Suppress -fstack-protector warning on hppa.
[official-gcc.git] / gcc / ifcvt.cc
blobeb8efb89a89770fb7c9e930bf282d8159824b905
1 /* If-conversion support.
2 Copyright (C) 2000-2022 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 3, or (at your option)
9 any later version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
13 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
14 License for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "cfghooks.h"
28 #include "df.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "expmed.h"
32 #include "optabs.h"
33 #include "regs.h"
34 #include "emit-rtl.h"
35 #include "recog.h"
37 #include "cfgrtl.h"
38 #include "cfganal.h"
39 #include "cfgcleanup.h"
40 #include "expr.h"
41 #include "output.h"
42 #include "cfgloop.h"
43 #include "tree-pass.h"
44 #include "dbgcnt.h"
45 #include "shrink-wrap.h"
46 #include "rtl-iter.h"
47 #include "ifcvt.h"
49 #ifndef MAX_CONDITIONAL_EXECUTE
50 #define MAX_CONDITIONAL_EXECUTE \
51 (BRANCH_COST (optimize_function_for_speed_p (cfun), false) \
52 + 1)
53 #endif
55 #define IFCVT_MULTIPLE_DUMPS 1
57 #define NULL_BLOCK ((basic_block) NULL)
59 /* True if after combine pass. */
60 static bool ifcvt_after_combine;
62 /* True if the target has the cbranchcc4 optab. */
63 static bool have_cbranchcc4;
65 /* # of IF-THEN or IF-THEN-ELSE blocks we looked at */
66 static int num_possible_if_blocks;
68 /* # of IF-THEN or IF-THEN-ELSE blocks were converted to conditional
69 execution. */
70 static int num_updated_if_blocks;
72 /* # of changes made. */
73 static int num_true_changes;
75 /* Whether conditional execution changes were made. */
76 static int cond_exec_changed_p;
78 /* Forward references. */
79 static int count_bb_insns (const_basic_block);
80 static bool cheap_bb_rtx_cost_p (const_basic_block, profile_probability, int);
81 static rtx_insn *first_active_insn (basic_block);
82 static rtx_insn *last_active_insn (basic_block, int);
83 static rtx_insn *find_active_insn_before (basic_block, rtx_insn *);
84 static rtx_insn *find_active_insn_after (basic_block, rtx_insn *);
85 static basic_block block_fallthru (basic_block);
86 static rtx cond_exec_get_condition (rtx_insn *, bool);
87 static rtx noce_get_condition (rtx_insn *, rtx_insn **, bool);
88 static int noce_operand_ok (const_rtx);
89 static void merge_if_block (ce_if_block *);
90 static int find_cond_trap (basic_block, edge, edge);
91 static basic_block find_if_header (basic_block, int);
92 static int block_jumps_and_fallthru_p (basic_block, basic_block);
93 static int noce_find_if_block (basic_block, edge, edge, int);
94 static int cond_exec_find_if_block (ce_if_block *);
95 static int find_if_case_1 (basic_block, edge, edge);
96 static int find_if_case_2 (basic_block, edge, edge);
97 static int dead_or_predicable (basic_block, basic_block, basic_block,
98 edge, int);
99 static void noce_emit_move_insn (rtx, rtx);
100 static rtx_insn *block_has_only_trap (basic_block);
101 static void need_cmov_or_rewire (basic_block, hash_set<rtx_insn *> *,
102 hash_map<rtx_insn *, int> *);
103 static bool noce_convert_multiple_sets_1 (struct noce_if_info *,
104 hash_set<rtx_insn *> *,
105 hash_map<rtx_insn *, int> *,
106 auto_vec<rtx> *,
107 auto_vec<rtx> *,
108 auto_vec<rtx_insn *> *, int *);
110 /* Count the number of non-jump active insns in BB. */
112 static int
113 count_bb_insns (const_basic_block bb)
115 int count = 0;
116 rtx_insn *insn = BB_HEAD (bb);
118 while (1)
120 if (active_insn_p (insn) && !JUMP_P (insn))
121 count++;
123 if (insn == BB_END (bb))
124 break;
125 insn = NEXT_INSN (insn);
128 return count;
131 /* Determine whether the total insn_cost on non-jump insns in
132 basic block BB is less than MAX_COST. This function returns
133 false if the cost of any instruction could not be estimated.
135 The cost of the non-jump insns in BB is scaled by REG_BR_PROB_BASE
136 as those insns are being speculated. MAX_COST is scaled with SCALE
137 plus a small fudge factor. */
139 static bool
140 cheap_bb_rtx_cost_p (const_basic_block bb,
141 profile_probability prob, int max_cost)
143 int count = 0;
144 rtx_insn *insn = BB_HEAD (bb);
145 bool speed = optimize_bb_for_speed_p (bb);
146 int scale = prob.initialized_p () ? prob.to_reg_br_prob_base ()
147 : REG_BR_PROB_BASE;
149 /* Set scale to REG_BR_PROB_BASE to void the identical scaling
150 applied to insn_cost when optimizing for size. Only do
151 this after combine because if-conversion might interfere with
152 passes before combine.
154 Use optimize_function_for_speed_p instead of the pre-defined
155 variable speed to make sure it is set to same value for all
156 basic blocks in one if-conversion transformation. */
157 if (!optimize_function_for_speed_p (cfun) && ifcvt_after_combine)
158 scale = REG_BR_PROB_BASE;
159 /* Our branch probability/scaling factors are just estimates and don't
160 account for cases where we can get speculation for free and other
161 secondary benefits. So we fudge the scale factor to make speculating
162 appear a little more profitable when optimizing for performance. */
163 else
164 scale += REG_BR_PROB_BASE / 8;
167 max_cost *= scale;
169 while (1)
171 if (NONJUMP_INSN_P (insn))
173 int cost = insn_cost (insn, speed) * REG_BR_PROB_BASE;
174 if (cost == 0)
175 return false;
177 /* If this instruction is the load or set of a "stack" register,
178 such as a floating point register on x87, then the cost of
179 speculatively executing this insn may need to include
180 the additional cost of popping its result off of the
181 register stack. Unfortunately, correctly recognizing and
182 accounting for this additional overhead is tricky, so for
183 now we simply prohibit such speculative execution. */
184 #ifdef STACK_REGS
186 rtx set = single_set (insn);
187 if (set && STACK_REG_P (SET_DEST (set)))
188 return false;
190 #endif
192 count += cost;
193 if (count >= max_cost)
194 return false;
196 else if (CALL_P (insn))
197 return false;
199 if (insn == BB_END (bb))
200 break;
201 insn = NEXT_INSN (insn);
204 return true;
207 /* Return the first non-jump active insn in the basic block. */
209 static rtx_insn *
210 first_active_insn (basic_block bb)
212 rtx_insn *insn = BB_HEAD (bb);
214 if (LABEL_P (insn))
216 if (insn == BB_END (bb))
217 return NULL;
218 insn = NEXT_INSN (insn);
221 while (NOTE_P (insn) || DEBUG_INSN_P (insn))
223 if (insn == BB_END (bb))
224 return NULL;
225 insn = NEXT_INSN (insn);
228 if (JUMP_P (insn))
229 return NULL;
231 return insn;
234 /* Return the last non-jump active (non-jump) insn in the basic block. */
236 static rtx_insn *
237 last_active_insn (basic_block bb, int skip_use_p)
239 rtx_insn *insn = BB_END (bb);
240 rtx_insn *head = BB_HEAD (bb);
242 while (NOTE_P (insn)
243 || JUMP_P (insn)
244 || DEBUG_INSN_P (insn)
245 || (skip_use_p
246 && NONJUMP_INSN_P (insn)
247 && GET_CODE (PATTERN (insn)) == USE))
249 if (insn == head)
250 return NULL;
251 insn = PREV_INSN (insn);
254 if (LABEL_P (insn))
255 return NULL;
257 return insn;
260 /* Return the active insn before INSN inside basic block CURR_BB. */
262 static rtx_insn *
263 find_active_insn_before (basic_block curr_bb, rtx_insn *insn)
265 if (!insn || insn == BB_HEAD (curr_bb))
266 return NULL;
268 while ((insn = PREV_INSN (insn)) != NULL_RTX)
270 if (NONJUMP_INSN_P (insn) || JUMP_P (insn) || CALL_P (insn))
271 break;
273 /* No other active insn all the way to the start of the basic block. */
274 if (insn == BB_HEAD (curr_bb))
275 return NULL;
278 return insn;
281 /* Return the active insn after INSN inside basic block CURR_BB. */
283 static rtx_insn *
284 find_active_insn_after (basic_block curr_bb, rtx_insn *insn)
286 if (!insn || insn == BB_END (curr_bb))
287 return NULL;
289 while ((insn = NEXT_INSN (insn)) != NULL_RTX)
291 if (NONJUMP_INSN_P (insn) || JUMP_P (insn) || CALL_P (insn))
292 break;
294 /* No other active insn all the way to the end of the basic block. */
295 if (insn == BB_END (curr_bb))
296 return NULL;
299 return insn;
302 /* Return the basic block reached by falling though the basic block BB. */
304 static basic_block
305 block_fallthru (basic_block bb)
307 edge e = find_fallthru_edge (bb->succs);
309 return (e) ? e->dest : NULL_BLOCK;
312 /* Return true if RTXs A and B can be safely interchanged. */
314 static bool
315 rtx_interchangeable_p (const_rtx a, const_rtx b)
317 if (!rtx_equal_p (a, b))
318 return false;
320 if (GET_CODE (a) != MEM)
321 return true;
323 /* A dead type-unsafe memory reference is legal, but a live type-unsafe memory
324 reference is not. Interchanging a dead type-unsafe memory reference with
325 a live type-safe one creates a live type-unsafe memory reference, in other
326 words, it makes the program illegal.
327 We check here conservatively whether the two memory references have equal
328 memory attributes. */
330 return mem_attrs_eq_p (get_mem_attrs (a), get_mem_attrs (b));
334 /* Go through a bunch of insns, converting them to conditional
335 execution format if possible. Return TRUE if all of the non-note
336 insns were processed. */
338 static int
339 cond_exec_process_insns (ce_if_block *ce_info ATTRIBUTE_UNUSED,
340 /* if block information */rtx_insn *start,
341 /* first insn to look at */rtx end,
342 /* last insn to look at */rtx test,
343 /* conditional execution test */profile_probability
344 prob_val,
345 /* probability of branch taken. */int mod_ok)
347 int must_be_last = FALSE;
348 rtx_insn *insn;
349 rtx xtest;
350 rtx pattern;
352 if (!start || !end)
353 return FALSE;
355 for (insn = start; ; insn = NEXT_INSN (insn))
357 /* dwarf2out can't cope with conditional prologues. */
358 if (NOTE_P (insn) && NOTE_KIND (insn) == NOTE_INSN_PROLOGUE_END)
359 return FALSE;
361 if (NOTE_P (insn) || DEBUG_INSN_P (insn))
362 goto insn_done;
364 gcc_assert (NONJUMP_INSN_P (insn) || CALL_P (insn));
366 /* dwarf2out can't cope with conditional unwind info. */
367 if (RTX_FRAME_RELATED_P (insn))
368 return FALSE;
370 /* Remove USE insns that get in the way. */
371 if (reload_completed && GET_CODE (PATTERN (insn)) == USE)
373 /* ??? Ug. Actually unlinking the thing is problematic,
374 given what we'd have to coordinate with our callers. */
375 SET_INSN_DELETED (insn);
376 goto insn_done;
379 /* Last insn wasn't last? */
380 if (must_be_last)
381 return FALSE;
383 if (modified_in_p (test, insn))
385 if (!mod_ok)
386 return FALSE;
387 must_be_last = TRUE;
390 /* Now build the conditional form of the instruction. */
391 pattern = PATTERN (insn);
392 xtest = copy_rtx (test);
394 /* If this is already a COND_EXEC, rewrite the test to be an AND of the
395 two conditions. */
396 if (GET_CODE (pattern) == COND_EXEC)
398 if (GET_MODE (xtest) != GET_MODE (COND_EXEC_TEST (pattern)))
399 return FALSE;
401 xtest = gen_rtx_AND (GET_MODE (xtest), xtest,
402 COND_EXEC_TEST (pattern));
403 pattern = COND_EXEC_CODE (pattern);
406 pattern = gen_rtx_COND_EXEC (VOIDmode, xtest, pattern);
408 /* If the machine needs to modify the insn being conditionally executed,
409 say for example to force a constant integer operand into a temp
410 register, do so here. */
411 #ifdef IFCVT_MODIFY_INSN
412 IFCVT_MODIFY_INSN (ce_info, pattern, insn);
413 if (! pattern)
414 return FALSE;
415 #endif
417 validate_change (insn, &PATTERN (insn), pattern, 1);
419 if (CALL_P (insn) && prob_val.initialized_p ())
420 validate_change (insn, &REG_NOTES (insn),
421 gen_rtx_INT_LIST ((machine_mode) REG_BR_PROB,
422 prob_val.to_reg_br_prob_note (),
423 REG_NOTES (insn)), 1);
425 insn_done:
426 if (insn == end)
427 break;
430 return TRUE;
433 /* Return the condition for a jump. Do not do any special processing. */
435 static rtx
436 cond_exec_get_condition (rtx_insn *jump, bool get_reversed = false)
438 rtx test_if, cond;
440 if (any_condjump_p (jump))
441 test_if = SET_SRC (pc_set (jump));
442 else
443 return NULL_RTX;
444 cond = XEXP (test_if, 0);
446 /* If this branches to JUMP_LABEL when the condition is false,
447 reverse the condition. */
448 if (get_reversed
449 || (GET_CODE (XEXP (test_if, 2)) == LABEL_REF
450 && label_ref_label (XEXP (test_if, 2))
451 == JUMP_LABEL (jump)))
453 enum rtx_code rev = reversed_comparison_code (cond, jump);
454 if (rev == UNKNOWN)
455 return NULL_RTX;
457 cond = gen_rtx_fmt_ee (rev, GET_MODE (cond), XEXP (cond, 0),
458 XEXP (cond, 1));
461 return cond;
464 /* Given a simple IF-THEN or IF-THEN-ELSE block, attempt to convert it
465 to conditional execution. Return TRUE if we were successful at
466 converting the block. */
468 static int
469 cond_exec_process_if_block (ce_if_block * ce_info,
470 /* if block information */int do_multiple_p)
472 basic_block test_bb = ce_info->test_bb; /* last test block */
473 basic_block then_bb = ce_info->then_bb; /* THEN */
474 basic_block else_bb = ce_info->else_bb; /* ELSE or NULL */
475 rtx test_expr; /* expression in IF_THEN_ELSE that is tested */
476 rtx_insn *then_start; /* first insn in THEN block */
477 rtx_insn *then_end; /* last insn + 1 in THEN block */
478 rtx_insn *else_start = NULL; /* first insn in ELSE block or NULL */
479 rtx_insn *else_end = NULL; /* last insn + 1 in ELSE block */
480 int max; /* max # of insns to convert. */
481 int then_mod_ok; /* whether conditional mods are ok in THEN */
482 rtx true_expr; /* test for else block insns */
483 rtx false_expr; /* test for then block insns */
484 profile_probability true_prob_val;/* probability of else block */
485 profile_probability false_prob_val;/* probability of then block */
486 rtx_insn *then_last_head = NULL; /* Last match at the head of THEN */
487 rtx_insn *else_last_head = NULL; /* Last match at the head of ELSE */
