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);
986 tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
987 unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype);
988 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
994 enum vect_def_type dt;
998 if (vect_print_dump_info (REPORT_DETAILS))
1000 fprintf (vect_dump, "vect_get_vec_def_for_operand: ");
1001 print_generic_expr (vect_dump, op, TDF_SLIM);
1004 is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def,
1006 gcc_assert (is_simple_use);
1007 if (vect_print_dump_info (REPORT_DETAILS))
1011 fprintf (vect_dump, "def = ");
1012 print_generic_expr (vect_dump, def, TDF_SLIM);
1016 fprintf (vect_dump, " def_stmt = ");
1017 print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM);
1023 /* Case 1: operand is a constant. */
1024 case vect_constant_def:
1026 vector_type = get_vectype_for_scalar_type (TREE_TYPE (op));
1027 gcc_assert (vector_type);
1032 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1033 if (vect_print_dump_info (REPORT_DETAILS))
1034 fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits);
1036 for (i = nunits - 1; i >= 0; --i)
1038 t = tree_cons (NULL_TREE, op, t);
1040 vec_cst = build_vector (vector_type, t);
1041 return vect_init_vector (stmt, vec_cst, vector_type, NULL);
1044 /* Case 2: operand is defined outside the loop - loop invariant. */
1045 case vect_external_def:
1047 vector_type = get_vectype_for_scalar_type (TREE_TYPE (def));
1048 gcc_assert (vector_type);
1049 nunits = TYPE_VECTOR_SUBPARTS (vector_type);
1054 /* Create 'vec_inv = {inv,inv,..,inv}' */
1055 if (vect_print_dump_info (REPORT_DETAILS))
1056 fprintf (vect_dump, "Create vector_inv.");
1058 for (i = nunits - 1; i >= 0; --i)
1060 t = tree_cons (NULL_TREE, def, t);
1063 /* FIXME: use build_constructor directly. */
1064 vec_inv = build_constructor_from_list (vector_type, t);
1065 return vect_init_vector (stmt, vec_inv, vector_type, NULL);
1068 /* Case 3: operand is defined inside the loop. */
1069 case vect_internal_def:
1072 *scalar_def = NULL/* FIXME tuples: def_stmt*/;
1074 /* Get the def from the vectorized stmt. */
1075 def_stmt_info = vinfo_for_stmt (def_stmt);
1076 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1077 gcc_assert (vec_stmt);
1078 if (gimple_code (vec_stmt) == GIMPLE_PHI)
1079 vec_oprnd = PHI_RESULT (vec_stmt);
1080 else if (is_gimple_call (vec_stmt))
1081 vec_oprnd = gimple_call_lhs (vec_stmt);
1083 vec_oprnd = gimple_assign_lhs (vec_stmt);
1087 /* Case 4: operand is defined by a loop header phi - reduction */
1088 case vect_reduction_def:
1089 case vect_double_reduction_def:
1090 case vect_nested_cycle:
1094 gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
1095 loop = (gimple_bb (def_stmt))->loop_father;
1097 /* Get the def before the loop */
1098 op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop));
1099 return get_initial_def_for_reduction (stmt, op, scalar_def);
1102 /* Case 5: operand is defined by loop-header phi - induction. */
1103 case vect_induction_def:
1105 gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
1107 /* Get the def from the vectorized stmt. */
1108 def_stmt_info = vinfo_for_stmt (def_stmt);
1109 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1110 gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI);
1111 vec_oprnd = PHI_RESULT (vec_stmt);
1121 /* Function vect_get_vec_def_for_stmt_copy
1123 Return a vector-def for an operand. This function is used when the
1124 vectorized stmt to be created (by the caller to this function) is a "copy"
1125 created in case the vectorized result cannot fit in one vector, and several
1126 copies of the vector-stmt are required. In this case the vector-def is
1127 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1128 of the stmt that defines VEC_OPRND.
1129 DT is the type of the vector def VEC_OPRND.
1132 In case the vectorization factor (VF) is bigger than the number
1133 of elements that can fit in a vectype (nunits), we have to generate
1134 more than one vector stmt to vectorize the scalar stmt. This situation
1135 arises when there are multiple data-types operated upon in the loop; the
1136 smallest data-type determines the VF, and as a result, when vectorizing
1137 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1138 vector stmt (each computing a vector of 'nunits' results, and together
1139 computing 'VF' results in each iteration). This function is called when
1140 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1141 which VF=16 and nunits=4, so the number of copies required is 4):
1143 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1145 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1146 VS1.1: vx.1 = memref1 VS1.2
1147 VS1.2: vx.2 = memref2 VS1.3
1148 VS1.3: vx.3 = memref3
1150 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1151 VSnew.1: vz1 = vx.1 + ... VSnew.2
1152 VSnew.2: vz2 = vx.2 + ... VSnew.3
1153 VSnew.3: vz3 = vx.3 + ...
1155 The vectorization of S1 is explained in vectorizable_load.
1156 The vectorization of S2:
1157 To create the first vector-stmt out of the 4 copies - VSnew.0 -
1158 the function 'vect_get_vec_def_for_operand' is called to
1159 get the relevant vector-def for each operand of S2. For operand x it
1160 returns the vector-def 'vx.0'.
1162 To create the remaining copies of the vector-stmt (VSnew.j), this
1163 function is called to get the relevant vector-def for each operand. It is
1164 obtained from the respective VS1.j stmt, which is recorded in the
1165 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
1167 For example, to obtain the vector-def 'vx.1' in order to create the
1168 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
1169 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
1170 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
1171 and return its def ('vx.1').
1172 Overall, to create the above sequence this function will be called 3 times:
1173 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
1174 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
1175 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
1178 vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd)
1180 gimple vec_stmt_for_operand;
1181 stmt_vec_info def_stmt_info;
1183 /* Do nothing; can reuse same def. */
1184 if (dt == vect_external_def || dt == vect_constant_def )
1187 vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd);
1188 def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand);
1189 gcc_assert (def_stmt_info);
1190 vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info);
1191 gcc_assert (vec_stmt_for_operand);
1192 vec_oprnd = gimple_get_lhs (vec_stmt_for_operand);
1193 if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI)
1194 vec_oprnd = PHI_RESULT (vec_stmt_for_operand);
1196 vec_oprnd = gimple_get_lhs (vec_stmt_for_operand);
1201 /* Get vectorized definitions for the operands to create a copy of an original
1202 stmt. See vect_get_vec_def_for_stmt_copy () for details. */
1205 vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt,
1206 VEC(tree,heap) **vec_oprnds0,
1207 VEC(tree,heap) **vec_oprnds1)
1209 tree vec_oprnd = VEC_pop (tree, *vec_oprnds0);
1211 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd);
1212 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
1214 if (vec_oprnds1 && *vec_oprnds1)
1216 vec_oprnd = VEC_pop (tree, *vec_oprnds1);
1217 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd);
1218 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
1223 /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not
1227 vect_get_vec_defs (tree op0, tree op1, gimple stmt,
1228 VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1,
1232 vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1, -1);
1237 *vec_oprnds0 = VEC_alloc (tree, heap, 1);
1238 vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL);
1239 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
1243 *vec_oprnds1 = VEC_alloc (tree, heap, 1);
1244 vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL);
1245 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
1251 /* Function vect_finish_stmt_generation.
1253 Insert a new stmt. */
1256 vect_finish_stmt_generation (gimple stmt, gimple vec_stmt,
1257 gimple_stmt_iterator *gsi)
1259 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1260 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1261 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
1263 gcc_assert (gimple_code (stmt) != GIMPLE_LABEL);
1265 gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT);
1267 set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo,
1270 if (vect_print_dump_info (REPORT_DETAILS))
1272 fprintf (vect_dump, "add new stmt: ");
1273 print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM);
1276 gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi)));
1279 /* Checks if CALL can be vectorized in type VECTYPE. Returns
1280 a function declaration if the target has a vectorized version
1281 of the function, or NULL_TREE if the function cannot be vectorized. */
1284 vectorizable_function (gimple call, tree vectype_out, tree vectype_in)
1286 tree fndecl = gimple_call_fndecl (call);
1288 /* We only handle functions that do not read or clobber memory -- i.e.
1289 const or novops ones. */
1290 if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS)))
1294 || TREE_CODE (fndecl) != FUNCTION_DECL
1295 || !DECL_BUILT_IN (fndecl))
1298 return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out,
1302 /* Function vectorizable_call.
1304 Check if STMT performs a function call that can be vectorized.
1305 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1306 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1307 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1310 vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt)
1315 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
1316 stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info;
1317 tree vectype_out, vectype_in;
1320 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1321 tree fndecl, new_temp, def, rhs_type;
1323 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1324 gimple new_stmt = NULL;
1326 VEC(tree, heap) *vargs = NULL;
1327 enum { NARROW, NONE, WIDEN } modifier;
1330 /* FORNOW: unsupported in basic block SLP. */
1331 gcc_assert (loop_vinfo);
1333 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1336 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1339 /* FORNOW: SLP not supported. */
1340 if (STMT_SLP_TYPE (stmt_info))
1343 /* Is STMT a vectorizable call? */
1344 if (!is_gimple_call (stmt))
1347 if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
1350 if (stmt_could_throw_p (stmt))
1353 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1355 /* Process function arguments. */
1356 rhs_type = NULL_TREE;
1357 vectype_in = NULL_TREE;
1358 nargs = gimple_call_num_args (stmt);
1360 /* Bail out if the function has more than two arguments, we
1361 do not have interesting builtin functions to vectorize with
1362 more than two arguments. No arguments is also not good. */
1363 if (nargs == 0 || nargs > 2)
1366 for (i = 0; i < nargs; i++)
1370 op = gimple_call_arg (stmt, i);
1372 /* We can only handle calls with arguments of the same type. */
1374 && !types_compatible_p (rhs_type, TREE_TYPE (op)))
1376 if (vect_print_dump_info (REPORT_DETAILS))
1377 fprintf (vect_dump, "argument types differ.");
1381 rhs_type = TREE_TYPE (op);
1383 if (!vect_is_simple_use_1 (op, loop_vinfo, NULL,
1384 &def_stmt, &def, &dt[i], &opvectype))
1386 if (vect_print_dump_info (REPORT_DETAILS))
1387 fprintf (vect_dump, "use not simple.");
1392 vectype_in = opvectype;
1394 && opvectype != vectype_in)
1396 if (vect_print_dump_info (REPORT_DETAILS))
1397 fprintf (vect_dump, "argument vector types differ.");
1401 /* If all arguments are external or constant defs use a vector type with
1402 the same size as the output vector type. */
1404 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1406 gcc_assert (vectype_in);
1409 if (vect_print_dump_info (REPORT_DETAILS))
1411 fprintf (vect_dump, "no vectype for scalar type ");
1412 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1419 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1420 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1421 if (nunits_in == nunits_out / 2)
1423 else if (nunits_out == nunits_in)
1425 else if (nunits_out == nunits_in / 2)
1430 /* For now, we only vectorize functions if a target specific builtin
1431 is available. TODO -- in some cases, it might be profitable to
1432 insert the calls for pieces of the vector, in order to be able
1433 to vectorize other operations in the loop. */
1434 fndecl = vectorizable_function (stmt, vectype_out, vectype_in);
1435 if (fndecl == NULL_TREE)
1437 if (vect_print_dump_info (REPORT_DETAILS))
1438 fprintf (vect_dump, "function is not vectorizable.");
1443 gcc_assert (!gimple_vuse (stmt));
1445 if (modifier == NARROW)
1446 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1448 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1450 /* Sanity check: make sure that at least one copy of the vectorized stmt
1451 needs to be generated. */
1452 gcc_assert (ncopies >= 1);
1454 if (!vec_stmt) /* transformation not required. */
1456 STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
1457 if (vect_print_dump_info (REPORT_DETAILS))
1458 fprintf (vect_dump, "=== vectorizable_call ===");
1459 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
1465 if (vect_print_dump_info (REPORT_DETAILS))
1466 fprintf (vect_dump, "transform operation.");
1469 scalar_dest = gimple_call_lhs (stmt);
1470 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1472 prev_stmt_info = NULL;
1476 for (j = 0; j < ncopies; ++j)
1478 /* Build argument list for the vectorized call. */
1480 vargs = VEC_alloc (tree, heap, nargs);
1482 VEC_truncate (tree, vargs, 0);
1484 for (i = 0; i < nargs; i++)
1486 op = gimple_call_arg (stmt, i);
1489 = vect_get_vec_def_for_operand (op, stmt, NULL);
1492 vec_oprnd0 = gimple_call_arg (new_stmt, i);
1494 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1497 VEC_quick_push (tree, vargs, vec_oprnd0);
1500 new_stmt = gimple_build_call_vec (fndecl, vargs);
1501 new_temp = make_ssa_name (vec_dest, new_stmt);
1502 gimple_call_set_lhs (new_stmt, new_temp);
1504 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1505 mark_symbols_for_renaming (new_stmt);
1508 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1510 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1512 prev_stmt_info = vinfo_for_stmt (new_stmt);
1518 for (j = 0; j < ncopies; ++j)
1520 /* Build argument list for the vectorized call. */
1522 vargs = VEC_alloc (tree, heap, nargs * 2);
1524 VEC_truncate (tree, vargs, 0);
1526 for (i = 0; i < nargs; i++)
1528 op = gimple_call_arg (stmt, i);
1532 = vect_get_vec_def_for_operand (op, stmt, NULL);
1534 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1538 vec_oprnd1 = gimple_call_arg (new_stmt, 2*i);
1540 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1);
1542 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1545 VEC_quick_push (tree, vargs, vec_oprnd0);
1546 VEC_quick_push (tree, vargs, vec_oprnd1);
1549 new_stmt = gimple_build_call_vec (fndecl, vargs);
1550 new_temp = make_ssa_name (vec_dest, new_stmt);
1551 gimple_call_set_lhs (new_stmt, new_temp);
1553 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1554 mark_symbols_for_renaming (new_stmt);
1557 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1559 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1561 prev_stmt_info = vinfo_for_stmt (new_stmt);
1564 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1569 /* No current target implements this case. */
1573 VEC_free (tree, heap, vargs);
1575 /* Update the exception handling table with the vector stmt if necessary. */
1576 if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt))
1577 gimple_purge_dead_eh_edges (gimple_bb (stmt));
1579 /* The call in STMT might prevent it from being removed in dce.
