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 cost_for_stmt (gimple stmt)
564 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
566 switch (STMT_VINFO_TYPE (stmt_info))
568 case load_vec_info_type:
569 return vect_get_stmt_cost (scalar_load);
570 case store_vec_info_type:
571 return vect_get_stmt_cost (scalar_store);
572 case op_vec_info_type:
573 case condition_vec_info_type:
574 case assignment_vec_info_type:
575 case reduc_vec_info_type:
576 case induc_vec_info_type:
577 case type_promotion_vec_info_type:
578 case type_demotion_vec_info_type:
579 case type_conversion_vec_info_type:
580 case call_vec_info_type:
581 return vect_get_stmt_cost (scalar_stmt);
582 case undef_vec_info_type:
588 /* Function vect_model_simple_cost.
590 Models cost for simple operations, i.e. those that only emit ncopies of a
591 single op. Right now, this does not account for multiple insns that could
592 be generated for the single vector op. We will handle that shortly. */
595 vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies,
596 enum vect_def_type *dt, slp_tree slp_node)
599 int inside_cost = 0, outside_cost = 0;
601 /* The SLP costs were already calculated during SLP tree build. */
602 if (PURE_SLP_STMT (stmt_info))
605 inside_cost = ncopies * vect_get_stmt_cost (vector_stmt);
607 /* FORNOW: Assuming maximum 2 args per stmts. */
608 for (i = 0; i < 2; i++)
610 if (dt[i] == vect_constant_def || dt[i] == vect_external_def)
611 outside_cost += vect_get_stmt_cost (vector_stmt);
614 if (vect_print_dump_info (REPORT_COST))
615 fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, "
616 "outside_cost = %d .", inside_cost, outside_cost);
618 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
619 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
620 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
624 /* Function vect_cost_strided_group_size
626 For strided load or store, return the group_size only if it is the first
627 load or store of a group, else return 1. This ensures that group size is
628 only returned once per group. */
631 vect_cost_strided_group_size (stmt_vec_info stmt_info)
633 gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info);
635 if (first_stmt == STMT_VINFO_STMT (stmt_info))
636 return DR_GROUP_SIZE (stmt_info);
642 /* Function vect_model_store_cost
644 Models cost for stores. In the case of strided accesses, one access
645 has the overhead of the strided access attributed to it. */
648 vect_model_store_cost (stmt_vec_info stmt_info, int ncopies,
649 enum vect_def_type dt, slp_tree slp_node)
652 unsigned int inside_cost = 0, outside_cost = 0;
653 struct data_reference *first_dr;
656 /* The SLP costs were already calculated during SLP tree build. */
657 if (PURE_SLP_STMT (stmt_info))
660 if (dt == vect_constant_def || dt == vect_external_def)
661 outside_cost = vect_get_stmt_cost (scalar_to_vec);
663 /* Strided access? */
664 if (DR_GROUP_FIRST_DR (stmt_info))
668 first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0);
673 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
674 group_size = vect_cost_strided_group_size (stmt_info);
677 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
679 /* Not a strided access. */
683 first_dr = STMT_VINFO_DATA_REF (stmt_info);
686 /* Is this an access in a group of stores, which provide strided access?
687 If so, add in the cost of the permutes. */
690 /* Uses a high and low interleave operation for each needed permute. */
691 inside_cost = ncopies * exact_log2(group_size) * group_size
692 * vect_get_stmt_cost (vector_stmt);
694 if (vect_print_dump_info (REPORT_COST))
695 fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .",
700 /* Costs of the stores. */
701 vect_get_store_cost (first_dr, ncopies, &inside_cost);
703 if (vect_print_dump_info (REPORT_COST))
704 fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, "
705 "outside_cost = %d .", inside_cost, outside_cost);
707 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
708 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
709 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
713 /* Calculate cost of DR's memory access. */
715 vect_get_store_cost (struct data_reference *dr, int ncopies,
716 unsigned int *inside_cost)
718 int alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
720 switch (alignment_support_scheme)
724 *inside_cost += ncopies * vect_get_stmt_cost (vector_store);
726 if (vect_print_dump_info (REPORT_COST))
727 fprintf (vect_dump, "vect_model_store_cost: aligned.");
732 case dr_unaligned_supported:
734 gimple stmt = DR_STMT (dr);
735 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
736 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
738 /* Here, we assign an additional cost for the unaligned store. */
739 *inside_cost += ncopies
740 * targetm.vectorize.builtin_vectorization_cost (unaligned_store,
741 vectype, DR_MISALIGNMENT (dr));
743 if (vect_print_dump_info (REPORT_COST))
744 fprintf (vect_dump, "vect_model_store_cost: unaligned supported by "
756 /* Function vect_model_load_cost
758 Models cost for loads. In the case of strided accesses, the last access
759 has the overhead of the strided access attributed to it. Since unaligned
760 accesses are supported for loads, we also account for the costs of the
761 access scheme chosen. */
764 vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node)
769 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
770 unsigned int inside_cost = 0, outside_cost = 0;
772 /* The SLP costs were already calculated during SLP tree build. */
773 if (PURE_SLP_STMT (stmt_info))
776 /* Strided accesses? */
777 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
778 if (first_stmt && !slp_node)
780 group_size = vect_cost_strided_group_size (stmt_info);
781 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
783 /* Not a strided access. */
790 /* Is this an access in a group of loads providing strided access?
791 If so, add in the cost of the permutes. */
794 /* Uses an even and odd extract operations for each needed permute. */
795 inside_cost = ncopies * exact_log2(group_size) * group_size
796 * vect_get_stmt_cost (vector_stmt);
798 if (vect_print_dump_info (REPORT_COST))
799 fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .",
803 /* The loads themselves. */
804 vect_get_load_cost (first_dr, ncopies,
805 ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node),
806 &inside_cost, &outside_cost);
808 if (vect_print_dump_info (REPORT_COST))
809 fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, "
810 "outside_cost = %d .", inside_cost, outside_cost);
812 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
813 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
814 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
818 /* Calculate cost of DR's memory access. */
820 vect_get_load_cost (struct data_reference *dr, int ncopies,
821 bool add_realign_cost, unsigned int *inside_cost,
822 unsigned int *outside_cost)
824 int alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
826 switch (alignment_support_scheme)
830 *inside_cost += ncopies * vect_get_stmt_cost (vector_load);
832 if (vect_print_dump_info (REPORT_COST))
833 fprintf (vect_dump, "vect_model_load_cost: aligned.");
837 case dr_unaligned_supported:
839 gimple stmt = DR_STMT (dr);
840 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
841 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
843 /* Here, we assign an additional cost for the unaligned load. */
844 *inside_cost += ncopies
845 * targetm.vectorize.builtin_vectorization_cost (unaligned_load,
846 vectype, DR_MISALIGNMENT (dr));
847 if (vect_print_dump_info (REPORT_COST))
848 fprintf (vect_dump, "vect_model_load_cost: unaligned supported by "
853 case dr_explicit_realign:
855 *inside_cost += ncopies * (2 * vect_get_stmt_cost (vector_load)
856 + vect_get_stmt_cost (vector_stmt));
858 /* FIXME: If the misalignment remains fixed across the iterations of
859 the containing loop, the following cost should be added to the
861 if (targetm.vectorize.builtin_mask_for_load)
862 *inside_cost += vect_get_stmt_cost (vector_stmt);
866 case dr_explicit_realign_optimized:
868 if (vect_print_dump_info (REPORT_COST))
869 fprintf (vect_dump, "vect_model_load_cost: unaligned software "
872 /* Unaligned software pipeline has a load of an address, an initial
873 load, and possibly a mask operation to "prime" the loop. However,
874 if this is an access in a group of loads, which provide strided
875 access, then the above cost should only be considered for one
876 access in the group. Inside the loop, there is a load op
877 and a realignment op. */
879 if (add_realign_cost)
881 *outside_cost = 2 * vect_get_stmt_cost (vector_stmt);
882 if (targetm.vectorize.builtin_mask_for_load)
883 *outside_cost += vect_get_stmt_cost (vector_stmt);
886 *inside_cost += ncopies * (vect_get_stmt_cost (vector_load)
887 + vect_get_stmt_cost (vector_stmt));
897 /* Function vect_init_vector.
899 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
900 the vector elements of VECTOR_VAR. Place the initialization at BSI if it
901 is not NULL. Otherwise, place the initialization at the loop preheader.
902 Return the DEF of INIT_STMT.
903 It will be used in the vectorization of STMT. */
906 vect_init_vector (gimple stmt, tree vector_var, tree vector_type,
907 gimple_stmt_iterator *gsi)
909 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
917 new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_");
918 add_referenced_var (new_var);
919 init_stmt = gimple_build_assign (new_var, vector_var);
920 new_temp = make_ssa_name (new_var, init_stmt);
921 gimple_assign_set_lhs (init_stmt, new_temp);
924 vect_finish_stmt_generation (stmt, init_stmt, gsi);
927 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
931 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
933 if (nested_in_vect_loop_p (loop, stmt))
936 pe = loop_preheader_edge (loop);
937 new_bb = gsi_insert_on_edge_immediate (pe, init_stmt);
938 gcc_assert (!new_bb);
942 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo);
944 gimple_stmt_iterator gsi_bb_start;
946 gcc_assert (bb_vinfo);
947 bb = BB_VINFO_BB (bb_vinfo);
948 gsi_bb_start = gsi_after_labels (bb);
949 gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT);
953 if (vect_print_dump_info (REPORT_DETAILS))
955 fprintf (vect_dump, "created new init_stmt: ");
956 print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM);
959 vec_oprnd = gimple_assign_lhs (init_stmt);
964 /* Function vect_get_vec_def_for_operand.
966 OP is an operand in STMT. This function returns a (vector) def that will be
967 used in the vectorized stmt for STMT.
969 In the case that OP is an SSA_NAME which is defined in the loop, then
970 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
972 In case OP is an invariant or constant, a new stmt that creates a vector def
973 needs to be introduced. */
976 vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def)
981 stmt_vec_info def_stmt_info = NULL;
982 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
983 tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
984 unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype);
985 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
991 enum vect_def_type dt;
995 if (vect_print_dump_info (REPORT_DETAILS))
997 fprintf (vect_dump, "vect_get_vec_def_for_operand: ");
998 print_generic_expr (vect_dump, op, TDF_SLIM);
1001 is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def,
1003 gcc_assert (is_simple_use);
1004 if (vect_print_dump_info (REPORT_DETAILS))
1008 fprintf (vect_dump, "def = ");
1009 print_generic_expr (vect_dump, def, TDF_SLIM);
1013 fprintf (vect_dump, " def_stmt = ");
1014 print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM);
1020 /* Case 1: operand is a constant. */
1021 case vect_constant_def:
1023 vector_type = get_vectype_for_scalar_type (TREE_TYPE (op));
1024 gcc_assert (vector_type);
1029 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1030 if (vect_print_dump_info (REPORT_DETAILS))
1031 fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits);
1033 for (i = nunits - 1; i >= 0; --i)
1035 t = tree_cons (NULL_TREE, op, t);
1037 vec_cst = build_vector (vector_type, t);
1038 return vect_init_vector (stmt, vec_cst, vector_type, NULL);
1041 /* Case 2: operand is defined outside the loop - loop invariant. */
1042 case vect_external_def:
1044 vector_type = get_vectype_for_scalar_type (TREE_TYPE (def));
1045 gcc_assert (vector_type);
1046 nunits = TYPE_VECTOR_SUBPARTS (vector_type);
1051 /* Create 'vec_inv = {inv,inv,..,inv}' */
1052 if (vect_print_dump_info (REPORT_DETAILS))
1053 fprintf (vect_dump, "Create vector_inv.");
1055 for (i = nunits - 1; i >= 0; --i)
1057 t = tree_cons (NULL_TREE, def, t);
1060 /* FIXME: use build_constructor directly. */
1061 vec_inv = build_constructor_from_list (vector_type, t);
1062 return vect_init_vector (stmt, vec_inv, vector_type, NULL);
1065 /* Case 3: operand is defined inside the loop. */
1066 case vect_internal_def:
1069 *scalar_def = NULL/* FIXME tuples: def_stmt*/;
1071 /* Get the def from the vectorized stmt. */
1072 def_stmt_info = vinfo_for_stmt (def_stmt);
1073 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1074 gcc_assert (vec_stmt);
1075 if (gimple_code (vec_stmt) == GIMPLE_PHI)
1076 vec_oprnd = PHI_RESULT (vec_stmt);
1077 else if (is_gimple_call (vec_stmt))
1078 vec_oprnd = gimple_call_lhs (vec_stmt);
1080 vec_oprnd = gimple_assign_lhs (vec_stmt);
1084 /* Case 4: operand is defined by a loop header phi - reduction */
1085 case vect_reduction_def:
1086 case vect_double_reduction_def:
1087 case vect_nested_cycle:
1091 gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
1092 loop = (gimple_bb (def_stmt))->loop_father;
1094 /* Get the def before the loop */
1095 op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop));
1096 return get_initial_def_for_reduction (stmt, op, scalar_def);
1099 /* Case 5: operand is defined by loop-header phi - induction. */
1100 case vect_induction_def:
1102 gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
1104 /* Get the def from the vectorized stmt. */
1105 def_stmt_info = vinfo_for_stmt (def_stmt);
1106 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1107 gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI);
1108 vec_oprnd = PHI_RESULT (vec_stmt);
1118 /* Function vect_get_vec_def_for_stmt_copy
1120 Return a vector-def for an operand. This function is used when the
1121 vectorized stmt to be created (by the caller to this function) is a "copy"
1122 created in case the vectorized result cannot fit in one vector, and several
1123 copies of the vector-stmt are required. In this case the vector-def is
1124 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1125 of the stmt that defines VEC_OPRND.
1126 DT is the type of the vector def VEC_OPRND.
1129 In case the vectorization factor (VF) is bigger than the number
1130 of elements that can fit in a vectype (nunits), we have to generate
1131 more than one vector stmt to vectorize the scalar stmt. This situation
1132 arises when there are multiple data-types operated upon in the loop; the
1133 smallest data-type determines the VF, and as a result, when vectorizing
1134 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1135 vector stmt (each computing a vector of 'nunits' results, and together
1136 computing 'VF' results in each iteration). This function is called when
1137 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1138 which VF=16 and nunits=4, so the number of copies required is 4):
1140 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1142 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1143 VS1.1: vx.1 = memref1 VS1.2
1144 VS1.2: vx.2 = memref2 VS1.3
1145 VS1.3: vx.3 = memref3
1147 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1148 VSnew.1: vz1 = vx.1 + ... VSnew.2
1149 VSnew.2: vz2 = vx.2 + ... VSnew.3
1150 VSnew.3: vz3 = vx.3 + ...
1152 The vectorization of S1 is explained in vectorizable_load.
1153 The vectorization of S2:
1154 To create the first vector-stmt out of the 4 copies - VSnew.0 -
1155 the function 'vect_get_vec_def_for_operand' is called to
1156 get the relevant vector-def for each operand of S2. For operand x it
1157 returns the vector-def 'vx.0'.