488 rtx_insn *then_first_tail = NULL; /* First match at the tail of THEN */
489 rtx_insn *else_first_tail = NULL; /* First match at the tail of ELSE */
490 int then_n_insns, else_n_insns, n_insns;
491 enum rtx_code false_code;
492 rtx note;
494 /* If test is comprised of && or || elements, and we've failed at handling
495 all of them together, just use the last test if it is the special case of
496 && elements without an ELSE block. */
497 if (!do_multiple_p && ce_info->num_multiple_test_blocks)
499 if (else_bb || ! ce_info->and_and_p)
500 return FALSE;
502 ce_info->test_bb = test_bb = ce_info->last_test_bb;
503 ce_info->num_multiple_test_blocks = 0;
504 ce_info->num_and_and_blocks = 0;
505 ce_info->num_or_or_blocks = 0;
508 /* Find the conditional jump to the ELSE or JOIN part, and isolate
509 the test. */
510 test_expr = cond_exec_get_condition (BB_END (test_bb));
511 if (! test_expr)
512 return FALSE;
514 /* If the conditional jump is more than just a conditional jump,
515 then we cannot do conditional execution conversion on this block. */
516 if (! onlyjump_p (BB_END (test_bb)))
517 return FALSE;
519 /* Collect the bounds of where we're to search, skipping any labels, jumps
520 and notes at the beginning and end of the block. Then count the total
521 number of insns and see if it is small enough to convert. */
522 then_start = first_active_insn (then_bb);
523 then_end = last_active_insn (then_bb, TRUE);
524 then_n_insns = ce_info->num_then_insns = count_bb_insns (then_bb);
525 n_insns = then_n_insns;
526 max = MAX_CONDITIONAL_EXECUTE;
528 if (else_bb)
530 int n_matching;
532 max *= 2;
533 else_start = first_active_insn (else_bb);
534 else_end = last_active_insn (else_bb, TRUE);
535 else_n_insns = ce_info->num_else_insns = count_bb_insns (else_bb);
536 n_insns += else_n_insns;
538 /* Look for matching sequences at the head and tail of the two blocks,
539 and limit the range of insns to be converted if possible. */
540 n_matching = flow_find_cross_jump (then_bb, else_bb,
541 &then_first_tail, &else_first_tail,
542 NULL);
543 if (then_first_tail == BB_HEAD (then_bb))
544 then_start = then_end = NULL;
545 if (else_first_tail == BB_HEAD (else_bb))
546 else_start = else_end = NULL;
548 if (n_matching > 0)
550 if (then_end)
551 then_end = find_active_insn_before (then_bb, then_first_tail);
552 if (else_end)
553 else_end = find_active_insn_before (else_bb, else_first_tail);
554 n_insns -= 2 * n_matching;
557 if (then_start
558 && else_start
559 && then_n_insns > n_matching
560 && else_n_insns > n_matching)
562 int longest_match = MIN (then_n_insns - n_matching,
563 else_n_insns - n_matching);
564 n_matching
565 = flow_find_head_matching_sequence (then_bb, else_bb,
566 &then_last_head,
567 &else_last_head,
568 longest_match);
570 if (n_matching > 0)
572 rtx_insn *insn;
574 /* We won't pass the insns in the head sequence to
575 cond_exec_process_insns, so we need to test them here
576 to make sure that they don't clobber the condition. */
577 for (insn = BB_HEAD (then_bb);
578 insn != NEXT_INSN (then_last_head);
579 insn = NEXT_INSN (insn))
580 if (!LABEL_P (insn) && !NOTE_P (insn)
581 && !DEBUG_INSN_P (insn)
582 && modified_in_p (test_expr, insn))
583 return FALSE;
586 if (then_last_head == then_end)
587 then_start = then_end = NULL;
588 if (else_last_head == else_end)
589 else_start = else_end = NULL;
591 if (n_matching > 0)
593 if (then_start)
594 then_start = find_active_insn_after (then_bb, then_last_head);
595 if (else_start)
596 else_start = find_active_insn_after (else_bb, else_last_head);
597 n_insns -= 2 * n_matching;
602 if (n_insns > max)
603 return FALSE;
605 /* Map test_expr/test_jump into the appropriate MD tests to use on
606 the conditionally executed code. */
608 true_expr = test_expr;
610 false_code = reversed_comparison_code (true_expr, BB_END (test_bb));
611 if (false_code != UNKNOWN)
612 false_expr = gen_rtx_fmt_ee (false_code, GET_MODE (true_expr),
613 XEXP (true_expr, 0), XEXP (true_expr, 1));
614 else
615 false_expr = NULL_RTX;
617 #ifdef IFCVT_MODIFY_TESTS
618 /* If the machine description needs to modify the tests, such as setting a
619 conditional execution register from a comparison, it can do so here. */
620 IFCVT_MODIFY_TESTS (ce_info, true_expr, false_expr);
622 /* See if the conversion failed. */
623 if (!true_expr || !false_expr)
624 goto fail;
625 #endif
627 note = find_reg_note (BB_END (test_bb), REG_BR_PROB, NULL_RTX);
628 if (note)
630 true_prob_val = profile_probability::from_reg_br_prob_note (XINT (note, 0));
631 false_prob_val = true_prob_val.invert ();
633 else
635 true_prob_val = profile_probability::uninitialized ();
636 false_prob_val = profile_probability::uninitialized ();
639 /* If we have && or || tests, do them here. These tests are in the adjacent
640 blocks after the first block containing the test. */
641 if (ce_info->num_multiple_test_blocks > 0)
643 basic_block bb = test_bb;
644 basic_block last_test_bb = ce_info->last_test_bb;
646 if (! false_expr)
647 goto fail;
651 rtx_insn *start, *end;
652 rtx t, f;
653 enum rtx_code f_code;
655 bb = block_fallthru (bb);
656 start = first_active_insn (bb);
657 end = last_active_insn (bb, TRUE);
658 if (start
659 && ! cond_exec_process_insns (ce_info, start, end, false_expr,
660 false_prob_val, FALSE))
661 goto fail;
663 /* If the conditional jump is more than just a conditional jump, then
664 we cannot do conditional execution conversion on this block. */
665 if (! onlyjump_p (BB_END (bb)))
666 goto fail;
668 /* Find the conditional jump and isolate the test. */
669 t = cond_exec_get_condition (BB_END (bb));
670 if (! t)
671 goto fail;
673 f_code = reversed_comparison_code (t, BB_END (bb));
674 if (f_code == UNKNOWN)
675 goto fail;
677 f = gen_rtx_fmt_ee (f_code, GET_MODE (t), XEXP (t, 0), XEXP (t, 1));
678 if (ce_info->and_and_p)
680 t = gen_rtx_AND (GET_MODE (t), true_expr, t);
681 f = gen_rtx_IOR (GET_MODE (t), false_expr, f);
683 else
685 t = gen_rtx_IOR (GET_MODE (t), true_expr, t);
686 f = gen_rtx_AND (GET_MODE (t), false_expr, f);
689 /* If the machine description needs to modify the tests, such as
690 setting a conditional execution register from a comparison, it can
691 do so here. */
692 #ifdef IFCVT_MODIFY_MULTIPLE_TESTS
693 IFCVT_MODIFY_MULTIPLE_TESTS (ce_info, bb, t, f);
695 /* See if the conversion failed. */
696 if (!t || !f)
697 goto fail;
698 #endif
700 true_expr = t;
701 false_expr = f;
703 while (bb != last_test_bb);
706 /* For IF-THEN-ELSE blocks, we don't allow modifications of the test
707 on then THEN block. */
708 then_mod_ok = (else_bb == NULL_BLOCK);
710 /* Go through the THEN and ELSE blocks converting the insns if possible
711 to conditional execution. */
713 if (then_end
714 && (! false_expr
715 || ! cond_exec_process_insns (ce_info, then_start, then_end,
716 false_expr, false_prob_val,
717 then_mod_ok)))
718 goto fail;
720 if (else_bb && else_end
721 && ! cond_exec_process_insns (ce_info, else_start, else_end,
722 true_expr, true_prob_val, TRUE))
723 goto fail;
725 /* If we cannot apply the changes, fail. Do not go through the normal fail
726 processing, since apply_change_group will call cancel_changes. */
727 if (! apply_change_group ())
729 #ifdef IFCVT_MODIFY_CANCEL
730 /* Cancel any machine dependent changes. */
731 IFCVT_MODIFY_CANCEL (ce_info);
732 #endif
733 return FALSE;
736 #ifdef IFCVT_MODIFY_FINAL
737 /* Do any machine dependent final modifications. */
738 IFCVT_MODIFY_FINAL (ce_info);
739 #endif
741 /* Conversion succeeded. */
742 if (dump_file)
743 fprintf (dump_file, "%d insn%s converted to conditional execution.\n",
744 n_insns, (n_insns == 1) ? " was" : "s were");
746 /* Merge the blocks! If we had matching sequences, make sure to delete one
747 copy at the appropriate location first: delete the copy in the THEN branch
748 for a tail sequence so that the remaining one is executed last for both
749 branches, and delete the copy in the ELSE branch for a head sequence so
750 that the remaining one is executed first for both branches. */
751 if (then_first_tail)
753 rtx_insn *from = then_first_tail;
754 if (!INSN_P (from))
755 from = find_active_insn_after (then_bb, from);
756 delete_insn_chain (from, get_last_bb_insn (then_bb), false);
758 if (else_last_head)
759 delete_insn_chain (first_active_insn (else_bb), else_last_head, false);
761 merge_if_block (ce_info);
762 cond_exec_changed_p = TRUE;
763 return TRUE;
765 fail:
766 #ifdef IFCVT_MODIFY_CANCEL
767 /* Cancel any machine dependent changes. */
768 IFCVT_MODIFY_CANCEL (ce_info);
769 #endif
771 cancel_changes (0);
772 return FALSE;
775 static rtx noce_emit_store_flag (struct noce_if_info *, rtx, int, int);
776 static int noce_try_move (struct noce_if_info *);
777 static int noce_try_ifelse_collapse (struct noce_if_info *);
778 static int noce_try_store_flag (struct noce_if_info *);
779 static int noce_try_addcc (struct noce_if_info *);
780 static int noce_try_store_flag_constants (struct noce_if_info *);
781 static int noce_try_store_flag_mask (struct noce_if_info *);
782 static rtx noce_emit_cmove (struct noce_if_info *, rtx, enum rtx_code, rtx,
783 rtx, rtx, rtx, rtx = NULL, rtx = NULL);
784 static int noce_try_cmove (struct noce_if_info *);
785 static int noce_try_cmove_arith (struct noce_if_info *);
786 static rtx noce_get_alt_condition (struct noce_if_info *, rtx, rtx_insn **);
787 static int noce_try_minmax (struct noce_if_info *);
788 static int noce_try_abs (struct noce_if_info *);
789 static int noce_try_sign_mask (struct noce_if_info *);
791 /* Return the comparison code for reversed condition for IF_INFO,
792 or UNKNOWN if reversing the condition is not possible. */
794 static inline enum rtx_code
795 noce_reversed_cond_code (struct noce_if_info *if_info)
797 if (if_info->rev_cond)
798 return GET_CODE (if_info->rev_cond);
799 return reversed_comparison_code (if_info->cond, if_info->jump);
802 /* Return true if SEQ is a good candidate as a replacement for the
803 if-convertible sequence described in IF_INFO.
804 This is the default implementation that targets can override
805 through a target hook. */
807 bool
808 default_noce_conversion_profitable_p (rtx_insn *seq,
809 struct noce_if_info *if_info)
811 bool speed_p = if_info->speed_p;
813 /* Cost up the new sequence. */
814 unsigned int cost = seq_cost (seq, speed_p);
816 if (cost <= if_info->original_cost)
817 return true;
819 /* When compiling for size, we can make a reasonably accurately guess
820 at the size growth. When compiling for speed, use the maximum. */
821 return speed_p && cost <= if_info->max_seq_cost;
824 /* Helper function for noce_try_store_flag*. */
826 static rtx
827 noce_emit_store_flag (struct noce_if_info *if_info, rtx x, int reversep,
828 int normalize)
830 rtx cond = if_info->cond;
831 int cond_complex;
832 enum rtx_code code;
834 cond_complex = (! general_operand (XEXP (cond, 0), VOIDmode)
835 || ! general_operand (XEXP (cond, 1), VOIDmode));
837 /* If earliest == jump, or when the condition is complex, try to
838 build the store_flag insn directly. */
840 if (cond_complex)
842 rtx set = pc_set (if_info->jump);
843 cond = XEXP (SET_SRC (set), 0);
844 if (GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
845 && label_ref_label (XEXP (SET_SRC (set), 2)) == JUMP_LABEL (if_info->jump))
846 reversep = !reversep;
847 if (if_info->then_else_reversed)
848 reversep = !reversep;
850 else if (reversep
851 && if_info->rev_cond
852 && general_operand (XEXP (if_info->rev_cond, 0), VOIDmode)
853 && general_operand (XEXP (if_info->rev_cond, 1), VOIDmode))
855 cond = if_info->rev_cond;
856 reversep = false;
859 if (reversep)
860 code = reversed_comparison_code (cond, if_info->jump);
861 else
862 code = GET_CODE (cond);
864 if ((if_info->cond_earliest == if_info->jump || cond_complex)
865 && (normalize == 0 || STORE_FLAG_VALUE == normalize))
867 rtx src = gen_rtx_fmt_ee (code, GET_MODE (x), XEXP (cond, 0),
868 XEXP (cond, 1));
869 rtx set = gen_rtx_SET (x, src);
871 start_sequence ();
872 rtx_insn *insn = emit_insn (set);
874 if (recog_memoized (insn) >= 0)
876 rtx_insn *seq = get_insns ();
877 end_sequence ();
878 emit_insn (seq);
880 if_info->cond_earliest = if_info->jump;
882 return x;
885 end_sequence ();
888 /* Don't even try if the comparison operands or the mode of X are weird. */
889 if (cond_complex || !SCALAR_INT_MODE_P (GET_MODE (x)))
890 return NULL_RTX;
892 return emit_store_flag (x, code, XEXP (cond, 0),
893 XEXP (cond, 1), VOIDmode,
894 (code == LTU || code == LEU
895 || code == GEU || code == GTU), normalize);
898 /* Return true if X can be safely forced into a register by copy_to_mode_reg
899 / force_operand. */
901 static bool
902 noce_can_force_operand (rtx x)
904 if (general_operand (x, VOIDmode))
905 return true;
906 if (SUBREG_P (x))
908 if (!noce_can_force_operand (SUBREG_REG (x)))
909 return false;
910 return true;
912 if (ARITHMETIC_P (x))
914 if (!noce_can_force_operand (XEXP (x, 0))
915 || !noce_can_force_operand (XEXP (x, 1)))
916 return false;
917 switch (GET_CODE (x))
919 case MULT:
920 case DIV:
921 case MOD:
922 case UDIV:
923 case UMOD:
924 return true;
925 default:
926 return code_to_optab (GET_CODE (x));
929 if (UNARY_P (x))
931 if (!noce_can_force_operand (XEXP (x, 0)))
932 return false;
933 switch (GET_CODE (x))
935 case ZERO_EXTEND:
936 case SIGN_EXTEND:
937 case TRUNCATE:
938 case FLOAT_EXTEND:
939 case FLOAT_TRUNCATE:
940 case FIX:
941 case UNSIGNED_FIX:
942 case FLOAT:
943 case UNSIGNED_FLOAT:
944 return true;
945 default:
946 return code_to_optab (GET_CODE (x));
949 return false;
952 /* Emit instruction to move an rtx, possibly into STRICT_LOW_PART.
