decl.c (value_annotation_hasher::handle_cache_entry): Delete.
[gcc.git] / gcc / loop-doloop.c
1 /* Perform doloop optimizations
2 Copyright (C) 2004-2015 Free Software Foundation, Inc.
3 Based on code by Michael P. Hayes (m.hayes@elec.canterbury.ac.nz)
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
11
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
20
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "rtl.h"
26 #include "flags.h"
27 #include "symtab.h"
28 #include "hard-reg-set.h"
29 #include "function.h"
30 #include "alias.h"
31 #include "tree.h"
32 #include "insn-config.h"
33 #include "expmed.h"
34 #include "dojump.h"
35 #include "explow.h"
36 #include "calls.h"
37 #include "emit-rtl.h"
38 #include "varasm.h"
39 #include "stmt.h"
40 #include "expr.h"
41 #include "diagnostic-core.h"
42 #include "tm_p.h"
43 #include "predict.h"
44 #include "dominance.h"
45 #include "cfg.h"
46 #include "cfgloop.h"
47 #include "cfgrtl.h"
48 #include "basic-block.h"
49 #include "params.h"
50 #include "target.h"
51 #include "dumpfile.h"
52 #include "loop-unroll.h"
53
54 /* This module is used to modify loops with a determinable number of
55 iterations to use special low-overhead looping instructions.
56
57 It first validates whether the loop is well behaved and has a
58 determinable number of iterations (either at compile or run-time).
59 It then modifies the loop to use a low-overhead looping pattern as
60 follows:
61
62 1. A pseudo register is allocated as the loop iteration counter.
63
64 2. The number of loop iterations is calculated and is stored
65 in the loop counter.
66
67 3. At the end of the loop, the jump insn is replaced by the
68 doloop_end pattern. The compare must remain because it might be
69 used elsewhere. If the loop-variable or condition register are
70 used elsewhere, they will be eliminated by flow.
71
72 4. An optional doloop_begin pattern is inserted at the top of the
73 loop.
74
75 TODO The optimization should only performed when either the biv used for exit
76 condition is unused at all except for the exit test, or if we do not have to
77 change its value, since otherwise we have to add a new induction variable,
78 which usually will not pay up (unless the cost of the doloop pattern is
79 somehow extremely lower than the cost of compare & jump, or unless the bct
80 register cannot be used for anything else but doloop -- ??? detect these
81 cases). */
82
83 #ifdef HAVE_doloop_end
84
85 /* Return the loop termination condition for PATTERN or zero
86 if it is not a decrement and branch jump insn. */
87
88 rtx
89 doloop_condition_get (rtx doloop_pat)
90 {
91 rtx cmp;
92 rtx inc;
93 rtx reg;
94 rtx inc_src;
95 rtx condition;
96 rtx pattern;
97 rtx cc_reg = NULL_RTX;
98 rtx reg_orig = NULL_RTX;
99
100 /* The canonical doloop pattern we expect has one of the following
101 forms:
102
103 1) (parallel [(set (pc) (if_then_else (condition)
104 (label_ref (label))
105 (pc)))
106 (set (reg) (plus (reg) (const_int -1)))
107 (additional clobbers and uses)])
108
109 The branch must be the first entry of the parallel (also required
110 by jump.c), and the second entry of the parallel must be a set of
111 the loop counter register. Some targets (IA-64) wrap the set of
112 the loop counter in an if_then_else too.
113
114 2) (set (reg) (plus (reg) (const_int -1))
115 (set (pc) (if_then_else (reg != 0)
116 (label_ref (label))
117 (pc))).
