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