1159 To create the remaining copies of the vector-stmt (VSnew.j), this
1160 function is called to get the relevant vector-def for each operand. It is
1161 obtained from the respective VS1.j stmt, which is recorded in the
1162 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
1164 For example, to obtain the vector-def 'vx.1' in order to create the
1165 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
1166 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
1167 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
1168 and return its def ('vx.1').
1169 Overall, to create the above sequence this function will be called 3 times:
1170 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
1171 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
1172 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
1175 vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd)
1177 gimple vec_stmt_for_operand;
1178 stmt_vec_info def_stmt_info;
1180 /* Do nothing; can reuse same def. */
1181 if (dt == vect_external_def || dt == vect_constant_def )
1184 vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd);
1185 def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand);
1186 gcc_assert (def_stmt_info);
1187 vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info);
1188 gcc_assert (vec_stmt_for_operand);
1189 vec_oprnd = gimple_get_lhs (vec_stmt_for_operand);
1190 if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI)
1191 vec_oprnd = PHI_RESULT (vec_stmt_for_operand);
1193 vec_oprnd = gimple_get_lhs (vec_stmt_for_operand);
1198 /* Get vectorized definitions for the operands to create a copy of an original
1199 stmt. See vect_get_vec_def_for_stmt_copy() for details. */
1202 vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt,
1203 VEC(tree,heap) **vec_oprnds0,
1204 VEC(tree,heap) **vec_oprnds1)
1206 tree vec_oprnd = VEC_pop (tree, *vec_oprnds0);
1208 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd);
1209 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
1211 if (vec_oprnds1 && *vec_oprnds1)
1213 vec_oprnd = VEC_pop (tree, *vec_oprnds1);
1214 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd);
1215 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
1220 /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */
1223 vect_get_vec_defs (tree op0, tree op1, gimple stmt,
1224 VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1,
1228 vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1, -1);
1233 *vec_oprnds0 = VEC_alloc (tree, heap, 1);
1234 vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL);
1235 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
1239 *vec_oprnds1 = VEC_alloc (tree, heap, 1);
1240 vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL);
1241 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
1247 /* Function vect_finish_stmt_generation.
1249 Insert a new stmt. */
1252 vect_finish_stmt_generation (gimple stmt, gimple vec_stmt,
1253 gimple_stmt_iterator *gsi)
1255 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1256 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1257 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
1259 gcc_assert (gimple_code (stmt) != GIMPLE_LABEL);
1261 gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT);
1263 set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo,
1266 if (vect_print_dump_info (REPORT_DETAILS))
1268 fprintf (vect_dump, "add new stmt: ");
1269 print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM);
1272 gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi)));
1275 /* Checks if CALL can be vectorized in type VECTYPE. Returns
1276 a function declaration if the target has a vectorized version
1277 of the function, or NULL_TREE if the function cannot be vectorized. */
1280 vectorizable_function (gimple call, tree vectype_out, tree vectype_in)
1282 tree fndecl = gimple_call_fndecl (call);
1284 /* We only handle functions that do not read or clobber memory -- i.e.
1285 const or novops ones. */
1286 if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS)))
1290 || TREE_CODE (fndecl) != FUNCTION_DECL
1291 || !DECL_BUILT_IN (fndecl))
1294 return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out,
1298 /* Function vectorizable_call.
1300 Check if STMT performs a function call that can be vectorized.
1301 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1302 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1303 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1306 vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt)
1311 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
1312 stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info;
1313 tree vectype_out, vectype_in;
1316 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1317 tree fndecl, new_temp, def, rhs_type;
1319 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1320 gimple new_stmt = NULL;
1322 VEC(tree, heap) *vargs = NULL;
1323 enum { NARROW, NONE, WIDEN } modifier;
1326 /* FORNOW: unsupported in basic block SLP. */
1327 gcc_assert (loop_vinfo);
1329 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1332 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1335 /* FORNOW: SLP not supported. */
1336 if (STMT_SLP_TYPE (stmt_info))
1339 /* Is STMT a vectorizable call? */
1340 if (!is_gimple_call (stmt))
1343 if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
1346 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1348 /* Process function arguments. */
1349 rhs_type = NULL_TREE;
1350 vectype_in = NULL_TREE;
1351 nargs = gimple_call_num_args (stmt);
1353 /* Bail out if the function has more than two arguments, we
1354 do not have interesting builtin functions to vectorize with
1355 more than two arguments. No arguments is also not good. */
1356 if (nargs == 0 || nargs > 2)
1359 for (i = 0; i < nargs; i++)
1363 op = gimple_call_arg (stmt, i);
1365 /* We can only handle calls with arguments of the same type. */
1367 && !types_compatible_p (rhs_type, TREE_TYPE (op)))
1369 if (vect_print_dump_info (REPORT_DETAILS))
1370 fprintf (vect_dump, "argument types differ.");
1374 rhs_type = TREE_TYPE (op);
1376 if (!vect_is_simple_use_1 (op, loop_vinfo, NULL,
1377 &def_stmt, &def, &dt[i], &opvectype))
1379 if (vect_print_dump_info (REPORT_DETAILS))
1380 fprintf (vect_dump, "use not simple.");
1385 vectype_in = opvectype;
1387 && opvectype != vectype_in)
1389 if (vect_print_dump_info (REPORT_DETAILS))
1390 fprintf (vect_dump, "argument vector types differ.");
1394 /* If all arguments are external or constant defs use a vector type with
1395 the same size as the output vector type. */
1397 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1399 gcc_assert (vectype_in);
1402 if (vect_print_dump_info (REPORT_DETAILS))
1404 fprintf (vect_dump, "no vectype for scalar type ");
1405 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1412 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1413 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1414 if (nunits_in == nunits_out / 2)
1416 else if (nunits_out == nunits_in)
1418 else if (nunits_out == nunits_in / 2)
1423 /* For now, we only vectorize functions if a target specific builtin
1424 is available. TODO -- in some cases, it might be profitable to
1425 insert the calls for pieces of the vector, in order to be able
1426 to vectorize other operations in the loop. */
1427 fndecl = vectorizable_function (stmt, vectype_out, vectype_in);
1428 if (fndecl == NULL_TREE)
1430 if (vect_print_dump_info (REPORT_DETAILS))
1431 fprintf (vect_dump, "function is not vectorizable.");
1436 gcc_assert (!gimple_vuse (stmt));
1438 if (modifier == NARROW)
1439 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1441 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1443 /* Sanity check: make sure that at least one copy of the vectorized stmt
1444 needs to be generated. */
1445 gcc_assert (ncopies >= 1);
1447 if (!vec_stmt) /* transformation not required. */
1449 STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
1450 if (vect_print_dump_info (REPORT_DETAILS))
1451 fprintf (vect_dump, "=== vectorizable_call ===");
1452 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
1458 if (vect_print_dump_info (REPORT_DETAILS))
1459 fprintf (vect_dump, "transform operation.");
1462 scalar_dest = gimple_call_lhs (stmt);
1463 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1465 prev_stmt_info = NULL;
1469 for (j = 0; j < ncopies; ++j)
1471 /* Build argument list for the vectorized call. */
1473 vargs = VEC_alloc (tree, heap, nargs);
1475 VEC_truncate (tree, vargs, 0);
1477 for (i = 0; i < nargs; i++)
1479 op = gimple_call_arg (stmt, i);
1482 = vect_get_vec_def_for_operand (op, stmt, NULL);
1485 vec_oprnd0 = gimple_call_arg (new_stmt, i);
1487 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1490 VEC_quick_push (tree, vargs, vec_oprnd0);
1493 new_stmt = gimple_build_call_vec (fndecl, vargs);
1494 new_temp = make_ssa_name (vec_dest, new_stmt);
1495 gimple_call_set_lhs (new_stmt, new_temp);
1497 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1498 mark_symbols_for_renaming (new_stmt);
1501 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1503 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1505 prev_stmt_info = vinfo_for_stmt (new_stmt);
1511 for (j = 0; j < ncopies; ++j)
1513 /* Build argument list for the vectorized call. */
1515 vargs = VEC_alloc (tree, heap, nargs * 2);
1517 VEC_truncate (tree, vargs, 0);
1519 for (i = 0; i < nargs; i++)
1521 op = gimple_call_arg (stmt, i);
1525 = vect_get_vec_def_for_operand (op, stmt, NULL);
1527 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1531 vec_oprnd1 = gimple_call_arg (new_stmt, 2*i);
1533 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1);
1535 = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0);
1538 VEC_quick_push (tree, vargs, vec_oprnd0);
1539 VEC_quick_push (tree, vargs, vec_oprnd1);
1542 new_stmt = gimple_build_call_vec (fndecl, vargs);
1543 new_temp = make_ssa_name (vec_dest, new_stmt);
1544 gimple_call_set_lhs (new_stmt, new_temp);
1546 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1547 mark_symbols_for_renaming (new_stmt);
1550 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1552 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1554 prev_stmt_info = vinfo_for_stmt (new_stmt);
1557 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1562 /* No current target implements this case. */
1566 VEC_free (tree, heap, vargs);
1568 /* Update the exception handling table with the vector stmt if necessary. */
1569 if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt))
1570 gimple_purge_dead_eh_edges (gimple_bb (stmt));
1572 /* The call in STMT might prevent it from being removed in dce.
1573 We however cannot remove it here, due to the way the ssa name
1574 it defines is mapped to the new definition. So just replace
1575 rhs of the statement with something harmless. */
1577 type = TREE_TYPE (scalar_dest);
1578 new_stmt = gimple_build_assign (gimple_call_lhs (stmt),
1579 fold_convert (type, integer_zero_node));
1580 set_vinfo_for_stmt (new_stmt, stmt_info);
1581 set_vinfo_for_stmt (stmt, NULL);
1582 STMT_VINFO_STMT (stmt_info) = new_stmt;
1583 gsi_replace (gsi, new_stmt, false);
1584 SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt;
1590 /* Function vect_gen_widened_results_half
1592 Create a vector stmt whose code, type, number of arguments, and result
1593 variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are
1594 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
1595 In the case that CODE is a CALL_EXPR, this means that a call to DECL
1596 needs to be created (DECL is a function-decl of a target-builtin).
1597 STMT is the original scalar stmt that we are vectorizing. */
1600 vect_gen_widened_results_half (enum tree_code code,
1602 tree vec_oprnd0, tree vec_oprnd1, int op_type,
1603 tree vec_dest, gimple_stmt_iterator *gsi,
1609 /* Generate half of the widened result: */
1610 if (code == CALL_EXPR)
1612 /* Target specific support */
1613 if (op_type == binary_op)
1614 new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1);
1616 new_stmt = gimple_build_call (decl, 1, vec_oprnd0);
1617 new_temp = make_ssa_name (vec_dest, new_stmt);
1618 gimple_call_set_lhs (new_stmt, new_temp);
1622 /* Generic support */
1623 gcc_assert (op_type == TREE_CODE_LENGTH (code));
1624 if (op_type != binary_op)
1626 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0,
1628 new_temp = make_ssa_name (vec_dest, new_stmt);
1629 gimple_assign_set_lhs (new_stmt, new_temp);
1631 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1637 /* Check if STMT performs a conversion operation, that can be vectorized.