1580 We however cannot remove it here, due to the way the ssa name
1581 it defines is mapped to the new definition. So just replace
1582 rhs of the statement with something harmless. */
1584 type = TREE_TYPE (scalar_dest);
1585 new_stmt = gimple_build_assign (gimple_call_lhs (stmt),
1586 fold_convert (type, integer_zero_node));
1587 set_vinfo_for_stmt (new_stmt, stmt_info);
1588 set_vinfo_for_stmt (stmt, NULL);
1589 STMT_VINFO_STMT (stmt_info) = new_stmt;
1590 gsi_replace (gsi, new_stmt, false);
1591 SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt;
1597 /* Function vect_gen_widened_results_half
1599 Create a vector stmt whose code, type, number of arguments, and result
1600 variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are
1601 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
1602 In the case that CODE is a CALL_EXPR, this means that a call to DECL
1603 needs to be created (DECL is a function-decl of a target-builtin).
1604 STMT is the original scalar stmt that we are vectorizing. */
1607 vect_gen_widened_results_half (enum tree_code code,
1609 tree vec_oprnd0, tree vec_oprnd1, int op_type,
1610 tree vec_dest, gimple_stmt_iterator *gsi,
1616 /* Generate half of the widened result: */
1617 if (code == CALL_EXPR)
1619 /* Target specific support */
1620 if (op_type == binary_op)
1621 new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1);
1623 new_stmt = gimple_build_call (decl, 1, vec_oprnd0);
1624 new_temp = make_ssa_name (vec_dest, new_stmt);
1625 gimple_call_set_lhs (new_stmt, new_temp);
1629 /* Generic support */
1630 gcc_assert (op_type == TREE_CODE_LENGTH (code));
1631 if (op_type != binary_op)
1633 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0,
1635 new_temp = make_ssa_name (vec_dest, new_stmt);
1636 gimple_assign_set_lhs (new_stmt, new_temp);
1638 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1644 /* Check if STMT performs a conversion operation, that can be vectorized.
1645 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1646 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1647 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1650 vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi,
1651 gimple *vec_stmt, slp_tree slp_node)
1656 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
1657 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1658 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1659 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
1660 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
1664 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1665 gimple new_stmt = NULL;
1666 stmt_vec_info prev_stmt_info;
1669 tree vectype_out, vectype_in;
1673 enum { NARROW, NONE, WIDEN } modifier;
1675 VEC(tree,heap) *vec_oprnds0 = NULL;
1677 VEC(tree,heap) *dummy = NULL;
1680 /* Is STMT a vectorizable conversion? */
1682 /* FORNOW: unsupported in basic block SLP. */
1683 gcc_assert (loop_vinfo);
1685 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1688 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1691 if (!is_gimple_assign (stmt))
1694 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
1697 code = gimple_assign_rhs_code (stmt);
1698 if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
1701 /* Check types of lhs and rhs. */
1702 scalar_dest = gimple_assign_lhs (stmt);
1703 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1705 op0 = gimple_assign_rhs1 (stmt);
1706 rhs_type = TREE_TYPE (op0);
1707 /* Check the operands of the operation. */
1708 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
1709 &def_stmt, &def, &dt[0], &vectype_in))
1711 if (vect_print_dump_info (REPORT_DETAILS))
1712 fprintf (vect_dump, "use not simple.");
1715 /* If op0 is an external or constant defs use a vector type of
1716 the same size as the output vector type. */
1718 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1720 gcc_assert (vectype_in);
1723 if (vect_print_dump_info (REPORT_DETAILS))
1725 fprintf (vect_dump, "no vectype for scalar type ");
1726 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1733 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1734 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1735 if (nunits_in == nunits_out / 2)
1737 else if (nunits_out == nunits_in)
1739 else if (nunits_out == nunits_in / 2)
1744 if (modifier == NARROW)
1745 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1747 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1749 /* Multiple types in SLP are handled by creating the appropriate number of
1750 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
1755 /* Sanity check: make sure that at least one copy of the vectorized stmt
1756 needs to be generated. */
1757 gcc_assert (ncopies >= 1);
1759 /* Supportable by target? */
1760 if ((modifier == NONE
1761 && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in))
1762 || (modifier == WIDEN
1763 && !supportable_widening_operation (code, stmt,
1764 vectype_out, vectype_in,
1767 &dummy_int, &dummy))
1768 || (modifier == NARROW
1769 && !supportable_narrowing_operation (code, vectype_out, vectype_in,
1770 &code1, &dummy_int, &dummy)))
1772 if (vect_print_dump_info (REPORT_DETAILS))
1773 fprintf (vect_dump, "conversion not supported by target.");
1777 if (modifier != NONE)
1779 /* FORNOW: SLP not supported. */
1780 if (STMT_SLP_TYPE (stmt_info))
1784 if (!vec_stmt) /* transformation not required. */
1786 STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type;
1791 if (vect_print_dump_info (REPORT_DETAILS))
1792 fprintf (vect_dump, "transform conversion.");
1795 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1797 if (modifier == NONE && !slp_node)
1798 vec_oprnds0 = VEC_alloc (tree, heap, 1);
1800 prev_stmt_info = NULL;
1804 for (j = 0; j < ncopies; j++)
1807 vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node);
1809 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL);
1812 targetm.vectorize.builtin_conversion (code,
1813 vectype_out, vectype_in);
1814 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
1816 /* Arguments are ready. create the new vector stmt. */
1817 new_stmt = gimple_build_call (builtin_decl, 1, vop0);
1818 new_temp = make_ssa_name (vec_dest, new_stmt);
1819 gimple_call_set_lhs (new_stmt, new_temp);
1820 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1822 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
1826 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1828 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1829 prev_stmt_info = vinfo_for_stmt (new_stmt);
1834 /* In case the vectorization factor (VF) is bigger than the number
1835 of elements that we can fit in a vectype (nunits), we have to
1836 generate more than one vector stmt - i.e - we need to "unroll"
1837 the vector stmt by a factor VF/nunits. */
1838 for (j = 0; j < ncopies; j++)
1841 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1843 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1845 /* Generate first half of the widened result: */
1847 = vect_gen_widened_results_half (code1, decl1,
1848 vec_oprnd0, vec_oprnd1,
1849 unary_op, vec_dest, gsi, stmt);
1851 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1853 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1854 prev_stmt_info = vinfo_for_stmt (new_stmt);
1856 /* Generate second half of the widened result: */
1858 = vect_gen_widened_results_half (code2, decl2,
1859 vec_oprnd0, vec_oprnd1,
1860 unary_op, vec_dest, gsi, stmt);
1861 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1862 prev_stmt_info = vinfo_for_stmt (new_stmt);
1867 /* In case the vectorization factor (VF) is bigger than the number
1868 of elements that we can fit in a vectype (nunits), we have to
1869 generate more than one vector stmt - i.e - we need to "unroll"
1870 the vector stmt by a factor VF/nunits. */
1871 for (j = 0; j < ncopies; j++)
1876 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1877 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1881 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
1882 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1885 /* Arguments are ready. Create the new vector stmt. */
1886 new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0,
1888 new_temp = make_ssa_name (vec_dest, new_stmt);
1889 gimple_assign_set_lhs (new_stmt, new_temp);
1890 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1893 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1895 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1897 prev_stmt_info = vinfo_for_stmt (new_stmt);
1900 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1904 VEC_free (tree, heap, vec_oprnds0);
1910 /* Function vectorizable_assignment.
1912 Check if STMT performs an assignment (copy) that can be vectorized.
1913 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1914 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1915 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1918 vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi,
1919 gimple *vec_stmt, slp_tree slp_node)
1924 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1925 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
1926 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1930 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1931 unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype);
1934 VEC(tree,heap) *vec_oprnds = NULL;
1936 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
1937 gimple new_stmt = NULL;
1938 stmt_vec_info prev_stmt_info = NULL;
1939 enum tree_code code;
1942 /* Multiple types in SLP are handled by creating the appropriate number of
1943 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
1948 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
1950 gcc_assert (ncopies >= 1);
1952 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
1955 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1958 /* Is vectorizable assignment? */
1959 if (!is_gimple_assign (stmt))
1962 scalar_dest = gimple_assign_lhs (stmt);
1963 if (TREE_CODE (scalar_dest) != SSA_NAME)
1966 code = gimple_assign_rhs_code (stmt);
1967 if (gimple_assign_single_p (stmt)
1968 || code == PAREN_EXPR
1969 || CONVERT_EXPR_CODE_P (code))
1970 op = gimple_assign_rhs1 (stmt);
1974 if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo,
1975 &def_stmt, &def, &dt[0], &vectype_in))
1977 if (vect_print_dump_info (REPORT_DETAILS))
1978 fprintf (vect_dump, "use not simple.");
1982 /* We can handle NOP_EXPR conversions that do not change the number
1983 of elements or the vector size. */
1984 if (CONVERT_EXPR_CODE_P (code)
1986 || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits
1987 || (GET_MODE_SIZE (TYPE_MODE (vectype))
1988 != GET_MODE_SIZE (TYPE_MODE (vectype_in)))))
1991 if (!vec_stmt) /* transformation not required. */
1993 STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
1994 if (vect_print_dump_info (REPORT_DETAILS))
1995 fprintf (vect_dump, "=== vectorizable_assignment ===");
1996 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2001 if (vect_print_dump_info (REPORT_DETAILS))
2002 fprintf (vect_dump, "transform assignment.");
2005 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2008 for (j = 0; j < ncopies; j++)
2012 vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node);
2014 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL);
2016 /* Arguments are ready. create the new vector stmt. */
2017 FOR_EACH_VEC_ELT (tree, vec_oprnds, i, vop)
2019 if (CONVERT_EXPR_CODE_P (code))
2020 vop = build1 (VIEW_CONVERT_EXPR, vectype, vop);
2021 new_stmt = gimple_build_assign (vec_dest, vop);
2022 new_temp = make_ssa_name (vec_dest, new_stmt);
2023 gimple_assign_set_lhs (new_stmt, new_temp);
2024 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2026 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2033 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2035 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2037 prev_stmt_info = vinfo_for_stmt (new_stmt);
2040 VEC_free (tree, heap, vec_oprnds);
2044 /* Function vectorizable_operation.
