[ARM] Add ACLE 2.0 predefined marco __ARM_FEATURE_IDIV
[gcc.git] / gcc / tree-ssa-loop-manip.c
1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2014 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "tree.h"
25 #include "tm_p.h"
26 #include "basic-block.h"
27 #include "tree-ssa-alias.h"
28 #include "internal-fn.h"
29 #include "gimple-expr.h"
30 #include "is-a.h"
31 #include "gimple.h"
32 #include "gimplify.h"
33 #include "gimple-iterator.h"
34 #include "gimplify-me.h"
35 #include "gimple-ssa.h"
36 #include "tree-cfg.h"
37 #include "tree-phinodes.h"
38 #include "ssa-iterators.h"
39 #include "stringpool.h"
40 #include "tree-ssanames.h"
41 #include "tree-ssa-loop-ivopts.h"
42 #include "tree-ssa-loop-manip.h"
43 #include "tree-ssa-loop-niter.h"
44 #include "tree-ssa-loop.h"
45 #include "tree-into-ssa.h"
46 #include "tree-ssa.h"
47 #include "dumpfile.h"
48 #include "gimple-pretty-print.h"
49 #include "cfgloop.h"
50 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
51 #include "tree-scalar-evolution.h"
52 #include "params.h"
53 #include "tree-inline.h"
54 #include "langhooks.h"
55
56 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
57 so that we can free them all at once. */
58 static bitmap_obstack loop_renamer_obstack;
59
60 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
61 It is expected that neither BASE nor STEP are shared with other expressions
62 (unless the sharing rules allow this). Use VAR as a base var_decl for it
63 (if NULL, a new temporary will be created). The increment will occur at
64 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
65 AFTER can be computed using standard_iv_increment_position. The ssa versions
66 of the variable before and after increment will be stored in VAR_BEFORE and
67 VAR_AFTER (unless they are NULL). */
68
69 void
70 create_iv (tree base, tree step, tree var, struct loop *loop,
71 gimple_stmt_iterator *incr_pos, bool after,
72 tree *var_before, tree *var_after)
73 {
74 gimple stmt;
75 tree initial, step1;
76 gimple_seq stmts;
77 tree vb, va;
78 enum tree_code incr_op = PLUS_EXPR;
79 edge pe = loop_preheader_edge (loop);
80
81 if (var != NULL_TREE)
82 {
83 vb = make_ssa_name (var, NULL);
84 va = make_ssa_name (var, NULL);
85 }
86 else
87 {
88 vb = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
89 va = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
90 }
91 if (var_before)
92 *var_before = vb;
93 if (var_after)
94 *var_after = va;
95
96 /* For easier readability of the created code, produce MINUS_EXPRs
97 when suitable. */
98 if (TREE_CODE (step) == INTEGER_CST)
99 {
100 if (TYPE_UNSIGNED (TREE_TYPE (step)))
101 {
102 step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
103 if (tree_int_cst_lt (step1, step))
104 {
105 incr_op = MINUS_EXPR;
106 step = step1;
107 }
108 }
109 else
110 {
111 bool ovf;
112
113 if (!tree_expr_nonnegative_warnv_p (step, &ovf)
114 && may_negate_without_overflow_p (step))
115 {
116 incr_op = MINUS_EXPR;
117 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
118 }
119 }
120 }
121 if (POINTER_TYPE_P (TREE_TYPE (base)))
122 {
123 if (TREE_CODE (base) == ADDR_EXPR)
124 mark_addressable (TREE_OPERAND (base, 0));
125 step = convert_to_ptrofftype (step);
126 if (incr_op == MINUS_EXPR)
127 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
128 incr_op = POINTER_PLUS_EXPR;
129 }
130 /* Gimplify the step if necessary. We put the computations in front of the
131 loop (i.e. the step should be loop invariant). */
132 step = force_gimple_operand (step, &stmts, true, NULL_TREE);
133 if (stmts)
134 gsi_insert_seq_on_edge_immediate (pe, stmts);
135
136 stmt = gimple_build_assign_with_ops (incr_op, va, vb, step);
137 if (after)
138 gsi_insert_after (incr_pos, stmt, GSI_NEW_STMT);
139 else
140 gsi_insert_before (incr_pos, stmt, GSI_NEW_STMT);
141
142 initial = force_gimple_operand (base, &stmts, true, var);
143 if (stmts)
144 gsi_insert_seq_on_edge_immediate (pe, stmts);
145
146 stmt = create_phi_node (vb, loop->header);
147 add_phi_arg (stmt, initial, loop_preheader_edge (loop), UNKNOWN_LOCATION);
148 add_phi_arg (stmt, va, loop_latch_edge (loop), UNKNOWN_LOCATION);
149 }
150
151 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
152 both DEF_LOOP and USE_LOOP. */
153
154 static inline struct loop *
155 find_sibling_superloop (struct loop *use_loop, struct loop *def_loop)
156 {
157 unsigned ud = loop_depth (use_loop);
158 unsigned dd = loop_depth (def_loop);
159 gcc_assert (ud > 0 && dd > 0);
160 if (ud > dd)
161 use_loop = superloop_at_depth (use_loop, dd);
162 if (ud < dd)
163 def_loop = superloop_at_depth (def_loop, ud);
164 while (loop_outer (use_loop) != loop_outer (def_loop))
165 {
166 use_loop = loop_outer (use_loop);
167 def_loop = loop_outer (def_loop);
168 gcc_assert (use_loop && def_loop);
169 }
170 return use_loop;
171 }
172
173 /* DEF_BB is a basic block containing a DEF that needs rewriting into
174 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
175 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
176 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
177 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
178
179 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
180 or one of its loop fathers, in which DEF is live. This set is returned
181 in the bitmap LIVE_EXITS.
