ifcvt.c (HAVE_cbranchcc4): Define.
[gcc.git] / gcc / ifcvt.c
1 /* If-conversion support.
2 Copyright (C) 2000-2014 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
8 the Free Software Foundation; either version 3, or (at your option)
9 any 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
13 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
14 License 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
25 #include "rtl.h"
26 #include "regs.h"
27 #include "hashtab.h"
28 #include "hash-set.h"
29 #include "vec.h"
30 #include "machmode.h"
31 #include "hard-reg-set.h"
32 #include "input.h"
33 #include "function.h"
34 #include "flags.h"
35 #include "insn-config.h"
36 #include "recog.h"
37 #include "except.h"
38 #include "predict.h"
39 #include "dominance.h"
40 #include "cfg.h"
41 #include "cfgrtl.h"
42 #include "cfganal.h"
43 #include "cfgcleanup.h"
44 #include "basic-block.h"
45 #include "expr.h"
46 #include "output.h"
47 #include "insn-codes.h"
48 #include "optabs.h"
49 #include "diagnostic-core.h"
50 #include "tm_p.h"
51 #include "cfgloop.h"
52 #include "target.h"
53 #include "tree-pass.h"
54 #include "df.h"
55 #include "dbgcnt.h"
56 #include "shrink-wrap.h"
57 #include "ifcvt.h"
58
59 #ifndef HAVE_conditional_move
60 #define HAVE_conditional_move 0
61 #endif
62 #ifndef HAVE_incscc
63 #define HAVE_incscc 0
64 #endif
65 #ifndef HAVE_decscc
66 #define HAVE_decscc 0
67 #endif
68 #ifndef HAVE_trap
69 #define HAVE_trap 0
70 #endif
71
72 #ifndef MAX_CONDITIONAL_EXECUTE
73 #define MAX_CONDITIONAL_EXECUTE \
74 (BRANCH_COST (optimize_function_for_speed_p (cfun), false) \
75 + 1)
76 #endif
77
78 #ifndef HAVE_cbranchcc4
79 #define HAVE_cbranchcc4 0
80 #endif
81
82 #define IFCVT_MULTIPLE_DUMPS 1
83
84 #define NULL_BLOCK ((basic_block) NULL)
85
86 /* True if after combine pass. */
87 static bool ifcvt_after_combine;
88
89 /* # of IF-THEN or IF-THEN-ELSE blocks we looked at */
90 static int num_possible_if_blocks;
91
92 /* # of IF-THEN or IF-THEN-ELSE blocks were converted to conditional
93 execution. */
94 static int num_updated_if_blocks;
95
96 /* # of changes made. */
97 static int num_true_changes;
98
99 /* Whether conditional execution changes were made. */
100 static int cond_exec_changed_p;
101
102 /* Forward references. */
103 static int count_bb_insns (const_basic_block);
104 static bool cheap_bb_rtx_cost_p (const_basic_block, int, int);
105 static rtx_insn *first_active_insn (basic_block);
106 static rtx_insn *last_active_insn (basic_block, int);
107 static rtx_insn *find_active_insn_before (basic_block, rtx_insn *);
108 static rtx_insn *find_active_insn_after (basic_block, rtx_insn *);
109 static basic_block block_fallthru (basic_block);
110 static int cond_exec_process_insns (ce_if_block *, rtx_insn *, rtx, rtx, int,
111 int);
112 static rtx cond_exec_get_condition (rtx_insn *);
113 static rtx noce_get_condition (rtx_insn *, rtx_insn **, bool);
114 static int noce_operand_ok (const_rtx);
115 static void merge_if_block (ce_if_block *);
116 static int find_cond_trap (basic_block, edge, edge);
117 static basic_block find_if_header (basic_block, int);
118 static int block_jumps_and_fallthru_p (basic_block, basic_block);
119 static int noce_find_if_block (basic_block, edge, edge, int);
120 static int cond_exec_find_if_block (ce_if_block *);
121 static int find_if_case_1 (basic_block, edge, edge);
122 static int find_if_case_2 (basic_block, edge, edge);
123 static int dead_or_predicable (basic_block, basic_block, basic_block,
124 edge, int);
125 static void noce_emit_move_insn (rtx, rtx);
126 static rtx_insn *block_has_only_trap (basic_block);
127 \f
128 /* Count the number of non-jump active insns in BB. */
129
130 static int
131 count_bb_insns (const_basic_block bb)
132 {
133 int count = 0;
134 rtx_insn *insn = BB_HEAD (bb);
135
136 while (1)
137 {
138 if (active_insn_p (insn) && !JUMP_P (insn))
139 count++;
140
141 if (insn == BB_END (bb))
142 break;
143 insn = NEXT_INSN (insn);
144 }
145
146 return count;
147 }
148
149 /* Determine whether the total insn_rtx_cost on non-jump insns in
150 basic block BB is less than MAX_COST. This function returns
151 false if the cost of any instruction could not be estimated.
152
153 The cost of the non-jump insns in BB is scaled by REG_BR_PROB_BASE
154 as those insns are being speculated. MAX_COST is scaled with SCALE
155 plus a small fudge factor. */
156
157 static bool
158 cheap_bb_rtx_cost_p (const_basic_block bb, int scale, int max_cost)
159 {
160 int count = 0;
161 rtx_insn *insn = BB_HEAD (bb);
162 bool speed = optimize_bb_for_speed_p (bb);
163
164 /* Set scale to REG_BR_PROB_BASE to void the identical scaling
165 applied to insn_rtx_cost when optimizing for size. Only do
166 this after combine because if-conversion might interfere with
167 passes before combine.
168
169 Use optimize_function_for_speed_p instead of the pre-defined
170 variable speed to make sure it is set to same value for all
171 basic blocks in one if-conversion transformation. */
172 if (!optimize_function_for_speed_p (cfun) && ifcvt_after_combine)
173 scale = REG_BR_PROB_BASE;
174 /* Our branch probability/scaling factors are just estimates and don't
175 account for cases where we can get speculation for free and other
176 secondary benefits. So we fudge the scale factor to make speculating
177 appear a little more profitable when optimizing for performance. */
178 else
179 scale += REG_BR_PROB_BASE / 8;
180
181
182 max_cost *= scale;
183
184 while (1)
185 {
186 if (NONJUMP_INSN_P (insn))
187 {
188 int cost = insn_rtx_cost (PATTERN (insn), speed) * REG_BR_PROB_BASE;
189 if (cost == 0)
190 return false;
191
192 /* If this instruction is the load or set of a "stack" register,
193 such as a floating point register on x87, then the cost of
194 speculatively executing this insn may need to include
195 the additional cost of popping its result off of the
196 register stack. Unfortunately, correctly recognizing and
197 accounting for this additional overhead is tricky, so for
198 now we simply prohibit such speculative execution. */
199 #ifdef STACK_REGS
200 {
201 rtx set = single_set (insn);
202 if (set && STACK_REG_P (SET_DEST (set)))
203 return false;
204 }
205 #endif
206
207 count += cost;
208 if (count >= max_cost)
209 return false;
210 }
211 else if (CALL_P (insn))
212 return false;
213
214 if (insn == BB_END (bb))
215 break;
216 insn = NEXT_INSN (insn);
217 }
218
219 return true;
220 }
221
222 /* Return the first non-jump active insn in the basic block. */
223
224 static rtx_insn *
225 first_active_insn (basic_block bb)
226 {
227 rtx_insn *insn = BB_HEAD (bb);
228
229 if (LABEL_P (insn))
230 {
231 if (insn == BB_END (bb))
232 return NULL;
233 insn = NEXT_INSN (insn);
234 }
235
236 while (NOTE_P (insn) || DEBUG_INSN_P (insn))
237 {
238 if (insn == BB_END (bb))
239 return NULL;
240 insn = NEXT_INSN (insn);
241 }
242
243 if (JUMP_P (insn))
244 return NULL;
245
246 return insn;
247 }
248
249 /* Return the last non-jump active (non-jump) insn in the basic block. */
250
251 static rtx_insn *
252 last_active_insn (basic_block bb, int skip_use_p)
253 {
254 rtx_insn *insn = BB_END (bb);
255 rtx_insn *head = BB_HEAD (bb);
256
257 while (NOTE_P (insn)
258 || JUMP_P (insn)
259 || DEBUG_INSN_P (insn)
260 || (skip_use_p
261 && NONJUMP_INSN_P (insn)
262 && GET_CODE (PATTERN (insn)) == USE))
263 {
264 if (insn == head)
265 return NULL;
266 insn = PREV_INSN (insn);
267 }
268
269 if (LABEL_P (insn))
270 return NULL;
271
272 return insn;
273 }
274
275 /* Return the active insn before INSN inside basic block CURR_BB. */
276
277 static rtx_insn *
278 find_active_insn_before (basic_block curr_bb, rtx_insn *insn)
279 {
280 if (!insn || insn == BB_HEAD (curr_bb))
281 return NULL;
282
283 while ((insn = PREV_INSN (insn)) != NULL_RTX)
284 {
285 if (NONJUMP_INSN_P (insn) || JUMP_P (insn) || CALL_P (insn))
286 break;
287
288 /* No other active insn all the way to the start of the basic block. */
289 if (insn == BB_HEAD (curr_bb))
290 return NULL;
291 }
292
293 return insn;
294 }
295
296 /* Return the active insn after INSN inside basic block CURR_BB. */
297
298 static rtx_insn *
299 find_active_insn_after (basic_block curr_bb, rtx_insn *insn)
300 {
301 if (!insn || insn == BB_END (curr_bb))
302 return NULL;
303
304 while ((insn = NEXT_INSN (insn)) != NULL_RTX)
305 {
306 if (NONJUMP_INSN_P (insn) || JUMP_P (insn) || CALL_P (insn))
307 break;
308
309 /* No other active insn all the way to the end of the basic block. */
310 if (insn == BB_END (curr_bb))
311 return NULL;
312 }
313
314 return insn;
315 }
316
317 /* Return the basic block reached by falling though the basic block BB. */
318
319 static basic_block
320 block_fallthru (basic_block bb)
321 {
322 edge e = find_fallthru_edge (bb->succs);
323
324 return (e) ? e->dest : NULL_BLOCK;
325 }
326
327 /* Return true if RTXs A and B can be safely interchanged. */
328
329 static bool
330 rtx_interchangeable_p (const_rtx a, const_rtx b)
331 {
332 if (!rtx_equal_p (a, b))
333 return false;
334
335 if (GET_CODE (a) != MEM)
336 return true;
337
338 /* A dead type-unsafe memory reference is legal, but a live type-unsafe memory
339 reference is not. Interchanging a dead type-unsafe memory reference with
340 a live type-safe one creates a live type-unsafe memory reference, in other
341 words, it makes the program illegal.
342 We check here conservatively whether the two memory references have equal
343 memory attributes. */
344
345 return mem_attrs_eq_p (get_mem_attrs (a), get_mem_attrs (b));
346 }
347
348 \f
349 /* Go through a bunch of insns, converting them to conditional
350 execution format if possible. Return TRUE if all of the non-note
351 insns were processed. */
352
353 static int
354 cond_exec_process_insns (ce_if_block *ce_info ATTRIBUTE_UNUSED,
355 /* if block information */rtx_insn *start,
356 /* first insn to look at */rtx end,
357 /* last insn to look at */rtx test,
358 /* conditional execution test */int prob_val,
359 /* probability of branch taken. */int mod_ok)
360 {
361 int must_be_last = FALSE;
362 rtx_insn *insn;
363 rtx xtest;
364 rtx pattern;
365
366 if (!start || !end)
367 return FALSE;
368
369 for (insn = start; ; insn = NEXT_INSN (insn))
370 {
371 /* dwarf2out can't cope with conditional prologues. */
372 if (NOTE_P (insn) && NOTE_KIND (insn) == NOTE_INSN_PROLOGUE_END)
373 return FALSE;
374
375 if (NOTE_P (insn) || DEBUG_INSN_P (insn))
376 goto insn_done;
377
378 gcc_assert (NONJUMP_INSN_P (insn) || CALL_P (insn));
379
380 /* dwarf2out can't cope with conditional unwind info. */
381 if (RTX_FRAME_RELATED_P (insn))
382 return FALSE;
383
384 /* Remove USE insns that get in the way. */
385 if (reload_completed && GET_CODE (PATTERN (insn)) == USE)
386 {
387 /* ??? Ug. Actually unlinking the thing is problematic,
388 given what we'd have to coordinate with our callers. */
389 SET_INSN_DELETED (insn);
390 goto insn_done;
391 }
392
393 /* Last insn wasn't last? */
394 if (must_be_last)
395 return FALSE;
396
397 if (modified_in_p (test, insn))
398 {
399 if (!mod_ok)
400 return FALSE;
401 must_be_last = TRUE;
402 }
403
404 /* Now build the conditional form of the instruction. */
405 pattern = PATTERN (insn);
406 xtest = copy_rtx (test);
407
408 /* If this is already a COND_EXEC, rewrite the test to be an AND of the
409 two conditions. */
410 if (GET_CODE (pattern) == COND_EXEC)
411 {
412 if (GET_MODE (xtest) != GET_MODE (COND_EXEC_TEST (pattern)))
413 return FALSE;
414
415 xtest = gen_rtx_AND (GET_MODE (xtest), xtest,
416 COND_EXEC_TEST (pattern));
417 pattern = COND_EXEC_CODE (pattern);
418 }
419
420 pattern = gen_rtx_COND_EXEC (VOIDmode, xtest, pattern);
421
422 /* If the machine needs to modify the insn being conditionally executed,
423 say for example to force a constant integer operand into a temp
424 register, do so here. */
425 #ifdef IFCVT_MODIFY_INSN
426 IFCVT_MODIFY_INSN (ce_info, pattern, insn);
427 if (! pattern)
428 return FALSE;
429 #endif
430
431 validate_change (insn, &PATTERN (insn), pattern, 1);
432
433 if (CALL_P (insn) && prob_val >= 0)
434 validate_change (insn, &REG_NOTES (insn),
435 gen_rtx_INT_LIST ((machine_mode) REG_BR_PROB,
436 prob_val, REG_NOTES (insn)), 1);
437
438 insn_done:
439 if (insn == end)
440 break;
441 }
442
443 return TRUE;
444 }
445
446 /* Return the condition for a jump. Do not do any special processing. */
447
448 static rtx
449 cond_exec_get_condition (rtx_insn *jump)
450 {
451 rtx test_if, cond;
452
453 if (any_condjump_p (jump))
454 test_if = SET_SRC (pc_set (jump));
455 else
456 return NULL_RTX;
457 cond = XEXP (test_if, 0);
458
459 /* If this branches to JUMP_LABEL when the condition is false,
460 reverse the condition. */
461 if (GET_CODE (XEXP (test_if, 2)) == LABEL_REF
462 && LABEL_REF_LABEL (XEXP (test_if, 2)) == JUMP_LABEL (jump))
463 {
464 enum rtx_code rev = reversed_comparison_code (cond, jump);
465 if (rev == UNKNOWN)
466 return NULL_RTX;
467
468 cond = gen_rtx_fmt_ee (rev, GET_MODE (cond), XEXP (cond, 0),
469 XEXP (cond, 1));
470 }
471
472 return cond;
473 }
474
475 /* Given a simple IF-THEN or IF-THEN-ELSE block, attempt to convert it
476 to conditional execution. Return TRUE if we were successful at
477 converting the block. */
478
479 static int
480 cond_exec_process_if_block (ce_if_block * ce_info,
481 /* if block information */int do_multiple_p)
482 {
483 basic_block test_bb = ce_info->test_bb; /* last test block */
484 basic_block then_bb = ce_info->then_bb; /* THEN */
485 basic_block else_bb = ce_info->else_bb; /* ELSE or NULL */
486 rtx test_expr; /* expression in IF_THEN_ELSE that is tested */
487 rtx_insn *then_start; /* first insn in THEN block */
488 rtx_insn *then_end; /* last insn + 1 in THEN block */
489 rtx_insn *else_start = NULL; /* first insn in ELSE block or NULL */
490 rtx_insn *else_end = NULL; /* last insn + 1 in ELSE block */
491 int max; /* max # of insns to convert. */
492 int then_mod_ok; /* whether conditional mods are ok in THEN */
493 rtx true_expr; /* test for else block insns */
494 rtx false_expr; /* test for then block insns */
495 int true_prob_val; /* probability of else block */
496 int false_prob_val; /* probability of then block */
497 rtx_insn *then_last_head = NULL; /* Last match at the head of THEN */
498 rtx_insn *else_last_head = NULL; /* Last match at the head of ELSE */
499 rtx_insn *then_first_tail = NULL; /* First match at the tail of THEN */
500 rtx_insn *else_first_tail = NULL; /* First match at the tail of ELSE */
501 int then_n_insns, else_n_insns, n_insns;
502 enum rtx_code false_code;
503 rtx note;
504
505 /* If test is comprised of && or || elements, and we've failed at handling
506 all of them together, just use the last test if it is the special case of
507 && elements without an ELSE block. */
508 if (!do_multiple_p && ce_info->num_multiple_test_blocks)
509 {
510 if (else_bb || ! ce_info->and_and_p)
511 return FALSE;
512
513 ce_info->test_bb = test_bb = ce_info->last_test_bb;
514 ce_info->num_multiple_test_blocks = 0;
515 ce_info->num_and_and_blocks = 0;
516 ce_info->num_or_or_blocks = 0;
517 }
518
519 /* Find the conditional jump to the ELSE or JOIN part, and isolate
520 the test. */
521 test_expr = cond_exec_get_condition (BB_END (test_bb));
522 if (! test_expr)
523 return FALSE;
524
525 /* If the conditional jump is more than just a conditional jump,
526 then we can not do conditional execution conversion on this block. */
527 if (! onlyjump_p (BB_END (test_bb)))
528 return FALSE;
529
530 /* Collect the bounds of where we're to search, skipping any labels, jumps
531 and notes at the beginning and end of the block. Then count the total
532 number of insns and see if it is small enough to convert. */
533 then_start = first_active_insn (then_bb);
534 then_end = last_active_insn (then_bb, TRUE);
535 then_n_insns = ce_info->num_then_insns = count_bb_insns (then_bb);
536 n_insns = then_n_insns;
537 max = MAX_CONDITIONAL_EXECUTE;
538
539 if (else_bb)
540 {
541 int n_matching;
542
543 max *= 2;
544 else_start = first_active_insn (else_bb);
545 else_end = last_active_insn (else_bb, TRUE);
546 else_n_insns = ce_info->num_else_insns = count_bb_insns (else_bb);
547 n_insns += else_n_insns;
548
549 /* Look for matching sequences at the head and tail of the two blocks,
550 and limit the range of insns to be converted if possible. */
551 n_matching = flow_find_cross_jump (then_bb, else_bb,
552 &then_first_tail, &else_first_tail,
553 NULL);
554 if (then_first_tail == BB_HEAD (then_bb))
555 then_start = then_end = NULL;
556 if (else_first_tail == BB_HEAD (else_bb))
557 else_start = else_end = NULL;
558
559 if (n_matching > 0)
560 {
561 if (then_end)
562 then_end = find_active_insn_before (then_bb, then_first_tail);
563 if (else_end)
564 else_end = find_active_insn_before (else_bb, else_first_tail);
565 n_insns -= 2 * n_matching;
566 }
567
568 if (then_start
569 && else_start
570 && then_n_insns > n_matching
571 && else_n_insns > n_matching)
572 {
573 int longest_match = MIN (then_n_insns - n_matching,
574 else_n_insns - n_matching);
575 n_matching
576 = flow_find_head_matching_sequence (then_bb, else_bb,
577 &then_last_head,
578 &else_last_head,
579 longest_match);
580
581 if (n_matching > 0)
582 {
583 rtx_insn *insn;
584
585 /* We won't pass the insns in the head sequence to
586 cond_exec_process_insns, so we need to test them here
587 to make sure that they don't clobber the condition. */
588 for (insn = BB_HEAD (then_bb);
589 insn != NEXT_INSN (then_last_head);
590 insn = NEXT_INSN (insn))
591 if (!LABEL_P (insn) && !NOTE_P (insn)
592 && !DEBUG_INSN_P (insn)
593 && modified_in_p (test_expr, insn))
594 return FALSE;
595 }
596
597 if (then_last_head == then_end)
598 then_start = then_end = NULL;
599 if (else_last_head == else_end)
600 else_start = else_end = NULL;
601
602 if (n_matching > 0)
603 {
604 if (then_start)
605 then_start = find_active_insn_after (then_bb, then_last_head);
606 if (else_start)
607 else_start = find_active_insn_after (else_bb, else_last_head);
608 n_insns -= 2 * n_matching;
609 }
610 }
611 }
612
613 if (n_insns > max)
614 return FALSE;
615
616 /* Map test_expr/test_jump into the appropriate MD tests to use on
617 the conditionally executed code. */
618
619 true_expr = test_expr;
620
621 false_code = reversed_comparison_code (true_expr, BB_END (test_bb));
622 if (false_code != UNKNOWN)
623 false_expr = gen_rtx_fmt_ee (false_code, GET_MODE (true_expr),
624 XEXP (true_expr, 0), XEXP (true_expr, 1));
625 else
626 false_expr = NULL_RTX;
627
628 #ifdef IFCVT_MODIFY_TESTS
629 /* If the machine description needs to modify the tests, such as setting a
630 conditional execution register from a comparison, it can do so here. */
631 IFCVT_MODIFY_TESTS (ce_info, true_expr, false_expr);
632
633 /* See if the conversion failed. */
634 if (!true_expr || !false_expr)
635 goto fail;
636 #endif
637
638 note = find_reg_note (BB_END (test_bb), REG_BR_PROB, NULL_RTX);
639 if (note)
640 {
641 true_prob_val = XINT (note, 0);
642 false_prob_val = REG_BR_PROB_BASE - true_prob_val;
643 }
644 else
645 {
646 true_prob_val = -1;
647 false_prob_val = -1;
648 }
649
650 /* If we have && or || tests, do them here. These tests are in the adjacent
651 blocks after the first block containing the test. */
652 if (ce_info->num_multiple_test_blocks > 0)
653 {
654 basic_block bb = test_bb;
655 basic_block last_test_bb = ce_info->last_test_bb;
656
657 if (! false_expr)
658 goto fail;
659
660 do
661 {
662 rtx_insn *start, *end;
663 rtx t, f;
664 enum rtx_code f_code;
665
666 bb = block_fallthru (bb);
667 start = first_active_insn (bb);
668 end = last_active_insn (bb, TRUE);
669 if (start
670 && ! cond_exec_process_insns (ce_info, start, end, false_expr,
671 false_prob_val, FALSE))
672 goto fail;
673
674 /* If the conditional jump is more than just a conditional jump, then
675 we can not do conditional execution conversion on this block. */
676 if (! onlyjump_p (BB_END (bb)))
677 goto fail;
678
679 /* Find the conditional jump and isolate the test. */
680 t = cond_exec_get_condition (BB_END (bb));
681 if (! t)
682 goto fail;
683
684 f_code = reversed_comparison_code (t, BB_END (bb));
685 if (f_code == UNKNOWN)
686 goto fail;
687
688 f = gen_rtx_fmt_ee (f_code, GET_MODE (t), XEXP (t, 0), XEXP (t, 1));
689 if (ce_info->and_and_p)
690 {
691 t = gen_rtx_AND (GET_MODE (t), true_expr, t);
692 f = gen_rtx_IOR (GET_MODE (t), false_expr, f);
693 }
694 else
695 {
696 t = gen_rtx_IOR (GET_MODE (t), true_expr, t);
697 f = gen_rtx_AND (GET_MODE (t), false_expr, f);
698 }
699
700 /* If the machine description needs to modify the tests, such as
701 setting a conditional execution register from a comparison, it can
702 do so here. */
703 #ifdef IFCVT_MODIFY_MULTIPLE_TESTS
704 IFCVT_MODIFY_MULTIPLE_TESTS (ce_info, bb, t, f);
705
706 /* See if the conversion failed. */
707 if (!t || !f)
708 goto fail;
709 #endif
710
711 true_expr = t;
712 false_expr = f;
713 }
714 while (bb != last_test_bb);
715 }
716
717 /* For IF-THEN-ELSE blocks, we don't allow modifications of the test
718 on then THEN block. */
719 then_mod_ok = (else_bb == NULL_BLOCK);
720
721 /* Go through the THEN and ELSE blocks converting the insns if possible
722 to conditional execution. */
723
724 if (then_end
725 && (! false_expr
726 || ! cond_exec_process_insns (ce_info, then_start, then_end,
727 false_expr, false_prob_val,
728 then_mod_ok)))
729 goto fail;
730
731 if (else_bb && else_end
732 && ! cond_exec_process_insns (ce_info, else_start, else_end,
733 true_expr, true_prob_val, TRUE))
734 goto fail;
735
736 /* If we cannot apply the changes, fail. Do not go through the normal fail
737 processing, since apply_change_group will call cancel_changes. */
738 if (! apply_change_group ())
739 {
740 #ifdef IFCVT_MODIFY_CANCEL
741 /* Cancel any machine dependent changes. */
742 IFCVT_MODIFY_CANCEL (ce_info);
743 #endif
744 return FALSE;
745 }
746
747 #ifdef IFCVT_MODIFY_FINAL
748 /* Do any machine dependent final modifications. */
749 IFCVT_MODIFY_FINAL (ce_info);
750 #endif
751
752 /* Conversion succeeded. */
753 if (dump_file)
754 fprintf (dump_file, "%d insn%s converted to conditional execution.\n",
755 n_insns, (n_insns == 1) ? " was" : "s were");
756
757 /* Merge the blocks! If we had matching sequences, make sure to delete one
758 copy at the appropriate location first: delete the copy in the THEN branch
759 for a tail sequence so that the remaining one is executed last for both
760 branches, and delete the copy in the ELSE branch for a head sequence so
761 that the remaining one is executed first for both branches. */
762 if (then_first_tail)
763 {
764 rtx_insn *from = then_first_tail;
765 if (!INSN_P (from))
766 from = find_active_insn_after (then_bb, from);
767 delete_insn_chain (from, BB_END (then_bb), false);
768 }
769 if (else_last_head)
770 delete_insn_chain (first_active_insn (else_bb), else_last_head, false);
771
772 merge_if_block (ce_info);
773 cond_exec_changed_p = TRUE;
774 return TRUE;
775
776 fail:
777 #ifdef IFCVT_MODIFY_CANCEL
778 /* Cancel any machine dependent changes. */
779 IFCVT_MODIFY_CANCEL (ce_info);
780 #endif
781
782 cancel_changes (0);
783 return FALSE;
784 }
785 \f
786 /* Used by noce_process_if_block to communicate with its subroutines.