118
119 Some targets (ARM) do the comparison before the branch, as in the
120 following form:
121
122 3) (parallel [(set (cc) (compare ((plus (reg) (const_int -1), 0)))
123 (set (reg) (plus (reg) (const_int -1)))])
124 (set (pc) (if_then_else (cc == NE)
125 (label_ref (label))
126 (pc))) */
127
128 pattern = PATTERN (doloop_pat);
129
130 if (GET_CODE (pattern) != PARALLEL)
131 {
132 rtx cond;
133 rtx_insn *prev_insn = prev_nondebug_insn (doloop_pat);
134 rtx cmp_arg1, cmp_arg2;
135 rtx cmp_orig;
136
137 /* In case the pattern is not PARALLEL we expect two forms
138 of doloop which are cases 2) and 3) above: in case 2) the
139 decrement immediately precedes the branch, while in case 3)
140 the compare and decrement instructions immediately precede
141 the branch. */
142
143 if (prev_insn == NULL_RTX || !INSN_P (prev_insn))
144 return 0;
145
146 cmp = pattern;
147 if (GET_CODE (PATTERN (prev_insn)) == PARALLEL)
148 {
149 /* The third case: the compare and decrement instructions
150 immediately precede the branch. */
151 cmp_orig = XVECEXP (PATTERN (prev_insn), 0, 0);
152 if (GET_CODE (cmp_orig) != SET)
153 return 0;
154 if (GET_CODE (SET_SRC (cmp_orig)) != COMPARE)
155 return 0;
156 cmp_arg1 = XEXP (SET_SRC (cmp_orig), 0);
157 cmp_arg2 = XEXP (SET_SRC (cmp_orig), 1);
158 if (cmp_arg2 != const0_rtx
159 || GET_CODE (cmp_arg1) != PLUS)
160 return 0;
161 reg_orig = XEXP (cmp_arg1, 0);
162 if (XEXP (cmp_arg1, 1) != GEN_INT (-1)
163 || !REG_P (reg_orig))
164 return 0;
165 cc_reg = SET_DEST (cmp_orig);
166
167 inc = XVECEXP (PATTERN (prev_insn), 0, 1);
168 }
169 else
170 inc = PATTERN (prev_insn);
171 /* We expect the condition to be of the form (reg != 0) */
172 cond = XEXP (SET_SRC (cmp), 0);
173 if (GET_CODE (cond) != NE || XEXP (cond, 1) != const0_rtx)
174 return 0;
175 }
176 else
177 {
178 cmp = XVECEXP (pattern, 0, 0);
179 inc = XVECEXP (pattern, 0, 1);
180 }
181
182 /* Check for (set (reg) (something)). */
183 if (GET_CODE (inc) != SET)
184 return 0;
185 reg = SET_DEST (inc);
186 if (! REG_P (reg))
187 return 0;
188
189 /* Check if something = (plus (reg) (const_int -1)).
190 On IA-64, this decrement is wrapped in an if_then_else. */
191 inc_src = SET_SRC (inc);
192 if (GET_CODE (inc_src) == IF_THEN_ELSE)
193 inc_src = XEXP (inc_src, 1);
194 if (GET_CODE (inc_src) != PLUS
195 || XEXP (inc_src, 0) != reg
196 || XEXP (inc_src, 1) != constm1_rtx)
197 return 0;
198
199 /* Check for (set (pc) (if_then_else (condition)
200 (label_ref (label))
201 (pc))). */
202 if (GET_CODE (cmp) != SET
203 || SET_DEST (cmp) != pc_rtx
204 || GET_CODE (SET_SRC (cmp)) != IF_THEN_ELSE
205 || GET_CODE (XEXP (SET_SRC (cmp), 1)) != LABEL_REF
206 || XEXP (SET_SRC (cmp), 2) != pc_rtx)
207 return 0;
208
209 /* Extract loop termination condition. */
210 condition = XEXP (SET_SRC (cmp), 0);
211
212 /* We expect a GE or NE comparison with 0 or 1. */
213 if ((GET_CODE (condition) != GE
214 && GET_CODE (condition) != NE)
215 || (XEXP (condition, 1) != const0_rtx
216 && XEXP (condition, 1) != const1_rtx))
217 return 0;
218
219 if ((XEXP (condition, 0) == reg)
220 /* For the third case: */
221 || ((cc_reg != NULL_RTX)
222 && (XEXP (condition, 0) == cc_reg)
223 && (reg_orig == reg))
224 || (GET_CODE (XEXP (condition, 0)) == PLUS
225 && XEXP (XEXP (condition, 0), 0) == reg))
226 {
227 if (GET_CODE (pattern) != PARALLEL)
228 /* For the second form we expect:
229
230 (set (reg) (plus (reg) (const_int -1))
231 (set (pc) (if_then_else (reg != 0)
232 (label_ref (label))
233 (pc))).
