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[3]
1312 = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type};
1313 gimple new_stmt = NULL;
1315 VEC(tree, heap) *vargs = NULL;
1316 enum { NARROW, NONE, WIDEN } modifier;
1319 /* FORNOW: unsupported in basic block SLP. */
1320 gcc_assert (loop_vinfo);
1322 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1325 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1328 /* FORNOW: SLP not supported. */
1329 if (STMT_SLP_TYPE (stmt_info))
1332 /* Is STMT a vectorizable call? */
1333 if (!is_gimple_call (stmt))
1336 if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
1339 if (stmt_could_throw_p (stmt))
1342 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1344 /* Process function arguments. */
1345 rhs_type = NULL_TREE;
1346 vectype_in = NULL_TREE;
1347 nargs = gimple_call_num_args (stmt);
1349 /* Bail out if the function has more than three arguments, we do not have
1350 interesting builtin functions to vectorize with more than two arguments
1351 except for fma. No arguments is also not good. */
1352 if (nargs == 0 || nargs > 3)
1355 for (i = 0; i < nargs; i++)
1359 op = gimple_call_arg (stmt, i);
1361 /* We can only handle calls with arguments of the same type. */
1363 && !types_compatible_p (rhs_type, TREE_TYPE (op)))
1365 if (vect_print_dump_info (REPORT_DETAILS))
1366 fprintf (vect_dump, "argument types differ.");
1370 rhs_type = TREE_TYPE (op);
1372 if (!vect_is_simple_use_1 (op, loop_vinfo, NULL,
1373 &def_stmt, &def, &dt[i], &opvectype))
1375 if (vect_print_dump_info (REPORT_DETAILS))
1376 fprintf (vect_dump, "use not simple.");
1381 vectype_in = opvectype;
1383 && opvectype != vectype_in)
1385 if (vect_print_dump_info (REPORT_DETAILS))
1386 fprintf (vect_dump, "argument vector types differ.");
1390 /* If all arguments are external or constant defs use a vector type with
1391 the same size as the output vector type. */
1393 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1395 gcc_assert (vectype_in);
1398 if (vect_print_dump_info (REPORT_DETAILS))
1400 fprintf (vect_dump, "no vectype for scalar type ");
1401 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1408 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1409 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1410 if (nunits_in == nunits_out / 2)
1412 else if (nunits_out == nunits_in)
1414 else if (nunits_out == nunits_in / 2)
1419 /* For now, we only vectorize functions if a target specific builtin
1420 is available. TODO -- in some cases, it might be profitable to
1421 insert the calls for pieces of the vector, in order to be able
1422 to vectorize other operations in the loop. */
1423 fndecl = vectorizable_function (stmt, vectype_out, vectype_in);
1424 if (fndecl == NULL_TREE)
1426 if (vect_print_dump_info (REPORT_DETAILS))
1427 fprintf (vect_dump, "function is not vectorizable.");
1432 gcc_assert (!gimple_vuse (stmt));
1434 if (modifier == NARROW)
1435 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1437 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1439 /* Sanity check: make sure that at least one copy of the vectorized stmt
1440 needs to be generated. */
1441 gcc_assert (ncopies >= 1);
1443 if (!vec_stmt) /* transformation not required. */
1445 STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
1446 if (vect_print_dump_info (REPORT_DETAILS))
1447 fprintf (vect_dump, "=== vectorizable_call ===");
1448 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
1454 if (vect_print_dump_info (REPORT_DETAILS))
1455 fprintf (vect_dump, "transform operation.");
1458 scalar_dest = gimple_call_lhs (stmt);
1459 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1461 prev_stmt_info = NULL;
1465 for (j = 0; j < ncopies; ++j)
1467 /* Build argument list for the vectorized call. */
1469 vargs = VEC_alloc (tree, heap, nargs);
1471 VEC_truncate (tree, vargs, 0);
1473 for (i = 0; i < nargs; i++)
1475 op = gimple_call_arg (stmt, i);
1478 = vect_get_vec_def_for_operand (op, stmt, NULL);
1481 vec_oprnd0 = gimple_call_arg (new_stmt, i);
1483 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1486 VEC_quick_push (tree, vargs, vec_oprnd0);
1489 new_stmt = gimple_build_call_vec (fndecl, vargs);
1490 new_temp = make_ssa_name (vec_dest, new_stmt);
1491 gimple_call_set_lhs (new_stmt, new_temp);
1493 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1494 mark_symbols_for_renaming (new_stmt);
1497 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1499 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1501 prev_stmt_info = vinfo_for_stmt (new_stmt);
1507 for (j = 0; j < ncopies; ++j)
1509 /* Build argument list for the vectorized call. */
1511 vargs = VEC_alloc (tree, heap, nargs * 2);
1513 VEC_truncate (tree, vargs, 0);
1515 for (i = 0; i < nargs; i++)
1517 op = gimple_call_arg (stmt, i);
1521 = vect_get_vec_def_for_operand (op, stmt, NULL);
1523 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1527 vec_oprnd1 = gimple_call_arg (new_stmt, 2*i);
1529 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1);
1531 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1534 VEC_quick_push (tree, vargs, vec_oprnd0);
1535 VEC_quick_push (tree, vargs, vec_oprnd1);
1538 new_stmt = gimple_build_call_vec (fndecl, vargs);
1539 new_temp = make_ssa_name (vec_dest, new_stmt);
1540 gimple_call_set_lhs (new_stmt, new_temp);
1542 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1543 mark_symbols_for_renaming (new_stmt);
1546 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1548 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1550 prev_stmt_info = vinfo_for_stmt (new_stmt);
1553 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1558 /* No current target implements this case. */
1562 VEC_free (tree, heap, vargs);
1564 /* Update the exception handling table with the vector stmt if necessary. */
1565 if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt))
1566 gimple_purge_dead_eh_edges (gimple_bb (stmt));
1568 /* The call in STMT might prevent it from being removed in dce.
1569 We however cannot remove it here, due to the way the ssa name
1570 it defines is mapped to the new definition. So just replace
1571 rhs of the statement with something harmless. */
1573 type = TREE_TYPE (scalar_dest);
1574 new_stmt = gimple_build_assign (gimple_call_lhs (stmt),
1575 fold_convert (type, integer_zero_node));
1576 set_vinfo_for_stmt (new_stmt, stmt_info);
1577 set_vinfo_for_stmt (stmt, NULL);
1578 STMT_VINFO_STMT (stmt_info) = new_stmt;
1579 gsi_replace (gsi, new_stmt, false);
1580 SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt;
1586 /* Function vect_gen_widened_results_half
1588 Create a vector stmt whose code, type, number of arguments, and result
1589 variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are
1590 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
1591 In the case that CODE is a CALL_EXPR, this means that a call to DECL
1592 needs to be created (DECL is a function-decl of a target-builtin).
1593 STMT is the original scalar stmt that we are vectorizing. */
1596 vect_gen_widened_results_half (enum tree_code code,
1598 tree vec_oprnd0, tree vec_oprnd1, int op_type,
1599 tree vec_dest, gimple_stmt_iterator *gsi,
1605 /* Generate half of the widened result: */
1606 if (code == CALL_EXPR)
1608 /* Target specific support */
1609 if (op_type == binary_op)
1610 new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1);
1612 new_stmt = gimple_build_call (decl, 1, vec_oprnd0);
1613 new_temp = make_ssa_name (vec_dest, new_stmt);
1614 gimple_call_set_lhs (new_stmt, new_temp);
1618 /* Generic support */
1619 gcc_assert (op_type == TREE_CODE_LENGTH (code));
1620 if (op_type != binary_op)
1622 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0,
1624 new_temp = make_ssa_name (vec_dest, new_stmt);
1625 gimple_assign_set_lhs (new_stmt, new_temp);
1627 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1633 /* Check if STMT performs a conversion operation, that can be vectorized.
1634 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1635 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1636 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1639 vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi,
1640 gimple *vec_stmt, slp_tree slp_node)
1645 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
1646 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1647 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1648 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
1649 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
1653 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1654 gimple new_stmt = NULL;
1655 stmt_vec_info prev_stmt_info;
1658 tree vectype_out, vectype_in;
1662 enum { NARROW, NONE, WIDEN } modifier;
1664 VEC(tree,heap) *vec_oprnds0 = NULL;
1666 VEC(tree,heap) *dummy = NULL;
1669 /* Is STMT a vectorizable conversion? */
1671 /* FORNOW: unsupported in basic block SLP. */
1672 gcc_assert (loop_vinfo);
1674 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1677 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1680 if (!is_gimple_assign (stmt))
1683 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
1686 code = gimple_assign_rhs_code (stmt);
1687 if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
1690 /* Check types of lhs and rhs. */
1691 scalar_dest = gimple_assign_lhs (stmt);
1692 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1694 op0 = gimple_assign_rhs1 (stmt);
1695 rhs_type = TREE_TYPE (op0);
1696 /* Check the operands of the operation. */
1697 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
1698 &def_stmt, &def, &dt[0], &vectype_in))
1700 if (vect_print_dump_info (REPORT_DETAILS))
1701 fprintf (vect_dump, "use not simple.");
1704 /* If op0 is an external or constant defs use a vector type of
1705 the same size as the output vector type. */
1707 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1709 gcc_assert (vectype_in);
1712 if (vect_print_dump_info (REPORT_DETAILS))
1714 fprintf (vect_dump, "no vectype for scalar type ");
1715 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1722 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1723 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1724 if (nunits_in == nunits_out / 2)
1726 else if (nunits_out == nunits_in)
1728 else if (nunits_out == nunits_in / 2)
1733 if (modifier == NARROW)
1734 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1736 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1738 /* Multiple types in SLP are handled by creating the appropriate number of
1739 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
1744 /* Sanity check: make sure that at least one copy of the vectorized stmt
1745 needs to be generated. */
1746 gcc_assert (ncopies >= 1);
1748 /* Supportable by target? */
1749 if ((modifier == NONE
1750 && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in))
1751 || (modifier == WIDEN
1752 && !supportable_widening_operation (code, stmt,
1753 vectype_out, vectype_in,
1756 &dummy_int, &dummy))
1757 || (modifier == NARROW
1758 && !supportable_narrowing_operation (code, vectype_out, vectype_in,
1759 &code1, &dummy_int, &dummy)))
1761 if (vect_print_dump_info (REPORT_DETAILS))
1762 fprintf (vect_dump, "conversion not supported by target.");
1766 if (modifier != NONE)
1768 /* FORNOW: SLP not supported. */
1769 if (STMT_SLP_TYPE (stmt_info))
1773 if (!vec_stmt) /* transformation not required. */
1775 STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type;
1780 if (vect_print_dump_info (REPORT_DETAILS))
1781 fprintf (vect_dump, "transform conversion.");
1784 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1786 if (modifier == NONE && !slp_node)
1787 vec_oprnds0 = VEC_alloc (tree, heap, 1);
1789 prev_stmt_info = NULL;
1793 for (j = 0; j < ncopies; j++)
1796 vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node);
1798 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL);
1801 targetm.vectorize.builtin_conversion (code,
1802 vectype_out, vectype_in);
1803 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
1805 /* Arguments are ready. create the new vector stmt. */
1806 new_stmt = gimple_build_call (builtin_decl, 1, vop0);
1807 new_temp = make_ssa_name (vec_dest, new_stmt);
1808 gimple_call_set_lhs (new_stmt, new_temp);
1809 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1811 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
1815 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1817 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1818 prev_stmt_info = vinfo_for_stmt (new_stmt);
1823 /* In case the vectorization factor (VF) is bigger than the number
1824 of elements that we can fit in a vectype (nunits), we have to
1825 generate more than one vector stmt - i.e - we need to "unroll"
1826 the vector stmt by a factor VF/nunits. */
1827 for (j = 0; j < ncopies; j++)
1830 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1832 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1834 /* Generate first half of the widened result: */
1836 = vect_gen_widened_results_half (code1, decl1,
1837 vec_oprnd0, vec_oprnd1,
1838 unary_op, vec_dest, gsi, stmt);
1840 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1842 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1843 prev_stmt_info = vinfo_for_stmt (new_stmt);
1845 /* Generate second half of the widened result: */
1847 = vect_gen_widened_results_half (code2, decl2,
1848 vec_oprnd0, vec_oprnd1,
1849 unary_op, vec_dest, gsi, stmt);
1850 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1851 prev_stmt_info = vinfo_for_stmt (new_stmt);
1856 /* In case the vectorization factor (VF) is bigger than the number
1857 of elements that we can fit in a vectype (nunits), we have to
1858 generate more than one vector stmt - i.e - we need to "unroll"
1859 the vector stmt by a factor VF/nunits. */
1860 for (j = 0; j < ncopies; j++)
1865 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1866 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1870 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
1871 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1874 /* Arguments are ready. Create the new vector stmt. */
1875 new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0,
1877 new_temp = make_ssa_name (vec_dest, new_stmt);
1878 gimple_assign_set_lhs (new_stmt, new_temp);
1879 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1882 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1884 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1886 prev_stmt_info = vinfo_for_stmt (new_stmt);
1889 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1893 VEC_free (tree, heap, vec_oprnds0);
1899 /* Function vectorizable_assignment.
