2 Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
22 /* Loop Vectorization Pass.
24 This pass tries to vectorize loops. This first implementation focuses on
25 simple inner-most loops, with no conditional control flow, and a set of
26 simple operations which vector form can be expressed using existing
27 tree codes (PLUS, MULT etc).
29 For example, the vectorizer transforms the following simple loop:
31 short a[N]; short b[N]; short c[N]; int i;
37 as if it was manually vectorized by rewriting the source code into:
39 typedef int __attribute__((mode(V8HI))) v8hi;
40 short a[N]; short b[N]; short c[N]; int i;
41 v8hi *pa = (v8hi*)a, *pb = (v8hi*)b, *pc = (v8hi*)c;
44 for (i=0; i<N/8; i++){
51 The main entry to this pass is vectorize_loops(), in which
52 the vectorizer applies a set of analyses on a given set of loops,
53 followed by the actual vectorization transformation for the loops that
54 had successfully passed the analysis phase.
56 Throughout this pass we make a distinction between two types of
57 data: scalars (which are represented by SSA_NAMES), and memory references
58 ("data-refs"). These two types of data require different handling both
59 during analysis and transformation. The types of data-refs that the
60 vectorizer currently supports are ARRAY_REFS which base is an array DECL
61 (not a pointer), and INDIRECT_REFS through pointers; both array and pointer
62 accesses are required to have a simple (consecutive) access pattern.
66 The driver for the analysis phase is vect_analyze_loop_nest().
67 It applies a set of analyses, some of which rely on the scalar evolution
68 analyzer (scev) developed by Sebastian Pop.
70 During the analysis phase the vectorizer records some information
71 per stmt in a "stmt_vec_info" struct which is attached to each stmt in the
72 loop, as well as general information about the loop as a whole, which is
73 recorded in a "loop_vec_info" struct attached to each loop.
77 The loop transformation phase scans all the stmts in the loop, and
78 creates a vector stmt (or a sequence of stmts) for each scalar stmt S in
79 the loop that needs to be vectorized. It insert the vector code sequence
80 just before the scalar stmt S, and records a pointer to the vector code
81 in STMT_VINFO_VEC_STMT (stmt_info) (stmt_info is the stmt_vec_info struct
82 attached to S). This pointer will be used for the vectorization of following
83 stmts which use the def of stmt S. Stmt S is removed if it writes to memory;
84 otherwise, we rely on dead code elimination for removing it.
86 For example, say stmt S1 was vectorized into stmt VS1:
89 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
92 To vectorize stmt S2, the vectorizer first finds the stmt that defines
93 the operand 'b' (S1), and gets the relevant vector def 'vb' from the
94 vector stmt VS1 pointed to by STMT_VINFO_VEC_STMT (stmt_info (S1)). The
95 resulting sequence would be:
98 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
100 S2: a = b; STMT_VINFO_VEC_STMT (stmt_info (S2)) = VS2
102 Operands that are not SSA_NAMEs, are data-refs that appear in
103 load/store operations (like 'x[i]' in S1), and are handled differently.
107 Currently the only target specific information that is used is the
108 size of the vector (in bytes) - "UNITS_PER_SIMD_WORD". Targets that can
109 support different sizes of vectors, for now will need to specify one value
110 for "UNITS_PER_SIMD_WORD". More flexibility will be added in the future.
112 Since we only vectorize operations which vector form can be
113 expressed using existing tree codes, to verify that an operation is
114 supported, the vectorizer checks the relevant optab at the relevant
115 machine_mode (e.g, add_optab->handlers[(int) V8HImode].insn_code). If
116 the value found is CODE_FOR_nothing, then there's no target support, and
117 we can't vectorize the stmt.
119 For additional information on this project see:
120 http://gcc.gnu.org/projects/tree-ssa/vectorization.html
125 #include "coretypes.h"
131 #include "basic-block.h"
132 #include "diagnostic.h"
133 #include "tree-flow.h"
134 #include "tree-dump.h"
137 #include "cfglayout.h"
141 #include "tree-chrec.h"
142 #include "tree-data-ref.h"
143 #include "tree-scalar-evolution.h"
145 #include "tree-vectorizer.h"
146 #include "tree-pass.h"
148 /*************************************************************************
149 Simple Loop Peeling Utilities
150 *************************************************************************/
151 static struct loop *slpeel_tree_duplicate_loop_to_edge_cfg
152 (struct loop *, struct loops *, edge);
153 static void slpeel_update_phis_for_duplicate_loop
154 (struct loop *, struct loop *, bool after);
155 static void slpeel_update_phi_nodes_for_guard1
156 (edge, struct loop *, bool, basic_block *, bitmap *);
157 static void slpeel_update_phi_nodes_for_guard2
158 (edge, struct loop *, bool, basic_block *);
159 static edge slpeel_add_loop_guard (basic_block, tree, basic_block, basic_block);
161 static void rename_use_op (use_operand_p);
162 static void rename_variables_in_bb (basic_block);
163 static void rename_variables_in_loop (struct loop *);
165 /*************************************************************************
166 General Vectorization Utilities
167 *************************************************************************/
168 static void vect_set_dump_settings (void);
170 /* vect_dump will be set to stderr or dump_file if exist. */
173 /* vect_verbosity_level set to an invalid value
174 to mark that it's uninitialized. */
175 enum verbosity_levels vect_verbosity_level = MAX_VERBOSITY_LEVEL;
177 /* Number of loops, at the beginning of vectorization. */
178 unsigned int vect_loops_num;
181 static LOC vect_loop_location;
183 /*************************************************************************
184 Simple Loop Peeling Utilities
186 Utilities to support loop peeling for vectorization purposes.
187 *************************************************************************/
190 /* Renames the use *OP_P. */
193 rename_use_op (use_operand_p op_p)
197 if (TREE_CODE (USE_FROM_PTR (op_p)) != SSA_NAME)
200 new_name = get_current_def (USE_FROM_PTR (op_p));
202 /* Something defined outside of the loop. */
206 /* An ordinary ssa name defined in the loop. */
208 SET_USE (op_p, new_name);
212 /* Renames the variables in basic block BB. */
215 rename_variables_in_bb (basic_block bb)
218 block_stmt_iterator bsi;
224 struct loop *loop = bb->loop_father;
226 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
228 stmt = bsi_stmt (bsi);
229 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter,
230 (SSA_OP_ALL_USES | SSA_OP_ALL_KILLS))
231 rename_use_op (use_p);
234 FOR_EACH_EDGE (e, ei, bb->succs)
236 if (!flow_bb_inside_loop_p (loop, e->dest))
238 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
239 rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e));
244 /* Renames variables in new generated LOOP. */
247 rename_variables_in_loop (struct loop *loop)
252 bbs = get_loop_body (loop);
254 for (i = 0; i < loop->num_nodes; i++)
255 rename_variables_in_bb (bbs[i]);
261 /* Update the PHI nodes of NEW_LOOP.
263 NEW_LOOP is a duplicate of ORIG_LOOP.
264 AFTER indicates whether NEW_LOOP executes before or after ORIG_LOOP:
265 AFTER is true if NEW_LOOP executes after ORIG_LOOP, and false if it
266 executes before it. */
269 slpeel_update_phis_for_duplicate_loop (struct loop *orig_loop,
270 struct loop *new_loop, bool after)
273 tree phi_new, phi_orig;
275 edge orig_loop_latch = loop_latch_edge (orig_loop);
276 edge orig_entry_e = loop_preheader_edge (orig_loop);
277 edge new_loop_exit_e = new_loop->single_exit;
278 edge new_loop_entry_e = loop_preheader_edge (new_loop);
279 edge entry_arg_e = (after ? orig_loop_latch : orig_entry_e);
282 step 1. For each loop-header-phi:
283 Add the first phi argument for the phi in NEW_LOOP
284 (the one associated with the entry of NEW_LOOP)
286 step 2. For each loop-header-phi:
287 Add the second phi argument for the phi in NEW_LOOP
288 (the one associated with the latch of NEW_LOOP)
290 step 3. Update the phis in the successor block of NEW_LOOP.