234
235 is equivalent to the following:
236
237 (parallel [(set (pc) (if_then_else (reg != 1)
238 (label_ref (label))
239 (pc)))
240 (set (reg) (plus (reg) (const_int -1)))
241 (additional clobbers and uses)])
242
243 For the third form we expect:
244
245 (parallel [(set (cc) (compare ((plus (reg) (const_int -1)), 0))
246 (set (reg) (plus (reg) (const_int -1)))])
247 (set (pc) (if_then_else (cc == NE)
248 (label_ref (label))
249 (pc)))
250
251 which is equivalent to the following:
252
253 (parallel [(set (cc) (compare (reg, 1))
254 (set (reg) (plus (reg) (const_int -1)))
255 (set (pc) (if_then_else (NE == cc)
256 (label_ref (label))
257 (pc))))])
258
259 So we return the second form instead for the two cases.
260
261 */
262 condition = gen_rtx_fmt_ee (NE, VOIDmode, inc_src, const1_rtx);
263
264 return condition;
265 }
266
267 /* ??? If a machine uses a funny comparison, we could return a
268 canonicalized form here. */
269
270 return 0;
271 }
272
273 /* Return nonzero if the loop specified by LOOP is suitable for
274 the use of special low-overhead looping instructions. DESC
275 describes the number of iterations of the loop. */
276
277 static bool
278 doloop_valid_p (struct loop *loop, struct niter_desc *desc)
279 {
280 basic_block *body = get_loop_body (loop), bb;
281 rtx_insn *insn;
282 unsigned i;
283 bool result = true;
284
285 /* Check for loops that may not terminate under special conditions. */
286 if (!desc->simple_p
287 || desc->assumptions
288 || desc->infinite)
289 {
290 /* There are some cases that would require a special attention.
291 For example if the comparison is LEU and the comparison value
292 is UINT_MAX then the loop will not terminate. Similarly, if the
293 comparison code is GEU and the comparison value is 0, the
294 loop will not terminate.
295
296 If the absolute increment is not 1, the loop can be infinite
297 even with LTU/GTU, e.g. for (i = 3; i > 0; i -= 2)
298
299 ??? We could compute these conditions at run-time and have a
300 additional jump around the loop to ensure an infinite loop.
301 However, it is very unlikely that this is the intended
302 behavior of the loop and checking for these rare boundary
303 conditions would pessimize all other code.
304
305 If the loop is executed only a few times an extra check to
306 restart the loop could use up most of the benefits of using a
307 count register loop. Note however, that normally, this
308 restart branch would never execute, so it could be predicted
309 well by the CPU. We should generate the pessimistic code by
310 default, and have an option, e.g. -funsafe-loops that would
311 enable count-register loops in this case. */
312 if (dump_file)
313 fprintf (dump_file, "Doloop: Possible infinite iteration case.\n");
314 result = false;
315 goto cleanup;
316 }
317
318 for (i = 0; i < loop->num_nodes; i++)
319 {
320 bb = body[i];
321
322 for (insn = BB_HEAD (bb);
323 insn != NEXT_INSN (BB_END (bb));
324 insn = NEXT_INSN (insn))
325 {
326 /* Different targets have different necessities for low-overhead
327 looping. Call the back end for each instruction within the loop
328 to let it decide whether the insn prohibits a low-overhead loop.