1901 Check if STMT performs an assignment (copy) that can be vectorized.
1902 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1903 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1904 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1907 vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi,
1908 gimple *vec_stmt, slp_tree slp_node)
1913 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1914 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
1915 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1919 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1920 unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype);
1923 VEC(tree,heap) *vec_oprnds = NULL;
1925 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
1926 gimple new_stmt = NULL;
1927 stmt_vec_info prev_stmt_info = NULL;
1928 enum tree_code code;
1931 /* Multiple types in SLP are handled by creating the appropriate number of
1932 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
1937 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
1939 gcc_assert (ncopies >= 1);
1941 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
1944 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1947 /* Is vectorizable assignment? */
1948 if (!is_gimple_assign (stmt))
1951 scalar_dest = gimple_assign_lhs (stmt);
1952 if (TREE_CODE (scalar_dest) != SSA_NAME)
1955 code = gimple_assign_rhs_code (stmt);
1956 if (gimple_assign_single_p (stmt)
1957 || code == PAREN_EXPR
1958 || CONVERT_EXPR_CODE_P (code))
1959 op = gimple_assign_rhs1 (stmt);
1963 if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo,
1964 &def_stmt, &def, &dt[0], &vectype_in))
1966 if (vect_print_dump_info (REPORT_DETAILS))
1967 fprintf (vect_dump, "use not simple.");
1971 /* We can handle NOP_EXPR conversions that do not change the number
1972 of elements or the vector size. */
1973 if (CONVERT_EXPR_CODE_P (code)
1975 || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits
1976 || (GET_MODE_SIZE (TYPE_MODE (vectype))
1977 != GET_MODE_SIZE (TYPE_MODE (vectype_in)))))
1980 if (!vec_stmt) /* transformation not required. */
1982 STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
1983 if (vect_print_dump_info (REPORT_DETAILS))
1984 fprintf (vect_dump, "=== vectorizable_assignment ===");
1985 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
1990 if (vect_print_dump_info (REPORT_DETAILS))
1991 fprintf (vect_dump, "transform assignment.");
1994 vec_dest = vect_create_destination_var (scalar_dest, vectype);
1997 for (j = 0; j < ncopies; j++)
2001 vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node);
2003 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL);
2005 /* Arguments are ready. create the new vector stmt. */
2006 FOR_EACH_VEC_ELT (tree, vec_oprnds, i, vop)
2008 if (CONVERT_EXPR_CODE_P (code))
2009 vop = build1 (VIEW_CONVERT_EXPR, vectype, vop);
2010 new_stmt = gimple_build_assign (vec_dest, vop);
2011 new_temp = make_ssa_name (vec_dest, new_stmt);
2012 gimple_assign_set_lhs (new_stmt, new_temp);
2013 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2015 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2022 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2024 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2026 prev_stmt_info = vinfo_for_stmt (new_stmt);
2029 VEC_free (tree, heap, vec_oprnds);
2034 /* Function vectorizable_shift.
2036 Check if STMT performs a shift operation that can be vectorized.
2037 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2038 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2039 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2042 vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi,
2043 gimple *vec_stmt, slp_tree slp_node)
2047 tree op0, op1 = NULL;
2048 tree vec_oprnd1 = NULL_TREE;
2049 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2051 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2052 enum tree_code code;
2053 enum machine_mode vec_mode;
2057 enum machine_mode optab_op2_mode;
2060 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2061 gimple new_stmt = NULL;
2062 stmt_vec_info prev_stmt_info;
2068 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2071 bool scalar_shift_arg = false;
2072 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
2075 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
2078 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2081 /* Is STMT a vectorizable binary/unary operation? */
2082 if (!is_gimple_assign (stmt))
2085 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2088 code = gimple_assign_rhs_code (stmt);
2090 if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR
2091 || code == RROTATE_EXPR))
2094 scalar_dest = gimple_assign_lhs (stmt);
2095 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2097 op0 = gimple_assign_rhs1 (stmt);
2098 if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo,
2099 &def_stmt, &def, &dt[0], &vectype))
2101 if (vect_print_dump_info (REPORT_DETAILS))
2102 fprintf (vect_dump, "use not simple.");
2105 /* If op0 is an external or constant def use a vector type with
2106 the same size as the output vector type. */
2108 vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2110 gcc_assert (vectype);
2113 if (vect_print_dump_info (REPORT_DETAILS))
2115 fprintf (vect_dump, "no vectype for scalar type ");
2116 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2122 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2123 nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
2124 if (nunits_out != nunits_in)
2127 op1 = gimple_assign_rhs2 (stmt);
2128 if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[1]))
2130 if (vect_print_dump_info (REPORT_DETAILS))
2131 fprintf (vect_dump, "use not simple.");
2136 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2140 /* Multiple types in SLP are handled by creating the appropriate number of
2141 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2146 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2148 gcc_assert (ncopies >= 1);
2150 /* Determine whether the shift amount is a vector, or scalar. If the
2151 shift/rotate amount is a vector, use the vector/vector shift optabs. */
2153 /* Vector shifted by vector. */
2154 if (dt[1] == vect_internal_def)
2156 optab = optab_for_tree_code (code, vectype, optab_vector);
2157 if (vect_print_dump_info (REPORT_DETAILS))
2158 fprintf (vect_dump, "vector/vector shift/rotate found.");
2160 /* See if the machine has a vector shifted by scalar insn and if not
2161 then see if it has a vector shifted by vector insn. */
2162 else if (dt[1] == vect_constant_def || dt[1] == vect_external_def)
2164 optab = optab_for_tree_code (code, vectype, optab_scalar);
2166 && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing)
2168 scalar_shift_arg = true;
2169 if (vect_print_dump_info (REPORT_DETAILS))
2170 fprintf (vect_dump, "vector/scalar shift/rotate found.");
2174 optab = optab_for_tree_code (code, vectype, optab_vector);
2176 && (optab_handler (optab, TYPE_MODE (vectype))
2177 != CODE_FOR_nothing))
2179 if (vect_print_dump_info (REPORT_DETAILS))
2180 fprintf (vect_dump, "vector/vector shift/rotate found.");
2182 /* Unlike the other binary operators, shifts/rotates have
2183 the rhs being int, instead of the same type as the lhs,
2184 so make sure the scalar is the right type if we are
2185 dealing with vectors of short/char. */
2186 if (dt[1] == vect_constant_def)
2187 op1 = fold_convert (TREE_TYPE (vectype), op1);
2193 if (vect_print_dump_info (REPORT_DETAILS))
2194 fprintf (vect_dump, "operand mode requires invariant argument.");
2198 /* Supportable by target? */
2201 if (vect_print_dump_info (REPORT_DETAILS))
2202 fprintf (vect_dump, "no optab.");
2205 vec_mode = TYPE_MODE (vectype);
2206 icode = (int) optab_handler (optab, vec_mode);
2207 if (icode == CODE_FOR_nothing)
2209 if (vect_print_dump_info (REPORT_DETAILS))
2210 fprintf (vect_dump, "op not supported by target.");
2211 /* Check only during analysis. */
2212 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2213 || (vf < vect_min_worthwhile_factor (code)
2216 if (vect_print_dump_info (REPORT_DETAILS))
2217 fprintf (vect_dump, "proceeding using word mode.");
2220 /* Worthwhile without SIMD support? Check only during analysis. */
2221 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2222 && vf < vect_min_worthwhile_factor (code)
2225 if (vect_print_dump_info (REPORT_DETAILS))
2226 fprintf (vect_dump, "not worthwhile without SIMD support.");
2230 if (!vec_stmt) /* transformation not required. */
2232 STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type;
2233 if (vect_print_dump_info (REPORT_DETAILS))
2234 fprintf (vect_dump, "=== vectorizable_shift ===");
2235 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2241 if (vect_print_dump_info (REPORT_DETAILS))
2242 fprintf (vect_dump, "transform binary/unary operation.");
2245 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2247 /* Allocate VECs for vector operands. In case of SLP, vector operands are
2248 created in the previous stages of the recursion, so no allocation is
2249 needed, except for the case of shift with scalar shift argument. In that
2250 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
2251 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
2252 In case of loop-based vectorization we allocate VECs of size 1. We
2253 allocate VEC_OPRNDS1 only in case of binary operation. */
2256 vec_oprnds0 = VEC_alloc (tree, heap, 1);
2257 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2259 else if (scalar_shift_arg)
2260 vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size);
2262 prev_stmt_info = NULL;
2263 for (j = 0; j < ncopies; j++)
2268 if (scalar_shift_arg)
2270 /* Vector shl and shr insn patterns can be defined with scalar
2271 operand 2 (shift operand). In this case, use constant or loop
2272 invariant op1 directly, without extending it to vector mode
2274 optab_op2_mode = insn_data[icode].operand[2].mode;
2275 if (!VECTOR_MODE_P (optab_op2_mode))
2277 if (vect_print_dump_info (REPORT_DETAILS))
2278 fprintf (vect_dump, "operand 1 using scalar mode.");
2280 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2283 /* Store vec_oprnd1 for every vector stmt to be created
2284 for SLP_NODE. We check during the analysis that all
2285 the shift arguments are the same.
2286 TODO: Allow different constants for different vector
2287 stmts generated for an SLP instance. */
2288 for (k = 0; k < slp_node->vec_stmts_size - 1; k++)
2289 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2294 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
2295 (a special case for certain kind of vector shifts); otherwise,
2296 operand 1 should be of a vector type (the usual case). */
2298 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
2301 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
2305 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
2307 /* Arguments are ready. Create the new vector stmt. */
2308 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
2310 vop1 = VEC_index (tree, vec_oprnds1, i);
2311 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2312 new_temp = make_ssa_name (vec_dest, new_stmt);
2313 gimple_assign_set_lhs (new_stmt, new_temp);
2314 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2316 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2323 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2325 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2326 prev_stmt_info = vinfo_for_stmt (new_stmt);
2329 VEC_free (tree, heap, vec_oprnds0);
2330 VEC_free (tree, heap, vec_oprnds1);
2336 /* Function vectorizable_operation.
2338 Check if STMT performs a binary or unary operation that can be vectorized.
2339 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2340 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2341 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2344 vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi,
2345 gimple *vec_stmt, slp_tree slp_node)
2349 tree op0, op1 = NULL;
2350 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2352 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2353 enum tree_code code;
2354 enum machine_mode vec_mode;
2361 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2362 gimple new_stmt = NULL;
2363 stmt_vec_info prev_stmt_info;
2369 VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2371 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
2374 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
2377 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2380 /* Is STMT a vectorizable binary/unary operation? */
2381 if (!is_gimple_assign (stmt))
2384 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2387 code = gimple_assign_rhs_code (stmt);
2389 /* For pointer addition, we should use the normal plus for
2390 the vector addition. */
2391 if (code == POINTER_PLUS_EXPR)
2394 /* Support only unary or binary operations. */
2395 op_type = TREE_CODE_LENGTH (code);
2396 if (op_type != unary_op && op_type != binary_op)
2398 if (vect_print_dump_info (REPORT_DETAILS))
2399 fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type);
2403 scalar_dest = gimple_assign_lhs (stmt);
2404 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2406 op0 = gimple_assign_rhs1 (stmt);
2407 if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo,
2408 &def_stmt, &def, &dt[0], &vectype))
2410 if (vect_print_dump_info (REPORT_DETAILS))
2411 fprintf (vect_dump, "use not simple.");
2414 /* If op0 is an external or constant def use a vector type with
2415 the same size as the output vector type. */
2417 vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2419 gcc_assert (vectype);
2422 if (vect_print_dump_info (REPORT_DETAILS))
2424 fprintf (vect_dump, "no vectype for scalar type ");
2425 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2431 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2432 nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
2433 if (nunits_out != nunits_in)
2436 if (op_type == binary_op)
2438 op1 = gimple_assign_rhs2 (stmt);
2439 if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def,
2442 if (vect_print_dump_info (REPORT_DETAILS))
2443 fprintf (vect_dump, "use not simple.");
2449 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2453 /* Multiple types in SLP are handled by creating the appropriate number of
2454 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2459 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2461 gcc_assert (ncopies >= 1);
2463 /* Shifts are handled in vectorizable_shift (). */
2464 if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR
2465 || code == RROTATE_EXPR)
2468 optab = optab_for_tree_code (code, vectype, optab_default);
2470 /* Supportable by target? */
2473 if (vect_print_dump_info (REPORT_DETAILS))
2474 fprintf (vect_dump, "no optab.");
2477 vec_mode = TYPE_MODE (vectype);
2478 icode = (int) optab_handler (optab, vec_mode);
2479 if (icode == CODE_FOR_nothing)
2481 if (vect_print_dump_info (REPORT_DETAILS))
2482 fprintf (vect_dump, "op not supported by target.");
2483 /* Check only during analysis. */
2484 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2485 || (vf < vect_min_worthwhile_factor (code)
2488 if (vect_print_dump_info (REPORT_DETAILS))
2489 fprintf (vect_dump, "proceeding using word mode.");
2492 /* Worthwhile without SIMD support? Check only during analysis. */
2493 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2494 && vf < vect_min_worthwhile_factor (code)
2497 if (vect_print_dump_info (REPORT_DETAILS))
2498 fprintf (vect_dump, "not worthwhile without SIMD support.");
2502 if (!vec_stmt) /* transformation not required. */
2504 STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
2505 if (vect_print_dump_info (REPORT_DETAILS))
2506 fprintf (vect_dump, "=== vectorizable_operation ===");
2507 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2513 if (vect_print_dump_info (REPORT_DETAILS))
2514 fprintf (vect_dump, "transform binary/unary operation.");
2517 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2519 /* Allocate VECs for vector operands. In case of SLP, vector operands are
2520 created in the previous stages of the recursion, so no allocation is
2521 needed, except for the case of shift with scalar shift argument. In that
2522 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
2523 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
2524 In case of loop-based vectorization we allocate VECs of size 1. We
2525 allocate VEC_OPRNDS1 only in case of binary operation. */
2528 vec_oprnds0 = VEC_alloc (tree, heap, 1);
2529 if (op_type == binary_op)
2530 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2533 /* In case the vectorization factor (VF) is bigger than the number
2534 of elements that we can fit in a vectype (nunits), we have to generate
2535 more than one vector stmt - i.e - we need to "unroll" the
2536 vector stmt by a factor VF/nunits. In doing so, we record a pointer
2537 from one copy of the vector stmt to the next, in the field
2538 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
2539 stages to find the correct vector defs to be used when vectorizing
2540 stmts that use the defs of the current stmt. The example below
2541 illustrates the vectorization process when VF=16 and nunits=4 (i.e.,
2542 we need to create 4 vectorized stmts):
2544 before vectorization:
2545 RELATED_STMT VEC_STMT
2549 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
2551 RELATED_STMT VEC_STMT
2552 VS1_0: vx0 = memref0 VS1_1 -
2553 VS1_1: vx1 = memref1 VS1_2 -
2554 VS1_2: vx2 = memref2 VS1_3 -
2555 VS1_3: vx3 = memref3 - -
2556 S1: x = load - VS1_0
2559 step2: vectorize stmt S2 (done here):