2046 Check if STMT performs a binary or unary operation that can be vectorized.
2047 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2048 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2049 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2052 vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi,
2053 gimple *vec_stmt, slp_tree slp_node)
2057 tree op0, op1 = NULL;
2058 tree vec_oprnd1 = NULL_TREE;
2059 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2061 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2062 enum tree_code code;
2063 enum machine_mode vec_mode;
2068 enum machine_mode optab_op2_mode;
2071 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2072 gimple new_stmt = NULL;
2073 stmt_vec_info prev_stmt_info;
2079 VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2082 bool scalar_shift_arg = false;
2083 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
2086 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
2089 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2092 /* Is STMT a vectorizable binary/unary operation? */
2093 if (!is_gimple_assign (stmt))
2096 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2099 code = gimple_assign_rhs_code (stmt);
2101 /* For pointer addition, we should use the normal plus for
2102 the vector addition. */
2103 if (code == POINTER_PLUS_EXPR)
2106 /* Support only unary or binary operations. */
2107 op_type = TREE_CODE_LENGTH (code);
2108 if (op_type != unary_op && op_type != binary_op)
2110 if (vect_print_dump_info (REPORT_DETAILS))
2111 fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type);
2115 scalar_dest = gimple_assign_lhs (stmt);
2116 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2118 op0 = gimple_assign_rhs1 (stmt);
2119 if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo,
2120 &def_stmt, &def, &dt[0], &vectype))
2122 if (vect_print_dump_info (REPORT_DETAILS))
2123 fprintf (vect_dump, "use not simple.");
2126 /* If op0 is an external or constant def use a vector type with
2127 the same size as the output vector type. */
2129 vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2131 gcc_assert (vectype);
2134 if (vect_print_dump_info (REPORT_DETAILS))
2136 fprintf (vect_dump, "no vectype for scalar type ");
2137 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2143 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2144 nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
2145 if (nunits_out != nunits_in)
2148 if (op_type == binary_op)
2150 op1 = gimple_assign_rhs2 (stmt);
2151 if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def,
2154 if (vect_print_dump_info (REPORT_DETAILS))
2155 fprintf (vect_dump, "use not simple.");
2161 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2165 /* Multiple types in SLP are handled by creating the appropriate number of
2166 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2171 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2173 gcc_assert (ncopies >= 1);
2175 /* If this is a shift/rotate, determine whether the shift amount is a vector,
2176 or scalar. If the shift/rotate amount is a vector, use the vector/vector
2178 if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR
2179 || code == RROTATE_EXPR)
2181 /* vector shifted by vector */
2182 if (dt[1] == vect_internal_def)
2184 optab = optab_for_tree_code (code, vectype, optab_vector);
2185 if (vect_print_dump_info (REPORT_DETAILS))
2186 fprintf (vect_dump, "vector/vector shift/rotate found.");
2189 /* See if the machine has a vector shifted by scalar insn and if not
2190 then see if it has a vector shifted by vector insn */
2191 else if (dt[1] == vect_constant_def || dt[1] == vect_external_def)
2193 optab = optab_for_tree_code (code, vectype, optab_scalar);
2195 && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing)
2197 scalar_shift_arg = true;
2198 if (vect_print_dump_info (REPORT_DETAILS))
2199 fprintf (vect_dump, "vector/scalar shift/rotate found.");
2203 optab = optab_for_tree_code (code, vectype, optab_vector);
2205 && (optab_handler (optab, TYPE_MODE (vectype))
2206 != CODE_FOR_nothing))
2208 if (vect_print_dump_info (REPORT_DETAILS))
2209 fprintf (vect_dump, "vector/vector shift/rotate found.");
2211 /* Unlike the other binary operators, shifts/rotates have
2212 the rhs being int, instead of the same type as the lhs,
2213 so make sure the scalar is the right type if we are
2214 dealing with vectors of short/char. */
2215 if (dt[1] == vect_constant_def)
2216 op1 = fold_convert (TREE_TYPE (vectype), op1);
2223 if (vect_print_dump_info (REPORT_DETAILS))
2224 fprintf (vect_dump, "operand mode requires invariant argument.");
2229 optab = optab_for_tree_code (code, vectype, optab_default);
2231 /* Supportable by target? */
2234 if (vect_print_dump_info (REPORT_DETAILS))
2235 fprintf (vect_dump, "no optab.");
2238 vec_mode = TYPE_MODE (vectype);
2239 icode = (int) optab_handler (optab, vec_mode);
2240 if (icode == CODE_FOR_nothing)
2242 if (vect_print_dump_info (REPORT_DETAILS))
2243 fprintf (vect_dump, "op not supported by target.");
2244 /* Check only during analysis. */
2245 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2246 || (vf < vect_min_worthwhile_factor (code)
2249 if (vect_print_dump_info (REPORT_DETAILS))
2250 fprintf (vect_dump, "proceeding using word mode.");
2253 /* Worthwhile without SIMD support? Check only during analysis. */
2254 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2255 && vf < vect_min_worthwhile_factor (code)
2258 if (vect_print_dump_info (REPORT_DETAILS))
2259 fprintf (vect_dump, "not worthwhile without SIMD support.");
2263 if (!vec_stmt) /* transformation not required. */
2265 STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
2266 if (vect_print_dump_info (REPORT_DETAILS))
2267 fprintf (vect_dump, "=== vectorizable_operation ===");
2268 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2274 if (vect_print_dump_info (REPORT_DETAILS))
2275 fprintf (vect_dump, "transform binary/unary operation.");
2278 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2280 /* Allocate VECs for vector operands. In case of SLP, vector operands are
2281 created in the previous stages of the recursion, so no allocation is
2282 needed, except for the case of shift with scalar shift argument. In that
2283 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
2284 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
2285 In case of loop-based vectorization we allocate VECs of size 1. We
2286 allocate VEC_OPRNDS1 only in case of binary operation. */
2289 vec_oprnds0 = VEC_alloc (tree, heap, 1);
2290 if (op_type == binary_op)
2291 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2293 else if (scalar_shift_arg)
2294 vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size);
2296 /* In case the vectorization factor (VF) is bigger than the number
2297 of elements that we can fit in a vectype (nunits), we have to generate
2298 more than one vector stmt - i.e - we need to "unroll" the
2299 vector stmt by a factor VF/nunits. In doing so, we record a pointer
2300 from one copy of the vector stmt to the next, in the field
2301 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
2302 stages to find the correct vector defs to be used when vectorizing
2303 stmts that use the defs of the current stmt. The example below
2304 illustrates the vectorization process when VF=16 and nunits=4 (i.e.,
2305 we need to create 4 vectorized stmts):
2307 before vectorization:
2308 RELATED_STMT VEC_STMT
2312 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
2314 RELATED_STMT VEC_STMT
2315 VS1_0: vx0 = memref0 VS1_1 -
2316 VS1_1: vx1 = memref1 VS1_2 -
2317 VS1_2: vx2 = memref2 VS1_3 -
2318 VS1_3: vx3 = memref3 - -
2319 S1: x = load - VS1_0
2322 step2: vectorize stmt S2 (done here):
2323 To vectorize stmt S2 we first need to find the relevant vector
2324 def for the first operand 'x'. This is, as usual, obtained from
2325 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
2326 that defines 'x' (S1). This way we find the stmt VS1_0, and the
2327 relevant vector def 'vx0'. Having found 'vx0' we can generate
2328 the vector stmt VS2_0, and as usual, record it in the
2329 STMT_VINFO_VEC_STMT of stmt S2.
2330 When creating the second copy (VS2_1), we obtain the relevant vector
2331 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
2332 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
2333 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
2334 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
2335 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
2336 chain of stmts and pointers:
2337 RELATED_STMT VEC_STMT
2338 VS1_0: vx0 = memref0 VS1_1 -
2339 VS1_1: vx1 = memref1 VS1_2 -
2340 VS1_2: vx2 = memref2 VS1_3 -
2341 VS1_3: vx3 = memref3 - -
2342 S1: x = load - VS1_0
2343 VS2_0: vz0 = vx0 + v1 VS2_1 -
2344 VS2_1: vz1 = vx1 + v1 VS2_2 -
2345 VS2_2: vz2 = vx2 + v1 VS2_3 -
2346 VS2_3: vz3 = vx3 + v1 - -
2347 S2: z = x + 1 - VS2_0 */
2349 prev_stmt_info = NULL;
2350 for (j = 0; j < ncopies; j++)
2355 if (op_type == binary_op && scalar_shift_arg)
2357 /* Vector shl and shr insn patterns can be defined with scalar
2358 operand 2 (shift operand). In this case, use constant or loop
2359 invariant op1 directly, without extending it to vector mode
2361 optab_op2_mode = insn_data[icode].operand[2].mode;
2362 if (!VECTOR_MODE_P (optab_op2_mode))
2364 if (vect_print_dump_info (REPORT_DETAILS))
2365 fprintf (vect_dump, "operand 1 using scalar mode.");
2367 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2370 /* Store vec_oprnd1 for every vector stmt to be created
2371 for SLP_NODE. We check during the analysis that all
2372 the shift arguments are the same.
2373 TODO: Allow different constants for different vector
2374 stmts generated for an SLP instance. */
2375 for (k = 0; k < slp_node->vec_stmts_size - 1; k++)
2376 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2381 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
2382 (a special case for certain kind of vector shifts); otherwise,
2383 operand 1 should be of a vector type (the usual case). */
2384 if (op_type == binary_op && !vec_oprnd1)
2385 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
2388 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
2392 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
2394 /* Arguments are ready. Create the new vector stmt. */
2395 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, vop0)
2397 vop1 = ((op_type == binary_op)
2398 ? VEC_index (tree, vec_oprnds1, i) : NULL);
2399 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2400 new_temp = make_ssa_name (vec_dest, new_stmt);
2401 gimple_assign_set_lhs (new_stmt, new_temp);
2402 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2404 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2411 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2413 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2414 prev_stmt_info = vinfo_for_stmt (new_stmt);
2417 VEC_free (tree, heap, vec_oprnds0);
2419 VEC_free (tree, heap, vec_oprnds1);
2425 /* Get vectorized definitions for loop-based vectorization. For the first
2426 operand we call vect_get_vec_def_for_operand() (with OPRND containing
2427 scalar operand), and for the rest we get a copy with
2428 vect_get_vec_def_for_stmt_copy() using the previous vector definition
2429 (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details.
2430 The vectors are collected into VEC_OPRNDS. */
2433 vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt,
2434 VEC (tree, heap) **vec_oprnds, int multi_step_cvt)
2438 /* Get first vector operand. */
2439 /* All the vector operands except the very first one (that is scalar oprnd)
2441 if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE)
2442 vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL);
2444 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd);
2446 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2448 /* Get second vector operand. */
2449 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd);
2450 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2454 /* For conversion in multiple steps, continue to get operands
2457 vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1);
2461 /* Create vectorized demotion statements for vector operands from VEC_OPRNDS.
2462 For multi-step conversions store the resulting vectors and call the function
2466 vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds,
2467 int multi_step_cvt, gimple stmt,
2468 VEC (tree, heap) *vec_dsts,
2469 gimple_stmt_iterator *gsi,
2470 slp_tree slp_node, enum tree_code code,
2471 stmt_vec_info *prev_stmt_info)
2474 tree vop0, vop1, new_tmp, vec_dest;
2476 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2478 vec_dest = VEC_pop (tree, vec_dsts);
2480 for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2)
2482 /* Create demotion operation. */
2483 vop0 = VEC_index (tree, *vec_oprnds, i);
2484 vop1 = VEC_index (tree, *vec_oprnds, i + 1);
2485 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2486 new_tmp = make_ssa_name (vec_dest, new_stmt);
2487 gimple_assign_set_lhs (new_stmt, new_tmp);
2488 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2491 /* Store the resulting vector for next recursive call. */
2492 VEC_replace (tree, *vec_oprnds, i/2, new_tmp);
2495 /* This is the last step of the conversion sequence. Store the
2496 vectors in SLP_NODE or in vector info of the scalar statement
2497 (or in STMT_VINFO_RELATED_STMT chain). */
2499 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2502 if (!*prev_stmt_info)
2503 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
2505 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt;
2507 *prev_stmt_info = vinfo_for_stmt (new_stmt);
2512 /* For multi-step demotion operations we first generate demotion operations
2513 from the source type to the intermediate types, and then combine the
2514 results (stored in VEC_OPRNDS) in demotion operation to the destination
2518 /* At each level of recursion we have have of the operands we had at the
2520 VEC_truncate (tree, *vec_oprnds, (i+1)/2);
2521 vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1,
2522 stmt, vec_dsts, gsi, slp_node,
2523 code, prev_stmt_info);
2528 /* Function vectorizable_type_demotion
2530 Check if STMT performs a binary or unary operation that involves
2531 type demotion, and if it can be vectorized.
2532 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2533 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2534 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2537 vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi,
2538 gimple *vec_stmt, slp_tree slp_node)
2543 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2544 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2545 enum tree_code code, code1 = ERROR_MARK;
2548 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2549 stmt_vec_info prev_stmt_info;
2556 int multi_step_cvt = 0;
2557 VEC (tree, heap) *vec_oprnds0 = NULL;
2558 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
2559 tree last_oprnd, intermediate_type;
2561 /* FORNOW: not supported by basic block SLP vectorization. */
2562 gcc_assert (loop_vinfo);
2564 if (!STMT_VINFO_RELEVANT_P (stmt_info))
2567 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2570 /* Is STMT a vectorizable type-demotion operation? */
2571 if (!is_gimple_assign (stmt))
2574 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2577 code = gimple_assign_rhs_code (stmt);
2578 if (!CONVERT_EXPR_CODE_P (code))
2581 scalar_dest = gimple_assign_lhs (stmt);
2582 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2584 /* Check the operands of the operation. */
2585 op0 = gimple_assign_rhs1 (stmt);
2586 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
2587 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
2588 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
2589 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
2590 && CONVERT_EXPR_CODE_P (code))))
2592 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
2593 &def_stmt, &def, &dt[0], &vectype_in))
2595 if (vect_print_dump_info (REPORT_DETAILS))
2596 fprintf (vect_dump, "use not simple.");
2599 /* If op0 is an external def use a vector type with the
2600 same size as the output vector type if possible. */
2602 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2604 gcc_assert (vectype_in);
2607 if (vect_print_dump_info (REPORT_DETAILS))
2609 fprintf (vect_dump, "no vectype for scalar type ");
2610 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2616 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
2617 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2618 if (nunits_in >= nunits_out)
2621 /* Multiple types in SLP are handled by creating the appropriate number of
2622 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2627 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
2628 gcc_assert (ncopies >= 1);
2630 /* Supportable by target? */
2631 if (!supportable_narrowing_operation (code, vectype_out, vectype_in,
2632 &code1, &multi_step_cvt, &interm_types))
2635 if (!vec_stmt) /* transformation not required. */
2637 STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type;
2638 if (vect_print_dump_info (REPORT_DETAILS))
2639 fprintf (vect_dump, "=== vectorizable_demotion ===");
2640 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2645 if (vect_print_dump_info (REPORT_DETAILS))
2646 fprintf (vect_dump, "transform type demotion operation. ncopies = %d.",
2649 /* In case of multi-step demotion, we first generate demotion operations to
2650 the intermediate types, and then from that types to the final one.
