mn10300: Use the STC bb-reorder algorithm at -Os
[official-gcc.git] / gcc / tree-ssa-loop-ivcanon.c
blob6599ffc743ebf118957235107d5bdf9d4203ffec
1 /* Induction variable canonicalization and loop peeling.
2 Copyright (C) 2004-2015 Free Software Foundation, Inc.
4 This file is part of GCC.
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.
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.
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/>. */
20 /* This pass detects the loops that iterate a constant number of times,
21 adds a canonical induction variable (step -1, tested against 0)
22 and replaces the exit test. This enables the less powerful rtl
23 level analysis to use this information.
25 This might spoil the code in some cases (by increasing register pressure).
26 Note that in the case the new variable is not needed, ivopts will get rid
27 of it, so it might only be a problem when there are no other linear induction
28 variables. In that case the created optimization possibilities are likely
29 to pay up.
31 We also perform
32 - complete unrolling (or peeling) when the loops is rolling few enough
33 times
34 - simple peeling (i.e. copying few initial iterations prior the loop)
35 when number of iteration estimate is known (typically by the profile
36 info). */
38 #include "config.h"
39 #include "system.h"
40 #include "coretypes.h"
41 #include "backend.h"
42 #include "cfghooks.h"
43 #include "tree.h"
44 #include "gimple.h"
45 #include "hard-reg-set.h"
46 #include "ssa.h"
47 #include "alias.h"
48 #include "fold-const.h"
49 #include "tm_p.h"
50 #include "profile.h"
51 #include "gimple-pretty-print.h"
52 #include "internal-fn.h"
53 #include "gimple-fold.h"
54 #include "tree-eh.h"
55 #include "gimple-iterator.h"
56 #include "cgraph.h"
57 #include "tree-cfg.h"
58 #include "tree-ssa-loop-manip.h"
59 #include "tree-ssa-loop-niter.h"
60 #include "tree-ssa-loop.h"
61 #include "tree-into-ssa.h"
62 #include "cfgloop.h"
63 #include "tree-pass.h"
64 #include "tree-chrec.h"
65 #include "tree-scalar-evolution.h"
66 #include "params.h"
67 #include "flags.h"
68 #include "tree-inline.h"
69 #include "target.h"
70 #include "tree-cfgcleanup.h"
71 #include "builtins.h"
73 /* Specifies types of loops that may be unrolled. */
75 enum unroll_level
77 UL_SINGLE_ITER, /* Only loops that exit immediately in the first
78 iteration. */
79 UL_NO_GROWTH, /* Only loops whose unrolling will not cause increase
80 of code size. */
81 UL_ALL /* All suitable loops. */
84 /* Adds a canonical induction variable to LOOP iterating NITER times. EXIT
85 is the exit edge whose condition is replaced. */
87 static void
88 create_canonical_iv (struct loop *loop, edge exit, tree niter)
90 edge in;
91 tree type, var;
92 gcond *cond;
93 gimple_stmt_iterator incr_at;
94 enum tree_code cmp;
96 if (dump_file && (dump_flags & TDF_DETAILS))
98 fprintf (dump_file, "Added canonical iv to loop %d, ", loop->num);
99 print_generic_expr (dump_file, niter, TDF_SLIM);
100 fprintf (dump_file, " iterations.\n");
103 cond = as_a <gcond *> (last_stmt (exit->src));
104 in = EDGE_SUCC (exit->src, 0);
105 if (in == exit)
106 in = EDGE_SUCC (exit->src, 1);
108 /* Note that we do not need to worry about overflows, since
109 type of niter is always unsigned and all comparisons are
110 just for equality/nonequality -- i.e. everything works
111 with a modulo arithmetics. */
113 type = TREE_TYPE (niter);
114 niter = fold_build2 (PLUS_EXPR, type,
115 niter,
116 build_int_cst (type, 1));
117 incr_at = gsi_last_bb (in->src);
118 create_iv (niter,
119 build_int_cst (type, -1),
120 NULL_TREE, loop,
121 &incr_at, false, NULL, &var);
123 cmp = (exit->flags & EDGE_TRUE_VALUE) ? EQ_EXPR : NE_EXPR;
124 gimple_cond_set_code (cond, cmp);
125 gimple_cond_set_lhs (cond, var);
126 gimple_cond_set_rhs (cond, build_int_cst (type, 0));
127 update_stmt (cond);
130 /* Describe size of loop as detected by tree_estimate_loop_size. */
131 struct loop_size
133 /* Number of instructions in the loop. */
134 int overall;
136 /* Number of instructions that will be likely optimized out in
137 peeled iterations of loop (i.e. computation based on induction
138 variable where induction variable starts at known constant.) */
139 int eliminated_by_peeling;
141 /* Same statistics for last iteration of loop: it is smaller because
142 instructions after exit are not executed. */
143 int last_iteration;
144 int last_iteration_eliminated_by_peeling;
146 /* If some IV computation will become constant. */
147 bool constant_iv;
149 /* Number of call stmts that are not a builtin and are pure or const
150 present on the hot path. */
151 int num_pure_calls_on_hot_path;
152 /* Number of call stmts that are not a builtin and are not pure nor const
153 present on the hot path. */
154 int num_non_pure_calls_on_hot_path;
155 /* Number of statements other than calls in the loop. */
156 int non_call_stmts_on_hot_path;
157 /* Number of branches seen on the hot path. */
158 int num_branches_on_hot_path;
161 /* Return true if OP in STMT will be constant after peeling LOOP. */
163 static bool
164 constant_after_peeling (tree op, gimple *stmt, struct loop *loop)
166 affine_iv iv;
168 if (is_gimple_min_invariant (op))
169 return true;
171 /* We can still fold accesses to constant arrays when index is known. */
172 if (TREE_CODE (op) != SSA_NAME)
174 tree base = op;
176 /* First make fast look if we see constant array inside. */
177 while (handled_component_p (base))
178 base = TREE_OPERAND (base, 0);
179 if ((DECL_P (base)
180 && ctor_for_folding (base) != error_mark_node)
181 || CONSTANT_CLASS_P (base))
183 /* If so, see if we understand all the indices. */
184 base = op;
185 while (handled_component_p (base))
187 if (TREE_CODE (base) == ARRAY_REF
188 && !constant_after_peeling (TREE_OPERAND (base, 1), stmt, loop))
189 return false;
190 base = TREE_OPERAND (base, 0);
192 return true;
194 return false;
197 /* Induction variables are constants. */
198 if (!simple_iv (loop, loop_containing_stmt (stmt), op, &iv, false))
199 return false;
200 if (!is_gimple_min_invariant (iv.base))
201 return false;
202 if (!is_gimple_min_invariant (iv.step))
203 return false;
204 return true;
207 /* Computes an estimated number of insns in LOOP.
208 EXIT (if non-NULL) is an exite edge that will be eliminated in all but last
209 iteration of the loop.