953 X is the destination/target and Y is the value to copy. */
955 static void
956 noce_emit_move_insn (rtx x, rtx y)
958 machine_mode outmode;
959 rtx outer, inner;
960 poly_int64 bitpos;
962 if (GET_CODE (x) != STRICT_LOW_PART)
964 rtx_insn *seq, *insn;
965 rtx target;
966 optab ot;
968 start_sequence ();
969 /* Check that the SET_SRC is reasonable before calling emit_move_insn,
970 otherwise construct a suitable SET pattern ourselves. */
971 insn = (OBJECT_P (y) || CONSTANT_P (y) || GET_CODE (y) == SUBREG)
972 ? emit_move_insn (x, y)
973 : emit_insn (gen_rtx_SET (x, y));
974 seq = get_insns ();
975 end_sequence ();
977 if (recog_memoized (insn) <= 0)
979 if (GET_CODE (x) == ZERO_EXTRACT)
981 rtx op = XEXP (x, 0);
982 unsigned HOST_WIDE_INT size = INTVAL (XEXP (x, 1));
983 unsigned HOST_WIDE_INT start = INTVAL (XEXP (x, 2));
985 /* store_bit_field expects START to be relative to
986 BYTES_BIG_ENDIAN and adjusts this value for machines with
987 BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN. In order to be able to
988 invoke store_bit_field again it is necessary to have the START
989 value from the first call. */
990 if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
992 if (MEM_P (op))
993 start = BITS_PER_UNIT - start - size;
994 else
996 gcc_assert (REG_P (op));
997 start = BITS_PER_WORD - start - size;
1001 gcc_assert (start < (MEM_P (op) ? BITS_PER_UNIT : BITS_PER_WORD));
1002 store_bit_field (op, size, start, 0, 0, GET_MODE (x), y, false,
1003 false);
1004 return;
1007 switch (GET_RTX_CLASS (GET_CODE (y)))
1009 case RTX_UNARY:
1010 ot = code_to_optab (GET_CODE (y));
1011 if (ot && noce_can_force_operand (XEXP (y, 0)))
1013 start_sequence ();
1014 target = expand_unop (GET_MODE (y), ot, XEXP (y, 0), x, 0);
1015 if (target != NULL_RTX)
1017 if (target != x)
1018 emit_move_insn (x, target);
1019 seq = get_insns ();
1021 end_sequence ();
1023 break;
1025 case RTX_BIN_ARITH:
1026 case RTX_COMM_ARITH:
1027 ot = code_to_optab (GET_CODE (y));
1028 if (ot
1029 && noce_can_force_operand (XEXP (y, 0))
1030 && noce_can_force_operand (XEXP (y, 1)))
1032 start_sequence ();
1033 target = expand_binop (GET_MODE (y), ot,
1034 XEXP (y, 0), XEXP (y, 1),
1035 x, 0, OPTAB_DIRECT);
1036 if (target != NULL_RTX)
1038 if (target != x)
1039 emit_move_insn (x, target);
1040 seq = get_insns ();
1042 end_sequence ();
1044 break;
1046 default:
1047 break;
1051 emit_insn (seq);
1052 return;
1055 outer = XEXP (x, 0);
1056 inner = XEXP (outer, 0);
1057 outmode = GET_MODE (outer);
1058 bitpos = SUBREG_BYTE (outer) * BITS_PER_UNIT;
1059 store_bit_field (inner, GET_MODE_BITSIZE (outmode), bitpos,
1060 0, 0, outmode, y, false, false);
1063 /* Return the CC reg if it is used in COND. */
1065 static rtx
1066 cc_in_cond (rtx cond)
1068 if (have_cbranchcc4 && cond
1069 && GET_MODE_CLASS (GET_MODE (XEXP (cond, 0))) == MODE_CC)
1070 return XEXP (cond, 0);
1072 return NULL_RTX;
1075 /* Return sequence of instructions generated by if conversion. This
1076 function calls end_sequence() to end the current stream, ensures
1077 that the instructions are unshared, recognizable non-jump insns.
1078 On failure, this function returns a NULL_RTX. */
1080 static rtx_insn *
1081 end_ifcvt_sequence (struct noce_if_info *if_info)
1083 rtx_insn *insn;
1084 rtx_insn *seq = get_insns ();
1085 rtx cc = cc_in_cond (if_info->cond);
1087 set_used_flags (if_info->x);
1088 set_used_flags (if_info->cond);
1089 set_used_flags (if_info->a);
1090 set_used_flags (if_info->b);
1092 for (insn = seq; insn; insn = NEXT_INSN (insn))
1093 set_used_flags (insn);
1095 unshare_all_rtl_in_chain (seq);
1096 end_sequence ();
1098 /* Make sure that all of the instructions emitted are recognizable,
1099 and that we haven't introduced a new jump instruction.
1100 As an exercise for the reader, build a general mechanism that
1101 allows proper placement of required clobbers. */
1102 for (insn = seq; insn; insn = NEXT_INSN (insn))
1103 if (JUMP_P (insn)
1104 || recog_memoized (insn) == -1
1105 /* Make sure new generated code does not clobber CC. */
1106 || (cc && set_of (cc, insn)))
1107 return NULL;
1109 return seq;
1112 /* Return true iff the then and else basic block (if it exists)
1113 consist of a single simple set instruction. */
1115 static bool
1116 noce_simple_bbs (struct noce_if_info *if_info)
1118 if (!if_info->then_simple)
1119 return false;
1121 if (if_info->else_bb)
1122 return if_info->else_simple;
1124 return true;
1127 /* Convert "if (a != b) x = a; else x = b" into "x = a" and
1128 "if (a == b) x = a; else x = b" into "x = b". */
1130 static int
1131 noce_try_move (struct noce_if_info *if_info)
1133 rtx cond = if_info->cond;
1134 enum rtx_code code = GET_CODE (cond);
1135 rtx y;
1136 rtx_insn *seq;
1138 if (code != NE && code != EQ)
1139 return FALSE;
1141 if (!noce_simple_bbs (if_info))
1142 return FALSE;
1144 /* This optimization isn't valid if either A or B could be a NaN
1145 or a signed zero. */
1146 if (HONOR_NANS (if_info->x)
1147 || HONOR_SIGNED_ZEROS (if_info->x))
1148 return FALSE;
1150 /* Check whether the operands of the comparison are A and in
1151 either order. */
1152 if ((rtx_equal_p (if_info->a, XEXP (cond, 0))
1153 && rtx_equal_p (if_info->b, XEXP (cond, 1)))
1154 || (rtx_equal_p (if_info->a, XEXP (cond, 1))
1155 && rtx_equal_p (if_info->b, XEXP (cond, 0))))
1157 if (!rtx_interchangeable_p (if_info->a, if_info->b))
1158 return FALSE;
1160 y = (code == EQ) ? if_info->a : if_info->b;
1162 /* Avoid generating the move if the source is the destination. */
1163 if (! rtx_equal_p (if_info->x, y))
1165 start_sequence ();
1166 noce_emit_move_insn (if_info->x, y);
1167 seq = end_ifcvt_sequence (if_info);
1168 if (!seq)
1169 return FALSE;
1171 emit_insn_before_setloc (seq, if_info->jump,
1172 INSN_LOCATION (if_info->insn_a));
1174 if_info->transform_name = "noce_try_move";
1175 return TRUE;
1177 return FALSE;
1180 /* Try forming an IF_THEN_ELSE (cond, b, a) and collapsing that
1181 through simplify_rtx. Sometimes that can eliminate the IF_THEN_ELSE.
1182 If that is the case, emit the result into x. */
1184 static int
1185 noce_try_ifelse_collapse (struct noce_if_info * if_info)
1187 if (!noce_simple_bbs (if_info))
1188 return FALSE;
1190 machine_mode mode = GET_MODE (if_info->x);
1191 rtx if_then_else = simplify_gen_ternary (IF_THEN_ELSE, mode, mode,
1192 if_info->cond, if_info->b,
1193 if_info->a);
1195 if (GET_CODE (if_then_else) == IF_THEN_ELSE)
1196 return FALSE;
1198 rtx_insn *seq;
1199 start_sequence ();
1200 noce_emit_move_insn (if_info->x, if_then_else);
1201 seq = end_ifcvt_sequence (if_info);
1202 if (!seq)
1203 return FALSE;
1205 emit_insn_before_setloc (seq, if_info->jump,
1206 INSN_LOCATION (if_info->insn_a));
1208 if_info->transform_name = "noce_try_ifelse_collapse";
1209 return TRUE;
1213 /* Convert "if (test) x = 1; else x = 0".
1215 Only try 0 and STORE_FLAG_VALUE here. Other combinations will be
1216 tried in noce_try_store_flag_constants after noce_try_cmove has had
1217 a go at the conversion. */
1219 static int
1220 noce_try_store_flag (struct noce_if_info *if_info)
1222 int reversep;
1223 rtx target;
1224 rtx_insn *seq;
1226 if (!noce_simple_bbs (if_info))
1227 return FALSE;
1229 if (CONST_INT_P (if_info->b)
1230 && INTVAL (if_info->b) == STORE_FLAG_VALUE
1231 && if_info->a == const0_rtx)
1232 reversep = 0;
1233 else if (if_info->b == const0_rtx
1234 && CONST_INT_P (if_info->a)
1235 && INTVAL (if_info->a) == STORE_FLAG_VALUE
1236 && noce_reversed_cond_code (if_info) != UNKNOWN)
1237 reversep = 1;
1238 else
1239 return FALSE;
1241 start_sequence ();
1243 target = noce_emit_store_flag (if_info, if_info->x, reversep, 0);
1244 if (target)
1246 if (target != if_info->x)
1247 noce_emit_move_insn (if_info->x, target);
1249 seq = end_ifcvt_sequence (if_info);
1250 if (! seq)
1251 return FALSE;
1253 emit_insn_before_setloc (seq, if_info->jump,
1254 INSN_LOCATION (if_info->insn_a));
1255 if_info->transform_name = "noce_try_store_flag";
1256 return TRUE;
1258 else
1260 end_sequence ();
1261 return FALSE;
1266 /* Convert "if (test) x = -A; else x = A" into
1267 x = A; if (test) x = -x if the machine can do the
1268 conditional negate form of this cheaply.
1269 Try this before noce_try_cmove that will just load the
1270 immediates into two registers and do a conditional select
1271 between them. If the target has a conditional negate or
1272 conditional invert operation we can save a potentially
1273 expensive constant synthesis. */
1275 static bool
1276 noce_try_inverse_constants (struct noce_if_info *if_info)
1278 if (!noce_simple_bbs (if_info))
1279 return false;
1281 if (!CONST_INT_P (if_info->a)
1282 || !CONST_INT_P (if_info->b)
1283 || !REG_P (if_info->x))
1284 return false;
1286 machine_mode mode = GET_MODE (if_info->x);
1288 HOST_WIDE_INT val_a = INTVAL (if_info->a);
1289 HOST_WIDE_INT val_b = INTVAL (if_info->b);
1291 rtx cond = if_info->cond;
1293 rtx x = if_info->x;
1294 rtx target;
1296 start_sequence ();
1298 rtx_code code;
1299 if (val_b != HOST_WIDE_INT_MIN && val_a == -val_b)
1300 code = NEG;
1301 else if (val_a == ~val_b)
1302 code = NOT;
1303 else
1305 end_sequence ();
1306 return false;
1309 rtx tmp = gen_reg_rtx (mode);
1310 noce_emit_move_insn (tmp, if_info->a);
1312 target = emit_conditional_neg_or_complement (x, code, mode, cond, tmp, tmp);
1314 if (target)
1316 rtx_insn *seq = get_insns ();
1318 if (!seq)
1320 end_sequence ();
1321 return false;
1324 if (target != if_info->x)
1325 noce_emit_move_insn (if_info->x, target);
1327 seq = end_ifcvt_sequence (if_info);
1329 if (!seq)
1330 return false;
1332 emit_insn_before_setloc (seq, if_info->jump,
1333 INSN_LOCATION (if_info->insn_a));
1334 if_info->transform_name = "noce_try_inverse_constants";
1335 return true;
1338 end_sequence ();
1339 return false;
1343 /* Convert "if (test) x = a; else x = b", for A and B constant.