1638 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1639 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1640 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1643 vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi,
1644 gimple *vec_stmt, slp_tree slp_node)
1649 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
1650 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1651 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1652 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
1653 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
1657 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1658 gimple new_stmt = NULL;
1659 stmt_vec_info prev_stmt_info;
1662 tree vectype_out, vectype_in;
1666 enum { NARROW, NONE, WIDEN } modifier;
1668 VEC(tree,heap) *vec_oprnds0 = NULL;
1670 VEC(tree,heap) *dummy = NULL;
1673 /* Is STMT a vectorizable conversion? */
1675 /* FORNOW: unsupported in basic block SLP. */
1676 gcc_assert (loop_vinfo);
1678 if (!STMT_VINFO_RELEVANT_P (stmt_info))
1681 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1684 if (!is_gimple_assign (stmt))
1687 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
1690 code = gimple_assign_rhs_code (stmt);
1691 if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
1694 /* Check types of lhs and rhs. */
1695 scalar_dest = gimple_assign_lhs (stmt);
1696 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
1698 op0 = gimple_assign_rhs1 (stmt);
1699 rhs_type = TREE_TYPE (op0);
1700 /* Check the operands of the operation. */
1701 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
1702 &def_stmt, &def, &dt[0], &vectype_in))
1704 if (vect_print_dump_info (REPORT_DETAILS))
1705 fprintf (vect_dump, "use not simple.");
1708 /* If op0 is an external or constant defs use a vector type of
1709 the same size as the output vector type. */
1711 vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
1713 gcc_assert (vectype_in);
1716 if (vect_print_dump_info (REPORT_DETAILS))
1718 fprintf (vect_dump, "no vectype for scalar type ");
1719 print_generic_expr (vect_dump, rhs_type, TDF_SLIM);
1726 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
1727 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
1728 if (nunits_in == nunits_out / 2)
1730 else if (nunits_out == nunits_in)
1732 else if (nunits_out == nunits_in / 2)
1737 if (modifier == NARROW)
1738 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
1740 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
1742 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
1743 this, so we can safely override NCOPIES with 1 here. */
1747 /* Sanity check: make sure that at least one copy of the vectorized stmt
1748 needs to be generated. */
1749 gcc_assert (ncopies >= 1);
1751 /* Supportable by target? */
1752 if ((modifier == NONE
1753 && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in))
1754 || (modifier == WIDEN
1755 && !supportable_widening_operation (code, stmt,
1756 vectype_out, vectype_in,
1759 &dummy_int, &dummy))
1760 || (modifier == NARROW
1761 && !supportable_narrowing_operation (code, vectype_out, vectype_in,
1762 &code1, &dummy_int, &dummy)))
1764 if (vect_print_dump_info (REPORT_DETAILS))
1765 fprintf (vect_dump, "conversion not supported by target.");
1769 if (modifier != NONE)
1771 /* FORNOW: SLP not supported. */
1772 if (STMT_SLP_TYPE (stmt_info))
1776 if (!vec_stmt) /* transformation not required. */
1778 STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type;
1783 if (vect_print_dump_info (REPORT_DETAILS))
1784 fprintf (vect_dump, "transform conversion.");
1787 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
1789 if (modifier == NONE && !slp_node)
1790 vec_oprnds0 = VEC_alloc (tree, heap, 1);
1792 prev_stmt_info = NULL;
1796 for (j = 0; j < ncopies; j++)
1799 vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node);
1801 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL);
1804 targetm.vectorize.builtin_conversion (code,
1805 vectype_out, vectype_in);
1806 for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++)
1808 /* Arguments are ready. create the new vector stmt. */
1809 new_stmt = gimple_build_call (builtin_decl, 1, vop0);
1810 new_temp = make_ssa_name (vec_dest, new_stmt);
1811 gimple_call_set_lhs (new_stmt, new_temp);
1812 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1814 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
1818 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
1820 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1821 prev_stmt_info = vinfo_for_stmt (new_stmt);
1826 /* In case the vectorization factor (VF) is bigger than the number
1827 of elements that we can fit in a vectype (nunits), we have to
1828 generate more than one vector stmt - i.e - we need to "unroll"
1829 the vector stmt by a factor VF/nunits. */
1830 for (j = 0; j < ncopies; j++)
1833 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1835 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1837 /* Generate first half of the widened result: */
1839 = vect_gen_widened_results_half (code1, decl1,
1840 vec_oprnd0, vec_oprnd1,
1841 unary_op, vec_dest, gsi, stmt);
1843 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1845 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1846 prev_stmt_info = vinfo_for_stmt (new_stmt);
1848 /* Generate second half of the widened result: */
1850 = vect_gen_widened_results_half (code2, decl2,
1851 vec_oprnd0, vec_oprnd1,
1852 unary_op, vec_dest, gsi, stmt);
1853 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1854 prev_stmt_info = vinfo_for_stmt (new_stmt);
1859 /* In case the vectorization factor (VF) is bigger than the number
1860 of elements that we can fit in a vectype (nunits), we have to
1861 generate more than one vector stmt - i.e - we need to "unroll"
1862 the vector stmt by a factor VF/nunits. */
1863 for (j = 0; j < ncopies; j++)
1868 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
1869 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1873 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
1874 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
1877 /* Arguments are ready. Create the new vector stmt. */
1878 new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0,
1880 new_temp = make_ssa_name (vec_dest, new_stmt);
1881 gimple_assign_set_lhs (new_stmt, new_temp);
1882 vect_finish_stmt_generation (stmt, new_stmt, gsi);
1885 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
1887 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
1889 prev_stmt_info = vinfo_for_stmt (new_stmt);
1892 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
1896 VEC_free (tree, heap, vec_oprnds0);
1900 /* Function vectorizable_assignment.
1902 Check if STMT performs an assignment (copy) that can be vectorized.
1903 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1904 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1905 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1908 vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi,
1909 gimple *vec_stmt, slp_tree slp_node)
1914 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1915 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
1916 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1920 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
1921 unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype);
1924 VEC(tree,heap) *vec_oprnds = NULL;
1926 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
1927 gimple new_stmt = NULL;
1928 stmt_vec_info prev_stmt_info = NULL;
1929 enum tree_code code;
1932 /* Multiple types in SLP are handled by creating the appropriate number of
1933 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
1938 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
1940 gcc_assert (ncopies >= 1);
1942 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
1945 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
1948 /* Is vectorizable assignment? */
1949 if (!is_gimple_assign (stmt))
1952 scalar_dest = gimple_assign_lhs (stmt);
1953 if (TREE_CODE (scalar_dest) != SSA_NAME)
1956 code = gimple_assign_rhs_code (stmt);
1957 if (gimple_assign_single_p (stmt)
1958 || code == PAREN_EXPR
1959 || CONVERT_EXPR_CODE_P (code))
1960 op = gimple_assign_rhs1 (stmt);
1964 if (!vect_is_simple_use_1 (op, loop_vinfo, bb_vinfo,
1965 &def_stmt, &def, &dt[0], &vectype_in))
1967 if (vect_print_dump_info (REPORT_DETAILS))
1968 fprintf (vect_dump, "use not simple.");
1972 /* We can handle NOP_EXPR conversions that do not change the number
1973 of elements or the vector size. */
1974 if (CONVERT_EXPR_CODE_P (code)
1976 || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits
1977 || (GET_MODE_SIZE (TYPE_MODE (vectype))
1978 != GET_MODE_SIZE (TYPE_MODE (vectype_in)))))
1981 if (!vec_stmt) /* transformation not required. */
1983 STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
1984 if (vect_print_dump_info (REPORT_DETAILS))
1985 fprintf (vect_dump, "=== vectorizable_assignment ===");
1986 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
1991 if (vect_print_dump_info (REPORT_DETAILS))
1992 fprintf (vect_dump, "transform assignment.");
1995 vec_dest = vect_create_destination_var (scalar_dest, vectype);
1998 for (j = 0; j < ncopies; j++)
2002 vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node);
2004 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL);
2006 /* Arguments are ready. create the new vector stmt. */
2007 for (i = 0; VEC_iterate (tree, vec_oprnds, i, vop); i++)
2009 if (CONVERT_EXPR_CODE_P (code))
2010 vop = build1 (VIEW_CONVERT_EXPR, vectype, vop);
2011 new_stmt = gimple_build_assign (vec_dest, vop);
2012 new_temp = make_ssa_name (vec_dest, new_stmt);
2013 gimple_assign_set_lhs (new_stmt, new_temp);
2014 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2016 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2023 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2025 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2027 prev_stmt_info = vinfo_for_stmt (new_stmt);
2030 VEC_free (tree, heap, vec_oprnds);
2034 /* Function vectorizable_operation.
2036 Check if STMT performs a binary or unary operation that can be vectorized.
2037 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2038 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2039 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2042 vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi,
2043 gimple *vec_stmt, slp_tree slp_node)
2047 tree op0, op1 = NULL;
2048 tree vec_oprnd1 = NULL_TREE;
2049 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2051 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2052 enum tree_code code;
2053 enum machine_mode vec_mode;
2058 enum machine_mode optab_op2_mode;
2061 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2062 gimple new_stmt = NULL;
2063 stmt_vec_info prev_stmt_info;
2069 VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2072 bool scalar_shift_arg = false;
2073 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
2076 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
2079 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2082 /* Is STMT a vectorizable binary/unary operation? */
2083 if (!is_gimple_assign (stmt))
2086 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2089 code = gimple_assign_rhs_code (stmt);
2091 /* For pointer addition, we should use the normal plus for
2092 the vector addition. */
2093 if (code == POINTER_PLUS_EXPR)
2096 /* Support only unary or binary operations. */
2097 op_type = TREE_CODE_LENGTH (code);
2098 if (op_type != unary_op && op_type != binary_op)
2100 if (vect_print_dump_info (REPORT_DETAILS))
2101 fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type);
2105 scalar_dest = gimple_assign_lhs (stmt);
2106 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2108 op0 = gimple_assign_rhs1 (stmt);
2109 if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo,
2110 &def_stmt, &def, &dt[0], &vectype))
2112 if (vect_print_dump_info (REPORT_DETAILS))
2113 fprintf (vect_dump, "use not simple.");
2116 /* If op0 is an external or constant def use a vector type with
2117 the same size as the output vector type. */
2119 vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2121 gcc_assert (vectype);
2124 if (vect_print_dump_info (REPORT_DETAILS))
2126 fprintf (vect_dump, "no vectype for scalar type ");
2127 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2133 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2134 nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
2135 if (nunits_out != nunits_in)
2138 if (op_type == binary_op)
2140 op1 = gimple_assign_rhs2 (stmt);
2141 if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def,
2144 if (vect_print_dump_info (REPORT_DETAILS))
2145 fprintf (vect_dump, "use not simple.");
2151 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2155 /* Multiple types in SLP are handled by creating the appropriate number of
2156 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2161 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2163 gcc_assert (ncopies >= 1);
2165 /* If this is a shift/rotate, determine whether the shift amount is a vector,
2166 or scalar. If the shift/rotate amount is a vector, use the vector/vector
2168 if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR
2169 || code == RROTATE_EXPR)
2171 /* vector shifted by vector */
2172 if (dt[1] == vect_internal_def)
2174 optab = optab_for_tree_code (code, vectype, optab_vector);
2175 if (vect_print_dump_info (REPORT_DETAILS))
2176 fprintf (vect_dump, "vector/vector shift/rotate found.");
2179 /* See if the machine has a vector shifted by scalar insn and if not
2180 then see if it has a vector shifted by vector insn */
2181 else if (dt[1] == vect_constant_def || dt[1] == vect_external_def)
2183 optab = optab_for_tree_code (code, vectype, optab_scalar);
2185 && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing)
2187 scalar_shift_arg = true;
2188 if (vect_print_dump_info (REPORT_DETAILS))
2189 fprintf (vect_dump, "vector/scalar shift/rotate found.");
2193 optab = optab_for_tree_code (code, vectype, optab_vector);
2195 && (optab_handler (optab, TYPE_MODE (vectype))
2196 != CODE_FOR_nothing))
2198 if (vect_print_dump_info (REPORT_DETAILS))
2199 fprintf (vect_dump, "vector/vector shift/rotate found.");
2201 /* Unlike the other binary operators, shifts/rotates have
2202 the rhs being int, instead of the same type as the lhs,
2203 so make sure the scalar is the right type if we are
2204 dealing with vectors of short/char. */
2205 if (dt[1] == vect_constant_def)
2206 op1 = fold_convert (TREE_TYPE (vectype), op1);
2213 if (vect_print_dump_info (REPORT_DETAILS))
2214 fprintf (vect_dump, "operand mode requires invariant argument.");
2219 optab = optab_for_tree_code (code, vectype, optab_default);
2221 /* Supportable by target? */
2224 if (vect_print_dump_info (REPORT_DETAILS))
2225 fprintf (vect_dump, "no optab.");
2228 vec_mode = TYPE_MODE (vectype);
2229 icode = (int) optab_handler (optab, vec_mode);
2230 if (icode == CODE_FOR_nothing)
2232 if (vect_print_dump_info (REPORT_DETAILS))
2233 fprintf (vect_dump, "op not supported by target.");
2234 /* Check only during analysis. */
2235 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2236 || (vf < vect_min_worthwhile_factor (code)
2239 if (vect_print_dump_info (REPORT_DETAILS))
2240 fprintf (vect_dump, "proceeding using word mode.");
2243 /* Worthwhile without SIMD support? Check only during analysis. */
2244 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2245 && vf < vect_min_worthwhile_factor (code)
2248 if (vect_print_dump_info (REPORT_DETAILS))
2249 fprintf (vect_dump, "not worthwhile without SIMD support.");
2253 if (!vec_stmt) /* transformation not required. */
2255 STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
2256 if (vect_print_dump_info (REPORT_DETAILS))
2257 fprintf (vect_dump, "=== vectorizable_operation ===");
2258 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2264 if (vect_print_dump_info (REPORT_DETAILS))
2265 fprintf (vect_dump, "transform binary/unary operation.");
2268 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2270 /* Allocate VECs for vector operands. In case of SLP, vector operands are
2271 created in the previous stages of the recursion, so no allocation is
2272 needed, except for the case of shift with scalar shift argument. In that
2273 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
2274 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
2275 In case of loop-based vectorization we allocate VECs of size 1. We
2276 allocate VEC_OPRNDS1 only in case of binary operation. */
2279 vec_oprnds0 = VEC_alloc (tree, heap, 1);
2280 if (op_type == binary_op)
2281 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2283 else if (scalar_shift_arg)
2284 vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size);
2286 /* In case the vectorization factor (VF) is bigger than the number
2287 of elements that we can fit in a vectype (nunits), we have to generate
2288 more than one vector stmt - i.e - we need to "unroll" the
2289 vector stmt by a factor VF/nunits. In doing so, we record a pointer
2290 from one copy of the vector stmt to the next, in the field
2291 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
2292 stages to find the correct vector defs to be used when vectorizing
2293 stmts that use the defs of the current stmt. The example below illustrates
2294 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
2295 4 vectorized stmts):
2297 before vectorization:
2298 RELATED_STMT VEC_STMT
2302 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
2304 RELATED_STMT VEC_STMT
2305 VS1_0: vx0 = memref0 VS1_1 -
2306 VS1_1: vx1 = memref1 VS1_2 -
2307 VS1_2: vx2 = memref2 VS1_3 -
2308 VS1_3: vx3 = memref3 - -
2309 S1: x = load - VS1_0
2312 step2: vectorize stmt S2 (done here):
2313 To vectorize stmt S2 we first need to find the relevant vector
2314 def for the first operand 'x'. This is, as usual, obtained from
2315 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
2316 that defines 'x' (S1). This way we find the stmt VS1_0, and the
2317 relevant vector def 'vx0'. Having found 'vx0' we can generate
2318 the vector stmt VS2_0, and as usual, record it in the
2319 STMT_VINFO_VEC_STMT of stmt S2.
2320 When creating the second copy (VS2_1), we obtain the relevant vector
2321 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
2322 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
2323 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
2324 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
2325 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
2326 chain of stmts and pointers:
2327 RELATED_STMT VEC_STMT
2328 VS1_0: vx0 = memref0 VS1_1 -
2329 VS1_1: vx1 = memref1 VS1_2 -
2330 VS1_2: vx2 = memref2 VS1_3 -
2331 VS1_3: vx3 = memref3 - -
2332 S1: x = load - VS1_0
2333 VS2_0: vz0 = vx0 + v1 VS2_1 -
2334 VS2_1: vz1 = vx1 + v1 VS2_2 -
2335 VS2_2: vz2 = vx2 + v1 VS2_3 -
2336 VS2_3: vz3 = vx3 + v1 - -
2337 S2: z = x + 1 - VS2_0 */
2339 prev_stmt_info = NULL;
2340 for (j = 0; j < ncopies; j++)
2345 if (op_type == binary_op && scalar_shift_arg)
2347 /* Vector shl and shr insn patterns can be defined with scalar
2348 operand 2 (shift operand). In this case, use constant or loop
2349 invariant op1 directly, without extending it to vector mode
2351 optab_op2_mode = insn_data[icode].operand[2].mode;
2352 if (!VECTOR_MODE_P (optab_op2_mode))
2354 if (vect_print_dump_info (REPORT_DETAILS))
2355 fprintf (vect_dump, "operand 1 using scalar mode.");
2357 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2360 /* Store vec_oprnd1 for every vector stmt to be created
2361 for SLP_NODE. We check during the analysis that all the
2362 shift arguments are the same.