182
183 Instead of computing the complete livein set of the def, we use the loop
184 nesting tree as a form of poor man's structure analysis. This greatly
185 speeds up the analysis, which is important because this function may be
186 called on all SSA names that need rewriting, one at a time. */
187
188 static void
189 compute_live_loop_exits (bitmap live_exits, bitmap use_blocks,
190 bitmap *loop_exits, basic_block def_bb)
191 {
192 unsigned i;
193 bitmap_iterator bi;
194 struct loop *def_loop = def_bb->loop_father;
195 unsigned def_loop_depth = loop_depth (def_loop);
196 bitmap def_loop_exits;
197
198 /* Normally the work list size is bounded by the number of basic
199 blocks in the largest loop. We don't know this number, but we
200 can be fairly sure that it will be relatively small. */
201 auto_vec<basic_block> worklist (MAX (8, n_basic_blocks_for_fn (cfun) / 128));
202
203 EXECUTE_IF_SET_IN_BITMAP (use_blocks, 0, i, bi)
204 {
205 basic_block use_bb = BASIC_BLOCK_FOR_FN (cfun, i);
206 struct loop *use_loop = use_bb->loop_father;
207 gcc_checking_assert (def_loop != use_loop
208 && ! flow_loop_nested_p (def_loop, use_loop));
209 if (! flow_loop_nested_p (use_loop, def_loop))
210 use_bb = find_sibling_superloop (use_loop, def_loop)->header;
211 if (bitmap_set_bit (live_exits, use_bb->index))
212 worklist.safe_push (use_bb);
213 }
214
215 /* Iterate until the worklist is empty. */
216 while (! worklist.is_empty ())
217 {
218 edge e;
219 edge_iterator ei;
220
221 /* Pull a block off the worklist. */
222 basic_block bb = worklist.pop ();
223
224 /* Make sure we have at least enough room in the work list
225 for all predecessors of this block. */
226 worklist.reserve (EDGE_COUNT (bb->preds));
227
228 /* For each predecessor block. */
229 FOR_EACH_EDGE (e, ei, bb->preds)
230 {
231 basic_block pred = e->src;
232 struct loop *pred_loop = pred->loop_father;
233 unsigned pred_loop_depth = loop_depth (pred_loop);
234 bool pred_visited;
235
236 /* We should have met DEF_BB along the way. */
237 gcc_assert (pred != ENTRY_BLOCK_PTR_FOR_FN (cfun));
238
239 if (pred_loop_depth >= def_loop_depth)
240 {
241 if (pred_loop_depth > def_loop_depth)
242 pred_loop = superloop_at_depth (pred_loop, def_loop_depth);
243 /* If we've reached DEF_LOOP, our train ends here. */
244 if (pred_loop == def_loop)
245 continue;
246 }
247 else if (! flow_loop_nested_p (pred_loop, def_loop))
248 pred = find_sibling_superloop (pred_loop, def_loop)->header;
249
250 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
251 we had already added PRED to LIVEIN before. */
252 pred_visited = !bitmap_set_bit (live_exits, pred->index);
253
254 /* If we have visited PRED before, don't add it to the worklist.
255 If BB dominates PRED, then we're probably looking at a loop.
256 We're only interested in looking up in the dominance tree
257 because DEF_BB dominates all the uses. */
258 if (pred_visited || dominated_by_p (CDI_DOMINATORS, pred, bb))
259 continue;
260
261 worklist.quick_push (pred);
262 }
263 }
264
265 def_loop_exits = BITMAP_ALLOC (&loop_renamer_obstack);
266 for (struct loop *loop = def_loop;
267 loop != current_loops->tree_root;
268 loop = loop_outer (loop))
269 bitmap_ior_into (def_loop_exits, loop_exits[loop->num]);
270 bitmap_and_into (live_exits, def_loop_exits);
271 BITMAP_FREE (def_loop_exits);
272 }
273
274 /* Add a loop-closing PHI for VAR in basic block EXIT. */
275
276 static void
277 add_exit_phi (basic_block exit, tree var)
278 {
279 gimple phi;
280 edge e;
281 edge_iterator ei;
282
283 #ifdef ENABLE_CHECKING
284 /* Check that at least one of the edges entering the EXIT block exits
285 the loop, or a superloop of that loop, that VAR is defined in. */
286 gimple def_stmt = SSA_NAME_DEF_STMT (var);
287 basic_block def_bb = gimple_bb (def_stmt);
288 FOR_EACH_EDGE (e, ei, exit->preds)
289 {
290 struct loop *aloop = find_common_loop (def_bb->loop_father,
291 e->src->loop_father);
292 if (!flow_bb_inside_loop_p (aloop, e->dest))
293 break;
294 }
295
296 gcc_checking_assert (e);
297 #endif
298
299 phi = create_phi_node (NULL_TREE, exit);
300 create_new_def_for (var, phi, gimple_phi_result_ptr (phi));
301 FOR_EACH_EDGE (e, ei, exit->preds)
302 add_phi_arg (phi, var, e, UNKNOWN_LOCATION);
303
304 if (dump_file && (dump_flags & TDF_DETAILS))
305 {
306 fprintf (dump_file, ";; Created LCSSA PHI: ");
307 print_gimple_stmt (dump_file, phi, 0, dump_flags);
308 }
309 }
310
311 /* Add exit phis for VAR that is used in LIVEIN.
312 Exits of the loops are stored in LOOP_EXITS. */
313
314 static void
315 add_exit_phis_var (tree var, bitmap use_blocks, bitmap *loop_exits)
316 {
317 unsigned index;
318 bitmap_iterator bi;
319 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var));
320 bitmap live_exits = BITMAP_ALLOC (&loop_renamer_obstack);
321
322 gcc_checking_assert (! bitmap_bit_p (use_blocks, def_bb->index));
323
324 compute_live_loop_exits (live_exits, use_blocks, loop_exits, def_bb);
325
326 EXECUTE_IF_SET_IN_BITMAP (live_exits, 0, index, bi)
327 {
328 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun, index), var);
329 }
330
331 BITMAP_FREE (live_exits);
332 }
333
334 /* Add exit phis for the names marked in NAMES_TO_RENAME.