2560 To vectorize stmt S2 we first need to find the relevant vector
2561 def for the first operand 'x'. This is, as usual, obtained from
2562 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
2563 that defines 'x' (S1). This way we find the stmt VS1_0, and the
2564 relevant vector def 'vx0'. Having found 'vx0' we can generate
2565 the vector stmt VS2_0, and as usual, record it in the
2566 STMT_VINFO_VEC_STMT of stmt S2.
2567 When creating the second copy (VS2_1), we obtain the relevant vector
2568 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
2569 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
2570 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
2571 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
2572 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
2573 chain of stmts and pointers:
2574 RELATED_STMT VEC_STMT
2575 VS1_0: vx0 = memref0 VS1_1 -
2576 VS1_1: vx1 = memref1 VS1_2 -
2577 VS1_2: vx2 = memref2 VS1_3 -
2578 VS1_3: vx3 = memref3 - -
2579 S1: x = load - VS1_0
2580 VS2_0: vz0 = vx0 + v1 VS2_1 -
2581 VS2_1: vz1 = vx1 + v1 VS2_2 -
2582 VS2_2: vz2 = vx2 + v1 VS2_3 -
2583 VS2_3: vz3 = vx3 + v1 - -
2584 S2: z = x + 1 - VS2_0 */
2586 prev_stmt_info = NULL;
2587 for (j = 0; j < ncopies; j++)
2592 if (op_type == binary_op)
2593 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
2596 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
2600 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
2602 /* Arguments are ready. Create the new vector stmt. */
2603 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
2605 vop1 = ((op_type == binary_op)
2606 ? VEC_index (tree, vec_oprnds1, i) : NULL);
2607 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2608 new_temp = make_ssa_name (vec_dest, new_stmt);
2609 gimple_assign_set_lhs (new_stmt, new_temp);
2610 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2612 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2619 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2621 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2622 prev_stmt_info = vinfo_for_stmt (new_stmt);
2625 VEC_free (tree, heap, vec_oprnds0);
2627 VEC_free (tree, heap, vec_oprnds1);
2633 /* Get vectorized definitions for loop-based vectorization. For the first
2634 operand we call vect_get_vec_def_for_operand() (with OPRND containing
2635 scalar operand), and for the rest we get a copy with
2636 vect_get_vec_def_for_stmt_copy() using the previous vector definition
2637 (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details.
2638 The vectors are collected into VEC_OPRNDS. */
2641 vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt,
2642 VEC (tree, heap) **vec_oprnds, int multi_step_cvt)
2646 /* Get first vector operand. */
2647 /* All the vector operands except the very first one (that is scalar oprnd)
2649 if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE)
2650 vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL);
2652 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd);
2654 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2656 /* Get second vector operand. */
2657 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd);
2658 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2662 /* For conversion in multiple steps, continue to get operands
2665 vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1);
2669 /* Create vectorized demotion statements for vector operands from VEC_OPRNDS.
2670 For multi-step conversions store the resulting vectors and call the function
2674 vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds,
2675 int multi_step_cvt, gimple stmt,
2676 VEC (tree, heap) *vec_dsts,
2677 gimple_stmt_iterator *gsi,
2678 slp_tree slp_node, enum tree_code code,
2679 stmt_vec_info *prev_stmt_info)
2682 tree vop0, vop1, new_tmp, vec_dest;
2684 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2686 vec_dest = VEC_pop (tree, vec_dsts);
2688 for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2)
2690 /* Create demotion operation. */
2691 vop0 = VEC_index (tree, *vec_oprnds, i);
2692 vop1 = VEC_index (tree, *vec_oprnds, i + 1);
2693 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2694 new_tmp = make_ssa_name (vec_dest, new_stmt);
2695 gimple_assign_set_lhs (new_stmt, new_tmp);
2696 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2699 /* Store the resulting vector for next recursive call. */
2700 VEC_replace (tree, *vec_oprnds, i/2, new_tmp);
2703 /* This is the last step of the conversion sequence. Store the
2704 vectors in SLP_NODE or in vector info of the scalar statement
2705 (or in STMT_VINFO_RELATED_STMT chain). */
2707 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2710 if (!*prev_stmt_info)
2711 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
2713 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt;
2715 *prev_stmt_info = vinfo_for_stmt (new_stmt);
2720 /* For multi-step demotion operations we first generate demotion operations
2721 from the source type to the intermediate types, and then combine the
2722 results (stored in VEC_OPRNDS) in demotion operation to the destination
2726 /* At each level of recursion we have have of the operands we had at the
2728 VEC_truncate (tree, *vec_oprnds, (i+1)/2);
2729 vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1,
2730 stmt, vec_dsts, gsi, slp_node,
2731 code, prev_stmt_info);
2736 /* Function vectorizable_type_demotion
2738 Check if STMT performs a binary or unary operation that involves
2739 type demotion, and if it can be vectorized.
2740 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2741 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2742 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2745 vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi,
2746 gimple *vec_stmt, slp_tree slp_node)
2751 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2752 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2753 enum tree_code code, code1 = ERROR_MARK;
2756 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2757 stmt_vec_info prev_stmt_info;
2764 int multi_step_cvt = 0;
2765 VEC (tree, heap) *vec_oprnds0 = NULL;
2766 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
2767 tree last_oprnd, intermediate_type;
2769 /* FORNOW: not supported by basic block SLP vectorization. */
2770 gcc_assert (loop_vinfo);
2772 if (!STMT_VINFO_RELEVANT_P (stmt_info))
2775 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2778 /* Is STMT a vectorizable type-demotion operation? */
2779 if (!is_gimple_assign (stmt))
2782 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2785 code = gimple_assign_rhs_code (stmt);
2786 if (!CONVERT_EXPR_CODE_P (code))
2789 scalar_dest = gimple_assign_lhs (stmt);
2790 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2792 /* Check the operands of the operation. */
2793 op0 = gimple_assign_rhs1 (stmt);
2794 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
2795 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
2796 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
2797 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
2798 && CONVERT_EXPR_CODE_P (code))))
2800 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
2801 &def_stmt, &def, &dt[0], &vectype_in))
2803 if (vect_print_dump_info (REPORT_DETAILS))
2804 fprintf (vect_dump, "use not simple.");
2807 /* If op0 is an external def use a vector type with the
2808 same size as the output vector type if possible. */
2810 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2812 gcc_assert (vectype_in);
2815 if (vect_print_dump_info (REPORT_DETAILS))
2817 fprintf (vect_dump, "no vectype for scalar type ");
2818 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2824 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
2825 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2826 if (nunits_in >= nunits_out)
2829 /* Multiple types in SLP are handled by creating the appropriate number of
2830 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2835 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
2836 gcc_assert (ncopies >= 1);
2838 /* Supportable by target? */
2839 if (!supportable_narrowing_operation (code, vectype_out, vectype_in,
2840 &code1, &multi_step_cvt, &interm_types))
2843 if (!vec_stmt) /* transformation not required. */
2845 STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type;
2846 if (vect_print_dump_info (REPORT_DETAILS))
2847 fprintf (vect_dump, "=== vectorizable_demotion ===");
2848 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2853 if (vect_print_dump_info (REPORT_DETAILS))
2854 fprintf (vect_dump, "transform type demotion operation. ncopies = %d.",
2857 /* In case of multi-step demotion, we first generate demotion operations to
2858 the intermediate types, and then from that types to the final one.
2859 We create vector destinations for the intermediate type (TYPES) received
2860 from supportable_narrowing_operation, and store them in the correct order
2861 for future use in vect_create_vectorized_demotion_stmts(). */
2863 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
2865 vec_dsts = VEC_alloc (tree, heap, 1);
2867 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
2868 VEC_quick_push (tree, vec_dsts, vec_dest);
2872 for (i = VEC_length (tree, interm_types) - 1;
2873 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
2875 vec_dest = vect_create_destination_var (scalar_dest,
2877 VEC_quick_push (tree, vec_dsts, vec_dest);
2881 /* In case the vectorization factor (VF) is bigger than the number
2882 of elements that we can fit in a vectype (nunits), we have to generate
2883 more than one vector stmt - i.e - we need to "unroll" the
2884 vector stmt by a factor VF/nunits. */
2886 prev_stmt_info = NULL;
2887 for (j = 0; j < ncopies; j++)
2891 vect_get_slp_defs (op0, NULL_TREE, slp_node, &vec_oprnds0, NULL, -1);
2894 VEC_free (tree, heap, vec_oprnds0);
2895 vec_oprnds0 = VEC_alloc (tree, heap,
2896 (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2));
2897 vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0,
2898 vect_pow2 (multi_step_cvt) - 1);
2901 /* Arguments are ready. Create the new vector stmts. */
2902 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
2903 vect_create_vectorized_demotion_stmts (&vec_oprnds0,
2904 multi_step_cvt, stmt, tmp_vec_dsts,
2905 gsi, slp_node, code1,
2909 VEC_free (tree, heap, vec_oprnds0);
2910 VEC_free (tree, heap, vec_dsts);
2911 VEC_free (tree, heap, tmp_vec_dsts);
2912 VEC_free (tree, heap, interm_types);
2914 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
2919 /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0
2920 and VEC_OPRNDS1 (for binary operations). For multi-step conversions store
2921 the resulting vectors and call the function recursively. */
2924 vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0,
2925 VEC (tree, heap) **vec_oprnds1,
2926 int multi_step_cvt, gimple stmt,
2927 VEC (tree, heap) *vec_dsts,
2928 gimple_stmt_iterator *gsi,
2929 slp_tree slp_node, enum tree_code code1,
2930 enum tree_code code2, tree decl1,
2931 tree decl2, int op_type,
2932 stmt_vec_info *prev_stmt_info)
2935 tree vop0, vop1, new_tmp1, new_tmp2, vec_dest;
2936 gimple new_stmt1, new_stmt2;
2937 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2938 VEC (tree, heap) *vec_tmp;
2940 vec_dest = VEC_pop (tree, vec_dsts);
2941 vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2);
2943 FOR_EACH_VEC_ELT (tree, *vec_oprnds0, i, vop0)
2945 if (op_type == binary_op)
2946 vop1 = VEC_index (tree, *vec_oprnds1, i);
2950 /* Generate the two halves of promotion operation. */
2951 new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1,
2952 op_type, vec_dest, gsi, stmt);
2953 new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1,
2954 op_type, vec_dest, gsi, stmt);
2955 if (is_gimple_call (new_stmt1))
2957 new_tmp1 = gimple_call_lhs (new_stmt1);
2958 new_tmp2 = gimple_call_lhs (new_stmt2);
2962 new_tmp1 = gimple_assign_lhs (new_stmt1);
2963 new_tmp2 = gimple_assign_lhs (new_stmt2);
2968 /* Store the results for the recursive call. */
2969 VEC_quick_push (tree, vec_tmp, new_tmp1);
2970 VEC_quick_push (tree, vec_tmp, new_tmp2);
2974 /* Last step of promotion sequience - store the results. */
2977 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1);
2978 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2);
2982 if (!*prev_stmt_info)
2983 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1;
2985 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1;
2987 *prev_stmt_info = vinfo_for_stmt (new_stmt1);
2988 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2;
2989 *prev_stmt_info = vinfo_for_stmt (new_stmt2);
2996 /* For multi-step promotion operation we first generate we call the
2997 function recurcively for every stage. We start from the input type,
2998 create promotion operations to the intermediate types, and then
2999 create promotions to the output type. */
3000 *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp);
3001 vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1,
3002 multi_step_cvt - 1, stmt,
3003 vec_dsts, gsi, slp_node, code1,
3004 code2, decl2, decl2, op_type,
3008 VEC_free (tree, heap, vec_tmp);
3012 /* Function vectorizable_type_promotion
3014 Check if STMT performs a binary or unary operation that involves
3015 type promotion, and if it can be vectorized.