787
788 The subroutines know that A and B may be evaluated freely. They
789 know that X is a register. They should insert new instructions
790 before cond_earliest. */
791
792 struct noce_if_info
793 {
794 /* The basic blocks that make up the IF-THEN-{ELSE-,}JOIN block. */
795 basic_block test_bb, then_bb, else_bb, join_bb;
796
797 /* The jump that ends TEST_BB. */
798 rtx_insn *jump;
799
800 /* The jump condition. */
801 rtx cond;
802
803 /* New insns should be inserted before this one. */
804 rtx_insn *cond_earliest;
805
806 /* Insns in the THEN and ELSE block. There is always just this
807 one insns in those blocks. The insns are single_set insns.
808 If there was no ELSE block, INSN_B is the last insn before
809 COND_EARLIEST, or NULL_RTX. In the former case, the insn
810 operands are still valid, as if INSN_B was moved down below
811 the jump. */
812 rtx_insn *insn_a, *insn_b;
813
814 /* The SET_SRC of INSN_A and INSN_B. */
815 rtx a, b;
816
817 /* The SET_DEST of INSN_A. */
818 rtx x;
819
820 /* True if this if block is not canonical. In the canonical form of
821 if blocks, the THEN_BB is the block reached via the fallthru edge
822 from TEST_BB. For the noce transformations, we allow the symmetric
823 form as well. */
824 bool then_else_reversed;
825
826 /* Estimated cost of the particular branch instruction. */
827 int branch_cost;
828 };
829
830 static rtx noce_emit_store_flag (struct noce_if_info *, rtx, int, int);
831 static int noce_try_move (struct noce_if_info *);
832 static int noce_try_store_flag (struct noce_if_info *);
833 static int noce_try_addcc (struct noce_if_info *);
834 static int noce_try_store_flag_constants (struct noce_if_info *);
835 static int noce_try_store_flag_mask (struct noce_if_info *);
836 static rtx noce_emit_cmove (struct noce_if_info *, rtx, enum rtx_code, rtx,
837 rtx, rtx, rtx);
838 static int noce_try_cmove (struct noce_if_info *);
839 static int noce_try_cmove_arith (struct noce_if_info *);
840 static rtx noce_get_alt_condition (struct noce_if_info *, rtx, rtx_insn **);
841 static int noce_try_minmax (struct noce_if_info *);
842 static int noce_try_abs (struct noce_if_info *);
843 static int noce_try_sign_mask (struct noce_if_info *);
844
845 /* Helper function for noce_try_store_flag*. */
846
847 static rtx
848 noce_emit_store_flag (struct noce_if_info *if_info, rtx x, int reversep,
849 int normalize)
850 {
851 rtx cond = if_info->cond;
852 int cond_complex;
853 enum rtx_code code;
854
855 cond_complex = (! general_operand (XEXP (cond, 0), VOIDmode)
856 || ! general_operand (XEXP (cond, 1), VOIDmode));
857
858 /* If earliest == jump, or when the condition is complex, try to
859 build the store_flag insn directly. */
860
861 if (cond_complex)
862 {
863 rtx set = pc_set (if_info->jump);
864 cond = XEXP (SET_SRC (set), 0);
865 if (GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
866 && LABEL_REF_LABEL (XEXP (SET_SRC (set), 2)) == JUMP_LABEL (if_info->jump))
867 reversep = !reversep;
868 if (if_info->then_else_reversed)
869 reversep = !reversep;
870 }
871
872 if (reversep)
873 code = reversed_comparison_code (cond, if_info->jump);
874 else
875 code = GET_CODE (cond);
876
877 if ((if_info->cond_earliest == if_info->jump || cond_complex)
878 && (normalize == 0 || STORE_FLAG_VALUE == normalize))
879 {
880 rtx src = gen_rtx_fmt_ee (code, GET_MODE (x), XEXP (cond, 0),
881 XEXP (cond, 1));
882 rtx set = gen_rtx_SET (VOIDmode, x, src);
883
884 start_sequence ();
885 rtx_insn *insn = emit_insn (set);
886
887 if (recog_memoized (insn) >= 0)
888 {
889 rtx_insn *seq = get_insns ();
890 end_sequence ();
891 emit_insn (seq);
892
893 if_info->cond_earliest = if_info->jump;
894
895 return x;
896 }
897
898 end_sequence ();
899 }
900
901 /* Don't even try if the comparison operands or the mode of X are weird. */
902 if (cond_complex || !SCALAR_INT_MODE_P (GET_MODE (x)))
903 return NULL_RTX;
904
905 return emit_store_flag (x, code, XEXP (cond, 0),
906 XEXP (cond, 1), VOIDmode,
907 (code == LTU || code == LEU
908 || code == GEU || code == GTU), normalize);
909 }
910
911 /* Emit instruction to move an rtx, possibly into STRICT_LOW_PART.
912 X is the destination/target and Y is the value to copy. */
913
914 static void
915 noce_emit_move_insn (rtx x, rtx y)
916 {
917 machine_mode outmode;
918 rtx outer, inner;
919 int bitpos;
920
921 if (GET_CODE (x) != STRICT_LOW_PART)
922 {
923 rtx_insn *seq, *insn;
924 rtx target;
925 optab ot;
926
927 start_sequence ();
928 /* Check that the SET_SRC is reasonable before calling emit_move_insn,
929 otherwise construct a suitable SET pattern ourselves. */
930 insn = (OBJECT_P (y) || CONSTANT_P (y) || GET_CODE (y) == SUBREG)
931 ? emit_move_insn (x, y)
932 : emit_insn (gen_rtx_SET (VOIDmode, x, y));
933 seq = get_insns ();
934 end_sequence ();
935
936 if (recog_memoized (insn) <= 0)
937 {
938 if (GET_CODE (x) == ZERO_EXTRACT)
939 {
940 rtx op = XEXP (x, 0);
941 unsigned HOST_WIDE_INT size = INTVAL (XEXP (x, 1));
942 unsigned HOST_WIDE_INT start = INTVAL (XEXP (x, 2));
943
944 /* store_bit_field expects START to be relative to
945 BYTES_BIG_ENDIAN and adjusts this value for machines with
946 BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN. In order to be able to
947 invoke store_bit_field again it is necessary to have the START
948 value from the first call. */
949 if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
950 {
951 if (MEM_P (op))
952 start = BITS_PER_UNIT - start - size;
953 else
954 {
955 gcc_assert (REG_P (op));
956 start = BITS_PER_WORD - start - size;
957 }
958 }
959
960 gcc_assert (start < (MEM_P (op) ? BITS_PER_UNIT : BITS_PER_WORD));
961 store_bit_field (op, size, start, 0, 0, GET_MODE (x), y);
962 return;
963 }
964
965 switch (GET_RTX_CLASS (GET_CODE (y)))
966 {
967 case RTX_UNARY:
968 ot = code_to_optab (GET_CODE (y));
969 if (ot)
970 {
971 start_sequence ();
972 target = expand_unop (GET_MODE (y), ot, XEXP (y, 0), x, 0);
973 if (target != NULL_RTX)
974 {
975 if (target != x)
976 emit_move_insn (x, target);
977 seq = get_insns ();
978 }
979 end_sequence ();
980 }
981 break;
982
983 case RTX_BIN_ARITH:
984 case RTX_COMM_ARITH:
985 ot = code_to_optab (GET_CODE (y));
986 if (ot)
987 {
988 start_sequence ();
989 target = expand_binop (GET_MODE (y), ot,
990 XEXP (y, 0), XEXP (y, 1),
991 x, 0, OPTAB_DIRECT);
992 if (target != NULL_RTX)
993 {
994 if (target != x)
995 emit_move_insn (x, target);
996 seq = get_insns ();
997 }
998 end_sequence ();
999 }
1000 break;
1001
1002 default:
1003 break;
1004 }
1005 }
1006
1007 emit_insn (seq);
1008 return;
1009 }
1010
1011 outer = XEXP (x, 0);
1012 inner = XEXP (outer, 0);
1013 outmode = GET_MODE (outer);
1014 bitpos = SUBREG_BYTE (outer) * BITS_PER_UNIT;
1015 store_bit_field (inner, GET_MODE_BITSIZE (outmode), bitpos,
1016 0, 0, outmode, y);
1017 }
1018
1019 /* Return sequence of instructions generated by if conversion. This
1020 function calls end_sequence() to end the current stream, ensures
1021 that are instructions are unshared, recognizable non-jump insns.
1022 On failure, this function returns a NULL_RTX. */
1023
1024 static rtx_insn *
1025 end_ifcvt_sequence (struct noce_if_info *if_info)
1026 {
1027 rtx_insn *insn;
1028 rtx_insn *seq = get_insns ();
1029
1030 set_used_flags (if_info->x);
1031 set_used_flags (if_info->cond);
1032 set_used_flags (if_info->a);
1033 set_used_flags (if_info->b);
1034 unshare_all_rtl_in_chain (seq);
1035 end_sequence ();
1036
1037 /* Make sure that all of the instructions emitted are recognizable,
1038 and that we haven't introduced a new jump instruction.
1039 As an exercise for the reader, build a general mechanism that
1040 allows proper placement of required clobbers. */
1041 for (insn = seq; insn; insn = NEXT_INSN (insn))
1042 if (JUMP_P (insn)
1043 || recog_memoized (insn) == -1)
1044 return NULL;
1045
1046 return seq;
1047 }
1048
1049 /* Convert "if (a != b) x = a; else x = b" into "x = a" and
1050 "if (a == b) x = a; else x = b" into "x = b". */
1051
1052 static int
1053 noce_try_move (struct noce_if_info *if_info)
1054 {
1055 rtx cond = if_info->cond;
1056 enum rtx_code code = GET_CODE (cond);
1057 rtx y;
1058 rtx_insn *seq;
1059
1060 if (code != NE && code != EQ)
1061 return FALSE;
1062
1063 /* This optimization isn't valid if either A or B could be a NaN
1064 or a signed zero. */
1065 if (HONOR_NANS (GET_MODE (if_info->x))
1066 || HONOR_SIGNED_ZEROS (GET_MODE (if_info->x)))
1067 return FALSE;
1068
1069 /* Check whether the operands of the comparison are A and in
1070 either order. */
1071 if ((rtx_equal_p (if_info->a, XEXP (cond, 0))
1072 && rtx_equal_p (if_info->b, XEXP (cond, 1)))
1073 || (rtx_equal_p (if_info->a, XEXP (cond, 1))
1074 && rtx_equal_p (if_info->b, XEXP (cond, 0))))
1075 {
1076 if (!rtx_interchangeable_p (if_info->a, if_info->b))
1077 return FALSE;
1078
1079 y = (code == EQ) ? if_info->a : if_info->b;
1080
1081 /* Avoid generating the move if the source is the destination. */
1082 if (! rtx_equal_p (if_info->x, y))
1083 {
1084 start_sequence ();
1085 noce_emit_move_insn (if_info->x, y);
1086 seq = end_ifcvt_sequence (if_info);
1087 if (!seq)
1088 return FALSE;
1089
1090 emit_insn_before_setloc (seq, if_info->jump,
1091 INSN_LOCATION (if_info->insn_a));
1092 }
1093 return TRUE;
1094 }
1095 return FALSE;
1096 }
1097
1098 /* Convert "if (test) x = 1; else x = 0".