1344 Also allow A = y + c1, B = y + c2, with a common y between A
1345 and B. */
1347 static int
1348 noce_try_store_flag_constants (struct noce_if_info *if_info)
1350 rtx target;
1351 rtx_insn *seq;
1352 bool reversep;
1353 HOST_WIDE_INT itrue, ifalse, diff, tmp;
1354 int normalize;
1355 bool can_reverse;
1356 machine_mode mode = GET_MODE (if_info->x);
1357 rtx common = NULL_RTX;
1359 rtx a = if_info->a;
1360 rtx b = if_info->b;
1362 /* Handle cases like x := test ? y + 3 : y + 4. */
1363 if (GET_CODE (a) == PLUS
1364 && GET_CODE (b) == PLUS
1365 && CONST_INT_P (XEXP (a, 1))
1366 && CONST_INT_P (XEXP (b, 1))
1367 && rtx_equal_p (XEXP (a, 0), XEXP (b, 0))
1368 /* Allow expressions that are not using the result or plain
1369 registers where we handle overlap below. */
1370 && (REG_P (XEXP (a, 0))
1371 || (noce_operand_ok (XEXP (a, 0))
1372 && ! reg_overlap_mentioned_p (if_info->x, XEXP (a, 0)))))
1374 common = XEXP (a, 0);
1375 a = XEXP (a, 1);
1376 b = XEXP (b, 1);
1379 if (!noce_simple_bbs (if_info))
1380 return FALSE;
1382 if (CONST_INT_P (a)
1383 && CONST_INT_P (b))
1385 ifalse = INTVAL (a);
1386 itrue = INTVAL (b);
1387 bool subtract_flag_p = false;
1389 diff = (unsigned HOST_WIDE_INT) itrue - ifalse;
1390 /* Make sure we can represent the difference between the two values. */
1391 if ((diff > 0)
1392 != ((ifalse < 0) != (itrue < 0) ? ifalse < 0 : ifalse < itrue))
1393 return FALSE;
1395 diff = trunc_int_for_mode (diff, mode);
1397 can_reverse = noce_reversed_cond_code (if_info) != UNKNOWN;
1398 reversep = false;
1399 if (diff == STORE_FLAG_VALUE || diff == -STORE_FLAG_VALUE)
1401 normalize = 0;
1402 /* We could collapse these cases but it is easier to follow the
1403 diff/STORE_FLAG_VALUE combinations when they are listed
1404 explicitly. */
1406 /* test ? 3 : 4
1407 => 4 + (test != 0). */
1408 if (diff < 0 && STORE_FLAG_VALUE < 0)
1409 reversep = false;
1410 /* test ? 4 : 3
1411 => can_reverse | 4 + (test == 0)
1412 !can_reverse | 3 - (test != 0). */
1413 else if (diff > 0 && STORE_FLAG_VALUE < 0)
1415 reversep = can_reverse;
1416 subtract_flag_p = !can_reverse;
1417 /* If we need to subtract the flag and we have PLUS-immediate
1418 A and B then it is unlikely to be beneficial to play tricks
1419 here. */
1420 if (subtract_flag_p && common)
1421 return FALSE;
1423 /* test ? 3 : 4
1424 => can_reverse | 3 + (test == 0)
1425 !can_reverse | 4 - (test != 0). */
1426 else if (diff < 0 && STORE_FLAG_VALUE > 0)
1428 reversep = can_reverse;
1429 subtract_flag_p = !can_reverse;
1430 /* If we need to subtract the flag and we have PLUS-immediate
1431 A and B then it is unlikely to be beneficial to play tricks
1432 here. */
1433 if (subtract_flag_p && common)
1434 return FALSE;
1436 /* test ? 4 : 3
1437 => 4 + (test != 0). */
1438 else if (diff > 0 && STORE_FLAG_VALUE > 0)
1439 reversep = false;
1440 else
1441 gcc_unreachable ();
1443 /* Is this (cond) ? 2^n : 0? */
1444 else if (ifalse == 0 && pow2p_hwi (itrue)
1445 && STORE_FLAG_VALUE == 1)
1446 normalize = 1;
1447 /* Is this (cond) ? 0 : 2^n? */
1448 else if (itrue == 0 && pow2p_hwi (ifalse) && can_reverse
1449 && STORE_FLAG_VALUE == 1)
1451 normalize = 1;
1452 reversep = true;
1454 /* Is this (cond) ? -1 : x? */
1455 else if (itrue == -1
1456 && STORE_FLAG_VALUE == -1)
1457 normalize = -1;
1458 /* Is this (cond) ? x : -1? */
1459 else if (ifalse == -1 && can_reverse
1460 && STORE_FLAG_VALUE == -1)
1462 normalize = -1;
1463 reversep = true;
1465 else
1466 return FALSE;
1468 if (reversep)
1470 std::swap (itrue, ifalse);
1471 diff = trunc_int_for_mode (-(unsigned HOST_WIDE_INT) diff, mode);
1474 start_sequence ();
1476 /* If we have x := test ? x + 3 : x + 4 then move the original
1477 x out of the way while we store flags. */
1478 if (common && rtx_equal_p (common, if_info->x))
1480 common = gen_reg_rtx (mode);
1481 noce_emit_move_insn (common, if_info->x);
1484 target = noce_emit_store_flag (if_info, if_info->x, reversep, normalize);
1485 if (! target)
1487 end_sequence ();
1488 return FALSE;
1491 /* if (test) x = 3; else x = 4;
1492 => x = 3 + (test == 0); */
1493 if (diff == STORE_FLAG_VALUE || diff == -STORE_FLAG_VALUE)
1495 /* Add the common part now. This may allow combine to merge this
1496 with the store flag operation earlier into some sort of conditional
1497 increment/decrement if the target allows it. */
1498 if (common)
1499 target = expand_simple_binop (mode, PLUS,
1500 target, common,
1501 target, 0, OPTAB_WIDEN);
1503 /* Always use ifalse here. It should have been swapped with itrue
1504 when appropriate when reversep is true. */
1505 target = expand_simple_binop (mode, subtract_flag_p ? MINUS : PLUS,
1506 gen_int_mode (ifalse, mode), target,
1507 if_info->x, 0, OPTAB_WIDEN);
1509 /* Other cases are not beneficial when the original A and B are PLUS
1510 expressions. */
1511 else if (common)
1513 end_sequence ();
1514 return FALSE;
1516 /* if (test) x = 8; else x = 0;
1517 => x = (test != 0) << 3; */
1518 else if (ifalse == 0 && (tmp = exact_log2 (itrue)) >= 0)
1520 target = expand_simple_binop (mode, ASHIFT,
1521 target, GEN_INT (tmp), if_info->x, 0,
1522 OPTAB_WIDEN);
1525 /* if (test) x = -1; else x = b;
1526 => x = -(test != 0) | b; */
1527 else if (itrue == -1)
1529 target = expand_simple_binop (mode, IOR,
1530 target, gen_int_mode (ifalse, mode),
1531 if_info->x, 0, OPTAB_WIDEN);
1533 else
1535 end_sequence ();
1536 return FALSE;
1539 if (! target)
1541 end_sequence ();
1542 return FALSE;
1545 if (target != if_info->x)
1546 noce_emit_move_insn (if_info->x, target);
1548 seq = end_ifcvt_sequence (if_info);
1549 if (!seq || !targetm.noce_conversion_profitable_p (seq, if_info))
1550 return FALSE;
1552 emit_insn_before_setloc (seq, if_info->jump,
1553 INSN_LOCATION (if_info->insn_a));
1554 if_info->transform_name = "noce_try_store_flag_constants";
1556 return TRUE;
1559 return FALSE;
1562 /* Convert "if (test) foo++" into "foo += (test != 0)", and
1563 similarly for "foo--". */
1565 static int
1566 noce_try_addcc (struct noce_if_info *if_info)
1568 rtx target;
1569 rtx_insn *seq;
1570 int subtract, normalize;
1572 if (!noce_simple_bbs (if_info))
1573 return FALSE;
1575 if (GET_CODE (if_info->a) == PLUS
1576 && rtx_equal_p (XEXP (if_info->a, 0), if_info->b)
1577 && noce_reversed_cond_code (if_info) != UNKNOWN)
1579 rtx cond = if_info->rev_cond;
1580 enum rtx_code code;
1582 if (cond == NULL_RTX)
1584 cond = if_info->cond;
1585 code = reversed_comparison_code (cond, if_info->jump);
1587 else
1588 code = GET_CODE (cond);
1590 /* First try to use addcc pattern. */
1591 if (general_operand (XEXP (cond, 0), VOIDmode)
1592 && general_operand (XEXP (cond, 1), VOIDmode))
1594 start_sequence ();
1595 target = emit_conditional_add (if_info->x, code,
1596 XEXP (cond, 0),
1597 XEXP (cond, 1),
1598 VOIDmode,
1599 if_info->b,
1600 XEXP (if_info->a, 1),
1601 GET_MODE (if_info->x),
1602 (code == LTU || code == GEU
1603 || code == LEU || code == GTU));
1604 if (target)
1606 if (target != if_info->x)
1607 noce_emit_move_insn (if_info->x, target);
1609 seq = end_ifcvt_sequence (if_info);
1610 if (!seq || !targetm.noce_conversion_profitable_p (seq, if_info))
1611 return FALSE;
1613 emit_insn_before_setloc (seq, if_info->jump,
1614 INSN_LOCATION (if_info->insn_a));
1615 if_info->transform_name = "noce_try_addcc";
1617 return TRUE;
1619 end_sequence ();
1622 /* If that fails, construct conditional increment or decrement using
1623 setcc. We're changing a branch and an increment to a comparison and
1624 an ADD/SUB. */
1625 if (XEXP (if_info->a, 1) == const1_rtx
1626 || XEXP (if_info->a, 1) == constm1_rtx)
1628 start_sequence ();
1629 if (STORE_FLAG_VALUE == INTVAL (XEXP (if_info->a, 1)))
1630 subtract = 0, normalize = 0;
1631 else if (-STORE_FLAG_VALUE == INTVAL (XEXP (if_info->a, 1)))
1632 subtract = 1, normalize = 0;
1633 else
1634 subtract = 0, normalize = INTVAL (XEXP (if_info->a, 1));
1637 target = noce_emit_store_flag (if_info,
1638 gen_reg_rtx (GET_MODE (if_info->x)),
1639 1, normalize);
1641 if (target)
1642 target = expand_simple_binop (GET_MODE (if_info->x),
1643 subtract ? MINUS : PLUS,
1644 if_info->b, target, if_info->x,
1645 0, OPTAB_WIDEN);
1646 if (target)
1648 if (target != if_info->x)
1649 noce_emit_move_insn (if_info->x, target);
1651 seq = end_ifcvt_sequence (if_info);
1652 if (!seq || !targetm.noce_conversion_profitable_p (seq, if_info))
1653 return FALSE;
1655 emit_insn_before_setloc (seq, if_info->jump,
1656 INSN_LOCATION (if_info->insn_a));
1657 if_info->transform_name = "noce_try_addcc";
1658 return TRUE;
1660 end_sequence ();
1664 return FALSE;
1667 /* Convert "if (test) x = 0;" to "x &= -(test == 0);" */
1669 static int
1670 noce_try_store_flag_mask (struct noce_if_info *if_info)
1672 rtx target;
1673 rtx_insn *seq;
1674 int reversep;
1676 if (!noce_simple_bbs (if_info))
1677 return FALSE;
1679 reversep = 0;
1681 if ((if_info->a == const0_rtx
1682 && (REG_P (if_info->b) || rtx_equal_p (if_info->b, if_info->x)))
1683 || ((reversep = (noce_reversed_cond_code (if_info) != UNKNOWN))
1684 && if_info->b == const0_rtx
1685 && (REG_P (if_info->a) || rtx_equal_p (if_info->a, if_info->x))))
1687 start_sequence ();
1688 target = noce_emit_store_flag (if_info,
1689 gen_reg_rtx (GET_MODE (if_info->x)),
1690 reversep, -1);
1691 if (target)
1692 target = expand_simple_binop (GET_MODE (if_info->x), AND,
1693 reversep ? if_info->a : if_info->b,
1694 target, if_info->x, 0,
1695 OPTAB_WIDEN);
1697 if (target)
1699 if (target != if_info->x)
1700 noce_emit_move_insn (if_info->x, target);
1702 seq = end_ifcvt_sequence (if_info);
1703 if (!seq || !targetm.noce_conversion_profitable_p (seq, if_info))
1704 return FALSE;
1706 emit_insn_before_setloc (seq, if_info->jump,
1707 INSN_LOCATION (if_info->insn_a));
1708 if_info->transform_name = "noce_try_store_flag_mask";
1710 return TRUE;
1713 end_sequence ();
1716 return FALSE;
1719 /* Helper function for noce_try_cmove and noce_try_cmove_arith. */
1721 static rtx
1722 noce_emit_cmove (struct noce_if_info *if_info, rtx x, enum rtx_code code,
1723 rtx cmp_a, rtx cmp_b, rtx vfalse, rtx vtrue, rtx cc_cmp,
1724 rtx rev_cc_cmp)
1726 rtx target ATTRIBUTE_UNUSED;
1727 int unsignedp ATTRIBUTE_UNUSED;
1729 /* If earliest == jump, try to build the cmove insn directly.
1730 This is helpful when combine has created some complex condition
1731 (like for alpha's cmovlbs) that we can't hope to regenerate
1732 through the normal interface. */
1734 if (if_info->cond_earliest == if_info->jump)
1736 rtx cond = gen_rtx_fmt_ee (code, GET_MODE (if_info->cond), cmp_a, cmp_b);
1737 rtx if_then_else = gen_rtx_IF_THEN_ELSE (GET_MODE (x),
1738 cond, vtrue, vfalse);
1739 rtx set = gen_rtx_SET (x, if_then_else);
1741 start_sequence ();
1742 rtx_insn *insn = emit_insn (set);
1744 if (recog_memoized (insn) >= 0)
1746 rtx_insn *seq = get_insns ();
1747 end_sequence ();
1748 emit_insn (seq);
1750 return x;
1753 end_sequence ();
1756 unsignedp = (code == LTU || code == GEU
1757 || code == LEU || code == GTU);
1759 if (cc_cmp != NULL_RTX && rev_cc_cmp != NULL_RTX)
1760 target = emit_conditional_move (x, cc_cmp, rev_cc_cmp,
1761 vtrue, vfalse, GET_MODE (x));
1762 else
1764 /* Don't even try if the comparison operands are weird
1765 except that the target supports cbranchcc4. */
1766 if (! general_operand (cmp_a, GET_MODE (cmp_a))
1767 || ! general_operand (cmp_b, GET_MODE (cmp_b)))
1769 if (!have_cbranchcc4
1770 || GET_MODE_CLASS (GET_MODE (cmp_a)) != MODE_CC
1771 || cmp_b != const0_rtx)
1772 return NULL_RTX;
1775 target = emit_conditional_move (x, { code, cmp_a, cmp_b, VOIDmode },
1776 vtrue, vfalse, GET_MODE (x),
1777 unsignedp);
1780 if (target)
1781 return target;
1783 /* We might be faced with a situation like:
1785 x = (reg:M TARGET)
1786 vtrue = (subreg:M (reg:N VTRUE) BYTE)
1787 vfalse = (subreg:M (reg:N VFALSE) BYTE)
1789 We can't do a conditional move in mode M, but it's possible that we
1790 could do a conditional move in mode N instead and take a subreg of
1791 the result.
1793 If we can't create new pseudos, though, don't bother. */
1794 if (reload_completed)
1795 return NULL_RTX;
1797 if (GET_CODE (vtrue) == SUBREG && GET_CODE (vfalse) == SUBREG)
1799 rtx reg_vtrue = SUBREG_REG (vtrue);
1800 rtx reg_vfalse = SUBREG_REG (vfalse);
1801 poly_uint64 byte_vtrue = SUBREG_BYTE (vtrue);
1802 poly_uint64 byte_vfalse = SUBREG_BYTE (vfalse);
1803 rtx promoted_target;
1805 if (GET_MODE (reg_vtrue) != GET_MODE (reg_vfalse)
1806 || maybe_ne (byte_vtrue, byte_vfalse)
1807 || (SUBREG_PROMOTED_VAR_P (vtrue)
1808 != SUBREG_PROMOTED_VAR_P (vfalse))
1809 || (SUBREG_PROMOTED_GET (vtrue)
1810 != SUBREG_PROMOTED_GET (vfalse)))
1811 return NULL_RTX;
1813 promoted_target = gen_reg_rtx (GET_MODE (reg_vtrue));
1815 target = emit_conditional_move (promoted_target,
1816 { code, cmp_a, cmp_b, VOIDmode },
1817 reg_vtrue, reg_vfalse,
1818 GET_MODE (reg_vtrue), unsignedp);
1819 /* Nope, couldn't do it in that mode either. */
1820 if (!target)
1821 return NULL_RTX;
1823 target = gen_rtx_SUBREG (GET_MODE (vtrue), promoted_target, byte_vtrue);
1824 SUBREG_PROMOTED_VAR_P (target) = SUBREG_PROMOTED_VAR_P (vtrue);
1825 SUBREG_PROMOTED_SET (target, SUBREG_PROMOTED_GET (vtrue));
1826 emit_move_insn (x, target);
1827 return x;
1829 else
1830 return NULL_RTX;
1833 /* Try only simple constants and registers here. More complex cases
1834 are handled in noce_try_cmove_arith after noce_try_store_flag_arith
1835 has had a go at it. */
1837 static int
1838 noce_try_cmove (struct noce_if_info *if_info)
1840 enum rtx_code code;
1841 rtx target;
1842 rtx_insn *seq;
1844 if (!noce_simple_bbs (if_info))
1845 return FALSE;
1847 if ((CONSTANT_P (if_info->a) || register_operand (if_info->a, VOIDmode))
1848 && (CONSTANT_P (if_info->b) || register_operand (if_info->b, VOIDmode)))
1850 start_sequence ();
1852 code = GET_CODE (if_info->cond);
1853 target = noce_emit_cmove (if_info, if_info->x, code,
1854 XEXP (if_info->cond, 0),
1855 XEXP (if_info->cond, 1),
1856 if_info->a, if_info->b);
1858 if (target)
1860 if (target != if_info->x)
1861 noce_emit_move_insn (if_info->x, target);
1863 seq = end_ifcvt_sequence (if_info);
1864 if (!seq || !targetm.noce_conversion_profitable_p (seq, if_info))
1865 return FALSE;
1867 emit_insn_before_setloc (seq, if_info->jump,
1868 INSN_LOCATION (if_info->insn_a));
1869 if_info->transform_name = "noce_try_cmove";
1871 return TRUE;
1873 /* If both a and b are constants try a last-ditch transformation:
1874 if (test) x = a; else x = b;
1875 => x = (-(test != 0) & (b - a)) + a;