335 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
336 names are used are stored in USE_BLOCKS. */
337
338 static void
339 add_exit_phis (bitmap names_to_rename, bitmap *use_blocks, bitmap *loop_exits)
340 {
341 unsigned i;
342 bitmap_iterator bi;
343
344 EXECUTE_IF_SET_IN_BITMAP (names_to_rename, 0, i, bi)
345 {
346 add_exit_phis_var (ssa_name (i), use_blocks[i], loop_exits);
347 }
348 }
349
350 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
351
352 static void
353 get_loops_exits (bitmap *loop_exits)
354 {
355 struct loop *loop;
356 unsigned j;
357 edge e;
358
359 FOR_EACH_LOOP (loop, 0)
360 {
361 vec<edge> exit_edges = get_loop_exit_edges (loop);
362 loop_exits[loop->num] = BITMAP_ALLOC (&loop_renamer_obstack);
363 FOR_EACH_VEC_ELT (exit_edges, j, e)
364 bitmap_set_bit (loop_exits[loop->num], e->dest->index);
365 exit_edges.release ();
366 }
367 }
368
369 /* For USE in BB, if it is used outside of the loop it is defined in,
370 mark it for rewrite. Record basic block BB where it is used
371 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
372
373 static void
374 find_uses_to_rename_use (basic_block bb, tree use, bitmap *use_blocks,
375 bitmap need_phis)
376 {
377 unsigned ver;
378 basic_block def_bb;
379 struct loop *def_loop;
380
381 if (TREE_CODE (use) != SSA_NAME)
382 return;
383
384 ver = SSA_NAME_VERSION (use);
385 def_bb = gimple_bb (SSA_NAME_DEF_STMT (use));
386 if (!def_bb)
387 return;
388 def_loop = def_bb->loop_father;
389
390 /* If the definition is not inside a loop, it is not interesting. */
391 if (!loop_outer (def_loop))
392 return;
393
394 /* If the use is not outside of the loop it is defined in, it is not
395 interesting. */
396 if (flow_bb_inside_loop_p (def_loop, bb))
397 return;
398
399 /* If we're seeing VER for the first time, we still have to allocate
400 a bitmap for its uses. */
401 if (bitmap_set_bit (need_phis, ver))
402 use_blocks[ver] = BITMAP_ALLOC (&loop_renamer_obstack);
403 bitmap_set_bit (use_blocks[ver], bb->index);
404 }
405
406 /* For uses in STMT, mark names that are used outside of the loop they are
407 defined to rewrite. Record the set of blocks in that the ssa
408 names are defined to USE_BLOCKS and the ssa names themselves to
409 NEED_PHIS. */
410
411 static void
412 find_uses_to_rename_stmt (gimple stmt, bitmap *use_blocks, bitmap need_phis)
413 {
414 ssa_op_iter iter;
415 tree var;
416 basic_block bb = gimple_bb (stmt);
417
418 if (is_gimple_debug (stmt))
419 return;
420
421 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_USE)
422 find_uses_to_rename_use (bb, var, use_blocks, need_phis);
423 }
424
425 /* Marks names that are used in BB and outside of the loop they are
426 defined in for rewrite. Records the set of blocks in that the ssa
427 names are defined to USE_BLOCKS. Record the SSA names that will
428 need exit PHIs in NEED_PHIS. */
429
430 static void
431 find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis)
432 {
433 gimple_stmt_iterator bsi;
434 edge e;
435 edge_iterator ei;
436
437 FOR_EACH_EDGE (e, ei, bb->succs)
438 for (bsi = gsi_start_phis (e->dest); !gsi_end_p (bsi); gsi_next (&bsi))
439 {
440 gimple phi = gsi_stmt (bsi);
441 if (! virtual_operand_p (gimple_phi_result (phi)))
442 find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e),
443 use_blocks, need_phis);
444 }
445
446 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
447 find_uses_to_rename_stmt (gsi_stmt (bsi), use_blocks, need_phis);
448 }
449
450 /* Marks names that are used outside of the loop they are defined in
451 for rewrite. Records the set of blocks in that the ssa
452 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
453 scan only blocks in this set. */
454
455 static void
456 find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis)
457 {
458 basic_block bb;
459 unsigned index;
460 bitmap_iterator bi;
461
462 if (changed_bbs)
463 EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi)
464 find_uses_to_rename_bb (BASIC_BLOCK_FOR_FN (cfun, index), use_blocks, need_phis);
465 else
466 FOR_EACH_BB_FN (bb, cfun)
467 find_uses_to_rename_bb (bb, use_blocks, need_phis);
468 }
469
470 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
471 phi nodes to ensure that no variable is used outside the loop it is
472 defined in.
473
474 This strengthening of the basic ssa form has several advantages:
475
476 1) Updating it during unrolling/peeling/versioning is trivial, since
477 we do not need to care about the uses outside of the loop.
478 The same applies to virtual operands which are also rewritten into
479 loop closed SSA form. Note that virtual operands are always live
480 until function exit.
481 2) The behavior of all uses of an induction variable is the same.
482 Without this, you need to distinguish the case when the variable
483 is used outside of the loop it is defined in, for example
484
485 for (i = 0; i < 100; i++)
486 {
487 for (j = 0; j < 100; j++)
488 {
489 k = i + j;
490 use1 (k);
491 }
492 use2 (k);
493 }
494
495 Looking from the outer loop with the normal SSA form, the first use of k
496 is not well-behaved, while the second one is an induction variable with
497 base 99 and step 1.
498
499 If CHANGED_BBS is not NULL, we look for uses outside loops only in
500 the basic blocks in this set.