2363 TODO: Allow different constants for different vector
2364 stmts generated for an SLP instance. */
2365 for (k = 0; k < slp_node->vec_stmts_size - 1; k++)
2366 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2371 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
2372 (a special case for certain kind of vector shifts); otherwise,
2373 operand 1 should be of a vector type (the usual case). */
2374 if (op_type == binary_op && !vec_oprnd1)
2375 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
2378 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
2382 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
2384 /* Arguments are ready. Create the new vector stmt. */
2385 for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++)
2387 vop1 = ((op_type == binary_op)
2388 ? VEC_index (tree, vec_oprnds1, i) : NULL);
2389 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2390 new_temp = make_ssa_name (vec_dest, new_stmt);
2391 gimple_assign_set_lhs (new_stmt, new_temp);
2392 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2394 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2401 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2403 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2404 prev_stmt_info = vinfo_for_stmt (new_stmt);
2407 VEC_free (tree, heap, vec_oprnds0);
2409 VEC_free (tree, heap, vec_oprnds1);
2415 /* Get vectorized definitions for loop-based vectorization. For the first
2416 operand we call vect_get_vec_def_for_operand() (with OPRND containing
2417 scalar operand), and for the rest we get a copy with
2418 vect_get_vec_def_for_stmt_copy() using the previous vector definition
2419 (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details.
2420 The vectors are collected into VEC_OPRNDS. */
2423 vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt,
2424 VEC (tree, heap) **vec_oprnds, int multi_step_cvt)
2428 /* Get first vector operand. */
2429 /* All the vector operands except the very first one (that is scalar oprnd)
2431 if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE)
2432 vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL);
2434 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd);
2436 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2438 /* Get second vector operand. */
2439 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd);
2440 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
2444 /* For conversion in multiple steps, continue to get operands
2447 vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1);
2451 /* Create vectorized demotion statements for vector operands from VEC_OPRNDS.
2452 For multi-step conversions store the resulting vectors and call the function
2456 vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds,
2457 int multi_step_cvt, gimple stmt,
2458 VEC (tree, heap) *vec_dsts,
2459 gimple_stmt_iterator *gsi,
2460 slp_tree slp_node, enum tree_code code,
2461 stmt_vec_info *prev_stmt_info)
2464 tree vop0, vop1, new_tmp, vec_dest;
2466 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2468 vec_dest = VEC_pop (tree, vec_dsts);
2470 for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2)
2472 /* Create demotion operation. */
2473 vop0 = VEC_index (tree, *vec_oprnds, i);
2474 vop1 = VEC_index (tree, *vec_oprnds, i + 1);
2475 new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1);
2476 new_tmp = make_ssa_name (vec_dest, new_stmt);
2477 gimple_assign_set_lhs (new_stmt, new_tmp);
2478 vect_finish_stmt_generation (stmt, new_stmt, gsi);
2481 /* Store the resulting vector for next recursive call. */
2482 VEC_replace (tree, *vec_oprnds, i/2, new_tmp);
2485 /* This is the last step of the conversion sequence. Store the
2486 vectors in SLP_NODE or in vector info of the scalar statement
2487 (or in STMT_VINFO_RELATED_STMT chain). */
2489 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
2492 if (!*prev_stmt_info)
2493 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
2495 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt;
2497 *prev_stmt_info = vinfo_for_stmt (new_stmt);
2502 /* For multi-step demotion operations we first generate demotion operations
2503 from the source type to the intermediate types, and then combine the
2504 results (stored in VEC_OPRNDS) in demotion operation to the destination
2508 /* At each level of recursion we have have of the operands we had at the
2510 VEC_truncate (tree, *vec_oprnds, (i+1)/2);
2511 vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1,
2512 stmt, vec_dsts, gsi, slp_node,
2513 code, prev_stmt_info);
2518 /* Function vectorizable_type_demotion
2520 Check if STMT performs a binary or unary operation that involves
2521 type demotion, and if it can be vectorized.
2522 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2523 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2524 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2527 vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi,
2528 gimple *vec_stmt, slp_tree slp_node)
2533 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2534 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2535 enum tree_code code, code1 = ERROR_MARK;
2538 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2539 stmt_vec_info prev_stmt_info;
2546 int multi_step_cvt = 0;
2547 VEC (tree, heap) *vec_oprnds0 = NULL;
2548 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
2549 tree last_oprnd, intermediate_type;
2551 /* FORNOW: not supported by basic block SLP vectorization. */
2552 gcc_assert (loop_vinfo);
2554 if (!STMT_VINFO_RELEVANT_P (stmt_info))
2557 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2560 /* Is STMT a vectorizable type-demotion operation? */
2561 if (!is_gimple_assign (stmt))
2564 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2567 code = gimple_assign_rhs_code (stmt);
2568 if (!CONVERT_EXPR_CODE_P (code))
2571 scalar_dest = gimple_assign_lhs (stmt);
2572 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2574 /* Check the operands of the operation. */
2575 op0 = gimple_assign_rhs1 (stmt);
2576 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
2577 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
2578 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
2579 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
2580 && CONVERT_EXPR_CODE_P (code))))
2582 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
2583 &def_stmt, &def, &dt[0], &vectype_in))
2585 if (vect_print_dump_info (REPORT_DETAILS))
2586 fprintf (vect_dump, "use not simple.");
2589 /* If op0 is an external def use a vector type with the
2590 same size as the output vector type if possible. */
2592 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2594 gcc_assert (vectype_in);
2597 if (vect_print_dump_info (REPORT_DETAILS))
2599 fprintf (vect_dump, "no vectype for scalar type ");
2600 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2606 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
2607 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2608 if (nunits_in >= nunits_out)
2611 /* Multiple types in SLP are handled by creating the appropriate number of
2612 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2617 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
2618 gcc_assert (ncopies >= 1);
2620 /* Supportable by target? */
2621 if (!supportable_narrowing_operation (code, vectype_out, vectype_in,
2622 &code1, &multi_step_cvt, &interm_types))
2625 if (!vec_stmt) /* transformation not required. */
2627 STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type;
2628 if (vect_print_dump_info (REPORT_DETAILS))
2629 fprintf (vect_dump, "=== vectorizable_demotion ===");
2630 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
2635 if (vect_print_dump_info (REPORT_DETAILS))
2636 fprintf (vect_dump, "transform type demotion operation. ncopies = %d.",
2639 /* In case of multi-step demotion, we first generate demotion operations to
2640 the intermediate types, and then from that types to the final one.
2641 We create vector destinations for the intermediate type (TYPES) received
2642 from supportable_narrowing_operation, and store them in the correct order
2643 for future use in vect_create_vectorized_demotion_stmts(). */
2645 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
2647 vec_dsts = VEC_alloc (tree, heap, 1);
2649 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
2650 VEC_quick_push (tree, vec_dsts, vec_dest);
2654 for (i = VEC_length (tree, interm_types) - 1;
2655 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
2657 vec_dest = vect_create_destination_var (scalar_dest,
2659 VEC_quick_push (tree, vec_dsts, vec_dest);
2663 /* In case the vectorization factor (VF) is bigger than the number
2664 of elements that we can fit in a vectype (nunits), we have to generate
2665 more than one vector stmt - i.e - we need to "unroll" the
2666 vector stmt by a factor VF/nunits. */
2668 prev_stmt_info = NULL;
2669 for (j = 0; j < ncopies; j++)
2673 vect_get_slp_defs (slp_node, &vec_oprnds0, NULL, -1);
2676 VEC_free (tree, heap, vec_oprnds0);
2677 vec_oprnds0 = VEC_alloc (tree, heap,
2678 (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2));
2679 vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0,
2680 vect_pow2 (multi_step_cvt) - 1);
2683 /* Arguments are ready. Create the new vector stmts. */
2684 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
2685 vect_create_vectorized_demotion_stmts (&vec_oprnds0,
2686 multi_step_cvt, stmt, tmp_vec_dsts,
2687 gsi, slp_node, code1,
2691 VEC_free (tree, heap, vec_oprnds0);
2692 VEC_free (tree, heap, vec_dsts);
2693 VEC_free (tree, heap, tmp_vec_dsts);
2694 VEC_free (tree, heap, interm_types);
2696 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
2701 /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0
2702 and VEC_OPRNDS1 (for binary operations). For multi-step conversions store
2703 the resulting vectors and call the function recursively. */
2706 vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0,
2707 VEC (tree, heap) **vec_oprnds1,
2708 int multi_step_cvt, gimple stmt,
2709 VEC (tree, heap) *vec_dsts,
2710 gimple_stmt_iterator *gsi,
2711 slp_tree slp_node, enum tree_code code1,
2712 enum tree_code code2, tree decl1,
2713 tree decl2, int op_type,
2714 stmt_vec_info *prev_stmt_info)
2717 tree vop0, vop1, new_tmp1, new_tmp2, vec_dest;
2718 gimple new_stmt1, new_stmt2;
2719 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2720 VEC (tree, heap) *vec_tmp;
2722 vec_dest = VEC_pop (tree, vec_dsts);
2723 vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2);
2725 for (i = 0; VEC_iterate (tree, *vec_oprnds0, i, vop0); i++)
2727 if (op_type == binary_op)
2728 vop1 = VEC_index (tree, *vec_oprnds1, i);
2732 /* Generate the two halves of promotion operation. */
2733 new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1,
2734 op_type, vec_dest, gsi, stmt);
2735 new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1,
2736 op_type, vec_dest, gsi, stmt);
2737 if (is_gimple_call (new_stmt1))
2739 new_tmp1 = gimple_call_lhs (new_stmt1);
2740 new_tmp2 = gimple_call_lhs (new_stmt2);
2744 new_tmp1 = gimple_assign_lhs (new_stmt1);
2745 new_tmp2 = gimple_assign_lhs (new_stmt2);
2750 /* Store the results for the recursive call. */
2751 VEC_quick_push (tree, vec_tmp, new_tmp1);
2752 VEC_quick_push (tree, vec_tmp, new_tmp2);
2756 /* Last step of promotion sequience - store the results. */
2759 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1);
2760 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2);
2764 if (!*prev_stmt_info)
2765 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1;
2767 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1;
2769 *prev_stmt_info = vinfo_for_stmt (new_stmt1);
2770 STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2;
2771 *prev_stmt_info = vinfo_for_stmt (new_stmt2);
2778 /* For multi-step promotion operation we first generate we call the
2779 function recurcively for every stage. We start from the input type,
2780 create promotion operations to the intermediate types, and then
2781 create promotions to the output type. */
2782 *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp);
2783 VEC_free (tree, heap, vec_tmp);
2784 vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1,
2785 multi_step_cvt - 1, stmt,
2786 vec_dsts, gsi, slp_node, code1,
2787 code2, decl2, decl2, op_type,
2793 /* Function vectorizable_type_promotion
2795 Check if STMT performs a binary or unary operation that involves
2796 type promotion, and if it can be vectorized.
2797 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2798 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2799 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2802 vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi,
2803 gimple *vec_stmt, slp_tree slp_node)
2807 tree op0, op1 = NULL;
2808 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
2809 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2810 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2811 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
2812 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
2816 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
2817 stmt_vec_info prev_stmt_info;
2824 tree intermediate_type = NULL_TREE;
2825 int multi_step_cvt = 0;
2826 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
2827 VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL;
2829 /* FORNOW: not supported by basic block SLP vectorization. */
2830 gcc_assert (loop_vinfo);
2832 if (!STMT_VINFO_RELEVANT_P (stmt_info))
2835 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
2838 /* Is STMT a vectorizable type-promotion operation? */
2839 if (!is_gimple_assign (stmt))
2842 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
2845 code = gimple_assign_rhs_code (stmt);
2846 if (!CONVERT_EXPR_CODE_P (code)
2847 && code != WIDEN_MULT_EXPR)
2850 scalar_dest = gimple_assign_lhs (stmt);
2851 vectype_out = STMT_VINFO_VECTYPE (stmt_info);
2853 /* Check the operands of the operation. */
2854 op0 = gimple_assign_rhs1 (stmt);
2855 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
2856 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
2857 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
2858 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
2859 && CONVERT_EXPR_CODE_P (code))))
2861 if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL,
2862 &def_stmt, &def, &dt[0], &vectype_in))
2864 if (vect_print_dump_info (REPORT_DETAILS))
2865 fprintf (vect_dump, "use not simple.");
2868 /* If op0 is an external or constant def use a vector type with
2869 the same size as the output vector type. */
2871 vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out);
2873 gcc_assert (vectype_in);
2876 if (vect_print_dump_info (REPORT_DETAILS))
2878 fprintf (vect_dump, "no vectype for scalar type ");
2879 print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM);
2885 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
2886 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
2887 if (nunits_in <= nunits_out)
2890 /* Multiple types in SLP are handled by creating the appropriate number of
2891 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2896 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
2898 gcc_assert (ncopies >= 1);
2900 op_type = TREE_CODE_LENGTH (code);
2901 if (op_type == binary_op)
2903 op1 = gimple_assign_rhs2 (stmt);
2904 if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1]))
2906 if (vect_print_dump_info (REPORT_DETAILS))
2907 fprintf (vect_dump, "use not simple.");
2912 /* Supportable by target? */
2913 if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in,
2914 &decl1, &decl2, &code1, &code2,
2915 &multi_step_cvt, &interm_types))
2918 /* Binary widening operation can only be supported directly by the
2920 gcc_assert (!(multi_step_cvt && op_type == binary_op));
2922 if (!vec_stmt) /* transformation not required. */
2924 STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type;
2925 if (vect_print_dump_info (REPORT_DETAILS))
2926 fprintf (vect_dump, "=== vectorizable_promotion ===");
2927 vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL);
2933 if (vect_print_dump_info (REPORT_DETAILS))
2934 fprintf (vect_dump, "transform type promotion operation. ncopies = %d.",
2938 /* In case of multi-step promotion, we first generate promotion operations
2939 to the intermediate types, and then from that types to the final one.