329 It will then return the cause for it to emit to the dump file. */
330 const char * invalid = targetm.invalid_within_doloop (insn);
331 if (invalid)
332 {
333 if (dump_file)
334 fprintf (dump_file, "Doloop: %s\n", invalid);
335 result = false;
336 goto cleanup;
337 }
338 }
339 }
340 result = true;
341
342 cleanup:
343 free (body);
344
345 return result;
346 }
347
348 /* Adds test of COND jumping to DEST on edge *E and set *E to the new fallthru
349 edge. If the condition is always false, do not do anything. If it is always
350 true, redirect E to DEST and return false. In all other cases, true is
351 returned. */
352
353 static bool
354 add_test (rtx cond, edge *e, basic_block dest)
355 {
356 rtx_insn *seq, *jump;
357 rtx_code_label *label;
358 machine_mode mode;
359 rtx op0 = XEXP (cond, 0), op1 = XEXP (cond, 1);
360 enum rtx_code code = GET_CODE (cond);
361 basic_block bb;
362
363 mode = GET_MODE (XEXP (cond, 0));
364 if (mode == VOIDmode)
365 mode = GET_MODE (XEXP (cond, 1));
366
367 start_sequence ();
368 op0 = force_operand (op0, NULL_RTX);
369 op1 = force_operand (op1, NULL_RTX);
370 label = block_label (dest);
371 do_compare_rtx_and_jump (op0, op1, code, 0, mode, NULL_RTX, NULL, label, -1);
372
373 jump = get_last_insn ();
374 if (!jump || !JUMP_P (jump))
375 {
376 /* The condition is always false and the jump was optimized out. */
377 end_sequence ();
378 return true;
379 }
380
381 seq = get_insns ();
382 end_sequence ();
383
384 /* There always is at least the jump insn in the sequence. */
385 gcc_assert (seq != NULL_RTX);
386
387 bb = split_edge_and_insert (*e, seq);
388 *e = single_succ_edge (bb);
389
390 if (any_uncondjump_p (jump))
391 {
392 /* The condition is always true. */
393 delete_insn (jump);
394 redirect_edge_and_branch_force (*e, dest);
395 return false;
396 }
397
398 JUMP_LABEL (jump) = label;
399
400 /* The jump is supposed to handle an unlikely special case. */
401 add_int_reg_note (jump, REG_BR_PROB, 0);
402
403 LABEL_NUSES (label)++;
404
405 make_edge (bb, dest, (*e)->flags & ~EDGE_FALLTHRU);
406 return true;
407 }
408
409 /* Modify the loop to use the low-overhead looping insn where LOOP
410 describes the loop, DESC describes the number of iterations of the
411 loop, and DOLOOP_INSN is the low-overhead looping insn to emit at the
412 end of the loop. CONDITION is the condition separated from the
413 DOLOOP_SEQ. COUNT is the number of iterations of the LOOP. */
414
415 static void
416 doloop_modify (struct loop *loop, struct niter_desc *desc,
417 rtx doloop_seq, rtx condition, rtx count)
418 {
419 rtx counter_reg;
420 rtx tmp, noloop = NULL_RTX;
421 rtx_insn *sequence;
422 rtx_insn *jump_insn;
423 rtx_code_label *jump_label;
424 int nonneg = 0;
425 bool increment_count;
426 basic_block loop_end = desc->out_edge->src;
427 machine_mode mode;
428 rtx true_prob_val;
429 widest_int iterations;
430
431 jump_insn = BB_END (loop_end);
432
433 if (dump_file)
434 {
435 fprintf (dump_file, "Doloop: Inserting doloop pattern (");
436 if (desc->const_iter)
437 fprintf (dump_file, "%" PRId64, desc->niter);