1876 Try this only if the target-specific expansion above has failed.
1877 The target-specific expander may want to generate sequences that
1878 we don't know about, so give them a chance before trying this
1879 approach. */
1880 else if (!targetm.have_conditional_execution ()
1881 && CONST_INT_P (if_info->a) && CONST_INT_P (if_info->b))
1883 machine_mode mode = GET_MODE (if_info->x);
1884 HOST_WIDE_INT ifalse = INTVAL (if_info->a);
1885 HOST_WIDE_INT itrue = INTVAL (if_info->b);
1886 rtx target = noce_emit_store_flag (if_info, if_info->x, false, -1);
1887 if (!target)
1889 end_sequence ();
1890 return FALSE;
1893 HOST_WIDE_INT diff = (unsigned HOST_WIDE_INT) itrue - ifalse;
1894 /* Make sure we can represent the difference
1895 between the two values. */
1896 if ((diff > 0)
1897 != ((ifalse < 0) != (itrue < 0) ? ifalse < 0 : ifalse < itrue))
1899 end_sequence ();
1900 return FALSE;
1903 diff = trunc_int_for_mode (diff, mode);
1904 target = expand_simple_binop (mode, AND,
1905 target, gen_int_mode (diff, mode),
1906 if_info->x, 0, OPTAB_WIDEN);
1907 if (target)
1908 target = expand_simple_binop (mode, PLUS,
1909 target, gen_int_mode (ifalse, mode),
1910 if_info->x, 0, OPTAB_WIDEN);
1911 if (target)
1913 if (target != if_info->x)
1914 noce_emit_move_insn (if_info->x, target);
1916 seq = end_ifcvt_sequence (if_info);
1917 if (!seq || !targetm.noce_conversion_profitable_p (seq, if_info))
1918 return FALSE;
1920 emit_insn_before_setloc (seq, if_info->jump,
1921 INSN_LOCATION (if_info->insn_a));
1922 if_info->transform_name = "noce_try_cmove";
1923 return TRUE;
1925 else
1927 end_sequence ();
1928 return FALSE;
1931 else
1932 end_sequence ();
1935 return FALSE;
1938 /* Return true if X contains a conditional code mode rtx. */
1940 static bool
1941 contains_ccmode_rtx_p (rtx x)
1943 subrtx_iterator::array_type array;
1944 FOR_EACH_SUBRTX (iter, array, x, ALL)
1945 if (GET_MODE_CLASS (GET_MODE (*iter)) == MODE_CC)
1946 return true;
1948 return false;
1951 /* Helper for bb_valid_for_noce_process_p. Validate that
1952 the rtx insn INSN is a single set that does not set
1953 the conditional register CC and is in general valid for
1954 if-conversion. */
1956 static bool
1957 insn_valid_noce_process_p (rtx_insn *insn, rtx cc)
1959 if (!insn
1960 || !NONJUMP_INSN_P (insn)
1961 || (cc && set_of (cc, insn)))
1962 return false;
1964 rtx sset = single_set (insn);
1966 /* Currently support only simple single sets in test_bb. */
1967 if (!sset
1968 || !noce_operand_ok (SET_DEST (sset))
1969 || contains_ccmode_rtx_p (SET_DEST (sset))
1970 || !noce_operand_ok (SET_SRC (sset)))
1971 return false;
1973 return true;
1977 /* Return true iff the registers that the insns in BB_A set do not get
1978 used in BB_B. If TO_RENAME is non-NULL then it is a location that will be
1979 renamed later by the caller and so conflicts on it should be ignored
1980 in this function. */
1982 static bool
1983 bbs_ok_for_cmove_arith (basic_block bb_a, basic_block bb_b, rtx to_rename)
1985 rtx_insn *a_insn;
1986 bitmap bba_sets = BITMAP_ALLOC (&reg_obstack);
1988 df_ref def;
1989 df_ref use;
1991 FOR_BB_INSNS (bb_a, a_insn)
1993 if (!active_insn_p (a_insn))
1994 continue;
1996 rtx sset_a = single_set (a_insn);
1998 if (!sset_a)
2000 BITMAP_FREE (bba_sets);
2001 return false;
2003 /* Record all registers that BB_A sets. */
2004 FOR_EACH_INSN_DEF (def, a_insn)
2005 if (!(to_rename && DF_REF_REG (def) == to_rename))
2006 bitmap_set_bit (bba_sets, DF_REF_REGNO (def));
2009 rtx_insn *b_insn;
2011 FOR_BB_INSNS (bb_b, b_insn)
2013 if (!active_insn_p (b_insn))
2014 continue;
2016 rtx sset_b = single_set (b_insn);
2018 if (!sset_b)
2020 BITMAP_FREE (bba_sets);
2021 return false;
2024 /* Make sure this is a REG and not some instance
2025 of ZERO_EXTRACT or SUBREG or other dangerous stuff.
2026 If we have a memory destination then we have a pair of simple
2027 basic blocks performing an operation of the form [addr] = c ? a : b.
2028 bb_valid_for_noce_process_p will have ensured that these are
2029 the only stores present. In that case [addr] should be the location
2030 to be renamed. Assert that the callers set this up properly. */
2031 if (MEM_P (SET_DEST (sset_b)))
2032 gcc_assert (rtx_equal_p (SET_DEST (sset_b), to_rename));
2033 else if (!REG_P (SET_DEST (sset_b)))
2035 BITMAP_FREE (bba_sets);
2036 return false;
2039 /* If the insn uses a reg set in BB_A return false. */
2040 FOR_EACH_INSN_USE (use, b_insn)
2042 if (bitmap_bit_p (bba_sets, DF_REF_REGNO (use)))
2044 BITMAP_FREE (bba_sets);
2045 return false;
2051 BITMAP_FREE (bba_sets);
2052 return true;
2055 /* Emit copies of all the active instructions in BB except the last.
2056 This is a helper for noce_try_cmove_arith. */
2058 static void
2059 noce_emit_all_but_last (basic_block bb)
2061 rtx_insn *last = last_active_insn (bb, FALSE);
2062 rtx_insn *insn;
2063 FOR_BB_INSNS (bb, insn)
2065 if (insn != last && active_insn_p (insn))
2067 rtx_insn *to_emit = as_a <rtx_insn *> (copy_rtx (insn));
2069 emit_insn (PATTERN (to_emit));
2074 /* Helper for noce_try_cmove_arith. Emit the pattern TO_EMIT and return
2075 the resulting insn or NULL if it's not a valid insn. */
2077 static rtx_insn *
2078 noce_emit_insn (rtx to_emit)
2080 gcc_assert (to_emit);
2081 rtx_insn *insn = emit_insn (to_emit);
2083 if (recog_memoized (insn) < 0)
2084 return NULL;
2086 return insn;
2089 /* Helper for noce_try_cmove_arith. Emit a copy of the insns up to
2090 and including the penultimate one in BB if it is not simple
2091 (as indicated by SIMPLE). Then emit LAST_INSN as the last
2092 insn in the block. The reason for that is that LAST_INSN may
2093 have been modified by the preparation in noce_try_cmove_arith. */
2095 static bool
2096 noce_emit_bb (rtx last_insn, basic_block bb, bool simple)
2098 if (bb && !simple)
2099 noce_emit_all_but_last (bb);
2101 if (last_insn && !noce_emit_insn (last_insn))
2102 return false;
2104 return true;
2107 /* Try more complex cases involving conditional_move. */
2109 static int
2110 noce_try_cmove_arith (struct noce_if_info *if_info)
2112 rtx a = if_info->a;
2113 rtx b = if_info->b;
2114 rtx x = if_info->x;
2115 rtx orig_a, orig_b;
2116 rtx_insn *insn_a, *insn_b;
2117 bool a_simple = if_info->then_simple;
2118 bool b_simple = if_info->else_simple;
2119 basic_block then_bb = if_info->then_bb;
2120 basic_block else_bb = if_info->else_bb;
2121 rtx target;
2122 int is_mem = 0;
2123 enum rtx_code code;
2124 rtx cond = if_info->cond;
2125 rtx_insn *ifcvt_seq;
2127 /* A conditional move from two memory sources is equivalent to a
2128 conditional on their addresses followed by a load. Don't do this
2129 early because it'll screw alias analysis. Note that we've
2130 already checked for no side effects. */
2131 if (cse_not_expected
2132 && MEM_P (a) && MEM_P (b)
2133 && MEM_ADDR_SPACE (a) == MEM_ADDR_SPACE (b))
2135 machine_mode address_mode = get_address_mode (a);
2137 a = XEXP (a, 0);
2138 b = XEXP (b, 0);
2139 x = gen_reg_rtx (address_mode);
2140 is_mem = 1;
2143 /* ??? We could handle this if we knew that a load from A or B could
2144 not trap or fault. This is also true if we've already loaded
2145 from the address along the path from ENTRY. */
2146 else if (may_trap_or_fault_p (a) || may_trap_or_fault_p (b))
2147 return FALSE;
2149 /* if (test) x = a + b; else x = c - d;
2150 => y = a + b;
2151 x = c - d;
2152 if (test)
2153 x = y;
2156 code = GET_CODE (cond);
2157 insn_a = if_info->insn_a;
2158 insn_b = if_info->insn_b;
2160 machine_mode x_mode = GET_MODE (x);
2162 if (!can_conditionally_move_p (x_mode))
2163 return FALSE;
2165 /* Possibly rearrange operands to make things come out more natural. */
2166 if (noce_reversed_cond_code (if_info) != UNKNOWN)
2168 int reversep = 0;
2169 if (rtx_equal_p (b, x))
2170 reversep = 1;
2171 else if (general_operand (b, GET_MODE (b)))
2172 reversep = 1;
2174 if (reversep)
2176 if (if_info->rev_cond)
2178 cond = if_info->rev_cond;
2179 code = GET_CODE (cond);
2181 else
2182 code = reversed_comparison_code (cond, if_info->jump);
2183 std::swap (a, b);
2184 std::swap (insn_a, insn_b);
2185 std::swap (a_simple, b_simple);
2186 std::swap (then_bb, else_bb);
2190 if (then_bb && else_bb
2191 && (!bbs_ok_for_cmove_arith (then_bb, else_bb, if_info->orig_x)
2192 || !bbs_ok_for_cmove_arith (else_bb, then_bb, if_info->orig_x)))
2193 return FALSE;
2195 start_sequence ();
2197 /* If one of the blocks is empty then the corresponding B or A value
2198 came from the test block. The non-empty complex block that we will
2199 emit might clobber the register used by B or A, so move it to a pseudo
2200 first. */
2202 rtx tmp_a = NULL_RTX;
2203 rtx tmp_b = NULL_RTX;
2205 if (b_simple || !else_bb)
2206 tmp_b = gen_reg_rtx (x_mode);
2208 if (a_simple || !then_bb)
2209 tmp_a = gen_reg_rtx (x_mode);
2211 orig_a = a;
2212 orig_b = b;
2214 rtx emit_a = NULL_RTX;
2215 rtx emit_b = NULL_RTX;
2216 rtx_insn *tmp_insn = NULL;
2217 bool modified_in_a = false;
2218 bool modified_in_b = false;
2219 /* If either operand is complex, load it into a register first.
2220 The best way to do this is to copy the original insn. In this
2221 way we preserve any clobbers etc that the insn may have had.
2222 This is of course not possible in the IS_MEM case. */
2224 if (! general_operand (a, GET_MODE (a)) || tmp_a)
2227 if (is_mem)
2229 rtx reg = gen_reg_rtx (GET_MODE (a));
2230 emit_a = gen_rtx_SET (reg, a);
2232 else
2234 if (insn_a)
2236 a = tmp_a ? tmp_a : gen_reg_rtx (GET_MODE (a));
2238 rtx_insn *copy_of_a = as_a <rtx_insn *> (copy_rtx (insn_a));
2239 rtx set = single_set (copy_of_a);
2240 SET_DEST (set) = a;
2242 emit_a = PATTERN (copy_of_a);
2244 else
2246 rtx tmp_reg = tmp_a ? tmp_a : gen_reg_rtx (GET_MODE (a));
2247 emit_a = gen_rtx_SET (tmp_reg, a);
2248 a = tmp_reg;
2253 if (! general_operand (b, GET_MODE (b)) || tmp_b)
2255 if (is_mem)
2257 rtx reg = gen_reg_rtx (GET_MODE (b));
2258 emit_b = gen_rtx_SET (reg, b);
2260 else
2262 if (insn_b)
2264 b = tmp_b ? tmp_b : gen_reg_rtx (GET_MODE (b));
2265 rtx_insn *copy_of_b = as_a <rtx_insn *> (copy_rtx (insn_b));
2266 rtx set = single_set (copy_of_b);
2268 SET_DEST (set) = b;
2269 emit_b = PATTERN (copy_of_b);
2271 else
2273 rtx tmp_reg = tmp_b ? tmp_b : gen_reg_rtx (GET_MODE (b));
2274 emit_b = gen_rtx_SET (tmp_reg, b);
2275 b = tmp_reg;
2280 modified_in_a = emit_a != NULL_RTX && modified_in_p (orig_b, emit_a);
2281 if (tmp_b && then_bb)
2283 FOR_BB_INSNS (then_bb, tmp_insn)
2284 /* Don't check inside insn_a. We will have changed it to emit_a
2285 with a destination that doesn't conflict. */
2286 if (!(insn_a && tmp_insn == insn_a)
2287 && modified_in_p (orig_b, tmp_insn))
2289 modified_in_a = true;
2290 break;
2295 modified_in_b = emit_b != NULL_RTX && modified_in_p (orig_a, emit_b);
2296 if (tmp_a && else_bb)
2298 FOR_BB_INSNS (else_bb, tmp_insn)
2299 /* Don't check inside insn_b. We will have changed it to emit_b
2300 with a destination that doesn't conflict. */
2301 if (!(insn_b && tmp_insn == insn_b)
2302 && modified_in_p (orig_a, tmp_insn))
2304 modified_in_b = true;
2305 break;
2309 /* If insn to set up A clobbers any registers B depends on, try to
2310 swap insn that sets up A with the one that sets up B. If even
2311 that doesn't help, punt. */
2312 if (modified_in_a && !modified_in_b)
2314 if (!noce_emit_bb (emit_b, else_bb, b_simple))
2315 goto end_seq_and_fail;
2317 if (!noce_emit_bb (emit_a, then_bb, a_simple))
2318 goto end_seq_and_fail;
2320 else if (!modified_in_a)
2322 if (!noce_emit_bb (emit_a, then_bb, a_simple))
2323 goto end_seq_and_fail;
2325 if (!noce_emit_bb (emit_b, else_bb, b_simple))
2326 goto end_seq_and_fail;
2328 else
2329 goto end_seq_and_fail;
2331 target = noce_emit_cmove (if_info, x, code, XEXP (cond, 0), XEXP (cond, 1),
2332 a, b);
2334 if (! target)
2335 goto end_seq_and_fail;
2337 /* If we're handling a memory for above, emit the load now. */
2338 if (is_mem)
2340 rtx mem = gen_rtx_MEM (GET_MODE (if_info->x), target);
2342 /* Copy over flags as appropriate. */
2343 if (MEM_VOLATILE_P (if_info->a) || MEM_VOLATILE_P (if_info->b))
2344 MEM_VOLATILE_P (mem) = 1;
2345 if (MEM_ALIAS_SET (if_info->a) == MEM_ALIAS_SET (if_info->b))
2346 set_mem_alias_set (mem, MEM_ALIAS_SET (if_info->a));
2347 set_mem_align (mem,
2348 MIN (MEM_ALIGN (if_info->a), MEM_ALIGN (if_info->b)));
2350 gcc_assert (MEM_ADDR_SPACE (if_info->a) == MEM_ADDR_SPACE (if_info->b));
2351 set_mem_addr_space (mem, MEM_ADDR_SPACE (if_info->a));
2353 noce_emit_move_insn (if_info->x, mem);
2355 else if (target != x)
2356 noce_emit_move_insn (x, target);
2358 ifcvt_seq = end_ifcvt_sequence (if_info);
2359 if (!ifcvt_seq || !targetm.noce_conversion_profitable_p (ifcvt_seq, if_info))
2360 return FALSE;
2362 emit_insn_before_setloc (ifcvt_seq, if_info->jump,
2363 INSN_LOCATION (if_info->insn_a));
2364 if_info->transform_name = "noce_try_cmove_arith";
2365 return TRUE;
2367 end_seq_and_fail:
2368 end_sequence ();
2369 return FALSE;
2372 /* For most cases, the simplified condition we found is the best
2373 choice, but this is not the case for the min/max/abs transforms.