210 EDGE_TO_CANCEL (if non-NULL) is an non-exit edge eliminated in the last iteration
211 of loop.
212 Return results in SIZE, estimate benefits for complete unrolling exiting by EXIT.
213 Stop estimating after UPPER_BOUND is met. Return true in this case. */
215 static bool
216 tree_estimate_loop_size (struct loop *loop, edge exit, edge edge_to_cancel, struct loop_size *size,
217 int upper_bound)
219 basic_block *body = get_loop_body (loop);
220 gimple_stmt_iterator gsi;
221 unsigned int i;
222 bool after_exit;
223 vec<basic_block> path = get_loop_hot_path (loop);
225 size->overall = 0;
226 size->eliminated_by_peeling = 0;
227 size->last_iteration = 0;
228 size->last_iteration_eliminated_by_peeling = 0;
229 size->num_pure_calls_on_hot_path = 0;
230 size->num_non_pure_calls_on_hot_path = 0;
231 size->non_call_stmts_on_hot_path = 0;
232 size->num_branches_on_hot_path = 0;
233 size->constant_iv = 0;
235 if (dump_file && (dump_flags & TDF_DETAILS))
236 fprintf (dump_file, "Estimating sizes for loop %i\n", loop->num);
237 for (i = 0; i < loop->num_nodes; i++)
239 if (edge_to_cancel && body[i] != edge_to_cancel->src
240 && dominated_by_p (CDI_DOMINATORS, body[i], edge_to_cancel->src))
241 after_exit = true;
242 else
243 after_exit = false;
244 if (dump_file && (dump_flags & TDF_DETAILS))
245 fprintf (dump_file, " BB: %i, after_exit: %i\n", body[i]->index, after_exit);
247 for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi); gsi_next (&gsi))
249 gimple *stmt = gsi_stmt (gsi);
250 int num = estimate_num_insns (stmt, &eni_size_weights);
251 bool likely_eliminated = false;
252 bool likely_eliminated_last = false;
253 bool likely_eliminated_peeled = false;
255 if (dump_file && (dump_flags & TDF_DETAILS))
257 fprintf (dump_file, " size: %3i ", num);
258 print_gimple_stmt (dump_file, gsi_stmt (gsi), 0, 0);
261 /* Look for reasons why we might optimize this stmt away. */
263 if (gimple_has_side_effects (stmt))
265 /* Exit conditional. */
266 else if (exit && body[i] == exit->src
267 && stmt == last_stmt (exit->src))
269 if (dump_file && (dump_flags & TDF_DETAILS))
270 fprintf (dump_file, " Exit condition will be eliminated "
271 "in peeled copies.\n");
272 likely_eliminated_peeled = true;
274 else if (edge_to_cancel && body[i] == edge_to_cancel->src
275 && stmt == last_stmt (edge_to_cancel->src))
277 if (dump_file && (dump_flags & TDF_DETAILS))
278 fprintf (dump_file, " Exit condition will be eliminated "
279 "in last copy.\n");
280 likely_eliminated_last = true;
282 /* Sets of IV variables */
283 else if (gimple_code (stmt) == GIMPLE_ASSIGN
284 && constant_after_peeling (gimple_assign_lhs (stmt), stmt, loop))
286 if (dump_file && (dump_flags & TDF_DETAILS))
287 fprintf (dump_file, " Induction variable computation will"
288 " be folded away.\n");
289 likely_eliminated = true;
291 /* Assignments of IV variables. */
292 else if (gimple_code (stmt) == GIMPLE_ASSIGN
293 && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME
294 && constant_after_peeling (gimple_assign_rhs1 (stmt), stmt, loop)
295 && (gimple_assign_rhs_class (stmt) != GIMPLE_BINARY_RHS
296 || constant_after_peeling (gimple_assign_rhs2 (stmt),
297 stmt, loop)))
299 size->constant_iv = true;
300 if (dump_file && (dump_flags & TDF_DETAILS))
301 fprintf (dump_file, " Constant expression will be folded away.\n");
302 likely_eliminated = true;
304 /* Conditionals. */
305 else if ((gimple_code (stmt) == GIMPLE_COND
306 && constant_after_peeling (gimple_cond_lhs (stmt), stmt, loop)
307 && constant_after_peeling (gimple_cond_rhs (stmt), stmt, loop))
308 || (gimple_code (stmt) == GIMPLE_SWITCH
309 && constant_after_peeling (gimple_switch_index (
310 as_a <gswitch *> (stmt)),
311 stmt, loop)))
313 if (dump_file && (dump_flags & TDF_DETAILS))
314 fprintf (dump_file, " Constant conditional.\n");
315 likely_eliminated = true;
318 size->overall += num;
319 if (likely_eliminated || likely_eliminated_peeled)
320 size->eliminated_by_peeling += num;
321 if (!after_exit)
323 size->last_iteration += num;
324 if (likely_eliminated || likely_eliminated_last)
325 size->last_iteration_eliminated_by_peeling += num;
327 if ((size->overall * 3 / 2 - size->eliminated_by_peeling
328 - size->last_iteration_eliminated_by_peeling) > upper_bound)
330 free (body);
331 path.release ();
332 return true;
336 while (path.length ())
338 basic_block bb = path.pop ();
339 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
341 gimple *stmt = gsi_stmt (gsi);
342 if (gimple_code (stmt) == GIMPLE_CALL)
344 int flags = gimple_call_flags (stmt);
345 tree decl = gimple_call_fndecl (stmt);
347 if (decl && DECL_IS_BUILTIN (decl)
348 && is_inexpensive_builtin (decl))
350 else if (flags & (ECF_PURE | ECF_CONST))
351 size->num_pure_calls_on_hot_path++;
352 else
353 size->num_non_pure_calls_on_hot_path++;
354 size->num_branches_on_hot_path ++;
356 else if (gimple_code (stmt) != GIMPLE_CALL
357 && gimple_code (stmt) != GIMPLE_DEBUG)
358 size->non_call_stmts_on_hot_path++;
359 if (((gimple_code (stmt) == GIMPLE_COND
360 && (!constant_after_peeling (gimple_cond_lhs (stmt), stmt, loop)
361 || constant_after_peeling (gimple_cond_rhs (stmt), stmt, loop)))
362 || (gimple_code (stmt) == GIMPLE_SWITCH
363 && !constant_after_peeling (gimple_switch_index (
364 as_a <gswitch *> (stmt)),
365 stmt, loop)))
366 && (!exit || bb != exit->src))
367 size->num_branches_on_hot_path++;
370 path.release ();
371 if (dump_file && (dump_flags & TDF_DETAILS))
372 fprintf (dump_file, "size: %i-%i, last_iteration: %i-%i\n", size->overall,
373 size->eliminated_by_peeling, size->last_iteration,
374 size->last_iteration_eliminated_by_peeling);
376 free (body);
377 return false;
380 /* Estimate number of insns of completely unrolled loop.