438 else
439 fputs ("runtime", dump_file);
440 fputs (" iterations).\n", dump_file);
441 }
442
443 /* Get the probability of the original branch. If it exists we would
444 need to update REG_BR_PROB of the new jump_insn. */
445 true_prob_val = find_reg_note (jump_insn, REG_BR_PROB, NULL_RTX);
446
447 /* Discard original jump to continue loop. The original compare
448 result may still be live, so it cannot be discarded explicitly. */
449 delete_insn (jump_insn);
450
451 counter_reg = XEXP (condition, 0);
452 if (GET_CODE (counter_reg) == PLUS)
453 counter_reg = XEXP (counter_reg, 0);
454 mode = GET_MODE (counter_reg);
455
456 increment_count = false;
457 switch (GET_CODE (condition))
458 {
459 case NE:
460 /* Currently only NE tests against zero and one are supported. */
461 noloop = XEXP (condition, 1);
462 if (noloop != const0_rtx)
463 {
464 gcc_assert (noloop == const1_rtx);
465 increment_count = true;
466 }
467 break;
468
469 case GE:
470 /* Currently only GE tests against zero are supported. */
471 gcc_assert (XEXP (condition, 1) == const0_rtx);
472
473 noloop = constm1_rtx;
474
475 /* The iteration count does not need incrementing for a GE test. */
476 increment_count = false;
477
478 /* Determine if the iteration counter will be non-negative.
479 Note that the maximum value loaded is iterations_max - 1. */
480 if (get_max_loop_iterations (loop, &iterations)
481 && wi::leu_p (iterations,
482 wi::set_bit_in_zero <widest_int>
483 (GET_MODE_PRECISION (mode) - 1)))
484 nonneg = 1;
485 break;
486
487 /* Abort if an invalid doloop pattern has been generated. */
488 default:
489 gcc_unreachable ();
490 }
491
492 if (increment_count)
493 count = simplify_gen_binary (PLUS, mode, count, const1_rtx);
494
495 /* Insert initialization of the count register into the loop header. */
496 start_sequence ();
497 tmp = force_operand (count, counter_reg);
498 convert_move (counter_reg, tmp, 1);
499 sequence = get_insns ();
500 end_sequence ();
501 emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src));
502
503 if (desc->noloop_assumptions)
504 {
505 rtx ass = copy_rtx (desc->noloop_assumptions);
506 basic_block preheader = loop_preheader_edge (loop)->src;
507 basic_block set_zero
508 = split_edge (loop_preheader_edge (loop));
509 basic_block new_preheader
510 = split_edge (loop_preheader_edge (loop));
511 edge te;
512
513 /* Expand the condition testing the assumptions and if it does not pass,
514 reset the count register to 0. */
515 redirect_edge_and_branch_force (single_succ_edge (preheader), new_preheader);
516 set_immediate_dominator (CDI_DOMINATORS, new_preheader, preheader);
517
518 set_zero->count = 0;
519 set_zero->frequency = 0;
520
521 te = single_succ_edge (preheader);
522 for (; ass; ass = XEXP (ass, 1))
523 if (!add_test (XEXP (ass, 0), &te, set_zero))
524 break;
525
526 if (ass)
527 {
528 /* We reached a condition that is always true. This is very hard to
529 reproduce (such a loop does not roll, and thus it would most
530 likely get optimized out by some of the preceding optimizations).