1099
1100 Only try 0 and STORE_FLAG_VALUE here. Other combinations will be
1101 tried in noce_try_store_flag_constants after noce_try_cmove has had
1102 a go at the conversion. */
1103
1104 static int
1105 noce_try_store_flag (struct noce_if_info *if_info)
1106 {
1107 int reversep;
1108 rtx target;
1109 rtx_insn *seq;
1110
1111 if (CONST_INT_P (if_info->b)
1112 && INTVAL (if_info->b) == STORE_FLAG_VALUE
1113 && if_info->a == const0_rtx)
1114 reversep = 0;
1115 else if (if_info->b == const0_rtx
1116 && CONST_INT_P (if_info->a)
1117 && INTVAL (if_info->a) == STORE_FLAG_VALUE
1118 && (reversed_comparison_code (if_info->cond, if_info->jump)
1119 != UNKNOWN))
1120 reversep = 1;
1121 else
1122 return FALSE;
1123
1124 start_sequence ();
1125
1126 target = noce_emit_store_flag (if_info, if_info->x, reversep, 0);
1127 if (target)
1128 {
1129 if (target != if_info->x)
1130 noce_emit_move_insn (if_info->x, target);
1131
1132 seq = end_ifcvt_sequence (if_info);
1133 if (! seq)
1134 return FALSE;
1135
1136 emit_insn_before_setloc (seq, if_info->jump,
1137 INSN_LOCATION (if_info->insn_a));
1138 return TRUE;
1139 }
1140 else
1141 {
1142 end_sequence ();
1143 return FALSE;
1144 }
1145 }
1146
1147 /* Convert "if (test) x = a; else x = b", for A and B constant. */
1148
1149 static int
1150 noce_try_store_flag_constants (struct noce_if_info *if_info)
1151 {
1152 rtx target;
1153 rtx_insn *seq;
1154 int reversep;
1155 HOST_WIDE_INT itrue, ifalse, diff, tmp;
1156 int normalize, can_reverse;
1157 machine_mode mode;
1158
1159 if (CONST_INT_P (if_info->a)
1160 && CONST_INT_P (if_info->b))
1161 {
1162 mode = GET_MODE (if_info->x);
1163 ifalse = INTVAL (if_info->a);
1164 itrue = INTVAL (if_info->b);
1165
1166 diff = (unsigned HOST_WIDE_INT) itrue - ifalse;
1167 /* Make sure we can represent the difference between the two values. */
1168 if ((diff > 0)
1169 != ((ifalse < 0) != (itrue < 0) ? ifalse < 0 : ifalse < itrue))
1170 return FALSE;
1171
1172 diff = trunc_int_for_mode (diff, mode);
1173
1174 can_reverse = (reversed_comparison_code (if_info->cond, if_info->jump)
1175 != UNKNOWN);
1176
1177 reversep = 0;
1178 if (diff == STORE_FLAG_VALUE || diff == -STORE_FLAG_VALUE)
1179 normalize = 0;
1180 else if (ifalse == 0 && exact_log2 (itrue) >= 0
1181 && (STORE_FLAG_VALUE == 1
1182 || if_info->branch_cost >= 2))
1183 normalize = 1;
1184 else if (itrue == 0 && exact_log2 (ifalse) >= 0 && can_reverse
1185 && (STORE_FLAG_VALUE == 1 || if_info->branch_cost >= 2))
1186 normalize = 1, reversep = 1;
1187 else if (itrue == -1
1188 && (STORE_FLAG_VALUE == -1
1189 || if_info->branch_cost >= 2))
1190 normalize = -1;
1191 else if (ifalse == -1 && can_reverse
1192 && (STORE_FLAG_VALUE == -1 || if_info->branch_cost >= 2))
1193 normalize = -1, reversep = 1;
1194 else if ((if_info->branch_cost >= 2 && STORE_FLAG_VALUE == -1)
1195 || if_info->branch_cost >= 3)
1196 normalize = -1;
1197 else
1198 return FALSE;
1199
1200 if (reversep)
1201 {
1202 tmp = itrue; itrue = ifalse; ifalse = tmp;
1203 diff = trunc_int_for_mode (-(unsigned HOST_WIDE_INT) diff, mode);
1204 }
1205
1206 start_sequence ();
1207 target = noce_emit_store_flag (if_info, if_info->x, reversep, normalize);
1208 if (! target)
1209 {
1210 end_sequence ();
1211 return FALSE;
1212 }
1213
1214 /* if (test) x = 3; else x = 4;
1215 => x = 3 + (test == 0); */
1216 if (diff == STORE_FLAG_VALUE || diff == -STORE_FLAG_VALUE)
1217 {
1218 target = expand_simple_binop (mode,
1219 (diff == STORE_FLAG_VALUE
1220 ? PLUS : MINUS),
1221 gen_int_mode (ifalse, mode), target,
1222 if_info->x, 0, OPTAB_WIDEN);
1223 }
1224
1225 /* if (test) x = 8; else x = 0;
1226 => x = (test != 0) << 3; */
1227 else if (ifalse == 0 && (tmp = exact_log2 (itrue)) >= 0)
1228 {
1229 target = expand_simple_binop (mode, ASHIFT,
1230 target, GEN_INT (tmp), if_info->x, 0,
1231 OPTAB_WIDEN);
1232 }
1233
1234 /* if (test) x = -1; else x = b;
1235 => x = -(test != 0) | b; */
1236 else if (itrue == -1)
1237 {
1238 target = expand_simple_binop (mode, IOR,
1239 target, gen_int_mode (ifalse, mode),
1240 if_info->x, 0, OPTAB_WIDEN);
1241 }
1242
1243 /* if (test) x = a; else x = b;
1244 => x = (-(test != 0) & (b - a)) + a; */
1245 else
1246 {
1247 target = expand_simple_binop (mode, AND,
1248 target, gen_int_mode (diff, mode),
1249 if_info->x, 0, OPTAB_WIDEN);
1250 if (target)
1251 target = expand_simple_binop (mode, PLUS,
1252 target, gen_int_mode (ifalse, mode),
1253 if_info->x, 0, OPTAB_WIDEN);
1254 }
1255
1256 if (! target)
1257 {
1258 end_sequence ();
1259 return FALSE;
1260 }
1261
1262 if (target != if_info->x)
1263 noce_emit_move_insn (if_info->x, target);
1264
1265 seq = end_ifcvt_sequence (if_info);
1266 if (!seq)
1267 return FALSE;
1268
1269 emit_insn_before_setloc (seq, if_info->jump,
1270 INSN_LOCATION (if_info->insn_a));
1271 return TRUE;
1272 }
1273
1274 return FALSE;
1275 }
1276
1277 /* Convert "if (test) foo++" into "foo += (test != 0)", and
1278 similarly for "foo--". */
1279
1280 static int
1281 noce_try_addcc (struct noce_if_info *if_info)
1282 {
1283 rtx target;
1284 rtx_insn *seq;
1285 int subtract, normalize;
1286
1287 if (GET_CODE (if_info->a) == PLUS
1288 && rtx_equal_p (XEXP (if_info->a, 0), if_info->b)
1289 && (reversed_comparison_code (if_info->cond, if_info->jump)
1290 != UNKNOWN))
1291 {
1292 rtx cond = if_info->cond;
1293 enum rtx_code code = reversed_comparison_code (cond, if_info->jump);
1294
1295 /* First try to use addcc pattern. */
1296 if (general_operand (XEXP (cond, 0), VOIDmode)
1297 && general_operand (XEXP (cond, 1), VOIDmode))
1298 {
1299 start_sequence ();
1300 target = emit_conditional_add (if_info->x, code,
1301 XEXP (cond, 0),
1302 XEXP (cond, 1),
1303 VOIDmode,
1304 if_info->b,
1305 XEXP (if_info->a, 1),
1306 GET_MODE (if_info->x),
1307 (code == LTU || code == GEU
1308 || code == LEU || code == GTU));
1309 if (target)
1310 {
1311 if (target != if_info->x)
1312 noce_emit_move_insn (if_info->x, target);
1313
1314 seq = end_ifcvt_sequence (if_info);
1315 if (!seq)
1316 return FALSE;
1317
1318 emit_insn_before_setloc (seq, if_info->jump,
1319 INSN_LOCATION (if_info->insn_a));
1320 return TRUE;
1321 }
1322 end_sequence ();
1323 }
1324
1325 /* If that fails, construct conditional increment or decrement using
1326 setcc. */
1327 if (if_info->branch_cost >= 2
1328 && (XEXP (if_info->a, 1) == const1_rtx
1329 || XEXP (if_info->a, 1) == constm1_rtx))
1330 {
1331 start_sequence ();
1332 if (STORE_FLAG_VALUE == INTVAL (XEXP (if_info->a, 1)))
1333 subtract = 0, normalize = 0;
1334 else if (-STORE_FLAG_VALUE == INTVAL (XEXP (if_info->a, 1)))
1335 subtract = 1, normalize = 0;
1336 else
1337 subtract = 0, normalize = INTVAL (XEXP (if_info->a, 1));
1338
1339
1340 target = noce_emit_store_flag (if_info,
1341 gen_reg_rtx (GET_MODE (if_info->x)),
1342 1, normalize);
1343
1344 if (target)
1345 target = expand_simple_binop (GET_MODE (if_info->x),
1346 subtract ? MINUS : PLUS,
1347 if_info->b, target, if_info->x,
1348 0, OPTAB_WIDEN);
1349 if (target)
1350 {
1351 if (target != if_info->x)
1352 noce_emit_move_insn (if_info->x, target);
1353
1354 seq = end_ifcvt_sequence (if_info);
1355 if (!seq)
1356 return FALSE;
1357
1358 emit_insn_before_setloc (seq, if_info->jump,
1359 INSN_LOCATION (if_info->insn_a));
1360 return TRUE;
1361 }
1362 end_sequence ();
1363 }
1364 }
1365
1366 return FALSE;
1367 }
1368
1369 /* Convert "if (test) x = 0;" to "x &= -(test == 0);" */
1370
1371 static int
1372 noce_try_store_flag_mask (struct noce_if_info *if_info)
1373 {
1374 rtx target;
1375 rtx_insn *seq;
1376 int reversep;
1377
1378 reversep = 0;
1379 if ((if_info->branch_cost >= 2
1380 || STORE_FLAG_VALUE == -1)
1381 && ((if_info->a == const0_rtx
1382 && rtx_equal_p (if_info->b, if_info->x))
1383 || ((reversep = (reversed_comparison_code (if_info->cond,
1384 if_info->jump)
1385 != UNKNOWN))
1386 && if_info->b == const0_rtx
1387 && rtx_equal_p (if_info->a, if_info->x))))
1388 {
1389 start_sequence ();
1390 target = noce_emit_store_flag (if_info,
1391 gen_reg_rtx (GET_MODE (if_info->x)),
1392 reversep, -1);
1393 if (target)
1394 target = expand_simple_binop (GET_MODE (if_info->x), AND,
1395 if_info->x,
1396 target, if_info->x, 0,
1397 OPTAB_WIDEN);
1398
1399 if (target)
1400 {
1401 int old_cost, new_cost, insn_cost;
1402 int speed_p;
1403
1404 if (target != if_info->x)
1405 noce_emit_move_insn (if_info->x, target);
1406
1407 seq = end_ifcvt_sequence (if_info);
1408 if (!seq)
1409 return FALSE;
1410
1411 speed_p = optimize_bb_for_speed_p (BLOCK_FOR_INSN (if_info->insn_a));
1412 insn_cost = insn_rtx_cost (PATTERN (if_info->insn_a), speed_p);
1413 old_cost = COSTS_N_INSNS (if_info->branch_cost) + insn_cost;
1414 new_cost = seq_cost (seq, speed_p);
1415
1416 if (new_cost > old_cost)
1417 return FALSE;
1418
1419 emit_insn_before_setloc (seq, if_info->jump,
1420 INSN_LOCATION (if_info->insn_a));
1421 return TRUE;
1422 }
1423
1424 end_sequence ();
1425 }
1426
1427 return FALSE;
1428 }
1429
1430 /* Helper function for noce_try_cmove and noce_try_cmove_arith. */
1431
1432 static rtx
1433 noce_emit_cmove (struct noce_if_info *if_info, rtx x, enum rtx_code code,
1434 rtx cmp_a, rtx cmp_b, rtx vfalse, rtx vtrue)
1435 {
1436 rtx target ATTRIBUTE_UNUSED;
1437 int unsignedp ATTRIBUTE_UNUSED;
1438
1439 /* If earliest == jump, try to build the cmove insn directly.
1440 This is helpful when combine has created some complex condition
1441 (like for alpha's cmovlbs) that we can't hope to regenerate
1442 through the normal interface. */
1443
1444 if (if_info->cond_earliest == if_info->jump)
1445 {
1446 rtx cond = gen_rtx_fmt_ee (code, GET_MODE (if_info->cond), cmp_a, cmp_b);
1447 rtx if_then_else = gen_rtx_IF_THEN_ELSE (GET_MODE (x),
1448 cond, vtrue, vfalse);
1449 rtx set = gen_rtx_SET (VOIDmode, x, if_then_else);
1450
1451 start_sequence ();
1452 rtx_insn *insn = emit_insn (set);
1453
1454 if (recog_memoized (insn) >= 0)
1455 {
1456 rtx_insn *seq = get_insns ();
1457 end_sequence ();
1458 emit_insn (seq);
1459
1460 return x;
1461 }
1462
1463 end_sequence ();
1464 }
1465
1466 /* Don't even try if the comparison operands are weird
1467 except that the target supports cbranchcc4. */
1468 if (! general_operand (cmp_a, GET_MODE (cmp_a))
1469 || ! general_operand (cmp_b, GET_MODE (cmp_b)))
1470 {
1471 if (!(HAVE_cbranchcc4)
1472 || GET_MODE_CLASS (GET_MODE (cmp_a)) != MODE_CC
1473 || cmp_b != const0_rtx)
1474 return NULL_RTX;
1475 }
1476
1477 #if HAVE_conditional_move
1478 unsignedp = (code == LTU || code == GEU
1479 || code == LEU || code == GTU);
1480
1481 target = emit_conditional_move (x, code, cmp_a, cmp_b, VOIDmode,
1482 vtrue, vfalse, GET_MODE (x),
1483 unsignedp);
1484 if (target)
1485 return target;
1486
1487 /* We might be faced with a situation like:
1488
1489 x = (reg:M TARGET)
1490 vtrue = (subreg:M (reg:N VTRUE) BYTE)
1491 vfalse = (subreg:M (reg:N VFALSE) BYTE)
1492
1493 We can't do a conditional move in mode M, but it's possible that we
1494 could do a conditional move in mode N instead and take a subreg of
1495 the result.
1496
1497 If we can't create new pseudos, though, don't bother. */
1498 if (reload_completed)
1499 return NULL_RTX;
1500
1501 if (GET_CODE (vtrue) == SUBREG && GET_CODE (vfalse) == SUBREG)
1502 {
1503 rtx reg_vtrue = SUBREG_REG (vtrue);
1504 rtx reg_vfalse = SUBREG_REG (vfalse);
1505 unsigned int byte_vtrue = SUBREG_BYTE (vtrue);
1506 unsigned int byte_vfalse = SUBREG_BYTE (vfalse);
1507 rtx promoted_target;
1508
1509 if (GET_MODE (reg_vtrue) != GET_MODE (reg_vfalse)
1510 || byte_vtrue != byte_vfalse
1511 || (SUBREG_PROMOTED_VAR_P (vtrue)
1512 != SUBREG_PROMOTED_VAR_P (vfalse))
1513 || (SUBREG_PROMOTED_GET (vtrue)
1514 != SUBREG_PROMOTED_GET (vfalse)))
1515 return NULL_RTX;
1516
1517 promoted_target = gen_reg_rtx (GET_MODE (reg_vtrue));
1518
1519 target = emit_conditional_move (promoted_target, code, cmp_a, cmp_b,
1520 VOIDmode, reg_vtrue, reg_vfalse,
1521 GET_MODE (reg_vtrue), unsignedp);
1522 /* Nope, couldn't do it in that mode either. */
1523 if (!target)
1524 return NULL_RTX;
1525
1526 target = gen_rtx_SUBREG (GET_MODE (vtrue), promoted_target, byte_vtrue);
1527 SUBREG_PROMOTED_VAR_P (target) = SUBREG_PROMOTED_VAR_P (vtrue);
1528 SUBREG_PROMOTED_SET (target, SUBREG_PROMOTED_GET (vtrue));
1529 emit_move_insn (x, target);
1530 return x;
1531 }
1532 else
1533 return NULL_RTX;
1534 #else
1535 /* We'll never get here, as noce_process_if_block doesn't call the
1536 functions involved. Ifdef code, however, should be discouraged
1537 because it leads to typos in the code not selected. However,
1538 emit_conditional_move won't exist either. */
1539 return NULL_RTX;
1540 #endif
1541 }
1542
1543 /* Try only simple constants and registers here. More complex cases
1544 are handled in noce_try_cmove_arith after noce_try_store_flag_arith
1545 has had a go at it. */
1546
1547 static int
1548 noce_try_cmove (struct noce_if_info *if_info)
1549 {
1550 enum rtx_code code;
1551 rtx target;
1552 rtx_insn *seq;
1553
1554 if ((CONSTANT_P (if_info->a) || register_operand (if_info->a, VOIDmode))
1555 && (CONSTANT_P (if_info->b) || register_operand (if_info->b, VOIDmode)))
1556 {
1557 start_sequence ();
1558
1559 code = GET_CODE (if_info->cond);
1560 target = noce_emit_cmove (if_info, if_info->x, code,
1561 XEXP (if_info->cond, 0),
1562 XEXP (if_info->cond, 1),
1563 if_info->a, if_info->b);
1564
1565 if (target)
1566 {
1567 if (target != if_info->x)
1568 noce_emit_move_insn (if_info->x, target);
1569
1570 seq = end_ifcvt_sequence (if_info);
1571 if (!seq)
1572 return FALSE;
1573
1574 emit_insn_before_setloc (seq, if_info->jump,
1575 INSN_LOCATION (if_info->insn_a));
1576 return TRUE;
1577 }
1578 else
1579 {
1580 end_sequence ();
1581 return FALSE;
1582 }
1583 }
1584
1585 return FALSE;
1586 }
1587
1588 /* Try more complex cases involving conditional_move. */
1589
1590 static int
1591 noce_try_cmove_arith (struct noce_if_info *if_info)
1592 {
1593 rtx a = if_info->a;
1594 rtx b = if_info->b;
1595 rtx x = if_info->x;
1596 rtx orig_a, orig_b;
1597 rtx_insn *insn_a, *insn_b;
1598 rtx target;
1599 int is_mem = 0;
1600 int insn_cost;
1601 enum rtx_code code;
1602 rtx_insn *ifcvt_seq;
1603
1604 /* A conditional move from two memory sources is equivalent to a
1605 conditional on their addresses followed by a load. Don't do this
1606 early because it'll screw alias analysis. Note that we've
1607 already checked for no side effects. */
1608 /* ??? FIXME: Magic number 5. */
1609 if (cse_not_expected
1610 && MEM_P (a) && MEM_P (b)
1611 && MEM_ADDR_SPACE (a) == MEM_ADDR_SPACE (b)
1612 && if_info->branch_cost >= 5)
1613 {
1614 machine_mode address_mode = get_address_mode (a);
1615
1616 a = XEXP (a, 0);
1617 b = XEXP (b, 0);
1618 x = gen_reg_rtx (address_mode);
1619 is_mem = 1;
1620 }
1621
1622 /* ??? We could handle this if we knew that a load from A or B could
1623 not trap or fault. This is also true if we've already loaded
1624 from the address along the path from ENTRY. */
1625 else if (may_trap_or_fault_p (a) || may_trap_or_fault_p (b))
1626 return FALSE;
1627
1628 /* if (test) x = a + b; else x = c - d;
1629 => y = a + b;
1630 x = c - d;
1631 if (test)
1632 x = y;
1633 */
1634
1635 code = GET_CODE (if_info->cond);
1636 insn_a = if_info->insn_a;
1637 insn_b = if_info->insn_b;
1638
1639 /* Total insn_rtx_cost should be smaller than branch cost. Exit
1640 if insn_rtx_cost can't be estimated. */
1641 if (insn_a)
1642 {
1643 insn_cost
1644 = insn_rtx_cost (PATTERN (insn_a),
1645 optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn_a)));
1646 if (insn_cost == 0 || insn_cost > COSTS_N_INSNS (if_info->branch_cost))
1647 return FALSE;
1648 }
1649 else
1650 insn_cost = 0;
1651
1652 if (insn_b)
1653 {
1654 insn_cost
1655 += insn_rtx_cost (PATTERN (insn_b),
1656 optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn_b)));
1657 if (insn_cost == 0 || insn_cost > COSTS_N_INSNS (if_info->branch_cost))
1658 return FALSE;
1659 }
1660
1661 /* Possibly rearrange operands to make things come out more natural. */
1662 if (reversed_comparison_code (if_info->cond, if_info->jump) != UNKNOWN)
1663 {
1664 int reversep = 0;
1665 if (rtx_equal_p (b, x))
1666 reversep = 1;
1667 else if (general_operand (b, GET_MODE (b)))
1668 reversep = 1;
1669
1670 if (reversep)
1671 {
1672 rtx tmp;
1673 rtx_insn *tmp_insn;
1674 code = reversed_comparison_code (if_info->cond, if_info->jump);
1675 tmp = a, a = b, b = tmp;
1676 tmp_insn = insn_a, insn_a = insn_b, insn_b = tmp_insn;
1677 }
1678 }
1679
1680 start_sequence ();
1681
1682 orig_a = a;
1683 orig_b = b;
1684
1685 /* If either operand is complex, load it into a register first.
1686 The best way to do this is to copy the original insn. In this
1687 way we preserve any clobbers etc that the insn may have had.
1688 This is of course not possible in the IS_MEM case. */
1689 if (! general_operand (a, GET_MODE (a)))
1690 {
1691 rtx_insn *insn;
1692
1693 if (is_mem)
1694 {
1695 rtx reg = gen_reg_rtx (GET_MODE (a));
1696 insn = emit_insn (gen_rtx_SET (VOIDmode, reg, a));
1697 }
1698 else if (! insn_a)
1699 goto end_seq_and_fail;
1700 else
1701 {
1702 a = gen_reg_rtx (GET_MODE (a));
1703 rtx_insn *copy_of_a = as_a <rtx_insn *> (copy_rtx (insn_a));
1704 rtx set = single_set (copy_of_a);
1705 SET_DEST (set) = a;
1706 insn = emit_insn (PATTERN (copy_of_a));
1707 }
1708 if (recog_memoized (insn) < 0)
1709 goto end_seq_and_fail;
1710 }
1711 if (! general_operand (b, GET_MODE (b)))
1712 {
1713 rtx pat;
1714 rtx_insn *last;
1715 rtx_insn *new_insn;
1716
1717 if (is_mem)
1718 {
1719 rtx reg = gen_reg_rtx (GET_MODE (b));
1720 pat = gen_rtx_SET (VOIDmode, reg, b);
1721 }
1722 else if (! insn_b)
1723 goto end_seq_and_fail;
1724 else
1725 {
1726 b = gen_reg_rtx (GET_MODE (b));
1727 rtx_insn *copy_of_insn_b = as_a <rtx_insn *> (copy_rtx (insn_b));
1728 rtx set = single_set (copy_of_insn_b);
1729 SET_DEST (set) = b;
1730 pat = PATTERN (copy_of_insn_b);
1731 }
1732
1733 /* If insn to set up A clobbers any registers B depends on, try to
1734 swap insn that sets up A with the one that sets up B. If even
1735 that doesn't help, punt. */
1736 last = get_last_insn ();
1737 if (last && modified_in_p (orig_b, last))
1738 {
1739 new_insn = emit_insn_before (pat, get_insns ());
1740 if (modified_in_p (orig_a, new_insn))
1741 goto end_seq_and_fail;
1742 }
1743 else
1744 new_insn = emit_insn (pat);
1745
1746 if (recog_memoized (new_insn) < 0)
1747 goto end_seq_and_fail;
1748 }
1749
1750 target = noce_emit_cmove (if_info, x, code, XEXP (if_info->cond, 0),
1751 XEXP (if_info->cond, 1), a, b);
1752
1753 if (! target)
1754 goto end_seq_and_fail;
1755
1756 /* If we're handling a memory for above, emit the load now. */
1757 if (is_mem)
1758 {
1759 rtx mem = gen_rtx_MEM (GET_MODE (if_info->x), target);
1760
1761 /* Copy over flags as appropriate. */
1762 if (MEM_VOLATILE_P (if_info->a) || MEM_VOLATILE_P (if_info->b))
1763 MEM_VOLATILE_P (mem) = 1;
1764 if (MEM_ALIAS_SET (if_info->a) == MEM_ALIAS_SET (if_info->b))
1765 set_mem_alias_set (mem, MEM_ALIAS_SET (if_info->a));
1766 set_mem_align (mem,
1767 MIN (MEM_ALIGN (if_info->a), MEM_ALIGN (if_info->b)));
1768
1769 gcc_assert (MEM_ADDR_SPACE (if_info->a) == MEM_ADDR_SPACE (if_info->b));
1770 set_mem_addr_space (mem, MEM_ADDR_SPACE (if_info->a));
1771
1772 noce_emit_move_insn (if_info->x, mem);
1773 }
1774 else if (target != x)
1775 noce_emit_move_insn (x, target);
1776
1777 ifcvt_seq = end_ifcvt_sequence (if_info);
1778 if (!ifcvt_seq)
1779 return FALSE;
1780
1781 emit_insn_before_setloc (ifcvt_seq, if_info->jump,
1782 INSN_LOCATION (if_info->insn_a));
1783 return TRUE;
1784
1785 end_seq_and_fail:
1786 end_sequence ();
1787 return FALSE;
1788 }
1789
1790 /* For most cases, the simplified condition we found is the best
1791 choice, but this is not the case for the min/max/abs transforms.