2940 We store vector destination in VEC_DSTS in the correct order for
2941 recursive creation of promotion operations in
2942 vect_create_vectorized_promotion_stmts(). Vector destinations are created
2943 according to TYPES recieved from supportable_widening_operation(). */
2945 vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1);
2947 vec_dsts = VEC_alloc (tree, heap, 1);
2949 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
2950 VEC_quick_push (tree, vec_dsts, vec_dest);
2954 for (i = VEC_length (tree, interm_types) - 1;
2955 VEC_iterate (tree, interm_types, i, intermediate_type); i--)
2957 vec_dest = vect_create_destination_var (scalar_dest,
2959 VEC_quick_push (tree, vec_dsts, vec_dest);
2965 vec_oprnds0 = VEC_alloc (tree, heap,
2966 (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1));
2967 if (op_type == binary_op)
2968 vec_oprnds1 = VEC_alloc (tree, heap, 1);
2971 /* In case the vectorization factor (VF) is bigger than the number
2972 of elements that we can fit in a vectype (nunits), we have to generate
2973 more than one vector stmt - i.e - we need to "unroll" the
2974 vector stmt by a factor VF/nunits. */
2976 prev_stmt_info = NULL;
2977 for (j = 0; j < ncopies; j++)
2983 vect_get_slp_defs (slp_node, &vec_oprnds0, &vec_oprnds1, -1);
2986 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
2987 VEC_quick_push (tree, vec_oprnds0, vec_oprnd0);
2988 if (op_type == binary_op)
2990 vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
2991 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
2997 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
2998 VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0);
2999 if (op_type == binary_op)
3001 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
3002 VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1);
3006 /* Arguments are ready. Create the new vector stmts. */
3007 tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts);
3008 vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1,
3009 multi_step_cvt, stmt,
3011 gsi, slp_node, code1, code2,
3012 decl1, decl2, op_type,
3016 VEC_free (tree, heap, vec_dsts);
3017 VEC_free (tree, heap, tmp_vec_dsts);
3018 VEC_free (tree, heap, interm_types);
3019 VEC_free (tree, heap, vec_oprnds0);
3020 VEC_free (tree, heap, vec_oprnds1);
3022 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3027 /* Function vectorizable_store.
3029 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
3031 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3032 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3033 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3036 vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3042 tree vec_oprnd = NULL_TREE;
3043 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3044 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL;
3045 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3046 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3047 struct loop *loop = NULL;
3048 enum machine_mode vec_mode;
3050 enum dr_alignment_support alignment_support_scheme;
3053 enum vect_def_type dt;
3054 stmt_vec_info prev_stmt_info = NULL;
3055 tree dataref_ptr = NULL_TREE;
3056 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3059 gimple next_stmt, first_stmt = NULL;
3060 bool strided_store = false;
3061 unsigned int group_size, i;
3062 VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL;
3064 VEC(tree,heap) *vec_oprnds = NULL;
3065 bool slp = (slp_node != NULL);
3066 unsigned int vec_num;
3067 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3070 loop = LOOP_VINFO_LOOP (loop_vinfo);
3072 /* Multiple types in SLP are handled by creating the appropriate number of
3073 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3078 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3080 gcc_assert (ncopies >= 1);
3082 /* FORNOW. This restriction should be relaxed. */
3083 if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3085 if (vect_print_dump_info (REPORT_DETAILS))
3086 fprintf (vect_dump, "multiple types in nested loop.");
3090 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3093 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3096 /* Is vectorizable store? */
3098 if (!is_gimple_assign (stmt))
3101 scalar_dest = gimple_assign_lhs (stmt);
3102 if (TREE_CODE (scalar_dest) != ARRAY_REF
3103 && TREE_CODE (scalar_dest) != INDIRECT_REF
3104 && TREE_CODE (scalar_dest) != COMPONENT_REF
3105 && TREE_CODE (scalar_dest) != IMAGPART_EXPR
3106 && TREE_CODE (scalar_dest) != REALPART_EXPR
3107 && TREE_CODE (scalar_dest) != MEM_REF)
3110 gcc_assert (gimple_assign_single_p (stmt));
3111 op = gimple_assign_rhs1 (stmt);
3112 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt))
3114 if (vect_print_dump_info (REPORT_DETAILS))
3115 fprintf (vect_dump, "use not simple.");
3119 /* The scalar rhs type needs to be trivially convertible to the vector
3120 component type. This should always be the case. */
3121 if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op)))
3123 if (vect_print_dump_info (REPORT_DETAILS))
3124 fprintf (vect_dump, "??? operands of different types");
3128 vec_mode = TYPE_MODE (vectype);
3129 /* FORNOW. In some cases can vectorize even if data-type not supported
3130 (e.g. - array initialization with 0). */
3131 if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing)
3134 if (!STMT_VINFO_DATA_REF (stmt_info))
3137 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3139 strided_store = true;
3140 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3141 if (!vect_strided_store_supported (vectype)
3142 && !PURE_SLP_STMT (stmt_info) && !slp)
3145 if (first_stmt == stmt)
3147 /* STMT is the leader of the group. Check the operands of all the
3148 stmts of the group. */
3149 next_stmt = DR_GROUP_NEXT_DR (stmt_info);
3152 gcc_assert (gimple_assign_single_p (next_stmt));
3153 op = gimple_assign_rhs1 (next_stmt);
3154 if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt,
3157 if (vect_print_dump_info (REPORT_DETAILS))
3158 fprintf (vect_dump, "use not simple.");
3161 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3166 if (!vec_stmt) /* transformation not required. */
3168 STMT_VINFO_TYPE (stmt_info) = store_vec_info_type;
3169 vect_model_store_cost (stmt_info, ncopies, dt, NULL);
3177 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3178 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3180 DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++;
3183 gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt));
3185 /* We vectorize all the stmts of the interleaving group when we
3186 reach the last stmt in the group. */
3187 if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))
3188 < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt))
3197 strided_store = false;
3198 /* VEC_NUM is the number of vect stmts to be created for this
3200 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3201 first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0);
3202 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3205 /* VEC_NUM is the number of vect stmts to be created for this
3207 vec_num = group_size;
3213 group_size = vec_num = 1;
3216 if (vect_print_dump_info (REPORT_DETAILS))
3217 fprintf (vect_dump, "transform store. ncopies = %d",ncopies);
3219 dr_chain = VEC_alloc (tree, heap, group_size);
3220 oprnds = VEC_alloc (tree, heap, group_size);
3222 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3223 gcc_assert (alignment_support_scheme);
3225 /* In case the vectorization factor (VF) is bigger than the number
3226 of elements that we can fit in a vectype (nunits), we have to generate
3227 more than one vector stmt - i.e - we need to "unroll" the
3228 vector stmt by a factor VF/nunits. For more details see documentation in
3229 vect_get_vec_def_for_copy_stmt. */
3231 /* In case of interleaving (non-unit strided access):
3238 We create vectorized stores starting from base address (the access of the
3239 first stmt in the chain (S2 in the above example), when the last store stmt
3240 of the chain (S4) is reached:
3243 VS2: &base + vec_size*1 = vx0
3244 VS3: &base + vec_size*2 = vx1
3245 VS4: &base + vec_size*3 = vx3
3247 Then permutation statements are generated:
3249 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
3250 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
3253 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3254 (the order of the data-refs in the output of vect_permute_store_chain
3255 corresponds to the order of scalar stmts in the interleaving chain - see
3256 the documentation of vect_permute_store_chain()).
3258 In case of both multiple types and interleaving, above vector stores and
3259 permutation stmts are created for every copy. The result vector stmts are
3260 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
3261 STMT_VINFO_RELATED_STMT for the next copies.
3264 prev_stmt_info = NULL;
3265 for (j = 0; j < ncopies; j++)
3274 /* Get vectorized arguments for SLP_NODE. */
3275 vect_get_slp_defs (slp_node, &vec_oprnds, NULL, -1);
3277 vec_oprnd = VEC_index (tree, vec_oprnds, 0);
3281 /* For interleaved stores we collect vectorized defs for all the
3282 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
3283 used as an input to vect_permute_store_chain(), and OPRNDS as
3284 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
3286 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3287 OPRNDS are of size 1. */
3288 next_stmt = first_stmt;
3289 for (i = 0; i < group_size; i++)
3291 /* Since gaps are not supported for interleaved stores,
3292 GROUP_SIZE is the exact number of stmts in the chain.
3293 Therefore, NEXT_STMT can't be NULL_TREE. In case that
3294 there is no interleaving, GROUP_SIZE is 1, and only one
3295 iteration of the loop will be executed. */
3296 gcc_assert (next_stmt
3297 && gimple_assign_single_p (next_stmt));
3298 op = gimple_assign_rhs1 (next_stmt);
3300 vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt,
3302 VEC_quick_push(tree, dr_chain, vec_oprnd);
3303 VEC_quick_push(tree, oprnds, vec_oprnd);
3304 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3308 /* We should have catched mismatched types earlier. */
3309 gcc_assert (useless_type_conversion_p (vectype,
3310 TREE_TYPE (vec_oprnd)));
3311 dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE,
3312 &dummy, &ptr_incr, false,
3314 gcc_assert (bb_vinfo || !inv_p);
3318 /* For interleaved stores we created vectorized defs for all the
3319 defs stored in OPRNDS in the previous iteration (previous copy).
3320 DR_CHAIN is then used as an input to vect_permute_store_chain(),
3321 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
3323 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3324 OPRNDS are of size 1. */
3325 for (i = 0; i < group_size; i++)
3327 op = VEC_index (tree, oprnds, i);
3328 vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def,
3330 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op);
3331 VEC_replace(tree, dr_chain, i, vec_oprnd);
3332 VEC_replace(tree, oprnds, i, vec_oprnd);
3335 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
3340 result_chain = VEC_alloc (tree, heap, group_size);
3342 if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi,
3347 next_stmt = first_stmt;
3348 for (i = 0; i < vec_num; i++)
3351 /* Bump the vector pointer. */
3352 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
3356 vec_oprnd = VEC_index (tree, vec_oprnds, i);
3357 else if (strided_store)
3358 /* For strided stores vectorized defs are interleaved in
3359 vect_permute_store_chain(). */
3360 vec_oprnd = VEC_index (tree, result_chain, i);
3362 if (aligned_access_p (first_dr))
3363 data_ref = build_simple_mem_ref (dataref_ptr);
3366 int mis = DR_MISALIGNMENT (first_dr);
3367 tree tmis = (mis == -1 ? size_zero_node : size_int (mis));
3368 tmis = size_binop (MULT_EXPR, tmis, size_int (BITS_PER_UNIT));
3369 data_ref = build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis);
3372 /* If accesses through a pointer to vectype do not alias the original
3373 memory reference we have a problem. This should never happen. */
3374 gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref),
3375 get_alias_set (gimple_assign_lhs (stmt))));
3377 /* Arguments are ready. Create the new vector stmt. */
3378 new_stmt = gimple_build_assign (data_ref, vec_oprnd);
3379 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3380 mark_symbols_for_renaming (new_stmt);
3386 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3388 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3390 prev_stmt_info = vinfo_for_stmt (new_stmt);
3391 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
3397 VEC_free (tree, heap, dr_chain);
3398 VEC_free (tree, heap, oprnds);
3400 VEC_free (tree, heap, result_chain);
3405 /* vectorizable_load.
3407 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
3409 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3410 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3411 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3414 vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt,
3415 slp_tree slp_node, slp_instance slp_node_instance)
3418 tree vec_dest = NULL;
3419 tree data_ref = NULL;
3420 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3421 stmt_vec_info prev_stmt_info;
3422 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3423 struct loop *loop = NULL;
3424 struct loop *containing_loop = (gimple_bb (stmt))->loop_father;
3425 bool nested_in_vect_loop = false;
3426 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
3427 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3429 enum machine_mode mode;
3430 gimple new_stmt = NULL;
3432 enum dr_alignment_support alignment_support_scheme;
3433 tree dataref_ptr = NULL_TREE;
3435 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3437 int i, j, group_size;
3438 tree msq = NULL_TREE, lsq;
3439 tree offset = NULL_TREE;
3440 tree realignment_token = NULL_TREE;
3442 VEC(tree,heap) *dr_chain = NULL;
3443 bool strided_load = false;
3447 bool compute_in_loop = false;
3448 struct loop *at_loop;
3450 bool slp = (slp_node != NULL);
3451 bool slp_perm = false;
3452 enum tree_code code;
3453 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3458 loop = LOOP_VINFO_LOOP (loop_vinfo);
3459 nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
3460 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
3465 /* Multiple types in SLP are handled by creating the appropriate number of
3466 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
3471 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3473 gcc_assert (ncopies >= 1);
3475 /* FORNOW. This restriction should be relaxed. */
3476 if (nested_in_vect_loop && ncopies > 1)
3478 if (vect_print_dump_info (REPORT_DETAILS))
3479 fprintf (vect_dump, "multiple types in nested loop.");
3483 if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
3486 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
3489 /* Is vectorizable load? */
3490 if (!is_gimple_assign (stmt))
3493 scalar_dest = gimple_assign_lhs (stmt);
3494 if (TREE_CODE (scalar_dest) != SSA_NAME)
3497 code = gimple_assign_rhs_code (stmt);
3498 if (code != ARRAY_REF
3499 && code != INDIRECT_REF
3500 && code != COMPONENT_REF
3501 && code != IMAGPART_EXPR
3502 && code != REALPART_EXPR
3506 if (!STMT_VINFO_DATA_REF (stmt_info))
3509 scalar_type = TREE_TYPE (DR_REF (dr));
3510 mode = TYPE_MODE (vectype);
3512 /* FORNOW. In some cases can vectorize even if data-type not supported
3513 (e.g. - data copies). */
3514 if (optab_handler (mov_optab, mode) == CODE_FOR_nothing)
3516 if (vect_print_dump_info (REPORT_DETAILS))
3517 fprintf (vect_dump, "Aligned load, but unsupported type.");
3521 /* The vector component type needs to be trivially convertible to the
3522 scalar lhs. This should always be the case. */
3523 if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype)))
3525 if (vect_print_dump_info (REPORT_DETAILS))
3526 fprintf (vect_dump, "??? operands of different types");
3530 /* Check if the load is a part of an interleaving chain. */
3531 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
3533 strided_load = true;
3535 gcc_assert (! nested_in_vect_loop);
3537 /* Check if interleaving is supported. */
3538 if (!vect_strided_load_supported (vectype)
3539 && !PURE_SLP_STMT (stmt_info) && !slp)
3543 if (!vec_stmt) /* transformation not required. */
3545 STMT_VINFO_TYPE (stmt_info) = load_vec_info_type;
3546 vect_model_load_cost (stmt_info, ncopies, NULL);
3550 if (vect_print_dump_info (REPORT_DETAILS))
3551 fprintf (vect_dump, "transform load.");
3557 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
3558 /* Check if the chain of loads is already vectorized. */
3559 if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)))
3561 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3564 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
3565 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
3567 /* VEC_NUM is the number of vect stmts to be created for this group. */
3570 strided_load = false;
3571 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
3572 if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance))
3576 vec_num = group_size;
3578 dr_chain = VEC_alloc (tree, heap, vec_num);
3584 group_size = vec_num = 1;
3587 alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false);
3588 gcc_assert (alignment_support_scheme);
3590 /* In case the vectorization factor (VF) is bigger than the number
3591 of elements that we can fit in a vectype (nunits), we have to generate
3592 more than one vector stmt - i.e - we need to "unroll" the
3593 vector stmt by a factor VF/nunits. In doing so, we record a pointer
3594 from one copy of the vector stmt to the next, in the field
3595 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
3596 stages to find the correct vector defs to be used when vectorizing
3597 stmts that use the defs of the current stmt. The example below illustrates
3598 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
3599 4 vectorized stmts):
3601 before vectorization:
3602 RELATED_STMT VEC_STMT
3606 step 1: vectorize stmt S1:
3607 We first create the vector stmt VS1_0, and, as usual, record a
3608 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
3609 Next, we create the vector stmt VS1_1, and record a pointer to
3610 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
3611 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
3613 RELATED_STMT VEC_STMT
3614 VS1_0: vx0 = memref0 VS1_1 -
3615 VS1_1: vx1 = memref1 VS1_2 -
3616 VS1_2: vx2 = memref2 VS1_3 -
3617 VS1_3: vx3 = memref3 - -
3618 S1: x = load - VS1_0
3621 See in documentation in vect_get_vec_def_for_stmt_copy for how the
3622 information we recorded in RELATED_STMT field is used to vectorize
3625 /* In case of interleaving (non-unit strided access):
3632 Vectorized loads are created in the order of memory accesses
3633 starting from the access of the first stmt of the chain:
3636 VS2: vx1 = &base + vec_size*1
3637 VS3: vx3 = &base + vec_size*2
3638 VS4: vx4 = &base + vec_size*3
3640 Then permutation statements are generated:
3642 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
3643 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
3646 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3647 (the order of the data-refs in the output of vect_permute_load_chain
3648 corresponds to the order of scalar stmts in the interleaving chain - see
3649 the documentation of vect_permute_load_chain()).