501
502 UPDATE_FLAG is used in the call to update_ssa. See
503 TODO_update_ssa* for documentation. */
504
505 void
506 rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag)
507 {
508 bitmap *use_blocks;
509 bitmap names_to_rename;
510
511 loops_state_set (LOOP_CLOSED_SSA);
512 if (number_of_loops (cfun) <= 1)
513 return;
514
515 /* If the pass has caused the SSA form to be out-of-date, update it
516 now. */
517 update_ssa (update_flag);
518
519 bitmap_obstack_initialize (&loop_renamer_obstack);
520
521 names_to_rename = BITMAP_ALLOC (&loop_renamer_obstack);
522
523 /* Uses of names to rename. We don't have to initialize this array,
524 because we know that we will only have entries for the SSA names
525 in NAMES_TO_RENAME. */
526 use_blocks = XNEWVEC (bitmap, num_ssa_names);
527
528 /* Find the uses outside loops. */
529 find_uses_to_rename (changed_bbs, use_blocks, names_to_rename);
530
531 if (!bitmap_empty_p (names_to_rename))
532 {
533 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
534 that are the destination of an edge exiting loop number I. */
535 bitmap *loop_exits = XNEWVEC (bitmap, number_of_loops (cfun));
536 get_loops_exits (loop_exits);
537
538 /* Add the PHI nodes on exits of the loops for the names we need to
539 rewrite. */
540 add_exit_phis (names_to_rename, use_blocks, loop_exits);
541
542 free (loop_exits);
543
544 /* Fix up all the names found to be used outside their original
545 loops. */
546 update_ssa (TODO_update_ssa);
547 }
548
549 bitmap_obstack_release (&loop_renamer_obstack);
550 free (use_blocks);
551 }
552
553 /* Check invariants of the loop closed ssa form for the USE in BB. */
554
555 static void
556 check_loop_closed_ssa_use (basic_block bb, tree use)
557 {
558 gimple def;
559 basic_block def_bb;
560
561 if (TREE_CODE (use) != SSA_NAME || virtual_operand_p (use))
562 return;
563
564 def = SSA_NAME_DEF_STMT (use);
565 def_bb = gimple_bb (def);
566 gcc_assert (!def_bb
567 || flow_bb_inside_loop_p (def_bb->loop_father, bb));
568 }
569
570 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
571
572 static void
573 check_loop_closed_ssa_stmt (basic_block bb, gimple stmt)
574 {
575 ssa_op_iter iter;
576 tree var;
577
578 if (is_gimple_debug (stmt))
579 return;
580
581 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_USE)
582 check_loop_closed_ssa_use (bb, var);
583 }
584
585 /* Checks that invariants of the loop closed ssa form are preserved.
586 Call verify_ssa when VERIFY_SSA_P is true. */
587
588 DEBUG_FUNCTION void
589 verify_loop_closed_ssa (bool verify_ssa_p)
590 {
591 basic_block bb;
592 gimple_stmt_iterator bsi;
593 gimple phi;
594 edge e;
595 edge_iterator ei;
596
597 if (number_of_loops (cfun) <= 1)
598 return;
599
600 if (verify_ssa_p)
601 verify_ssa (false, true);
602
603 timevar_push (TV_VERIFY_LOOP_CLOSED);
604
605 FOR_EACH_BB_FN (bb, cfun)
606 {
607 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
608 {
609 phi = gsi_stmt (bsi);
610 FOR_EACH_EDGE (e, ei, bb->preds)
611 check_loop_closed_ssa_use (e->src,
612 PHI_ARG_DEF_FROM_EDGE (phi, e));
613 }
614
615 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
616 check_loop_closed_ssa_stmt (bb, gsi_stmt (bsi));
617 }
618
619 timevar_pop (TV_VERIFY_LOOP_CLOSED);
620 }
621
622 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
623 preserve the loop closed ssa form. The newly created block is returned. */
624
625 basic_block
626 split_loop_exit_edge (edge exit)
627 {
628 basic_block dest = exit->dest;
629 basic_block bb = split_edge (exit);
630 gimple phi, new_phi;
631 tree new_name, name;
632 use_operand_p op_p;
633 gimple_stmt_iterator psi;
634 source_location locus;
635
636 for (psi = gsi_start_phis (dest); !gsi_end_p (psi); gsi_next (&psi))
637 {
638 phi = gsi_stmt (psi);
639 op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb));
640 locus = gimple_phi_arg_location_from_edge (phi, single_succ_edge (bb));
641
642 name = USE_FROM_PTR (op_p);
643
644 /* If the argument of the PHI node is a constant, we do not need
645 to keep it inside loop. */
646 if (TREE_CODE (name) != SSA_NAME)
647 continue;
648
649 /* Otherwise create an auxiliary phi node that will copy the value
650 of the SSA name out of the loop. */
651 new_name = duplicate_ssa_name (name, NULL);
652 new_phi = create_phi_node (new_name, bb);
653 add_phi_arg (new_phi, name, exit, locus);
654 SET_USE (op_p, new_name);
655 }
656
657 return bb;
658 }
659
660 /* Returns the basic block in that statements should be emitted for induction
661 variables incremented at the end of the LOOP. */
662
663 basic_block
664 ip_end_pos (struct loop *loop)
665 {
666 return loop->latch;
667 }
668
669 /* Returns the basic block in that statements should be emitted for induction
670 variables incremented just before exit condition of a LOOP. */
671
672 basic_block
673 ip_normal_pos (struct loop *loop)
674 {
675 gimple last;
676 basic_block bb;
677 edge exit;
678
679 if (!single_pred_p (loop->latch))
680 return NULL;
681
682 bb = single_pred (loop->latch);
683 last = last_stmt (bb);
684 if (!last
685 || gimple_code (last) != GIMPLE_COND)
686 return NULL;
687
688 exit = EDGE_SUCC (bb, 0);
689 if (exit->dest == loop->latch)
690 exit = EDGE_SUCC (bb, 1);
691
692 if (flow_bb_inside_loop_p (loop, exit->dest))
693 return NULL;
694
695 return bb;
696 }
697
698 /* Stores the standard position for induction variable increment in LOOP
699 (just before the exit condition if it is available and latch block is empty,
700 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
701 the increment should be inserted after *BSI. */
702
703 void
704 standard_iv_increment_position (struct loop *loop, gimple_stmt_iterator *bsi,
705 bool *insert_after)
706 {
707 basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop);
708 gimple last = last_stmt (latch);
709
710 if (!bb
711 || (last && gimple_code (last) != GIMPLE_LABEL))
712 {
713 *bsi = gsi_last_bb (latch);
714 *insert_after = true;
715 }
716 else
717 {
718 *bsi = gsi_last_bb (bb);
719 *insert_after = false;
720 }
721 }
722
723 /* Copies phi node arguments for duplicated blocks. The index of the first
724 duplicated block is FIRST_NEW_BLOCK. */
725
726 static void
727 copy_phi_node_args (unsigned first_new_block)
728 {
729 unsigned i;
730
731 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
732 BASIC_BLOCK_FOR_FN (cfun, i)->flags |= BB_DUPLICATED;
733
734 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
735 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun, i));
736
737 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
738 BASIC_BLOCK_FOR_FN (cfun, i)->flags &= ~BB_DUPLICATED;
739 }
740
741
742 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
743 updates the PHI nodes at start of the copied region. In order to
744 achieve this, only loops whose exits all lead to the same location
745 are handled.