1792 For these we wish to know that it is A or B in the condition. */
1793
1794 static rtx
1795 noce_get_alt_condition (struct noce_if_info *if_info, rtx target,
1796 rtx_insn **earliest)
1797 {
1798 rtx cond, set;
1799 rtx_insn *insn;
1800 int reverse;
1801
1802 /* If target is already mentioned in the known condition, return it. */
1803 if (reg_mentioned_p (target, if_info->cond))
1804 {
1805 *earliest = if_info->cond_earliest;
1806 return if_info->cond;
1807 }
1808
1809 set = pc_set (if_info->jump);
1810 cond = XEXP (SET_SRC (set), 0);
1811 reverse
1812 = GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
1813 && LABEL_REF_LABEL (XEXP (SET_SRC (set), 2)) == JUMP_LABEL (if_info->jump);
1814 if (if_info->then_else_reversed)
1815 reverse = !reverse;
1816
1817 /* If we're looking for a constant, try to make the conditional
1818 have that constant in it. There are two reasons why it may
1819 not have the constant we want:
1820
1821 1. GCC may have needed to put the constant in a register, because
1822 the target can't compare directly against that constant. For
1823 this case, we look for a SET immediately before the comparison
1824 that puts a constant in that register.
1825
1826 2. GCC may have canonicalized the conditional, for example
1827 replacing "if x < 4" with "if x <= 3". We can undo that (or
1828 make equivalent types of changes) to get the constants we need
1829 if they're off by one in the right direction. */
1830
1831 if (CONST_INT_P (target))
1832 {
1833 enum rtx_code code = GET_CODE (if_info->cond);
1834 rtx op_a = XEXP (if_info->cond, 0);
1835 rtx op_b = XEXP (if_info->cond, 1);
1836 rtx prev_insn;
1837
1838 /* First, look to see if we put a constant in a register. */
1839 prev_insn = prev_nonnote_insn (if_info->cond_earliest);
1840 if (prev_insn
1841 && BLOCK_FOR_INSN (prev_insn)
1842 == BLOCK_FOR_INSN (if_info->cond_earliest)
1843 && INSN_P (prev_insn)
1844 && GET_CODE (PATTERN (prev_insn)) == SET)
1845 {
1846 rtx src = find_reg_equal_equiv_note (prev_insn);
1847 if (!src)
1848 src = SET_SRC (PATTERN (prev_insn));
1849 if (CONST_INT_P (src))
1850 {
1851 if (rtx_equal_p (op_a, SET_DEST (PATTERN (prev_insn))))
1852 op_a = src;
1853 else if (rtx_equal_p (op_b, SET_DEST (PATTERN (prev_insn))))
1854 op_b = src;
1855
1856 if (CONST_INT_P (op_a))
1857 {
1858 rtx tmp = op_a;
1859 op_a = op_b;
1860 op_b = tmp;
1861 code = swap_condition (code);
1862 }
1863 }
1864 }
1865
1866 /* Now, look to see if we can get the right constant by
1867 adjusting the conditional. */
1868 if (CONST_INT_P (op_b))
1869 {
1870 HOST_WIDE_INT desired_val = INTVAL (target);
1871 HOST_WIDE_INT actual_val = INTVAL (op_b);
1872
1873 switch (code)
1874 {
1875 case LT:
1876 if (actual_val == desired_val + 1)
1877 {
1878 code = LE;
1879 op_b = GEN_INT (desired_val);
1880 }
1881 break;
1882 case LE:
1883 if (actual_val == desired_val - 1)
1884 {
1885 code = LT;
1886 op_b = GEN_INT (desired_val);
1887 }
1888 break;
1889 case GT:
1890 if (actual_val == desired_val - 1)
1891 {
1892 code = GE;
1893 op_b = GEN_INT (desired_val);
1894 }
1895 break;
1896 case GE:
1897 if (actual_val == desired_val + 1)
1898 {
1899 code = GT;
1900 op_b = GEN_INT (desired_val);
1901 }
1902 break;
1903 default:
1904 break;
1905 }
1906 }
1907
1908 /* If we made any changes, generate a new conditional that is
1909 equivalent to what we started with, but has the right
1910 constants in it. */
1911 if (code != GET_CODE (if_info->cond)
1912 || op_a != XEXP (if_info->cond, 0)
1913 || op_b != XEXP (if_info->cond, 1))
1914 {
1915 cond = gen_rtx_fmt_ee (code, GET_MODE (cond), op_a, op_b);
1916 *earliest = if_info->cond_earliest;
1917 return cond;
1918 }
1919 }
1920
1921 cond = canonicalize_condition (if_info->jump, cond, reverse,
1922 earliest, target, HAVE_cbranchcc4, true);
1923 if (! cond || ! reg_mentioned_p (target, cond))
1924 return NULL;
1925
1926 /* We almost certainly searched back to a different place.
1927 Need to re-verify correct lifetimes. */
1928
1929 /* X may not be mentioned in the range (cond_earliest, jump]. */
1930 for (insn = if_info->jump; insn != *earliest; insn = PREV_INSN (insn))
1931 if (INSN_P (insn) && reg_overlap_mentioned_p (if_info->x, PATTERN (insn)))
1932 return NULL;
1933
1934 /* A and B may not be modified in the range [cond_earliest, jump). */
1935 for (insn = *earliest; insn != if_info->jump; insn = NEXT_INSN (insn))
1936 if (INSN_P (insn)
1937 && (modified_in_p (if_info->a, insn)
1938 || modified_in_p (if_info->b, insn)))
1939 return NULL;
1940
1941 return cond;
1942 }
1943
1944 /* Convert "if (a < b) x = a; else x = b;" to "x = min(a, b);", etc. */
1945
1946 static int
1947 noce_try_minmax (struct noce_if_info *if_info)
1948 {
1949 rtx cond, target;
1950 rtx_insn *earliest, *seq;
1951 enum rtx_code code, op;
1952 int unsignedp;
1953
1954 /* ??? Reject modes with NaNs or signed zeros since we don't know how
1955 they will be resolved with an SMIN/SMAX. It wouldn't be too hard
1956 to get the target to tell us... */
1957 if (HONOR_SIGNED_ZEROS (GET_MODE (if_info->x))
1958 || HONOR_NANS (GET_MODE (if_info->x)))
1959 return FALSE;
1960
1961 cond = noce_get_alt_condition (if_info, if_info->a, &earliest);
1962 if (!cond)
1963 return FALSE;
1964
1965 /* Verify the condition is of the form we expect, and canonicalize
1966 the comparison code. */
1967 code = GET_CODE (cond);
1968 if (rtx_equal_p (XEXP (cond, 0), if_info->a))
1969 {
1970 if (! rtx_equal_p (XEXP (cond, 1), if_info->b))
1971 return FALSE;
1972 }
1973 else if (rtx_equal_p (XEXP (cond, 1), if_info->a))
1974 {
1975 if (! rtx_equal_p (XEXP (cond, 0), if_info->b))
1976 return FALSE;
1977 code = swap_condition (code);
1978 }
1979 else
1980 return FALSE;
1981
1982 /* Determine what sort of operation this is. Note that the code is for
1983 a taken branch, so the code->operation mapping appears backwards. */
1984 switch (code)
1985 {
1986 case LT:
1987 case LE:
1988 case UNLT:
1989 case UNLE:
1990 op = SMAX;
1991 unsignedp = 0;
1992 break;
1993 case GT:
1994 case GE:
1995 case UNGT:
1996 case UNGE:
1997 op = SMIN;
1998 unsignedp = 0;
1999 break;
2000 case LTU:
2001 case LEU:
2002 op = UMAX;
2003 unsignedp = 1;
2004 break;
2005 case GTU:
2006 case GEU:
2007 op = UMIN;
2008 unsignedp = 1;
2009 break;
2010 default:
2011 return FALSE;
2012 }
2013
2014 start_sequence ();
2015
2016 target = expand_simple_binop (GET_MODE (if_info->x), op,
2017 if_info->a, if_info->b,
2018 if_info->x, unsignedp, OPTAB_WIDEN);
2019 if (! target)
2020 {
2021 end_sequence ();
2022 return FALSE;
2023 }
2024 if (target != if_info->x)
2025 noce_emit_move_insn (if_info->x, target);
2026
2027 seq = end_ifcvt_sequence (if_info);
2028 if (!seq)
2029 return FALSE;
2030
2031 emit_insn_before_setloc (seq, if_info->jump, INSN_LOCATION (if_info->insn_a));
2032 if_info->cond = cond;
2033 if_info->cond_earliest = earliest;
2034
2035 return TRUE;
2036 }
2037
2038 /* Convert "if (a < 0) x = -a; else x = a;" to "x = abs(a);",
2039 "if (a < 0) x = ~a; else x = a;" to "x = one_cmpl_abs(a);",
2040 etc. */
2041
2042 static int
2043 noce_try_abs (struct noce_if_info *if_info)
2044 {
2045 rtx cond, target, a, b, c;
2046 rtx_insn *earliest, *seq;
2047 int negate;
2048 bool one_cmpl = false;
2049
2050 /* Reject modes with signed zeros. */
2051 if (HONOR_SIGNED_ZEROS (GET_MODE (if_info->x)))
2052 return FALSE;
2053
2054 /* Recognize A and B as constituting an ABS or NABS. The canonical
2055 form is a branch around the negation, taken when the object is the
2056 first operand of a comparison against 0 that evaluates to true. */
2057 a = if_info->a;
2058 b = if_info->b;
2059 if (GET_CODE (a) == NEG && rtx_equal_p (XEXP (a, 0), b))
2060 negate = 0;
2061 else if (GET_CODE (b) == NEG && rtx_equal_p (XEXP (b, 0), a))
2062 {
2063 c = a; a = b; b = c;
2064 negate = 1;
2065 }
2066 else if (GET_CODE (a) == NOT && rtx_equal_p (XEXP (a, 0), b))
2067 {
2068 negate = 0;
2069 one_cmpl = true;
2070 }
2071 else if (GET_CODE (b) == NOT && rtx_equal_p (XEXP (b, 0), a))
2072 {
2073 c = a; a = b; b = c;
2074 negate = 1;
2075 one_cmpl = true;
2076 }
2077 else
2078 return FALSE;
2079
2080 cond = noce_get_alt_condition (if_info, b, &earliest);
2081 if (!cond)
2082 return FALSE;
2083
2084 /* Verify the condition is of the form we expect. */
2085 if (rtx_equal_p (XEXP (cond, 0), b))
2086 c = XEXP (cond, 1);
2087 else if (rtx_equal_p (XEXP (cond, 1), b))
2088 {
2089 c = XEXP (cond, 0);
2090 negate = !negate;
2091 }
2092 else
2093 return FALSE;
2094
2095 /* Verify that C is zero. Search one step backward for a
2096 REG_EQUAL note or a simple source if necessary. */
2097 if (REG_P (c))
2098 {
2099 rtx set;
2100 rtx_insn *insn = prev_nonnote_insn (earliest);
2101 if (insn
2102 && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (earliest)
2103 && (set = single_set (insn))
2104 && rtx_equal_p (SET_DEST (set), c))
2105 {
2106 rtx note = find_reg_equal_equiv_note (insn);
2107 if (note)
2108 c = XEXP (note, 0);
2109 else
2110 c = SET_SRC (set);
2111 }
2112 else
2113 return FALSE;
2114 }
2115 if (MEM_P (c)
2116 && GET_CODE (XEXP (c, 0)) == SYMBOL_REF
2117 && CONSTANT_POOL_ADDRESS_P (XEXP (c, 0)))
2118 c = get_pool_constant (XEXP (c, 0));
2119
2120 /* Work around funny ideas get_condition has wrt canonicalization.
2121 Note that these rtx constants are known to be CONST_INT, and
2122 therefore imply integer comparisons. */
2123 if (c == constm1_rtx && GET_CODE (cond) == GT)
2124 ;
2125 else if (c == const1_rtx && GET_CODE (cond) == LT)
2126 ;
2127 else if (c != CONST0_RTX (GET_MODE (b)))
2128 return FALSE;
2129
2130 /* Determine what sort of operation this is. */
2131 switch (GET_CODE (cond))
2132 {
2133 case LT:
2134 case LE:
2135 case UNLT:
2136 case UNLE:
2137 negate = !negate;
2138 break;
2139 case GT:
2140 case GE:
2141 case UNGT:
2142 case UNGE:
2143 break;
2144 default:
2145 return FALSE;
2146 }
2147
2148 start_sequence ();
2149 if (one_cmpl)
2150 target = expand_one_cmpl_abs_nojump (GET_MODE (if_info->x), b,
2151 if_info->x);
2152 else
2153 target = expand_abs_nojump (GET_MODE (if_info->x), b, if_info->x, 1);
2154
2155 /* ??? It's a quandary whether cmove would be better here, especially
2156 for integers. Perhaps combine will clean things up. */
2157 if (target && negate)
2158 {
2159 if (one_cmpl)
2160 target = expand_simple_unop (GET_MODE (target), NOT, target,
2161 if_info->x, 0);
2162 else
2163 target = expand_simple_unop (GET_MODE (target), NEG, target,
2164 if_info->x, 0);
2165 }
2166
2167 if (! target)
2168 {
2169 end_sequence ();
2170 return FALSE;
2171 }
2172
2173 if (target != if_info->x)
2174 noce_emit_move_insn (if_info->x, target);
2175
2176 seq = end_ifcvt_sequence (if_info);
2177 if (!seq)
2178 return FALSE;
2179
2180 emit_insn_before_setloc (seq, if_info->jump, INSN_LOCATION (if_info->insn_a));
2181 if_info->cond = cond;
2182 if_info->cond_earliest = earliest;
2183
2184 return TRUE;
2185 }
2186
2187 /* Convert "if (m < 0) x = b; else x = 0;" to "x = (m >> C) & b;". */
2188
2189 static int
2190 noce_try_sign_mask (struct noce_if_info *if_info)
2191 {
2192 rtx cond, t, m, c;
2193 rtx_insn *seq;
2194 machine_mode mode;
2195 enum rtx_code code;
2196 bool t_unconditional;
2197
2198 cond = if_info->cond;
2199 code = GET_CODE (cond);
2200 m = XEXP (cond, 0);
2201 c = XEXP (cond, 1);
2202
2203 t = NULL_RTX;
2204 if (if_info->a == const0_rtx)
2205 {
2206 if ((code == LT && c == const0_rtx)
2207 || (code == LE && c == constm1_rtx))
2208 t = if_info->b;
2209 }
2210 else if (if_info->b == const0_rtx)
2211 {
2212 if ((code == GE && c == const0_rtx)
2213 || (code == GT && c == constm1_rtx))
2214 t = if_info->a;
2215 }
2216
2217 if (! t || side_effects_p (t))
2218 return FALSE;
2219
2220 /* We currently don't handle different modes. */
2221 mode = GET_MODE (t);
2222 if (GET_MODE (m) != mode)
2223 return FALSE;
2224
2225 /* This is only profitable if T is unconditionally executed/evaluated in the
2226 original insn sequence or T is cheap. The former happens if B is the
2227 non-zero (T) value and if INSN_B was taken from TEST_BB, or there was no
2228 INSN_B which can happen for e.g. conditional stores to memory. For the
2229 cost computation use the block TEST_BB where the evaluation will end up
2230 after the transformation. */
2231 t_unconditional =
2232 (t == if_info->b
2233 && (if_info->insn_b == NULL_RTX
2234 || BLOCK_FOR_INSN (if_info->insn_b) == if_info->test_bb));
2235 if (!(t_unconditional
2236 || (set_src_cost (t, optimize_bb_for_speed_p (if_info->test_bb))
2237 < COSTS_N_INSNS (2))))
2238 return FALSE;
2239
2240 start_sequence ();
2241 /* Use emit_store_flag to generate "m < 0 ? -1 : 0" instead of expanding
2242 "(signed) m >> 31" directly. This benefits targets with specialized
2243 insns to obtain the signmask, but still uses ashr_optab otherwise. */
2244 m = emit_store_flag (gen_reg_rtx (mode), LT, m, const0_rtx, mode, 0, -1);
2245 t = m ? expand_binop (mode, and_optab, m, t, NULL_RTX, 0, OPTAB_DIRECT)
2246 : NULL_RTX;
2247
2248 if (!t)
2249 {
2250 end_sequence ();
2251 return FALSE;
2252 }
2253
2254 noce_emit_move_insn (if_info->x, t);
2255
2256 seq = end_ifcvt_sequence (if_info);
2257 if (!seq)
2258 return FALSE;
2259
2260 emit_insn_before_setloc (seq, if_info->jump, INSN_LOCATION (if_info->insn_a));
2261 return TRUE;
2262 }
2263
2264
2265 /* Optimize away "if (x & C) x |= C" and similar bit manipulation
2266 transformations. */
2267
2268 static int
2269 noce_try_bitop (struct noce_if_info *if_info)
2270 {
2271 rtx cond, x, a, result;
2272 rtx_insn *seq;
2273 machine_mode mode;
2274 enum rtx_code code;
2275 int bitnum;
2276
2277 x = if_info->x;
2278 cond = if_info->cond;
2279 code = GET_CODE (cond);
2280
2281 /* Check for no else condition. */
2282 if (! rtx_equal_p (x, if_info->b))
2283 return FALSE;
2284
2285 /* Check for a suitable condition. */
2286 if (code != NE && code != EQ)
2287 return FALSE;
2288 if (XEXP (cond, 1) != const0_rtx)
2289 return FALSE;
2290 cond = XEXP (cond, 0);
2291
2292 /* ??? We could also handle AND here. */
2293 if (GET_CODE (cond) == ZERO_EXTRACT)
2294 {
2295 if (XEXP (cond, 1) != const1_rtx
2296 || !CONST_INT_P (XEXP (cond, 2))
2297 || ! rtx_equal_p (x, XEXP (cond, 0)))
2298 return FALSE;
2299 bitnum = INTVAL (XEXP (cond, 2));
2300 mode = GET_MODE (x);
2301 if (BITS_BIG_ENDIAN)
2302 bitnum = GET_MODE_BITSIZE (mode) - 1 - bitnum;
2303 if (bitnum < 0 || bitnum >= HOST_BITS_PER_WIDE_INT)
2304 return FALSE;
2305 }
2306 else
2307 return FALSE;
2308
2309 a = if_info->a;
2310 if (GET_CODE (a) == IOR || GET_CODE (a) == XOR)
2311 {
2312 /* Check for "if (X & C) x = x op C". */
2313 if (! rtx_equal_p (x, XEXP (a, 0))
2314 || !CONST_INT_P (XEXP (a, 1))
2315 || (INTVAL (XEXP (a, 1)) & GET_MODE_MASK (mode))
2316 != (unsigned HOST_WIDE_INT) 1 << bitnum)
2317 return FALSE;
2318
2319 /* if ((x & C) == 0) x |= C; is transformed to x |= C. */
2320 /* if ((x & C) != 0) x |= C; is transformed to nothing. */
2321 if (GET_CODE (a) == IOR)
2322 result = (code == NE) ? a : NULL_RTX;
2323 else if (code == NE)
2324 {
2325 /* if ((x & C) == 0) x ^= C; is transformed to x |= C. */
2326 result = gen_int_mode ((HOST_WIDE_INT) 1 << bitnum, mode);
2327 result = simplify_gen_binary (IOR, mode, x, result);
2328 }
2329 else
2330 {
2331 /* if ((x & C) != 0) x ^= C; is transformed to x &= ~C. */
2332 result = gen_int_mode (~((HOST_WIDE_INT) 1 << bitnum), mode);
2333 result = simplify_gen_binary (AND, mode, x, result);
2334 }
2335 }
2336 else if (GET_CODE (a) == AND)
2337 {
2338 /* Check for "if (X & C) x &= ~C". */
2339 if (! rtx_equal_p (x, XEXP (a, 0))
2340 || !CONST_INT_P (XEXP (a, 1))
2341 || (INTVAL (XEXP (a, 1)) & GET_MODE_MASK (mode))
2342 != (~((HOST_WIDE_INT) 1 << bitnum) & GET_MODE_MASK (mode)))
2343 return FALSE;
2344
2345 /* if ((x & C) == 0) x &= ~C; is transformed to nothing. */
2346 /* if ((x & C) != 0) x &= ~C; is transformed to x &= ~C. */
2347 result = (code == EQ) ? a : NULL_RTX;
2348 }
2349 else
2350 return FALSE;
2351
2352 if (result)
2353 {
2354 start_sequence ();
2355 noce_emit_move_insn (x, result);
2356 seq = end_ifcvt_sequence (if_info);
2357 if (!seq)
2358 return FALSE;
2359
2360 emit_insn_before_setloc (seq, if_info->jump,
2361 INSN_LOCATION (if_info->insn_a));
2362 }
2363 return TRUE;
2364 }
2365
2366
2367 /* Similar to get_condition, only the resulting condition must be
2368 valid at JUMP, instead of at EARLIEST.