2374 For these we wish to know that it is A or B in the condition. */
2376 static rtx
2377 noce_get_alt_condition (struct noce_if_info *if_info, rtx target,
2378 rtx_insn **earliest)
2380 rtx cond, set;
2381 rtx_insn *insn;
2382 int reverse;
2384 /* If target is already mentioned in the known condition, return it. */
2385 if (reg_mentioned_p (target, if_info->cond))
2387 *earliest = if_info->cond_earliest;
2388 return if_info->cond;
2391 set = pc_set (if_info->jump);
2392 cond = XEXP (SET_SRC (set), 0);
2393 reverse
2394 = GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
2395 && label_ref_label (XEXP (SET_SRC (set), 2)) == JUMP_LABEL (if_info->jump);
2396 if (if_info->then_else_reversed)
2397 reverse = !reverse;
2399 /* If we're looking for a constant, try to make the conditional
2400 have that constant in it. There are two reasons why it may
2401 not have the constant we want:
2403 1. GCC may have needed to put the constant in a register, because
2404 the target can't compare directly against that constant. For
2405 this case, we look for a SET immediately before the comparison
2406 that puts a constant in that register.
2408 2. GCC may have canonicalized the conditional, for example
2409 replacing "if x < 4" with "if x <= 3". We can undo that (or
2410 make equivalent types of changes) to get the constants we need
2411 if they're off by one in the right direction. */
2413 if (CONST_INT_P (target))
2415 enum rtx_code code = GET_CODE (if_info->cond);
2416 rtx op_a = XEXP (if_info->cond, 0);
2417 rtx op_b = XEXP (if_info->cond, 1);
2418 rtx_insn *prev_insn;
2420 /* First, look to see if we put a constant in a register. */
2421 prev_insn = prev_nonnote_nondebug_insn (if_info->cond_earliest);
2422 if (prev_insn
2423 && BLOCK_FOR_INSN (prev_insn)
2424 == BLOCK_FOR_INSN (if_info->cond_earliest)
2425 && INSN_P (prev_insn)
2426 && GET_CODE (PATTERN (prev_insn)) == SET)
2428 rtx src = find_reg_equal_equiv_note (prev_insn);
2429 if (!src)
2430 src = SET_SRC (PATTERN (prev_insn));
2431 if (CONST_INT_P (src))
2433 if (rtx_equal_p (op_a, SET_DEST (PATTERN (prev_insn))))
2434 op_a = src;
2435 else if (rtx_equal_p (op_b, SET_DEST (PATTERN (prev_insn))))
2436 op_b = src;
2438 if (CONST_INT_P (op_a))
2440 std::swap (op_a, op_b);
2441 code = swap_condition (code);
2446 /* Now, look to see if we can get the right constant by
2447 adjusting the conditional. */
2448 if (CONST_INT_P (op_b))
2450 HOST_WIDE_INT desired_val = INTVAL (target);
2451 HOST_WIDE_INT actual_val = INTVAL (op_b);
2453 switch (code)
2455 case LT:
2456 if (desired_val != HOST_WIDE_INT_MAX
2457 && actual_val == desired_val + 1)
2459 code = LE;
2460 op_b = GEN_INT (desired_val);
2462 break;
2463 case LE:
2464 if (desired_val != HOST_WIDE_INT_MIN
2465 && actual_val == desired_val - 1)
2467 code = LT;
2468 op_b = GEN_INT (desired_val);
2470 break;
2471 case GT:
2472 if (desired_val != HOST_WIDE_INT_MIN
2473 && actual_val == desired_val - 1)
2475 code = GE;
2476 op_b = GEN_INT (desired_val);
2478 break;
2479 case GE:
2480 if (desired_val != HOST_WIDE_INT_MAX
2481 && actual_val == desired_val + 1)
2483 code = GT;
2484 op_b = GEN_INT (desired_val);
2486 break;
2487 default:
2488 break;
2492 /* If we made any changes, generate a new conditional that is
2493 equivalent to what we started with, but has the right
2494 constants in it. */
2495 if (code != GET_CODE (if_info->cond)
2496 || op_a != XEXP (if_info->cond, 0)
2497 || op_b != XEXP (if_info->cond, 1))
2499 cond = gen_rtx_fmt_ee (code, GET_MODE (cond), op_a, op_b);
2500 *earliest = if_info->cond_earliest;
2501 return cond;
2505 cond = canonicalize_condition (if_info->jump, cond, reverse,
2506 earliest, target, have_cbranchcc4, true);
2507 if (! cond || ! reg_mentioned_p (target, cond))
2508 return NULL;
2510 /* We almost certainly searched back to a different place.
2511 Need to re-verify correct lifetimes. */
2513 /* X may not be mentioned in the range (cond_earliest, jump]. */
2514 for (insn = if_info->jump; insn != *earliest; insn = PREV_INSN (insn))
2515 if (INSN_P (insn) && reg_overlap_mentioned_p (if_info->x, PATTERN (insn)))
2516 return NULL;
2518 /* A and B may not be modified in the range [cond_earliest, jump). */
2519 for (insn = *earliest; insn != if_info->jump; insn = NEXT_INSN (insn))
2520 if (INSN_P (insn)
2521 && (modified_in_p (if_info->a, insn)
2522 || modified_in_p (if_info->b, insn)))
2523 return NULL;
2525 return cond;
2528 /* Convert "if (a < b) x = a; else x = b;" to "x = min(a, b);", etc. */
2530 static int
2531 noce_try_minmax (struct noce_if_info *if_info)
2533 rtx cond, target;
2534 rtx_insn *earliest, *seq;
2535 enum rtx_code code, op;
2536 int unsignedp;
2538 if (!noce_simple_bbs (if_info))
2539 return FALSE;
2541 /* ??? Reject modes with NaNs or signed zeros since we don't know how
2542 they will be resolved with an SMIN/SMAX. It wouldn't be too hard
2543 to get the target to tell us... */
2544 if (HONOR_SIGNED_ZEROS (if_info->x)
2545 || HONOR_NANS (if_info->x))
2546 return FALSE;
2548 cond = noce_get_alt_condition (if_info, if_info->a, &earliest);
2549 if (!cond)
2550 return FALSE;
2552 /* Verify the condition is of the form we expect, and canonicalize
2553 the comparison code. */
2554 code = GET_CODE (cond);
2555 if (rtx_equal_p (XEXP (cond, 0), if_info->a))
2557 if (! rtx_equal_p (XEXP (cond, 1), if_info->b))
2558 return FALSE;
2560 else if (rtx_equal_p (XEXP (cond, 1), if_info->a))
2562 if (! rtx_equal_p (XEXP (cond, 0), if_info->b))
2563 return FALSE;
2564 code = swap_condition (code);
2566 else
2567 return FALSE;
2569 /* Determine what sort of operation this is. Note that the code is for
2570 a taken branch, so the code->operation mapping appears backwards. */
2571 switch (code)
2573 case LT:
2574 case LE:
2575 case UNLT:
2576 case UNLE:
2577 op = SMAX;
2578 unsignedp = 0;
2579 break;
2580 case GT:
2581 case GE:
2582 case UNGT:
2583 case UNGE:
2584 op = SMIN;
2585 unsignedp = 0;
2586 break;
2587 case LTU:
2588 case LEU:
2589 op = UMAX;
2590 unsignedp = 1;
2591 break;
2592 case GTU:
2593 case GEU:
2594 op = UMIN;
2595 unsignedp = 1;
2596 break;
2597 default:
2598 return FALSE;
2601 start_sequence ();
2603 target = expand_simple_binop (GET_MODE (if_info->x), op,
2604 if_info->a, if_info->b,
2605 if_info->x, unsignedp, OPTAB_WIDEN);
2606 if (! target)
2608 end_sequence ();
2609 return FALSE;
2611 if (target != if_info->x)
2612 noce_emit_move_insn (if_info->x, target);
2614 seq = end_ifcvt_sequence (if_info);
2615 if (!seq)
2616 return FALSE;
2618 emit_insn_before_setloc (seq, if_info->jump, INSN_LOCATION (if_info->insn_a));
2619 if_info->cond = cond;
2620 if_info->cond_earliest = earliest;
2621 if_info->rev_cond = NULL_RTX;
2622 if_info->transform_name = "noce_try_minmax";
2624 return TRUE;
2627 /* Convert "if (a < 0) x = -a; else x = a;" to "x = abs(a);",
2628 "if (a < 0) x = ~a; else x = a;" to "x = one_cmpl_abs(a);",
2629 etc. */
2631 static int
2632 noce_try_abs (struct noce_if_info *if_info)
2634 rtx cond, target, a, b, c;
2635 rtx_insn *earliest, *seq;
2636 int negate;
2637 bool one_cmpl = false;
2639 if (!noce_simple_bbs (if_info))
2640 return FALSE;
2642 /* Reject modes with signed zeros. */
2643 if (HONOR_SIGNED_ZEROS (if_info->x))
2644 return FALSE;
2646 /* Recognize A and B as constituting an ABS or NABS. The canonical
2647 form is a branch around the negation, taken when the object is the
2648 first operand of a comparison against 0 that evaluates to true. */
2649 a = if_info->a;
2650 b = if_info->b;
2651 if (GET_CODE (a) == NEG && rtx_equal_p (XEXP (a, 0), b))
2652 negate = 0;
2653 else if (GET_CODE (b) == NEG && rtx_equal_p (XEXP (b, 0), a))
2655 std::swap (a, b);
2656 negate = 1;
2658 else if (GET_CODE (a) == NOT && rtx_equal_p (XEXP (a, 0), b))
2660 negate = 0;
2661 one_cmpl = true;
2663 else if (GET_CODE (b) == NOT && rtx_equal_p (XEXP (b, 0), a))
2665 std::swap (a, b);
2666 negate = 1;
2667 one_cmpl = true;
2669 else
2670 return FALSE;
2672 cond = noce_get_alt_condition (if_info, b, &earliest);
2673 if (!cond)
2674 return FALSE;
2676 /* Verify the condition is of the form we expect. */
2677 if (rtx_equal_p (XEXP (cond, 0), b))
2678 c = XEXP (cond, 1);
2679 else if (rtx_equal_p (XEXP (cond, 1), b))
2681 c = XEXP (cond, 0);
2682 negate = !negate;
2684 else
2685 return FALSE;
2687 /* Verify that C is zero. Search one step backward for a
2688 REG_EQUAL note or a simple source if necessary. */
2689 if (REG_P (c))
2691 rtx set;
2692 rtx_insn *insn = prev_nonnote_nondebug_insn (earliest);
2693 if (insn
2694 && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (earliest)
2695 && (set = single_set (insn))
2696 && rtx_equal_p (SET_DEST (set), c))
2698 rtx note = find_reg_equal_equiv_note (insn);
2699 if (note)
2700 c = XEXP (note, 0);
2701 else
2702 c = SET_SRC (set);
2704 else
2705 return FALSE;
2707 if (MEM_P (c)
2708 && GET_CODE (XEXP (c, 0)) == SYMBOL_REF
2709 && CONSTANT_POOL_ADDRESS_P (XEXP (c, 0)))
2710 c = get_pool_constant (XEXP (c, 0));
2712 /* Work around funny ideas get_condition has wrt canonicalization.
2713 Note that these rtx constants are known to be CONST_INT, and
2714 therefore imply integer comparisons.