381 It is (NUNROLL + 1) * size of loop body with taking into account
382 the fact that in last copy everything after exit conditional
383 is dead and that some instructions will be eliminated after
384 peeling.
386 Loop body is likely going to simplify further, this is difficult
387 to guess, we just decrease the result by 1/3. */
389 static unsigned HOST_WIDE_INT
390 estimated_unrolled_size (struct loop_size *size,
391 unsigned HOST_WIDE_INT nunroll)
393 HOST_WIDE_INT unr_insns = ((nunroll)
394 * (HOST_WIDE_INT) (size->overall
395 - size->eliminated_by_peeling));
396 if (!nunroll)
397 unr_insns = 0;
398 unr_insns += size->last_iteration - size->last_iteration_eliminated_by_peeling;
400 unr_insns = unr_insns * 2 / 3;
401 if (unr_insns <= 0)
402 unr_insns = 1;
404 return unr_insns;
407 /* Loop LOOP is known to not loop. See if there is an edge in the loop
408 body that can be remove to make the loop to always exit and at
409 the same time it does not make any code potentially executed
410 during the last iteration dead.
412 After complete unrolling we still may get rid of the conditional
413 on the exit in the last copy even if we have no idea what it does.
414 This is quite common case for loops of form
416 int a[5];
417 for (i=0;i<b;i++)
418 a[i]=0;
420 Here we prove the loop to iterate 5 times but we do not know
421 it from induction variable.
423 For now we handle only simple case where there is exit condition
424 just before the latch block and the latch block contains no statements
425 with side effect that may otherwise terminate the execution of loop
426 (such as by EH or by terminating the program or longjmp).
428 In the general case we may want to cancel the paths leading to statements
429 loop-niter identified as having undefined effect in the last iteration.
430 The other cases are hopefully rare and will be cleaned up later. */
432 static edge
433 loop_edge_to_cancel (struct loop *loop)
435 vec<edge> exits;
436 unsigned i;
437 edge edge_to_cancel;
438 gimple_stmt_iterator gsi;
440 /* We want only one predecestor of the loop. */
441 if (EDGE_COUNT (loop->latch->preds) > 1)
442 return NULL;
444 exits = get_loop_exit_edges (loop);
446 FOR_EACH_VEC_ELT (exits, i, edge_to_cancel)
448 /* Find the other edge than the loop exit
449 leaving the conditoinal. */
450 if (EDGE_COUNT (edge_to_cancel->src->succs) != 2)
451 continue;
452 if (EDGE_SUCC (edge_to_cancel->src, 0) == edge_to_cancel)
453 edge_to_cancel = EDGE_SUCC (edge_to_cancel->src, 1);
454 else
455 edge_to_cancel = EDGE_SUCC (edge_to_cancel->src, 0);
457 /* We only can handle conditionals. */
458 if (!(edge_to_cancel->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
459 continue;
461 /* We should never have conditionals in the loop latch. */
462 gcc_assert (edge_to_cancel->dest != loop->header);
464 /* Check that it leads to loop latch. */
465 if (edge_to_cancel->dest != loop->latch)
466 continue;
468 exits.release ();
470 /* Verify that the code in loop latch does nothing that may end program
471 execution without really reaching the exit. This may include
472 non-pure/const function calls, EH statements, volatile ASMs etc. */
473 for (gsi = gsi_start_bb (loop->latch); !gsi_end_p (gsi); gsi_next (&gsi))
474 if (gimple_has_side_effects (gsi_stmt (gsi)))
475 return NULL;
476 return edge_to_cancel;
478 exits.release ();
479 return NULL;
482 /* Remove all tests for exits that are known to be taken after LOOP was
483 peeled NPEELED times. Put gcc_unreachable before every statement
484 known to not be executed. */
486 static bool
487 remove_exits_and_undefined_stmts (struct loop *loop, unsigned int npeeled)
489 struct nb_iter_bound *elt;
490 bool changed = false;
492 for (elt = loop->bounds; elt; elt = elt->next)
494 /* If statement is known to be undefined after peeling, turn it
495 into unreachable (or trap when debugging experience is supposed
496 to be good). */
497 if (!elt->is_exit
498 && wi::ltu_p (elt->bound, npeeled))
500 gimple_stmt_iterator gsi = gsi_for_stmt (elt->stmt);
501 gcall *stmt = gimple_build_call
502 (builtin_decl_implicit (BUILT_IN_UNREACHABLE), 0);
503 gimple_set_location (stmt, gimple_location (elt->stmt));
504 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
505 split_block (gimple_bb (stmt), stmt);
506 changed = true;
507 if (dump_file && (dump_flags & TDF_DETAILS))
509 fprintf (dump_file, "Forced statement unreachable: ");
510 print_gimple_stmt (dump_file, elt->stmt, 0, 0);
513 /* If we know the exit will be taken after peeling, update. */
514 else if (elt->is_exit
515 && wi::leu_p (elt->bound, npeeled))
517 basic_block bb = gimple_bb (elt->stmt);
518 edge exit_edge = EDGE_SUCC (bb, 0);
520 if (dump_file && (dump_flags & TDF_DETAILS))
522 fprintf (dump_file, "Forced exit to be taken: ");
523 print_gimple_stmt (dump_file, elt->stmt, 0, 0);
525 if (!loop_exit_edge_p (loop, exit_edge))
526 exit_edge = EDGE_SUCC (bb, 1);
527 gcc_checking_assert (loop_exit_edge_p (loop, exit_edge));
528 gcond *cond_stmt = as_a <gcond *> (elt->stmt);
529 if (exit_edge->flags & EDGE_TRUE_VALUE)
530 gimple_cond_make_true (cond_stmt);
531 else
532 gimple_cond_make_false (cond_stmt);
533 update_stmt (cond_stmt);
534 changed = true;
537 return changed;
540 /* Remove all exits that are known to be never taken because of the loop bound
541 discovered. */
543 static bool
544 remove_redundant_iv_tests (struct loop *loop)
546 struct nb_iter_bound *elt;
547 bool changed = false;
549 if (!loop->any_upper_bound)
550 return false;
551 for (elt = loop->bounds; elt; elt = elt->next)
553 /* Exit is pointless if it won't be taken before loop reaches
554 upper bound. */
555 if (elt->is_exit && loop->any_upper_bound
556 && wi::ltu_p (loop->nb_iterations_upper_bound, elt->bound))
558 basic_block bb = gimple_bb (elt->stmt);
559 edge exit_edge = EDGE_SUCC (bb, 0);
560 struct tree_niter_desc niter;
562 if (!loop_exit_edge_p (loop, exit_edge))
563 exit_edge = EDGE_SUCC (bb, 1);
565 /* Only when we know the actual number of iterations, not
566 just a bound, we can remove the exit. */
567 if (!number_of_iterations_exit (loop, exit_edge,
568 &niter, false, false)
569 || !integer_onep (niter.assumptions)
570 || !integer_zerop (niter.may_be_zero)
571 || !niter.niter
572 || TREE_CODE (niter.niter) != INTEGER_CST
573 || !wi::ltu_p (loop->nb_iterations_upper_bound,
574 wi::to_widest (niter.niter)))
575 continue;
577 if (dump_file && (dump_flags & TDF_DETAILS))
579 fprintf (dump_file, "Removed pointless exit: ");
580 print_gimple_stmt (dump_file, elt->stmt, 0, 0);
582 gcond *cond_stmt = as_a <gcond *> (elt->stmt);
583 if (exit_edge->flags & EDGE_TRUE_VALUE)
584 gimple_cond_make_false (cond_stmt);
585 else
586 gimple_cond_make_true (cond_stmt);
587 update_stmt (cond_stmt);
588 changed = true;
591 return changed;
594 /* Stores loops that will be unlooped after we process whole loop tree. */
595 static vec<loop_p> loops_to_unloop;
596 static vec<int> loops_to_unloop_nunroll;
598 /* Cancel all fully unrolled loops by putting __builtin_unreachable
599 on the latch edge.