3650 The generation of permutation stmts and recording them in
3651 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
3653 In case of both multiple types and interleaving, the vector loads and
3654 permutation stmts above are created for every copy. The result vector stmts
3655 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
3656 STMT_VINFO_RELATED_STMT for the next copies. */
3658 /* If the data reference is aligned (dr_aligned) or potentially unaligned
3659 on a target that supports unaligned accesses (dr_unaligned_supported)
3660 we generate the following code:
3664 p = p + indx * vectype_size;
3669 Otherwise, the data reference is potentially unaligned on a target that
3670 does not support unaligned accesses (dr_explicit_realign_optimized) -
3671 then generate the following code, in which the data in each iteration is
3672 obtained by two vector loads, one from the previous iteration, and one
3673 from the current iteration:
3675 msq_init = *(floor(p1))
3676 p2 = initial_addr + VS - 1;
3677 realignment_token = call target_builtin;
3680 p2 = p2 + indx * vectype_size
3682 vec_dest = realign_load (msq, lsq, realignment_token)
3687 /* If the misalignment remains the same throughout the execution of the
3688 loop, we can create the init_addr and permutation mask at the loop
3689 preheader. Otherwise, it needs to be created inside the loop.
3690 This can only occur when vectorizing memory accesses in the inner-loop
3691 nested within an outer-loop that is being vectorized. */
3693 if (loop && nested_in_vect_loop_p (loop, stmt)
3694 && (TREE_INT_CST_LOW (DR_STEP (dr))
3695 % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0))
3697 gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized);
3698 compute_in_loop = true;
3701 if ((alignment_support_scheme == dr_explicit_realign_optimized
3702 || alignment_support_scheme == dr_explicit_realign)
3703 && !compute_in_loop)
3705 msq = vect_setup_realignment (first_stmt, gsi, &realignment_token,
3706 alignment_support_scheme, NULL_TREE,
3708 if (alignment_support_scheme == dr_explicit_realign_optimized)
3710 phi = SSA_NAME_DEF_STMT (msq);
3711 offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
3717 prev_stmt_info = NULL;
3718 for (j = 0; j < ncopies; j++)
3720 /* 1. Create the vector pointer update chain. */
3722 dataref_ptr = vect_create_data_ref_ptr (first_stmt,
3724 &dummy, &ptr_incr, false,
3728 bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE);
3730 for (i = 0; i < vec_num; i++)
3733 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
3736 /* 2. Create the vector-load in the loop. */
3737 switch (alignment_support_scheme)
3740 gcc_assert (aligned_access_p (first_dr));
3741 data_ref = build_simple_mem_ref (dataref_ptr);
3743 case dr_unaligned_supported:
3745 int mis = DR_MISALIGNMENT (first_dr);
3746 tree tmis = (mis == -1 ? size_zero_node : size_int (mis));
3748 tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT));
3750 build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis);
3753 case dr_explicit_realign:
3756 tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
3758 if (compute_in_loop)
3759 msq = vect_setup_realignment (first_stmt, gsi,
3761 dr_explicit_realign,
3764 new_stmt = gimple_build_assign_with_ops
3765 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
3767 (TREE_TYPE (dataref_ptr),
3768 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
3769 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
3770 gimple_assign_set_lhs (new_stmt, ptr);
3771 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3772 data_ref = build_simple_mem_ref (ptr);
3773 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3774 new_stmt = gimple_build_assign (vec_dest, data_ref);
3775 new_temp = make_ssa_name (vec_dest, new_stmt);
3776 gimple_assign_set_lhs (new_stmt, new_temp);
3777 gimple_set_vdef (new_stmt, gimple_vdef (stmt));
3778 gimple_set_vuse (new_stmt, gimple_vuse (stmt));
3779 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3782 bump = size_binop (MULT_EXPR, vs_minus_1,
3783 TYPE_SIZE_UNIT (scalar_type));
3784 ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump);
3785 new_stmt = gimple_build_assign_with_ops
3786 (BIT_AND_EXPR, NULL_TREE, ptr,
3789 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
3790 ptr = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
3791 gimple_assign_set_lhs (new_stmt, ptr);
3792 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3793 data_ref = build_simple_mem_ref (ptr);
3796 case dr_explicit_realign_optimized:
3797 new_stmt = gimple_build_assign_with_ops
3798 (BIT_AND_EXPR, NULL_TREE, dataref_ptr,
3800 (TREE_TYPE (dataref_ptr),
3801 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
3802 new_temp = make_ssa_name (SSA_NAME_VAR (dataref_ptr), new_stmt);
3803 gimple_assign_set_lhs (new_stmt, new_temp);
3804 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3805 data_ref = build_simple_mem_ref (new_temp);
3810 /* If accesses through a pointer to vectype do not alias the original
3811 memory reference we have a problem. This should never happen. */
3812 gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref),
3813 get_alias_set (gimple_assign_rhs1 (stmt))));
3814 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3815 new_stmt = gimple_build_assign (vec_dest, data_ref);
3816 new_temp = make_ssa_name (vec_dest, new_stmt);
3817 gimple_assign_set_lhs (new_stmt, new_temp);
3818 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3819 mark_symbols_for_renaming (new_stmt);
3821 /* 3. Handle explicit realignment if necessary/supported. Create in
3822 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
3823 if (alignment_support_scheme == dr_explicit_realign_optimized
3824 || alignment_support_scheme == dr_explicit_realign)
3828 lsq = gimple_assign_lhs (new_stmt);
3829 if (!realignment_token)
3830 realignment_token = dataref_ptr;
3831 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3832 tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq,
3834 new_stmt = gimple_build_assign (vec_dest, tmp);
3835 new_temp = make_ssa_name (vec_dest, new_stmt);
3836 gimple_assign_set_lhs (new_stmt, new_temp);
3837 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3839 if (alignment_support_scheme == dr_explicit_realign_optimized)
3842 if (i == vec_num - 1 && j == ncopies - 1)
3843 add_phi_arg (phi, lsq, loop_latch_edge (containing_loop),
3849 /* 4. Handle invariant-load. */
3850 if (inv_p && !bb_vinfo)
3852 gcc_assert (!strided_load);
3853 gcc_assert (nested_in_vect_loop_p (loop, stmt));
3858 tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type);
3860 /* CHECKME: bitpos depends on endianess? */
3861 bitpos = bitsize_zero_node;
3862 vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp,
3865 vect_create_destination_var (scalar_dest, NULL_TREE);
3866 new_stmt = gimple_build_assign (vec_dest, vec_inv);
3867 new_temp = make_ssa_name (vec_dest, new_stmt);
3868 gimple_assign_set_lhs (new_stmt, new_temp);
3869 vect_finish_stmt_generation (stmt, new_stmt, gsi);
3871 for (k = nunits - 1; k >= 0; --k)
3872 t = tree_cons (NULL_TREE, new_temp, t);
3873 /* FIXME: use build_constructor directly. */
3874 vec_inv = build_constructor_from_list (vectype, t);
3875 new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi);
3876 new_stmt = SSA_NAME_DEF_STMT (new_temp);
3879 gcc_unreachable (); /* FORNOW. */
3882 /* Collect vector loads and later create their permutation in
3883 vect_transform_strided_load (). */
3884 if (strided_load || slp_perm)
3885 VEC_quick_push (tree, dr_chain, new_temp);
3887 /* Store vector loads in the corresponding SLP_NODE. */
3888 if (slp && !slp_perm)
3889 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
3892 if (slp && !slp_perm)
3897 if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf,
3898 slp_node_instance, false))
3900 VEC_free (tree, heap, dr_chain);
3908 if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi))
3911 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3912 VEC_free (tree, heap, dr_chain);
3913 dr_chain = VEC_alloc (tree, heap, group_size);
3918 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3920 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3921 prev_stmt_info = vinfo_for_stmt (new_stmt);
3927 VEC_free (tree, heap, dr_chain);
3932 /* Function vect_is_simple_cond.
3935 LOOP - the loop that is being vectorized.
3936 COND - Condition that is checked for simple use.
3938 Returns whether a COND can be vectorized. Checks whether
3939 condition operands are supportable using vec_is_simple_use. */
3942 vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo)
3946 enum vect_def_type dt;
3948 if (!COMPARISON_CLASS_P (cond))
3951 lhs = TREE_OPERAND (cond, 0);
3952 rhs = TREE_OPERAND (cond, 1);
3954 if (TREE_CODE (lhs) == SSA_NAME)
3956 gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs);
3957 if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def,
3961 else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST
3962 && TREE_CODE (lhs) != FIXED_CST)
3965 if (TREE_CODE (rhs) == SSA_NAME)
3967 gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs);
3968 if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def,
3972 else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST
3973 && TREE_CODE (rhs) != FIXED_CST)
3979 /* vectorizable_condition.
3981 Check if STMT is conditional modify expression that can be vectorized.
3982 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3983 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
3986 When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable
3987 to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in
3988 else caluse if it is 2).
3990 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3993 vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi,
3994 gimple *vec_stmt, tree reduc_def, int reduc_index)
3996 tree scalar_dest = NULL_TREE;
3997 tree vec_dest = NULL_TREE;
3998 tree op = NULL_TREE;
3999 tree cond_expr, then_clause, else_clause;
4000 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4001 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4002 tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause;
4003 tree vec_compare, vec_cond_expr;
4005 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4006 enum machine_mode vec_mode;
4008 enum vect_def_type dt;
4009 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
4010 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
4011 enum tree_code code;
4013 /* FORNOW: unsupported in basic block SLP. */
4014 gcc_assert (loop_vinfo);
4016 gcc_assert (ncopies >= 1);
4018 return false; /* FORNOW */
4020 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4023 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
4024 && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle
4028 /* FORNOW: SLP not supported. */
4029 if (STMT_SLP_TYPE (stmt_info))
4032 /* FORNOW: not yet supported. */
4033 if (STMT_VINFO_LIVE_P (stmt_info))
4035 if (vect_print_dump_info (REPORT_DETAILS))
4036 fprintf (vect_dump, "value used after loop.");
4040 /* Is vectorizable conditional operation? */
4041 if (!is_gimple_assign (stmt))
4044 code = gimple_assign_rhs_code (stmt);
4046 if (code != COND_EXPR)
4049 gcc_assert (gimple_assign_single_p (stmt));
4050 op = gimple_assign_rhs1 (stmt);
4051 cond_expr = TREE_OPERAND (op, 0);
4052 then_clause = TREE_OPERAND (op, 1);
4053 else_clause = TREE_OPERAND (op, 2);
4055 if (!vect_is_simple_cond (cond_expr, loop_vinfo))
4058 /* We do not handle two different vector types for the condition
4060 if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)),
4061 TREE_TYPE (vectype)))
4064 if (TREE_CODE (then_clause) == SSA_NAME)
4066 gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause);
4067 if (!vect_is_simple_use (then_clause, loop_vinfo, NULL,
4068 &then_def_stmt, &def, &dt))
4071 else if (TREE_CODE (then_clause) != INTEGER_CST
4072 && TREE_CODE (then_clause) != REAL_CST
4073 && TREE_CODE (then_clause) != FIXED_CST)
4076 if (TREE_CODE (else_clause) == SSA_NAME)
4078 gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause);
4079 if (!vect_is_simple_use (else_clause, loop_vinfo, NULL,
4080 &else_def_stmt, &def, &dt))
4083 else if (TREE_CODE (else_clause) != INTEGER_CST
4084 && TREE_CODE (else_clause) != REAL_CST
4085 && TREE_CODE (else_clause) != FIXED_CST)
4089 vec_mode = TYPE_MODE (vectype);
4093 STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type;
4094 return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode);
4100 scalar_dest = gimple_assign_lhs (stmt);
4101 vec_dest = vect_create_destination_var (scalar_dest, vectype);
4103 /* Handle cond expr. */
4105 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL);
4107 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL);
4108 if (reduc_index == 1)
4109 vec_then_clause = reduc_def;
4111 vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL);
4112 if (reduc_index == 2)
4113 vec_else_clause = reduc_def;
4115 vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL);
4117 /* Arguments are ready. Create the new vector stmt. */
4118 vec_compare = build2 (TREE_CODE (cond_expr), vectype,
4119 vec_cond_lhs, vec_cond_rhs);
4120 vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
4121 vec_compare, vec_then_clause, vec_else_clause);
4123 *vec_stmt = gimple_build_assign (vec_dest, vec_cond_expr);
4124 new_temp = make_ssa_name (vec_dest, *vec_stmt);
4125 gimple_assign_set_lhs (*vec_stmt, new_temp);
4126 vect_finish_stmt_generation (stmt, *vec_stmt, gsi);
4132 /* Make sure the statement is vectorizable. */
4135 vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node)
4137 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4138 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
4139 enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info);
4141 tree scalar_type, vectype;
4143 if (vect_print_dump_info (REPORT_DETAILS))
4145 fprintf (vect_dump, "==> examining statement: ");
4146 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4149 if (gimple_has_volatile_ops (stmt))
4151 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4152 fprintf (vect_dump, "not vectorized: stmt has volatile operands");
4157 /* Skip stmts that do not need to be vectorized. In loops this is expected
4159 - the COND_EXPR which is the loop exit condition
4160 - any LABEL_EXPRs in the loop
4161 - computations that are used only for array indexing or loop control.