3016 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3017 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3018 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3021 vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi,
3022 gimple *vec_stmt, slp_tree slp_node)
3026 tree op0, op1 = NULL;
3027 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
3028 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3029 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3030 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
3031 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
3035 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3036 stmt_vec_info prev_stmt_info;
3043 tree intermediate_type = NULL_TREE;
3044 int multi_step_cvt = 0;
3045 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
3046 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
3048 /* FORNOW: not supported by basic block SLP vectorization. */
3049 gcc_assert (loop_vinfo);
3051 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3054 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3057 /* Is STMT a vectorizable type-promotion operation? */
3058 if (!is_gimple_assign (stmt))
3061 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
3064 code = gimple_assign_rhs_code (stmt);
3065 if (!CONVERT_EXPR_CODE_P (code)
3066 && code != WIDEN_MULT_EXPR)
3069 scalar_dest = gimple_assign_lhs (stmt);
3070 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
3072 /* Check the operands of the operation. */
3073 op0 = gimple_assign_rhs1 (stmt);
3074 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
3075 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
3076 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
3077 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
3078 && CONVERT_EXPR_CODE_P (code))))
3080 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
3081 &def_stmt, &def, &dt[0], &vectype_in))
3083 if (vect_print_dump_info (REPORT_DETAILS))
3084 fprintf (vect_dump, "use not simple.");
3087 /* If op0 is an external or constant def use a vector type with
3088 the same size as the output vector type. */
3090 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
3092 gcc_assert (vectype_in);
3095 if (vect_print_dump_info (REPORT_DETAILS))
3097 fprintf (vect_dump, "no vectype for scalar type ");
3098 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
3104 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
3105 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3106 if (nunits_in <= nunits_out)
3109 /* Multiple types in SLP are handled by creating the appropriate number of
3110 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3115 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3117 gcc_assert (ncopies >= 1);
3119 op_type = TREE_CODE_LENGTH (code);
3120 if (op_type == binary_op)
3122 op1 = gimple_assign_rhs2 (stmt);
3123 if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1]))
3125 if (vect_print_dump_info (REPORT_DETAILS))
3126 fprintf (vect_dump, "use not simple.");
3131 /* Supportable by target? */
3132 if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in,
3133 &decl1, &decl2, &code1, &code2,
3134 &multi_step_cvt, &interm_types))
3137 /* Binary widening operation can only be supported directly by the
3139 gcc_assert (!(multi_step_cvt && op_type == binary_op));
3141 if (!vec_stmt) /* transformation not required. */
3143 STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type;
3144 if (vect_print_dump_info (REPORT_DETAILS))
3145 fprintf (vect_dump, "=== vectorizable_promotion ===");
3146 vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL);
3152 if (vect_print_dump_info (REPORT_DETAILS))
3153 fprintf (vect_dump, "transform type promotion operation. ncopies = %d.",
3157 /* In case of multi-step promotion, we first generate promotion operations
3158 to the intermediate types, and then from that types to the final one.
3159 We store vector destination in VEC_DSTS in the correct order for
3160 recursive creation of promotion operations in
3161 vect_create_vectorized_promotion_stmts(). Vector destinations are created
3162 according to TYPES recieved from supportable_widening_operation(). */
3164 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
3166 vec_dsts = VEC_alloc (tree, heap, 1);
3168 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
3169 VEC_quick_push (tree, vec_dsts, vec_dest);
3173 for (i = VEC_length (tree, interm_types) - 1;
3174 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
3176 vec_dest = vect_create_destination_var (scalar_dest,
3178 VEC_quick_push (tree, vec_dsts, vec_dest);
3184 vec_oprnds0 = VEC_alloc (tree, heap,
3185 (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1));
3186 if (op_type == binary_op)
3187 vec_oprnds1 = VEC_alloc (tree, heap, 1);
3190 /* In case the vectorization factor (VF) is bigger than the number
3191 of elements that we can fit in a vectype (nunits), we have to generate
3192 more than one vector stmt - i.e - we need to "unroll" the
3193 vector stmt by a factor VF/nunits. */
3195 prev_stmt_info = NULL;
3196 for (j = 0; j < ncopies; j++)
3202 vect_get_slp_defs (op0, op1, slp_node, &vec_oprnds0,
3206 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
3207 VEC_quick_push (tree, vec_oprnds0, vec_oprnd0);
3208 if (op_type == binary_op)
3210 vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
3211 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
3217 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3218 VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0);
3219 if (op_type == binary_op)
3221 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
3222 VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1);
3226 /* Arguments are ready. Create the new vector stmts. */
3227 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
3228 vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1,
3229 multi_step_cvt, stmt,
3231 gsi, slp_node, code1, code2,
3232 decl1, decl2, op_type,
3236 VEC_free (tree, heap, vec_dsts);
3237 VEC_free (tree, heap, tmp_vec_dsts);
3238 VEC_free (tree, heap, interm_types);
3239 VEC_free (tree, heap, vec_oprnds0);
3240 VEC_free (tree, heap, vec_oprnds1);
3242 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3247 /* Function vectorizable_store.
3249 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
3251 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3252 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3253 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3256 vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3262 tree vec_oprnd = NULL_TREE;
3263 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3264 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL;
3265 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3266 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3267 struct loop *loop = NULL;
3268 enum machine_mode vec_mode;
3270 enum dr_alignment_support alignment_support_scheme;
3273 enum vect_def_type dt;
3274 stmt_vec_info prev_stmt_info = NULL;
3275 tree dataref_ptr = NULL_TREE;
3276 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3279 gimple next_stmt, first_stmt = NULL;
3280 bool strided_store = false;
3281 unsigned int group_size, i;
3282 VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL;
3284 VEC(tree,heap) *vec_oprnds = NULL;
3285 bool slp = (slp_node != NULL);
3286 unsigned int vec_num;
3287 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3290 loop = LOOP_VINFO_LOOP (loop_vinfo);
3292 /* Multiple types in SLP are handled by creating the appropriate number of
3293 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3298 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3300 gcc_assert (ncopies >= 1);
3302 /* FORNOW. This restriction should be relaxed. */
3303 if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3305 if (vect_print_dump_info (REPORT_DETAILS))
3306 fprintf (vect_dump, "multiple types in nested loop.");
3310 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3313 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3316 /* Is vectorizable store? */
3318 if (!is_gimple_assign (stmt))
3321 scalar_dest = gimple_assign_lhs (stmt);
3322 if (TREE_CODE (scalar_dest) != ARRAY_REF
3323 && TREE_CODE (scalar_dest) != INDIRECT_REF
3324 && TREE_CODE (scalar_dest) != COMPONENT_REF
3325 && TREE_CODE (scalar_dest) != IMAGPART_EXPR
3326 && TREE_CODE (scalar_dest) != REALPART_EXPR
3327 && TREE_CODE (scalar_dest) != MEM_REF)
3330 gcc_assert (gimple_assign_single_p (stmt));
3331 op = gimple_assign_rhs1 (stmt);
3332 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt))
3334 if (vect_print_dump_info (REPORT_DETAILS))
3335 fprintf (vect_dump, "use not simple.");
3339 /* The scalar rhs type needs to be trivially convertible to the vector
3340 component type. This should always be the case. */
3341 if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op)))
3343 if (vect_print_dump_info (REPORT_DETAILS))
3344 fprintf (vect_dump, "??? operands of different types");
3348 vec_mode = TYPE_MODE (vectype);
3349 /* FORNOW. In some cases can vectorize even if data-type not supported
3350 (e.g. - array initialization with 0). */
3351 if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing)
3354 if (!STMT_VINFO_DATA_REF (stmt_info))
3357 if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0)
3359 if (vect_print_dump_info (REPORT_DETAILS))
3360 fprintf (vect_dump, "negative step for store.");
3364 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3366 strided_store = true;
3367 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3368 if (!vect_strided_store_supported (vectype)
3369 && !PURE_SLP_STMT (stmt_info) && !slp)
3372 if (first_stmt == stmt)
3374 /* STMT is the leader of the group. Check the operands of all the
3375 stmts of the group. */
3376 next_stmt = DR_GROUP_NEXT_DR (stmt_info);
3379 gcc_assert (gimple_assign_single_p (next_stmt));
3380 op = gimple_assign_rhs1 (next_stmt);
3381 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt,
3384 if (vect_print_dump_info (REPORT_DETAILS))
3385 fprintf (vect_dump, "use not simple.");
3388 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3393 if (!vec_stmt) /* transformation not required. */
3395 STMT_VINFO_TYPE (stmt_info) = store_vec_info_type;
3396 vect_model_store_cost (stmt_info, ncopies, dt, NULL);
3404 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3405 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3407 DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++;
3410 gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt));
3412 /* We vectorize all the stmts of the interleaving group when we
3413 reach the last stmt in the group. */
3414 if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))
3415 < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt))
3424 strided_store = false;
3425 /* VEC_NUM is the number of vect stmts to be created for this
3427 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3428 first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0);
3429 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3432 /* VEC_NUM is the number of vect stmts to be created for this
3434 vec_num = group_size;
3440 group_size = vec_num = 1;
3443 if (vect_print_dump_info (REPORT_DETAILS))
3444 fprintf (vect_dump, "transform store. ncopies = %d",ncopies);
3446 dr_chain = VEC_alloc (tree, heap, group_size);
3447 oprnds = VEC_alloc (tree, heap, group_size);
3449 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3450 gcc_assert (alignment_support_scheme);
3452 /* In case the vectorization factor (VF) is bigger than the number
3453 of elements that we can fit in a vectype (nunits), we have to generate
3454 more than one vector stmt - i.e - we need to "unroll" the
3455 vector stmt by a factor VF/nunits. For more details see documentation in
3456 vect_get_vec_def_for_copy_stmt. */
3458 /* In case of interleaving (non-unit strided access):
3465 We create vectorized stores starting from base address (the access of the
3466 first stmt in the chain (S2 in the above example), when the last store stmt
3467 of the chain (S4) is reached:
3470 VS2: &base + vec_size*1 = vx0
3471 VS3: &base + vec_size*2 = vx1
3472 VS4: &base + vec_size*3 = vx3
3474 Then permutation statements are generated:
3476 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
3477 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
3480 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3481 (the order of the data-refs in the output of vect_permute_store_chain
3482 corresponds to the order of scalar stmts in the interleaving chain - see
3483 the documentation of vect_permute_store_chain()).
3485 In case of both multiple types and interleaving, above vector stores and
3486 permutation stmts are created for every copy. The result vector stmts are
3487 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
3488 STMT_VINFO_RELATED_STMT for the next copies.
3491 prev_stmt_info = NULL;
3492 for (j = 0; j < ncopies; j++)
3501 /* Get vectorized arguments for SLP_NODE. */
3502 vect_get_slp_defs (NULL_TREE, NULL_TREE, slp_node, &vec_oprnds,
3505 vec_oprnd = VEC_index (tree, vec_oprnds, 0);
3509 /* For interleaved stores we collect vectorized defs for all the
3510 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
3511 used as an input to vect_permute_store_chain(), and OPRNDS as
3512 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
3514 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3515 OPRNDS are of size 1. */
3516 next_stmt = first_stmt;
3517 for (i = 0; i < group_size; i++)
3519 /* Since gaps are not supported for interleaved stores,
3520 GROUP_SIZE is the exact number of stmts in the chain.
3521 Therefore, NEXT_STMT can't be NULL_TREE. In case that
3522 there is no interleaving, GROUP_SIZE is 1, and only one
3523 iteration of the loop will be executed. */
3524 gcc_assert (next_stmt
3525 && gimple_assign_single_p (next_stmt));
3526 op = gimple_assign_rhs1 (next_stmt);
3528 vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt,
3530 VEC_quick_push(tree, dr_chain, vec_oprnd);
3531 VEC_quick_push(tree, oprnds, vec_oprnd);
3532 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3536 /* We should have catched mismatched types earlier. */
3537 gcc_assert (useless_type_conversion_p (vectype,
3538 TREE_TYPE (vec_oprnd)));
3539 dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE,
3540 &dummy, &ptr_incr, false,
3542 gcc_assert (bb_vinfo || !inv_p);
3546 /* For interleaved stores we created vectorized defs for all the
3547 defs stored in OPRNDS in the previous iteration (previous copy).