2651 We create vector destinations for the intermediate type (TYPES) received
2652 from supportable_narrowing_operation, and store them in the correct order
2653 for future use in vect_create_vectorized_demotion_stmts(). */
2655 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
2657 vec_dsts = VEC_alloc (tree, heap, 1);
2659 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
2660 VEC_quick_push (tree, vec_dsts, vec_dest);
2664 for (i = VEC_length (tree, interm_types) - 1;
2665 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
2667 vec_dest = vect_create_destination_var (scalar_dest,
2669 VEC_quick_push (tree, vec_dsts, vec_dest);
2673 /* In case the vectorization factor (VF) is bigger than the number
2674 of elements that we can fit in a vectype (nunits), we have to generate
2675 more than one vector stmt - i.e - we need to "unroll" the
2676 vector stmt by a factor VF/nunits. */
2678 prev_stmt_info = NULL;
2679 for (j = 0; j < ncopies; j++)
2683 vect_get_slp_defs (slp_node, &vec_oprnds0, NULL, -1);
2686 VEC_free (tree, heap, vec_oprnds0);
2687 vec_oprnds0 = VEC_alloc (tree, heap,
2688 (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2));
2689 vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0,
2690 vect_pow2 (multi_step_cvt) - 1);
2693 /* Arguments are ready. Create the new vector stmts. */
2694 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
2695 vect_create_vectorized_demotion_stmts (&vec_oprnds0,
2696 multi_step_cvt, stmt, tmp_vec_dsts,
2697 gsi, slp_node, code1,
2701 VEC_free (tree, heap, vec_oprnds0);
2702 VEC_free (tree, heap, vec_dsts);
2703 VEC_free (tree, heap, tmp_vec_dsts);
2704 VEC_free (tree, heap, interm_types);
2706 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
2711 /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0
2712 and VEC_OPRNDS1 (for binary operations). For multi-step conversions store
2713 the resulting vectors and call the function recursively. */
2716 vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0,
2717 VEC (tree, heap) **vec_oprnds1,
2718 int multi_step_cvt, gimple stmt,
2719 VEC (tree, heap) *vec_dsts,
2720 gimple_stmt_iterator *gsi,
2721 slp_tree slp_node, enum tree_code code1,
2722 enum tree_code code2, tree decl1,
2723 tree decl2, int op_type,
2724 stmt_vec_info *prev_stmt_info)
2727 tree vop0, vop1, new_tmp1, new_tmp2, vec_dest;
2728 gimple new_stmt1, new_stmt2;
2729 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2730 VEC (tree, heap) *vec_tmp;
2732 vec_dest = VEC_pop (tree, vec_dsts);
2733 vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2);
2735 FOR_EACH_VEC_ELT (tree, *vec_oprnds0, i, vop0)
2737 if (op_type == binary_op)
2738 vop1 = VEC_index (tree, *vec_oprnds1, i);
2742 /* Generate the two halves of promotion operation. */
2743 new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1,
2744 op_type, vec_dest, gsi, stmt);
2745 new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1,
2746 op_type, vec_dest, gsi, stmt);
2747 if (is_gimple_call (new_stmt1))
2749 new_tmp1 = gimple_call_lhs (new_stmt1);
2750 new_tmp2 = gimple_call_lhs (new_stmt2);
2754 new_tmp1 = gimple_assign_lhs (new_stmt1);
2755 new_tmp2 = gimple_assign_lhs (new_stmt2);
2760 /* Store the results for the recursive call. */
2761 VEC_quick_push (tree, vec_tmp, new_tmp1);
2762 VEC_quick_push (tree, vec_tmp, new_tmp2);
2766 /* Last step of promotion sequience - store the results. */
2769 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1);
2770 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2);
2774 if (!*prev_stmt_info)
2775 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1;
2777 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1;
2779 *prev_stmt_info = vinfo_for_stmt (new_stmt1);
2780 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2;
2781 *prev_stmt_info = vinfo_for_stmt (new_stmt2);
2788 /* For multi-step promotion operation we first generate we call the
2789 function recurcively for every stage. We start from the input type,
2790 create promotion operations to the intermediate types, and then
2791 create promotions to the output type. */
2792 *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp);
2793 vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1,
2794 multi_step_cvt - 1, stmt,
2795 vec_dsts, gsi, slp_node, code1,
2796 code2, decl2, decl2, op_type,
2800 VEC_free (tree, heap, vec_tmp);
2804 /* Function vectorizable_type_promotion
2806 Check if STMT performs a binary or unary operation that involves
2807 type promotion, and if it can be vectorized.
2808 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2809 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2810 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2813 vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi,
2814 gimple *vec_stmt, slp_tree slp_node)
2818 tree op0, op1 = NULL;
2819 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
2820 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2821 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2822 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
2823 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
2827 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2828 stmt_vec_info prev_stmt_info;
2835 tree intermediate_type = NULL_TREE;
2836 int multi_step_cvt = 0;
2837 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2838 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
2840 /* FORNOW: not supported by basic block SLP vectorization. */
2841 gcc_assert (loop_vinfo);
2843 if (!STMT_VINFO_RELEVANT_P (stmt_info))
2846 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2849 /* Is STMT a vectorizable type-promotion operation? */
2850 if (!is_gimple_assign (stmt))
2853 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2856 code = gimple_assign_rhs_code (stmt);
2857 if (!CONVERT_EXPR_CODE_P (code)
2858 && code != WIDEN_MULT_EXPR)
2861 scalar_dest = gimple_assign_lhs (stmt);
2862 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2864 /* Check the operands of the operation. */
2865 op0 = gimple_assign_rhs1 (stmt);
2866 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
2867 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
2868 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
2869 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
2870 && CONVERT_EXPR_CODE_P (code))))
2872 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
2873 &def_stmt, &def, &dt[0], &vectype_in))
2875 if (vect_print_dump_info (REPORT_DETAILS))
2876 fprintf (vect_dump, "use not simple.");
2879 /* If op0 is an external or constant def use a vector type with
2880 the same size as the output vector type. */
2882 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2884 gcc_assert (vectype_in);
2887 if (vect_print_dump_info (REPORT_DETAILS))
2889 fprintf (vect_dump, "no vectype for scalar type ");
2890 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2896 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
2897 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2898 if (nunits_in <= nunits_out)
2901 /* Multiple types in SLP are handled by creating the appropriate number of
2902 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2907 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2909 gcc_assert (ncopies >= 1);
2911 op_type = TREE_CODE_LENGTH (code);
2912 if (op_type == binary_op)
2914 op1 = gimple_assign_rhs2 (stmt);
2915 if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1]))
2917 if (vect_print_dump_info (REPORT_DETAILS))
2918 fprintf (vect_dump, "use not simple.");
2923 /* Supportable by target? */
2924 if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in,
2925 &decl1, &decl2, &code1, &code2,
2926 &multi_step_cvt, &interm_types))
2929 /* Binary widening operation can only be supported directly by the
2931 gcc_assert (!(multi_step_cvt && op_type == binary_op));
2933 if (!vec_stmt) /* transformation not required. */
2935 STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type;
2936 if (vect_print_dump_info (REPORT_DETAILS))
2937 fprintf (vect_dump, "=== vectorizable_promotion ===");
2938 vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL);
2944 if (vect_print_dump_info (REPORT_DETAILS))
2945 fprintf (vect_dump, "transform type promotion operation. ncopies = %d.",
2949 /* In case of multi-step promotion, we first generate promotion operations
2950 to the intermediate types, and then from that types to the final one.
2951 We store vector destination in VEC_DSTS in the correct order for
2952 recursive creation of promotion operations in
2953 vect_create_vectorized_promotion_stmts(). Vector destinations are created
2954 according to TYPES recieved from supportable_widening_operation(). */
2956 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
2958 vec_dsts = VEC_alloc (tree, heap, 1);
2960 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
2961 VEC_quick_push (tree, vec_dsts, vec_dest);
2965 for (i = VEC_length (tree, interm_types) - 1;
2966 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
2968 vec_dest = vect_create_destination_var (scalar_dest,
2970 VEC_quick_push (tree, vec_dsts, vec_dest);
2976 vec_oprnds0 = VEC_alloc (tree, heap,
2977 (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1));
2978 if (op_type == binary_op)
2979 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2982 /* In case the vectorization factor (VF) is bigger than the number
2983 of elements that we can fit in a vectype (nunits), we have to generate
2984 more than one vector stmt - i.e - we need to "unroll" the
2985 vector stmt by a factor VF/nunits. */
2987 prev_stmt_info = NULL;
2988 for (j = 0; j < ncopies; j++)
2994 vect_get_slp_defs (slp_node, &vec_oprnds0, &vec_oprnds1, -1);
2997 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
2998 VEC_quick_push (tree, vec_oprnds0, vec_oprnd0);
2999 if (op_type == binary_op)
3001 vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
3002 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
3008 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3009 VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0);
3010 if (op_type == binary_op)
3012 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
3013 VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1);
3017 /* Arguments are ready. Create the new vector stmts. */
3018 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
3019 vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1,
3020 multi_step_cvt, stmt,
3022 gsi, slp_node, code1, code2,
3023 decl1, decl2, op_type,
3027 VEC_free (tree, heap, vec_dsts);
3028 VEC_free (tree, heap, tmp_vec_dsts);
3029 VEC_free (tree, heap, interm_types);
3030 VEC_free (tree, heap, vec_oprnds0);
3031 VEC_free (tree, heap, vec_oprnds1);
3033 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3038 /* Function vectorizable_store.
3040 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
3042 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3043 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3044 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3047 vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3053 tree vec_oprnd = NULL_TREE;
3054 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3055 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL;
3056 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3057 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3058 struct loop *loop = NULL;
3059 enum machine_mode vec_mode;
3061 enum dr_alignment_support alignment_support_scheme;
3064 enum vect_def_type dt;
3065 stmt_vec_info prev_stmt_info = NULL;
3066 tree dataref_ptr = NULL_TREE;
3067 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3070 gimple next_stmt, first_stmt = NULL;
3071 bool strided_store = false;
3072 unsigned int group_size, i;
3073 VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL;
3075 VEC(tree,heap) *vec_oprnds = NULL;
3076 bool slp = (slp_node != NULL);
3077 unsigned int vec_num;
3078 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3081 loop = LOOP_VINFO_LOOP (loop_vinfo);
3083 /* Multiple types in SLP are handled by creating the appropriate number of
3084 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3089 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3091 gcc_assert (ncopies >= 1);
3093 /* FORNOW. This restriction should be relaxed. */
3094 if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3096 if (vect_print_dump_info (REPORT_DETAILS))
3097 fprintf (vect_dump, "multiple types in nested loop.");
3101 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3104 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3107 /* Is vectorizable store? */
3109 if (!is_gimple_assign (stmt))
3112 scalar_dest = gimple_assign_lhs (stmt);
3113 if (TREE_CODE (scalar_dest) != ARRAY_REF
3114 && TREE_CODE (scalar_dest) != INDIRECT_REF
3115 && TREE_CODE (scalar_dest) != COMPONENT_REF
3116 && TREE_CODE (scalar_dest) != IMAGPART_EXPR
3117 && TREE_CODE (scalar_dest) != REALPART_EXPR
3118 && TREE_CODE (scalar_dest) != MEM_REF)
3121 gcc_assert (gimple_assign_single_p (stmt));
3122 op = gimple_assign_rhs1 (stmt);
3123 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt))
3125 if (vect_print_dump_info (REPORT_DETAILS))
3126 fprintf (vect_dump, "use not simple.");
3130 /* The scalar rhs type needs to be trivially convertible to the vector
3131 component type. This should always be the case. */
3132 if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op)))
3134 if (vect_print_dump_info (REPORT_DETAILS))
3135 fprintf (vect_dump, "??? operands of different types");
3139 vec_mode = TYPE_MODE (vectype);
3140 /* FORNOW. In some cases can vectorize even if data-type not supported
3141 (e.g. - array initialization with 0). */
3142 if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing)
3145 if (!STMT_VINFO_DATA_REF (stmt_info))
3148 if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0)
3150 if (vect_print_dump_info (REPORT_DETAILS))
3151 fprintf (vect_dump, "negative step for store.");
3155 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3157 strided_store = true;
3158 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3159 if (!vect_strided_store_supported (vectype)
3160 && !PURE_SLP_STMT (stmt_info) && !slp)
3163 if (first_stmt == stmt)
3165 /* STMT is the leader of the group. Check the operands of all the
3166 stmts of the group. */
3167 next_stmt = DR_GROUP_NEXT_DR (stmt_info);
3170 gcc_assert (gimple_assign_single_p (next_stmt));
3171 op = gimple_assign_rhs1 (next_stmt);
3172 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt,
3175 if (vect_print_dump_info (REPORT_DETAILS))
3176 fprintf (vect_dump, "use not simple.");
3179 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3184 if (!vec_stmt) /* transformation not required. */
3186 STMT_VINFO_TYPE (stmt_info) = store_vec_info_type;
3187 vect_model_store_cost (stmt_info, ncopies, dt, NULL);
3195 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3196 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3198 DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++;
3201 gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt));
3203 /* We vectorize all the stmts of the interleaving group when we
3204 reach the last stmt in the group. */
3205 if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))
3206 < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt))
3215 strided_store = false;
3216 /* VEC_NUM is the number of vect stmts to be created for this
3218 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3219 first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0);
3220 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3223 /* VEC_NUM is the number of vect stmts to be created for this
3225 vec_num = group_size;
3231 group_size = vec_num = 1;
3234 if (vect_print_dump_info (REPORT_DETAILS))
3235 fprintf (vect_dump, "transform store. ncopies = %d",ncopies);
3237 dr_chain = VEC_alloc (tree, heap, group_size);
3238 oprnds = VEC_alloc (tree, heap, group_size);
3240 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3241 gcc_assert (alignment_support_scheme);
3243 /* In case the vectorization factor (VF) is bigger than the number
3244 of elements that we can fit in a vectype (nunits), we have to generate
3245 more than one vector stmt - i.e - we need to "unroll" the
3246 vector stmt by a factor VF/nunits. For more details see documentation in
3247 vect_get_vec_def_for_copy_stmt. */
3249 /* In case of interleaving (non-unit strided access):
3256 We create vectorized stores starting from base address (the access of the
3257 first stmt in the chain (S2 in the above example), when the last store stmt
3258 of the chain (S4) is reached:
3261 VS2: &base + vec_size*1 = vx0
3262 VS3: &base + vec_size*2 = vx1
3263 VS4: &base + vec_size*3 = vx3
3265 Then permutation statements are generated:
3267 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
3268 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
3271 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3272 (the order of the data-refs in the output of vect_permute_store_chain
3273 corresponds to the order of scalar stmts in the interleaving chain - see
3274 the documentation of vect_permute_store_chain()).