600 We do it after all unrolling since unlooping moves basic blocks
601 across loop boundaries trashing loop closed SSA form as well
602 as SCEV info needed to be intact during unrolling.
604 IRRED_INVALIDATED is used to bookkeep if information about
605 irreducible regions may become invalid as a result
606 of the transformation.
607 LOOP_CLOSED_SSA_INVALIDATED is used to bookkepp the case
608 when we need to go into loop closed SSA form. */
610 static void
611 unloop_loops (bitmap loop_closed_ssa_invalidated,
612 bool *irred_invalidated)
614 while (loops_to_unloop.length ())
616 struct loop *loop = loops_to_unloop.pop ();
617 int n_unroll = loops_to_unloop_nunroll.pop ();
618 basic_block latch = loop->latch;
619 edge latch_edge = loop_latch_edge (loop);
620 int flags = latch_edge->flags;
621 location_t locus = latch_edge->goto_locus;
622 gcall *stmt;
623 gimple_stmt_iterator gsi;
625 remove_exits_and_undefined_stmts (loop, n_unroll);
627 /* Unloop destroys the latch edge. */
628 unloop (loop, irred_invalidated, loop_closed_ssa_invalidated);
630 /* Create new basic block for the latch edge destination and wire
631 it in. */
632 stmt = gimple_build_call (builtin_decl_implicit (BUILT_IN_UNREACHABLE), 0);
633 latch_edge = make_edge (latch, create_basic_block (NULL, NULL, latch), flags);
634 latch_edge->probability = 0;
635 latch_edge->count = 0;
636 latch_edge->flags |= flags;
637 latch_edge->goto_locus = locus;
639 latch_edge->dest->loop_father = current_loops->tree_root;
640 latch_edge->dest->count = 0;
641 latch_edge->dest->frequency = 0;
642 set_immediate_dominator (CDI_DOMINATORS, latch_edge->dest, latch_edge->src);
644 gsi = gsi_start_bb (latch_edge->dest);
645 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
647 loops_to_unloop.release ();
648 loops_to_unloop_nunroll.release ();
651 /* Tries to unroll LOOP completely, i.e. NITER times.
652 UL determines which loops we are allowed to unroll.
653 EXIT is the exit of the loop that should be eliminated.
654 MAXITER specfy bound on number of iterations, -1 if it is
655 not known or too large for HOST_WIDE_INT. The location
656 LOCUS corresponding to the loop is used when emitting
657 a summary of the unroll to the dump file. */
659 static bool
660 try_unroll_loop_completely (struct loop *loop,
661 edge exit, tree niter,
662 enum unroll_level ul,
663 HOST_WIDE_INT maxiter,
664 location_t locus)
666 unsigned HOST_WIDE_INT n_unroll = 0, ninsns, unr_insns;
667 struct loop_size size;
668 bool n_unroll_found = false;
669 edge edge_to_cancel = NULL;
670 int report_flags = MSG_OPTIMIZED_LOCATIONS | TDF_RTL | TDF_DETAILS;
672 /* See if we proved number of iterations to be low constant.
674 EXIT is an edge that will be removed in all but last iteration of
675 the loop.
677 EDGE_TO_CACNEL is an edge that will be removed from the last iteration
678 of the unrolled sequence and is expected to make the final loop not
679 rolling.
681 If the number of execution of loop is determined by standard induction
682 variable test, then EXIT and EDGE_TO_CANCEL are the two edges leaving
683 from the iv test. */
684 if (tree_fits_uhwi_p (niter))
686 n_unroll = tree_to_uhwi (niter);
687 n_unroll_found = true;
688 edge_to_cancel = EDGE_SUCC (exit->src, 0);
689 if (edge_to_cancel == exit)
690 edge_to_cancel = EDGE_SUCC (exit->src, 1);
692 /* We do not know the number of iterations and thus we can not eliminate
693 the EXIT edge. */
694 else
695 exit = NULL;
697 /* See if we can improve our estimate by using recorded loop bounds. */
698 if (maxiter >= 0
699 && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
701 n_unroll = maxiter;
702 n_unroll_found = true;
703 /* Loop terminates before the IV variable test, so we can not
704 remove it in the last iteration. */
705 edge_to_cancel = NULL;
708 if (!n_unroll_found)
709 return false;
711 if (n_unroll > (unsigned) PARAM_VALUE (PARAM_MAX_COMPLETELY_PEEL_TIMES))
713 if (dump_file && (dump_flags & TDF_DETAILS))
714 fprintf (dump_file, "Not unrolling loop %d "
715 "(--param max-completely-peeled-times limit reached).\n",
716 loop->num);
717 return false;
720 if (!edge_to_cancel)
721 edge_to_cancel = loop_edge_to_cancel (loop);
723 if (n_unroll)
725 sbitmap wont_exit;
726 edge e;
727 unsigned i;
728 bool large;
729 vec<edge> to_remove = vNULL;
730 if (ul == UL_SINGLE_ITER)
731 return false;
733 large = tree_estimate_loop_size
734 (loop, exit, edge_to_cancel, &size,
735 PARAM_VALUE (PARAM_MAX_COMPLETELY_PEELED_INSNS));
736 ninsns = size.overall;
737 if (large)
739 if (dump_file && (dump_flags & TDF_DETAILS))
740 fprintf (dump_file, "Not unrolling loop %d: it is too large.\n",
741 loop->num);
742 return false;
745 unr_insns = estimated_unrolled_size (&size, n_unroll);
746 if (dump_file && (dump_flags & TDF_DETAILS))
748 fprintf (dump_file, " Loop size: %d\n", (int) ninsns);
749 fprintf (dump_file, " Estimated size after unrolling: %d\n",
750 (int) unr_insns);
753 /* If the code is going to shrink, we don't need to be extra cautious
754 on guessing if the unrolling is going to be profitable. */
755 if (unr_insns
756 /* If there is IV variable that will become constant, we save
757 one instruction in the loop prologue we do not account
758 otherwise. */
759 <= ninsns + (size.constant_iv != false))
761 /* We unroll only inner loops, because we do not consider it profitable
762 otheriwse. We still can cancel loopback edge of not rolling loop;