2715 The one_cmpl case is more complicated, as we want to handle
2716 only x < 0 ? ~x : x or x >= 0 ? x : ~x to one_cmpl_abs (x)
2717 and x < 0 ? x : ~x or x >= 0 ? ~x : x to ~one_cmpl_abs (x),
2718 but not other cases (x > -1 is equivalent of x >= 0). */
2719 if (c == constm1_rtx && GET_CODE (cond) == GT)
2721 else if (c == const1_rtx && GET_CODE (cond) == LT)
2723 if (one_cmpl)
2724 return FALSE;
2726 else if (c == CONST0_RTX (GET_MODE (b)))
2728 if (one_cmpl
2729 && GET_CODE (cond) != GE
2730 && GET_CODE (cond) != LT)
2731 return FALSE;
2733 else
2734 return FALSE;
2736 /* Determine what sort of operation this is. */
2737 switch (GET_CODE (cond))
2739 case LT:
2740 case LE:
2741 case UNLT:
2742 case UNLE:
2743 negate = !negate;
2744 break;
2745 case GT:
2746 case GE:
2747 case UNGT:
2748 case UNGE:
2749 break;
2750 default:
2751 return FALSE;
2754 start_sequence ();
2755 if (one_cmpl)
2756 target = expand_one_cmpl_abs_nojump (GET_MODE (if_info->x), b,
2757 if_info->x);
2758 else
2759 target = expand_abs_nojump (GET_MODE (if_info->x), b, if_info->x, 1);
2761 /* ??? It's a quandary whether cmove would be better here, especially
2762 for integers. Perhaps combine will clean things up. */
2763 if (target && negate)
2765 if (one_cmpl)
2766 target = expand_simple_unop (GET_MODE (target), NOT, target,
2767 if_info->x, 0);
2768 else
2769 target = expand_simple_unop (GET_MODE (target), NEG, target,
2770 if_info->x, 0);
2773 if (! target)
2775 end_sequence ();
2776 return FALSE;
2779 if (target != if_info->x)
2780 noce_emit_move_insn (if_info->x, target);
2782 seq = end_ifcvt_sequence (if_info);
2783 if (!seq)
2784 return FALSE;
2786 emit_insn_before_setloc (seq, if_info->jump, INSN_LOCATION (if_info->insn_a));
2787 if_info->cond = cond;
2788 if_info->cond_earliest = earliest;
2789 if_info->rev_cond = NULL_RTX;
2790 if_info->transform_name = "noce_try_abs";
2792 return TRUE;
2795 /* Convert "if (m < 0) x = b; else x = 0;" to "x = (m >> C) & b;". */
2797 static int
2798 noce_try_sign_mask (struct noce_if_info *if_info)
2800 rtx cond, t, m, c;
2801 rtx_insn *seq;
2802 machine_mode mode;
2803 enum rtx_code code;
2804 bool t_unconditional;
2806 if (!noce_simple_bbs (if_info))
2807 return FALSE;
2809 cond = if_info->cond;
2810 code = GET_CODE (cond);
2811 m = XEXP (cond, 0);
2812 c = XEXP (cond, 1);
2814 t = NULL_RTX;
2815 if (if_info->a == const0_rtx)
2817 if ((code == LT && c == const0_rtx)
2818 || (code == LE && c == constm1_rtx))
2819 t = if_info->b;
2821 else if (if_info->b == const0_rtx)
2823 if ((code == GE && c == const0_rtx)
2824 || (code == GT && c == constm1_rtx))
2825 t = if_info->a;
2828 if (! t || side_effects_p (t))
2829 return FALSE;
2831 /* We currently don't handle different modes. */
2832 mode = GET_MODE (t);
2833 if (GET_MODE (m) != mode)
2834 return FALSE;
2836 /* This is only profitable if T is unconditionally executed/evaluated in the
2837 original insn sequence or T is cheap and can't trap or fault. The former
2838 happens if B is the non-zero (T) value and if INSN_B was taken from
2839 TEST_BB, or there was no INSN_B which can happen for e.g. conditional
2840 stores to memory. For the cost computation use the block TEST_BB where
2841 the evaluation will end up after the transformation. */
2842 t_unconditional
2843 = (t == if_info->b
2844 && (if_info->insn_b == NULL_RTX
2845 || BLOCK_FOR_INSN (if_info->insn_b) == if_info->test_bb));
2846 if (!(t_unconditional
2847 || ((set_src_cost (t, mode, if_info->speed_p)
2848 < COSTS_N_INSNS (2))
2849 && !may_trap_or_fault_p (t))))
2850 return FALSE;
2852 if (!noce_can_force_operand (t))
2853 return FALSE;
2855 start_sequence ();
2856 /* Use emit_store_flag to generate "m < 0 ? -1 : 0" instead of expanding
2857 "(signed) m >> 31" directly. This benefits targets with specialized
2858 insns to obtain the signmask, but still uses ashr_optab otherwise. */
2859 m = emit_store_flag (gen_reg_rtx (mode), LT, m, const0_rtx, mode, 0, -1);
2860 t = m ? expand_binop (mode, and_optab, m, t, NULL_RTX, 0, OPTAB_DIRECT)
2861 : NULL_RTX;
2863 if (!t)
2865 end_sequence ();
2866 return FALSE;
2869 noce_emit_move_insn (if_info->x, t);
2871 seq = end_ifcvt_sequence (if_info);
2872 if (!seq)
2873 return FALSE;
2875 emit_insn_before_setloc (seq, if_info->jump, INSN_LOCATION (if_info->insn_a));
2876 if_info->transform_name = "noce_try_sign_mask";
2878 return TRUE;
2882 /* Optimize away "if (x & C) x |= C" and similar bit manipulation
2883 transformations. */
2885 static int
2886 noce_try_bitop (struct noce_if_info *if_info)
2888 rtx cond, x, a, result;
2889 rtx_insn *seq;
2890 scalar_int_mode mode;
2891 enum rtx_code code;
2892 int bitnum;
2894 x = if_info->x;
2895 cond = if_info->cond;
2896 code = GET_CODE (cond);
2898 /* Check for an integer operation. */
2899 if (!is_a <scalar_int_mode> (GET_MODE (x), &mode))
2900 return FALSE;
2902 if (!noce_simple_bbs (if_info))
2903 return FALSE;
2905 /* Check for no else condition. */
2906 if (! rtx_equal_p (x, if_info->b))
2907 return FALSE;
2909 /* Check for a suitable condition. */
2910 if (code != NE && code != EQ)
2911 return FALSE;
2912 if (XEXP (cond, 1) != const0_rtx)
2913 return FALSE;
2914 cond = XEXP (cond, 0);
2916 /* ??? We could also handle AND here. */
2917 if (GET_CODE (cond) == ZERO_EXTRACT)
2919 if (XEXP (cond, 1) != const1_rtx
2920 || !CONST_INT_P (XEXP (cond, 2))
2921 || ! rtx_equal_p (x, XEXP (cond, 0)))
2922 return FALSE;
2923 bitnum = INTVAL (XEXP (cond, 2));
2924 if (BITS_BIG_ENDIAN)
2925 bitnum = GET_MODE_BITSIZE (mode) - 1 - bitnum;
2926 if (bitnum < 0 || bitnum >= HOST_BITS_PER_WIDE_INT)
2927 return FALSE;
2929 else
2930 return FALSE;
2932 a = if_info->a;
2933 if (GET_CODE (a) == IOR || GET_CODE (a) == XOR)
2935 /* Check for "if (X & C) x = x op C". */
2936 if (! rtx_equal_p (x, XEXP (a, 0))
2937 || !CONST_INT_P (XEXP (a, 1))
2938 || (INTVAL (XEXP (a, 1)) & GET_MODE_MASK (mode))
2939 != HOST_WIDE_INT_1U << bitnum)
2940 return FALSE;
2942 /* if ((x & C) == 0) x |= C; is transformed to x |= C. */
2943 /* if ((x & C) != 0) x |= C; is transformed to nothing. */
2944 if (GET_CODE (a) == IOR)
2945 result = (code == NE) ? a : NULL_RTX;
2946 else if (code == NE)
2948 /* if ((x & C) == 0) x ^= C; is transformed to x |= C. */
2949 result = gen_int_mode (HOST_WIDE_INT_1 << bitnum, mode);
2950 result = simplify_gen_binary (IOR, mode, x, result);
2952 else
2954 /* if ((x & C) != 0) x ^= C; is transformed to x &= ~C. */
2955 result = gen_int_mode (~(HOST_WIDE_INT_1 << bitnum), mode);
2956 result = simplify_gen_binary (AND, mode, x, result);
2959 else if (GET_CODE (a) == AND)
2961 /* Check for "if (X & C) x &= ~C". */
2962 if (! rtx_equal_p (x, XEXP (a, 0))
2963 || !CONST_INT_P (XEXP (a, 1))
2964 || (INTVAL (XEXP (a, 1)) & GET_MODE_MASK (mode))
2965 != (~(HOST_WIDE_INT_1 << bitnum) & GET_MODE_MASK (mode)))
2966 return FALSE;
2968 /* if ((x & C) == 0) x &= ~C; is transformed to nothing. */
2969 /* if ((x & C) != 0) x &= ~C; is transformed to x &= ~C. */
2970 result = (code == EQ) ? a : NULL_RTX;
2972 else
2973 return FALSE;
2975 if (result)
2977 start_sequence ();
2978 noce_emit_move_insn (x, result);
2979 seq = end_ifcvt_sequence (if_info);
2980 if (!seq)
2981 return FALSE;
2983 emit_insn_before_setloc (seq, if_info->jump,
2984 INSN_LOCATION (if_info->insn_a));
2986 if_info->transform_name = "noce_try_bitop";
2987 return TRUE;
2991 /* Similar to get_condition, only the resulting condition must be
2992 valid at JUMP, instead of at EARLIEST.
2994 If THEN_ELSE_REVERSED is true, the fallthrough does not go to the
2995 THEN block of the caller, and we have to reverse the condition. */
2997 static rtx
2998 noce_get_condition (rtx_insn *jump, rtx_insn **earliest, bool then_else_reversed)
3000 rtx cond, set, tmp;
3001 bool reverse;
3003 if (! any_condjump_p (jump))
3004 return NULL_RTX;
3006 set = pc_set (jump);
3008 /* If this branches to JUMP_LABEL when the condition is false,
3009 reverse the condition. */
3010 reverse = (GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
3011 && label_ref_label (XEXP (SET_SRC (set), 2)) == JUMP_LABEL (jump));
3013 /* We may have to reverse because the caller's if block is not canonical,
3014 i.e. the THEN block isn't the fallthrough block for the TEST block
3015 (see find_if_header). */
3016 if (then_else_reversed)
3017 reverse = !reverse;
3019 /* If the condition variable is a register and is MODE_INT, accept it. */
3021 cond = XEXP (SET_SRC (set), 0);
3022 tmp = XEXP (cond, 0);
3023 if (REG_P (tmp) && GET_MODE_CLASS (GET_MODE (tmp)) == MODE_INT
3024 && (GET_MODE (tmp) != BImode
3025 || !targetm.small_register_classes_for_mode_p (BImode)))
3027 *earliest = jump;
3029 if (reverse)
3030 cond = gen_rtx_fmt_ee (reverse_condition (GET_CODE (cond)),
3031 GET_MODE (cond), tmp, XEXP (cond, 1));
3032 return cond;
3035 /* Otherwise, fall back on canonicalize_condition to do the dirty
3036 work of manipulating MODE_CC values and COMPARE rtx codes. */
3037 tmp = canonicalize_condition (jump, cond, reverse, earliest,
3038 NULL_RTX, have_cbranchcc4, true);
3040 /* We don't handle side-effects in the condition, like handling
3041 REG_INC notes and making sure no duplicate conditions are emitted. */
3042 if (tmp != NULL_RTX && side_effects_p (tmp))
3043 return NULL_RTX;
3045 return tmp;
3048 /* Return true if OP is ok for if-then-else processing. */
3050 static int
3051 noce_operand_ok (const_rtx op)
3053 if (side_effects_p (op))
3054 return FALSE;
3056 /* We special-case memories, so handle any of them with
3057 no address side effects. */
3058 if (MEM_P (op))
3059 return ! side_effects_p (XEXP (op, 0));
3061 return ! may_trap_p (op);
3064 /* Return true iff basic block TEST_BB is valid for noce if-conversion.
3065 The condition used in this if-conversion is in COND.
3066 In practice, check that TEST_BB ends with a single set
3067 x := a and all previous computations
3068 in TEST_BB don't produce any values that are live after TEST_BB.
3069 In other words, all the insns in TEST_BB are there only
3070 to compute a value for x. Add the rtx cost of the insns
3071 in TEST_BB to COST. Record whether TEST_BB is a single simple
3072 set instruction in SIMPLE_P. */
3074 static bool
3075 bb_valid_for_noce_process_p (basic_block test_bb, rtx cond,
3076 unsigned int *cost, bool *simple_p)
3078 if (!test_bb)
3079 return false;
3081 rtx_insn *last_insn = last_active_insn (test_bb, FALSE);
3082 rtx last_set = NULL_RTX;
3084 rtx cc = cc_in_cond (cond);
3086 if (!insn_valid_noce_process_p (last_insn, cc))
3087 return false;
3089 /* Punt on blocks ending with asm goto or jumps with other side-effects,
3090 last_active_insn ignores JUMP_INSNs. */
3091 if (JUMP_P (BB_END (test_bb)) && !onlyjump_p (BB_END (test_bb)))
3092 return false;
3094 last_set = single_set (last_insn);
3096 rtx x = SET_DEST (last_set);
3097 rtx_insn *first_insn = first_active_insn (test_bb);
3098 rtx first_set = single_set (first_insn);
3100 if (!first_set)
3101 return false;
3103 /* We have a single simple set, that's okay. */
3104 bool speed_p = optimize_bb_for_speed_p (test_bb);
3106 if (first_insn == last_insn)
3108 *simple_p = noce_operand_ok (SET_DEST (first_set));
3109 *cost += pattern_cost (first_set, speed_p);
3110 return *simple_p;
3113 rtx_insn *prev_last_insn = PREV_INSN (last_insn);
3114 gcc_assert (prev_last_insn);
3116 /* For now, disallow setting x multiple times in test_bb. */
3117 if (REG_P (x) && reg_set_between_p (x, first_insn, prev_last_insn))
3118 return false;
3120 bitmap test_bb_temps = BITMAP_ALLOC (&reg_obstack);
3122 /* The regs that are live out of test_bb. */
3123 bitmap test_bb_live_out = df_get_live_out (test_bb);
3125 int potential_cost = pattern_cost (last_set, speed_p);
3126 rtx_insn *insn;
3127 FOR_BB_INSNS (test_bb, insn)
3129 if (insn != last_insn)
3131 if (!active_insn_p (insn))
3132 continue;
3134 if (!insn_valid_noce_process_p (insn, cc))
3135 goto free_bitmap_and_fail;
3137 rtx sset = single_set (insn);
3138 gcc_assert (sset);
3140 if (contains_mem_rtx_p (SET_SRC (sset))
3141 || !REG_P (SET_DEST (sset))
3142 || reg_overlap_mentioned_p (SET_DEST (sset), cond))
3143 goto free_bitmap_and_fail;
3145 potential_cost += pattern_cost (sset, speed_p);
3146 bitmap_set_bit (test_bb_temps, REGNO (SET_DEST (sset)));
3150 /* If any of the intermediate results in test_bb are live after test_bb
3151 then fail. */
3152 if (bitmap_intersect_p (test_bb_live_out, test_bb_temps))
3153 goto free_bitmap_and_fail;
3155 BITMAP_FREE (test_bb_temps);
3156 *cost += potential_cost;
3157 *simple_p = false;
3158 return true;
3160 free_bitmap_and_fail:
3161 BITMAP_FREE (test_bb_temps);
3162 return false;
3165 /* Helper function to emit a cmov sequence encapsulated in
3166 start_sequence () and end_sequence (). If NEED_CMOV is true
3167 we call noce_emit_cmove to create a cmove sequence. Otherwise emit
3168 a simple move. If successful, store the first instruction of the
3169 sequence in TEMP_DEST and the sequence costs in SEQ_COST. */
3171 static rtx_insn*
3172 try_emit_cmove_seq (struct noce_if_info *if_info, rtx temp,
3173 rtx cond, rtx new_val, rtx old_val, bool need_cmov,
3174 unsigned *cost, rtx *temp_dest,
3175 rtx cc_cmp = NULL, rtx rev_cc_cmp = NULL)
3177 rtx_insn *seq = NULL;
3178 *cost = 0;
3180 rtx x = XEXP (cond, 0);
3181 rtx y = XEXP (cond, 1);
3182 rtx_code cond_code = GET_CODE (cond);
3184 start_sequence ();
3186 if (need_cmov)
3187 *temp_dest = noce_emit_cmove (if_info, temp, cond_code,
3188 x, y, new_val, old_val, cc_cmp, rev_cc_cmp);
3189 else
3191 *temp_dest = temp;
3192 if (if_info->then_else_reversed)
3193 noce_emit_move_insn (temp, old_val);
3194 else
3195 noce_emit_move_insn (temp, new_val);
3198 if (*temp_dest != NULL_RTX)
3200 seq = get_insns ();
3201 *cost = seq_cost (seq, if_info->speed_p);
3204 end_sequence ();
3206 return seq;
3209 /* We have something like:
3211 if (x > y)
3212 { i = a; j = b; k = c; }
3214 Make it:
3216 tmp_i = (x > y) ? a : i;
3217 tmp_j = (x > y) ? b : j;
3218 tmp_k = (x > y) ? c : k;
3219 i = tmp_i;
3220 j = tmp_j;
3221 k = tmp_k;
3223 Subsequent passes are expected to clean up the extra moves.
3225 Look for special cases such as writes to one register which are
3226 read back in another SET, as might occur in a swap idiom or
3227 similar.
3229 These look like:
3231 if (x > y)
3232 i = a;
3233 j = i;
3235 Which we want to rewrite to:
3237 tmp_i = (x > y) ? a : i;
3238 tmp_j = (x > y) ? tmp_i : j;
3239 i = tmp_i;
3240 j = tmp_j;
3242 We can catch these when looking at (SET x y) by keeping a list of the
3243 registers we would have targeted before if-conversion and looking back
3244 through it for an overlap with Y. If we find one, we rewire the
3245 conditional set to use the temporary we introduced earlier.