292 case 1: NEW_LOOP was placed before ORIG_LOOP:
293 The successor block of NEW_LOOP is the header of ORIG_LOOP.
294 Updating the phis in the successor block can therefore be done
295 along with the scanning of the loop header phis, because the
296 header blocks of ORIG_LOOP and NEW_LOOP have exactly the same
297 phi nodes, organized in the same order.
299 case 2: NEW_LOOP was placed after ORIG_LOOP:
300 The successor block of NEW_LOOP is the original exit block of
301 ORIG_LOOP - the phis to be updated are the loop-closed-ssa phis.
302 We postpone updating these phis to a later stage (when
303 loop guards are added).
307 /* Scan the phis in the headers of the old and new loops
308 (they are organized in exactly the same order). */
310 for (phi_new = phi_nodes (new_loop->header),
311 phi_orig = phi_nodes (orig_loop->header);
313 phi_new = PHI_CHAIN (phi_new), phi_orig = PHI_CHAIN (phi_orig))
316 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, entry_arg_e);
317 add_phi_arg (phi_new, def, new_loop_entry_e);
320 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_loop_latch);
321 if (TREE_CODE (def) != SSA_NAME)
324 new_ssa_name = get_current_def (def);
327 /* This only happens if there are no definitions
328 inside the loop. use the phi_result in this case. */
329 new_ssa_name = PHI_RESULT (phi_new);
332 /* An ordinary ssa name defined in the loop. */
333 add_phi_arg (phi_new, new_ssa_name, loop_latch_edge (new_loop));
335 /* step 3 (case 1). */
338 gcc_assert (new_loop_exit_e == orig_entry_e);
339 SET_PHI_ARG_DEF (phi_orig,
340 new_loop_exit_e->dest_idx,
347 /* Update PHI nodes for a guard of the LOOP.
350 - LOOP, GUARD_EDGE: LOOP is a loop for which we added guard code that
351 controls whether LOOP is to be executed. GUARD_EDGE is the edge that
352 originates from the guard-bb, skips LOOP and reaches the (unique) exit
353 bb of LOOP. This loop-exit-bb is an empty bb with one successor.
354 We denote this bb NEW_MERGE_BB because before the guard code was added
355 it had a single predecessor (the LOOP header), and now it became a merge
356 point of two paths - the path that ends with the LOOP exit-edge, and
357 the path that ends with GUARD_EDGE.
358 - NEW_EXIT_BB: New basic block that is added by this function between LOOP
359 and NEW_MERGE_BB. It is used to place loop-closed-ssa-form exit-phis.
361 ===> The CFG before the guard-code was added:
364 if (exit_loop) goto update_bb
365 else goto LOOP_header_bb
368 ==> The CFG after the guard-code was added:
370 if (LOOP_guard_condition) goto new_merge_bb
371 else goto LOOP_header_bb
374 if (exit_loop_condition) goto new_merge_bb
375 else goto LOOP_header_bb
380 ==> The CFG after this function:
382 if (LOOP_guard_condition) goto new_merge_bb
383 else goto LOOP_header_bb
386 if (exit_loop_condition) goto new_exit_bb
387 else goto LOOP_header_bb
394 1. creates and updates the relevant phi nodes to account for the new
395 incoming edge (GUARD_EDGE) into NEW_MERGE_BB. This involves:
396 1.1. Create phi nodes at NEW_MERGE_BB.
397 1.2. Update the phi nodes at the successor of NEW_MERGE_BB (denoted
398 UPDATE_BB). UPDATE_BB was the exit-bb of LOOP before NEW_MERGE_BB
399 2. preserves loop-closed-ssa-form by creating the required phi nodes
400 at the exit of LOOP (i.e, in NEW_EXIT_BB).
402 There are two flavors to this function:
404 slpeel_update_phi_nodes_for_guard1:
405 Here the guard controls whether we enter or skip LOOP, where LOOP is a
406 prolog_loop (loop1 below), and the new phis created in NEW_MERGE_BB are
407 for variables that have phis in the loop header.
409 slpeel_update_phi_nodes_for_guard2:
410 Here the guard controls whether we enter or skip LOOP, where LOOP is an
411 epilog_loop (loop2 below), and the new phis created in NEW_MERGE_BB are
412 for variables that have phis in the loop exit.
414 I.E., the overall structure is:
417 guard1 (goto loop1/merg1_bb)
420 guard2 (goto merge1_bb/merge2_bb)
427 slpeel_update_phi_nodes_for_guard1 takes care of creating phis in
428 loop1_exit_bb and merge1_bb. These are entry phis (phis for the vars
429 that have phis in loop1->header).
431 slpeel_update_phi_nodes_for_guard2 takes care of creating phis in
432 loop2_exit_bb and merge2_bb. These are exit phis (phis for the vars
433 that have phis in next_bb). It also adds some of these phis to
436 slpeel_update_phi_nodes_for_guard1 is always called before
437 slpeel_update_phi_nodes_for_guard2. They are both needed in order
438 to create correct data-flow and loop-closed-ssa-form.
440 Generally slpeel_update_phi_nodes_for_guard1 creates phis for variables
441 that change between iterations of a loop (and therefore have a phi-node
442 at the loop entry), whereas slpeel_update_phi_nodes_for_guard2 creates
443 phis for variables that are used out of the loop (and therefore have
444 loop-closed exit phis). Some variables may be both updated between
445 iterations and used after the loop. This is why in loop1_exit_bb we
446 may need both entry_phis (created by slpeel_update_phi_nodes_for_guard1)
447 and exit phis (created by slpeel_update_phi_nodes_for_guard2).
449 - IS_NEW_LOOP: if IS_NEW_LOOP is true, then LOOP is a newly created copy of
450 an original loop. i.e., we have:
453 guard_bb (goto LOOP/new_merge)
459 If IS_NEW_LOOP is false, then LOOP is an original loop, in which case we
463 guard_bb (goto LOOP/new_merge)
469 The SSA names defined in the original loop have a current
470 reaching definition that that records the corresponding new
471 ssa-name used in the new duplicated loop copy.
474 /* Function slpeel_update_phi_nodes_for_guard1
477 - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above.
478 - DEFS - a bitmap of ssa names to mark new names for which we recorded
481 In the context of the overall structure, we have:
484 guard1 (goto loop1/merg1_bb)
487 guard2 (goto merge1_bb/merge2_bb)
494 For each name updated between loop iterations (i.e - for each name that has
495 an entry (loop-header) phi in LOOP) we create a new phi in:
496 1. merge1_bb (to account for the edge from guard1)
497 2. loop1_exit_bb (an exit-phi to keep LOOP in loop-closed form)
501 slpeel_update_phi_nodes_for_guard1 (edge guard_edge, struct loop *loop,
502 bool is_new_loop, basic_block *new_exit_bb,
505 tree orig_phi, new_phi;
506 tree update_phi, update_phi2;
507 tree guard_arg, loop_arg;
508 basic_block new_merge_bb = guard_edge->dest;
509 edge e = EDGE_SUCC (new_merge_bb, 0);
510 basic_block update_bb = e->dest;
511 basic_block orig_bb = loop->header;
513 tree current_new_name;
515 /* Create new bb between loop and new_merge_bb. */
516 *new_exit_bb = split_edge (loop->single_exit);
517 add_bb_to_loop (*new_exit_bb, loop->outer);
519 new_exit_e = EDGE_SUCC (*new_exit_bb, 0);
521 for (orig_phi = phi_nodes (orig_bb), update_phi = phi_nodes (update_bb);
522 orig_phi && update_phi;
523 orig_phi = PHI_CHAIN (orig_phi), update_phi = PHI_CHAIN (update_phi))
525 /** 1. Handle new-merge-point phis **/
527 /* 1.1. Generate new phi node in NEW_MERGE_BB: */
528 new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)),
531 /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge
532 of LOOP. Set the two phi args in NEW_PHI for these edges: */
533 loop_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, EDGE_SUCC (loop->latch, 0));
534 guard_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, loop_preheader_edge (loop));
536 add_phi_arg (new_phi, loop_arg, new_exit_e);
537 add_phi_arg (new_phi, guard_arg, guard_edge);
539 /* 1.3. Update phi in successor block. */
540 gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == loop_arg
541 || PHI_ARG_DEF_FROM_EDGE (update_phi, e) == guard_arg);
542 SET_PHI_ARG_DEF (update_phi, e->dest_idx, PHI_RESULT (new_phi));
543 update_phi2 = new_phi;
546 /** 2. Handle loop-closed-ssa-form phis **/
548 /* 2.1. Generate new phi node in NEW_EXIT_BB: */
549 new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)),
552 /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */
553 add_phi_arg (new_phi, loop_arg, loop->single_exit);
555 /* 2.3. Update phi in successor of NEW_EXIT_BB: */
556 gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg);
557 SET_PHI_ARG_DEF (update_phi2, new_exit_e->dest_idx, PHI_RESULT (new_phi));
559 /* 2.4. Record the newly created name with set_current_def.