3548 DR_CHAIN is then used as an input to vect_permute_store_chain(),
3549 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
3551 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3552 OPRNDS are of size 1. */
3553 for (i = 0; i < group_size; i++)
3555 op = VEC_index (tree, oprnds, i);
3556 vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def,
3558 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op);
3559 VEC_replace(tree, dr_chain, i, vec_oprnd);
3560 VEC_replace(tree, oprnds, i, vec_oprnd);
3563 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
3568 result_chain = VEC_alloc (tree, heap, group_size);
3570 if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi,
3575 next_stmt = first_stmt;
3576 for (i = 0; i < vec_num; i++)
3578 struct ptr_info_def *pi;
3581 /* Bump the vector pointer. */
3582 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
3586 vec_oprnd = VEC_index (tree, vec_oprnds, i);
3587 else if (strided_store)
3588 /* For strided stores vectorized defs are interleaved in
3589 vect_permute_store_chain(). */
3590 vec_oprnd = VEC_index (tree, result_chain, i);
3592 data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr,
3593 build_int_cst (reference_alias_ptr_type
3594 (DR_REF (first_dr)), 0));
3595 pi = get_ptr_info (dataref_ptr);
3596 pi->align = TYPE_ALIGN_UNIT (vectype);
3597 if (aligned_access_p (first_dr))
3599 else if (DR_MISALIGNMENT (first_dr) == -1)
3601 TREE_TYPE (data_ref)
3602 = build_aligned_type (TREE_TYPE (data_ref),
3603 TYPE_ALIGN (TREE_TYPE (vectype)));
3604 pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype));
3609 TREE_TYPE (data_ref)
3610 = build_aligned_type (TREE_TYPE (data_ref),
3611 TYPE_ALIGN (TREE_TYPE (vectype)));
3612 pi->misalign = DR_MISALIGNMENT (first_dr);
3615 /* Arguments are ready. Create the new vector stmt. */
3616 new_stmt = gimple_build_assign (data_ref, vec_oprnd);
3617 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3618 mark_symbols_for_renaming (new_stmt);
3624 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3626 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3628 prev_stmt_info = vinfo_for_stmt (new_stmt);
3629 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3635 VEC_free (tree, heap, dr_chain);
3636 VEC_free (tree, heap, oprnds);
3638 VEC_free (tree, heap, result_chain);
3640 VEC_free (tree, heap, vec_oprnds);
3645 /* Given a vector type VECTYPE returns a builtin DECL to be used
3646 for vector permutation and stores a mask into *MASK that implements
3647 reversal of the vector elements. If that is impossible to do
3648 returns NULL (and *MASK is unchanged). */
3651 perm_mask_for_reverse (tree vectype, tree *mask)
3654 tree mask_element_type, mask_type;
3655 tree mask_vec = NULL;
3658 if (!targetm.vectorize.builtin_vec_perm)
3661 builtin_decl = targetm.vectorize.builtin_vec_perm (vectype,
3662 &mask_element_type);
3663 if (!builtin_decl || !mask_element_type)
3666 mask_type = get_vectype_for_scalar_type (mask_element_type);
3667 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3669 || TYPE_VECTOR_SUBPARTS (vectype) != TYPE_VECTOR_SUBPARTS (mask_type))
3672 for (i = 0; i < nunits; i++)
3673 mask_vec = tree_cons (NULL, build_int_cst (mask_element_type, i), mask_vec);
3674 mask_vec = build_vector (mask_type, mask_vec);
3676 if (!targetm.vectorize.builtin_vec_perm_ok (vectype, mask_vec))
3680 return builtin_decl;
3683 /* Given a vector variable X, that was generated for the scalar LHS of
3684 STMT, generate instructions to reverse the vector elements of X,
3685 insert them a *GSI and return the permuted vector variable. */
3688 reverse_vec_elements (tree x, gimple stmt, gimple_stmt_iterator *gsi)
3690 tree vectype = TREE_TYPE (x);
3691 tree mask_vec, builtin_decl;
3692 tree perm_dest, data_ref;
3695 builtin_decl = perm_mask_for_reverse (vectype, &mask_vec);
3697 perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype);
3699 /* Generate the permute statement. */
3700 perm_stmt = gimple_build_call (builtin_decl, 3, x, x, mask_vec);
3701 data_ref = make_ssa_name (perm_dest, perm_stmt);
3702 gimple_call_set_lhs (perm_stmt, data_ref);
3703 vect_finish_stmt_generation (stmt, perm_stmt, gsi);
3708 /* vectorizable_load.
3710 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
3712 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3713 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3714 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3717 vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3718 slp_tree slp_node, slp_instance slp_node_instance)
3721 tree vec_dest = NULL;
3722 tree data_ref = NULL;
3723 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3724 stmt_vec_info prev_stmt_info;
3725 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3726 struct loop *loop = NULL;
3727 struct loop *containing_loop = (gimple_bb (stmt))->loop_father;
3728 bool nested_in_vect_loop = false;
3729 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
3730 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3732 enum machine_mode mode;
3733 gimple new_stmt = NULL;
3735 enum dr_alignment_support alignment_support_scheme;
3736 tree dataref_ptr = NULL_TREE;
3738 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3740 int i, j, group_size;
3741 tree msq = NULL_TREE, lsq;
3742 tree offset = NULL_TREE;
3743 tree realignment_token = NULL_TREE;
3745 VEC(tree,heap) *dr_chain = NULL;
3746 bool strided_load = false;
3751 bool compute_in_loop = false;
3752 struct loop *at_loop;
3754 bool slp = (slp_node != NULL);
3755 bool slp_perm = false;
3756 enum tree_code code;
3757 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3762 loop = LOOP_VINFO_LOOP (loop_vinfo);
3763 nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
3764 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
3769 /* Multiple types in SLP are handled by creating the appropriate number of
3770 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3775 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3777 gcc_assert (ncopies >= 1);
3779 /* FORNOW. This restriction should be relaxed. */
3780 if (nested_in_vect_loop && ncopies > 1)
3782 if (vect_print_dump_info (REPORT_DETAILS))
3783 fprintf (vect_dump, "multiple types in nested loop.");
3787 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3790 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3793 /* Is vectorizable load? */
3794 if (!is_gimple_assign (stmt))
3797 scalar_dest = gimple_assign_lhs (stmt);
3798 if (TREE_CODE (scalar_dest) != SSA_NAME)
3801 code = gimple_assign_rhs_code (stmt);
3802 if (code != ARRAY_REF
3803 && code != INDIRECT_REF
3804 && code != COMPONENT_REF
3805 && code != IMAGPART_EXPR
3806 && code != REALPART_EXPR
3810 if (!STMT_VINFO_DATA_REF (stmt_info))
3813 negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0;
3814 if (negative && ncopies > 1)
3816 if (vect_print_dump_info (REPORT_DETAILS))
3817 fprintf (vect_dump, "multiple types with negative step.");
3821 scalar_type = TREE_TYPE (DR_REF (dr));
3822 mode = TYPE_MODE (vectype);
3824 /* FORNOW. In some cases can vectorize even if data-type not supported
3825 (e.g. - data copies). */
3826 if (optab_handler (mov_optab, mode) == CODE_FOR_nothing)
3828 if (vect_print_dump_info (REPORT_DETAILS))
3829 fprintf (vect_dump, "Aligned load, but unsupported type.");
3833 /* The vector component type needs to be trivially convertible to the
3834 scalar lhs. This should always be the case. */
3835 if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype)))
3837 if (vect_print_dump_info (REPORT_DETAILS))
3838 fprintf (vect_dump, "??? operands of different types");
3842 /* Check if the load is a part of an interleaving chain. */
3843 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3845 strided_load = true;
3847 gcc_assert (! nested_in_vect_loop);
3849 /* Check if interleaving is supported. */
3850 if (!vect_strided_load_supported (vectype)
3851 && !PURE_SLP_STMT (stmt_info) && !slp)
3857 gcc_assert (!strided_load);
3858 alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
3859 if (alignment_support_scheme != dr_aligned
3860 && alignment_support_scheme != dr_unaligned_supported)
3862 if (vect_print_dump_info (REPORT_DETAILS))
3863 fprintf (vect_dump, "negative step but alignment required.");
3866 if (!perm_mask_for_reverse (vectype, NULL))
3868 if (vect_print_dump_info (REPORT_DETAILS))
3869 fprintf (vect_dump, "negative step and reversing not supported.");
3874 if (!vec_stmt) /* transformation not required. */
3876 STMT_VINFO_TYPE (stmt_info) = load_vec_info_type;
3877 vect_model_load_cost (stmt_info, ncopies, NULL);
3881 if (vect_print_dump_info (REPORT_DETAILS))
3882 fprintf (vect_dump, "transform load.");
3888 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3889 /* Check if the chain of loads is already vectorized. */
3890 if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)))
3892 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3895 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3896 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3898 /* VEC_NUM is the number of vect stmts to be created for this group. */
3901 strided_load = false;
3902 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3903 if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance))
3907 vec_num = group_size;
3909 dr_chain = VEC_alloc (tree, heap, vec_num);
3915 group_size = vec_num = 1;
3918 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3919 gcc_assert (alignment_support_scheme);
3921 /* In case the vectorization factor (VF) is bigger than the number
3922 of elements that we can fit in a vectype (nunits), we have to generate
3923 more than one vector stmt - i.e - we need to "unroll" the
3924 vector stmt by a factor VF/nunits. In doing so, we record a pointer
3925 from one copy of the vector stmt to the next, in the field
3926 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
3927 stages to find the correct vector defs to be used when vectorizing
3928 stmts that use the defs of the current stmt. The example below
3929 illustrates the vectorization process when VF=16 and nunits=4 (i.e., we
3930 need to create 4 vectorized stmts):
3932 before vectorization:
3933 RELATED_STMT VEC_STMT
3937 step 1: vectorize stmt S1:
3938 We first create the vector stmt VS1_0, and, as usual, record a
3939 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
3940 Next, we create the vector stmt VS1_1, and record a pointer to
3941 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
3942 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
3944 RELATED_STMT VEC_STMT
3945 VS1_0: vx0 = memref0 VS1_1 -
3946 VS1_1: vx1 = memref1 VS1_2 -
3947 VS1_2: vx2 = memref2 VS1_3 -
3948 VS1_3: vx3 = memref3 - -
3949 S1: x = load - VS1_0
3952 See in documentation in vect_get_vec_def_for_stmt_copy for how the
3953 information we recorded in RELATED_STMT field is used to vectorize
3956 /* In case of interleaving (non-unit strided access):
3963 Vectorized loads are created in the order of memory accesses
3964 starting from the access of the first stmt of the chain:
3967 VS2: vx1 = &base + vec_size*1
3968 VS3: vx3 = &base + vec_size*2
3969 VS4: vx4 = &base + vec_size*3
3971 Then permutation statements are generated:
3973 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
3974 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
3977 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3978 (the order of the data-refs in the output of vect_permute_load_chain
3979 corresponds to the order of scalar stmts in the interleaving chain - see
3980 the documentation of vect_permute_load_chain()).
3981 The generation of permutation stmts and recording them in
3982 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
3984 In case of both multiple types and interleaving, the vector loads and
3985 permutation stmts above are created for every copy. The result vector
3986 stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the
3987 corresponding STMT_VINFO_RELATED_STMT for the next copies. */
3989 /* If the data reference is aligned (dr_aligned) or potentially unaligned
3990 on a target that supports unaligned accesses (dr_unaligned_supported)
3991 we generate the following code:
3995 p = p + indx * vectype_size;
4000 Otherwise, the data reference is potentially unaligned on a target that
4001 does not support unaligned accesses (dr_explicit_realign_optimized) -
4002 then generate the following code, in which the data in each iteration is
4003 obtained by two vector loads, one from the previous iteration, and one
4004 from the current iteration:
4006 msq_init = *(floor(p1))
4007 p2 = initial_addr + VS - 1;
4008 realignment_token = call target_builtin;
4011 p2 = p2 + indx * vectype_size
4013 vec_dest = realign_load (msq, lsq, realignment_token)
4018 /* If the misalignment remains the same throughout the execution of the
4019 loop, we can create the init_addr and permutation mask at the loop
4020 preheader. Otherwise, it needs to be created inside the loop.