3276 In case of both multiple types and interleaving, above vector stores and
3277 permutation stmts are created for every copy. The result vector stmts are
3278 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
3279 STMT_VINFO_RELATED_STMT for the next copies.
3282 prev_stmt_info = NULL;
3283 for (j = 0; j < ncopies; j++)
3292 /* Get vectorized arguments for SLP_NODE. */
3293 vect_get_slp_defs (slp_node, &vec_oprnds, NULL, -1);
3295 vec_oprnd = VEC_index (tree, vec_oprnds, 0);
3299 /* For interleaved stores we collect vectorized defs for all the
3300 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
3301 used as an input to vect_permute_store_chain(), and OPRNDS as
3302 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
3304 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3305 OPRNDS are of size 1. */
3306 next_stmt = first_stmt;
3307 for (i = 0; i < group_size; i++)
3309 /* Since gaps are not supported for interleaved stores,
3310 GROUP_SIZE is the exact number of stmts in the chain.
3311 Therefore, NEXT_STMT can't be NULL_TREE. In case that
3312 there is no interleaving, GROUP_SIZE is 1, and only one
3313 iteration of the loop will be executed. */
3314 gcc_assert (next_stmt
3315 && gimple_assign_single_p (next_stmt));
3316 op = gimple_assign_rhs1 (next_stmt);
3318 vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt,
3320 VEC_quick_push(tree, dr_chain, vec_oprnd);
3321 VEC_quick_push(tree, oprnds, vec_oprnd);
3322 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3326 /* We should have catched mismatched types earlier. */
3327 gcc_assert (useless_type_conversion_p (vectype,
3328 TREE_TYPE (vec_oprnd)));
3329 dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE,
3330 &dummy, &ptr_incr, false,
3332 gcc_assert (bb_vinfo || !inv_p);
3336 /* For interleaved stores we created vectorized defs for all the
3337 defs stored in OPRNDS in the previous iteration (previous copy).
3338 DR_CHAIN is then used as an input to vect_permute_store_chain(),
3339 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
3341 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3342 OPRNDS are of size 1. */
3343 for (i = 0; i < group_size; i++)
3345 op = VEC_index (tree, oprnds, i);
3346 vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def,
3348 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op);
3349 VEC_replace(tree, dr_chain, i, vec_oprnd);
3350 VEC_replace(tree, oprnds, i, vec_oprnd);
3353 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
3358 result_chain = VEC_alloc (tree, heap, group_size);
3360 if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi,
3365 next_stmt = first_stmt;
3366 for (i = 0; i < vec_num; i++)
3368 struct ptr_info_def *pi;
3371 /* Bump the vector pointer. */
3372 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
3376 vec_oprnd = VEC_index (tree, vec_oprnds, i);
3377 else if (strided_store)
3378 /* For strided stores vectorized defs are interleaved in
3379 vect_permute_store_chain(). */
3380 vec_oprnd = VEC_index (tree, result_chain, i);
3382 data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr,
3383 build_int_cst (reference_alias_ptr_type
3384 (DR_REF (first_dr)), 0));
3385 pi = get_ptr_info (dataref_ptr);
3386 pi->align = TYPE_ALIGN_UNIT (vectype);
3387 if (aligned_access_p (first_dr))
3389 else if (DR_MISALIGNMENT (first_dr) == -1)
3391 TREE_TYPE (data_ref)
3392 = build_aligned_type (TREE_TYPE (data_ref),
3393 TYPE_ALIGN (TREE_TYPE (vectype)));
3394 pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype));
3399 TREE_TYPE (data_ref)
3400 = build_aligned_type (TREE_TYPE (data_ref),
3401 TYPE_ALIGN (TREE_TYPE (vectype)));
3402 pi->misalign = DR_MISALIGNMENT (first_dr);
3405 /* Arguments are ready. Create the new vector stmt. */
3406 new_stmt = gimple_build_assign (data_ref, vec_oprnd);
3407 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3408 mark_symbols_for_renaming (new_stmt);
3414 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3416 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3418 prev_stmt_info = vinfo_for_stmt (new_stmt);
3419 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3425 VEC_free (tree, heap, dr_chain);
3426 VEC_free (tree, heap, oprnds);
3428 VEC_free (tree, heap, result_chain);
3430 VEC_free (tree, heap, vec_oprnds);
3435 /* Given a vector type VECTYPE returns a builtin DECL to be used
3436 for vector permutation and stores a mask into *MASK that implements
3437 reversal of the vector elements. If that is impossible to do
3438 returns NULL (and *MASK is unchanged). */
3441 perm_mask_for_reverse (tree vectype, tree *mask)
3444 tree mask_element_type, mask_type;
3445 tree mask_vec = NULL;
3448 if (!targetm.vectorize.builtin_vec_perm)
3451 builtin_decl = targetm.vectorize.builtin_vec_perm (vectype,
3452 &mask_element_type);
3453 if (!builtin_decl || !mask_element_type)
3456 mask_type = get_vectype_for_scalar_type (mask_element_type);
3457 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3458 if (TYPE_VECTOR_SUBPARTS (vectype) != TYPE_VECTOR_SUBPARTS (mask_type))
3461 for (i = 0; i < nunits; i++)
3462 mask_vec = tree_cons (NULL, build_int_cst (mask_element_type, i), mask_vec);
3463 mask_vec = build_vector (mask_type, mask_vec);
3465 if (!targetm.vectorize.builtin_vec_perm_ok (vectype, mask_vec))
3469 return builtin_decl;
3472 /* Given a vector variable X, that was generated for the scalar LHS of
3473 STMT, generate instructions to reverse the vector elements of X,
3474 insert them a *GSI and return the permuted vector variable. */
3477 reverse_vec_elements (tree x, gimple stmt, gimple_stmt_iterator *gsi)
3479 tree vectype = TREE_TYPE (x);
3480 tree mask_vec, builtin_decl;
3481 tree perm_dest, data_ref;
3484 builtin_decl = perm_mask_for_reverse (vectype, &mask_vec);
3486 perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype);
3488 /* Generate the permute statement. */
3489 perm_stmt = gimple_build_call (builtin_decl, 3, x, x, mask_vec);
3490 data_ref = make_ssa_name (perm_dest, perm_stmt);
3491 gimple_call_set_lhs (perm_stmt, data_ref);
3492 vect_finish_stmt_generation (stmt, perm_stmt, gsi);
3497 /* vectorizable_load.
3499 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
3501 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3502 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3503 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3506 vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3507 slp_tree slp_node, slp_instance slp_node_instance)
3510 tree vec_dest = NULL;
3511 tree data_ref = NULL;
3512 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3513 stmt_vec_info prev_stmt_info;
3514 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3515 struct loop *loop = NULL;
3516 struct loop *containing_loop = (gimple_bb (stmt))->loop_father;
3517 bool nested_in_vect_loop = false;
3518 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
3519 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3521 enum machine_mode mode;
3522 gimple new_stmt = NULL;
3524 enum dr_alignment_support alignment_support_scheme;
3525 tree dataref_ptr = NULL_TREE;
3527 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3529 int i, j, group_size;
3530 tree msq = NULL_TREE, lsq;
3531 tree offset = NULL_TREE;
3532 tree realignment_token = NULL_TREE;
3534 VEC(tree,heap) *dr_chain = NULL;
3535 bool strided_load = false;
3540 bool compute_in_loop = false;
3541 struct loop *at_loop;
3543 bool slp = (slp_node != NULL);
3544 bool slp_perm = false;
3545 enum tree_code code;
3546 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3551 loop = LOOP_VINFO_LOOP (loop_vinfo);
3552 nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
3553 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
3558 /* Multiple types in SLP are handled by creating the appropriate number of
3559 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3564 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3566 gcc_assert (ncopies >= 1);
3568 /* FORNOW. This restriction should be relaxed. */
3569 if (nested_in_vect_loop && ncopies > 1)
3571 if (vect_print_dump_info (REPORT_DETAILS))
3572 fprintf (vect_dump, "multiple types in nested loop.");
3576 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3579 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3582 /* Is vectorizable load? */
3583 if (!is_gimple_assign (stmt))
3586 scalar_dest = gimple_assign_lhs (stmt);
3587 if (TREE_CODE (scalar_dest) != SSA_NAME)
3590 code = gimple_assign_rhs_code (stmt);
3591 if (code != ARRAY_REF
3592 && code != INDIRECT_REF
3593 && code != COMPONENT_REF
3594 && code != IMAGPART_EXPR
3595 && code != REALPART_EXPR
3599 if (!STMT_VINFO_DATA_REF (stmt_info))
3602 negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0;
3603 if (negative && ncopies > 1)
3605 if (vect_print_dump_info (REPORT_DETAILS))
3606 fprintf (vect_dump, "multiple types with negative step.");
3610 scalar_type = TREE_TYPE (DR_REF (dr));
3611 mode = TYPE_MODE (vectype);
3613 /* FORNOW. In some cases can vectorize even if data-type not supported
3614 (e.g. - data copies). */
3615 if (optab_handler (mov_optab, mode) == CODE_FOR_nothing)
3617 if (vect_print_dump_info (REPORT_DETAILS))
3618 fprintf (vect_dump, "Aligned load, but unsupported type.");
3622 /* The vector component type needs to be trivially convertible to the
3623 scalar lhs. This should always be the case. */
3624 if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype)))
3626 if (vect_print_dump_info (REPORT_DETAILS))
3627 fprintf (vect_dump, "??? operands of different types");
3631 /* Check if the load is a part of an interleaving chain. */
3632 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3634 strided_load = true;
3636 gcc_assert (! nested_in_vect_loop);
3638 /* Check if interleaving is supported. */
3639 if (!vect_strided_load_supported (vectype)
3640 && !PURE_SLP_STMT (stmt_info) && !slp)
3646 gcc_assert (!strided_load);
3647 alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
3648 if (alignment_support_scheme != dr_aligned
3649 && alignment_support_scheme != dr_unaligned_supported)
3651 if (vect_print_dump_info (REPORT_DETAILS))
3652 fprintf (vect_dump, "negative step but alignment required.");
3655 if (!perm_mask_for_reverse (vectype, NULL))
3657 if (vect_print_dump_info (REPORT_DETAILS))
3658 fprintf (vect_dump, "negative step and reversing not supported.");
3663 if (!vec_stmt) /* transformation not required. */
3665 STMT_VINFO_TYPE (stmt_info) = load_vec_info_type;
3666 vect_model_load_cost (stmt_info, ncopies, NULL);
3670 if (vect_print_dump_info (REPORT_DETAILS))
3671 fprintf (vect_dump, "transform load.");
3677 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3678 /* Check if the chain of loads is already vectorized. */
3679 if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)))
3681 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3684 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3685 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3687 /* VEC_NUM is the number of vect stmts to be created for this group. */
3690 strided_load = false;
3691 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3692 if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance))
3696 vec_num = group_size;
3698 dr_chain = VEC_alloc (tree, heap, vec_num);
3704 group_size = vec_num = 1;
3707 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3708 gcc_assert (alignment_support_scheme);
3710 /* In case the vectorization factor (VF) is bigger than the number
3711 of elements that we can fit in a vectype (nunits), we have to generate
3712 more than one vector stmt - i.e - we need to "unroll" the
3713 vector stmt by a factor VF/nunits. In doing so, we record a pointer
3714 from one copy of the vector stmt to the next, in the field
3715 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
3716 stages to find the correct vector defs to be used when vectorizing
3717 stmts that use the defs of the current stmt. The example below
3718 illustrates the vectorization process when VF=16 and nunits=4 (i.e., we
3719 need to create 4 vectorized stmts):
3721 before vectorization:
3722 RELATED_STMT VEC_STMT
3726 step 1: vectorize stmt S1:
3727 We first create the vector stmt VS1_0, and, as usual, record a
3728 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
3729 Next, we create the vector stmt VS1_1, and record a pointer to
3730 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
3731 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
3733 RELATED_STMT VEC_STMT
3734 VS1_0: vx0 = memref0 VS1_1 -
3735 VS1_1: vx1 = memref1 VS1_2 -
3736 VS1_2: vx2 = memref2 VS1_3 -
3737 VS1_3: vx3 = memref3 - -
3738 S1: x = load - VS1_0
3741 See in documentation in vect_get_vec_def_for_stmt_copy for how the
3742 information we recorded in RELATED_STMT field is used to vectorize
3745 /* In case of interleaving (non-unit strided access):
3752 Vectorized loads are created in the order of memory accesses
3753 starting from the access of the first stmt of the chain:
3756 VS2: vx1 = &base + vec_size*1
3757 VS3: vx3 = &base + vec_size*2
3758 VS4: vx4 = &base + vec_size*3
3760 Then permutation statements are generated:
3762 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
3763 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
3766 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3767 (the order of the data-refs in the output of vect_permute_load_chain
3768 corresponds to the order of scalar stmts in the interleaving chain - see
3769 the documentation of vect_permute_load_chain()).
3770 The generation of permutation stmts and recording them in
3771 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
3773 In case of both multiple types and interleaving, the vector loads and
3774 permutation stmts above are created for every copy. The result vector
3775 stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the
3776 corresponding STMT_VINFO_RELATED_STMT for the next copies. */
3778 /* If the data reference is aligned (dr_aligned) or potentially unaligned
3779 on a target that supports unaligned accesses (dr_unaligned_supported)
3780 we generate the following code:
3784 p = p + indx * vectype_size;
3789 Otherwise, the data reference is potentially unaligned on a target that
3790 does not support unaligned accesses (dr_explicit_realign_optimized) -
3791 then generate the following code, in which the data in each iteration is
3792 obtained by two vector loads, one from the previous iteration, and one
3793 from the current iteration:
3795 msq_init = *(floor(p1))
3796 p2 = initial_addr + VS - 1;
3797 realignment_token = call target_builtin;
3800 p2 = p2 + indx * vectype_size
3802 vec_dest = realign_load (msq, lsq, realignment_token)
3807 /* If the misalignment remains the same throughout the execution of the
3808 loop, we can create the init_addr and permutation mask at the loop
3809 preheader. Otherwise, it needs to be created inside the loop.