560 We want to find a name such that
561 name = get_current_def (orig_loop_name)
562 and to set its current definition as follows:
563 set_current_def (name, new_phi_name)
565 If LOOP is a new loop then loop_arg is already the name we're
566 looking for. If LOOP is the original loop, then loop_arg is
567 the orig_loop_name and the relevant name is recorded in its
568 current reaching definition. */
570 current_new_name = loop_arg;
573 current_new_name = get_current_def (loop_arg);
574 /* current_def is not available only if the variable does not
575 change inside the loop, in which case we also don't care
576 about recording a current_def for it because we won't be
577 trying to create loop-exit-phis for it. */
578 if (!current_new_name)
581 gcc_assert (get_current_def (current_new_name) == NULL_TREE);
583 set_current_def (current_new_name, PHI_RESULT (new_phi));
584 bitmap_set_bit (*defs, SSA_NAME_VERSION (current_new_name));
587 set_phi_nodes (new_merge_bb, phi_reverse (phi_nodes (new_merge_bb)));
591 /* Function slpeel_update_phi_nodes_for_guard2
594 - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above.
596 In the context of the overall structure, we have:
599 guard1 (goto loop1/merg1_bb)
602 guard2 (goto merge1_bb/merge2_bb)
609 For each name used out side the loop (i.e - for each name that has an exit
610 phi in next_bb) we create a new phi in:
611 1. merge2_bb (to account for the edge from guard_bb)
612 2. loop2_exit_bb (an exit-phi to keep LOOP in loop-closed form)
613 3. guard2 bb (an exit phi to keep the preceding loop in loop-closed form),
614 if needed (if it wasn't handled by slpeel_update_phis_nodes_for_phi1).
618 slpeel_update_phi_nodes_for_guard2 (edge guard_edge, struct loop *loop,
619 bool is_new_loop, basic_block *new_exit_bb)
621 tree orig_phi, new_phi;
622 tree update_phi, update_phi2;
623 tree guard_arg, loop_arg;
624 basic_block new_merge_bb = guard_edge->dest;
625 edge e = EDGE_SUCC (new_merge_bb, 0);
626 basic_block update_bb = e->dest;
628 tree orig_def, orig_def_new_name;
629 tree new_name, new_name2;
632 /* Create new bb between loop and new_merge_bb. */
633 *new_exit_bb = split_edge (loop->single_exit);
634 add_bb_to_loop (*new_exit_bb, loop->outer);
636 new_exit_e = EDGE_SUCC (*new_exit_bb, 0);
638 for (update_phi = phi_nodes (update_bb); update_phi;
639 update_phi = PHI_CHAIN (update_phi))
641 orig_phi = update_phi;
642 orig_def = PHI_ARG_DEF_FROM_EDGE (orig_phi, e);
643 /* This loop-closed-phi actually doesn't represent a use
644 out of the loop - the phi arg is a constant. */
645 if (TREE_CODE (orig_def) != SSA_NAME)
647 orig_def_new_name = get_current_def (orig_def);
650 /** 1. Handle new-merge-point phis **/
652 /* 1.1. Generate new phi node in NEW_MERGE_BB: */
653 new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)),
656 /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge
657 of LOOP. Set the two PHI args in NEW_PHI for these edges: */
659 new_name2 = NULL_TREE;
660 if (orig_def_new_name)
662 new_name = orig_def_new_name;
663 /* Some variables have both loop-entry-phis and loop-exit-phis.
664 Such variables were given yet newer names by phis placed in
665 guard_bb by slpeel_update_phi_nodes_for_guard1. I.e:
666 new_name2 = get_current_def (get_current_def (orig_name)). */
667 new_name2 = get_current_def (new_name);
672 guard_arg = orig_def;
677 guard_arg = new_name;
681 guard_arg = new_name2;
683 add_phi_arg (new_phi, loop_arg, new_exit_e);
684 add_phi_arg (new_phi, guard_arg, guard_edge);
686 /* 1.3. Update phi in successor block. */
687 gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == orig_def);
688 SET_PHI_ARG_DEF (update_phi, e->dest_idx, PHI_RESULT (new_phi));
689 update_phi2 = new_phi;
692 /** 2. Handle loop-closed-ssa-form phis **/
694 /* 2.1. Generate new phi node in NEW_EXIT_BB: */
695 new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)),
698 /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */
699 add_phi_arg (new_phi, loop_arg, loop->single_exit);
701 /* 2.3. Update phi in successor of NEW_EXIT_BB: */
702 gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg);
703 SET_PHI_ARG_DEF (update_phi2, new_exit_e->dest_idx, PHI_RESULT (new_phi));
706 /** 3. Handle loop-closed-ssa-form phis for first loop **/
708 /* 3.1. Find the relevant names that need an exit-phi in
709 GUARD_BB, i.e. names for which
710 slpeel_update_phi_nodes_for_guard1 had not already created a
711 phi node. This is the case for names that are used outside
712 the loop (and therefore need an exit phi) but are not updated
713 across loop iterations (and therefore don't have a
716 slpeel_update_phi_nodes_for_guard1 is responsible for
717 creating loop-exit phis in GUARD_BB for names that have a
718 loop-header-phi. When such a phi is created we also record
719 the new name in its current definition. If this new name
720 exists, then guard_arg was set to this new name (see 1.2
721 above). Therefore, if guard_arg is not this new name, this
722 is an indication that an exit-phi in GUARD_BB was not yet
723 created, so we take care of it here. */
724 if (guard_arg == new_name2)
728 /* 3.2. Generate new phi node in GUARD_BB: */
729 new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)),
732 /* 3.3. GUARD_BB has one incoming edge: */
733 gcc_assert (EDGE_COUNT (guard_edge->src->preds) == 1);
734 add_phi_arg (new_phi, arg, EDGE_PRED (guard_edge->src, 0));
736 /* 3.4. Update phi in successor of GUARD_BB: */
737 gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, guard_edge)
739 SET_PHI_ARG_DEF (update_phi2, guard_edge->dest_idx, PHI_RESULT (new_phi));
742 set_phi_nodes (new_merge_bb, phi_reverse (phi_nodes (new_merge_bb)));
746 /* Make the LOOP iterate NITERS times. This is done by adding a new IV
747 that starts at zero, increases by one and its limit is NITERS.