4021 This can only occur when vectorizing memory accesses in the inner-loop
4022 nested within an outer-loop that is being vectorized. */
4024 if (loop && nested_in_vect_loop_p (loop, stmt)
4025 && (TREE_INT_CST_LOW (DR_STEP (dr))
4026 % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0))
4028 gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized);
4029 compute_in_loop = true;
4032 if ((alignment_support_scheme == dr_explicit_realign_optimized
4033 || alignment_support_scheme == dr_explicit_realign)
4034 && !compute_in_loop)
4036 msq = vect_setup_realignment (first_stmt, gsi, &realignment_token,
4037 alignment_support_scheme, NULL_TREE,
4039 if (alignment_support_scheme == dr_explicit_realign_optimized)
4041 phi = SSA_NAME_DEF_STMT (msq);
4042 offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
4049 offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1);
4051 prev_stmt_info = NULL;
4052 for (j = 0; j < ncopies; j++)
4054 /* 1. Create the vector pointer update chain. */
4056 dataref_ptr = vect_create_data_ref_ptr (first_stmt,
4058 &dummy, &ptr_incr, false,
4062 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
4064 for (i = 0; i < vec_num; i++)
4067 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
4070 /* 2. Create the vector-load in the loop. */
4071 switch (alignment_support_scheme)
4074 case dr_unaligned_supported:
4076 struct ptr_info_def *pi;
4078 = build2 (MEM_REF, vectype, dataref_ptr,
4079 build_int_cst (reference_alias_ptr_type
4080 (DR_REF (first_dr)), 0));
4081 pi = get_ptr_info (dataref_ptr);
4082 pi->align = TYPE_ALIGN_UNIT (vectype);
4083 if (alignment_support_scheme == dr_aligned)
4085 gcc_assert (aligned_access_p (first_dr));
4088 else if (DR_MISALIGNMENT (first_dr) == -1)
4090 TREE_TYPE (data_ref)
4091 = build_aligned_type (TREE_TYPE (data_ref),
4092 TYPE_ALIGN (TREE_TYPE (vectype)));
4093 pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype));
4098 TREE_TYPE (data_ref)
4099 = build_aligned_type (TREE_TYPE (data_ref),
4100 TYPE_ALIGN (TREE_TYPE (vectype)));
4101 pi->misalign = DR_MISALIGNMENT (first_dr);
4105 case dr_explicit_realign:
4108 tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
4110 if (compute_in_loop)
4111 msq = vect_setup_realignment (first_stmt, gsi,
4113 dr_explicit_realign,
4116 new_stmt = gimple_build_assign_with_ops
4117 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
4119 (TREE_TYPE (dataref_ptr),
4120 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
4121 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
4122 gimple_assign_set_lhs (new_stmt, ptr);
4123 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4125 = build2 (MEM_REF, vectype, ptr,
4126 build_int_cst (reference_alias_ptr_type
4127 (DR_REF (first_dr)), 0));
4128 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4129 new_stmt = gimple_build_assign (vec_dest, data_ref);
4130 new_temp = make_ssa_name (vec_dest, new_stmt);
4131 gimple_assign_set_lhs (new_stmt, new_temp);
4132 gimple_set_vdef (new_stmt, gimple_vdef (stmt));
4133 gimple_set_vuse (new_stmt, gimple_vuse (stmt));
4134 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4137 bump = size_binop (MULT_EXPR, vs_minus_1,
4138 TYPE_SIZE_UNIT (scalar_type));
4139 ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump);
4140 new_stmt = gimple_build_assign_with_ops
4141 (BIT_AND_EXPR, NULL_TREE, ptr,
4144 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
4145 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
4146 gimple_assign_set_lhs (new_stmt, ptr);
4147 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4149 = build2 (MEM_REF, vectype, ptr,
4150 build_int_cst (reference_alias_ptr_type
4151 (DR_REF (first_dr)), 0));
4154 case dr_explicit_realign_optimized:
4155 new_stmt = gimple_build_assign_with_ops
4156 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
4158 (TREE_TYPE (dataref_ptr),
4159 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
4160 new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
4161 gimple_assign_set_lhs (new_stmt, new_temp);
4162 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4164 = build2 (MEM_REF, vectype, new_temp,
4165 build_int_cst (reference_alias_ptr_type
4166 (DR_REF (first_dr)), 0));
4171 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4172 new_stmt = gimple_build_assign (vec_dest, data_ref);
4173 new_temp = make_ssa_name (vec_dest, new_stmt);
4174 gimple_assign_set_lhs (new_stmt, new_temp);
4175 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4176 mark_symbols_for_renaming (new_stmt);
4178 /* 3. Handle explicit realignment if necessary/supported. Create in
4179 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
4180 if (alignment_support_scheme == dr_explicit_realign_optimized
4181 || alignment_support_scheme == dr_explicit_realign)
4185 lsq = gimple_assign_lhs (new_stmt);
4186 if (!realignment_token)
4187 realignment_token = dataref_ptr;
4188 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4189 tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq,
4191 new_stmt = gimple_build_assign (vec_dest, tmp);
4192 new_temp = make_ssa_name (vec_dest, new_stmt);
4193 gimple_assign_set_lhs (new_stmt, new_temp);
4194 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4196 if (alignment_support_scheme == dr_explicit_realign_optimized)
4199 if (i == vec_num - 1 && j == ncopies - 1)
4200 add_phi_arg (phi, lsq, loop_latch_edge (containing_loop),
4206 /* 4. Handle invariant-load. */
4207 if (inv_p && !bb_vinfo)
4209 gcc_assert (!strided_load);
4210 gcc_assert (nested_in_vect_loop_p (loop, stmt));
4215 tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type);
4217 /* CHECKME: bitpos depends on endianess? */
4218 bitpos = bitsize_zero_node;
4219 vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp,
4222 vect_create_destination_var (scalar_dest, NULL_TREE);
4223 new_stmt = gimple_build_assign (vec_dest, vec_inv);
4224 new_temp = make_ssa_name (vec_dest, new_stmt);
4225 gimple_assign_set_lhs (new_stmt, new_temp);
4226 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4228 for (k = nunits - 1; k >= 0; --k)
4229 t = tree_cons (NULL_TREE, new_temp, t);
4230 /* FIXME: use build_constructor directly. */
4231 vec_inv = build_constructor_from_list (vectype, t);
4232 new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi);
4233 new_stmt = SSA_NAME_DEF_STMT (new_temp);
4236 gcc_unreachable (); /* FORNOW. */
4241 new_temp = reverse_vec_elements (new_temp, stmt, gsi);
4242 new_stmt = SSA_NAME_DEF_STMT (new_temp);
4245 /* Collect vector loads and later create their permutation in
4246 vect_transform_strided_load (). */
4247 if (strided_load || slp_perm)
4248 VEC_quick_push (tree, dr_chain, new_temp);
4250 /* Store vector loads in the corresponding SLP_NODE. */
4251 if (slp && !slp_perm)
4252 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
4255 if (slp && !slp_perm)
4260 if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf,
4261 slp_node_instance, false))
4263 VEC_free (tree, heap, dr_chain);
4271 if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi))
4274 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
4275 VEC_free (tree, heap, dr_chain);
4276 dr_chain = VEC_alloc (tree, heap, group_size);
4281 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4283 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4284 prev_stmt_info = vinfo_for_stmt (new_stmt);
4290 VEC_free (tree, heap, dr_chain);
4295 /* Function vect_is_simple_cond.
4298 LOOP - the loop that is being vectorized.
4299 COND - Condition that is checked for simple use.
4301 Returns whether a COND can be vectorized. Checks whether
4302 condition operands are supportable using vec_is_simple_use. */
4305 vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo)
4309 enum vect_def_type dt;
4311 if (!COMPARISON_CLASS_P (cond))
4314 lhs = TREE_OPERAND (cond, 0);
4315 rhs = TREE_OPERAND (cond, 1);
4317 if (TREE_CODE (lhs) == SSA_NAME)
4319 gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs);
4320 if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def,
4324 else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST
4325 && TREE_CODE (lhs) != FIXED_CST)
4328 if (TREE_CODE (rhs) == SSA_NAME)
4330 gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs);
4331 if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def,
4335 else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST
4336 && TREE_CODE (rhs) != FIXED_CST)
4342 /* vectorizable_condition.
4344 Check if STMT is conditional modify expression that can be vectorized.
4345 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4346 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
4349 When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable
4350 to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in
4351 else caluse if it is 2).
4353 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4356 vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi,
4357 gimple *vec_stmt, tree reduc_def, int reduc_index)
4359 tree scalar_dest = NULL_TREE;
4360 tree vec_dest = NULL_TREE;
4361 tree op = NULL_TREE;
4362 tree cond_expr, then_clause, else_clause;
4363 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4364 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4365 tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE;
4366 tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE;
4367 tree vec_compare, vec_cond_expr;
4369 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4370 enum machine_mode vec_mode;
4372 enum vect_def_type dt, dts[4];
4373 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
4374 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
4375 enum tree_code code;
4376 stmt_vec_info prev_stmt_info = NULL;
4379 /* FORNOW: unsupported in basic block SLP. */
4380 gcc_assert (loop_vinfo);
4382 gcc_assert (ncopies >= 1);
4383 if (reduc_index && ncopies > 1)
4384 return false; /* FORNOW */
4386 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4389 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
4390 && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle
4394 /* FORNOW: SLP not supported. */
4395 if (STMT_SLP_TYPE (stmt_info))
4398 /* FORNOW: not yet supported. */
4399 if (STMT_VINFO_LIVE_P (stmt_info))
4401 if (vect_print_dump_info (REPORT_DETAILS))
4402 fprintf (vect_dump, "value used after loop.");
4406 /* Is vectorizable conditional operation? */
4407 if (!is_gimple_assign (stmt))
4410 code = gimple_assign_rhs_code (stmt);
4412 if (code != COND_EXPR)
4415 gcc_assert (gimple_assign_single_p (stmt));
4416 op = gimple_assign_rhs1 (stmt);
4417 cond_expr = TREE_OPERAND (op, 0);
4418 then_clause = TREE_OPERAND (op, 1);
4419 else_clause = TREE_OPERAND (op, 2);
4421 if (!vect_is_simple_cond (cond_expr, loop_vinfo))
4424 /* We do not handle two different vector types for the condition
4426 if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)),
4427 TREE_TYPE (vectype)))
4430 if (TREE_CODE (then_clause) == SSA_NAME)
4432 gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause);
4433 if (!vect_is_simple_use (then_clause, loop_vinfo, NULL,
4434 &then_def_stmt, &def, &dt))
4437 else if (TREE_CODE (then_clause) != INTEGER_CST
4438 && TREE_CODE (then_clause) != REAL_CST
4439 && TREE_CODE (then_clause) != FIXED_CST)
4442 if (TREE_CODE (else_clause) == SSA_NAME)
4444 gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause);
4445 if (!vect_is_simple_use (else_clause, loop_vinfo, NULL,
4446 &else_def_stmt, &def, &dt))
4449 else if (TREE_CODE (else_clause) != INTEGER_CST
4450 && TREE_CODE (else_clause) != REAL_CST
4451 && TREE_CODE (else_clause) != FIXED_CST)
4455 vec_mode = TYPE_MODE (vectype);
4459 STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type;
4460 return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode);
4466 scalar_dest = gimple_assign_lhs (stmt);
4467 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4469 /* Handle cond expr. */
4470 for (j = 0; j < ncopies; j++)
4477 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0),
4479 vect_is_simple_use (TREE_OPERAND (cond_expr, 0), loop_vinfo,
4480 NULL, >emp, &def, &dts[0]);
4482 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1),
4484 vect_is_simple_use (TREE_OPERAND (cond_expr, 1), loop_vinfo,
4485 NULL, >emp, &def, &dts[1]);
4486 if (reduc_index == 1)
4487 vec_then_clause = reduc_def;
4490 vec_then_clause = vect_get_vec_def_for_operand (then_clause,
4492 vect_is_simple_use (then_clause, loop_vinfo,
4493 NULL, >emp, &def, &dts[2]);
4495 if (reduc_index == 2)
4496 vec_else_clause = reduc_def;
4499 vec_else_clause = vect_get_vec_def_for_operand (else_clause,
4501 vect_is_simple_use (else_clause, loop_vinfo,
4502 NULL, >emp, &def, &dts[3]);
4507 vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], vec_cond_lhs);
4508 vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], vec_cond_rhs);
4509 vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2],
4511 vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3],
4515 /* Arguments are ready. Create the new vector stmt. */
4516 vec_compare = build2 (TREE_CODE (cond_expr), vectype,
4517 vec_cond_lhs, vec_cond_rhs);
4518 vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
4519 vec_compare, vec_then_clause, vec_else_clause);
4521 new_stmt = gimple_build_assign (vec_dest, vec_cond_expr);
4522 new_temp = make_ssa_name (vec_dest, new_stmt);
4523 gimple_assign_set_lhs (new_stmt, new_temp);
4524 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4526 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4528 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4530 prev_stmt_info = vinfo_for_stmt (new_stmt);
4537 /* Make sure the statement is vectorizable. */
4540 vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node)
4542 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4543 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
4544 enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info);
4546 tree scalar_type, vectype;
4548 if (vect_print_dump_info (REPORT_DETAILS))
4550 fprintf (vect_dump, "==> examining statement: ");
4551 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4554 if (gimple_has_volatile_ops (stmt))
4556 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4557 fprintf (vect_dump, "not vectorized: stmt has volatile operands");
4562 /* Skip stmts that do not need to be vectorized. In loops this is expected
4564 - the COND_EXPR which is the loop exit condition
4565 - any LABEL_EXPRs in the loop
4566 - computations that are used only for array indexing or loop control.