746
747 Notice that we do not completely update the SSA web after
748 duplication. The caller is responsible for calling update_ssa
749 after the loop has been duplicated. */
750
751 bool
752 gimple_duplicate_loop_to_header_edge (struct loop *loop, edge e,
753 unsigned int ndupl, sbitmap wont_exit,
754 edge orig, vec<edge> *to_remove,
755 int flags)
756 {
757 unsigned first_new_block;
758
759 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES))
760 return false;
761 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS))
762 return false;
763
764 first_new_block = last_basic_block_for_fn (cfun);
765 if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit,
766 orig, to_remove, flags))
767 return false;
768
769 /* Readd the removed phi args for e. */
770 flush_pending_stmts (e);
771
772 /* Copy the phi node arguments. */
773 copy_phi_node_args (first_new_block);
774
775 scev_reset ();
776
777 return true;
778 }
779
780 /* Returns true if we can unroll LOOP FACTOR times. Number
781 of iterations of the loop is returned in NITER. */
782
783 bool
784 can_unroll_loop_p (struct loop *loop, unsigned factor,
785 struct tree_niter_desc *niter)
786 {
787 edge exit;
788
789 /* Check whether unrolling is possible. We only want to unroll loops
790 for that we are able to determine number of iterations. We also
791 want to split the extra iterations of the loop from its end,
792 therefore we require that the loop has precisely one
793 exit. */
794
795 exit = single_dom_exit (loop);
796 if (!exit)
797 return false;
798
799 if (!number_of_iterations_exit (loop, exit, niter, false)
800 || niter->cmp == ERROR_MARK
801 /* Scalar evolutions analysis might have copy propagated
802 the abnormal ssa names into these expressions, hence
803 emitting the computations based on them during loop
804 unrolling might create overlapping life ranges for
805 them, and failures in out-of-ssa. */
806 || contains_abnormal_ssa_name_p (niter->may_be_zero)
807 || contains_abnormal_ssa_name_p (niter->control.base)
808 || contains_abnormal_ssa_name_p (niter->control.step)
809 || contains_abnormal_ssa_name_p (niter->bound))
810 return false;
811
812 /* And of course, we must be able to duplicate the loop. */
813 if (!can_duplicate_loop_p (loop))
814 return false;
815
816 /* The final loop should be small enough. */
817 if (tree_num_loop_insns (loop, &eni_size_weights) * factor
818 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS))
819 return false;
820
821 return true;
822 }
823
824 /* Determines the conditions that control execution of LOOP unrolled FACTOR
825 times. DESC is number of iterations of LOOP. ENTER_COND is set to
826 condition that must be true if the main loop can be entered.
827 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
828 how the exit from the unrolled loop should be controlled. */
829
830 static void
831 determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc,
832 unsigned factor, tree *enter_cond,
833 tree *exit_base, tree *exit_step,
834 enum tree_code *exit_cmp, tree *exit_bound)
835 {
836 gimple_seq stmts;
837 tree base = desc->control.base;
838 tree step = desc->control.step;
839 tree bound = desc->bound;
840 tree type = TREE_TYPE (step);
841 tree bigstep, delta;
842 tree min = lower_bound_in_type (type, type);
843 tree max = upper_bound_in_type (type, type);
844 enum tree_code cmp = desc->cmp;
845 tree cond = boolean_true_node, assum;
846
847 /* For pointers, do the arithmetics in the type of step. */
848 base = fold_convert (type, base);
849 bound = fold_convert (type, bound);
850
851 *enter_cond = boolean_false_node;
852 *exit_base = NULL_TREE;
853 *exit_step = NULL_TREE;
854 *exit_cmp = ERROR_MARK;
855 *exit_bound = NULL_TREE;
856 gcc_assert (cmp != ERROR_MARK);
857
858 /* We only need to be correct when we answer question
859 "Do at least FACTOR more iterations remain?" in the unrolled loop.
860 Thus, transforming BASE + STEP * i <> BOUND to
861 BASE + STEP * i < BOUND is ok. */
862 if (cmp == NE_EXPR)
863 {
864 if (tree_int_cst_sign_bit (step))
865 cmp = GT_EXPR;
866 else
867 cmp = LT_EXPR;
868 }
869 else if (cmp == LT_EXPR)
870 {
871 gcc_assert (!tree_int_cst_sign_bit (step));
872 }
873 else if (cmp == GT_EXPR)
874 {
875 gcc_assert (tree_int_cst_sign_bit (step));
876 }
877 else
878 gcc_unreachable ();
879
880 /* The main body of the loop may be entered iff:
881
882 1) desc->may_be_zero is false.
883 2) it is possible to check that there are at least FACTOR iterations
884 of the loop, i.e., BOUND - step * FACTOR does not overflow.