749 Assumption: the exit-condition of LOOP is the last stmt in the loop. */
752 slpeel_make_loop_iterate_ntimes (struct loop *loop, tree niters)
754 tree indx_before_incr, indx_after_incr, cond_stmt, cond;
756 edge exit_edge = loop->single_exit;
757 block_stmt_iterator loop_cond_bsi;
758 block_stmt_iterator incr_bsi;
760 tree begin_label = tree_block_label (loop->latch);
761 tree exit_label = tree_block_label (loop->single_exit->dest);
762 tree init = build_int_cst (TREE_TYPE (niters), 0);
763 tree step = build_int_cst (TREE_TYPE (niters), 1);
768 orig_cond = get_loop_exit_condition (loop);
769 gcc_assert (orig_cond);
770 loop_cond_bsi = bsi_for_stmt (orig_cond);
772 standard_iv_increment_position (loop, &incr_bsi, &insert_after);
773 create_iv (init, step, NULL_TREE, loop,
774 &incr_bsi, insert_after, &indx_before_incr, &indx_after_incr);
776 if (exit_edge->flags & EDGE_TRUE_VALUE) /* 'then' edge exits the loop. */
778 cond = build2 (GE_EXPR, boolean_type_node, indx_after_incr, niters);
779 then_label = build1 (GOTO_EXPR, void_type_node, exit_label);
780 else_label = build1 (GOTO_EXPR, void_type_node, begin_label);
782 else /* 'then' edge loops back. */
784 cond = build2 (LT_EXPR, boolean_type_node, indx_after_incr, niters);
785 then_label = build1 (GOTO_EXPR, void_type_node, begin_label);
786 else_label = build1 (GOTO_EXPR, void_type_node, exit_label);
789 cond_stmt = build3 (COND_EXPR, TREE_TYPE (orig_cond), cond,
790 then_label, else_label);
791 bsi_insert_before (&loop_cond_bsi, cond_stmt, BSI_SAME_STMT);
793 /* Remove old loop exit test: */
794 bsi_remove (&loop_cond_bsi);
796 loop_loc = find_loop_location (loop);
797 if (dump_file && (dump_flags & TDF_DETAILS))
799 if (loop_loc != UNKNOWN_LOC)
800 fprintf (dump_file, "\nloop at %s:%d: ",
801 LOC_FILE (loop_loc), LOC_LINE (loop_loc));
802 print_generic_expr (dump_file, cond_stmt, TDF_SLIM);
805 loop->nb_iterations = niters;
809 /* Given LOOP this function generates a new copy of it and puts it
810 on E which is either the entry or exit of LOOP. */
813 slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *loop, struct loops *loops,
816 struct loop *new_loop;
817 basic_block *new_bbs, *bbs;
820 basic_block exit_dest;
823 at_exit = (e == loop->single_exit);
824 if (!at_exit && e != loop_preheader_edge (loop))
827 bbs = get_loop_body (loop);
829 /* Check whether duplication is possible. */
830 if (!can_copy_bbs_p (bbs, loop->num_nodes))
836 /* Generate new loop structure. */
837 new_loop = duplicate_loop (loops, loop, loop->outer);
844 exit_dest = loop->single_exit->dest;
845 was_imm_dom = (get_immediate_dominator (CDI_DOMINATORS,
846 exit_dest) == loop->header ?
849 new_bbs = xmalloc (sizeof (basic_block) * loop->num_nodes);
851 copy_bbs (bbs, loop->num_nodes, new_bbs,
852 &loop->single_exit, 1, &new_loop->single_exit, NULL);
854 /* Duplicating phi args at exit bbs as coming
855 also from exit of duplicated loop. */
856 for (phi = phi_nodes (exit_dest); phi; phi = PHI_CHAIN (phi))
858 phi_arg = PHI_ARG_DEF_FROM_EDGE (phi, loop->single_exit);
861 edge new_loop_exit_edge;
863 if (EDGE_SUCC (new_loop->header, 0)->dest == new_loop->latch)
864 new_loop_exit_edge = EDGE_SUCC (new_loop->header, 1);
866 new_loop_exit_edge = EDGE_SUCC (new_loop->header, 0);
868 add_phi_arg (phi, phi_arg, new_loop_exit_edge);
872 if (at_exit) /* Add the loop copy at exit. */
874 redirect_edge_and_branch_force (e, new_loop->header);
875 set_immediate_dominator (CDI_DOMINATORS, new_loop->header, e->src);
877 set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_loop->header);
879 else /* Add the copy at entry. */
882 edge entry_e = loop_preheader_edge (loop);
883 basic_block preheader = entry_e->src;
885 if (!flow_bb_inside_loop_p (new_loop,
886 EDGE_SUCC (new_loop->header, 0)->dest))
887 new_exit_e = EDGE_SUCC (new_loop->header, 0);
889 new_exit_e = EDGE_SUCC (new_loop->header, 1);
891 redirect_edge_and_branch_force (new_exit_e, loop->header);
892 set_immediate_dominator (CDI_DOMINATORS, loop->header,
895 /* We have to add phi args to the loop->header here as coming
896 from new_exit_e edge. */
897 for (phi = phi_nodes (loop->header); phi; phi = PHI_CHAIN (phi))
899 phi_arg = PHI_ARG_DEF_FROM_EDGE (phi, entry_e);
901 add_phi_arg (phi, phi_arg, new_exit_e);
904 redirect_edge_and_branch_force (entry_e, new_loop->header);
905 set_immediate_dominator (CDI_DOMINATORS, new_loop->header, preheader);
915 /* Given the condition statement COND, put it as the last statement
916 of GUARD_BB; EXIT_BB is the basic block to skip the loop;
917 Assumes that this is the single exit of the guarded loop.
918 Returns the skip edge. */
921 slpeel_add_loop_guard (basic_block guard_bb, tree cond, basic_block exit_bb,
924 block_stmt_iterator bsi;
926 tree cond_stmt, then_label, else_label;
928 enter_e = EDGE_SUCC (guard_bb, 0);
929 enter_e->flags &= ~EDGE_FALLTHRU;
930 enter_e->flags |= EDGE_FALSE_VALUE;
931 bsi = bsi_last (guard_bb);
933 then_label = build1 (GOTO_EXPR, void_type_node,
934 tree_block_label (exit_bb));
935 else_label = build1 (GOTO_EXPR, void_type_node,
936 tree_block_label (enter_e->dest));
937 cond_stmt = build3 (COND_EXPR, void_type_node, cond,
938 then_label, else_label);
939 bsi_insert_after (&bsi, cond_stmt, BSI_NEW_STMT);
940 /* Add new edge to connect guard block to the merge/loop-exit block. */
941 new_e = make_edge (guard_bb, exit_bb, EDGE_TRUE_VALUE);
942 set_immediate_dominator (CDI_DOMINATORS, exit_bb, dom_bb);
947 /* This function verifies that the following restrictions apply to LOOP:
949 (2) it consists of exactly 2 basic blocks - header, and an empty latch.
950 (3) it is single entry, single exit
951 (4) its exit condition is the last stmt in the header
952 (5) E is the entry/exit edge of LOOP.