4567 In basic blocks we only analyze statements that are a part of some SLP
4568 instance, therefore, all the statements are relevant. */
4570 if (!STMT_VINFO_RELEVANT_P (stmt_info)
4571 && !STMT_VINFO_LIVE_P (stmt_info))
4573 if (vect_print_dump_info (REPORT_DETAILS))
4574 fprintf (vect_dump, "irrelevant.");
4579 switch (STMT_VINFO_DEF_TYPE (stmt_info))
4581 case vect_internal_def:
4584 case vect_reduction_def:
4585 case vect_nested_cycle:
4586 gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer
4587 || relevance == vect_used_in_outer_by_reduction
4588 || relevance == vect_unused_in_scope));
4591 case vect_induction_def:
4592 case vect_constant_def:
4593 case vect_external_def:
4594 case vect_unknown_def_type:
4601 gcc_assert (PURE_SLP_STMT (stmt_info));
4603 scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
4604 if (vect_print_dump_info (REPORT_DETAILS))
4606 fprintf (vect_dump, "get vectype for scalar type: ");
4607 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4610 vectype = get_vectype_for_scalar_type (scalar_type);
4613 if (vect_print_dump_info (REPORT_DETAILS))
4615 fprintf (vect_dump, "not SLPed: unsupported data-type ");
4616 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4621 if (vect_print_dump_info (REPORT_DETAILS))
4623 fprintf (vect_dump, "vectype: ");
4624 print_generic_expr (vect_dump, vectype, TDF_SLIM);
4627 STMT_VINFO_VECTYPE (stmt_info) = vectype;
4630 if (STMT_VINFO_RELEVANT_P (stmt_info))
4632 gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt))));
4633 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
4634 *need_to_vectorize = true;
4639 && (STMT_VINFO_RELEVANT_P (stmt_info)
4640 || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def))
4641 ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL)
4642 || vectorizable_type_demotion (stmt, NULL, NULL, NULL)
4643 || vectorizable_conversion (stmt, NULL, NULL, NULL)
4644 || vectorizable_shift (stmt, NULL, NULL, NULL)
4645 || vectorizable_operation (stmt, NULL, NULL, NULL)
4646 || vectorizable_assignment (stmt, NULL, NULL, NULL)
4647 || vectorizable_load (stmt, NULL, NULL, NULL, NULL)
4648 || vectorizable_call (stmt, NULL, NULL)
4649 || vectorizable_store (stmt, NULL, NULL, NULL)
4650 || vectorizable_reduction (stmt, NULL, NULL, NULL)
4651 || vectorizable_condition (stmt, NULL, NULL, NULL, 0));
4655 ok = (vectorizable_shift (stmt, NULL, NULL, node)
4656 || vectorizable_operation (stmt, NULL, NULL, node)
4657 || vectorizable_assignment (stmt, NULL, NULL, node)
4658 || vectorizable_load (stmt, NULL, NULL, node, NULL)
4659 || vectorizable_store (stmt, NULL, NULL, node));
4664 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4666 fprintf (vect_dump, "not vectorized: relevant stmt not ");
4667 fprintf (vect_dump, "supported: ");
4668 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4677 /* Stmts that are (also) "live" (i.e. - that are used out of the loop)
4678 need extra handling, except for vectorizable reductions. */
4679 if (STMT_VINFO_LIVE_P (stmt_info)
4680 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4681 ok = vectorizable_live_operation (stmt, NULL, NULL);
4685 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4687 fprintf (vect_dump, "not vectorized: live stmt not ");
4688 fprintf (vect_dump, "supported: ");
4689 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4695 if (!PURE_SLP_STMT (stmt_info))
4697 /* Groups of strided accesses whose size is not a power of 2 are not
4698 vectorizable yet using loop-vectorization. Therefore, if this stmt
4699 feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and
4700 loop-based vectorized), the loop cannot be vectorized. */
4701 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
4702 && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt (
4703 DR_GROUP_FIRST_DR (stmt_info)))) == -1)
4705 if (vect_print_dump_info (REPORT_DETAILS))
4707 fprintf (vect_dump, "not vectorized: the size of group "
4708 "of strided accesses is not a power of 2");
4709 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4720 /* Function vect_transform_stmt.
4722 Create a vectorized stmt to replace STMT, and insert it at BSI. */
4725 vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi,
4726 bool *strided_store, slp_tree slp_node,
4727 slp_instance slp_node_instance)
4729 bool is_store = false;
4730 gimple vec_stmt = NULL;
4731 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4732 gimple orig_stmt_in_pattern, orig_scalar_stmt = stmt;
4735 switch (STMT_VINFO_TYPE (stmt_info))
4737 case type_demotion_vec_info_type:
4738 done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node);
4742 case type_promotion_vec_info_type:
4743 done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node);
4747 case type_conversion_vec_info_type:
4748 done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node);
4752 case induc_vec_info_type:
4753 gcc_assert (!slp_node);
4754 done = vectorizable_induction (stmt, gsi, &vec_stmt);
4758 case shift_vec_info_type:
4759 done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node);
4763 case op_vec_info_type:
4764 done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node);
4768 case assignment_vec_info_type:
4769 done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node);
4773 case load_vec_info_type:
4774 done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node,
4779 case store_vec_info_type:
4780 done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node);
4782 if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node)
4784 /* In case of interleaving, the whole chain is vectorized when the
4785 last store in the chain is reached. Store stmts before the last
4786 one are skipped, and there vec_stmt_info shouldn't be freed
4788 *strided_store = true;
4789 if (STMT_VINFO_VEC_STMT (stmt_info))
4796 case condition_vec_info_type:
4797 gcc_assert (!slp_node);
4798 done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0);
4802 case call_vec_info_type:
4803 gcc_assert (!slp_node);
4804 done = vectorizable_call (stmt, gsi, &vec_stmt);
4805 stmt = gsi_stmt (*gsi);
4808 case reduc_vec_info_type:
4809 done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node);
4814 if (!STMT_VINFO_LIVE_P (stmt_info))
4816 if (vect_print_dump_info (REPORT_DETAILS))
4817 fprintf (vect_dump, "stmt not supported.");
4822 /* Handle inner-loop stmts whose DEF is used in the loop-nest that
4823 is being vectorized, but outside the immediately enclosing loop. */
4825 && STMT_VINFO_LOOP_VINFO (stmt_info)
4826 && nested_in_vect_loop_p (LOOP_VINFO_LOOP (
4827 STMT_VINFO_LOOP_VINFO (stmt_info)), stmt)
4828 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type
4829 && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer
4830 || STMT_VINFO_RELEVANT (stmt_info) ==
4831 vect_used_in_outer_by_reduction))
4833 struct loop *innerloop = LOOP_VINFO_LOOP (
4834 STMT_VINFO_LOOP_VINFO (stmt_info))->inner;
4835 imm_use_iterator imm_iter;
4836 use_operand_p use_p;
4840 if (vect_print_dump_info (REPORT_DETAILS))
4841 fprintf (vect_dump, "Record the vdef for outer-loop vectorization.");
4843 /* Find the relevant loop-exit phi-node, and reord the vec_stmt there
4844 (to be used when vectorizing outer-loop stmts that use the DEF of
4846 if (gimple_code (stmt) == GIMPLE_PHI)
4847 scalar_dest = PHI_RESULT (stmt);
4849 scalar_dest = gimple_assign_lhs (stmt);
4851 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4853 if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p))))
4855 exit_phi = USE_STMT (use_p);
4856 STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt;
4861 /* Handle stmts whose DEF is used outside the loop-nest that is
4862 being vectorized. */
4863 if (STMT_VINFO_LIVE_P (stmt_info)
4864 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4866 done = vectorizable_live_operation (stmt, gsi, &vec_stmt);
4872 STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
4873 orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info);
4874 if (orig_stmt_in_pattern)
4876 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern);
4877 /* STMT was inserted by the vectorizer to replace a computation idiom.
4878 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
4879 computed this idiom. We need to record a pointer to VEC_STMT in
4880 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
4881 documentation of vect_pattern_recog. */
4882 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
4884 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo)
4885 == orig_scalar_stmt);
4886 STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
4895 /* Remove a group of stores (for SLP or interleaving), free their
4899 vect_remove_stores (gimple first_stmt)
4901 gimple next = first_stmt;
4903 gimple_stmt_iterator next_si;
4907 /* Free the attached stmt_vec_info and remove the stmt. */
4908 next_si = gsi_for_stmt (next);
4909 gsi_remove (&next_si, true);
4910 tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next));
4911 free_stmt_vec_info (next);
4917 /* Function new_stmt_vec_info.
4919 Create and initialize a new stmt_vec_info struct for STMT. */
4922 new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo,
4923 bb_vec_info bb_vinfo)
4926 res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info));
4928 STMT_VINFO_TYPE (res) = undef_vec_info_type;
4929 STMT_VINFO_STMT (res) = stmt;
4930 STMT_VINFO_LOOP_VINFO (res) = loop_vinfo;
4931 STMT_VINFO_BB_VINFO (res) = bb_vinfo;
4932 STMT_VINFO_RELEVANT (res) = vect_unused_in_scope;
4933 STMT_VINFO_LIVE_P (res) = false;
4934 STMT_VINFO_VECTYPE (res) = NULL;
4935 STMT_VINFO_VEC_STMT (res) = NULL;
4936 STMT_VINFO_VECTORIZABLE (res) = true;
4937 STMT_VINFO_IN_PATTERN_P (res) = false;
4938 STMT_VINFO_RELATED_STMT (res) = NULL;
4939 STMT_VINFO_DATA_REF (res) = NULL;
4941 STMT_VINFO_DR_BASE_ADDRESS (res) = NULL;
4942 STMT_VINFO_DR_OFFSET (res) = NULL;
4943 STMT_VINFO_DR_INIT (res) = NULL;
4944 STMT_VINFO_DR_STEP (res) = NULL;
4945 STMT_VINFO_DR_ALIGNED_TO (res) = NULL;
4947 if (gimple_code (stmt) == GIMPLE_PHI
4948 && is_loop_header_bb_p (gimple_bb (stmt)))
4949 STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
4951 STMT_VINFO_DEF_TYPE (res) = vect_internal_def;
4953 STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5);
4954 STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0;
4955 STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0;
4956 STMT_SLP_TYPE (res) = loop_vect;
4957 DR_GROUP_FIRST_DR (res) = NULL;
4958 DR_GROUP_NEXT_DR (res) = NULL;
4959 DR_GROUP_SIZE (res) = 0;
4960 DR_GROUP_STORE_COUNT (res) = 0;
4961 DR_GROUP_GAP (res) = 0;
4962 DR_GROUP_SAME_DR_STMT (res) = NULL;
4963 DR_GROUP_READ_WRITE_DEPENDENCE (res) = false;
4969 /* Create a hash table for stmt_vec_info. */
4972 init_stmt_vec_info_vec (void)
4974 gcc_assert (!stmt_vec_info_vec);
4975 stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50);
4979 /* Free hash table for stmt_vec_info. */
4982 free_stmt_vec_info_vec (void)
4984 gcc_assert (stmt_vec_info_vec);
4985 VEC_free (vec_void_p, heap, stmt_vec_info_vec);
4989 /* Free stmt vectorization related info. */
4992 free_stmt_vec_info (gimple stmt)
4994 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4999 VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info));
5000 set_vinfo_for_stmt (stmt, NULL);
5005 /* Function get_vectype_for_scalar_type_and_size.
5007 Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported
5011 get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size)
5013 enum machine_mode inner_mode = TYPE_MODE (scalar_type);
5014 enum machine_mode simd_mode;
5015 unsigned int nbytes = GET_MODE_SIZE (inner_mode);
5022 /* We can't build a vector type of elements with alignment bigger than
5024 if (nbytes < TYPE_ALIGN_UNIT (scalar_type))
5027 /* If we'd build a vector type of elements whose mode precision doesn't
5028 match their types precision we'll get mismatched types on vector
5029 extracts via BIT_FIELD_REFs. This effectively means we disable
5030 vectorization of bool and/or enum types in some languages. */
5031 if (INTEGRAL_TYPE_P (scalar_type)
5032 && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type))
5035 if (GET_MODE_CLASS (inner_mode) != MODE_INT
5036 && GET_MODE_CLASS (inner_mode) != MODE_FLOAT)
5039 /* If no size was supplied use the mode the target prefers. Otherwise
5040 lookup a vector mode of the specified size. */
5042 simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode);
5044 simd_mode = mode_for_vector (inner_mode, size / nbytes);
5045 nunits = GET_MODE_SIZE (simd_mode) / nbytes;
5049 vectype = build_vector_type (scalar_type, nunits);
5050 if (vect_print_dump_info (REPORT_DETAILS))
5052 fprintf (vect_dump, "get vectype with %d units of type ", nunits);
5053 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
5059 if (vect_print_dump_info (REPORT_DETAILS))
5061 fprintf (vect_dump, "vectype: ");
5062 print_generic_expr (vect_dump, vectype, TDF_SLIM);
5065 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
5066 && !INTEGRAL_MODE_P (TYPE_MODE (vectype)))
5068 if (vect_print_dump_info (REPORT_DETAILS))
5069 fprintf (vect_dump, "mode not supported by target.");
5076 unsigned int current_vector_size;
5078 /* Function get_vectype_for_scalar_type.
5080 Returns the vector type corresponding to SCALAR_TYPE as supported
5084 get_vectype_for_scalar_type (tree scalar_type)
5087 vectype = get_vectype_for_scalar_type_and_size (scalar_type,
5088 current_vector_size);
5090 && current_vector_size == 0)
5091 current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype));
5095 /* Function get_same_sized_vectype
5097 Returns a vector type corresponding to SCALAR_TYPE of size
5098 VECTOR_TYPE if supported by the target. */
5101 get_same_sized_vectype (tree scalar_type, tree vector_type)
5103 return get_vectype_for_scalar_type_and_size
5104 (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type)));
5107 /* Function vect_is_simple_use.
5110 LOOP_VINFO - the vect info of the loop that is being vectorized.
5111 BB_VINFO - the vect info of the basic block that is being vectorized.
5112 OPERAND - operand of a stmt in the loop or bb.
5113 DEF - the defining stmt in case OPERAND is an SSA_NAME.
5115 Returns whether a stmt with OPERAND can be vectorized.
5116 For loops, supportable operands are constants, loop invariants, and operands
5117 that are defined by the current iteration of the loop. Unsupportable
5118 operands are those that are defined by a previous iteration of the loop (as
5119 is the case in reduction/induction computations).