531 In fact, I do not have any testcase for it. However, it would
532 also be very hard to show that it is impossible, so we must
533 handle this case. */
534 set_zero->count = preheader->count;
535 set_zero->frequency = preheader->frequency;
536 }
537
538 if (EDGE_COUNT (set_zero->preds) == 0)
539 {
540 /* All the conditions were simplified to false, remove the
541 unreachable set_zero block. */
542 delete_basic_block (set_zero);
543 }
544 else
545 {
546 /* Reset the counter to zero in the set_zero block. */
547 start_sequence ();
548 convert_move (counter_reg, noloop, 0);
549 sequence = get_insns ();
550 end_sequence ();
551 emit_insn_after (sequence, BB_END (set_zero));
552
553 set_immediate_dominator (CDI_DOMINATORS, set_zero,
554 recompute_dominator (CDI_DOMINATORS,
555 set_zero));
556 }
557
558 set_immediate_dominator (CDI_DOMINATORS, new_preheader,
559 recompute_dominator (CDI_DOMINATORS,
560 new_preheader));
561 }
562
563 /* Some targets (eg, C4x) need to initialize special looping
564 registers. */
565 #ifdef HAVE_doloop_begin
566 {
567 rtx init;
568
569 init = gen_doloop_begin (counter_reg, doloop_seq);
570 if (init)
571 {
572 start_sequence ();
573 emit_insn (init);
574 sequence = get_insns ();
575 end_sequence ();
576 emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src));
577 }
578 }
579 #endif
580
581 /* Insert the new low-overhead looping insn. */
582 emit_jump_insn_after (doloop_seq, BB_END (loop_end));
583 jump_insn = BB_END (loop_end);
584 jump_label = block_label (desc->in_edge->dest);
585 JUMP_LABEL (jump_insn) = jump_label;
586 LABEL_NUSES (jump_label)++;
587
588 /* Ensure the right fallthru edge is marked, for case we have reversed
589 the condition. */
590 desc->in_edge->flags &= ~EDGE_FALLTHRU;
591 desc->out_edge->flags |= EDGE_FALLTHRU;
592
593 /* Add a REG_NONNEG note if the actual or estimated maximum number
594 of iterations is non-negative. */
595 if (nonneg)
596 add_reg_note (jump_insn, REG_NONNEG, NULL_RTX);
597
598 /* Update the REG_BR_PROB note. */
599 if (true_prob_val)
600 {
601 /* Seems safer to use the branch probability. */
602 add_int_reg_note (jump_insn, REG_BR_PROB, desc->in_edge->probability);
603 }
604 }
605
606 /* Process loop described by LOOP validating that the loop is suitable for
607 conversion to use a low overhead looping instruction, replacing the jump
608 insn where suitable. Returns true if the loop was successfully
609 modified. */
610
611 static bool
612 doloop_optimize (struct loop *loop)
613 {
614 machine_mode mode;
615 rtx doloop_seq, doloop_pat, doloop_reg;
616 rtx count;
617 widest_int iterations, iterations_max;
618 rtx_code_label *start_label;
619 rtx condition;
620 unsigned level, est_niter;
621 int max_cost;
622 struct niter_desc *desc;
623 unsigned word_mode_size;
624 unsigned HOST_WIDE_INT word_mode_max;
625 int entered_at_top;
626
627 if (dump_file)
628 fprintf (dump_file, "Doloop: Processing loop %d.\n", loop->num);
629
630 iv_analysis_loop_init (loop);
631
632 /* Find the simple exit of a LOOP. */
633 desc = get_simple_loop_desc (loop);
634
635 /* Check that loop is a candidate for a low-overhead looping insn. */
636 if (!doloop_valid_p (loop, desc))
637 {
638 if (dump_file)
639 fprintf (dump_file,
640 "Doloop: The loop is not suitable.\n");
641 return false;
642 }
643 mode = desc->mode;
644
645 est_niter = 3;
646 if (desc->const_iter)
647 est_niter = desc->niter;
648 /* If the estimate on number of iterations is reliable (comes from profile
649 feedback), use it. Do not use it normally, since the expected number