885 3) # of iterations is at least FACTOR */
886
887 if (!integer_zerop (desc->may_be_zero))
888 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
889 invert_truthvalue (desc->may_be_zero),
890 cond);
891
892 bigstep = fold_build2 (MULT_EXPR, type, step,
893 build_int_cst_type (type, factor));
894 delta = fold_build2 (MINUS_EXPR, type, bigstep, step);
895 if (cmp == LT_EXPR)
896 assum = fold_build2 (GE_EXPR, boolean_type_node,
897 bound,
898 fold_build2 (PLUS_EXPR, type, min, delta));
899 else
900 assum = fold_build2 (LE_EXPR, boolean_type_node,
901 bound,
902 fold_build2 (PLUS_EXPR, type, max, delta));
903 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
904
905 bound = fold_build2 (MINUS_EXPR, type, bound, delta);
906 assum = fold_build2 (cmp, boolean_type_node, base, bound);
907 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
908
909 cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE);
910 if (stmts)
911 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
912 /* cond now may be a gimple comparison, which would be OK, but also any
913 other gimple rhs (say a && b). In this case we need to force it to
914 operand. */
915 if (!is_gimple_condexpr (cond))
916 {
917 cond = force_gimple_operand (cond, &stmts, true, NULL_TREE);
918 if (stmts)
919 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
920 }
921 *enter_cond = cond;
922
923 base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE);
924 if (stmts)
925 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
926 bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE);
927 if (stmts)
928 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
929
930 *exit_base = base;
931 *exit_step = bigstep;
932 *exit_cmp = cmp;
933 *exit_bound = bound;
934 }
935
936 /* Scales the frequencies of all basic blocks in LOOP that are strictly
937 dominated by BB by NUM/DEN. */
938
939 static void
940 scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb,
941 int num, int den)
942 {
943 basic_block son;
944
945 if (den == 0)
946 return;
947
948 for (son = first_dom_son (CDI_DOMINATORS, bb);
949 son;
950 son = next_dom_son (CDI_DOMINATORS, son))
951 {
952 if (!flow_bb_inside_loop_p (loop, son))
953 continue;
954 scale_bbs_frequencies_int (&son, 1, num, den);
955 scale_dominated_blocks_in_loop (loop, son, num, den);
956 }
957 }
958
959 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
960 EXIT is the exit of the loop to that DESC corresponds.
961
962 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
963 under that loop exits in the first iteration even if N != 0,
964
965 while (1)
966 {
967 x = phi (init, next);
968
969 pre;
970 if (st)
971 break;
972 post;
973 }
974
975 becomes (with possibly the exit conditions formulated a bit differently,
976 avoiding the need to create a new iv):
977
978 if (MAY_BE_ZERO || N < FACTOR)
979 goto rest;
980
981 do
982 {
983 x = phi (init, next);
984
985 pre;
986 post;
987 pre;
988 post;
989 ...
990 pre;
991 post;
992 N -= FACTOR;
993
994 } while (N >= FACTOR);
995
996 rest:
997 init' = phi (init, x);
998
999 while (1)
1000 {
1001 x = phi (init', next);
1002
1003 pre;
1004 if (st)
1005 break;
1006 post;
1007 }
1008
1009 Before the loop is unrolled, TRANSFORM is called for it (only for the
1010 unrolled loop, but not for its versioned copy). DATA is passed to
1011 TRANSFORM. */
1012
1013 /* Probability in % that the unrolled loop is entered. Just a guess. */
1014 #define PROB_UNROLLED_LOOP_ENTERED 90
1015
1016 void
1017 tree_transform_and_unroll_loop (struct loop *loop, unsigned factor,
1018 edge exit, struct tree_niter_desc *desc,
1019 transform_callback transform,
1020 void *data)
1021 {
1022 gimple exit_if;
1023 tree ctr_before, ctr_after;
1024 tree enter_main_cond, exit_base, exit_step, exit_bound;
1025 enum tree_code exit_cmp;
1026 gimple phi_old_loop, phi_new_loop, phi_rest;
1027 gimple_stmt_iterator psi_old_loop, psi_new_loop;
1028 tree init, next, new_init;
1029 struct loop *new_loop;
1030 basic_block rest, exit_bb;
1031 edge old_entry, new_entry, old_latch, precond_edge, new_exit;
1032 edge new_nonexit, e;
1033 gimple_stmt_iterator bsi;
1034 use_operand_p op;
1035 bool ok;
1036 unsigned est_niter, prob_entry, scale_unrolled, scale_rest, freq_e, freq_h;
1037 unsigned new_est_niter, i, prob;
1038 unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP;
1039 sbitmap wont_exit;
1040 auto_vec<edge> to_remove;
1041
1042 est_niter = expected_loop_iterations (loop);
1043 determine_exit_conditions (loop, desc, factor,
1044 &enter_main_cond, &exit_base, &exit_step,
1045 &exit_cmp, &exit_bound);
1046
1047 /* Let us assume that the unrolled loop is quite likely to be entered. */
1048 if (integer_nonzerop (enter_main_cond))
1049 prob_entry = REG_BR_PROB_BASE;
1050 else
1051 prob_entry = PROB_UNROLLED_LOOP_ENTERED * REG_BR_PROB_BASE / 100;
1052
1053 /* The values for scales should keep profile consistent, and somewhat close
1054 to correct.
1055
1056 TODO: The current value of SCALE_REST makes it appear that the loop that
1057 is created by splitting the remaining iterations of the unrolled loop is
1058 executed the same number of times as the original loop, and with the same
1059 frequencies, which is obviously wrong. This does not appear to cause
1060 problems, so we do not bother with fixing it for now. To make the profile
1061 correct, we would need to change the probability of the exit edge of the
1062 loop, and recompute the distribution of frequencies in its body because
1063 of this change (scale the frequencies of blocks before and after the exit
1064 by appropriate factors). */
1065 scale_unrolled = prob_entry;
1066 scale_rest = REG_BR_PROB_BASE;
1067