956 slpeel_can_duplicate_loop_p (struct loop *loop, edge e)
958 edge exit_e = loop->single_exit;
959 edge entry_e = loop_preheader_edge (loop);
960 tree orig_cond = get_loop_exit_condition (loop);
961 block_stmt_iterator loop_exit_bsi = bsi_last (exit_e->src);
963 if (need_ssa_update_p ())
967 /* All loops have an outer scope; the only case loop->outer is NULL is for
968 the function itself. */
970 || loop->num_nodes != 2
971 || !empty_block_p (loop->latch)
972 || !loop->single_exit
973 /* Verify that new loop exit condition can be trivially modified. */
974 || (!orig_cond || orig_cond != bsi_stmt (loop_exit_bsi))
975 || (e != exit_e && e != entry_e))
981 #ifdef ENABLE_CHECKING
983 slpeel_verify_cfg_after_peeling (struct loop *first_loop,
984 struct loop *second_loop)
986 basic_block loop1_exit_bb = first_loop->single_exit->dest;
987 basic_block loop2_entry_bb = loop_preheader_edge (second_loop)->src;
988 basic_block loop1_entry_bb = loop_preheader_edge (first_loop)->src;
990 /* A guard that controls whether the second_loop is to be executed or skipped
991 is placed in first_loop->exit. first_loopt->exit therefore has two
992 successors - one is the preheader of second_loop, and the other is a bb
995 gcc_assert (EDGE_COUNT (loop1_exit_bb->succs) == 2);
997 /* 1. Verify that one of the successors of first_loopt->exit is the preheader
1000 /* The preheader of new_loop is expected to have two predecessors:
1001 first_loop->exit and the block that precedes first_loop. */
1003 gcc_assert (EDGE_COUNT (loop2_entry_bb->preds) == 2
1004 && ((EDGE_PRED (loop2_entry_bb, 0)->src == loop1_exit_bb
1005 && EDGE_PRED (loop2_entry_bb, 1)->src == loop1_entry_bb)
1006 || (EDGE_PRED (loop2_entry_bb, 1)->src == loop1_exit_bb
1007 && EDGE_PRED (loop2_entry_bb, 0)->src == loop1_entry_bb)));
1009 /* Verify that the other successor of first_loopt->exit is after the
1015 /* Function slpeel_tree_peel_loop_to_edge.
1017 Peel the first (last) iterations of LOOP into a new prolog (epilog) loop
1018 that is placed on the entry (exit) edge E of LOOP. After this transformation
1019 we have two loops one after the other - first-loop iterates FIRST_NITERS
1020 times, and second-loop iterates the remainder NITERS - FIRST_NITERS times.
1023 - LOOP: the loop to be peeled.
1024 - E: the exit or entry edge of LOOP.
1025 If it is the entry edge, we peel the first iterations of LOOP. In this
1026 case first-loop is LOOP, and second-loop is the newly created loop.
1027 If it is the exit edge, we peel the last iterations of LOOP. In this
1028 case, first-loop is the newly created loop, and second-loop is LOOP.
1029 - NITERS: the number of iterations that LOOP iterates.
1030 - FIRST_NITERS: the number of iterations that the first-loop should iterate.
1031 - UPDATE_FIRST_LOOP_COUNT: specified whether this function is responsible
1032 for updating the loop bound of the first-loop to FIRST_NITERS. If it
1033 is false, the caller of this function may want to take care of this
1034 (this can be useful if we don't want new stmts added to first-loop).
1037 The function returns a pointer to the new loop-copy, or NULL if it failed
1038 to perform the transformation.
1040 The function generates two if-then-else guards: one before the first loop,
1041 and the other before the second loop:
1043 if (FIRST_NITERS == 0) then skip the first loop,
1044 and go directly to the second loop.
1045 The second guard is:
1046 if (FIRST_NITERS == NITERS) then skip the second loop.
1048 FORNOW only simple loops are supported (see slpeel_can_duplicate_loop_p).
1049 FORNOW the resulting code will not be in loop-closed-ssa form.
1053 slpeel_tree_peel_loop_to_edge (struct loop *loop, struct loops *loops,
1054 edge e, tree first_niters,
1055 tree niters, bool update_first_loop_count)
1057 struct loop *new_loop = NULL, *first_loop, *second_loop;
1061 basic_block bb_before_second_loop, bb_after_second_loop;
1062 basic_block bb_before_first_loop;
1063 basic_block bb_between_loops;
1064 basic_block new_exit_bb;
1065 edge exit_e = loop->single_exit;
1068 if (!slpeel_can_duplicate_loop_p (loop, e))
1071 /* We have to initialize cfg_hooks. Then, when calling
1072 cfg_hooks->split_edge, the function tree_split_edge
1073 is actually called and, when calling cfg_hooks->duplicate_block,
1074 the function tree_duplicate_bb is called. */
1075 tree_register_cfg_hooks ();
1078 /* 1. Generate a copy of LOOP and put it on E (E is the entry/exit of LOOP).
1079 Resulting CFG would be:
1092 if (!(new_loop = slpeel_tree_duplicate_loop_to_edge_cfg (loop, loops, e)))
1094 loop_loc = find_loop_location (loop);
1095 if (dump_file && (dump_flags & TDF_DETAILS))
1097 if (loop_loc != UNKNOWN_LOC)
1098 fprintf (dump_file, "\n%s:%d: note: ",
1099 LOC_FILE (loop_loc), LOC_LINE (loop_loc));
1100 fprintf (dump_file, "tree_duplicate_loop_to_edge_cfg failed.\n");
1107 /* NEW_LOOP was placed after LOOP. */
1109 second_loop = new_loop;
1113 /* NEW_LOOP was placed before LOOP. */
1114 first_loop = new_loop;
1118 definitions = ssa_names_to_replace ();
1119 slpeel_update_phis_for_duplicate_loop (loop, new_loop, e == exit_e);
1120 rename_variables_in_loop (new_loop);
1123 /* 2. Add the guard that controls whether the first loop is executed.
1124 Resulting CFG would be:
1126 bb_before_first_loop:
1127 if (FIRST_NITERS == 0) GOTO bb_before_second_loop
1134 bb_before_second_loop:
1143 bb_before_first_loop = split_edge (loop_preheader_edge (first_loop));
1144 add_bb_to_loop (bb_before_first_loop, first_loop->outer);
1145 bb_before_second_loop = split_edge (first_loop->single_exit);
1146 add_bb_to_loop (bb_before_second_loop, first_loop->outer);
1149 fold_build2 (LE_EXPR, boolean_type_node, first_niters,
1150 build_int_cst (TREE_TYPE (first_niters), 0));
1151 skip_e = slpeel_add_loop_guard (bb_before_first_loop, pre_condition,
1152 bb_before_second_loop, bb_before_first_loop);
1153 slpeel_update_phi_nodes_for_guard1 (skip_e, first_loop,
1154 first_loop == new_loop,
1155 &new_exit_bb, &definitions);
1158 /* 3. Add the guard that controls whether the second loop is executed.
1159 Resulting CFG would be:
1161 bb_before_first_loop:
1162 if (FIRST_NITERS == 0) GOTO bb_before_second_loop (skip first loop)
1170 if (FIRST_NITERS == NITERS) GOTO bb_after_second_loop (skip second loop)
1171 GOTO bb_before_second_loop
1173 bb_before_second_loop:
1179 bb_after_second_loop:
1184 bb_between_loops = new_exit_bb;
1185 bb_after_second_loop = split_edge (second_loop->single_exit);
1186 add_bb_to_loop (bb_after_second_loop, second_loop->outer);
1189 fold_build2 (EQ_EXPR, boolean_type_node, first_niters, niters);
1190 skip_e = slpeel_add_loop_guard (bb_between_loops, pre_condition,
1191 bb_after_second_loop, bb_before_first_loop);
1192 slpeel_update_phi_nodes_for_guard2 (skip_e, second_loop,
1193 second_loop == new_loop, &new_exit_bb);
1195 /* 4. Make first-loop iterate FIRST_NITERS times, if requested.
1197 if (update_first_loop_count)
1198 slpeel_make_loop_iterate_ntimes (first_loop, first_niters);
1200 BITMAP_FREE (definitions);
1201 delete_update_ssa ();
1206 /* Function vect_get_loop_location.
1208 Extract the location of the loop in the source code.
1209 If the loop is not well formed for vectorization, an estimated
1210 location is calculated.