3810 This can only occur when vectorizing memory accesses in the inner-loop
3811 nested within an outer-loop that is being vectorized. */
3813 if (loop && nested_in_vect_loop_p (loop, stmt)
3814 && (TREE_INT_CST_LOW (DR_STEP (dr))
3815 % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0))
3817 gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized);
3818 compute_in_loop = true;
3821 if ((alignment_support_scheme == dr_explicit_realign_optimized
3822 || alignment_support_scheme == dr_explicit_realign)
3823 && !compute_in_loop)
3825 msq = vect_setup_realignment (first_stmt, gsi, &realignment_token,
3826 alignment_support_scheme, NULL_TREE,
3828 if (alignment_support_scheme == dr_explicit_realign_optimized)
3830 phi = SSA_NAME_DEF_STMT (msq);
3831 offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
3838 offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1);
3840 prev_stmt_info = NULL;
3841 for (j = 0; j < ncopies; j++)
3843 /* 1. Create the vector pointer update chain. */
3845 dataref_ptr = vect_create_data_ref_ptr (first_stmt,
3847 &dummy, &ptr_incr, false,
3851 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
3853 for (i = 0; i < vec_num; i++)
3856 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
3859 /* 2. Create the vector-load in the loop. */
3860 switch (alignment_support_scheme)
3863 case dr_unaligned_supported:
3865 struct ptr_info_def *pi;
3867 = build2 (MEM_REF, vectype, dataref_ptr,
3868 build_int_cst (reference_alias_ptr_type
3869 (DR_REF (first_dr)), 0));
3870 pi = get_ptr_info (dataref_ptr);
3871 pi->align = TYPE_ALIGN_UNIT (vectype);
3872 if (alignment_support_scheme == dr_aligned)
3874 gcc_assert (aligned_access_p (first_dr));
3877 else if (DR_MISALIGNMENT (first_dr) == -1)
3879 TREE_TYPE (data_ref)
3880 = build_aligned_type (TREE_TYPE (data_ref),
3881 TYPE_ALIGN (TREE_TYPE (vectype)));
3882 pi->align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype));
3887 TREE_TYPE (data_ref)
3888 = build_aligned_type (TREE_TYPE (data_ref),
3889 TYPE_ALIGN (TREE_TYPE (vectype)));
3890 pi->misalign = DR_MISALIGNMENT (first_dr);
3894 case dr_explicit_realign:
3897 tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
3899 if (compute_in_loop)
3900 msq = vect_setup_realignment (first_stmt, gsi,
3902 dr_explicit_realign,
3905 new_stmt = gimple_build_assign_with_ops
3906 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
3908 (TREE_TYPE (dataref_ptr),
3909 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
3910 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
3911 gimple_assign_set_lhs (new_stmt, ptr);
3912 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3914 = build2 (MEM_REF, vectype, ptr,
3915 build_int_cst (reference_alias_ptr_type
3916 (DR_REF (first_dr)), 0));
3917 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3918 new_stmt = gimple_build_assign (vec_dest, data_ref);
3919 new_temp = make_ssa_name (vec_dest, new_stmt);
3920 gimple_assign_set_lhs (new_stmt, new_temp);
3921 gimple_set_vdef (new_stmt, gimple_vdef (stmt));
3922 gimple_set_vuse (new_stmt, gimple_vuse (stmt));
3923 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3926 bump = size_binop (MULT_EXPR, vs_minus_1,
3927 TYPE_SIZE_UNIT (scalar_type));
3928 ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump);
3929 new_stmt = gimple_build_assign_with_ops
3930 (BIT_AND_EXPR, NULL_TREE, ptr,
3933 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
3934 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
3935 gimple_assign_set_lhs (new_stmt, ptr);
3936 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3938 = build2 (MEM_REF, vectype, ptr,
3939 build_int_cst (reference_alias_ptr_type
3940 (DR_REF (first_dr)), 0));
3943 case dr_explicit_realign_optimized:
3944 new_stmt = gimple_build_assign_with_ops
3945 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
3947 (TREE_TYPE (dataref_ptr),
3948 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
3949 new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
3950 gimple_assign_set_lhs (new_stmt, new_temp);
3951 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3953 = build2 (MEM_REF, vectype, new_temp,
3954 build_int_cst (reference_alias_ptr_type
3955 (DR_REF (first_dr)), 0));
3960 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3961 new_stmt = gimple_build_assign (vec_dest, data_ref);
3962 new_temp = make_ssa_name (vec_dest, new_stmt);
3963 gimple_assign_set_lhs (new_stmt, new_temp);
3964 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3965 mark_symbols_for_renaming (new_stmt);
3967 /* 3. Handle explicit realignment if necessary/supported. Create in
3968 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
3969 if (alignment_support_scheme == dr_explicit_realign_optimized
3970 || alignment_support_scheme == dr_explicit_realign)
3974 lsq = gimple_assign_lhs (new_stmt);
3975 if (!realignment_token)
3976 realignment_token = dataref_ptr;
3977 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3978 tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq,
3980 new_stmt = gimple_build_assign (vec_dest, tmp);
3981 new_temp = make_ssa_name (vec_dest, new_stmt);
3982 gimple_assign_set_lhs (new_stmt, new_temp);
3983 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3985 if (alignment_support_scheme == dr_explicit_realign_optimized)
3988 if (i == vec_num - 1 && j == ncopies - 1)
3989 add_phi_arg (phi, lsq, loop_latch_edge (containing_loop),
3995 /* 4. Handle invariant-load. */
3996 if (inv_p && !bb_vinfo)
3998 gcc_assert (!strided_load);
3999 gcc_assert (nested_in_vect_loop_p (loop, stmt));
4004 tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type);
4006 /* CHECKME: bitpos depends on endianess? */
4007 bitpos = bitsize_zero_node;
4008 vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp,
4011 vect_create_destination_var (scalar_dest, NULL_TREE);
4012 new_stmt = gimple_build_assign (vec_dest, vec_inv);
4013 new_temp = make_ssa_name (vec_dest, new_stmt);
4014 gimple_assign_set_lhs (new_stmt, new_temp);
4015 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4017 for (k = nunits - 1; k >= 0; --k)
4018 t = tree_cons (NULL_TREE, new_temp, t);
4019 /* FIXME: use build_constructor directly. */
4020 vec_inv = build_constructor_from_list (vectype, t);
4021 new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi);
4022 new_stmt = SSA_NAME_DEF_STMT (new_temp);
4025 gcc_unreachable (); /* FORNOW. */
4030 new_temp = reverse_vec_elements (new_temp, stmt, gsi);
4031 new_stmt = SSA_NAME_DEF_STMT (new_temp);
4034 /* Collect vector loads and later create their permutation in
4035 vect_transform_strided_load (). */
4036 if (strided_load || slp_perm)
4037 VEC_quick_push (tree, dr_chain, new_temp);
4039 /* Store vector loads in the corresponding SLP_NODE. */
4040 if (slp && !slp_perm)
4041 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
4044 if (slp && !slp_perm)
4049 if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf,
4050 slp_node_instance, false))
4052 VEC_free (tree, heap, dr_chain);
4060 if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi))
4063 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
4064 VEC_free (tree, heap, dr_chain);
4065 dr_chain = VEC_alloc (tree, heap, group_size);
4070 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4072 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4073 prev_stmt_info = vinfo_for_stmt (new_stmt);
4079 VEC_free (tree, heap, dr_chain);
4084 /* Function vect_is_simple_cond.
4087 LOOP - the loop that is being vectorized.
4088 COND - Condition that is checked for simple use.
4090 Returns whether a COND can be vectorized. Checks whether
4091 condition operands are supportable using vec_is_simple_use. */
4094 vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo)
4098 enum vect_def_type dt;
4100 if (!COMPARISON_CLASS_P (cond))
4103 lhs = TREE_OPERAND (cond, 0);
4104 rhs = TREE_OPERAND (cond, 1);
4106 if (TREE_CODE (lhs) == SSA_NAME)
4108 gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs);
4109 if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def,
4113 else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST
4114 && TREE_CODE (lhs) != FIXED_CST)
4117 if (TREE_CODE (rhs) == SSA_NAME)
4119 gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs);
4120 if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def,
4124 else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST
4125 && TREE_CODE (rhs) != FIXED_CST)
4131 /* vectorizable_condition.
4133 Check if STMT is conditional modify expression that can be vectorized.
4134 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4135 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
4138 When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable
4139 to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in
4140 else caluse if it is 2).
4142 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4145 vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi,
4146 gimple *vec_stmt, tree reduc_def, int reduc_index)
4148 tree scalar_dest = NULL_TREE;
4149 tree vec_dest = NULL_TREE;
4150 tree op = NULL_TREE;
4151 tree cond_expr, then_clause, else_clause;
4152 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4153 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4154 tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE;
4155 tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE;
4156 tree vec_compare, vec_cond_expr;
4158 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4159 enum machine_mode vec_mode;
4161 enum vect_def_type dt, dts[4];
4162 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
4163 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
4164 enum tree_code code;
4165 stmt_vec_info prev_stmt_info = NULL;
4168 /* FORNOW: unsupported in basic block SLP. */
4169 gcc_assert (loop_vinfo);
4171 gcc_assert (ncopies >= 1);
4172 if (reduc_index && ncopies > 1)
4173 return false; /* FORNOW */
4175 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4178 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
4179 && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle
4183 /* FORNOW: SLP not supported. */
4184 if (STMT_SLP_TYPE (stmt_info))
4187 /* FORNOW: not yet supported. */
4188 if (STMT_VINFO_LIVE_P (stmt_info))
4190 if (vect_print_dump_info (REPORT_DETAILS))
4191 fprintf (vect_dump, "value used after loop.");
4195 /* Is vectorizable conditional operation? */
4196 if (!is_gimple_assign (stmt))
4199 code = gimple_assign_rhs_code (stmt);
4201 if (code != COND_EXPR)
4204 gcc_assert (gimple_assign_single_p (stmt));
4205 op = gimple_assign_rhs1 (stmt);
4206 cond_expr = TREE_OPERAND (op, 0);
4207 then_clause = TREE_OPERAND (op, 1);
4208 else_clause = TREE_OPERAND (op, 2);
4210 if (!vect_is_simple_cond (cond_expr, loop_vinfo))
4213 /* We do not handle two different vector types for the condition
4215 if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)),
4216 TREE_TYPE (vectype)))
4219 if (TREE_CODE (then_clause) == SSA_NAME)
4221 gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause);
4222 if (!vect_is_simple_use (then_clause, loop_vinfo, NULL,
4223 &then_def_stmt, &def, &dt))
4226 else if (TREE_CODE (then_clause) != INTEGER_CST
4227 && TREE_CODE (then_clause) != REAL_CST
4228 && TREE_CODE (then_clause) != FIXED_CST)
4231 if (TREE_CODE (else_clause) == SSA_NAME)
4233 gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause);
4234 if (!vect_is_simple_use (else_clause, loop_vinfo, NULL,
4235 &else_def_stmt, &def, &dt))
4238 else if (TREE_CODE (else_clause) != INTEGER_CST
4239 && TREE_CODE (else_clause) != REAL_CST
4240 && TREE_CODE (else_clause) != FIXED_CST)
4244 vec_mode = TYPE_MODE (vectype);
4248 STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type;
4249 return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode);
4255 scalar_dest = gimple_assign_lhs (stmt);
4256 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4258 /* Handle cond expr. */
4259 for (j = 0; j < ncopies; j++)
4266 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0),
4268 vect_is_simple_use (TREE_OPERAND (cond_expr, 0), loop_vinfo,
4269 NULL, >emp, &def, &dts[0]);
4271 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1),
4273 vect_is_simple_use (TREE_OPERAND (cond_expr, 1), loop_vinfo,
4274 NULL, >emp, &def, &dts[1]);
4275 if (reduc_index == 1)
4276 vec_then_clause = reduc_def;
4279 vec_then_clause = vect_get_vec_def_for_operand (then_clause,
4281 vect_is_simple_use (then_clause, loop_vinfo,
4282 NULL, >emp, &def, &dts[2]);
4284 if (reduc_index == 2)
4285 vec_else_clause = reduc_def;
4288 vec_else_clause = vect_get_vec_def_for_operand (else_clause,
4290 vect_is_simple_use (else_clause, loop_vinfo,
4291 NULL, >emp, &def, &dts[3]);
4296 vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], vec_cond_lhs);
4297 vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], vec_cond_rhs);
4298 vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2],
4300 vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3],
4304 /* Arguments are ready. Create the new vector stmt. */
4305 vec_compare = build2 (TREE_CODE (cond_expr), vectype,
4306 vec_cond_lhs, vec_cond_rhs);
4307 vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
4308 vec_compare, vec_then_clause, vec_else_clause);
4310 new_stmt = gimple_build_assign (vec_dest, vec_cond_expr);
4311 new_temp = make_ssa_name (vec_dest, new_stmt);
4312 gimple_assign_set_lhs (new_stmt, new_temp);
4313 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4315 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4317 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4319 prev_stmt_info = vinfo_for_stmt (new_stmt);
4326 /* Make sure the statement is vectorizable. */
4329 vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node)
4331 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4332 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
4333 enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info);
4335 tree scalar_type, vectype;
4337 if (vect_print_dump_info (REPORT_DETAILS))
4339 fprintf (vect_dump, "==> examining statement: ");
4340 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4343 if (gimple_has_volatile_ops (stmt))
4345 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4346 fprintf (vect_dump, "not vectorized: stmt has volatile operands");
4351 /* Skip stmts that do not need to be vectorized. In loops this is expected
4353 - the COND_EXPR which is the loop exit condition
4354 - any LABEL_EXPRs in the loop
4355 - computations that are used only for array indexing or loop control.