2369
2370 If THEN_ELSE_REVERSED is true, the fallthrough does not go to the
2371 THEN block of the caller, and we have to reverse the condition. */
2372
2373 static rtx
2374 noce_get_condition (rtx_insn *jump, rtx_insn **earliest, bool then_else_reversed)
2375 {
2376 rtx cond, set, tmp;
2377 bool reverse;
2378
2379 if (! any_condjump_p (jump))
2380 return NULL_RTX;
2381
2382 set = pc_set (jump);
2383
2384 /* If this branches to JUMP_LABEL when the condition is false,
2385 reverse the condition. */
2386 reverse = (GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
2387 && LABEL_REF_LABEL (XEXP (SET_SRC (set), 2)) == JUMP_LABEL (jump));
2388
2389 /* We may have to reverse because the caller's if block is not canonical,
2390 i.e. the THEN block isn't the fallthrough block for the TEST block
2391 (see find_if_header). */
2392 if (then_else_reversed)
2393 reverse = !reverse;
2394
2395 /* If the condition variable is a register and is MODE_INT, accept it. */
2396
2397 cond = XEXP (SET_SRC (set), 0);
2398 tmp = XEXP (cond, 0);
2399 if (REG_P (tmp) && GET_MODE_CLASS (GET_MODE (tmp)) == MODE_INT
2400 && (GET_MODE (tmp) != BImode
2401 || !targetm.small_register_classes_for_mode_p (BImode)))
2402 {
2403 *earliest = jump;
2404
2405 if (reverse)
2406 cond = gen_rtx_fmt_ee (reverse_condition (GET_CODE (cond)),
2407 GET_MODE (cond), tmp, XEXP (cond, 1));
2408 return cond;
2409 }
2410
2411 /* Otherwise, fall back on canonicalize_condition to do the dirty
2412 work of manipulating MODE_CC values and COMPARE rtx codes. */
2413 tmp = canonicalize_condition (jump, cond, reverse, earliest,
2414 NULL_RTX, HAVE_cbranchcc4, true);
2415
2416 /* We don't handle side-effects in the condition, like handling
2417 REG_INC notes and making sure no duplicate conditions are emitted. */
2418 if (tmp != NULL_RTX && side_effects_p (tmp))
2419 return NULL_RTX;
2420
2421 return tmp;
2422 }
2423
2424 /* Return true if OP is ok for if-then-else processing. */
2425
2426 static int
2427 noce_operand_ok (const_rtx op)
2428 {
2429 if (side_effects_p (op))
2430 return FALSE;
2431
2432 /* We special-case memories, so handle any of them with
2433 no address side effects. */
2434 if (MEM_P (op))
2435 return ! side_effects_p (XEXP (op, 0));
2436
2437 return ! may_trap_p (op);
2438 }
2439
2440 /* Return true if a write into MEM may trap or fault. */
2441
2442 static bool
2443 noce_mem_write_may_trap_or_fault_p (const_rtx mem)
2444 {
2445 rtx addr;
2446
2447 if (MEM_READONLY_P (mem))
2448 return true;
2449
2450 if (may_trap_or_fault_p (mem))
2451 return true;
2452
2453 addr = XEXP (mem, 0);
2454
2455 /* Call target hook to avoid the effects of -fpic etc.... */
2456 addr = targetm.delegitimize_address (addr);
2457
2458 while (addr)
2459 switch (GET_CODE (addr))
2460 {
2461 case CONST:
2462 case PRE_DEC:
2463 case PRE_INC:
2464 case POST_DEC:
2465 case POST_INC:
2466 case POST_MODIFY:
2467 addr = XEXP (addr, 0);
2468 break;
2469 case LO_SUM:
2470 case PRE_MODIFY:
2471 addr = XEXP (addr, 1);
2472 break;
2473 case PLUS:
2474 if (CONST_INT_P (XEXP (addr, 1)))
2475 addr = XEXP (addr, 0);
2476 else
2477 return false;
2478 break;
2479 case LABEL_REF:
2480 return true;
2481 case SYMBOL_REF:
2482 if (SYMBOL_REF_DECL (addr)
2483 && decl_readonly_section (SYMBOL_REF_DECL (addr), 0))
2484 return true;
2485 return false;
2486 default:
2487 return false;
2488 }
2489
2490 return false;
2491 }
2492
2493 /* Return whether we can use store speculation for MEM. TOP_BB is the
2494 basic block above the conditional block where we are considering
2495 doing the speculative store. We look for whether MEM is set
2496 unconditionally later in the function. */
2497
2498 static bool
2499 noce_can_store_speculate_p (basic_block top_bb, const_rtx mem)
2500 {
2501 basic_block dominator;
2502
2503 for (dominator = get_immediate_dominator (CDI_POST_DOMINATORS, top_bb);
2504 dominator != NULL;
2505 dominator = get_immediate_dominator (CDI_POST_DOMINATORS, dominator))
2506 {
2507 rtx_insn *insn;
2508
2509 FOR_BB_INSNS (dominator, insn)
2510 {
2511 /* If we see something that might be a memory barrier, we
2512 have to stop looking. Even if the MEM is set later in
2513 the function, we still don't want to set it
2514 unconditionally before the barrier. */
2515 if (INSN_P (insn)
2516 && (volatile_insn_p (PATTERN (insn))
2517 || (CALL_P (insn) && (!RTL_CONST_CALL_P (insn)))))
2518 return false;
2519
2520 if (memory_must_be_modified_in_insn_p (mem, insn))
2521 return true;
2522 if (modified_in_p (XEXP (mem, 0), insn))
2523 return false;
2524
2525 }
2526 }
2527
2528 return false;
2529 }
2530
2531 /* Given a simple IF-THEN-JOIN or IF-THEN-ELSE-JOIN block, attempt to convert
2532 it without using conditional execution. Return TRUE if we were successful
2533 at converting the block. */
2534
2535 static int
2536 noce_process_if_block (struct noce_if_info *if_info)
2537 {
2538 basic_block test_bb = if_info->test_bb; /* test block */
2539 basic_block then_bb = if_info->then_bb; /* THEN */
2540 basic_block else_bb = if_info->else_bb; /* ELSE or NULL */
2541 basic_block join_bb = if_info->join_bb; /* JOIN */
2542 rtx_insn *jump = if_info->jump;
2543 rtx cond = if_info->cond;
2544 rtx_insn *insn_a, *insn_b;
2545 rtx set_a, set_b;
2546 rtx orig_x, x, a, b;
2547
2548 /* We're looking for patterns of the form
2549
2550 (1) if (...) x = a; else x = b;
2551 (2) x = b; if (...) x = a;
2552 (3) if (...) x = a; // as if with an initial x = x.
2553
2554 The later patterns require jumps to be more expensive.
2555
2556 ??? For future expansion, look for multiple X in such patterns. */
2557
2558 /* Look for one of the potential sets. */
2559 insn_a = first_active_insn (then_bb);
2560 if (! insn_a
2561 || insn_a != last_active_insn (then_bb, FALSE)
2562 || (set_a = single_set (insn_a)) == NULL_RTX)
2563 return FALSE;
2564
2565 x = SET_DEST (set_a);
2566 a = SET_SRC (set_a);
2567
2568 /* Look for the other potential set. Make sure we've got equivalent
2569 destinations. */
2570 /* ??? This is overconservative. Storing to two different mems is
2571 as easy as conditionally computing the address. Storing to a
2572 single mem merely requires a scratch memory to use as one of the
2573 destination addresses; often the memory immediately below the
2574 stack pointer is available for this. */
2575 set_b = NULL_RTX;
2576 if (else_bb)
2577 {
2578 insn_b = first_active_insn (else_bb);
2579 if (! insn_b
2580 || insn_b != last_active_insn (else_bb, FALSE)
2581 || (set_b = single_set (insn_b)) == NULL_RTX
2582 || ! rtx_interchangeable_p (x, SET_DEST (set_b)))
2583 return FALSE;
2584 }
2585 else
2586 {
2587 insn_b = prev_nonnote_nondebug_insn (if_info->cond_earliest);
2588 /* We're going to be moving the evaluation of B down from above
2589 COND_EARLIEST to JUMP. Make sure the relevant data is still
2590 intact. */
2591 if (! insn_b
2592 || BLOCK_FOR_INSN (insn_b) != BLOCK_FOR_INSN (if_info->cond_earliest)
2593 || !NONJUMP_INSN_P (insn_b)
2594 || (set_b = single_set (insn_b)) == NULL_RTX
2595 || ! rtx_interchangeable_p (x, SET_DEST (set_b))
2596 || ! noce_operand_ok (SET_SRC (set_b))
2597 || reg_overlap_mentioned_p (x, SET_SRC (set_b))
2598 || modified_between_p (SET_SRC (set_b), insn_b, jump)
2599 /* Avoid extending the lifetime of hard registers on small
2600 register class machines. */
2601 || (REG_P (SET_SRC (set_b))
2602 && HARD_REGISTER_P (SET_SRC (set_b))
2603 && targetm.small_register_classes_for_mode_p
2604 (GET_MODE (SET_SRC (set_b))))
2605 /* Likewise with X. In particular this can happen when
2606 noce_get_condition looks farther back in the instruction
2607 stream than one might expect. */
2608 || reg_overlap_mentioned_p (x, cond)
2609 || reg_overlap_mentioned_p (x, a)
2610 || modified_between_p (x, insn_b, jump))
2611 {
2612 insn_b = NULL;
2613 set_b = NULL_RTX;
2614 }
2615 }
2616
2617 /* If x has side effects then only the if-then-else form is safe to
2618 convert. But even in that case we would need to restore any notes
2619 (such as REG_INC) at then end. That can be tricky if
2620 noce_emit_move_insn expands to more than one insn, so disable the
2621 optimization entirely for now if there are side effects. */
2622 if (side_effects_p (x))
2623 return FALSE;
2624
2625 b = (set_b ? SET_SRC (set_b) : x);
2626
2627 /* Only operate on register destinations, and even then avoid extending
2628 the lifetime of hard registers on small register class machines. */
2629 orig_x = x;
2630 if (!REG_P (x)
2631 || (HARD_REGISTER_P (x)
2632 && targetm.small_register_classes_for_mode_p (GET_MODE (x))))
2633 {
2634 if (GET_MODE (x) == BLKmode)
2635 return FALSE;
2636
2637 if (GET_CODE (x) == ZERO_EXTRACT
2638 && (!CONST_INT_P (XEXP (x, 1))
2639 || !CONST_INT_P (XEXP (x, 2))))
2640 return FALSE;
2641
2642 x = gen_reg_rtx (GET_MODE (GET_CODE (x) == STRICT_LOW_PART
2643 ? XEXP (x, 0) : x));
2644 }
2645
2646 /* Don't operate on sources that may trap or are volatile. */
2647 if (! noce_operand_ok (a) || ! noce_operand_ok (b))
2648 return FALSE;
2649
2650 retry:
2651 /* Set up the info block for our subroutines. */
2652 if_info->insn_a = insn_a;
2653 if_info->insn_b = insn_b;
2654 if_info->x = x;
2655 if_info->a = a;
2656 if_info->b = b;
2657
2658 /* Try optimizations in some approximation of a useful order. */
2659 /* ??? Should first look to see if X is live incoming at all. If it
2660 isn't, we don't need anything but an unconditional set. */
2661
2662 /* Look and see if A and B are really the same. Avoid creating silly
2663 cmove constructs that no one will fix up later. */
2664 if (rtx_interchangeable_p (a, b))
2665 {
2666 /* If we have an INSN_B, we don't have to create any new rtl. Just
2667 move the instruction that we already have. If we don't have an
2668 INSN_B, that means that A == X, and we've got a noop move. In
2669 that case don't do anything and let the code below delete INSN_A. */
2670 if (insn_b && else_bb)
2671 {
2672 rtx note;
2673
2674 if (else_bb && insn_b == BB_END (else_bb))
2675 BB_END (else_bb) = PREV_INSN (insn_b);
2676 reorder_insns (insn_b, insn_b, PREV_INSN (jump));
2677
2678 /* If there was a REG_EQUAL note, delete it since it may have been
2679 true due to this insn being after a jump. */
2680 if ((note = find_reg_note (insn_b, REG_EQUAL, NULL_RTX)) != 0)
2681 remove_note (insn_b, note);
2682
2683 insn_b = NULL;
2684 }
2685 /* If we have "x = b; if (...) x = a;", and x has side-effects, then
2686 x must be executed twice. */
2687 else if (insn_b && side_effects_p (orig_x))
2688 return FALSE;
2689
2690 x = orig_x;
2691 goto success;
2692 }
2693
2694 if (!set_b && MEM_P (orig_x))
2695 {
2696 /* Disallow the "if (...) x = a;" form (implicit "else x = x;")
2697 for optimizations if writing to x may trap or fault,
2698 i.e. it's a memory other than a static var or a stack slot,
2699 is misaligned on strict aligned machines or is read-only. If
2700 x is a read-only memory, then the program is valid only if we
2701 avoid the store into it. If there are stores on both the
2702 THEN and ELSE arms, then we can go ahead with the conversion;
2703 either the program is broken, or the condition is always
2704 false such that the other memory is selected. */
2705 if (noce_mem_write_may_trap_or_fault_p (orig_x))
2706 return FALSE;
2707
2708 /* Avoid store speculation: given "if (...) x = a" where x is a
2709 MEM, we only want to do the store if x is always set
2710 somewhere in the function. This avoids cases like
2711 if (pthread_mutex_trylock(mutex))
2712 ++global_variable;
2713 where we only want global_variable to be changed if the mutex
2714 is held. FIXME: This should ideally be expressed directly in
2715 RTL somehow. */
2716 if (!noce_can_store_speculate_p (test_bb, orig_x))
2717 return FALSE;
2718 }
2719
2720 if (noce_try_move (if_info))
2721 goto success;
2722 if (noce_try_store_flag (if_info))
2723 goto success;
2724 if (noce_try_bitop (if_info))
2725 goto success;
2726 if (noce_try_minmax (if_info))
2727 goto success;
2728 if (noce_try_abs (if_info))
2729 goto success;
2730 if (HAVE_conditional_move
2731 && noce_try_cmove (if_info))
2732 goto success;
2733 if (! targetm.have_conditional_execution ())
2734 {
2735 if (noce_try_store_flag_constants (if_info))
2736 goto success;
2737 if (noce_try_addcc (if_info))
2738 goto success;
2739 if (noce_try_store_flag_mask (if_info))
2740 goto success;
2741 if (HAVE_conditional_move
2742 && noce_try_cmove_arith (if_info))
2743 goto success;
2744 if (noce_try_sign_mask (if_info))
2745 goto success;
2746 }
2747
2748 if (!else_bb && set_b)
2749 {
2750 insn_b = NULL;
2751 set_b = NULL_RTX;
2752 b = orig_x;
2753 goto retry;
2754 }
2755
2756 return FALSE;
2757
2758 success:
2759
2760 /* If we used a temporary, fix it up now. */
2761 if (orig_x != x)
2762 {
2763 rtx_insn *seq;
2764
2765 start_sequence ();
2766 noce_emit_move_insn (orig_x, x);
2767 seq = get_insns ();
2768 set_used_flags (orig_x);
2769 unshare_all_rtl_in_chain (seq);
2770 end_sequence ();
2771
2772 emit_insn_before_setloc (seq, BB_END (test_bb), INSN_LOCATION (insn_a));
2773 }
2774
2775 /* The original THEN and ELSE blocks may now be removed. The test block
2776 must now jump to the join block. If the test block and the join block
2777 can be merged, do so. */
2778 if (else_bb)
2779 {
2780 delete_basic_block (else_bb);
2781 num_true_changes++;
2782 }
2783 else
2784 remove_edge (find_edge (test_bb, join_bb));
2785
2786 remove_edge (find_edge (then_bb, join_bb));
2787 redirect_edge_and_branch_force (single_succ_edge (test_bb), join_bb);
2788 delete_basic_block (then_bb);
2789 num_true_changes++;
2790
2791 if (can_merge_blocks_p (test_bb, join_bb))
2792 {
2793 merge_blocks (test_bb, join_bb);
2794 num_true_changes++;
2795 }
2796
2797 num_updated_if_blocks++;
2798 return TRUE;
2799 }
2800
2801 /* Check whether a block is suitable for conditional move conversion.
2802 Every insn must be a simple set of a register to a constant or a
2803 register. For each assignment, store the value in the pointer map
2804 VALS, keyed indexed by register pointer, then store the register
2805 pointer in REGS. COND is the condition we will test. */
2806
2807 static int
2808 check_cond_move_block (basic_block bb,
2809 hash_map<rtx, rtx> *vals,
2810 vec<rtx> *regs,
2811 rtx cond)
2812 {
2813 rtx_insn *insn;
2814
2815 /* We can only handle simple jumps at the end of the basic block.
2816 It is almost impossible to update the CFG otherwise. */
2817 insn = BB_END (bb);
2818 if (JUMP_P (insn) && !onlyjump_p (insn))
2819 return FALSE;
2820
2821 FOR_BB_INSNS (bb, insn)
2822 {
2823 rtx set, dest, src;
2824
2825 if (!NONDEBUG_INSN_P (insn) || JUMP_P (insn))
2826 continue;
2827 set = single_set (insn);
2828 if (!set)
2829 return FALSE;
2830
2831 dest = SET_DEST (set);
2832 src = SET_SRC (set);
2833 if (!REG_P (dest)
2834 || (HARD_REGISTER_P (dest)
2835 && targetm.small_register_classes_for_mode_p (GET_MODE (dest))))
2836 return FALSE;
2837
2838 if (!CONSTANT_P (src) && !register_operand (src, VOIDmode))
2839 return FALSE;
2840
2841 if (side_effects_p (src) || side_effects_p (dest))
2842 return FALSE;
2843
2844 if (may_trap_p (src) || may_trap_p (dest))
2845 return FALSE;
2846
2847 /* Don't try to handle this if the source register was
2848 modified earlier in the block. */
2849 if ((REG_P (src)
2850 && vals->get (src))
2851 || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
2852 && vals->get (SUBREG_REG (src))))
2853 return FALSE;
2854
2855 /* Don't try to handle this if the destination register was
2856 modified earlier in the block. */
2857 if (vals->get (dest))
2858 return FALSE;
2859
2860 /* Don't try to handle this if the condition uses the
2861 destination register. */
2862 if (reg_overlap_mentioned_p (dest, cond))
2863 return FALSE;
2864
2865 /* Don't try to handle this if the source register is modified
2866 later in the block. */
2867 if (!CONSTANT_P (src)
2868 && modified_between_p (src, insn, NEXT_INSN (BB_END (bb))))
2869 return FALSE;
2870
2871 vals->put (dest, src);
2872
2873 regs->safe_push (dest);
2874 }
2875
2876 return TRUE;
2877 }
2878
2879 /* Given a basic block BB suitable for conditional move conversion,
2880 a condition COND, and pointer maps THEN_VALS and ELSE_VALS containing
2881 the register values depending on COND, emit the insns in the block as
2882 conditional moves. If ELSE_BLOCK is true, THEN_BB was already
2883 processed. The caller has started a sequence for the conversion.