1211 Return the loop location if succeed and NULL if not. */
1214 find_loop_location (struct loop *loop)
1216 tree node = NULL_TREE;
1218 block_stmt_iterator si;
1223 node = get_loop_exit_condition (loop);
1225 if (node && EXPR_P (node) && EXPR_HAS_LOCATION (node)
1226 && EXPR_FILENAME (node) && EXPR_LINENO (node))
1227 return EXPR_LOC (node);
1229 /* If we got here the loop is probably not "well formed",
1230 try to estimate the loop location */
1237 for (si = bsi_start (bb); !bsi_end_p (si); bsi_next (&si))
1239 node = bsi_stmt (si);
1240 if (node && EXPR_P (node) && EXPR_HAS_LOCATION (node))
1241 return EXPR_LOC (node);
1248 /*************************************************************************
1249 Vectorization Debug Information.
1250 *************************************************************************/
1252 /* Function vect_set_verbosity_level.
1254 Called from toplev.c upon detection of the
1255 -ftree-vectorizer-verbose=N option. */
1258 vect_set_verbosity_level (const char *val)
1263 if (vl < MAX_VERBOSITY_LEVEL)
1264 vect_verbosity_level = vl;
1266 vect_verbosity_level = MAX_VERBOSITY_LEVEL - 1;
1270 /* Function vect_set_dump_settings.
1272 Fix the verbosity level of the vectorizer if the
1273 requested level was not set explicitly using the flag
1274 -ftree-vectorizer-verbose=N.
1275 Decide where to print the debugging information (dump_file/stderr).
1276 If the user defined the verbosity level, but there is no dump file,
1277 print to stderr, otherwise print to the dump file. */
1280 vect_set_dump_settings (void)
1282 vect_dump = dump_file;
1284 /* Check if the verbosity level was defined by the user: */
1285 if (vect_verbosity_level != MAX_VERBOSITY_LEVEL)
1287 /* If there is no dump file, print to stderr. */
1293 /* User didn't specify verbosity level: */
1294 if (dump_file && (dump_flags & TDF_DETAILS))
1295 vect_verbosity_level = REPORT_DETAILS;
1296 else if (dump_file && (dump_flags & TDF_STATS))
1297 vect_verbosity_level = REPORT_UNVECTORIZED_LOOPS;
1299 vect_verbosity_level = REPORT_NONE;
1301 gcc_assert (dump_file || vect_verbosity_level == REPORT_NONE);
1305 /* Function debug_loop_details.
1307 For vectorization debug dumps. */
1310 vect_print_dump_info (enum verbosity_levels vl)
1312 if (vl > vect_verbosity_level)
1315 if (vect_loop_location == UNKNOWN_LOC)
1316 fprintf (vect_dump, "\n%s:%d: note: ",
1317 DECL_SOURCE_FILE (current_function_decl),
1318 DECL_SOURCE_LINE (current_function_decl));
1320 fprintf (vect_dump, "\n%s:%d: note: ",
1321 LOC_FILE (vect_loop_location), LOC_LINE (vect_loop_location));
1328 /*************************************************************************
1329 Vectorization Utilities.
1330 *************************************************************************/
1332 /* Function new_stmt_vec_info.
1334 Create and initialize a new stmt_vec_info struct for STMT. */
1337 new_stmt_vec_info (tree stmt, loop_vec_info loop_vinfo)
1340 res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info));
1342 STMT_VINFO_TYPE (res) = undef_vec_info_type;
1343 STMT_VINFO_STMT (res) = stmt;
1344 STMT_VINFO_LOOP_VINFO (res) = loop_vinfo;
1345 STMT_VINFO_RELEVANT_P (res) = 0;
1346 STMT_VINFO_LIVE_P (res) = 0;
1347 STMT_VINFO_VECTYPE (res) = NULL;
1348 STMT_VINFO_VEC_STMT (res) = NULL;
1349 STMT_VINFO_DATA_REF (res) = NULL;
1350 if (TREE_CODE (stmt) == PHI_NODE)
1351 STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type;
1353 STMT_VINFO_DEF_TYPE (res) = vect_loop_def;
1354 STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5);
1360 /* Function new_loop_vec_info.
1362 Create and initialize a new loop_vec_info struct for LOOP, as well as
1363 stmt_vec_info structs for all the stmts in LOOP. */
1366 new_loop_vec_info (struct loop *loop)
1370 block_stmt_iterator si;
1373 res = (loop_vec_info) xcalloc (1, sizeof (struct _loop_vec_info));
1375 bbs = get_loop_body (loop);
1377 /* Create stmt_info for all stmts in the loop. */
1378 for (i = 0; i < loop->num_nodes; i++)
1380 basic_block bb = bbs[i];
1383 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1385 tree_ann_t ann = get_tree_ann (phi);
1386 set_stmt_info (ann, new_stmt_vec_info (phi, res));
1389 for (si = bsi_start (bb); !bsi_end_p (si); bsi_next (&si))
1391 tree stmt = bsi_stmt (si);
1394 ann = stmt_ann (stmt);
1395 set_stmt_info ((tree_ann_t)ann, new_stmt_vec_info (stmt, res));
1399 LOOP_VINFO_LOOP (res) = loop;
1400 LOOP_VINFO_BBS (res) = bbs;
1401 LOOP_VINFO_EXIT_COND (res) = NULL;
1402 LOOP_VINFO_NITERS (res) = NULL;
1403 LOOP_VINFO_VECTORIZABLE_P (res) = 0;
1404 LOOP_PEELING_FOR_ALIGNMENT (res) = 0;
1405 LOOP_VINFO_VECT_FACTOR (res) = 0;
1406 VARRAY_GENERIC_PTR_INIT (LOOP_VINFO_DATAREFS (res), 20, "loop_datarefs");
1407 VARRAY_GENERIC_PTR_INIT (LOOP_VINFO_DDRS (res), 20, "loop_ddrs");
1408 LOOP_VINFO_UNALIGNED_DR (res) = NULL;
1414 /* Function destroy_loop_vec_info.
1416 Free LOOP_VINFO struct, as well as all the stmt_vec_info structs of all the
1417 stmts in the loop. */
1420 destroy_loop_vec_info (loop_vec_info loop_vinfo)
1425 block_stmt_iterator si;
1431 loop = LOOP_VINFO_LOOP (loop_vinfo);
1433 bbs = LOOP_VINFO_BBS (loop_vinfo);
1434 nbbs = loop->num_nodes;
1436 for (j = 0; j < nbbs; j++)
1438 basic_block bb = bbs[j];
1440 stmt_vec_info stmt_info;
1442 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1444 tree_ann_t ann = get_tree_ann (phi);
1446 stmt_info = vinfo_for_stmt (phi);
1448 set_stmt_info (ann, NULL);
1451 for (si = bsi_start (bb); !bsi_end_p (si); bsi_next (&si))
1453 tree stmt = bsi_stmt (si);
1454 stmt_ann_t ann = stmt_ann (stmt);
1455 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1459 VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info));
1461 set_stmt_info ((tree_ann_t)ann, NULL);
1466 free (LOOP_VINFO_BBS (loop_vinfo));
1467 varray_clear (LOOP_VINFO_DATAREFS (loop_vinfo));
1468 varray_clear (LOOP_VINFO_DDRS (loop_vinfo));
1474 /* Function vect_force_dr_alignment_p.
1476 Returns whether the alignment of a DECL can be forced to be aligned
1477 on ALIGNMENT bit boundary. */
1480 vect_can_force_dr_alignment_p (tree decl, unsigned int alignment)
1482 if (TREE_CODE (decl) != VAR_DECL)
1485 if (DECL_EXTERNAL (decl))
1488 if (TREE_ASM_WRITTEN (decl))
1491 if (TREE_STATIC (decl))
1492 return (alignment <= MAX_OFILE_ALIGNMENT);
1494 /* This is not 100% correct. The absolute correct stack alignment
1495 is STACK_BOUNDARY. We're supposed to hope, but not assume, that
1496 PREFERRED_STACK_BOUNDARY is honored by all translation units.
1497 However, until someone implements forced stack alignment, SSE
1498 isn't really usable without this. */
1499 return (alignment <= PREFERRED_STACK_BOUNDARY);
1503 /* Function get_vectype_for_scalar_type.