1068 new_loop = loop_version (loop, enter_main_cond, NULL,
1069 prob_entry, scale_unrolled, scale_rest, true);
1070 gcc_assert (new_loop != NULL);
1071 update_ssa (TODO_update_ssa);
1072
1073 /* Determine the probability of the exit edge of the unrolled loop. */
1074 new_est_niter = est_niter / factor;
1075
1076 /* Without profile feedback, loops for that we do not know a better estimate
1077 are assumed to roll 10 times. When we unroll such loop, it appears to
1078 roll too little, and it may even seem to be cold. To avoid this, we
1079 ensure that the created loop appears to roll at least 5 times (but at
1080 most as many times as before unrolling). */
1081 if (new_est_niter < 5)
1082 {
1083 if (est_niter < 5)
1084 new_est_niter = est_niter;
1085 else
1086 new_est_niter = 5;
1087 }
1088
1089 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1090 loop latch (and make its condition dummy, for the moment). */
1091 rest = loop_preheader_edge (new_loop)->src;
1092 precond_edge = single_pred_edge (rest);
1093 split_edge (loop_latch_edge (loop));
1094 exit_bb = single_pred (loop->latch);
1095
1096 /* Since the exit edge will be removed, the frequency of all the blocks
1097 in the loop that are dominated by it must be scaled by
1098 1 / (1 - exit->probability). */
1099 scale_dominated_blocks_in_loop (loop, exit->src,
1100 REG_BR_PROB_BASE,
1101 REG_BR_PROB_BASE - exit->probability);
1102
1103 bsi = gsi_last_bb (exit_bb);
1104 exit_if = gimple_build_cond (EQ_EXPR, integer_zero_node,
1105 integer_zero_node,
1106 NULL_TREE, NULL_TREE);
1107
1108 gsi_insert_after (&bsi, exit_if, GSI_NEW_STMT);
1109 new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr);
1110 rescan_loop_exit (new_exit, true, false);
1111
1112 /* Set the probability of new exit to the same of the old one. Fix
1113 the frequency of the latch block, by scaling it back by
1114 1 - exit->probability. */
1115 new_exit->count = exit->count;
1116 new_exit->probability = exit->probability;
1117 new_nonexit = single_pred_edge (loop->latch);
1118 new_nonexit->probability = REG_BR_PROB_BASE - exit->probability;
1119 new_nonexit->flags = EDGE_TRUE_VALUE;
1120 new_nonexit->count -= exit->count;
1121 if (new_nonexit->count < 0)
1122 new_nonexit->count = 0;
1123 scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1124 REG_BR_PROB_BASE);
1125
1126 old_entry = loop_preheader_edge (loop);
1127 new_entry = loop_preheader_edge (new_loop);
1128 old_latch = loop_latch_edge (loop);
1129 for (psi_old_loop = gsi_start_phis (loop->header),
1130 psi_new_loop = gsi_start_phis (new_loop->header);
1131 !gsi_end_p (psi_old_loop);
1132 gsi_next (&psi_old_loop), gsi_next (&psi_new_loop))
1133 {
1134 phi_old_loop = gsi_stmt (psi_old_loop);
1135 phi_new_loop = gsi_stmt (psi_new_loop);
1136
1137 init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry);
1138 op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry);
1139 gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op)));
1140 next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch);
1141
1142 /* Prefer using original variable as a base for the new ssa name.
1143 This is necessary for virtual ops, and useful in order to avoid
1144 losing debug info for real ops. */
1145 if (TREE_CODE (next) == SSA_NAME
1146 && useless_type_conversion_p (TREE_TYPE (next),
1147 TREE_TYPE (init)))
1148 new_init = copy_ssa_name (next, NULL);
1149 else if (TREE_CODE (init) == SSA_NAME
1150 && useless_type_conversion_p (TREE_TYPE (init),
1151 TREE_TYPE (next)))
1152 new_init = copy_ssa_name (init, NULL);
1153 else if (useless_type_conversion_p (TREE_TYPE (next), TREE_TYPE (init)))
1154 new_init = make_temp_ssa_name (TREE_TYPE (next), NULL, "unrinittmp");
1155 else
1156 new_init = make_temp_ssa_name (TREE_TYPE (init), NULL, "unrinittmp");
1157
1158 phi_rest = create_phi_node (new_init, rest);
1159
1160 add_phi_arg (phi_rest, init, precond_edge, UNKNOWN_LOCATION);
1161 add_phi_arg (phi_rest, next, new_exit, UNKNOWN_LOCATION);
1162 SET_USE (op, new_init);
1163 }
1164
1165 remove_path (exit);
1166
1167 /* Transform the loop. */
1168 if (transform)
1169 (*transform) (loop, data);
1170
1171 /* Unroll the loop and remove the exits in all iterations except for the
1172 last one. */
1173 wont_exit = sbitmap_alloc (factor);
1174 bitmap_ones (wont_exit);
1175 bitmap_clear_bit (wont_exit, factor - 1);
1176
1177 ok = gimple_duplicate_loop_to_header_edge
1178 (loop, loop_latch_edge (loop), factor - 1,
1179 wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ);
1180 free (wont_exit);
1181 gcc_assert (ok);
1182
1183 FOR_EACH_VEC_ELT (to_remove, i, e)
1184 {
1185 ok = remove_path (e);
1186 gcc_assert (ok);
1187 }
1188 update_ssa (TODO_update_ssa);
1189
1190 /* Ensure that the frequencies in the loop match the new estimated
1191 number of iterations, and change the probability of the new
1192 exit edge. */
1193 freq_h = loop->header->frequency;
1194 freq_e = EDGE_FREQUENCY (loop_preheader_edge (loop));
1195 if (freq_h != 0)
1196 scale_loop_frequencies (loop, freq_e * (new_est_niter + 1), freq_h);
1197
1198 exit_bb = single_pred (loop->latch);
1199 new_exit = find_edge (exit_bb, rest);
1200 new_exit->count = loop_preheader_edge (loop)->count;
1201 new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1);
1202
1203 rest->count += new_exit->count;
1204 rest->frequency += EDGE_FREQUENCY (new_exit);
1205
1206 new_nonexit = single_pred_edge (loop->latch);
1207 prob = new_nonexit->probability;
1208 new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability;
1209 new_nonexit->count = exit_bb->count - new_exit->count;
1210 if (new_nonexit->count < 0)
1211 new_nonexit->count = 0;
1212 if (prob > 0)
1213 scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1214 prob);