2884 Return true if successful, false if something goes wrong. */
2885
2886 static bool
2887 cond_move_convert_if_block (struct noce_if_info *if_infop,
2888 basic_block bb, rtx cond,
2889 hash_map<rtx, rtx> *then_vals,
2890 hash_map<rtx, rtx> *else_vals,
2891 bool else_block_p)
2892 {
2893 enum rtx_code code;
2894 rtx_insn *insn;
2895 rtx cond_arg0, cond_arg1;
2896
2897 code = GET_CODE (cond);
2898 cond_arg0 = XEXP (cond, 0);
2899 cond_arg1 = XEXP (cond, 1);
2900
2901 FOR_BB_INSNS (bb, insn)
2902 {
2903 rtx set, target, dest, t, e;
2904
2905 /* ??? Maybe emit conditional debug insn? */
2906 if (!NONDEBUG_INSN_P (insn) || JUMP_P (insn))
2907 continue;
2908 set = single_set (insn);
2909 gcc_assert (set && REG_P (SET_DEST (set)));
2910
2911 dest = SET_DEST (set);
2912
2913 rtx *then_slot = then_vals->get (dest);
2914 rtx *else_slot = else_vals->get (dest);
2915 t = then_slot ? *then_slot : NULL_RTX;
2916 e = else_slot ? *else_slot : NULL_RTX;
2917
2918 if (else_block_p)
2919 {
2920 /* If this register was set in the then block, we already
2921 handled this case there. */
2922 if (t)
2923 continue;
2924 t = dest;
2925 gcc_assert (e);
2926 }
2927 else
2928 {
2929 gcc_assert (t);
2930 if (!e)
2931 e = dest;
2932 }
2933
2934 target = noce_emit_cmove (if_infop, dest, code, cond_arg0, cond_arg1,
2935 t, e);
2936 if (!target)
2937 return false;
2938
2939 if (target != dest)
2940 noce_emit_move_insn (dest, target);
2941 }
2942
2943 return true;
2944 }
2945
2946 /* Given a simple IF-THEN-JOIN or IF-THEN-ELSE-JOIN block, attempt to convert
2947 it using only conditional moves. Return TRUE if we were successful at
2948 converting the block. */
2949
2950 static int
2951 cond_move_process_if_block (struct noce_if_info *if_info)
2952 {
2953 basic_block test_bb = if_info->test_bb;
2954 basic_block then_bb = if_info->then_bb;
2955 basic_block else_bb = if_info->else_bb;
2956 basic_block join_bb = if_info->join_bb;
2957 rtx_insn *jump = if_info->jump;
2958 rtx cond = if_info->cond;
2959 rtx_insn *seq, *loc_insn;
2960 rtx reg;
2961 int c;
2962 vec<rtx> then_regs = vNULL;
2963 vec<rtx> else_regs = vNULL;
2964 unsigned int i;
2965 int success_p = FALSE;
2966
2967 /* Build a mapping for each block to the value used for each
2968 register. */
2969 hash_map<rtx, rtx> then_vals;
2970 hash_map<rtx, rtx> else_vals;
2971
2972 /* Make sure the blocks are suitable. */
2973 if (!check_cond_move_block (then_bb, &then_vals, &then_regs, cond)
2974 || (else_bb
2975 && !check_cond_move_block (else_bb, &else_vals, &else_regs, cond)))
2976 goto done;
2977
2978 /* Make sure the blocks can be used together. If the same register
2979 is set in both blocks, and is not set to a constant in both
2980 cases, then both blocks must set it to the same register. We
2981 have already verified that if it is set to a register, that the
2982 source register does not change after the assignment. Also count
2983 the number of registers set in only one of the blocks. */
2984 c = 0;
2985 FOR_EACH_VEC_ELT (then_regs, i, reg)
2986 {
2987 rtx *then_slot = then_vals.get (reg);
2988 rtx *else_slot = else_vals.get (reg);
2989
2990 gcc_checking_assert (then_slot);
2991 if (!else_slot)
2992 ++c;
2993 else
2994 {
2995 rtx then_val = *then_slot;
2996 rtx else_val = *else_slot;
2997 if (!CONSTANT_P (then_val) && !CONSTANT_P (else_val)
2998 && !rtx_equal_p (then_val, else_val))
2999 goto done;
3000 }
3001 }
3002
3003 /* Finish off c for MAX_CONDITIONAL_EXECUTE. */
3004 FOR_EACH_VEC_ELT (else_regs, i, reg)
3005 {
3006 gcc_checking_assert (else_vals.get (reg));
3007 if (!then_vals.get (reg))
3008 ++c;
3009 }
3010
3011 /* Make sure it is reasonable to convert this block. What matters
3012 is the number of assignments currently made in only one of the
3013 branches, since if we convert we are going to always execute
3014 them. */
3015 if (c > MAX_CONDITIONAL_EXECUTE)
3016 goto done;
3017
3018 /* Try to emit the conditional moves. First do the then block,
3019 then do anything left in the else blocks. */
3020 start_sequence ();
3021 if (!cond_move_convert_if_block (if_info, then_bb, cond,
3022 &then_vals, &else_vals, false)
3023 || (else_bb
3024 && !cond_move_convert_if_block (if_info, else_bb, cond,
3025 &then_vals, &else_vals, true)))
3026 {
3027 end_sequence ();
3028 goto done;
3029 }
3030 seq = end_ifcvt_sequence (if_info);
3031 if (!seq)
3032 goto done;
3033
3034 loc_insn = first_active_insn (then_bb);
3035 if (!loc_insn)
3036 {
3037 loc_insn = first_active_insn (else_bb);
3038 gcc_assert (loc_insn);
3039 }
3040 emit_insn_before_setloc (seq, jump, INSN_LOCATION (loc_insn));
3041
3042 if (else_bb)
3043 {
3044 delete_basic_block (else_bb);
3045 num_true_changes++;
3046 }
3047 else
3048 remove_edge (find_edge (test_bb, join_bb));
3049
3050 remove_edge (find_edge (then_bb, join_bb));
3051 redirect_edge_and_branch_force (single_succ_edge (test_bb), join_bb);
3052 delete_basic_block (then_bb);
3053 num_true_changes++;
3054
3055 if (can_merge_blocks_p (test_bb, join_bb))
3056 {
3057 merge_blocks (test_bb, join_bb);
3058 num_true_changes++;
3059 }
3060
3061 num_updated_if_blocks++;
3062
3063 success_p = TRUE;
3064
3065 done:
3066 then_regs.release ();
3067 else_regs.release ();
3068 return success_p;
3069 }
3070
3071 \f
3072 /* Determine if a given basic block heads a simple IF-THEN-JOIN or an
3073 IF-THEN-ELSE-JOIN block.
3074
3075 If so, we'll try to convert the insns to not require the branch,
3076 using only transformations that do not require conditional execution.
3077
3078 Return TRUE if we were successful at converting the block. */
3079
3080 static int
3081 noce_find_if_block (basic_block test_bb, edge then_edge, edge else_edge,
3082 int pass)
3083 {
3084 basic_block then_bb, else_bb, join_bb;
3085 bool then_else_reversed = false;
3086 rtx_insn *jump;
3087 rtx cond;
3088 rtx_insn *cond_earliest;
3089 struct noce_if_info if_info;
3090
3091 /* We only ever should get here before reload. */
3092 gcc_assert (!reload_completed);
3093
3094 /* Recognize an IF-THEN-ELSE-JOIN block. */
3095 if (single_pred_p (then_edge->dest)
3096 && single_succ_p (then_edge->dest)
3097 && single_pred_p (else_edge->dest)
3098 && single_succ_p (else_edge->dest)
3099 && single_succ (then_edge->dest) == single_succ (else_edge->dest))
3100 {
3101 then_bb = then_edge->dest;
3102 else_bb = else_edge->dest;
3103 join_bb = single_succ (then_bb);
3104 }
3105 /* Recognize an IF-THEN-JOIN block. */
3106 else if (single_pred_p (then_edge->dest)
3107 && single_succ_p (then_edge->dest)
3108 && single_succ (then_edge->dest) == else_edge->dest)
3109 {
3110 then_bb = then_edge->dest;
3111 else_bb = NULL_BLOCK;
3112 join_bb = else_edge->dest;
3113 }
3114 /* Recognize an IF-ELSE-JOIN block. We can have those because the order
3115 of basic blocks in cfglayout mode does not matter, so the fallthrough
3116 edge can go to any basic block (and not just to bb->next_bb, like in
3117 cfgrtl mode). */
3118 else if (single_pred_p (else_edge->dest)
3119 && single_succ_p (else_edge->dest)
3120 && single_succ (else_edge->dest) == then_edge->dest)
3121 {
3122 /* The noce transformations do not apply to IF-ELSE-JOIN blocks.
3123 To make this work, we have to invert the THEN and ELSE blocks
3124 and reverse the jump condition. */
3125 then_bb = else_edge->dest;
3126 else_bb = NULL_BLOCK;
3127 join_bb = single_succ (then_bb);
3128 then_else_reversed = true;
3129 }
3130 else
3131 /* Not a form we can handle. */
3132 return FALSE;
3133
3134 /* The edges of the THEN and ELSE blocks cannot have complex edges. */
3135 if (single_succ_edge (then_bb)->flags & EDGE_COMPLEX)
3136 return FALSE;
3137 if (else_bb
3138 && single_succ_edge (else_bb)->flags & EDGE_COMPLEX)
3139 return FALSE;
3140
3141 num_possible_if_blocks++;
3142
3143 if (dump_file)
3144 {
3145 fprintf (dump_file,
3146 "\nIF-THEN%s-JOIN block found, pass %d, test %d, then %d",
3147 (else_bb) ? "-ELSE" : "",
3148 pass, test_bb->index, then_bb->index);
3149
3150 if (else_bb)
3151 fprintf (dump_file, ", else %d", else_bb->index);
3152
3153 fprintf (dump_file, ", join %d\n", join_bb->index);
3154 }
3155
3156 /* If the conditional jump is more than just a conditional
3157 jump, then we can not do if-conversion on this block. */
3158 jump = BB_END (test_bb);
3159 if (! onlyjump_p (jump))
3160 return FALSE;
3161
3162 /* If this is not a standard conditional jump, we can't parse it. */
3163 cond = noce_get_condition (jump, &cond_earliest, then_else_reversed);
3164 if (!cond)
3165 return FALSE;
3166
3167 /* We must be comparing objects whose modes imply the size. */
3168 if (GET_MODE (XEXP (cond, 0)) == BLKmode)
3169 return FALSE;
3170
3171 /* Initialize an IF_INFO struct to pass around. */
3172 memset (&if_info, 0, sizeof if_info);
3173 if_info.test_bb = test_bb;
3174 if_info.then_bb = then_bb;
3175 if_info.else_bb = else_bb;
3176 if_info.join_bb = join_bb;
3177 if_info.cond = cond;
3178 if_info.cond_earliest = cond_earliest;
3179 if_info.jump = jump;
3180 if_info.then_else_reversed = then_else_reversed;
3181 if_info.branch_cost = BRANCH_COST (optimize_bb_for_speed_p (test_bb),
3182 predictable_edge_p (then_edge));
3183
3184 /* Do the real work. */
3185
3186 if (noce_process_if_block (&if_info))
3187 return TRUE;
3188
3189 if (HAVE_conditional_move
3190 && cond_move_process_if_block (&if_info))
3191 return TRUE;
3192
3193 return FALSE;
3194 }
3195 \f
3196
3197 /* Merge the blocks and mark for local life update. */
3198
3199 static void
3200 merge_if_block (struct ce_if_block * ce_info)
3201 {
3202 basic_block test_bb = ce_info->test_bb; /* last test block */
3203 basic_block then_bb = ce_info->then_bb; /* THEN */
3204 basic_block else_bb = ce_info->else_bb; /* ELSE or NULL */
3205 basic_block join_bb = ce_info->join_bb; /* join block */
3206 basic_block combo_bb;
3207
3208 /* All block merging is done into the lower block numbers. */
3209
3210 combo_bb = test_bb;
3211 df_set_bb_dirty (test_bb);
3212
3213 /* Merge any basic blocks to handle && and || subtests. Each of
3214 the blocks are on the fallthru path from the predecessor block. */
3215 if (ce_info->num_multiple_test_blocks > 0)
3216 {
3217 basic_block bb = test_bb;
3218 basic_block last_test_bb = ce_info->last_test_bb;
3219 basic_block fallthru = block_fallthru (bb);
3220
3221 do
3222 {
3223 bb = fallthru;
3224 fallthru = block_fallthru (bb);
3225 merge_blocks (combo_bb, bb);
3226 num_true_changes++;
3227 }
3228 while (bb != last_test_bb);
3229 }
3230
3231 /* Merge TEST block into THEN block. Normally the THEN block won't have a
3232 label, but it might if there were || tests. That label's count should be
3233 zero, and it normally should be removed. */
3234
3235 if (then_bb)
3236 {
3237 /* If THEN_BB has no successors, then there's a BARRIER after it.
3238 If COMBO_BB has more than one successor (THEN_BB), then that BARRIER
3239 is no longer needed, and in fact it is incorrect to leave it in
3240 the insn stream. */
3241 if (EDGE_COUNT (then_bb->succs) == 0
3242 && EDGE_COUNT (combo_bb->succs) > 1)
3243 {
3244 rtx_insn *end = NEXT_INSN (BB_END (then_bb));
3245 while (end && NOTE_P (end) && !NOTE_INSN_BASIC_BLOCK_P (end))
3246 end = NEXT_INSN (end);
3247
3248 if (end && BARRIER_P (end))
3249 delete_insn (end);
3250 }
3251 merge_blocks (combo_bb, then_bb);
3252 num_true_changes++;
3253 }
3254
3255 /* The ELSE block, if it existed, had a label. That label count
3256 will almost always be zero, but odd things can happen when labels
3257 get their addresses taken. */
3258 if (else_bb)
3259 {
3260 /* If ELSE_BB has no successors, then there's a BARRIER after it.
3261 If COMBO_BB has more than one successor (ELSE_BB), then that BARRIER
3262 is no longer needed, and in fact it is incorrect to leave it in
3263 the insn stream. */
3264 if (EDGE_COUNT (else_bb->succs) == 0
3265 && EDGE_COUNT (combo_bb->succs) > 1)
3266 {
3267 rtx_insn *end = NEXT_INSN (BB_END (else_bb));
3268 while (end && NOTE_P (end) && !NOTE_INSN_BASIC_BLOCK_P (end))
3269 end = NEXT_INSN (end);
3270
3271 if (end && BARRIER_P (end))
3272 delete_insn (end);
3273 }
3274 merge_blocks (combo_bb, else_bb);
3275 num_true_changes++;
3276 }
3277
3278 /* If there was no join block reported, that means it was not adjacent
3279 to the others, and so we cannot merge them. */
3280
3281 if (! join_bb)
3282 {
3283 rtx_insn *last = BB_END (combo_bb);
3284
3285 /* The outgoing edge for the current COMBO block should already
3286 be correct. Verify this. */
3287 if (EDGE_COUNT (combo_bb->succs) == 0)
3288 gcc_assert (find_reg_note (last, REG_NORETURN, NULL)
3289 || (NONJUMP_INSN_P (last)
3290 && GET_CODE (PATTERN (last)) == TRAP_IF
3291 && (TRAP_CONDITION (PATTERN (last))
3292 == const_true_rtx)));
3293
3294 else
3295 /* There should still be something at the end of the THEN or ELSE
3296 blocks taking us to our final destination. */
3297 gcc_assert (JUMP_P (last)
3298 || (EDGE_SUCC (combo_bb, 0)->dest
3299 == EXIT_BLOCK_PTR_FOR_FN (cfun)
3300 && CALL_P (last)
3301 && SIBLING_CALL_P (last))
3302 || ((EDGE_SUCC (combo_bb, 0)->flags & EDGE_EH)
3303 && can_throw_internal (last)));
3304 }
3305
3306 /* The JOIN block may have had quite a number of other predecessors too.
3307 Since we've already merged the TEST, THEN and ELSE blocks, we should
3308 have only one remaining edge from our if-then-else diamond. If there
3309 is more than one remaining edge, it must come from elsewhere. There
3310 may be zero incoming edges if the THEN block didn't actually join
3311 back up (as with a call to a non-return function). */
3312 else if (EDGE_COUNT (join_bb->preds) < 2
3313 && join_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
3314 {
3315 /* We can merge the JOIN cleanly and update the dataflow try
3316 again on this pass.*/
3317 merge_blocks (combo_bb, join_bb);
3318 num_true_changes++;
3319 }
3320 else
3321 {
3322 /* We cannot merge the JOIN. */
3323
3324 /* The outgoing edge for the current COMBO block should already
3325 be correct. Verify this. */
3326 gcc_assert (single_succ_p (combo_bb)
3327 && single_succ (combo_bb) == join_bb);
3328
3329 /* Remove the jump and cruft from the end of the COMBO block. */
3330 if (join_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
3331 tidy_fallthru_edge (single_succ_edge (combo_bb));
3332 }
3333
3334 num_updated_if_blocks++;
3335 }
3336 \f
3337 /* Find a block ending in a simple IF condition and try to transform it
3338 in some way. When converting a multi-block condition, put the new code
3339 in the first such block and delete the rest. Return a pointer to this
3340 first block if some transformation was done. Return NULL otherwise. */
3341
3342 static basic_block
3343 find_if_header (basic_block test_bb, int pass)
3344 {
3345 ce_if_block ce_info;
3346 edge then_edge;
3347 edge else_edge;
3348
3349 /* The kind of block we're looking for has exactly two successors. */
3350 if (EDGE_COUNT (test_bb->succs) != 2)
3351 return NULL;
3352
3353 then_edge = EDGE_SUCC (test_bb, 0);
3354 else_edge = EDGE_SUCC (test_bb, 1);
3355
3356 if (df_get_bb_dirty (then_edge->dest))
3357 return NULL;
3358 if (df_get_bb_dirty (else_edge->dest))
3359 return NULL;
3360
3361 /* Neither edge should be abnormal. */
3362 if ((then_edge->flags & EDGE_COMPLEX)
3363 || (else_edge->flags & EDGE_COMPLEX))
3364 return NULL;
3365
3366 /* Nor exit the loop. */
3367 if ((then_edge->flags & EDGE_LOOP_EXIT)
3368 || (else_edge->flags & EDGE_LOOP_EXIT))
3369 return NULL;
3370
3371 /* The THEN edge is canonically the one that falls through. */
3372 if (then_edge->flags & EDGE_FALLTHRU)
3373 ;
3374 else if (else_edge->flags & EDGE_FALLTHRU)
3375 {
3376 edge e = else_edge;
3377 else_edge = then_edge;
3378 then_edge = e;
3379 }
3380 else
3381 /* Otherwise this must be a multiway branch of some sort. */
3382 return NULL;
3383
3384 memset (&ce_info, 0, sizeof (ce_info));
3385 ce_info.test_bb = test_bb;
3386 ce_info.then_bb = then_edge->dest;
3387 ce_info.else_bb = else_edge->dest;
3388 ce_info.pass = pass;
3389
3390 #ifdef IFCVT_MACHDEP_INIT
3391 IFCVT_MACHDEP_INIT (&ce_info);
3392 #endif
3393
3394 if (!reload_completed
3395 && noce_find_if_block (test_bb, then_edge, else_edge, pass))
3396 goto success;
3397
3398 if (reload_completed
3399 && targetm.have_conditional_execution ()
3400 && cond_exec_find_if_block (&ce_info))
3401 goto success;
3402
3403 if (HAVE_trap
3404 && optab_handler (ctrap_optab, word_mode) != CODE_FOR_nothing
3405 && find_cond_trap (test_bb, then_edge, else_edge))
3406 goto success;
3407
3408 if (dom_info_state (CDI_POST_DOMINATORS) >= DOM_NO_FAST_QUERY
3409 && (reload_completed || !targetm.have_conditional_execution ()))
3410 {
3411 if (find_if_case_1 (test_bb, then_edge, else_edge))
3412 goto success;
3413 if (find_if_case_2 (test_bb, then_edge, else_edge))
3414 goto success;
3415 }
3416
3417 return NULL;
3418
3419 success:
3420 if (dump_file)
3421 fprintf (dump_file, "Conversion succeeded on pass %d.\n", pass);
3422 /* Set this so we continue looking. */
3423 cond_exec_changed_p = TRUE;
3424 return ce_info.test_bb;
3425 }
3426
3427 /* Return true if a block has two edges, one of which falls through to the next
3428 block, and the other jumps to a specific block, so that we can tell if the
3429 block is part of an && test or an || test. Returns either -1 or the number
3430 of non-note, non-jump, non-USE/CLOBBER insns in the block. */
3431
3432 static int
3433 block_jumps_and_fallthru_p (basic_block cur_bb, basic_block target_bb)
3434 {
3435 edge cur_edge;
3436 int fallthru_p = FALSE;
3437 int jump_p = FALSE;
3438 rtx_insn *insn;
3439 rtx_insn *end;
3440 int n_insns = 0;
3441 edge_iterator ei;
3442
3443 if (!cur_bb || !target_bb)
3444 return -1;
3445
3446 /* If no edges, obviously it doesn't jump or fallthru. */
3447 if (EDGE_COUNT (cur_bb->succs) == 0)
3448 return FALSE;
3449
3450 FOR_EACH_EDGE (cur_edge, ei, cur_bb->succs)
3451 {
3452 if (cur_edge->flags & EDGE_COMPLEX)
3453 /* Anything complex isn't what we want. */
3454 return -1;
3455
3456 else if (cur_edge->flags & EDGE_FALLTHRU)
3457 fallthru_p = TRUE;
3458
3459 else if (cur_edge->dest == target_bb)
3460 jump_p = TRUE;
3461
3462 else
3463 return -1;
3464 }
3465
3466 if ((jump_p & fallthru_p) == 0)
3467 return -1;
3468
3469 /* Don't allow calls in the block, since this is used to group && and ||
3470 together for conditional execution support. ??? we should support
3471 conditional execution support across calls for IA-64 some day, but
3472 for now it makes the code simpler. */
3473 end = BB_END (cur_bb);
3474 insn = BB_HEAD (cur_bb);
3475
3476 while (insn != NULL_RTX)
3477 {
3478 if (CALL_P (insn))
3479 return -1;
3480
3481 if (INSN_P (insn)
3482 && !JUMP_P (insn)
3483 && !DEBUG_INSN_P (insn)
3484 && GET_CODE (PATTERN (insn)) != USE
3485 && GET_CODE (PATTERN (insn)) != CLOBBER)
3486 n_insns++;
3487
3488 if (insn == end)
3489 break;
3490
3491 insn = NEXT_INSN (insn);
3492 }
3493
3494 return n_insns;
3495 }
3496
3497 /* Determine if a given basic block heads a simple IF-THEN or IF-THEN-ELSE
3498 block. If so, we'll try to convert the insns to not require the branch.