1505 Returns the vector type corresponding to SCALAR_TYPE as supported
1509 get_vectype_for_scalar_type (tree scalar_type)
1511 enum machine_mode inner_mode = TYPE_MODE (scalar_type);
1512 int nbytes = GET_MODE_SIZE (inner_mode);
1516 if (nbytes == 0 || nbytes >= UNITS_PER_SIMD_WORD)
1519 /* FORNOW: Only a single vector size per target (UNITS_PER_SIMD_WORD)
1521 nunits = UNITS_PER_SIMD_WORD / nbytes;
1523 vectype = build_vector_type (scalar_type, nunits);
1524 if (vect_print_dump_info (REPORT_DETAILS))
1526 fprintf (vect_dump, "get vectype with %d units of type ", nunits);
1527 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
1533 if (vect_print_dump_info (REPORT_DETAILS))
1535 fprintf (vect_dump, "vectype: ");
1536 print_generic_expr (vect_dump, vectype, TDF_SLIM);
1539 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
1540 && !INTEGRAL_MODE_P (TYPE_MODE (vectype)))
1542 if (vect_print_dump_info (REPORT_DETAILS))
1543 fprintf (vect_dump, "mode not supported by target.");
1551 /* Function vect_supportable_dr_alignment
1553 Return whether the data reference DR is supported with respect to its
1556 enum dr_alignment_support
1557 vect_supportable_dr_alignment (struct data_reference *dr)
1559 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr)));
1560 enum machine_mode mode = (int) TYPE_MODE (vectype);
1562 if (aligned_access_p (dr))
1565 /* Possibly unaligned access. */
1567 if (DR_IS_READ (dr))
1569 if (vec_realign_load_optab->handlers[mode].insn_code != CODE_FOR_nothing
1570 && (!targetm.vectorize.builtin_mask_for_load
1571 || targetm.vectorize.builtin_mask_for_load ()))
1572 return dr_unaligned_software_pipeline;
1574 if (movmisalign_optab->handlers[mode].insn_code != CODE_FOR_nothing)
1575 /* Can't software pipeline the loads, but can at least do them. */
1576 return dr_unaligned_supported;
1580 return dr_unaligned_unsupported;
1584 /* Function vect_is_simple_use.
1587 LOOP - the loop that is being vectorized.
1588 OPERAND - operand of a stmt in LOOP.
1589 DEF - the defining stmt in case OPERAND is an SSA_NAME.
1591 Returns whether a stmt with OPERAND can be vectorized.
1592 Supportable operands are constants, loop invariants, and operands that are
1593 defined by the current iteration of the loop. Unsupportable operands are
1594 those that are defined by a previous iteration of the loop (as is the case
1595 in reduction/induction computations). */
1598 vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, tree *def_stmt,
1599 tree *def, enum vect_def_type *dt)
1602 stmt_vec_info stmt_vinfo;
1603 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1605 *def_stmt = NULL_TREE;
1608 if (vect_print_dump_info (REPORT_DETAILS))
1610 fprintf (vect_dump, "vect_is_simple_use: operand ");
1611 print_generic_expr (vect_dump, operand, TDF_SLIM);
1614 if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST)
1616 *dt = vect_constant_def;
1620 if (TREE_CODE (operand) != SSA_NAME)
1622 if (vect_print_dump_info (REPORT_DETAILS))
1623 fprintf (vect_dump, "not ssa-name.");
1627 *def_stmt = SSA_NAME_DEF_STMT (operand);
1628 if (*def_stmt == NULL_TREE )
1630 if (vect_print_dump_info (REPORT_DETAILS))
1631 fprintf (vect_dump, "no def_stmt.");
1635 if (vect_print_dump_info (REPORT_DETAILS))
1637 fprintf (vect_dump, "def_stmt: ");
1638 print_generic_expr (vect_dump, *def_stmt, TDF_SLIM);
1641 /* empty stmt is expected only in case of a function argument.
1642 (Otherwise - we expect a phi_node or a modify_expr). */
1643 if (IS_EMPTY_STMT (*def_stmt))
1645 tree arg = TREE_OPERAND (*def_stmt, 0);
1646 if (TREE_CODE (arg) == INTEGER_CST || TREE_CODE (arg) == REAL_CST)
1649 *dt = vect_invariant_def;
1653 if (vect_print_dump_info (REPORT_DETAILS))
1654 fprintf (vect_dump, "Unexpected empty stmt.");
1658 bb = bb_for_stmt (*def_stmt);
1659 if (!flow_bb_inside_loop_p (loop, bb))
1660 *dt = vect_invariant_def;
1663 stmt_vinfo = vinfo_for_stmt (*def_stmt);
1664 *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo);
1667 if (*dt == vect_unknown_def_type)
1669 if (vect_print_dump_info (REPORT_DETAILS))
1670 fprintf (vect_dump, "Unsupported pattern.");
1674 /* stmts inside the loop that have been identified as performing
1675 a reduction operation cannot have uses in the loop. */
1676 if (*dt == vect_reduction_def && TREE_CODE (*def_stmt) != PHI_NODE)
1678 if (vect_print_dump_info (REPORT_DETAILS))
1679 fprintf (vect_dump, "reduction used in loop.");
1683 if (vect_print_dump_info (REPORT_DETAILS))
1684 fprintf (vect_dump, "type of def: %d.",*dt);
1686 switch (TREE_CODE (*def_stmt))
1689 *def = PHI_RESULT (*def_stmt);
1690 gcc_assert (*dt == vect_induction_def || *dt == vect_reduction_def
1691 || *dt == vect_invariant_def);
1695 *def = TREE_OPERAND (*def_stmt, 0);
1696 gcc_assert (*dt == vect_loop_def || *dt == vect_invariant_def);
1700 if (vect_print_dump_info (REPORT_DETAILS))
1701 fprintf (vect_dump, "unsupported defining stmt: ");
1705 if (*dt == vect_induction_def)
1707 if (vect_print_dump_info (REPORT_DETAILS))
1708 fprintf (vect_dump, "induction not supported.");
1716 /* Function reduction_code_for_scalar_code
1719 CODE - tree_code of a reduction operations.
1722 REDUC_CODE - the corresponding tree-code to be used to reduce the
1723 vector of partial results into a single scalar result (which
1724 will also reside in a vector).
1726 Return TRUE if a corresponding REDUC_CODE was found, FALSE otherwise. */
1729 reduction_code_for_scalar_code (enum tree_code code,
1730 enum tree_code *reduc_code)
1735 *reduc_code = REDUC_MAX_EXPR;
1739 *reduc_code = REDUC_MIN_EXPR;
1743 *reduc_code = REDUC_PLUS_EXPR;
1752 /* Function vect_is_simple_reduction
1754 Detect a cross-iteration def-use cucle that represents a simple
1755 reduction computation. We look for the following pattern:
1760 a2 = operation (a3, a1)
1763 1. operation is commutative and associative and it is safe to
1764 change the order of the computation.
1765 2. no uses for a2 in the loop (a2 is used out of the loop)
1766 3. no uses of a1 in the loop besides the reduction operation.