650 of iterations of an unrolled loop is 2. */
651 if (loop->header->count)
652 est_niter = expected_loop_iterations (loop);
653
654 if (est_niter < 3)
655 {
656 if (dump_file)
657 fprintf (dump_file,
658 "Doloop: Too few iterations (%u) to be profitable.\n",
659 est_niter);
660 return false;
661 }
662
663 max_cost
664 = COSTS_N_INSNS (PARAM_VALUE (PARAM_MAX_ITERATIONS_COMPUTATION_COST));
665 if (set_src_cost (desc->niter_expr, optimize_loop_for_speed_p (loop))
666 > max_cost)
667 {
668 if (dump_file)
669 fprintf (dump_file,
670 "Doloop: number of iterations too costly to compute.\n");
671 return false;
672 }
673
674 if (desc->const_iter)
675 iterations = widest_int::from (std::make_pair (desc->niter_expr, mode),
676 UNSIGNED);
677 else
678 iterations = 0;
679 if (!get_max_loop_iterations (loop, &iterations_max))
680 iterations_max = 0;
681 level = get_loop_level (loop) + 1;
682 entered_at_top = (loop->latch == desc->in_edge->dest
683 && contains_no_active_insn_p (loop->latch));
684 if (!targetm.can_use_doloop_p (iterations, iterations_max, level,
685 entered_at_top))
686 {
687 if (dump_file)
688 fprintf (dump_file, "Loop rejected by can_use_doloop_p.\n");
689 return false;
690 }
691
692 /* Generate looping insn. If the pattern FAILs then give up trying
693 to modify the loop since there is some aspect the back-end does
694 not like. */
695 count = copy_rtx (desc->niter_expr);
696 start_label = block_label (desc->in_edge->dest);
697 doloop_reg = gen_reg_rtx (mode);
698 doloop_seq = gen_doloop_end (doloop_reg, start_label);
699
700 word_mode_size = GET_MODE_PRECISION (word_mode);
701 word_mode_max
702 = ((unsigned HOST_WIDE_INT) 1 << (word_mode_size - 1) << 1) - 1;
703 if (! doloop_seq
704 && mode != word_mode
705 /* Before trying mode different from the one in that # of iterations is
706 computed, we must be sure that the number of iterations fits into
707 the new mode. */
708 && (word_mode_size >= GET_MODE_PRECISION (mode)
709 || wi::leu_p (iterations_max, word_mode_max)))
710 {
711 if (word_mode_size > GET_MODE_PRECISION (mode))
712 count = simplify_gen_unary (ZERO_EXTEND, word_mode, count, mode);
713 else
714 count = lowpart_subreg (word_mode, count, mode);
715 PUT_MODE (doloop_reg, word_mode);
716 doloop_seq = gen_doloop_end (doloop_reg, start_label);
717 }
718 if (! doloop_seq)
719 {
720 if (dump_file)
721 fprintf (dump_file,
722 "Doloop: Target unwilling to use doloop pattern!\n");
723 return false;
724 }
725
726 /* If multiple instructions were created, the last must be the
727 jump instruction. Also, a raw define_insn may yield a plain
728 pattern. */
729 doloop_pat = doloop_seq;
730 if (INSN_P (doloop_pat))
731 {
732 rtx_insn *doloop_insn = as_a <rtx_insn *> (doloop_pat);
733 while (NEXT_INSN (doloop_insn) != NULL_RTX)
734 doloop_insn = NEXT_INSN (doloop_insn);
735 if (!JUMP_P (doloop_insn))
736 doloop_insn = NULL;
737 doloop_pat = doloop_insn;
738 }
739
740 if (! doloop_pat
741 || ! (condition = doloop_condition_get (doloop_pat)))
742 {
743 if (dump_file)
744 fprintf (dump_file, "Doloop: Unrecognizable doloop pattern!\n");
745 return false;
746 }
747
748 doloop_modify (loop, desc, doloop_seq, condition, count);
749 return true;
750 }
751
752 /* This is the main entry point. Process all loops using doloop_optimize. */
753
754 void
755 doloop_optimize_loops (void)
756 {
757 struct loop *loop;
758
759 FOR_EACH_LOOP (loop, 0)
760 {
761 doloop_optimize (loop);
762 }
763
764 iv_analysis_done ();
765
766 #ifdef ENABLE_CHECKING
767 verify_loop_structure ();
768 #endif
769 }
770 #endif /* HAVE_doloop_end */
771