3499 Return TRUE if we were successful at converting the block. */
3500
3501 static int
3502 cond_exec_find_if_block (struct ce_if_block * ce_info)
3503 {
3504 basic_block test_bb = ce_info->test_bb;
3505 basic_block then_bb = ce_info->then_bb;
3506 basic_block else_bb = ce_info->else_bb;
3507 basic_block join_bb = NULL_BLOCK;
3508 edge cur_edge;
3509 basic_block next;
3510 edge_iterator ei;
3511
3512 ce_info->last_test_bb = test_bb;
3513
3514 /* We only ever should get here after reload,
3515 and if we have conditional execution. */
3516 gcc_assert (reload_completed && targetm.have_conditional_execution ());
3517
3518 /* Discover if any fall through predecessors of the current test basic block
3519 were && tests (which jump to the else block) or || tests (which jump to
3520 the then block). */
3521 if (single_pred_p (test_bb)
3522 && single_pred_edge (test_bb)->flags == EDGE_FALLTHRU)
3523 {
3524 basic_block bb = single_pred (test_bb);
3525 basic_block target_bb;
3526 int max_insns = MAX_CONDITIONAL_EXECUTE;
3527 int n_insns;
3528
3529 /* Determine if the preceding block is an && or || block. */
3530 if ((n_insns = block_jumps_and_fallthru_p (bb, else_bb)) >= 0)
3531 {
3532 ce_info->and_and_p = TRUE;
3533 target_bb = else_bb;
3534 }
3535 else if ((n_insns = block_jumps_and_fallthru_p (bb, then_bb)) >= 0)
3536 {
3537 ce_info->and_and_p = FALSE;
3538 target_bb = then_bb;
3539 }
3540 else
3541 target_bb = NULL_BLOCK;
3542
3543 if (target_bb && n_insns <= max_insns)
3544 {
3545 int total_insns = 0;
3546 int blocks = 0;
3547
3548 ce_info->last_test_bb = test_bb;
3549
3550 /* Found at least one && or || block, look for more. */
3551 do
3552 {
3553 ce_info->test_bb = test_bb = bb;
3554 total_insns += n_insns;
3555 blocks++;
3556
3557 if (!single_pred_p (bb))
3558 break;
3559
3560 bb = single_pred (bb);
3561 n_insns = block_jumps_and_fallthru_p (bb, target_bb);
3562 }
3563 while (n_insns >= 0 && (total_insns + n_insns) <= max_insns);
3564
3565 ce_info->num_multiple_test_blocks = blocks;
3566 ce_info->num_multiple_test_insns = total_insns;
3567
3568 if (ce_info->and_and_p)
3569 ce_info->num_and_and_blocks = blocks;
3570 else
3571 ce_info->num_or_or_blocks = blocks;
3572 }
3573 }
3574
3575 /* The THEN block of an IF-THEN combo must have exactly one predecessor,
3576 other than any || blocks which jump to the THEN block. */
3577 if ((EDGE_COUNT (then_bb->preds) - ce_info->num_or_or_blocks) != 1)
3578 return FALSE;
3579
3580 /* The edges of the THEN and ELSE blocks cannot have complex edges. */
3581 FOR_EACH_EDGE (cur_edge, ei, then_bb->preds)
3582 {
3583 if (cur_edge->flags & EDGE_COMPLEX)
3584 return FALSE;
3585 }
3586
3587 FOR_EACH_EDGE (cur_edge, ei, else_bb->preds)
3588 {
3589 if (cur_edge->flags & EDGE_COMPLEX)
3590 return FALSE;
3591 }
3592
3593 /* The THEN block of an IF-THEN combo must have zero or one successors. */
3594 if (EDGE_COUNT (then_bb->succs) > 0
3595 && (!single_succ_p (then_bb)
3596 || (single_succ_edge (then_bb)->flags & EDGE_COMPLEX)
3597 || (epilogue_completed
3598 && tablejump_p (BB_END (then_bb), NULL, NULL))))
3599 return FALSE;
3600
3601 /* If the THEN block has no successors, conditional execution can still
3602 make a conditional call. Don't do this unless the ELSE block has
3603 only one incoming edge -- the CFG manipulation is too ugly otherwise.
3604 Check for the last insn of the THEN block being an indirect jump, which
3605 is listed as not having any successors, but confuses the rest of the CE
3606 code processing. ??? we should fix this in the future. */
3607 if (EDGE_COUNT (then_bb->succs) == 0)
3608 {
3609 if (single_pred_p (else_bb) && else_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
3610 {
3611 rtx_insn *last_insn = BB_END (then_bb);
3612
3613 while (last_insn
3614 && NOTE_P (last_insn)
3615 && last_insn != BB_HEAD (then_bb))
3616 last_insn = PREV_INSN (last_insn);
3617
3618 if (last_insn
3619 && JUMP_P (last_insn)
3620 && ! simplejump_p (last_insn))
3621 return FALSE;
3622
3623 join_bb = else_bb;
3624 else_bb = NULL_BLOCK;
3625 }
3626 else
3627 return FALSE;
3628 }
3629
3630 /* If the THEN block's successor is the other edge out of the TEST block,
3631 then we have an IF-THEN combo without an ELSE. */
3632 else if (single_succ (then_bb) == else_bb)
3633 {
3634 join_bb = else_bb;
3635 else_bb = NULL_BLOCK;
3636 }
3637
3638 /* If the THEN and ELSE block meet in a subsequent block, and the ELSE
3639 has exactly one predecessor and one successor, and the outgoing edge
3640 is not complex, then we have an IF-THEN-ELSE combo. */
3641 else if (single_succ_p (else_bb)
3642 && single_succ (then_bb) == single_succ (else_bb)
3643 && single_pred_p (else_bb)
3644 && !(single_succ_edge (else_bb)->flags & EDGE_COMPLEX)
3645 && !(epilogue_completed
3646 && tablejump_p (BB_END (else_bb), NULL, NULL)))
3647 join_bb = single_succ (else_bb);
3648
3649 /* Otherwise it is not an IF-THEN or IF-THEN-ELSE combination. */
3650 else
3651 return FALSE;
3652
3653 num_possible_if_blocks++;
3654
3655 if (dump_file)
3656 {
3657 fprintf (dump_file,
3658 "\nIF-THEN%s block found, pass %d, start block %d "
3659 "[insn %d], then %d [%d]",
3660 (else_bb) ? "-ELSE" : "",
3661 ce_info->pass,
3662 test_bb->index,
3663 BB_HEAD (test_bb) ? (int)INSN_UID (BB_HEAD (test_bb)) : -1,
3664 then_bb->index,
3665 BB_HEAD (then_bb) ? (int)INSN_UID (BB_HEAD (then_bb)) : -1);
3666
3667 if (else_bb)
3668 fprintf (dump_file, ", else %d [%d]",
3669 else_bb->index,
3670 BB_HEAD (else_bb) ? (int)INSN_UID (BB_HEAD (else_bb)) : -1);
3671
3672 fprintf (dump_file, ", join %d [%d]",
3673 join_bb->index,
3674 BB_HEAD (join_bb) ? (int)INSN_UID (BB_HEAD (join_bb)) : -1);
3675
3676 if (ce_info->num_multiple_test_blocks > 0)
3677 fprintf (dump_file, ", %d %s block%s last test %d [%d]",
3678 ce_info->num_multiple_test_blocks,
3679 (ce_info->and_and_p) ? "&&" : "||",
3680 (ce_info->num_multiple_test_blocks == 1) ? "" : "s",
3681 ce_info->last_test_bb->index,
3682 ((BB_HEAD (ce_info->last_test_bb))
3683 ? (int)INSN_UID (BB_HEAD (ce_info->last_test_bb))
3684 : -1));
3685
3686 fputc ('\n', dump_file);
3687 }
3688
3689 /* Make sure IF, THEN, and ELSE, blocks are adjacent. Actually, we get the
3690 first condition for free, since we've already asserted that there's a
3691 fallthru edge from IF to THEN. Likewise for the && and || blocks, since
3692 we checked the FALLTHRU flag, those are already adjacent to the last IF
3693 block. */
3694 /* ??? As an enhancement, move the ELSE block. Have to deal with
3695 BLOCK notes, if by no other means than backing out the merge if they
3696 exist. Sticky enough I don't want to think about it now. */
3697 next = then_bb;
3698 if (else_bb && (next = next->next_bb) != else_bb)
3699 return FALSE;
3700 if ((next = next->next_bb) != join_bb
3701 && join_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
3702 {
3703 if (else_bb)
3704 join_bb = NULL;
3705 else
3706 return FALSE;
3707 }
3708
3709 /* Do the real work. */
3710
3711 ce_info->else_bb = else_bb;
3712 ce_info->join_bb = join_bb;
3713
3714 /* If we have && and || tests, try to first handle combining the && and ||
3715 tests into the conditional code, and if that fails, go back and handle
3716 it without the && and ||, which at present handles the && case if there
3717 was no ELSE block. */
3718 if (cond_exec_process_if_block (ce_info, TRUE))
3719 return TRUE;
3720
3721 if (ce_info->num_multiple_test_blocks)
3722 {
3723 cancel_changes (0);
3724
3725 if (cond_exec_process_if_block (ce_info, FALSE))
3726 return TRUE;
3727 }
3728
3729 return FALSE;
3730 }
3731
3732 /* Convert a branch over a trap, or a branch
3733 to a trap, into a conditional trap. */
3734
3735 static int
3736 find_cond_trap (basic_block test_bb, edge then_edge, edge else_edge)
3737 {
3738 basic_block then_bb = then_edge->dest;
3739 basic_block else_bb = else_edge->dest;
3740 basic_block other_bb, trap_bb;
3741 rtx_insn *trap, *jump;
3742 rtx cond, seq;
3743 rtx_insn *cond_earliest;
3744 enum rtx_code code;
3745
3746 /* Locate the block with the trap instruction. */
3747 /* ??? While we look for no successors, we really ought to allow
3748 EH successors. Need to fix merge_if_block for that to work. */
3749 if ((trap = block_has_only_trap (then_bb)) != NULL)
3750 trap_bb = then_bb, other_bb = else_bb;
3751 else if ((trap = block_has_only_trap (else_bb)) != NULL)
3752 trap_bb = else_bb, other_bb = then_bb;
3753 else
3754 return FALSE;
3755
3756 if (dump_file)
3757 {
3758 fprintf (dump_file, "\nTRAP-IF block found, start %d, trap %d\n",
3759 test_bb->index, trap_bb->index);
3760 }
3761
3762 /* If this is not a standard conditional jump, we can't parse it. */
3763 jump = BB_END (test_bb);
3764 cond = noce_get_condition (jump, &cond_earliest, false);
3765 if (! cond)
3766 return FALSE;
3767
3768 /* If the conditional jump is more than just a conditional jump, then
3769 we can not do if-conversion on this block. */
3770 if (! onlyjump_p (jump))
3771 return FALSE;
3772
3773 /* We must be comparing objects whose modes imply the size. */
3774 if (GET_MODE (XEXP (cond, 0)) == BLKmode)
3775 return FALSE;
3776
3777 /* Reverse the comparison code, if necessary. */
3778 code = GET_CODE (cond);
3779 if (then_bb == trap_bb)
3780 {
3781 code = reversed_comparison_code (cond, jump);
3782 if (code == UNKNOWN)
3783 return FALSE;
3784 }
3785
3786 /* Attempt to generate the conditional trap. */
3787 seq = gen_cond_trap (code, copy_rtx (XEXP (cond, 0)),
3788 copy_rtx (XEXP (cond, 1)),
3789 TRAP_CODE (PATTERN (trap)));
3790 if (seq == NULL)
3791 return FALSE;
3792
3793 /* Emit the new insns before cond_earliest. */
3794 emit_insn_before_setloc (seq, cond_earliest, INSN_LOCATION (trap));
3795
3796 /* Delete the trap block if possible. */
3797 remove_edge (trap_bb == then_bb ? then_edge : else_edge);
3798 df_set_bb_dirty (test_bb);
3799 df_set_bb_dirty (then_bb);
3800 df_set_bb_dirty (else_bb);
3801
3802 if (EDGE_COUNT (trap_bb->preds) == 0)
3803 {
3804 delete_basic_block (trap_bb);
3805 num_true_changes++;
3806 }
3807
3808 /* Wire together the blocks again. */
3809 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3810 single_succ_edge (test_bb)->flags |= EDGE_FALLTHRU;
3811 else if (trap_bb == then_bb)
3812 {
3813 rtx lab;
3814 rtx_insn *newjump;
3815
3816 lab = JUMP_LABEL (jump);
3817 newjump = emit_jump_insn_after (gen_jump (lab), jump);
3818 LABEL_NUSES (lab) += 1;
3819 JUMP_LABEL (newjump) = lab;
3820 emit_barrier_after (newjump);
3821 }
3822 delete_insn (jump);
3823
3824 if (can_merge_blocks_p (test_bb, other_bb))
3825 {
3826 merge_blocks (test_bb, other_bb);
3827 num_true_changes++;
3828 }
3829
3830 num_updated_if_blocks++;
3831 return TRUE;
3832 }
3833
3834 /* Subroutine of find_cond_trap: if BB contains only a trap insn,
3835 return it. */
3836
3837 static rtx_insn *
3838 block_has_only_trap (basic_block bb)
3839 {
3840 rtx_insn *trap;
3841
3842 /* We're not the exit block. */
3843 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
3844 return NULL;
3845
3846 /* The block must have no successors. */
3847 if (EDGE_COUNT (bb->succs) > 0)
3848 return NULL;
3849
3850 /* The only instruction in the THEN block must be the trap. */
3851 trap = first_active_insn (bb);
3852 if (! (trap == BB_END (bb)
3853 && GET_CODE (PATTERN (trap)) == TRAP_IF
3854 && TRAP_CONDITION (PATTERN (trap)) == const_true_rtx))
3855 return NULL;
3856
3857 return trap;
3858 }
3859
3860 /* Look for IF-THEN-ELSE cases in which one of THEN or ELSE is
3861 transformable, but not necessarily the other. There need be no
3862 JOIN block.
3863
3864 Return TRUE if we were successful at converting the block.
3865
3866 Cases we'd like to look at:
3867
3868 (1)
3869 if (test) goto over; // x not live
3870 x = a;
3871 goto label;
3872 over:
3873
3874 becomes
3875
3876 x = a;
3877 if (! test) goto label;
3878
3879 (2)
3880 if (test) goto E; // x not live
3881 x = big();
3882 goto L;
3883 E:
3884 x = b;
3885 goto M;
3886
3887 becomes
3888
3889 x = b;
3890 if (test) goto M;
3891 x = big();
3892 goto L;
3893
3894 (3) // This one's really only interesting for targets that can do
3895 // multiway branching, e.g. IA-64 BBB bundles. For other targets
3896 // it results in multiple branches on a cache line, which often
3897 // does not sit well with predictors.
3898
3899 if (test1) goto E; // predicted not taken
3900 x = a;
3901 if (test2) goto F;
3902 ...
3903 E:
3904 x = b;
3905 J:
3906
3907 becomes
3908
3909 x = a;
3910 if (test1) goto E;
3911 if (test2) goto F;
3912
3913 Notes:
3914
3915 (A) Don't do (2) if the branch is predicted against the block we're
3916 eliminating. Do it anyway if we can eliminate a branch; this requires
3917 that the sole successor of the eliminated block postdominate the other
3918 side of the if.
3919
3920 (B) With CE, on (3) we can steal from both sides of the if, creating
3921
3922 if (test1) x = a;
3923 if (!test1) x = b;
3924 if (test1) goto J;
3925 if (test2) goto F;
3926 ...
3927 J:
3928
3929 Again, this is most useful if J postdominates.
3930
3931 (C) CE substitutes for helpful life information.
3932
3933 (D) These heuristics need a lot of work. */
3934
3935 /* Tests for case 1 above. */
3936
3937 static int
3938 find_if_case_1 (basic_block test_bb, edge then_edge, edge else_edge)
3939 {
3940 basic_block then_bb = then_edge->dest;
3941 basic_block else_bb = else_edge->dest;
3942 basic_block new_bb;
3943 int then_bb_index, then_prob;
3944 rtx else_target = NULL_RTX;
3945
3946 /* If we are partitioning hot/cold basic blocks, we don't want to
3947 mess up unconditional or indirect jumps that cross between hot
3948 and cold sections.
3949
3950 Basic block partitioning may result in some jumps that appear to
3951 be optimizable (or blocks that appear to be mergeable), but which really
3952 must be left untouched (they are required to make it safely across
3953 partition boundaries). See the comments at the top of
3954 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
3955
3956 if ((BB_END (then_bb)
3957 && JUMP_P (BB_END (then_bb))
3958 && CROSSING_JUMP_P (BB_END (then_bb)))
3959 || (BB_END (test_bb)
3960 && JUMP_P (BB_END (test_bb))
3961 && CROSSING_JUMP_P (BB_END (test_bb)))
3962 || (BB_END (else_bb)
3963 && JUMP_P (BB_END (else_bb))
3964 && CROSSING_JUMP_P (BB_END (else_bb))))
3965 return FALSE;
3966
3967 /* THEN has one successor. */
3968 if (!single_succ_p (then_bb))
3969 return FALSE;
3970
3971 /* THEN does not fall through, but is not strange either. */
3972 if (single_succ_edge (then_bb)->flags & (EDGE_COMPLEX | EDGE_FALLTHRU))
3973 return FALSE;
3974
3975 /* THEN has one predecessor. */
3976 if (!single_pred_p (then_bb))
3977 return FALSE;
3978
3979 /* THEN must do something. */
3980 if (forwarder_block_p (then_bb))
3981 return FALSE;
3982
3983 num_possible_if_blocks++;
3984 if (dump_file)
3985 fprintf (dump_file,
3986 "\nIF-CASE-1 found, start %d, then %d\n",
3987 test_bb->index, then_bb->index);
3988
3989 if (then_edge->probability)
3990 then_prob = REG_BR_PROB_BASE - then_edge->probability;
3991 else
3992 then_prob = REG_BR_PROB_BASE / 2;
3993
3994 /* We're speculating from the THEN path, we want to make sure the cost
3995 of speculation is within reason. */
3996 if (! cheap_bb_rtx_cost_p (then_bb, then_prob,
3997 COSTS_N_INSNS (BRANCH_COST (optimize_bb_for_speed_p (then_edge->src),
3998 predictable_edge_p (then_edge)))))
3999 return FALSE;
4000
4001 if (else_bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
4002 {
4003 rtx_insn *jump = BB_END (else_edge->src);
4004 gcc_assert (JUMP_P (jump));
4005 else_target = JUMP_LABEL (jump);
4006 }
4007
4008 /* Registers set are dead, or are predicable. */
4009 if (! dead_or_predicable (test_bb, then_bb, else_bb,
4010 single_succ_edge (then_bb), 1))
4011 return FALSE;
4012
4013 /* Conversion went ok, including moving the insns and fixing up the
4014 jump. Adjust the CFG to match. */
4015
4016 /* We can avoid creating a new basic block if then_bb is immediately
4017 followed by else_bb, i.e. deleting then_bb allows test_bb to fall
4018 through to else_bb. */
4019
4020 if (then_bb->next_bb == else_bb
4021 && then_bb->prev_bb == test_bb
4022 && else_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
4023 {
4024 redirect_edge_succ (FALLTHRU_EDGE (test_bb), else_bb);
4025 new_bb = 0;
4026 }
4027 else if (else_bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
4028 new_bb = force_nonfallthru_and_redirect (FALLTHRU_EDGE (test_bb),
4029 else_bb, else_target);
4030 else
4031 new_bb = redirect_edge_and_branch_force (FALLTHRU_EDGE (test_bb),
4032 else_bb);
4033
4034 df_set_bb_dirty (test_bb);
4035 df_set_bb_dirty (else_bb);
4036
4037 then_bb_index = then_bb->index;
4038 delete_basic_block (then_bb);
4039
4040 /* Make rest of code believe that the newly created block is the THEN_BB
4041 block we removed. */
4042 if (new_bb)
4043 {
4044 df_bb_replace (then_bb_index, new_bb);
4045 /* This should have been done above via force_nonfallthru_and_redirect
4046 (possibly called from redirect_edge_and_branch_force). */
4047 gcc_checking_assert (BB_PARTITION (new_bb) == BB_PARTITION (test_bb));
4048 }
4049
4050 num_true_changes++;
4051 num_updated_if_blocks++;
4052
4053 return TRUE;
4054 }
4055
4056 /* Test for case 2 above. */
4057
4058 static int
4059 find_if_case_2 (basic_block test_bb, edge then_edge, edge else_edge)
4060 {
4061 basic_block then_bb = then_edge->dest;
4062 basic_block else_bb = else_edge->dest;
4063 edge else_succ;
4064 int then_prob, else_prob;
4065
4066 /* We do not want to speculate (empty) loop latches. */
4067 if (current_loops
4068 && else_bb->loop_father->latch == else_bb)
4069 return FALSE;
4070
4071 /* If we are partitioning hot/cold basic blocks, we don't want to
4072 mess up unconditional or indirect jumps that cross between hot
4073 and cold sections.
4074
4075 Basic block partitioning may result in some jumps that appear to
4076 be optimizable (or blocks that appear to be mergeable), but which really
4077 must be left untouched (they are required to make it safely across
4078 partition boundaries). See the comments at the top of
4079 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
4080
4081 if ((BB_END (then_bb)
4082 && JUMP_P (BB_END (then_bb))
4083 && CROSSING_JUMP_P (BB_END (then_bb)))
4084 || (BB_END (test_bb)
4085 && JUMP_P (BB_END (test_bb))
4086 && CROSSING_JUMP_P (BB_END (test_bb)))
4087 || (BB_END (else_bb)
4088 && JUMP_P (BB_END (else_bb))
4089 && CROSSING_JUMP_P (BB_END (else_bb))))
4090 return FALSE;
4091
4092 /* ELSE has one successor. */
4093 if (!single_succ_p (else_bb))
4094 return FALSE;
4095 else
4096 else_succ = single_succ_edge (else_bb);
4097
4098 /* ELSE outgoing edge is not complex. */
4099 if (else_succ->flags & EDGE_COMPLEX)
4100 return FALSE;
4101
4102 /* ELSE has one predecessor. */
4103 if (!single_pred_p (else_bb))
4104 return FALSE;
4105
4106 /* THEN is not EXIT. */
4107 if (then_bb->index < NUM_FIXED_BLOCKS)
4108 return FALSE;
4109
4110 if (else_edge->probability)
4111 {
4112 else_prob = else_edge->probability;
4113 then_prob = REG_BR_PROB_BASE - else_prob;
4114 }
4115 else
4116 {
4117 else_prob = REG_BR_PROB_BASE / 2;
4118 then_prob = REG_BR_PROB_BASE / 2;
4119 }
4120
4121 /* ELSE is predicted or SUCC(ELSE) postdominates THEN. */
4122 if (else_prob > then_prob)
4123 ;
4124 else if (else_succ->dest->index < NUM_FIXED_BLOCKS
4125 || dominated_by_p (CDI_POST_DOMINATORS, then_bb,
4126 else_succ->dest))
4127 ;
4128 else
4129 return FALSE;
4130
4131 num_possible_if_blocks++;
4132 if (dump_file)
4133 fprintf (dump_file,
4134 "\nIF-CASE-2 found, start %d, else %d\n",
4135 test_bb->index, else_bb->index);
4136
4137 /* We're speculating from the ELSE path, we want to make sure the cost
4138 of speculation is within reason. */
4139 if (! cheap_bb_rtx_cost_p (else_bb, else_prob,
4140 COSTS_N_INSNS (BRANCH_COST (optimize_bb_for_speed_p (else_edge->src),
4141 predictable_edge_p (else_edge)))))
4142 return FALSE;
4143
4144 /* Registers set are dead, or are predicable. */
4145 if (! dead_or_predicable (test_bb, else_bb, then_bb, else_succ, 0))
4146 return FALSE;
4147
4148 /* Conversion went ok, including moving the insns and fixing up the
4149 jump. Adjust the CFG to match. */
4150
4151 df_set_bb_dirty (test_bb);
4152 df_set_bb_dirty (then_bb);
4153 delete_basic_block (else_bb);
4154
4155 num_true_changes++;
4156 num_updated_if_blocks++;
4157
4158 /* ??? We may now fallthru from one of THEN's successors into a join
4159 block. Rerun cleanup_cfg? Examine things manually? Wait? */
4160
4161 return TRUE;
4162 }
4163
4164 /* Used by the code above to perform the actual rtl transformations.