4162 In basic blocks we only analyze statements that are a part of some SLP
4163 instance, therefore, all the statements are relevant. */
4165 if (!STMT_VINFO_RELEVANT_P (stmt_info)
4166 && !STMT_VINFO_LIVE_P (stmt_info))
4168 if (vect_print_dump_info (REPORT_DETAILS))
4169 fprintf (vect_dump, "irrelevant.");
4174 switch (STMT_VINFO_DEF_TYPE (stmt_info))
4176 case vect_internal_def:
4179 case vect_reduction_def:
4180 case vect_nested_cycle:
4181 gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer
4182 || relevance == vect_used_in_outer_by_reduction
4183 || relevance == vect_unused_in_scope));
4186 case vect_induction_def:
4187 case vect_constant_def:
4188 case vect_external_def:
4189 case vect_unknown_def_type:
4196 gcc_assert (PURE_SLP_STMT (stmt_info));
4198 scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
4199 if (vect_print_dump_info (REPORT_DETAILS))
4201 fprintf (vect_dump, "get vectype for scalar type: ");
4202 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4205 vectype = get_vectype_for_scalar_type (scalar_type);
4208 if (vect_print_dump_info (REPORT_DETAILS))
4210 fprintf (vect_dump, "not SLPed: unsupported data-type ");
4211 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4216 if (vect_print_dump_info (REPORT_DETAILS))
4218 fprintf (vect_dump, "vectype: ");
4219 print_generic_expr (vect_dump, vectype, TDF_SLIM);
4222 STMT_VINFO_VECTYPE (stmt_info) = vectype;
4225 if (STMT_VINFO_RELEVANT_P (stmt_info))
4227 gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt))));
4228 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
4229 *need_to_vectorize = true;
4234 && (STMT_VINFO_RELEVANT_P (stmt_info)
4235 || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def))
4236 ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL)
4237 || vectorizable_type_demotion (stmt, NULL, NULL, NULL)
4238 || vectorizable_conversion (stmt, NULL, NULL, NULL)
4239 || vectorizable_operation (stmt, NULL, NULL, NULL)
4240 || vectorizable_assignment (stmt, NULL, NULL, NULL)
4241 || vectorizable_load (stmt, NULL, NULL, NULL, NULL)
4242 || vectorizable_call (stmt, NULL, NULL)
4243 || vectorizable_store (stmt, NULL, NULL, NULL)
4244 || vectorizable_reduction (stmt, NULL, NULL, NULL)
4245 || vectorizable_condition (stmt, NULL, NULL, NULL, 0));
4249 ok = (vectorizable_operation (stmt, NULL, NULL, node)
4250 || vectorizable_assignment (stmt, NULL, NULL, node)
4251 || vectorizable_load (stmt, NULL, NULL, node, NULL)
4252 || vectorizable_store (stmt, NULL, NULL, node));
4257 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4259 fprintf (vect_dump, "not vectorized: relevant stmt not ");
4260 fprintf (vect_dump, "supported: ");
4261 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4270 /* Stmts that are (also) "live" (i.e. - that are used out of the loop)
4271 need extra handling, except for vectorizable reductions. */
4272 if (STMT_VINFO_LIVE_P (stmt_info)
4273 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4274 ok = vectorizable_live_operation (stmt, NULL, NULL);
4278 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
4280 fprintf (vect_dump, "not vectorized: live stmt not ");
4281 fprintf (vect_dump, "supported: ");
4282 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4288 if (!PURE_SLP_STMT (stmt_info))
4290 /* Groups of strided accesses whose size is not a power of 2 are not
4291 vectorizable yet using loop-vectorization. Therefore, if this stmt
4292 feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and
4293 loop-based vectorized), the loop cannot be vectorized. */
4294 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
4295 && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt (
4296 DR_GROUP_FIRST_DR (stmt_info)))) == -1)
4298 if (vect_print_dump_info (REPORT_DETAILS))
4300 fprintf (vect_dump, "not vectorized: the size of group "
4301 "of strided accesses is not a power of 2");
4302 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
4313 /* Function vect_transform_stmt.
4315 Create a vectorized stmt to replace STMT, and insert it at BSI. */
4318 vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi,
4319 bool *strided_store, slp_tree slp_node,
4320 slp_instance slp_node_instance)
4322 bool is_store = false;
4323 gimple vec_stmt = NULL;
4324 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4325 gimple orig_stmt_in_pattern;
4328 switch (STMT_VINFO_TYPE (stmt_info))
4330 case type_demotion_vec_info_type:
4331 done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node);
4335 case type_promotion_vec_info_type:
4336 done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node);
4340 case type_conversion_vec_info_type:
4341 done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node);
4345 case induc_vec_info_type:
4346 gcc_assert (!slp_node);
4347 done = vectorizable_induction (stmt, gsi, &vec_stmt);
4351 case op_vec_info_type:
4352 done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node);
4356 case assignment_vec_info_type:
4357 done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node);
4361 case load_vec_info_type:
4362 done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node,
4367 case store_vec_info_type:
4368 done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node);
4370 if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node)
4372 /* In case of interleaving, the whole chain is vectorized when the
4373 last store in the chain is reached. Store stmts before the last
4374 one are skipped, and there vec_stmt_info shouldn't be freed
4376 *strided_store = true;
4377 if (STMT_VINFO_VEC_STMT (stmt_info))
4384 case condition_vec_info_type:
4385 gcc_assert (!slp_node);
4386 done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0);
4390 case call_vec_info_type:
4391 gcc_assert (!slp_node);
4392 done = vectorizable_call (stmt, gsi, &vec_stmt);
4395 case reduc_vec_info_type:
4396 done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node);
4401 if (!STMT_VINFO_LIVE_P (stmt_info))
4403 if (vect_print_dump_info (REPORT_DETAILS))
4404 fprintf (vect_dump, "stmt not supported.");
4409 /* Handle inner-loop stmts whose DEF is used in the loop-nest that
4410 is being vectorized, but outside the immediately enclosing loop. */
4412 && STMT_VINFO_LOOP_VINFO (stmt_info)
4413 && nested_in_vect_loop_p (LOOP_VINFO_LOOP (
4414 STMT_VINFO_LOOP_VINFO (stmt_info)), stmt)
4415 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type
4416 && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer
4417 || STMT_VINFO_RELEVANT (stmt_info) ==
4418 vect_used_in_outer_by_reduction))
4420 struct loop *innerloop = LOOP_VINFO_LOOP (
4421 STMT_VINFO_LOOP_VINFO (stmt_info))->inner;
4422 imm_use_iterator imm_iter;
4423 use_operand_p use_p;
4427 if (vect_print_dump_info (REPORT_DETAILS))
4428 fprintf (vect_dump, "Record the vdef for outer-loop vectorization.");
4430 /* Find the relevant loop-exit phi-node, and reord the vec_stmt there
4431 (to be used when vectorizing outer-loop stmts that use the DEF of
4433 if (gimple_code (stmt) == GIMPLE_PHI)
4434 scalar_dest = PHI_RESULT (stmt);
4436 scalar_dest = gimple_assign_lhs (stmt);
4438 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4440 if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p))))
4442 exit_phi = USE_STMT (use_p);
4443 STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt;
4448 /* Handle stmts whose DEF is used outside the loop-nest that is
4449 being vectorized. */
4450 if (STMT_VINFO_LIVE_P (stmt_info)
4451 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
4453 done = vectorizable_live_operation (stmt, gsi, &vec_stmt);
4459 STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
4460 orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info);
4461 if (orig_stmt_in_pattern)
4463 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern);
4464 /* STMT was inserted by the vectorizer to replace a computation idiom.
4465 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
4466 computed this idiom. We need to record a pointer to VEC_STMT in
4467 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
4468 documentation of vect_pattern_recog. */
4469 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
4471 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
4472 STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
4481 /* Remove a group of stores (for SLP or interleaving), free their
4485 vect_remove_stores (gimple first_stmt)
4487 gimple next = first_stmt;
4489 gimple_stmt_iterator next_si;
4493 /* Free the attached stmt_vec_info and remove the stmt. */
4494 next_si = gsi_for_stmt (next);
4495 gsi_remove (&next_si, true);
4496 tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next));
4497 free_stmt_vec_info (next);
4503 /* Function new_stmt_vec_info.
4505 Create and initialize a new stmt_vec_info struct for STMT. */
4508 new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo,
4509 bb_vec_info bb_vinfo)
4512 res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info));
4514 STMT_VINFO_TYPE (res) = undef_vec_info_type;
4515 STMT_VINFO_STMT (res) = stmt;
4516 STMT_VINFO_LOOP_VINFO (res) = loop_vinfo;
4517 STMT_VINFO_BB_VINFO (res) = bb_vinfo;
4518 STMT_VINFO_RELEVANT (res) = vect_unused_in_scope;
4519 STMT_VINFO_LIVE_P (res) = false;
4520 STMT_VINFO_VECTYPE (res) = NULL;
4521 STMT_VINFO_VEC_STMT (res) = NULL;
4522 STMT_VINFO_VECTORIZABLE (res) = true;
4523 STMT_VINFO_IN_PATTERN_P (res) = false;
4524 STMT_VINFO_RELATED_STMT (res) = NULL;
4525 STMT_VINFO_DATA_REF (res) = NULL;
4527 STMT_VINFO_DR_BASE_ADDRESS (res) = NULL;
4528 STMT_VINFO_DR_OFFSET (res) = NULL;
4529 STMT_VINFO_DR_INIT (res) = NULL;
4530 STMT_VINFO_DR_STEP (res) = NULL;
4531 STMT_VINFO_DR_ALIGNED_TO (res) = NULL;
4533 if (gimple_code (stmt) == GIMPLE_PHI
4534 && is_loop_header_bb_p (gimple_bb (stmt)))
4535 STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
4537 STMT_VINFO_DEF_TYPE (res) = vect_internal_def;
4539 STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5);
4540 STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0;
4541 STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0;
4542 STMT_SLP_TYPE (res) = loop_vect;
4543 DR_GROUP_FIRST_DR (res) = NULL;
4544 DR_GROUP_NEXT_DR (res) = NULL;
4545 DR_GROUP_SIZE (res) = 0;
4546 DR_GROUP_STORE_COUNT (res) = 0;
4547 DR_GROUP_GAP (res) = 0;
4548 DR_GROUP_SAME_DR_STMT (res) = NULL;
4549 DR_GROUP_READ_WRITE_DEPENDENCE (res) = false;
4555 /* Create a hash table for stmt_vec_info. */
4558 init_stmt_vec_info_vec (void)
4560 gcc_assert (!stmt_vec_info_vec);
4561 stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50);
4565 /* Free hash table for stmt_vec_info. */
4568 free_stmt_vec_info_vec (void)
4570 gcc_assert (stmt_vec_info_vec);
4571 VEC_free (vec_void_p, heap, stmt_vec_info_vec);
4575 /* Free stmt vectorization related info. */
4578 free_stmt_vec_info (gimple stmt)
4580 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4585 VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info));
4586 set_vinfo_for_stmt (stmt, NULL);
4591 /* Function get_vectype_for_scalar_type.
4593 Returns the vector type corresponding to SCALAR_TYPE as supported
4597 get_vectype_for_scalar_type (tree scalar_type)
4599 enum machine_mode inner_mode = TYPE_MODE (scalar_type);
4600 unsigned int nbytes = GET_MODE_SIZE (inner_mode);
4604 if (nbytes == 0 || nbytes >= UNITS_PER_SIMD_WORD (inner_mode))
4607 /* We can't build a vector type of elements with alignment bigger than
4609 if (nbytes < TYPE_ALIGN_UNIT (scalar_type))
4612 /* If we'd build a vector type of elements whose mode precision doesn't
4613 match their types precision we'll get mismatched types on vector
4614 extracts via BIT_FIELD_REFs. This effectively means we disable
4615 vectorization of bool and/or enum types in some languages. */
4616 if (INTEGRAL_TYPE_P (scalar_type)
4617 && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type))
4620 /* FORNOW: Only a single vector size per mode (UNITS_PER_SIMD_WORD)
4622 nunits = UNITS_PER_SIMD_WORD (inner_mode) / nbytes;
4624 vectype = build_vector_type (scalar_type, nunits);
4625 if (vect_print_dump_info (REPORT_DETAILS))
4627 fprintf (vect_dump, "get vectype with %d units of type ", nunits);
4628 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
4634 if (vect_print_dump_info (REPORT_DETAILS))
4636 fprintf (vect_dump, "vectype: ");
4637 print_generic_expr (vect_dump, vectype, TDF_SLIM);
4640 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
4641 && !INTEGRAL_MODE_P (TYPE_MODE (vectype)))
4643 if (vect_print_dump_info (REPORT_DETAILS))
4644 fprintf (vect_dump, "mode not supported by target.");
4651 /* Function get_same_sized_vectype
4653 Returns a vector type corresponding to SCALAR_TYPE of size
4654 VECTOR_TYPE if supported by the target. */
4657 get_same_sized_vectype (tree scalar_type, tree vector_type ATTRIBUTE_UNUSED)
4659 return get_vectype_for_scalar_type (scalar_type);
4662 /* Function vect_is_simple_use.
