPE_IN
5505 and producing a result of type VECTYPE_OUT).
5507 Narrowing operations we currently support are NOP (CONVERT) and
5508 FIX_TRUNC. This function checks if these operations are supported by
5509 the target platform directly via vector tree-codes.
5512 - CODE1 is the code of a vector operation to be used when
5513 vectorizing the operation, if available.
5514 - MULTI_STEP_CVT determines the number of required intermediate steps in
5515 case of multi-step conversion (like int->short->char - in that case
5516 MULTI_STEP_CVT will be 1).
5517 - INTERM_TYPES contains the intermediate type required to perform the
5518 narrowing operation (short in the above example). */
5521 supportable_narrowing_operation (enum tree_code code,
5522 tree vectype_out, tree vectype_in,
5523 enum tree_code *code1, int *multi_step_cvt,
5524 VEC (tree, heap) **interm_types)
5526 enum machine_mode vec_mode;
5527 enum insn_code icode1;
5528 optab optab1, interm_optab;
5529 tree vectype = vectype_in;
5530 tree narrow_vectype = vectype_out;
5532 tree intermediate_type, prev_type;
5538 c1 = VEC_PACK_TRUNC_EXPR;
5541 case FIX_TRUNC_EXPR:
5542 c1 = VEC_PACK_FIX_TRUNC_EXPR;
5546 /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR
5547 tree code and optabs used for computing the operation. */
5554 if (code == FIX_TRUNC_EXPR)
5555 /* The signedness is determined from output operand. */
5556 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
5558 optab1 = optab_for_tree_code (c1, vectype, optab_default);
5563 vec_mode = TYPE_MODE (vectype);
5564 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing)
5567 /* Check if it's a multi-step conversion that can be done using intermediate
5569 if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype))
5571 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5574 prev_type = vectype;
5575 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5576 intermediate steps in promotion sequence. We try
5577 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
5579 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5580 for (i = 0; i < 3; i++)
5582 intermediate_mode = insn_data[icode1].operand[0].mode;
5583 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5584 TYPE_UNSIGNED (prev_type));
5585 interm_optab = optab_for_tree_code (c1, intermediate_type,
5588 || ((icode1 = optab_handler (optab1, prev_mode))
5589 == CODE_FOR_nothing)
5590 || insn_data[icode1].operand[0].mode != intermediate_mode
5591 || ((icode1 = optab_handler (interm_optab, intermediate_mode))
5592 == CODE_FOR_nothing))
5595 VEC_quick_push (tree, *interm_types, intermediate_type);
5596 (*multi_step_cvt)++;
5598 if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype))
5601 prev_type = intermediate_type;
5602 prev_mode = intermediate_mode;