4356 In basic blocks we only analyze statements that are a part of some SLP
4357 instance, therefore, all the statements are relevant. */
4359 if (!STMT_VINFO_RELEVANT_P (stmt_info)
4360 && !STMT_VINFO_LIVE_P (stmt_info))
4362 if (vect_print_dump_info (REPORT_DETAILS))
4363 fprintf (vect_dump, "irrelevant.");
4368 switch (STMT_VINFO_DEF_TYPE (stmt_info))
4370 case vect_internal_def:
4373 case vect_reduction_def:
4374 case vect_nested_cycle:
4375 gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer
4376 || relevance == vect_used_in_outer_by_reduction
4377 || relevance == vect_unused_in_scope));
4380 case vect_induction_def:
4381 case vect_constant_def:
4382 case vect_external_def:
4383 case vect_unknown_def_type:
4390 gcc_assert (PURE_SLP_STMT (stmt_info));
4392 scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
4393 if (vect_print_dump_info (REPORT_DETAILS))
4395 fprintf (vect_dump, "get vectype for scalar type: ");
4396 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4399 vectype = get_vectype_for_scalar_type (scalar_type);
4402 if (vect_print_dump_info (REPORT_DETAILS))
4404 fprintf (vect_dump, "not SLPed: unsupported data-type ");
4405 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4410 if (vect_print_dump_info (REPORT_DETAILS))
4412 fprintf (vect_dump, "vectype: ");
4413 print_generic_expr (vect_dump, vectype, TDF_SLIM);
4416 STMT_VINFO_VECTYPE (stmt_info) = vectype;
4419 if (STMT_VINFO_RELEVANT_P (stmt_info))
4421 gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt))));
4422 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
4423 *need_to_vectorize = true;
4428 && (STMT_VINFO_RELEVANT_P (stmt_info)
4429 || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def))
4430 ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL)
4431 || vectorizable_type_demotion (stmt, NULL, NULL, NULL)
4432 || vectorizable_conversion (stmt, NULL, NULL, NULL)
4433 || vectorizable_operation (stmt, NULL, NULL, NULL)
4434 || vectorizable_assignment (stmt, NULL, NULL, NULL)
4435 || vectorizable_load (stmt, NULL, NULL, NULL, NULL)
4436 || vectorizable_call (stmt, NULL, NULL)
4437 || vectorizable_store (stmt, NULL, NULL, NULL)
4438 || vectorizable_reduction (stmt, NULL, NULL, NULL)
4439 || vectorizable_condition (stmt, NULL, NULL, NULL, 0));
4443 ok = (vectorizable_operation (stmt, NULL, NULL, node)
4444 || vectorizable_assignment (stmt, NULL, NULL, node)
4445 || vectorizable_load (stmt, NULL, NULL, node, NULL)
4446 || vectorizable_store (stmt, NULL, NULL, node));
4451 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4453 fprintf (vect_dump, "not vectorized: relevant stmt not ");
4454 fprintf (vect_dump, "supported: ");
4455 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4464 /* Stmts that are (also) "live" (i.e. - that are used out of the loop)
4465 need extra handling, except for vectorizable reductions. */
4466 if (STMT_VINFO_LIVE_P (stmt_info)
4467 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4468 ok = vectorizable_live_operation (stmt, NULL, NULL);
4472 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4474 fprintf (vect_dump, "not vectorized: live stmt not ");
4475 fprintf (vect_dump, "supported: ");
4476 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4482 if (!PURE_SLP_STMT (stmt_info))
4484 /* Groups of strided accesses whose size is not a power of 2 are not
4485 vectorizable yet using loop-vectorization. Therefore, if this stmt
4486 feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and
4487 loop-based vectorized), the loop cannot be vectorized. */
4488 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
4489 && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt (
4490 DR_GROUP_FIRST_DR (stmt_info)))) == -1)
4492 if (vect_print_dump_info (REPORT_DETAILS))
4494 fprintf (vect_dump, "not vectorized: the size of group "
4495 "of strided accesses is not a power of 2");
4496 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4507 /* Function vect_transform_stmt.
4509 Create a vectorized stmt to replace STMT, and insert it at BSI. */
4512 vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi,
4513 bool *strided_store, slp_tree slp_node,
4514 slp_instance slp_node_instance)
4516 bool is_store = false;
4517 gimple vec_stmt = NULL;
4518 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4519 gimple orig_stmt_in_pattern, orig_scalar_stmt = stmt;
4522 switch (STMT_VINFO_TYPE (stmt_info))
4524 case type_demotion_vec_info_type:
4525 done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node);
4529 case type_promotion_vec_info_type:
4530 done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node);
4534 case type_conversion_vec_info_type:
4535 done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node);
4539 case induc_vec_info_type:
4540 gcc_assert (!slp_node);
4541 done = vectorizable_induction (stmt, gsi, &vec_stmt);
4545 case op_vec_info_type:
4546 done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node);
4550 case assignment_vec_info_type:
4551 done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node);
4555 case load_vec_info_type:
4556 done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node,
4561 case store_vec_info_type:
4562 done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node);
4564 if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node)
4566 /* In case of interleaving, the whole chain is vectorized when the
4567 last store in the chain is reached. Store stmts before the last
4568 one are skipped, and there vec_stmt_info shouldn't be freed
4570 *strided_store = true;
4571 if (STMT_VINFO_VEC_STMT (stmt_info))
4578 case condition_vec_info_type:
4579 gcc_assert (!slp_node);
4580 done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0);
4584 case call_vec_info_type:
4585 gcc_assert (!slp_node);
4586 done = vectorizable_call (stmt, gsi, &vec_stmt);
4587 stmt = gsi_stmt (*gsi);
4590 case reduc_vec_info_type:
4591 done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node);
4596 if (!STMT_VINFO_LIVE_P (stmt_info))
4598 if (vect_print_dump_info (REPORT_DETAILS))
4599 fprintf (vect_dump, "stmt not supported.");
4604 /* Handle inner-loop stmts whose DEF is used in the loop-nest that
4605 is being vectorized, but outside the immediately enclosing loop. */
4607 && STMT_VINFO_LOOP_VINFO (stmt_info)
4608 && nested_in_vect_loop_p (LOOP_VINFO_LOOP (
4609 STMT_VINFO_LOOP_VINFO (stmt_info)), stmt)
4610 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type
4611 && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer
4612 || STMT_VINFO_RELEVANT (stmt_info) ==
4613 vect_used_in_outer_by_reduction))
4615 struct loop *innerloop = LOOP_VINFO_LOOP (
4616 STMT_VINFO_LOOP_VINFO (stmt_info))->inner;
4617 imm_use_iterator imm_iter;
4618 use_operand_p use_p;
4622 if (vect_print_dump_info (REPORT_DETAILS))
4623 fprintf (vect_dump, "Record the vdef for outer-loop vectorization.");
4625 /* Find the relevant loop-exit phi-node, and reord the vec_stmt there
4626 (to be used when vectorizing outer-loop stmts that use the DEF of
4628 if (gimple_code (stmt) == GIMPLE_PHI)
4629 scalar_dest = PHI_RESULT (stmt);
4631 scalar_dest = gimple_assign_lhs (stmt);
4633 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4635 if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p))))
4637 exit_phi = USE_STMT (use_p);
4638 STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt;
4643 /* Handle stmts whose DEF is used outside the loop-nest that is
4644 being vectorized. */
4645 if (STMT_VINFO_LIVE_P (stmt_info)
4646 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4648 done = vectorizable_live_operation (stmt, gsi, &vec_stmt);
4654 STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
4655 orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info);
4656 if (orig_stmt_in_pattern)
4658 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern);
4659 /* STMT was inserted by the vectorizer to replace a computation idiom.
4660 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
4661 computed this idiom. We need to record a pointer to VEC_STMT in
4662 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
4663 documentation of vect_pattern_recog. */
4664 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
4666 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo)
4667 == orig_scalar_stmt);
4668 STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
4677 /* Remove a group of stores (for SLP or interleaving), free their
4681 vect_remove_stores (gimple first_stmt)
4683 gimple next = first_stmt;
4685 gimple_stmt_iterator next_si;
4689 /* Free the attached stmt_vec_info and remove the stmt. */
4690 next_si = gsi_for_stmt (next);
4691 gsi_remove (&next_si, true);
4692 tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next));
4693 free_stmt_vec_info (next);
4699 /* Function new_stmt_vec_info.
4701 Create and initialize a new stmt_vec_info struct for STMT. */
4704 new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo,
4705 bb_vec_info bb_vinfo)
4708 res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info));
4710 STMT_VINFO_TYPE (res) = undef_vec_info_type;
4711 STMT_VINFO_STMT (res) = stmt;
4712 STMT_VINFO_LOOP_VINFO (res) = loop_vinfo;
4713 STMT_VINFO_BB_VINFO (res) = bb_vinfo;
4714 STMT_VINFO_RELEVANT (res) = vect_unused_in_scope;
4715 STMT_VINFO_LIVE_P (res) = false;
4716 STMT_VINFO_VECTYPE (res) = NULL;
4717 STMT_VINFO_VEC_STMT (res) = NULL;
4718 STMT_VINFO_VECTORIZABLE (res) = true;
4719 STMT_VINFO_IN_PATTERN_P (res) = false;
4720 STMT_VINFO_RELATED_STMT (res) = NULL;
4721 STMT_VINFO_DATA_REF (res) = NULL;
4723 STMT_VINFO_DR_BASE_ADDRESS (res) = NULL;
4724 STMT_VINFO_DR_OFFSET (res) = NULL;
4725 STMT_VINFO_DR_INIT (res) = NULL;
4726 STMT_VINFO_DR_STEP (res) = NULL;
4727 STMT_VINFO_DR_ALIGNED_TO (res) = NULL;
4729 if (gimple_code (stmt) == GIMPLE_PHI
4730 && is_loop_header_bb_p (gimple_bb (stmt)))
4731 STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
4733 STMT_VINFO_DEF_TYPE (res) = vect_internal_def;
4735 STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5);
4736 STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0;
4737 STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0;
4738 STMT_SLP_TYPE (res) = loop_vect;
4739 DR_GROUP_FIRST_DR (res) = NULL;
4740 DR_GROUP_NEXT_DR (res) = NULL;
4741 DR_GROUP_SIZE (res) = 0;
4742 DR_GROUP_STORE_COUNT (res) = 0;
4743 DR_GROUP_GAP (res) = 0;
4744 DR_GROUP_SAME_DR_STMT (res) = NULL;
4745 DR_GROUP_READ_WRITE_DEPENDENCE (res) = false;
4751 /* Create a hash table for stmt_vec_info. */
4754 init_stmt_vec_info_vec (void)
4756 gcc_assert (!stmt_vec_info_vec);
4757 stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50);
4761 /* Free hash table for stmt_vec_info. */
4764 free_stmt_vec_info_vec (void)
4766 gcc_assert (stmt_vec_info_vec);
4767 VEC_free (vec_void_p, heap, stmt_vec_info_vec);
4771 /* Free stmt vectorization related info. */
4774 free_stmt_vec_info (gimple stmt)
4776 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4781 VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info));
4782 set_vinfo_for_stmt (stmt, NULL);
4787 /* Function get_vectype_for_scalar_type.
4789 Returns the vector type corresponding to SCALAR_TYPE as supported
4793 get_vectype_for_scalar_type (tree scalar_type)
4795 enum machine_mode inner_mode = TYPE_MODE (scalar_type);
4796 unsigned int nbytes = GET_MODE_SIZE (inner_mode);
4801 || (nbytes >= targetm.vectorize.units_per_simd_word (inner_mode)))
4804 /* We can't build a vector type of elements with alignment bigger than
4806 if (nbytes < TYPE_ALIGN_UNIT (scalar_type))
4809 /* If we'd build a vector type of elements whose mode precision doesn't
4810 match their types precision we'll get mismatched types on vector
4811 extracts via BIT_FIELD_REFs. This effectively means we disable
4812 vectorization of bool and/or enum types in some languages. */
4813 if (INTEGRAL_TYPE_P (scalar_type)
4814 && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type))
4817 /* FORNOW: Only a single vector size per mode
4818 (TARGET_VECTORIZE_UNITS_PER_SIMD_WORD) is expected. */
4819 nunits = targetm.vectorize.units_per_simd_word (inner_mode) / nbytes;
4821 vectype = build_vector_type (scalar_type, nunits);
4822 if (vect_print_dump_info (REPORT_DETAILS))
4824 fprintf (vect_dump, "get vectype with %d units of type ", nunits);
4825 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4831 if (vect_print_dump_info (REPORT_DETAILS))
4833 fprintf (vect_dump, "vectype: ");
4834 print_generic_expr (vect_dump, vectype, TDF_SLIM);
4837 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
4838 && !INTEGRAL_MODE_P (TYPE_MODE (vectype)))
4840 if (vect_print_dump_info (REPORT_DETAILS))
4841 fprintf (vect_dump, "mode not supported by target.");
4848 /* Function get_same_sized_vectype
4850 Returns a vector type corresponding to SCALAR_TYPE of size
4851 VECTOR_TYPE if supported by the target. */
4854 get_same_sized_vectype (tree scalar_type, tree vector_type ATTRIBUTE_UNUSED)
4856 return get_vectype_for_scalar_type (scalar_type);
4859 /* Function vect_is_simple_use.
4862 LOOP_VINFO - the vect info of the loop that is being vectorized.
4863 BB_VINFO - the vect info of the basic block that is being vectorized.
4864 OPERAND - operand of a stmt in the loop or bb.
4865 DEF - the defining stmt in case OPERAND is an SSA_NAME.
4867 Returns whether a stmt with OPERAND can be vectorized.
4868 For loops, supportable operands are constants, loop invariants, and operands
4869 that are defined by the current iteration of the loop. Unsupportable
4870 operands are those that are defined by a previous iteration of the loop (as
4871 is the case in reduction/induction computations).