763 this is always a good idea. */
764 else if (ul == UL_NO_GROWTH)
766 if (dump_file && (dump_flags & TDF_DETAILS))
767 fprintf (dump_file, "Not unrolling loop %d: size would grow.\n",
768 loop->num);
769 return false;
771 /* Outer loops tend to be less interesting candidates for complete
772 unrolling unless we can do a lot of propagation into the inner loop
773 body. For now we disable outer loop unrolling when the code would
774 grow. */
775 else if (loop->inner)
777 if (dump_file && (dump_flags & TDF_DETAILS))
778 fprintf (dump_file, "Not unrolling loop %d: "
779 "it is not innermost and code would grow.\n",
780 loop->num);
781 return false;
783 /* If there is call on a hot path through the loop, then
784 there is most probably not much to optimize. */
785 else if (size.num_non_pure_calls_on_hot_path)
787 if (dump_file && (dump_flags & TDF_DETAILS))
788 fprintf (dump_file, "Not unrolling loop %d: "
789 "contains call and code would grow.\n",
790 loop->num);
791 return false;
793 /* If there is pure/const call in the function, then we
794 can still optimize the unrolled loop body if it contains
795 some other interesting code than the calls and code
796 storing or cumulating the return value. */
797 else if (size.num_pure_calls_on_hot_path
798 /* One IV increment, one test, one ivtmp store
799 and one useful stmt. That is about minimal loop
800 doing pure call. */
801 && (size.non_call_stmts_on_hot_path
802 <= 3 + size.num_pure_calls_on_hot_path))
804 if (dump_file && (dump_flags & TDF_DETAILS))
805 fprintf (dump_file, "Not unrolling loop %d: "
806 "contains just pure calls and code would grow.\n",
807 loop->num);
808 return false;
810 /* Complette unrolling is major win when control flow is removed and
811 one big basic block is created. If the loop contains control flow
812 the optimization may still be a win because of eliminating the loop
813 overhead but it also may blow the branch predictor tables.
814 Limit number of branches on the hot path through the peeled
815 sequence. */
816 else if (size.num_branches_on_hot_path * (int)n_unroll
817 > PARAM_VALUE (PARAM_MAX_PEEL_BRANCHES))
819 if (dump_file && (dump_flags & TDF_DETAILS))
820 fprintf (dump_file, "Not unrolling loop %d: "
821 " number of branches on hot path in the unrolled sequence"
822 " reach --param max-peel-branches limit.\n",
823 loop->num);
824 return false;
826 else if (unr_insns
827 > (unsigned) PARAM_VALUE (PARAM_MAX_COMPLETELY_PEELED_INSNS))
829 if (dump_file && (dump_flags & TDF_DETAILS))
830 fprintf (dump_file, "Not unrolling loop %d: "
831 "(--param max-completely-peeled-insns limit reached).\n",
832 loop->num);
833 return false;
835 dump_printf_loc (report_flags, locus,
836 "loop turned into non-loop; it never loops.\n");
838 initialize_original_copy_tables ();
839 wont_exit = sbitmap_alloc (n_unroll + 1);
840 bitmap_ones (wont_exit);
841 bitmap_clear_bit (wont_exit, 0);
843 if (!gimple_duplicate_loop_to_header_edge (loop, loop_preheader_edge (loop),
844 n_unroll, wont_exit,
845 exit, &to_remove,
846 DLTHE_FLAG_UPDATE_FREQ
847 | DLTHE_FLAG_COMPLETTE_PEEL))
849 free_original_copy_tables ();
850 free (wont_exit);
851 if (dump_file && (dump_flags & TDF_DETAILS))
852 fprintf (dump_file, "Failed to duplicate the loop\n");
853 return false;
856 FOR_EACH_VEC_ELT (to_remove, i, e)
858 bool ok = remove_path (e);
859 gcc_assert (ok);
862 to_remove.release ();
863 free (wont_exit);
864 free_original_copy_tables ();
868 /* Remove the conditional from the last copy of the loop. */
869 if (edge_to_cancel)
871 gcond *cond = as_a <gcond *> (last_stmt (edge_to_cancel->src));
872 if (edge_to_cancel->flags & EDGE_TRUE_VALUE)
873 gimple_cond_make_false (cond);
874 else
875 gimple_cond_make_true (cond);
876 update_stmt (cond);
877 /* Do not remove the path. Doing so may remove outer loop
878 and confuse bookkeeping code in tree_unroll_loops_completelly. */
881 /* Store the loop for later unlooping and exit removal. */
882 loops_to_unloop.safe_push (loop);
883 loops_to_unloop_nunroll.safe_push (n_unroll);
885 if (dump_enabled_p ())
887 if (!n_unroll)
888 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, locus,
889 "loop turned into non-loop; it never loops\n");
890 else
892 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, locus,
893 "loop with %d iterations completely unrolled",
894 (int) (n_unroll + 1));
895 if (profile_info)
896 dump_printf (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS,
897 " (header execution count %d)",
898 (int)loop->header->count);
899 dump_printf (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, "\n");
903 if (dump_file && (dump_flags & TDF_DETAILS))
905 if (exit)
906 fprintf (dump_file, "Exit condition of peeled iterations was "
907 "eliminated.\n");
908 if (edge_to_cancel)
909 fprintf (dump_file, "Last iteration exit edge was proved true.\n");
910 else
911 fprintf (dump_file, "Latch of last iteration was marked by "
912 "__builtin_unreachable ().\n");
915 return true;
918 /* Return number of instructions after peeling. */
919 static unsigned HOST_WIDE_INT
920 estimated_peeled_sequence_size (struct loop_size *size,
921 unsigned HOST_WIDE_INT npeel)
923 return MAX (npeel * (HOST_WIDE_INT) (size->overall
924 - size->eliminated_by_peeling), 1);
927 /* If the loop is expected to iterate N times and is
928 small enough, duplicate the loop body N+1 times before
929 the loop itself. This way the hot path will never
930 enter the loop.