1215
1216 /* Finally create the new counter for number of iterations and add the new
1217 exit instruction. */
1218 bsi = gsi_last_nondebug_bb (exit_bb);
1219 exit_if = gsi_stmt (bsi);
1220 create_iv (exit_base, exit_step, NULL_TREE, loop,
1221 &bsi, false, &ctr_before, &ctr_after);
1222 gimple_cond_set_code (exit_if, exit_cmp);
1223 gimple_cond_set_lhs (exit_if, ctr_after);
1224 gimple_cond_set_rhs (exit_if, exit_bound);
1225 update_stmt (exit_if);
1226
1227 #ifdef ENABLE_CHECKING
1228 verify_flow_info ();
1229 verify_loop_structure ();
1230 verify_loop_closed_ssa (true);
1231 #endif
1232 }
1233
1234 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1235 want to transform the loop before unrolling. The meaning
1236 of the arguments is the same as for tree_transform_and_unroll_loop. */
1237
1238 void
1239 tree_unroll_loop (struct loop *loop, unsigned factor,
1240 edge exit, struct tree_niter_desc *desc)
1241 {
1242 tree_transform_and_unroll_loop (loop, factor, exit, desc,
1243 NULL, NULL);
1244 }
1245
1246 /* Rewrite the phi node at position PSI in function of the main
1247 induction variable MAIN_IV and insert the generated code at GSI. */
1248
1249 static void
1250 rewrite_phi_with_iv (loop_p loop,
1251 gimple_stmt_iterator *psi,
1252 gimple_stmt_iterator *gsi,
1253 tree main_iv)
1254 {
1255 affine_iv iv;
1256 gimple stmt, phi = gsi_stmt (*psi);
1257 tree atype, mtype, val, res = PHI_RESULT (phi);
1258
1259 if (virtual_operand_p (res) || res == main_iv)
1260 {
1261 gsi_next (psi);
1262 return;
1263 }
1264
1265 if (!simple_iv (loop, loop, res, &iv, true))
1266 {
1267 gsi_next (psi);
1268 return;
1269 }
1270
1271 remove_phi_node (psi, false);
1272
1273 atype = TREE_TYPE (res);
1274 mtype = POINTER_TYPE_P (atype) ? sizetype : atype;
1275 val = fold_build2 (MULT_EXPR, mtype, unshare_expr (iv.step),
1276 fold_convert (mtype, main_iv));
1277 val = fold_build2 (POINTER_TYPE_P (atype)
1278 ? POINTER_PLUS_EXPR : PLUS_EXPR,
1279 atype, unshare_expr (iv.base), val);
1280 val = force_gimple_operand_gsi (gsi, val, false, NULL_TREE, true,
1281 GSI_SAME_STMT);
1282 stmt = gimple_build_assign (res, val);
1283 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1284 }
1285
1286 /* Rewrite all the phi nodes of LOOP in function of the main induction
1287 variable MAIN_IV. */
1288
1289 static void
1290 rewrite_all_phi_nodes_with_iv (loop_p loop, tree main_iv)
1291 {
1292 unsigned i;
1293 basic_block *bbs = get_loop_body_in_dom_order (loop);
1294 gimple_stmt_iterator psi;
1295
1296 for (i = 0; i < loop->num_nodes; i++)
1297 {
1298 basic_block bb = bbs[i];
1299 gimple_stmt_iterator gsi = gsi_after_labels (bb);
1300
1301 if (bb->loop_father != loop)
1302 continue;
1303
1304 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); )
1305 rewrite_phi_with_iv (loop, &psi, &gsi, main_iv);
1306 }
1307
1308 free (bbs);
1309 }
1310
1311 /* Bases all the induction variables in LOOP on a single induction
1312 variable (unsigned with base 0 and step 1), whose final value is
1313 compared with *NIT. When the IV type precision has to be larger
1314 than *NIT type precision, *NIT is converted to the larger type, the
1315 conversion code is inserted before the loop, and *NIT is updated to
1316 the new definition. When BUMP_IN_LATCH is true, the induction
1317 variable is incremented in the loop latch, otherwise it is
1318 incremented in the loop header. Return the induction variable that
1319 was created. */
1320
1321 tree
1322 canonicalize_loop_ivs (struct loop *loop, tree *nit, bool bump_in_latch)
1323 {
1324 unsigned precision = TYPE_PRECISION (TREE_TYPE (*nit));
1325 unsigned original_precision = precision;
1326 tree type, var_before;
1327 gimple_stmt_iterator gsi, psi;
1328 gimple stmt;
1329 edge exit = single_dom_exit (loop);
1330 gimple_seq stmts;
1331 enum machine_mode mode;
1332 bool unsigned_p = false;
1333
1334 for (psi = gsi_start_phis (loop->header);
1335 !gsi_end_p (psi); gsi_next (&psi))
1336 {
1337 gimple phi = gsi_stmt (psi);
1338 tree res = PHI_RESULT (phi);
1339 bool uns;
1340
1341 type = TREE_TYPE (res);
1342 if (virtual_operand_p (res)
1343 || (!INTEGRAL_TYPE_P (type)
1344 && !POINTER_TYPE_P (type))
1345 || TYPE_PRECISION (type) < precision)
1346 continue;
1347
1348 uns = POINTER_TYPE_P (type) | TYPE_UNSIGNED (type);
1349
1350 if (TYPE_PRECISION (type) > precision)
1351 unsigned_p = uns;
1352 else
1353 unsigned_p |= uns;
1354
1355 precision = TYPE_PRECISION (type);
1356 }
1357
1358 mode = smallest_mode_for_size (precision, MODE_INT);
1359 precision = GET_MODE_PRECISION (mode);
1360 type = build_nonstandard_integer_type (precision, unsigned_p);
1361
1362 if (original_precision != precision)
1363 {
1364 *nit = fold_convert (type, *nit);
1365 *nit = force_gimple_operand (*nit, &stmts, true, NULL_TREE);
1366 if (stmts)
1367 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1368 }
1369
1370 if (bump_in_latch)
1371 gsi = gsi_last_bb (loop->latch);
1372 else
1373 gsi = gsi_last_nondebug_bb (loop->header);
1374 create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
1375 loop, &gsi, bump_in_latch, &var_before, NULL);
1376
1377 rewrite_all_phi_nodes_with_iv (loop, var_before);
1378
1379 stmt = last_stmt (exit->src);
1380 /* Make the loop exit if the control condition is not satisfied. */
1381 if (exit->flags & EDGE_TRUE_VALUE)
1382 {
1383 edge te, fe;
1384
1385 extract_true_false_edges_from_block (exit->src, &te, &fe);
1386 te->flags = EDGE_FALSE_VALUE;
1387 fe->flags = EDGE_TRUE_VALUE;
1388 }
1389 gimple_cond_set_code (stmt, LT_EXPR);
1390 gimple_cond_set_lhs (stmt, var_before);
1391 gimple_cond_set_rhs (stmt, *nit);
1392 update_stmt (stmt);
1393
1394 return var_before;
1395 }