4165 Return TRUE if successful.
4166
4167 TEST_BB is the block containing the conditional branch. MERGE_BB
4168 is the block containing the code to manipulate. DEST_EDGE is an
4169 edge representing a jump to the join block; after the conversion,
4170 TEST_BB should be branching to its destination.
4171 REVERSEP is true if the sense of the branch should be reversed. */
4172
4173 static int
4174 dead_or_predicable (basic_block test_bb, basic_block merge_bb,
4175 basic_block other_bb, edge dest_edge, int reversep)
4176 {
4177 basic_block new_dest = dest_edge->dest;
4178 rtx_insn *head, *end, *jump;
4179 rtx_insn *earliest = NULL;
4180 rtx old_dest;
4181 bitmap merge_set = NULL;
4182 /* Number of pending changes. */
4183 int n_validated_changes = 0;
4184 rtx new_dest_label = NULL_RTX;
4185
4186 jump = BB_END (test_bb);
4187
4188 /* Find the extent of the real code in the merge block. */
4189 head = BB_HEAD (merge_bb);
4190 end = BB_END (merge_bb);
4191
4192 while (DEBUG_INSN_P (end) && end != head)
4193 end = PREV_INSN (end);
4194
4195 /* If merge_bb ends with a tablejump, predicating/moving insn's
4196 into test_bb and then deleting merge_bb will result in the jumptable
4197 that follows merge_bb being removed along with merge_bb and then we
4198 get an unresolved reference to the jumptable. */
4199 if (tablejump_p (end, NULL, NULL))
4200 return FALSE;
4201
4202 if (LABEL_P (head))
4203 head = NEXT_INSN (head);
4204 while (DEBUG_INSN_P (head) && head != end)
4205 head = NEXT_INSN (head);
4206 if (NOTE_P (head))
4207 {
4208 if (head == end)
4209 {
4210 head = end = NULL;
4211 goto no_body;
4212 }
4213 head = NEXT_INSN (head);
4214 while (DEBUG_INSN_P (head) && head != end)
4215 head = NEXT_INSN (head);
4216 }
4217
4218 if (JUMP_P (end))
4219 {
4220 if (!onlyjump_p (end))
4221 return FALSE;
4222 if (head == end)
4223 {
4224 head = end = NULL;
4225 goto no_body;
4226 }
4227 end = PREV_INSN (end);
4228 while (DEBUG_INSN_P (end) && end != head)
4229 end = PREV_INSN (end);
4230 }
4231
4232 /* Don't move frame-related insn across the conditional branch. This
4233 can lead to one of the paths of the branch having wrong unwind info. */
4234 if (epilogue_completed)
4235 {
4236 rtx_insn *insn = head;
4237 while (1)
4238 {
4239 if (INSN_P (insn) && RTX_FRAME_RELATED_P (insn))
4240 return FALSE;
4241 if (insn == end)
4242 break;
4243 insn = NEXT_INSN (insn);
4244 }
4245 }
4246
4247 /* Disable handling dead code by conditional execution if the machine needs
4248 to do anything funny with the tests, etc. */
4249 #ifndef IFCVT_MODIFY_TESTS
4250 if (targetm.have_conditional_execution ())
4251 {
4252 /* In the conditional execution case, we have things easy. We know
4253 the condition is reversible. We don't have to check life info
4254 because we're going to conditionally execute the code anyway.
4255 All that's left is making sure the insns involved can actually
4256 be predicated. */
4257
4258 rtx cond;
4259
4260 cond = cond_exec_get_condition (jump);
4261 if (! cond)
4262 return FALSE;
4263
4264 rtx note = find_reg_note (jump, REG_BR_PROB, NULL_RTX);
4265 int prob_val = (note ? XINT (note, 0) : -1);
4266
4267 if (reversep)
4268 {
4269 enum rtx_code rev = reversed_comparison_code (cond, jump);
4270 if (rev == UNKNOWN)
4271 return FALSE;
4272 cond = gen_rtx_fmt_ee (rev, GET_MODE (cond), XEXP (cond, 0),
4273 XEXP (cond, 1));
4274 if (prob_val >= 0)
4275 prob_val = REG_BR_PROB_BASE - prob_val;
4276 }
4277
4278 if (cond_exec_process_insns (NULL, head, end, cond, prob_val, 0)
4279 && verify_changes (0))
4280 n_validated_changes = num_validated_changes ();
4281 else
4282 cancel_changes (0);
4283
4284 earliest = jump;
4285 }
4286 #endif
4287
4288 /* If we allocated new pseudos (e.g. in the conditional move
4289 expander called from noce_emit_cmove), we must resize the
4290 array first. */
4291 if (max_regno < max_reg_num ())
4292 max_regno = max_reg_num ();
4293
4294 /* Try the NCE path if the CE path did not result in any changes. */
4295 if (n_validated_changes == 0)
4296 {
4297 rtx cond;
4298 rtx_insn *insn;
4299 regset live;
4300 bool success;
4301
4302 /* In the non-conditional execution case, we have to verify that there
4303 are no trapping operations, no calls, no references to memory, and
4304 that any registers modified are dead at the branch site. */
4305
4306 if (!any_condjump_p (jump))
4307 return FALSE;
4308
4309 /* Find the extent of the conditional. */
4310 cond = noce_get_condition (jump, &earliest, false);
4311 if (!cond)
4312 return FALSE;
4313
4314 live = BITMAP_ALLOC (&reg_obstack);
4315 simulate_backwards_to_point (merge_bb, live, end);
4316 success = can_move_insns_across (head, end, earliest, jump,
4317 merge_bb, live,
4318 df_get_live_in (other_bb), NULL);
4319 BITMAP_FREE (live);
4320 if (!success)
4321 return FALSE;
4322
4323 /* Collect the set of registers set in MERGE_BB. */
4324 merge_set = BITMAP_ALLOC (&reg_obstack);
4325
4326 FOR_BB_INSNS (merge_bb, insn)
4327 if (NONDEBUG_INSN_P (insn))
4328 df_simulate_find_defs (insn, merge_set);
4329
4330 /* If shrink-wrapping, disable this optimization when test_bb is
4331 the first basic block and merge_bb exits. The idea is to not
4332 move code setting up a return register as that may clobber a
4333 register used to pass function parameters, which then must be
4334 saved in caller-saved regs. A caller-saved reg requires the
4335 prologue, killing a shrink-wrap opportunity. */
4336 if ((SHRINK_WRAPPING_ENABLED && !epilogue_completed)
4337 && ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb == test_bb
4338 && single_succ_p (new_dest)
4339 && single_succ (new_dest) == EXIT_BLOCK_PTR_FOR_FN (cfun)
4340 && bitmap_intersect_p (df_get_live_in (new_dest), merge_set))
4341 {
4342 regset return_regs;
4343 unsigned int i;
4344
4345 return_regs = BITMAP_ALLOC (&reg_obstack);
4346
4347 /* Start off with the intersection of regs used to pass
4348 params and regs used to return values. */
4349 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
4350 if (FUNCTION_ARG_REGNO_P (i)
4351 && targetm.calls.function_value_regno_p (i))
4352 bitmap_set_bit (return_regs, INCOMING_REGNO (i));
4353
4354 bitmap_and_into (return_regs,
4355 df_get_live_out (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
4356 bitmap_and_into (return_regs,
4357 df_get_live_in (EXIT_BLOCK_PTR_FOR_FN (cfun)));
4358 if (!bitmap_empty_p (return_regs))
4359 {
4360 FOR_BB_INSNS_REVERSE (new_dest, insn)
4361 if (NONDEBUG_INSN_P (insn))
4362 {
4363 df_ref def;
4364
4365 /* If this insn sets any reg in return_regs, add all
4366 reg uses to the set of regs we're interested in. */
4367 FOR_EACH_INSN_DEF (def, insn)
4368 if (bitmap_bit_p (return_regs, DF_REF_REGNO (def)))
4369 {
4370 df_simulate_uses (insn, return_regs);
4371 break;
4372 }
4373 }
4374 if (bitmap_intersect_p (merge_set, return_regs))
4375 {
4376 BITMAP_FREE (return_regs);
4377 BITMAP_FREE (merge_set);
4378 return FALSE;
4379 }
4380 }
4381 BITMAP_FREE (return_regs);
4382 }
4383 }
4384
4385 no_body:
4386 /* We don't want to use normal invert_jump or redirect_jump because
4387 we don't want to delete_insn called. Also, we want to do our own
4388 change group management. */
4389
4390 old_dest = JUMP_LABEL (jump);
4391 if (other_bb != new_dest)
4392 {
4393 if (!any_condjump_p (jump))
4394 goto cancel;
4395
4396 if (JUMP_P (BB_END (dest_edge->src)))
4397 new_dest_label = JUMP_LABEL (BB_END (dest_edge->src));
4398 else if (new_dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
4399 new_dest_label = ret_rtx;
4400 else
4401 new_dest_label = block_label (new_dest);
4402
4403 if (reversep
4404 ? ! invert_jump_1 (jump, new_dest_label)
4405 : ! redirect_jump_1 (jump, new_dest_label))
4406 goto cancel;
4407 }
4408
4409 if (verify_changes (n_validated_changes))
4410 confirm_change_group ();
4411 else
4412 goto cancel;
4413
4414 if (other_bb != new_dest)
4415 {
4416 redirect_jump_2 (jump, old_dest, new_dest_label, 0, reversep);
4417
4418 redirect_edge_succ (BRANCH_EDGE (test_bb), new_dest);
4419 if (reversep)
4420 {
4421 gcov_type count, probability;
4422 count = BRANCH_EDGE (test_bb)->count;
4423 BRANCH_EDGE (test_bb)->count = FALLTHRU_EDGE (test_bb)->count;
4424 FALLTHRU_EDGE (test_bb)->count = count;
4425 probability = BRANCH_EDGE (test_bb)->probability;
4426 BRANCH_EDGE (test_bb)->probability
4427 = FALLTHRU_EDGE (test_bb)->probability;
4428 FALLTHRU_EDGE (test_bb)->probability = probability;
4429 update_br_prob_note (test_bb);
4430 }
4431 }
4432
4433 /* Move the insns out of MERGE_BB to before the branch. */
4434 if (head != NULL)
4435 {
4436 rtx_insn *insn;
4437
4438 if (end == BB_END (merge_bb))
4439 BB_END (merge_bb) = PREV_INSN (head);
4440
4441 /* PR 21767: when moving insns above a conditional branch, the REG_EQUAL
4442 notes being moved might become invalid. */
4443 insn = head;
4444 do
4445 {
4446 rtx note;
4447
4448 if (! INSN_P (insn))
4449 continue;
4450 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4451 if (! note)
4452 continue;
4453 remove_note (insn, note);
4454 } while (insn != end && (insn = NEXT_INSN (insn)));
4455
4456 /* PR46315: when moving insns above a conditional branch, the REG_EQUAL
4457 notes referring to the registers being set might become invalid. */
4458 if (merge_set)
4459 {
4460 unsigned i;
4461 bitmap_iterator bi;
4462
4463 EXECUTE_IF_SET_IN_BITMAP (merge_set, 0, i, bi)
4464 remove_reg_equal_equiv_notes_for_regno (i);
4465
4466 BITMAP_FREE (merge_set);
4467 }
4468
4469 reorder_insns (head, end, PREV_INSN (earliest));
4470 }
4471
4472 /* Remove the jump and edge if we can. */
4473 if (other_bb == new_dest)
4474 {
4475 delete_insn (jump);
4476 remove_edge (BRANCH_EDGE (test_bb));
4477 /* ??? Can't merge blocks here, as then_bb is still in use.
4478 At minimum, the merge will get done just before bb-reorder. */
4479 }
4480
4481 return TRUE;
4482
4483 cancel:
4484 cancel_changes (0);
4485
4486 if (merge_set)
4487 BITMAP_FREE (merge_set);
4488
4489 return FALSE;
4490 }
4491 \f
4492 /* Main entry point for all if-conversion. AFTER_COMBINE is true if
4493 we are after combine pass. */
4494
4495 static void
4496 if_convert (bool after_combine)
4497 {
4498 basic_block bb;
4499 int pass;
4500
4501 if (optimize == 1)
4502 {
4503 df_live_add_problem ();
4504 df_live_set_all_dirty ();
4505 }
4506
4507 /* Record whether we are after combine pass. */
4508 ifcvt_after_combine = after_combine;
4509 num_possible_if_blocks = 0;
4510 num_updated_if_blocks = 0;
4511 num_true_changes = 0;
4512
4513 loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
4514 mark_loop_exit_edges ();
4515 loop_optimizer_finalize ();
4516 free_dominance_info (CDI_DOMINATORS);
4517
4518 /* Compute postdominators. */
4519 calculate_dominance_info (CDI_POST_DOMINATORS);
4520
4521 df_set_flags (DF_LR_RUN_DCE);
4522
4523 /* Go through each of the basic blocks looking for things to convert. If we
4524 have conditional execution, we make multiple passes to allow us to handle
4525 IF-THEN{-ELSE} blocks within other IF-THEN{-ELSE} blocks. */
4526 pass = 0;
4527 do
4528 {
4529 df_analyze ();
4530 /* Only need to do dce on the first pass. */
4531 df_clear_flags (DF_LR_RUN_DCE);
4532 cond_exec_changed_p = FALSE;
4533 pass++;
4534
4535 #ifdef IFCVT_MULTIPLE_DUMPS
4536 if (dump_file && pass > 1)
4537 fprintf (dump_file, "\n\n========== Pass %d ==========\n", pass);
4538 #endif
4539
4540 FOR_EACH_BB_FN (bb, cfun)
4541 {
4542 basic_block new_bb;
4543 while (!df_get_bb_dirty (bb)
4544 && (new_bb = find_if_header (bb, pass)) != NULL)
4545 bb = new_bb;
4546 }
4547
4548 #ifdef IFCVT_MULTIPLE_DUMPS
4549 if (dump_file && cond_exec_changed_p)
4550 print_rtl_with_bb (dump_file, get_insns (), dump_flags);
4551 #endif
4552 }
4553 while (cond_exec_changed_p);
4554
4555 #ifdef IFCVT_MULTIPLE_DUMPS
4556 if (dump_file)
4557 fprintf (dump_file, "\n\n========== no more changes\n");
4558 #endif
4559
4560 free_dominance_info (CDI_POST_DOMINATORS);
4561
4562 if (dump_file)
4563 fflush (dump_file);
4564
4565 clear_aux_for_blocks ();
4566
4567 /* If we allocated new pseudos, we must resize the array for sched1. */
4568 if (max_regno < max_reg_num ())
4569 max_regno = max_reg_num ();
4570
4571 /* Write the final stats. */
4572 if (dump_file && num_possible_if_blocks > 0)
4573 {
4574 fprintf (dump_file,
4575 "\n%d possible IF blocks searched.\n",
4576 num_possible_if_blocks);
4577 fprintf (dump_file,
4578 "%d IF blocks converted.\n",
4579 num_updated_if_blocks);
4580 fprintf (dump_file,
4581 "%d true changes made.\n\n\n",
4582 num_true_changes);
4583 }
4584
4585 if (optimize == 1)
4586 df_remove_problem (df_live);
4587
4588 #ifdef ENABLE_CHECKING
4589 verify_flow_info ();
4590 #endif
4591 }
4592 \f
4593 /* If-conversion and CFG cleanup. */
4594 static unsigned int
4595 rest_of_handle_if_conversion (void)
4596 {
4597 if (flag_if_conversion)
4598 {
4599 if (dump_file)
4600 {
4601 dump_reg_info (dump_file);
4602 dump_flow_info (dump_file, dump_flags);
4603 }
4604 cleanup_cfg (CLEANUP_EXPENSIVE);
4605 if_convert (false);
4606 }
4607
4608 cleanup_cfg (0);
4609 return 0;
4610 }
4611
4612 namespace {
4613
4614 const pass_data pass_data_rtl_ifcvt =
4615 {
4616 RTL_PASS, /* type */
4617 "ce1", /* name */
4618 OPTGROUP_NONE, /* optinfo_flags */
4619 TV_IFCVT, /* tv_id */
4620 0, /* properties_required */
4621 0, /* properties_provided */
4622 0, /* properties_destroyed */
4623 0, /* todo_flags_start */
4624 TODO_df_finish, /* todo_flags_finish */
4625 };
4626
4627 class pass_rtl_ifcvt : public rtl_opt_pass
4628 {
4629 public:
4630 pass_rtl_ifcvt (gcc::context *ctxt)
4631 : rtl_opt_pass (pass_data_rtl_ifcvt, ctxt)
4632 {}
4633
4634 /* opt_pass methods: */
4635 virtual bool gate (function *)
4636 {
4637 return (optimize > 0) && dbg_cnt (if_conversion);
4638 }
4639
4640 virtual unsigned int execute (function *)
4641 {
4642 return rest_of_handle_if_conversion ();
4643 }
4644
4645 }; // class pass_rtl_ifcvt
4646
4647 } // anon namespace
4648
4649 rtl_opt_pass *
4650 make_pass_rtl_ifcvt (gcc::context *ctxt)
4651 {
4652 return new pass_rtl_ifcvt (ctxt);
4653 }
4654
4655
4656 /* Rerun if-conversion, as combine may have simplified things enough
4657 to now meet sequence length restrictions. */
4658
4659 namespace {
4660
4661 const pass_data pass_data_if_after_combine =
4662 {
4663 RTL_PASS, /* type */
4664 "ce2", /* name */
4665 OPTGROUP_NONE, /* optinfo_flags */
4666 TV_IFCVT, /* tv_id */
4667 0, /* properties_required */
4668 0, /* properties_provided */
4669 0, /* properties_destroyed */
4670 0, /* todo_flags_start */
4671 TODO_df_finish, /* todo_flags_finish */
4672 };
4673
4674 class pass_if_after_combine : public rtl_opt_pass
4675 {
4676 public:
4677 pass_if_after_combine (gcc::context *ctxt)
4678 : rtl_opt_pass (pass_data_if_after_combine, ctxt)
4679 {}
4680
4681 /* opt_pass methods: */
4682 virtual bool gate (function *)
4683 {
4684 return optimize > 0 && flag_if_conversion
4685 && dbg_cnt (if_after_combine);
4686 }
4687
4688 virtual unsigned int execute (function *)
4689 {
4690 if_convert (true);
4691 return 0;
4692 }
4693
4694 }; // class pass_if_after_combine
4695
4696 } // anon namespace
4697
4698 rtl_opt_pass *
4699 make_pass_if_after_combine (gcc::context *ctxt)
4700 {
4701 return new pass_if_after_combine (ctxt);
4702 }
4703
4704
4705 namespace {
4706
4707 const pass_data pass_data_if_after_reload =
4708 {
4709 RTL_PASS, /* type */
4710 "ce3", /* name */
4711 OPTGROUP_NONE, /* optinfo_flags */
4712 TV_IFCVT2, /* tv_id */
4713 0, /* properties_required */
4714 0, /* properties_provided */
4715 0, /* properties_destroyed */
4716 0, /* todo_flags_start */
4717 TODO_df_finish, /* todo_flags_finish */
4718 };
4719
4720 class pass_if_after_reload : public rtl_opt_pass
4721 {
4722 public:
4723 pass_if_after_reload (gcc::context *ctxt)
4724 : rtl_opt_pass (pass_data_if_after_reload, ctxt)
4725 {}
4726
4727 /* opt_pass methods: */
4728 virtual bool gate (function *)
4729 {
4730 return optimize > 0 && flag_if_conversion2
4731 && dbg_cnt (if_after_reload);
4732 }
4733
4734 virtual unsigned int execute (function *)
4735 {
4736 if_convert (true);
4737 return 0;
4738 }
4739
4740 }; // class pass_if_after_reload
4741
4742 } // anon namespace
4743
4744 rtl_opt_pass *
4745 make_pass_if_after_reload (gcc::context *ctxt)
4746 {
4747 return new pass_if_after_reload (ctxt);
4748 }