931 Parameters are the same as for try_unroll_loops_completely */
933 static bool
934 try_peel_loop (struct loop *loop,
935 edge exit, tree niter,
936 HOST_WIDE_INT maxiter)
938 int npeel;
939 struct loop_size size;
940 int peeled_size;
941 sbitmap wont_exit;
942 unsigned i;
943 vec<edge> to_remove = vNULL;
944 edge e;
946 /* If the iteration bound is known and large, then we can safely eliminate
947 the check in peeled copies. */
948 if (TREE_CODE (niter) != INTEGER_CST)
949 exit = NULL;
951 if (!flag_peel_loops || PARAM_VALUE (PARAM_MAX_PEEL_TIMES) <= 0)
952 return false;
954 /* Peel only innermost loops. */
955 if (loop->inner)
957 if (dump_file)
958 fprintf (dump_file, "Not peeling: outer loop\n");
959 return false;
962 if (!optimize_loop_for_speed_p (loop))
964 if (dump_file)
965 fprintf (dump_file, "Not peeling: cold loop\n");
966 return false;
969 /* Check if there is an estimate on the number of iterations. */
970 npeel = estimated_loop_iterations_int (loop);
971 if (npeel < 0)
973 if (dump_file)
974 fprintf (dump_file, "Not peeling: number of iterations is not "
975 "estimated\n");
976 return false;
978 if (maxiter >= 0 && maxiter <= npeel)
980 if (dump_file)
981 fprintf (dump_file, "Not peeling: upper bound is known so can "
982 "unroll completely\n");
983 return false;
986 /* We want to peel estimated number of iterations + 1 (so we never
987 enter the loop on quick path). Check against PARAM_MAX_PEEL_TIMES
988 and be sure to avoid overflows. */
989 if (npeel > PARAM_VALUE (PARAM_MAX_PEEL_TIMES) - 1)
991 if (dump_file)
992 fprintf (dump_file, "Not peeling: rolls too much "
993 "(%i + 1 > --param max-peel-times)\n", npeel);
994 return false;
996 npeel++;
998 /* Check peeled loops size. */
999 tree_estimate_loop_size (loop, exit, NULL, &size,
1000 PARAM_VALUE (PARAM_MAX_PEELED_INSNS));
1001 if ((peeled_size = estimated_peeled_sequence_size (&size, npeel))
1002 > PARAM_VALUE (PARAM_MAX_PEELED_INSNS))
1004 if (dump_file)
1005 fprintf (dump_file, "Not peeling: peeled sequence size is too large "
1006 "(%i insns > --param max-peel-insns)", peeled_size);
1007 return false;
1010 /* Duplicate possibly eliminating the exits. */
1011 initialize_original_copy_tables ();
1012 wont_exit = sbitmap_alloc (npeel + 1);
1013 bitmap_ones (wont_exit);
1014 bitmap_clear_bit (wont_exit, 0);
1015 if (!gimple_duplicate_loop_to_header_edge (loop, loop_preheader_edge (loop),
1016 npeel, wont_exit,
1017 exit, &to_remove,
1018 DLTHE_FLAG_UPDATE_FREQ
1019 | DLTHE_FLAG_COMPLETTE_PEEL))
1021 free_original_copy_tables ();
1022 free (wont_exit);
1023 return false;
1025 FOR_EACH_VEC_ELT (to_remove, i, e)
1027 bool ok = remove_path (e);
1028 gcc_assert (ok);
1030 free (wont_exit);
1031 free_original_copy_tables ();
1032 if (dump_file && (dump_flags & TDF_DETAILS))
1034 fprintf (dump_file, "Peeled loop %d, %i times.\n",
1035 loop->num, npeel);
1037 if (loop->any_upper_bound)
1038 loop->nb_iterations_upper_bound -= npeel;
1039 loop->nb_iterations_estimate = 0;
1040 /* Make sure to mark loop cold so we do not try to peel it more. */
1041 scale_loop_profile (loop, 1, 0);
1042 loop->header->count = 0;
1043 return true;
1045 /* Adds a canonical induction variable to LOOP if suitable.
1046 CREATE_IV is true if we may create a new iv. UL determines
1047 which loops we are allowed to completely unroll. If TRY_EVAL is true, we try
1048 to determine the number of iterations of a loop by direct evaluation.
1049 Returns true if cfg is changed. */
1051 static bool
1052 canonicalize_loop_induction_variables (struct loop *loop,
1053 bool create_iv, enum unroll_level ul,
1054 bool try_eval)
1056 edge exit = NULL;
1057 tree niter;
1058 HOST_WIDE_INT maxiter;
1059 bool modified = false;
1060 location_t locus = UNKNOWN_LOCATION;
1062 niter = number_of_latch_executions (loop);
1063 exit = single_exit (loop);
1064 if (TREE_CODE (niter) == INTEGER_CST)
1065 locus = gimple_location (last_stmt (exit->src));
1066 else
1068 /* If the loop has more than one exit, try checking all of them
1069 for # of iterations determinable through scev. */
1070 if (!exit)
1071 niter = find_loop_niter (loop, &exit);
1073 /* Finally if everything else fails, try brute force evaluation. */
1074 if (try_eval
1075 && (chrec_contains_undetermined (niter)
1076 || TREE_CODE (niter) != INTEGER_CST))
1077 niter = find_loop_niter_by_eval (loop, &exit);
1079 if (exit)
1080 locus = gimple_location (last_stmt (exit->src));
1082 if (TREE_CODE (niter) != INTEGER_CST)
1083 exit = NULL;
1086 /* We work exceptionally hard here to estimate the bound
1087 by find_loop_niter_by_eval. Be sure to keep it for future. */
1088 if (niter && TREE_CODE (niter) == INTEGER_CST)
1090 record_niter_bound (loop, wi::to_widest (niter),
1091 exit == single_likely_exit (loop), true);
1094 /* Force re-computation of loop bounds so we can remove redundant exits. */
1095 maxiter = max_loop_iterations_int (loop);
1097 if (dump_file && (dump_flags & TDF_DETAILS)
1098 && TREE_CODE (niter) == INTEGER_CST)
1100 fprintf (dump_file, "Loop %d iterates ", loop->num);
1101 print_generic_expr (dump_file, niter, TDF_SLIM);
1102 fprintf (dump_file, " times.\n");
1104 if (dump_file && (dump_flags & TDF_DETAILS)
1105 && maxiter >= 0)
1107 fprintf (dump_file, "Loop %d iterates at most %i times.\n", loop->num,
1108 (int)maxiter);
1111 /* Remove exits that are known to be never taken based on loop bound.