1768 Condition 1 is tested here.
1769 Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized. */
1772 vect_is_simple_reduction (struct loop *loop ATTRIBUTE_UNUSED,
1773 tree phi ATTRIBUTE_UNUSED)
1775 edge latch_e = loop_latch_edge (loop);
1776 tree loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e);
1777 tree def_stmt, def1, def2;
1778 enum tree_code code;
1780 tree operation, op1, op2;
1783 if (TREE_CODE (loop_arg) != SSA_NAME)
1785 if (vect_print_dump_info (REPORT_DETAILS))
1787 fprintf (vect_dump, "reduction: not ssa_name: ");
1788 print_generic_expr (vect_dump, loop_arg, TDF_SLIM);
1793 def_stmt = SSA_NAME_DEF_STMT (loop_arg);
1796 if (vect_print_dump_info (REPORT_DETAILS))
1797 fprintf (vect_dump, "reduction: no def_stmt.");
1801 if (TREE_CODE (def_stmt) != MODIFY_EXPR)
1803 if (vect_print_dump_info (REPORT_DETAILS))
1805 print_generic_expr (vect_dump, def_stmt, TDF_SLIM);
1810 operation = TREE_OPERAND (def_stmt, 1);
1811 code = TREE_CODE (operation);
1812 if (!commutative_tree_code (code) || !associative_tree_code (code))
1814 if (vect_print_dump_info (REPORT_DETAILS))
1816 fprintf (vect_dump, "reduction: not commutative/associative: ");
1817 print_generic_expr (vect_dump, operation, TDF_SLIM);
1822 op_type = TREE_CODE_LENGTH (code);
1823 if (op_type != binary_op)
1825 if (vect_print_dump_info (REPORT_DETAILS))
1827 fprintf (vect_dump, "reduction: not binary operation: ");
1828 print_generic_expr (vect_dump, operation, TDF_SLIM);
1833 op1 = TREE_OPERAND (operation, 0);
1834 op2 = TREE_OPERAND (operation, 1);
1835 if (TREE_CODE (op1) != SSA_NAME || TREE_CODE (op2) != SSA_NAME)
1837 if (vect_print_dump_info (REPORT_DETAILS))
1839 fprintf (vect_dump, "reduction: uses not ssa_names: ");
1840 print_generic_expr (vect_dump, operation, TDF_SLIM);
1845 /* Check that it's ok to change the order of the computation. */
1846 type = TREE_TYPE (operation);
1847 if (TYPE_MAIN_VARIANT (type) != TYPE_MAIN_VARIANT (TREE_TYPE (op1))
1848 || TYPE_MAIN_VARIANT (type) != TYPE_MAIN_VARIANT (TREE_TYPE (op2)))
1850 if (vect_print_dump_info (REPORT_DETAILS))
1852 fprintf (vect_dump, "reduction: multiple types: operation type: ");
1853 print_generic_expr (vect_dump, type, TDF_SLIM);
1854 fprintf (vect_dump, ", operands types: ");
1855 print_generic_expr (vect_dump, TREE_TYPE (op1), TDF_SLIM);
1856 fprintf (vect_dump, ",");
1857 print_generic_expr (vect_dump, TREE_TYPE (op2), TDF_SLIM);
1862 /* CHECKME: check for !flag_finite_math_only too? */
1863 if (SCALAR_FLOAT_TYPE_P (type) && !flag_unsafe_math_optimizations)
1865 /* Changing the order of operations changes the sematics. */
1866 if (vect_print_dump_info (REPORT_DETAILS))
1868 fprintf (vect_dump, "reduction: unsafe fp math optimization: ");
1869 print_generic_expr (vect_dump, operation, TDF_SLIM);
1873 else if (INTEGRAL_TYPE_P (type) && !TYPE_UNSIGNED (type) && flag_trapv)
1875 /* Changing the order of operations changes the sematics. */
1876 if (vect_print_dump_info (REPORT_DETAILS))
1878 fprintf (vect_dump, "reduction: unsafe int math optimization: ");
1879 print_generic_expr (vect_dump, operation, TDF_SLIM);
1884 /* reduction is safe. we're dealing with one of the following:
1885 1) integer arithmetic and no trapv
1886 2) floating point arithmetic, and special flags permit this optimization.
1888 def1 = SSA_NAME_DEF_STMT (op1);
1889 def2 = SSA_NAME_DEF_STMT (op2);
1892 if (vect_print_dump_info (REPORT_DETAILS))
1894 fprintf (vect_dump, "reduction: no defs for operands: ");
1895 print_generic_expr (vect_dump, operation, TDF_SLIM);
1900 if (TREE_CODE (def1) == MODIFY_EXPR
1901 && flow_bb_inside_loop_p (loop, bb_for_stmt (def1))
1904 if (vect_print_dump_info (REPORT_DETAILS))
1906 fprintf (vect_dump, "detected reduction:");
1907 print_generic_expr (vect_dump, operation, TDF_SLIM);
1911 else if (TREE_CODE (def2) == MODIFY_EXPR
1912 && flow_bb_inside_loop_p (loop, bb_for_stmt (def2))
1918 /* Swap operands (just for simplicity - so that the rest of the code
1919 can assume that the reduction variable is always the last (second)
1921 if (vect_print_dump_info (REPORT_DETAILS))
1923 fprintf (vect_dump, "detected reduction: need to swap operands:");
1924 print_generic_expr (vect_dump, operation, TDF_SLIM);
1928 FOR_EACH_SSA_USE_OPERAND (use, def_stmt, iter, SSA_OP_USE)
1930 tree tuse = USE_FROM_PTR (use);
1933 else if (tuse == op2)
1940 if (vect_print_dump_info (REPORT_DETAILS))
1942 fprintf (vect_dump, "reduction: unknown pattern.");
1943 print_generic_expr (vect_dump, operation, TDF_SLIM);
1950 /* Function vect_is_simple_iv_evolution.
1952 FORNOW: A simple evolution of an induction variables in the loop is
1953 considered a polynomial evolution with constant step. */
1956 vect_is_simple_iv_evolution (unsigned loop_nb, tree access_fn, tree * init,
1962 tree evolution_part = evolution_part_in_loop_num (access_fn, loop_nb);
1964 /* When there is no evolution in this loop, the evolution function
1966 if (evolution_part == NULL_TREE)
1969 /* When the evolution is a polynomial of degree >= 2
1970 the evolution function is not "simple". */
1971 if (tree_is_chrec (evolution_part))
1974 step_expr = evolution_part;
1975 init_expr = unshare_expr (initial_condition_in_loop_num (access_fn,
1978 if (vect_print_dump_info (REPORT_DETAILS))
1980 fprintf (vect_dump, "step: ");
1981 print_generic_expr (vect_dump, step_expr, TDF_SLIM);
1982 fprintf (vect_dump, ", init: ");
1983 print_generic_expr (vect_dump, init_expr, TDF_SLIM);
1989 if (TREE_CODE (step_expr) != INTEGER_CST)
1991 if (vect_print_dump_info (REPORT_DETAILS))
1992 fprintf (vect_dump, "step unknown.");
2000 /* Function vectorize_loops.
2002 Entry Point to loop vectorization phase. */
2005 vectorize_loops (struct loops *loops)
2008 unsigned int num_vectorized_loops = 0;
2010 /* Fix the verbosity level if not defined explicitly by the user. */
2011 vect_set_dump_settings ();
2013 /* ----------- Analyze loops. ----------- */
2015 /* If some loop was duplicated, it gets bigger number
2016 than all previously defined loops. This fact allows us to run
2017 only over initial loops skipping newly generated ones. */
2018 vect_loops_num = loops->num;
2019 for (i = 1; i < vect_loops_num; i++)
2021 loop_vec_info loop_vinfo;
2022 struct loop *loop = loops->parray[i];
2027 vect_loop_location = find_loop_location (loop);
2028 loop_vinfo = vect_analyze_loop (loop);
2029 loop->aux = loop_vinfo;
2031 if (!loop_vinfo || !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo))
2034 vect_transform_loop (loop_vinfo, loops);
2035 num_vectorized_loops++;
2038 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS))
2039 fprintf (vect_dump, "vectorized %u loops in function.\n",
2040 num_vectorized_loops);
2042 /* ----------- Finalize. ----------- */
2044 for (i = 1; i < vect_loops_num; i++)
2046 struct loop *loop = loops->parray[i];
2047 loop_vec_info loop_vinfo;
2051 loop_vinfo = loop->aux;
2052 destroy_loop_vec_info (loop_vinfo);