5120 For basic blocks, supportable operands are constants and bb invariants.
5121 For now, operands defined outside the basic block are not supported. */
5124 vect_is_simple_use (tree operand, loop_vec_info loop_vinfo,
5125 bb_vec_info bb_vinfo, gimple *def_stmt,
5126 tree *def, enum vect_def_type *dt)
5129 stmt_vec_info stmt_vinfo;
5130 struct loop *loop = NULL;
5133 loop = LOOP_VINFO_LOOP (loop_vinfo);
5138 if (vect_print_dump_info (REPORT_DETAILS))
5140 fprintf (vect_dump, "vect_is_simple_use: operand ");
5141 print_generic_expr (vect_dump, operand, TDF_SLIM);
5144 if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST)
5146 *dt = vect_constant_def;
5150 if (is_gimple_min_invariant (operand))
5153 *dt = vect_external_def;
5157 if (TREE_CODE (operand) == PAREN_EXPR)
5159 if (vect_print_dump_info (REPORT_DETAILS))
5160 fprintf (vect_dump, "non-associatable copy.");
5161 operand = TREE_OPERAND (operand, 0);
5164 if (TREE_CODE (operand) != SSA_NAME)
5166 if (vect_print_dump_info (REPORT_DETAILS))
5167 fprintf (vect_dump, "not ssa-name.");
5171 *def_stmt = SSA_NAME_DEF_STMT (operand);
5172 if (*def_stmt == NULL)
5174 if (vect_print_dump_info (REPORT_DETAILS))
5175 fprintf (vect_dump, "no def_stmt.");
5179 if (vect_print_dump_info (REPORT_DETAILS))
5181 fprintf (vect_dump, "def_stmt: ");
5182 print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM);
5185 /* Empty stmt is expected only in case of a function argument.
5186 (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */
5187 if (gimple_nop_p (*def_stmt))
5190 *dt = vect_external_def;
5194 bb = gimple_bb (*def_stmt);
5196 if ((loop && !flow_bb_inside_loop_p (loop, bb))
5197 || (!loop && bb != BB_VINFO_BB (bb_vinfo))
5198 || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI))
5199 *dt = vect_external_def;
5202 stmt_vinfo = vinfo_for_stmt (*def_stmt);
5203 *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo);
5206 if (*dt == vect_unknown_def_type)
5208 if (vect_print_dump_info (REPORT_DETAILS))
5209 fprintf (vect_dump, "Unsupported pattern.");
5213 if (vect_print_dump_info (REPORT_DETAILS))
5214 fprintf (vect_dump, "type of def: %d.",*dt);
5216 switch (gimple_code (*def_stmt))
5219 *def = gimple_phi_result (*def_stmt);
5223 *def = gimple_assign_lhs (*def_stmt);
5227 *def = gimple_call_lhs (*def_stmt);
5232 if (vect_print_dump_info (REPORT_DETAILS))
5233 fprintf (vect_dump, "unsupported defining stmt: ");
5240 /* Function vect_is_simple_use_1.
5242 Same as vect_is_simple_use_1 but also determines the vector operand
5243 type of OPERAND and stores it to *VECTYPE. If the definition of
5244 OPERAND is vect_uninitialized_def, vect_constant_def or
5245 vect_external_def *VECTYPE will be set to NULL_TREE and the caller
5246 is responsible to compute the best suited vector type for the
5250 vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo,
5251 bb_vec_info bb_vinfo, gimple *def_stmt,
5252 tree *def, enum vect_def_type *dt, tree *vectype)
5254 if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt))
5257 /* Now get a vector type if the def is internal, otherwise supply
5258 NULL_TREE and leave it up to the caller to figure out a proper
5259 type for the use stmt. */
5260 if (*dt == vect_internal_def
5261 || *dt == vect_induction_def
5262 || *dt == vect_reduction_def
5263 || *dt == vect_double_reduction_def
5264 || *dt == vect_nested_cycle)
5266 stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt);
5267 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
5268 stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info));
5269 *vectype = STMT_VINFO_VECTYPE (stmt_info);
5270 gcc_assert (*vectype != NULL_TREE);
5272 else if (*dt == vect_uninitialized_def
5273 || *dt == vect_constant_def
5274 || *dt == vect_external_def)
5275 *vectype = NULL_TREE;
5283 /* Function supportable_widening_operation
5285 Check whether an operation represented by the code CODE is a
5286 widening operation that is supported by the target platform in
5287 vector form (i.e., when operating on arguments of type VECTYPE_IN
5288 producing a result of type VECTYPE_OUT).
5290 Widening operations we currently support are NOP (CONVERT), FLOAT
5291 and WIDEN_MULT. This function checks if these operations are supported
5292 by the target platform either directly (via vector tree-codes), or via
5296 - CODE1 and CODE2 are codes of vector operations to be used when
5297 vectorizing the operation, if available.
5298 - DECL1 and DECL2 are decls of target builtin functions to be used
5299 when vectorizing the operation, if available. In this case,
5300 CODE1 and CODE2 are CALL_EXPR.
5301 - MULTI_STEP_CVT determines the number of required intermediate steps in
5302 case of multi-step conversion (like char->short->int - in that case
5303 MULTI_STEP_CVT will be 1).
5304 - INTERM_TYPES contains the intermediate type required to perform the
5305 widening operation (short in the above example). */
5308 supportable_widening_operation (enum tree_code code, gimple stmt,
5309 tree vectype_out, tree vectype_in,
5310 tree *decl1, tree *decl2,
5311 enum tree_code *code1, enum tree_code *code2,
5312 int *multi_step_cvt,
5313 VEC (tree, heap) **interm_types)
5315 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5316 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info);
5317 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
5319 enum machine_mode vec_mode;
5320 enum insn_code icode1, icode2;
5321 optab optab1, optab2;
5322 tree vectype = vectype_in;
5323 tree wide_vectype = vectype_out;
5324 enum tree_code c1, c2;
5326 /* The result of a vectorized widening operation usually requires two vectors
5327 (because the widened results do not fit int one vector). The generated
5328 vector results would normally be expected to be generated in the same
5329 order as in the original scalar computation, i.e. if 8 results are
5330 generated in each vector iteration, they are to be organized as follows:
5331 vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8].
5333 However, in the special case that the result of the widening operation is
5334 used in a reduction computation only, the order doesn't matter (because
5335 when vectorizing a reduction we change the order of the computation).
5336 Some targets can take advantage of this and generate more efficient code.
5337 For example, targets like Altivec, that support widen_mult using a sequence
5338 of {mult_even,mult_odd} generate the following vectors:
5339 vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8].
5341 When vectorizing outer-loops, we execute the inner-loop sequentially
5342 (each vectorized inner-loop iteration contributes to VF outer-loop
5343 iterations in parallel). We therefore don't allow to change the order
5344 of the computation in the inner-loop during outer-loop vectorization. */
5346 if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction
5347 && !nested_in_vect_loop_p (vect_loop, stmt))
5353 && code == WIDEN_MULT_EXPR
5354 && targetm.vectorize.builtin_mul_widen_even
5355 && targetm.vectorize.builtin_mul_widen_even (vectype)
5356 && targetm.vectorize.builtin_mul_widen_odd
5357 && targetm.vectorize.builtin_mul_widen_odd (vectype))
5359 if (vect_print_dump_info (REPORT_DETAILS))
5360 fprintf (vect_dump, "Unordered widening operation detected.");
5362 *code1 = *code2 = CALL_EXPR;
5363 *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype);
5364 *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype);
5370 case WIDEN_MULT_EXPR:
5371 if (BYTES_BIG_ENDIAN)
5373 c1 = VEC_WIDEN_MULT_HI_EXPR;
5374 c2 = VEC_WIDEN_MULT_LO_EXPR;
5378 c2 = VEC_WIDEN_MULT_HI_EXPR;
5379 c1 = VEC_WIDEN_MULT_LO_EXPR;
5384 if (BYTES_BIG_ENDIAN)
5386 c1 = VEC_UNPACK_HI_EXPR;
5387 c2 = VEC_UNPACK_LO_EXPR;
5391 c2 = VEC_UNPACK_HI_EXPR;
5392 c1 = VEC_UNPACK_LO_EXPR;
5397 if (BYTES_BIG_ENDIAN)
5399 c1 = VEC_UNPACK_FLOAT_HI_EXPR;
5400 c2 = VEC_UNPACK_FLOAT_LO_EXPR;
5404 c2 = VEC_UNPACK_FLOAT_HI_EXPR;
5405 c1 = VEC_UNPACK_FLOAT_LO_EXPR;
5409 case FIX_TRUNC_EXPR:
5410 /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/
5411 VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for
5412 computing the operation. */
5419 if (code == FIX_TRUNC_EXPR)
5421 /* The signedness is determined from output operand. */
5422 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
5423 optab2 = optab_for_tree_code (c2, vectype_out, optab_default);
5427 optab1 = optab_for_tree_code (c1, vectype, optab_default);
5428 optab2 = optab_for_tree_code (c2, vectype, optab_default);
5431 if (!optab1 || !optab2)
5434 vec_mode = TYPE_MODE (vectype);
5435 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing
5436 || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing)
5439 /* Check if it's a multi-step conversion that can be done using intermediate
5441 if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype)
5442 || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype))
5445 tree prev_type = vectype, intermediate_type;
5446 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5447 optab optab3, optab4;
5449 if (!CONVERT_EXPR_CODE_P (code))
5455 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5456 intermediate steps in promotion sequence. We try
5457 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
5459 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5460 for (i = 0; i < 3; i++)
5462 intermediate_mode = insn_data[icode1].operand[0].mode;
5463 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5464 TYPE_UNSIGNED (prev_type));
5465 optab3 = optab_for_tree_code (c1, intermediate_type, optab_default);
5466 optab4 = optab_for_tree_code (c2, intermediate_type, optab_default);
5468 if (!optab3 || !optab4
5469 || ((icode1 = optab_handler (optab1, prev_mode))
5470 == CODE_FOR_nothing)
5471 || insn_data[icode1].operand[0].mode != intermediate_mode
5472 || ((icode2 = optab_handler (optab2, prev_mode))
5473 == CODE_FOR_nothing)
5474 || insn_data[icode2].operand[0].mode != intermediate_mode
5475 || ((icode1 = optab_handler (optab3, intermediate_mode))
5476 == CODE_FOR_nothing)
5477 || ((icode2 = optab_handler (optab4, intermediate_mode))
5478 == CODE_FOR_nothing))
5481 VEC_quick_push (tree, *interm_types, intermediate_type);
5482 (*multi_step_cvt)++;
5484 if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype)
5485 && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype))
5488 prev_type = intermediate_type;
5489 prev_mode = intermediate_mode;
5501 /* Function supportable_narrowing_operation
5503 Check whether an operation represented by the code CODE is a
5504 narrowing operation that is supported by the target platform in
5505 vector form (i.e., when operating on arguments of type VECTYPE_IN
5506 and producing a result of type VECTYPE_OUT).
5508 Narrowing operations we currently support are NOP (CONVERT) and
5509 FIX_TRUNC. This function checks if these operations are supported by
5510 the target platform directly via vector tree-codes.
5513 - CODE1 is the code of a vector operation to be used when
5514 vectorizing the operation, if available.
5515 - MULTI_STEP_CVT determines the number of required intermediate steps in
5516 case of multi-step conversion (like int->short->char - in that case
5517 MULTI_STEP_CVT will be 1).
5518 - INTERM_TYPES contains the intermediate type required to perform the
5519 narrowing operation (short in the above example). */
5522 supportable_narrowing_operation (enum tree_code code,
5523 tree vectype_out, tree vectype_in,
5524 enum tree_code *code1, int *multi_step_cvt,
5525 VEC (tree, heap) **interm_types)
5527 enum machine_mode vec_mode;
5528 enum insn_code icode1;
5529 optab optab1, interm_optab;
5530 tree vectype = vectype_in;
5531 tree narrow_vectype = vectype_out;
5533 tree intermediate_type, prev_type;
5539 c1 = VEC_PACK_TRUNC_EXPR;
5542 case FIX_TRUNC_EXPR:
5543 c1 = VEC_PACK_FIX_TRUNC_EXPR;
5547 /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR
5548 tree code and optabs used for computing the operation. */
5555 if (code == FIX_TRUNC_EXPR)
5556 /* The signedness is determined from output operand. */
5557 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
5559 optab1 = optab_for_tree_code (c1, vectype, optab_default);
5564 vec_mode = TYPE_MODE (vectype);
5565 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing)
5568 /* Check if it's a multi-step conversion that can be done using intermediate
5570 if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype))
5572 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5575 prev_type = vectype;
5576 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5577 intermediate steps in promotion sequence. We try
5578 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
5580 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5581 for (i = 0; i < 3; i++)
5583 intermediate_mode = insn_data[icode1].operand[0].mode;
5584 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5585 TYPE_UNSIGNED (prev_type));
5586 interm_optab = optab_for_tree_code (c1, intermediate_type,
5589 || ((icode1 = optab_handler (optab1, prev_mode))
5590 == CODE_FOR_nothing)
5591 || insn_data[icode1].operand[0].mode != intermediate_mode
5592 || ((icode1 = optab_handler (interm_optab, intermediate_mode))
5593 == CODE_FOR_nothing))
5596 VEC_quick_push (tree, *interm_types, intermediate_type);
5597 (*multi_step_cvt)++;
5599 if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype))
5602 prev_type = intermediate_type;
5603 prev_mode = intermediate_mode;