1112 Needs to be called after compilation of max_loop_iterations_int that
1113 populates the loop bounds. */
1114 modified |= remove_redundant_iv_tests (loop);
1116 if (try_unroll_loop_completely (loop, exit, niter, ul, maxiter, locus))
1117 return true;
1119 if (create_iv
1120 && niter && !chrec_contains_undetermined (niter)
1121 && exit && just_once_each_iteration_p (loop, exit->src))
1122 create_canonical_iv (loop, exit, niter);
1124 if (ul == UL_ALL)
1125 modified |= try_peel_loop (loop, exit, niter, maxiter);
1127 return modified;
1130 /* The main entry point of the pass. Adds canonical induction variables
1131 to the suitable loops. */
1133 unsigned int
1134 canonicalize_induction_variables (void)
1136 struct loop *loop;
1137 bool changed = false;
1138 bool irred_invalidated = false;
1139 bitmap loop_closed_ssa_invalidated = BITMAP_ALLOC (NULL);
1141 free_numbers_of_iterations_estimates ();
1142 estimate_numbers_of_iterations ();
1144 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
1146 changed |= canonicalize_loop_induction_variables (loop,
1147 true, UL_SINGLE_ITER,
1148 true);
1150 gcc_assert (!need_ssa_update_p (cfun));
1152 unloop_loops (loop_closed_ssa_invalidated, &irred_invalidated);
1153 if (irred_invalidated
1154 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS))
1155 mark_irreducible_loops ();
1157 /* Clean up the information about numbers of iterations, since brute force
1158 evaluation could reveal new information. */
1159 scev_reset ();
1161 if (!bitmap_empty_p (loop_closed_ssa_invalidated))
1163 gcc_checking_assert (loops_state_satisfies_p (LOOP_CLOSED_SSA));
1164 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1166 BITMAP_FREE (loop_closed_ssa_invalidated);
1168 if (changed)
1169 return TODO_cleanup_cfg;
1170 return 0;
1173 /* Propagate VAL into all uses of SSA_NAME. */
1175 static void
1176 propagate_into_all_uses (tree ssa_name, tree val)
1178 imm_use_iterator iter;
1179 gimple *use_stmt;
1181 FOR_EACH_IMM_USE_STMT (use_stmt, iter, ssa_name)
1183 gimple_stmt_iterator use_stmt_gsi = gsi_for_stmt (use_stmt);
1184 use_operand_p use;
1186 FOR_EACH_IMM_USE_ON_STMT (use, iter)
1187 SET_USE (use, val);
1189 if (is_gimple_assign (use_stmt)
1190 && get_gimple_rhs_class (gimple_assign_rhs_code (use_stmt))
1191 == GIMPLE_SINGLE_RHS)
1193 tree rhs = gimple_assign_rhs1 (use_stmt);
1195 if (TREE_CODE (rhs) == ADDR_EXPR)
1196 recompute_tree_invariant_for_addr_expr (rhs);
1199 fold_stmt_inplace (&use_stmt_gsi);
1200 update_stmt (use_stmt);
1201 maybe_clean_or_replace_eh_stmt (use_stmt, use_stmt);
1205 /* Propagate constant SSA_NAMEs defined in basic block BB. */
1207 static void
1208 propagate_constants_for_unrolling (basic_block bb)
1210 /* Look for degenerate PHI nodes with constant argument. */
1211 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); )
1213 gphi *phi = gsi.phi ();
1214 tree result = gimple_phi_result (phi);
1215 tree arg = gimple_phi_arg_def (phi, 0);
1217 if (gimple_phi_num_args (phi) == 1 && TREE_CODE (arg) == INTEGER_CST)
1219 propagate_into_all_uses (result, arg);
1220 gsi_remove (&gsi, true);
1221 release_ssa_name (result);
1223 else
1224 gsi_next (&gsi);
1227 /* Look for assignments to SSA names with constant RHS. */
1228 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
1230 gimple *stmt = gsi_stmt (gsi);
1231 tree lhs;
1233 if (is_gimple_assign (stmt)
1234 && gimple_assign_rhs_code (stmt) == INTEGER_CST
1235 && (lhs = gimple_assign_lhs (stmt), TREE_CODE (lhs) == SSA_NAME)
1236 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
1238 propagate_into_all_uses (lhs, gimple_assign_rhs1 (stmt));
1239 gsi_remove (&gsi, true);
1240 release_ssa_name (lhs);
1242 else
1243 gsi_next (&gsi);
1247 /* Process loops from innermost to outer, stopping at the innermost
1248 loop we unrolled. */
1250 static bool
1251 tree_unroll_loops_completely_1 (bool may_increase_size, bool unroll_outer,
1252 vec<loop_p, va_heap>& father_stack,
1253 struct loop *loop)
1255 struct loop *loop_father;
1256 bool changed = false;
1257 struct loop *inner;
1258 enum unroll_level ul;
1260 /* Process inner loops first. */
1261 for (inner = loop->inner; inner != NULL; inner = inner->next)
1262 changed |= tree_unroll_loops_completely_1 (may_increase_size,
1263 unroll_outer, father_stack,
1264 inner);
1266 /* If we changed an inner loop we cannot process outer loops in this
1267 iteration because SSA form is not up-to-date. Continue with
1268 siblings of outer loops instead. */
1269 if (changed)
1270 return true;
1272 /* Don't unroll #pragma omp simd loops until the vectorizer
1273 attempts to vectorize those. */
1274 if (loop->force_vectorize)
1275 return false;
1277 /* Try to unroll this loop. */
1278 loop_father = loop_outer (loop);
1279 if (!loop_father)
1280 return false;
1282 if (may_increase_size && optimize_loop_nest_for_speed_p (loop)
1283 /* Unroll outermost loops only if asked to do so or they do
1284 not cause code growth. */
1285 && (unroll_outer || loop_outer (loop_father)))
1286 ul = UL_ALL;
1287 else
1288 ul = UL_NO_GROWTH;
1290 if (canonicalize_loop_induction_variables
1291 (loop, false, ul, !flag_tree_loop_ivcanon))
1293 /* If we'll continue unrolling, we need to propagate constants
1294 within the new basic blocks to fold away induction variable
1295 computations; otherwise, the size might blow up before the
1296 iteration is complete and the IR eventually cleaned up. */
1297 if (loop_outer (loop_father) && !loop_father->aux)
1299 father_stack.safe_push (loop_father);
1300 loop_father->aux = loop_father;
1303 return true;
1306 return false;
1309 /* Unroll LOOPS completely if they iterate just few times. Unless
1310 MAY_INCREASE_SIZE is true, perform the unrolling only if the
1311 size of the code does not increase. */
1313 unsigned int
1314 tree_unroll_loops_completely (bool may_increase_size, bool unroll_outer)
1316 auto_vec<loop_p, 16> father_stack;