4872 For basic blocks, supportable operands are constants and bb invariants.
4873 For now, operands defined outside the basic block are not supported. */
4876 vect_is_simple_use (tree operand, loop_vec_info loop_vinfo,
4877 bb_vec_info bb_vinfo, gimple *def_stmt,
4878 tree *def, enum vect_def_type *dt)
4881 stmt_vec_info stmt_vinfo;
4882 struct loop *loop = NULL;
4885 loop = LOOP_VINFO_LOOP (loop_vinfo);
4890 if (vect_print_dump_info (REPORT_DETAILS))
4892 fprintf (vect_dump, "vect_is_simple_use: operand ");
4893 print_generic_expr (vect_dump, operand, TDF_SLIM);
4896 if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST)
4898 *dt = vect_constant_def;
4902 if (is_gimple_min_invariant (operand))
4905 *dt = vect_external_def;
4909 if (TREE_CODE (operand) == PAREN_EXPR)
4911 if (vect_print_dump_info (REPORT_DETAILS))
4912 fprintf (vect_dump, "non-associatable copy.");
4913 operand = TREE_OPERAND (operand, 0);
4916 if (TREE_CODE (operand) != SSA_NAME)
4918 if (vect_print_dump_info (REPORT_DETAILS))
4919 fprintf (vect_dump, "not ssa-name.");
4923 *def_stmt = SSA_NAME_DEF_STMT (operand);
4924 if (*def_stmt == NULL)
4926 if (vect_print_dump_info (REPORT_DETAILS))
4927 fprintf (vect_dump, "no def_stmt.");
4931 if (vect_print_dump_info (REPORT_DETAILS))
4933 fprintf (vect_dump, "def_stmt: ");
4934 print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM);
4937 /* Empty stmt is expected only in case of a function argument.
4938 (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */
4939 if (gimple_nop_p (*def_stmt))
4942 *dt = vect_external_def;
4946 bb = gimple_bb (*def_stmt);
4948 if ((loop && !flow_bb_inside_loop_p (loop, bb))
4949 || (!loop && bb != BB_VINFO_BB (bb_vinfo))
4950 || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI))
4951 *dt = vect_external_def;
4954 stmt_vinfo = vinfo_for_stmt (*def_stmt);
4955 *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo);
4958 if (*dt == vect_unknown_def_type)
4960 if (vect_print_dump_info (REPORT_DETAILS))
4961 fprintf (vect_dump, "Unsupported pattern.");
4965 if (vect_print_dump_info (REPORT_DETAILS))
4966 fprintf (vect_dump, "type of def: %d.",*dt);
4968 switch (gimple_code (*def_stmt))
4971 *def = gimple_phi_result (*def_stmt);
4975 *def = gimple_assign_lhs (*def_stmt);
4979 *def = gimple_call_lhs (*def_stmt);
4984 if (vect_print_dump_info (REPORT_DETAILS))
4985 fprintf (vect_dump, "unsupported defining stmt: ");
4992 /* Function vect_is_simple_use_1.
4994 Same as vect_is_simple_use_1 but also determines the vector operand
4995 type of OPERAND and stores it to *VECTYPE. If the definition of
4996 OPERAND is vect_uninitialized_def, vect_constant_def or
4997 vect_external_def *VECTYPE will be set to NULL_TREE and the caller
4998 is responsible to compute the best suited vector type for the
5002 vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo,
5003 bb_vec_info bb_vinfo, gimple *def_stmt,
5004 tree *def, enum vect_def_type *dt, tree *vectype)
5006 if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt))
5009 /* Now get a vector type if the def is internal, otherwise supply
5010 NULL_TREE and leave it up to the caller to figure out a proper
5011 type for the use stmt. */
5012 if (*dt == vect_internal_def
5013 || *dt == vect_induction_def
5014 || *dt == vect_reduction_def
5015 || *dt == vect_double_reduction_def
5016 || *dt == vect_nested_cycle)
5018 stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt);
5019 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
5020 stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info));
5021 *vectype = STMT_VINFO_VECTYPE (stmt_info);
5022 gcc_assert (*vectype != NULL_TREE);
5024 else if (*dt == vect_uninitialized_def
5025 || *dt == vect_constant_def
5026 || *dt == vect_external_def)
5027 *vectype = NULL_TREE;
5035 /* Function supportable_widening_operation
5037 Check whether an operation represented by the code CODE is a
5038 widening operation that is supported by the target platform in
5039 vector form (i.e., when operating on arguments of type VECTYPE_IN
5040 producing a result of type VECTYPE_OUT).
5042 Widening operations we currently support are NOP (CONVERT), FLOAT
5043 and WIDEN_MULT. This function checks if these operations are supported
5044 by the target platform either directly (via vector tree-codes), or via
5048 - CODE1 and CODE2 are codes of vector operations to be used when
5049 vectorizing the operation, if available.
5050 - DECL1 and DECL2 are decls of target builtin functions to be used
5051 when vectorizing the operation, if available. In this case,
5052 CODE1 and CODE2 are CALL_EXPR.
5053 - MULTI_STEP_CVT determines the number of required intermediate steps in
5054 case of multi-step conversion (like char->short->int - in that case
5055 MULTI_STEP_CVT will be 1).
5056 - INTERM_TYPES contains the intermediate type required to perform the
5057 widening operation (short in the above example). */
5060 supportable_widening_operation (enum tree_code code, gimple stmt,
5061 tree vectype_out, tree vectype_in,
5062 tree *decl1, tree *decl2,
5063 enum tree_code *code1, enum tree_code *code2,
5064 int *multi_step_cvt,
5065 VEC (tree, heap) **interm_types)
5067 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5068 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info);
5069 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
5071 enum machine_mode vec_mode;
5072 enum insn_code icode1, icode2;
5073 optab optab1, optab2;
5074 tree vectype = vectype_in;
5075 tree wide_vectype = vectype_out;
5076 enum tree_code c1, c2;
5078 /* The result of a vectorized widening operation usually requires two vectors
5079 (because the widened results do not fit int one vector). The generated
5080 vector results would normally be expected to be generated in the same
5081 order as in the original scalar computation, i.e. if 8 results are
5082 generated in each vector iteration, they are to be organized as follows:
5083 vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8].
5085 However, in the special case that the result of the widening operation is
5086 used in a reduction computation only, the order doesn't matter (because
5087 when vectorizing a reduction we change the order of the computation).
5088 Some targets can take advantage of this and generate more efficient code.
5089 For example, targets like Altivec, that support widen_mult using a sequence
5090 of {mult_even,mult_odd} generate the following vectors:
5091 vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8].
5093 When vectorizing outer-loops, we execute the inner-loop sequentially
5094 (each vectorized inner-loop iteration contributes to VF outer-loop
5095 iterations in parallel). We therefore don't allow to change the order
5096 of the computation in the inner-loop during outer-loop vectorization. */
5098 if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction
5099 && !nested_in_vect_loop_p (vect_loop, stmt))
5105 && code == WIDEN_MULT_EXPR
5106 && targetm.vectorize.builtin_mul_widen_even
5107 && targetm.vectorize.builtin_mul_widen_even (vectype)
5108 && targetm.vectorize.builtin_mul_widen_odd
5109 && targetm.vectorize.builtin_mul_widen_odd (vectype))
5111 if (vect_print_dump_info (REPORT_DETAILS))
5112 fprintf (vect_dump, "Unordered widening operation detected.");
5114 *code1 = *code2 = CALL_EXPR;
5115 *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype);
5116 *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype);
5122 case WIDEN_MULT_EXPR:
5123 if (BYTES_BIG_ENDIAN)
5125 c1 = VEC_WIDEN_MULT_HI_EXPR;
5126 c2 = VEC_WIDEN_MULT_LO_EXPR;
5130 c2 = VEC_WIDEN_MULT_HI_EXPR;
5131 c1 = VEC_WIDEN_MULT_LO_EXPR;
5136 if (BYTES_BIG_ENDIAN)
5138 c1 = VEC_UNPACK_HI_EXPR;
5139 c2 = VEC_UNPACK_LO_EXPR;
5143 c2 = VEC_UNPACK_HI_EXPR;
5144 c1 = VEC_UNPACK_LO_EXPR;
5149 if (BYTES_BIG_ENDIAN)
5151 c1 = VEC_UNPACK_FLOAT_HI_EXPR;
5152 c2 = VEC_UNPACK_FLOAT_LO_EXPR;
5156 c2 = VEC_UNPACK_FLOAT_HI_EXPR;
5157 c1 = VEC_UNPACK_FLOAT_LO_EXPR;
5161 case FIX_TRUNC_EXPR:
5162 /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/
5163 VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for
5164 computing the operation. */
5171 if (code == FIX_TRUNC_EXPR)
5173 /* The signedness is determined from output operand. */
5174 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
5175 optab2 = optab_for_tree_code (c2, vectype_out, optab_default);
5179 optab1 = optab_for_tree_code (c1, vectype, optab_default);
5180 optab2 = optab_for_tree_code (c2, vectype, optab_default);
5183 if (!optab1 || !optab2)
5186 vec_mode = TYPE_MODE (vectype);
5187 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing
5188 || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing)
5191 /* Check if it's a multi-step conversion that can be done using intermediate
5193 if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype)
5194 || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype))
5197 tree prev_type = vectype, intermediate_type;
5198 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5199 optab optab3, optab4;
5201 if (!CONVERT_EXPR_CODE_P (code))
5207 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5208 intermediate steps in promotion sequence. We try
5209 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
5211 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5212 for (i = 0; i < 3; i++)
5214 intermediate_mode = insn_data[icode1].operand[0].mode;
5215 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5216 TYPE_UNSIGNED (prev_type));
5217 optab3 = optab_for_tree_code (c1, intermediate_type, optab_default);
5218 optab4 = optab_for_tree_code (c2, intermediate_type, optab_default);
5220 if (!optab3 || !optab4
5221 || ((icode1 = optab_handler (optab1, prev_mode))
5222 == CODE_FOR_nothing)
5223 || insn_data[icode1].operand[0].mode != intermediate_mode
5224 || ((icode2 = optab_handler (optab2, prev_mode))
5225 == CODE_FOR_nothing)
5226 || insn_data[icode2].operand[0].mode != intermediate_mode
5227 || ((icode1 = optab_handler (optab3, intermediate_mode))
5228 == CODE_FOR_nothing)
5229 || ((icode2 = optab_handler (optab4, intermediate_mode))
5230 == CODE_FOR_nothing))
5233 VEC_quick_push (tree, *interm_types, intermediate_type);
5234 (*multi_step_cvt)++;
5236 if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype)
5237 && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype))
5240 prev_type = intermediate_type;
5241 prev_mode = intermediate_mode;
5253 /* Function supportable_narrowing_operation
5255 Check whether an operation represented by the code CODE is a
5256 narrowing operation that is supported by the target platform in
5257 vector form (i.e., when operating on arguments of type VECTYPE_IN
5258 and producing a result of type VECTYPE_OUT).
5260 Narrowing operations we currently support are NOP (CONVERT) and
5261 FIX_TRUNC. This function checks if these operations are supported by
5262 the target platform directly via vector tree-codes.
5265 - CODE1 is the code of a vector operation to be used when
5266 vectorizing the operation, if available.
5267 - MULTI_STEP_CVT determines the number of required intermediate steps in
5268 case of multi-step conversion (like int->short->char - in that case
5269 MULTI_STEP_CVT will be 1).
5270 - INTERM_TYPES contains the intermediate type required to perform the
5271 narrowing operation (short in the above example). */
5274 supportable_narrowing_operation (enum tree_code code,
5275 tree vectype_out, tree vectype_in,
5276 enum tree_code *code1, int *multi_step_cvt,
5277 VEC (tree, heap) **interm_types)
5279 enum machine_mode vec_mode;
5280 enum insn_code icode1;
5281 optab optab1, interm_optab;
5282 tree vectype = vectype_in;
5283 tree narrow_vectype = vectype_out;
5285 tree intermediate_type, prev_type;
5291 c1 = VEC_PACK_TRUNC_EXPR;
5294 case FIX_TRUNC_EXPR:
5295 c1 = VEC_PACK_FIX_TRUNC_EXPR;
5299 /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR
5300 tree code and optabs used for computing the operation. */
5307 if (code == FIX_TRUNC_EXPR)
5308 /* The signedness is determined from output operand. */
5309 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
5311 optab1 = optab_for_tree_code (c1, vectype, optab_default);
5316 vec_mode = TYPE_MODE (vectype);
5317 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing)
5320 /* Check if it's a multi-step conversion that can be done using intermediate
5322 if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype))
5324 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5327 prev_type = vectype;
5328 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5329 intermediate steps in promotion sequence. We try
5330 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
5332 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5333 for (i = 0; i < 3; i++)
5335 intermediate_mode = insn_data[icode1].operand[0].mode;
5336 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5337 TYPE_UNSIGNED (prev_type));
5338 interm_optab = optab_for_tree_code (c1, intermediate_type,
5341 || ((icode1 = optab_handler (optab1, prev_mode))
5342 == CODE_FOR_nothing)
5343 || insn_data[icode1].operand[0].mode != intermediate_mode
5344 || ((icode1 = optab_handler (interm_optab, intermediate_mode))
5345 == CODE_FOR_nothing))
5348 VEC_quick_push (tree, *interm_types, intermediate_type);
5349 (*multi_step_cvt)++;
5351 if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype))
5354 prev_type = intermediate_type;
5355 prev_mode = intermediate_mode;