4665 LOOP_VINFO - the vect info of the loop that is being vectorized.
4666 BB_VINFO - the vect info of the basic block that is being vectorized.
4667 OPERAND - operand of a stmt in the loop or bb.
4668 DEF - the defining stmt in case OPERAND is an SSA_NAME.
4670 Returns whether a stmt with OPERAND can be vectorized.
4671 For loops, supportable operands are constants, loop invariants, and operands
4672 that are defined by the current iteration of the loop. Unsupportable
4673 operands are those that are defined by a previous iteration of the loop (as
4674 is the case in reduction/induction computations).
4675 For basic blocks, supportable operands are constants and bb invariants.
4676 For now, operands defined outside the basic block are not supported. */
4679 vect_is_simple_use (tree operand, loop_vec_info loop_vinfo,
4680 bb_vec_info bb_vinfo, gimple *def_stmt,
4681 tree *def, enum vect_def_type *dt)
4684 stmt_vec_info stmt_vinfo;
4685 struct loop *loop = NULL;
4688 loop = LOOP_VINFO_LOOP (loop_vinfo);
4693 if (vect_print_dump_info (REPORT_DETAILS))
4695 fprintf (vect_dump, "vect_is_simple_use: operand ");
4696 print_generic_expr (vect_dump, operand, TDF_SLIM);
4699 if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST)
4701 *dt = vect_constant_def;
4705 if (is_gimple_min_invariant (operand))
4708 *dt = vect_external_def;
4712 if (TREE_CODE (operand) == PAREN_EXPR)
4714 if (vect_print_dump_info (REPORT_DETAILS))
4715 fprintf (vect_dump, "non-associatable copy.");
4716 operand = TREE_OPERAND (operand, 0);
4719 if (TREE_CODE (operand) != SSA_NAME)
4721 if (vect_print_dump_info (REPORT_DETAILS))
4722 fprintf (vect_dump, "not ssa-name.");
4726 *def_stmt = SSA_NAME_DEF_STMT (operand);
4727 if (*def_stmt == NULL)
4729 if (vect_print_dump_info (REPORT_DETAILS))
4730 fprintf (vect_dump, "no def_stmt.");
4734 if (vect_print_dump_info (REPORT_DETAILS))
4736 fprintf (vect_dump, "def_stmt: ");
4737 print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM);
4740 /* Empty stmt is expected only in case of a function argument.
4741 (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */
4742 if (gimple_nop_p (*def_stmt))
4745 *dt = vect_external_def;
4749 bb = gimple_bb (*def_stmt);
4751 if ((loop && !flow_bb_inside_loop_p (loop, bb))
4752 || (!loop && bb != BB_VINFO_BB (bb_vinfo))
4753 || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI))
4754 *dt = vect_external_def;
4757 stmt_vinfo = vinfo_for_stmt (*def_stmt);
4758 *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo);
4761 if (*dt == vect_unknown_def_type)
4763 if (vect_print_dump_info (REPORT_DETAILS))
4764 fprintf (vect_dump, "Unsupported pattern.");
4768 if (vect_print_dump_info (REPORT_DETAILS))
4769 fprintf (vect_dump, "type of def: %d.",*dt);
4771 switch (gimple_code (*def_stmt))
4774 *def = gimple_phi_result (*def_stmt);
4778 *def = gimple_assign_lhs (*def_stmt);
4782 *def = gimple_call_lhs (*def_stmt);
4787 if (vect_print_dump_info (REPORT_DETAILS))
4788 fprintf (vect_dump, "unsupported defining stmt: ");
4795 /* Function vect_is_simple_use_1.
4797 Same as vect_is_simple_use_1 but also determines the vector operand
4798 type of OPERAND and stores it to *VECTYPE. If the definition of
4799 OPERAND is vect_uninitialized_def, vect_constant_def or
4800 vect_external_def *VECTYPE will be set to NULL_TREE and the caller
4801 is responsible to compute the best suited vector type for the
4805 vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo,
4806 bb_vec_info bb_vinfo, gimple *def_stmt,
4807 tree *def, enum vect_def_type *dt, tree *vectype)
4809 if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt))
4812 /* Now get a vector type if the def is internal, otherwise supply
4813 NULL_TREE and leave it up to the caller to figure out a proper
4814 type for the use stmt. */
4815 if (*dt == vect_internal_def
4816 || *dt == vect_induction_def
4817 || *dt == vect_reduction_def
4818 || *dt == vect_double_reduction_def
4819 || *dt == vect_nested_cycle)
4821 stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt);
4822 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
4823 stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info));
4824 *vectype = STMT_VINFO_VECTYPE (stmt_info);
4825 gcc_assert (*vectype != NULL_TREE);
4827 else if (*dt == vect_uninitialized_def
4828 || *dt == vect_constant_def
4829 || *dt == vect_external_def)
4830 *vectype = NULL_TREE;
4838 /* Function supportable_widening_operation
4840 Check whether an operation represented by the code CODE is a
4841 widening operation that is supported by the target platform in
4842 vector form (i.e., when operating on arguments of type VECTYPE_IN
4843 producing a result of type VECTYPE_OUT).
4845 Widening operations we currently support are NOP (CONVERT), FLOAT
4846 and WIDEN_MULT. This function checks if these operations are supported
4847 by the target platform either directly (via vector tree-codes), or via
4851 - CODE1 and CODE2 are codes of vector operations to be used when
4852 vectorizing the operation, if available.
4853 - DECL1 and DECL2 are decls of target builtin functions to be used
4854 when vectorizing the operation, if available. In this case,
4855 CODE1 and CODE2 are CALL_EXPR.
4856 - MULTI_STEP_CVT determines the number of required intermediate steps in
4857 case of multi-step conversion (like char->short->int - in that case
4858 MULTI_STEP_CVT will be 1).
4859 - INTERM_TYPES contains the intermediate type required to perform the
4860 widening operation (short in the above example). */
4863 supportable_widening_operation (enum tree_code code, gimple stmt,
4864 tree vectype_out, tree vectype_in,
4865 tree *decl1, tree *decl2,
4866 enum tree_code *code1, enum tree_code *code2,
4867 int *multi_step_cvt,
4868 VEC (tree, heap) **interm_types)
4870 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4871 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info);
4872 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
4874 enum machine_mode vec_mode;
4875 enum insn_code icode1, icode2;
4876 optab optab1, optab2;
4877 tree vectype = vectype_in;
4878 tree wide_vectype = vectype_out;
4879 enum tree_code c1, c2;
4881 /* The result of a vectorized widening operation usually requires two vectors
4882 (because the widened results do not fit int one vector). The generated
4883 vector results would normally be expected to be generated in the same
4884 order as in the original scalar computation, i.e. if 8 results are
4885 generated in each vector iteration, they are to be organized as follows:
4886 vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8].
4888 However, in the special case that the result of the widening operation is
4889 used in a reduction computation only, the order doesn't matter (because
4890 when vectorizing a reduction we change the order of the computation).
4891 Some targets can take advantage of this and generate more efficient code.
4892 For example, targets like Altivec, that support widen_mult using a sequence
4893 of {mult_even,mult_odd} generate the following vectors:
4894 vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8].
4896 When vectorizing outer-loops, we execute the inner-loop sequentially
4897 (each vectorized inner-loop iteration contributes to VF outer-loop
4898 iterations in parallel). We therefore don't allow to change the order
4899 of the computation in the inner-loop during outer-loop vectorization. */
4901 if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction
4902 && !nested_in_vect_loop_p (vect_loop, stmt))
4908 && code == WIDEN_MULT_EXPR
4909 && targetm.vectorize.builtin_mul_widen_even
4910 && targetm.vectorize.builtin_mul_widen_even (vectype)
4911 && targetm.vectorize.builtin_mul_widen_odd
4912 && targetm.vectorize.builtin_mul_widen_odd (vectype))
4914 if (vect_print_dump_info (REPORT_DETAILS))
4915 fprintf (vect_dump, "Unordered widening operation detected.");
4917 *code1 = *code2 = CALL_EXPR;
4918 *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype);
4919 *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype);
4925 case WIDEN_MULT_EXPR:
4926 if (BYTES_BIG_ENDIAN)
4928 c1 = VEC_WIDEN_MULT_HI_EXPR;
4929 c2 = VEC_WIDEN_MULT_LO_EXPR;
4933 c2 = VEC_WIDEN_MULT_HI_EXPR;
4934 c1 = VEC_WIDEN_MULT_LO_EXPR;
4939 if (BYTES_BIG_ENDIAN)
4941 c1 = VEC_UNPACK_HI_EXPR;
4942 c2 = VEC_UNPACK_LO_EXPR;
4946 c2 = VEC_UNPACK_HI_EXPR;
4947 c1 = VEC_UNPACK_LO_EXPR;
4952 if (BYTES_BIG_ENDIAN)
4954 c1 = VEC_UNPACK_FLOAT_HI_EXPR;
4955 c2 = VEC_UNPACK_FLOAT_LO_EXPR;
4959 c2 = VEC_UNPACK_FLOAT_HI_EXPR;
4960 c1 = VEC_UNPACK_FLOAT_LO_EXPR;
4964 case FIX_TRUNC_EXPR:
4965 /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/
4966 VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for
4967 computing the operation. */
4974 if (code == FIX_TRUNC_EXPR)
4976 /* The signedness is determined from output operand. */
4977 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
4978 optab2 = optab_for_tree_code (c2, vectype_out, optab_default);
4982 optab1 = optab_for_tree_code (c1, vectype, optab_default);
4983 optab2 = optab_for_tree_code (c2, vectype, optab_default);
4986 if (!optab1 || !optab2)
4989 vec_mode = TYPE_MODE (vectype);
4990 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing
4991 || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing)
4994 /* Check if it's a multi-step conversion that can be done using intermediate
4996 if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype)
4997 || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype))
5000 tree prev_type = vectype, intermediate_type;
5001 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5002 optab optab3, optab4;
5004 if (!CONVERT_EXPR_CODE_P (code))
5010 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5011 intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS
5012 to get to NARROW_VECTYPE, and fail if we do not. */
5013 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5014 for (i = 0; i < 3; i++)
5016 intermediate_mode = insn_data[icode1].operand[0].mode;
5017 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5018 TYPE_UNSIGNED (prev_type));
5019 optab3 = optab_for_tree_code (c1, intermediate_type, optab_default);
5020 optab4 = optab_for_tree_code (c2, intermediate_type, optab_default);
5022 if (!optab3 || !optab4
5023 || ((icode1 = optab_handler (optab1, prev_mode))
5024 == CODE_FOR_nothing)
5025 || insn_data[icode1].operand[0].mode != intermediate_mode
5026 || ((icode2 = optab_handler (optab2, prev_mode))
5027 == CODE_FOR_nothing)
5028 || insn_data[icode2].operand[0].mode != intermediate_mode
5029 || ((icode1 = optab_handler (optab3, intermediate_mode))
5030 == CODE_FOR_nothing)
5031 || ((icode2 = optab_handler (optab4, intermediate_mode))
5032 == CODE_FOR_nothing))
5035 VEC_quick_push (tree, *interm_types, intermediate_type);
5036 (*multi_step_cvt)++;
5038 if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype)
5039 && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype))
5042 prev_type = intermediate_type;
5043 prev_mode = intermediate_mode;
5055 /* Function supportable_narrowing_operation
5057 Check whether an operation represented by the code CODE is a
5058 narrowing operation that is supported by the target platform in
5059 vector form (i.e., when operating on arguments of type VECTYPE_IN
5060 and producing a result of type VECTYPE_OUT).
5062 Narrowing operations we currently support are NOP (CONVERT) and
5063 FIX_TRUNC. This function checks if these operations are supported by
5064 the target platform directly via vector tree-codes.
5067 - CODE1 is the code of a vector operation to be used when
5068 vectorizing the operation, if available.
5069 - MULTI_STEP_CVT determines the number of required intermediate steps in
5070 case of multi-step conversion (like int->short->char - in that case
5071 MULTI_STEP_CVT will be 1).
5072 - INTERM_TYPES contains the intermediate type required to perform the
5073 narrowing operation (short in the above example). */
5076 supportable_narrowing_operation (enum tree_code code,
5077 tree vectype_out, tree vectype_in,
5078 enum tree_code *code1, int *multi_step_cvt,
5079 VEC (tree, heap) **interm_types)
5081 enum machine_mode vec_mode;
5082 enum insn_code icode1;
5083 optab optab1, interm_optab;
5084 tree vectype = vectype_in;
5085 tree narrow_vectype = vectype_out;
5087 tree intermediate_type, prev_type;
5093 c1 = VEC_PACK_TRUNC_EXPR;
5096 case FIX_TRUNC_EXPR:
5097 c1 = VEC_PACK_FIX_TRUNC_EXPR;
5101 /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR
5102 tree code and optabs used for computing the operation. */
5109 if (code == FIX_TRUNC_EXPR)
5110 /* The signedness is determined from output operand. */
5111 optab1 = optab_for_tree_code (c1, vectype_out, optab_default);
5113 optab1 = optab_for_tree_code (c1, vectype, optab_default);
5118 vec_mode = TYPE_MODE (vectype);
5119 if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing)
5122 /* Check if it's a multi-step conversion that can be done using intermediate
5124 if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype))
5126 enum machine_mode intermediate_mode, prev_mode = vec_mode;
5129 prev_type = vectype;
5130 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
5131 intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS
5132 to get to NARROW_VECTYPE, and fail if we do not. */
5133 *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS);
5134 for (i = 0; i < 3; i++)
5136 intermediate_mode = insn_data[icode1].operand[0].mode;
5137 intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode,
5138 TYPE_UNSIGNED (prev_type));
5139 interm_optab = optab_for_tree_code (c1, intermediate_type,
5142 || ((icode1 = optab_handler (optab1, prev_mode))
5143 == CODE_FOR_nothing)
5144 || insn_data[icode1].operand[0].mode != intermediate_mode
5145 || ((icode1 = optab_handler (interm_optab, intermediate_mode))
5146 == CODE_FOR_nothing))
5149 VEC_quick_push (tree, *interm_types, intermediate_type);
5150 (*multi_step_cvt)++;
5152 if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype))
5155 prev_type = intermediate_type;
5156 prev_mode = intermediate_mode;