1317 bool changed;
1318 int iteration = 0;
1319 bool irred_invalidated = false;
1323 changed = false;
1324 bitmap loop_closed_ssa_invalidated = NULL;
1326 if (loops_state_satisfies_p (LOOP_CLOSED_SSA))
1327 loop_closed_ssa_invalidated = BITMAP_ALLOC (NULL);
1329 free_numbers_of_iterations_estimates ();
1330 estimate_numbers_of_iterations ();
1332 changed = tree_unroll_loops_completely_1 (may_increase_size,
1333 unroll_outer, father_stack,
1334 current_loops->tree_root);
1335 if (changed)
1337 struct loop **iter;
1338 unsigned i;
1340 /* Be sure to skip unlooped loops while procesing father_stack
1341 array. */
1342 FOR_EACH_VEC_ELT (loops_to_unloop, i, iter)
1343 (*iter)->aux = NULL;
1344 FOR_EACH_VEC_ELT (father_stack, i, iter)
1345 if (!(*iter)->aux)
1346 *iter = NULL;
1347 unloop_loops (loop_closed_ssa_invalidated, &irred_invalidated);
1349 /* We can not use TODO_update_ssa_no_phi because VOPS gets confused. */
1350 if (loop_closed_ssa_invalidated
1351 && !bitmap_empty_p (loop_closed_ssa_invalidated))
1352 rewrite_into_loop_closed_ssa (loop_closed_ssa_invalidated,
1353 TODO_update_ssa);
1354 else
1355 update_ssa (TODO_update_ssa);
1357 /* Propagate the constants within the new basic blocks. */
1358 FOR_EACH_VEC_ELT (father_stack, i, iter)
1359 if (*iter)
1361 unsigned j;
1362 basic_block *body = get_loop_body_in_dom_order (*iter);
1363 for (j = 0; j < (*iter)->num_nodes; j++)
1364 propagate_constants_for_unrolling (body[j]);
1365 free (body);
1366 (*iter)->aux = NULL;
1368 father_stack.truncate (0);
1370 /* This will take care of removing completely unrolled loops
1371 from the loop structures so we can continue unrolling now
1372 innermost loops. */
1373 if (cleanup_tree_cfg ())
1374 update_ssa (TODO_update_ssa_only_virtuals);
1376 /* Clean up the information about numbers of iterations, since
1377 complete unrolling might have invalidated it. */
1378 scev_reset ();
1379 #ifdef ENABLE_CHECKING
1380 if (loops_state_satisfies_p (LOOP_CLOSED_SSA))
1381 verify_loop_closed_ssa (true);
1382 #endif
1384 if (loop_closed_ssa_invalidated)
1385 BITMAP_FREE (loop_closed_ssa_invalidated);
1387 while (changed
1388 && ++iteration <= PARAM_VALUE (PARAM_MAX_UNROLL_ITERATIONS));
1390 father_stack.release ();
1392 if (irred_invalidated
1393 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS))
1394 mark_irreducible_loops ();
1396 return 0;
1399 /* Canonical induction variable creation pass. */
1401 namespace {
1403 const pass_data pass_data_iv_canon =
1405 GIMPLE_PASS, /* type */
1406 "ivcanon", /* name */
1407 OPTGROUP_LOOP, /* optinfo_flags */
1408 TV_TREE_LOOP_IVCANON, /* tv_id */
1409 ( PROP_cfg | PROP_ssa ), /* properties_required */
1410 0, /* properties_provided */
1411 0, /* properties_destroyed */
1412 0, /* todo_flags_start */
1413 0, /* todo_flags_finish */
1416 class pass_iv_canon : public gimple_opt_pass
1418 public:
1419 pass_iv_canon (gcc::context *ctxt)
1420 : gimple_opt_pass (pass_data_iv_canon, ctxt)
1423 /* opt_pass methods: */
1424 virtual bool gate (function *) { return flag_tree_loop_ivcanon != 0; }
1425 virtual unsigned int execute (function *fun);
1427 }; // class pass_iv_canon
1429 unsigned int
1430 pass_iv_canon::execute (function *fun)
1432 if (number_of_loops (fun) <= 1)
1433 return 0;
1435 return canonicalize_induction_variables ();
1438 } // anon namespace
1440 gimple_opt_pass *
1441 make_pass_iv_canon (gcc::context *ctxt)
1443 return new pass_iv_canon (ctxt);
1446 /* Complete unrolling of loops. */
1448 namespace {
1450 const pass_data pass_data_complete_unroll =
1452 GIMPLE_PASS, /* type */
1453 "cunroll", /* name */
1454 OPTGROUP_LOOP, /* optinfo_flags */
1455 TV_COMPLETE_UNROLL, /* tv_id */
1456 ( PROP_cfg | PROP_ssa ), /* properties_required */
1457 0, /* properties_provided */
1458 0, /* properties_destroyed */
1459 0, /* todo_flags_start */
1460 0, /* todo_flags_finish */
1463 class pass_complete_unroll : public gimple_opt_pass
1465 public:
1466 pass_complete_unroll (gcc::context *ctxt)
1467 : gimple_opt_pass (pass_data_complete_unroll, ctxt)
1470 /* opt_pass methods: */
1471 virtual unsigned int execute (function *);
1473 }; // class pass_complete_unroll
1475 unsigned int
1476 pass_complete_unroll::execute (function *fun)
1478 if (number_of_loops (fun) <= 1)
1479 return 0;
1481 return tree_unroll_loops_completely (flag_unroll_loops
1482 || flag_peel_loops
1483 || optimize >= 3, true);
1486 } // anon namespace
1488 gimple_opt_pass *
1489 make_pass_complete_unroll (gcc::context *ctxt)
1491 return new pass_complete_unroll (ctxt);
1494 /* Complete unrolling of inner loops. */
1496 namespace {
1498 const pass_data pass_data_complete_unrolli =
1500 GIMPLE_PASS, /* type */
1501 "cunrolli", /* name */
1502 OPTGROUP_LOOP, /* optinfo_flags */
1503 TV_COMPLETE_UNROLL, /* tv_id */
1504 ( PROP_cfg | PROP_ssa ), /* properties_required */
1505 0, /* properties_provided */
1506 0, /* properties_destroyed */
1507 0, /* todo_flags_start */
1508 0, /* todo_flags_finish */
1511 class pass_complete_unrolli : public gimple_opt_pass
1513 public:
1514 pass_complete_unrolli (gcc::context *ctxt)
1515 : gimple_opt_pass (pass_data_complete_unrolli, ctxt)
1518 /* opt_pass methods: */
1519 virtual bool gate (function *) { return optimize >= 2; }
1520 virtual unsigned int execute (function *);
1522 }; // class pass_complete_unrolli
1524 unsigned int
1525 pass_complete_unrolli::execute (function *fun)
1527 unsigned ret = 0;
1529 loop_optimizer_init (LOOPS_NORMAL
1530 | LOOPS_HAVE_RECORDED_EXITS);
1531 if (number_of_loops (fun) > 1)
1533 scev_initialize ();
1534 ret = tree_unroll_loops_completely (optimize >= 3, false);
1535 free_numbers_of_iterations_estimates ();
1536 scev_finalize ();
1538 loop_optimizer_finalize ();
1540 return ret;
1543 } // anon namespace
1545 gimple_opt_pass *
1546 make_pass_complete_unrolli (gcc::context *ctxt)
1548 return new pass_complete_unrolli (ctxt);