3247 IF_INFO contains the useful information about the block structure and
3248 jump instructions. */
3250 static int
3251 noce_convert_multiple_sets (struct noce_if_info *if_info)
3253 basic_block test_bb = if_info->test_bb;
3254 basic_block then_bb = if_info->then_bb;
3255 basic_block join_bb = if_info->join_bb;
3256 rtx_insn *jump = if_info->jump;
3257 rtx_insn *cond_earliest;
3258 rtx_insn *insn;
3260 start_sequence ();
3262 /* Decompose the condition attached to the jump. */
3263 rtx cond = noce_get_condition (jump, &cond_earliest, false);
3264 rtx x = XEXP (cond, 0);
3265 rtx y = XEXP (cond, 1);
3267 /* The true targets for a conditional move. */
3268 auto_vec<rtx> targets;
3269 /* The temporaries introduced to allow us to not consider register
3270 overlap. */
3271 auto_vec<rtx> temporaries;
3272 /* The insns we've emitted. */
3273 auto_vec<rtx_insn *> unmodified_insns;
3275 hash_set<rtx_insn *> need_no_cmov;
3276 hash_map<rtx_insn *, int> rewired_src;
3278 need_cmov_or_rewire (then_bb, &need_no_cmov, &rewired_src);
3280 int last_needs_comparison = -1;
3282 bool ok = noce_convert_multiple_sets_1
3283 (if_info, &need_no_cmov, &rewired_src, &targets, &temporaries,
3284 &unmodified_insns, &last_needs_comparison);
3285 if (!ok)
3286 return false;
3288 /* If there are insns that overwrite part of the initial
3289 comparison, we can still omit creating temporaries for
3290 the last of them.
3291 As the second try will always create a less expensive,
3292 valid sequence, we do not need to compare and can discard
3293 the first one. */
3294 if (last_needs_comparison != -1)
3296 end_sequence ();
3297 start_sequence ();
3298 ok = noce_convert_multiple_sets_1
3299 (if_info, &need_no_cmov, &rewired_src, &targets, &temporaries,
3300 &unmodified_insns, &last_needs_comparison);
3301 /* Actually we should not fail anymore if we reached here,
3302 but better still check. */
3303 if (!ok)
3304 return false;
3307 /* We must have seen some sort of insn to insert, otherwise we were
3308 given an empty BB to convert, and we can't handle that. */
3309 gcc_assert (!unmodified_insns.is_empty ());
3311 /* Now fixup the assignments. */
3312 for (unsigned i = 0; i < targets.length (); i++)
3313 if (targets[i] != temporaries[i])
3314 noce_emit_move_insn (targets[i], temporaries[i]);
3316 /* Actually emit the sequence if it isn't too expensive. */
3317 rtx_insn *seq = get_insns ();
3319 if (!targetm.noce_conversion_profitable_p (seq, if_info))
3321 end_sequence ();
3322 return FALSE;
3325 for (insn = seq; insn; insn = NEXT_INSN (insn))
3326 set_used_flags (insn);
3328 /* Mark all our temporaries and targets as used. */
3329 for (unsigned i = 0; i < targets.length (); i++)
3331 set_used_flags (temporaries[i]);
3332 set_used_flags (targets[i]);
3335 set_used_flags (cond);
3336 set_used_flags (x);
3337 set_used_flags (y);
3339 unshare_all_rtl_in_chain (seq);
3340 end_sequence ();
3342 if (!seq)
3343 return FALSE;
3345 for (insn = seq; insn; insn = NEXT_INSN (insn))
3346 if (JUMP_P (insn)
3347 || recog_memoized (insn) == -1)
3348 return FALSE;
3350 emit_insn_before_setloc (seq, if_info->jump,
3351 INSN_LOCATION (unmodified_insns.last ()));
3353 /* Clean up THEN_BB and the edges in and out of it. */
3354 remove_edge (find_edge (test_bb, join_bb));
3355 remove_edge (find_edge (then_bb, join_bb));
3356 redirect_edge_and_branch_force (single_succ_edge (test_bb), join_bb);
3357 delete_basic_block (then_bb);
3358 num_true_changes++;
3360 /* Maybe merge blocks now the jump is simple enough. */
3361 if (can_merge_blocks_p (test_bb, join_bb))
3363 merge_blocks (test_bb, join_bb);
3364 num_true_changes++;
3367 num_updated_if_blocks++;
3368 if_info->transform_name = "noce_convert_multiple_sets";
3369 return TRUE;
3372 /* Helper function for noce_convert_multiple_sets_1. If store to
3373 DEST can affect P[0] or P[1], clear P[0]. Called via note_stores. */
3375 static void
3376 check_for_cc_cmp_clobbers (rtx dest, const_rtx, void *p0)
3378 rtx *p = (rtx *) p0;
3379 if (p[0] == NULL_RTX)
3380 return;
3381 if (reg_overlap_mentioned_p (dest, p[0])
3382 || (p[1] && reg_overlap_mentioned_p (dest, p[1])))
3383 p[0] = NULL_RTX;
3386 /* This goes through all relevant insns of IF_INFO->then_bb and tries to
3387 create conditional moves. In case a simple move sufficis the insn
3388 should be listed in NEED_NO_CMOV. The rewired-src cases should be
3389 specified via REWIRED_SRC. TARGETS, TEMPORARIES and UNMODIFIED_INSNS
3390 are specified and used in noce_convert_multiple_sets and should be passed
3391 to this function.. */
3393 static bool
3394 noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
3395 hash_set<rtx_insn *> *need_no_cmov,
3396 hash_map<rtx_insn *, int> *rewired_src,
3397 auto_vec<rtx> *targets,
3398 auto_vec<rtx> *temporaries,
3399 auto_vec<rtx_insn *> *unmodified_insns,
3400 int *last_needs_comparison)
3402 basic_block then_bb = if_info->then_bb;
3403 rtx_insn *jump = if_info->jump;
3404 rtx_insn *cond_earliest;
3406 /* Decompose the condition attached to the jump. */
3407 rtx cond = noce_get_condition (jump, &cond_earliest, false);
3409 rtx cc_cmp = cond_exec_get_condition (jump);
3410 if (cc_cmp)
3411 cc_cmp = copy_rtx (cc_cmp);
3412 rtx rev_cc_cmp = cond_exec_get_condition (jump, /* get_reversed */ true);
3413 if (rev_cc_cmp)
3414 rev_cc_cmp = copy_rtx (rev_cc_cmp);
3416 rtx_insn *insn;
3417 int count = 0;
3419 targets->truncate (0);
3420 temporaries->truncate (0);
3421 unmodified_insns->truncate (0);
3423 bool second_try = *last_needs_comparison != -1;
3425 FOR_BB_INSNS (then_bb, insn)
3427 /* Skip over non-insns. */
3428 if (!active_insn_p (insn))
3429 continue;
3431 rtx set = single_set (insn);
3432 gcc_checking_assert (set);
3434 rtx target = SET_DEST (set);
3435 rtx temp;
3437 rtx new_val = SET_SRC (set);
3438 if (int *ii = rewired_src->get (insn))
3439 new_val = simplify_replace_rtx (new_val, (*targets)[*ii],
3440 (*temporaries)[*ii]);
3441 rtx old_val = target;
3443 /* As we are transforming
3444 if (x > y)
3446 a = b;
3447 c = d;
3449 into
3450 a = (x > y) ...
3451 c = (x > y) ...
3453 we potentially check x > y before every set.
3454 Even though the check might be removed by subsequent passes, this means
3455 that we cannot transform
3456 if (x > y)
3458 x = y;
3461 into
3462 x = (x > y) ...
3464 since this would invalidate x and the following to-be-removed checks.
3465 Therefore we introduce a temporary every time we are about to
3466 overwrite a variable used in the check. Costing of a sequence with
3467 these is going to be inaccurate so only use temporaries when
3468 needed.
3470 If performing a second try, we know how many insns require a
3471 temporary. For the last of these, we can omit creating one. */
3472 if (reg_overlap_mentioned_p (target, cond)
3473 && (!second_try || count < *last_needs_comparison))
3474 temp = gen_reg_rtx (GET_MODE (target));
3475 else
3476 temp = target;
3478 /* We have identified swap-style idioms before. A normal
3479 set will need to be a cmov while the first instruction of a swap-style
3480 idiom can be a regular move. This helps with costing. */
3481 bool need_cmov = !need_no_cmov->contains (insn);
3483 /* If we had a non-canonical conditional jump (i.e. one where
3484 the fallthrough is to the "else" case) we need to reverse
3485 the conditional select. */
3486 if (if_info->then_else_reversed)
3487 std::swap (old_val, new_val);
3490 /* We allow simple lowpart register subreg SET sources in
3491 bb_ok_for_noce_convert_multiple_sets. Be careful when processing
3492 sequences like:
3493 (set (reg:SI r1) (reg:SI r2))
3494 (set (reg:HI r3) (subreg:HI (r1)))
3495 For the second insn new_val or old_val (r1 in this example) will be
3496 taken from the temporaries and have the wider mode which will not
3497 match with the mode of the other source of the conditional move, so
3498 we'll end up trying to emit r4:HI = cond ? (r1:SI) : (r3:HI).
3499 Wrap the two cmove operands into subregs if appropriate to prevent
3500 that. */
3502 if (!CONSTANT_P (new_val)
3503 && GET_MODE (new_val) != GET_MODE (temp))
3505 machine_mode src_mode = GET_MODE (new_val);
3506 machine_mode dst_mode = GET_MODE (temp);
3507 if (!partial_subreg_p (dst_mode, src_mode))
3509 end_sequence ();
3510 return FALSE;
3512 new_val = lowpart_subreg (dst_mode, new_val, src_mode);
3514 if (!CONSTANT_P (old_val)
3515 && GET_MODE (old_val) != GET_MODE (temp))
3517 machine_mode src_mode = GET_MODE (old_val);
3518 machine_mode dst_mode = GET_MODE (temp);
3519 if (!partial_subreg_p (dst_mode, src_mode))
3521 end_sequence ();
3522 return FALSE;
3524 old_val = lowpart_subreg (dst_mode, old_val, src_mode);
3527 /* Try emitting a conditional move passing the backend the
3528 canonicalized comparison. The backend is then able to
3529 recognize expressions like
3531 if (x > y)
3532 y = x;
3534 as min/max and emit an insn, accordingly. */
3535 unsigned cost1 = 0, cost2 = 0;
3536 rtx_insn *seq, *seq1, *seq2 = NULL;
3537 rtx temp_dest = NULL_RTX, temp_dest1 = NULL_RTX, temp_dest2 = NULL_RTX;
3538 bool read_comparison = false;
3540 seq1 = try_emit_cmove_seq (if_info, temp, cond,
3541 new_val, old_val, need_cmov,
3542 &cost1, &temp_dest1);
3544 /* Here, we try to pass the backend a non-canonicalized cc comparison
3545 as well. This allows the backend to emit a cmov directly without
3546 creating an additional compare for each. If successful, costing
3547 is easier and this sequence is usually preferred. */
3548 if (cc_cmp)
3549 seq2 = try_emit_cmove_seq (if_info, temp, cond,
3550 new_val, old_val, need_cmov,
3551 &cost2, &temp_dest2, cc_cmp, rev_cc_cmp);
3553 /* The backend might have created a sequence that uses the
3554 condition. Check this. */
3555 rtx_insn *walk = seq2;
3556 while (walk)
3558 rtx set = single_set (walk);
3560 if (!set || !SET_SRC (set))
3562 walk = NEXT_INSN (walk);
3563 continue;
3566 rtx src = SET_SRC (set);
3568 if (XEXP (set, 1) && GET_CODE (XEXP (set, 1)) == IF_THEN_ELSE)
3569 ; /* We assume that this is the cmove created by the backend that
3570 naturally uses the condition. Therefore we ignore it. */
3571 else
3573 if (reg_mentioned_p (XEXP (cond, 0), src)
3574 || reg_mentioned_p (XEXP (cond, 1), src))
3576 read_comparison = true;
3577 break;
3581 walk = NEXT_INSN (walk);
3584 /* Check which version is less expensive. */
3585 if (seq1 != NULL_RTX && (cost1 <= cost2 || seq2 == NULL_RTX))
3587 seq = seq1;
3588 temp_dest = temp_dest1;
3589 if (!second_try)
3590 *last_needs_comparison = count;
3592 else if (seq2 != NULL_RTX)
3594 seq = seq2;
3595 temp_dest = temp_dest2;
3596 if (!second_try && read_comparison)
3597 *last_needs_comparison = count;
3599 else
3601 /* Nothing worked, bail out. */
3602 end_sequence ();
3603 return FALSE;
3606 if (cc_cmp)
3608 /* Check if SEQ can clobber registers mentioned in
3609 cc_cmp and/or rev_cc_cmp. If yes, we need to use
3610 only seq1 from that point on. */
3611 rtx cc_cmp_pair[2] = { cc_cmp, rev_cc_cmp };
3612 for (walk = seq; walk; walk = NEXT_INSN (walk))
3614 note_stores (walk, check_for_cc_cmp_clobbers, cc_cmp_pair);
3615 if (cc_cmp_pair[0] == NULL_RTX)
3617 cc_cmp = NULL_RTX;
3618 rev_cc_cmp = NULL_RTX;
3619 break;
3624 /* End the sub sequence and emit to the main sequence. */
3625 emit_insn (seq);
3627 /* Bookkeeping. */
3628 count++;
3629 targets->safe_push (target);
3630 temporaries->safe_push (temp_dest);
3631 unmodified_insns->safe_push (insn);
3634 /* Even if we did not actually need the comparison, we want to make sure
3635 to try a second time in order to get rid of the temporaries. */
3636 if (*last_needs_comparison == -1)
3637 *last_needs_comparison = 0;
3640 return true;
3645 /* Return true iff basic block TEST_BB is comprised of only
3646 (SET (REG) (REG)) insns suitable for conversion to a series
3647 of conditional moves. Also check that we have more than one set
3648 (other routines can handle a single set better than we would), and
3649 fewer than PARAM_MAX_RTL_IF_CONVERSION_INSNS sets. While going
3650 through the insns store the sum of their potential costs in COST. */
3652 static bool
3653 bb_ok_for_noce_convert_multiple_sets (basic_block test_bb, unsigned *cost)
3655 rtx_insn *insn;
3656 unsigned count = 0;
3657 unsigned param = param_max_rtl_if_conversion_insns;
3658 bool speed_p = optimize_bb_for_speed_p (test_bb);
3659 unsigned potential_cost = 0;
3661 FOR_BB_INSNS (test_bb, insn)
3663 /* Skip over notes etc. */
3664 if (!active_insn_p (insn))
3665 continue;
3667 /* We only handle SET insns. */
3668 rtx set = single_set (insn);
3669 if (set == NULL_RTX)
3670 return false;
3672 rtx dest = SET_DEST (set);
3673 rtx src = SET_SRC (set);
3675 /* We can possibly relax this, but for now only handle REG to REG
3676 (including subreg) moves. This avoids any issues that might come
3677 from introducing loads/stores that might violate data-race-freedom
3678 guarantees. */
3679 if (!REG_P (dest))
3680 return false;
3682 if (!((REG_P (src) || CONSTANT_P (src))
3683 || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
3684 && subreg_lowpart_p (src))))
3685 return false;
3687 /* Destination must be appropriate for a conditional write. */
3688 if (!noce_operand_ok (dest))
3689 return false;
3691 /* We must be able to conditionally move in this mode. */
3692 if (!can_conditionally_move_p (GET_MODE (dest)))
3693 return false;
3695 potential_cost += insn_cost (insn, speed_p);
3697 count++;
3700 *cost += potential_cost;
3702 /* If we would only put out one conditional move, the other strategies
3703 this pass tries are better optimized and will be more appropriate.
3704 Some targets want to strictly limit the number of conditional moves
3705 that are emitted, they set this through PARAM, we need to respect
3706 that. */
3707 return count > 1 && count <= param;
3710 /* Compute average of two given costs weighted by relative probabilities
3711 of respective basic blocks in an IF-THEN-ELSE. E is the IF-THEN edge.
3712 With P as the probability to take the IF-THEN branch, return
3713 P * THEN_COST + (1 - P) * ELSE_COST. */
3714 static unsigned
3715 average_cost (unsigned then_cost, unsigned else_cost, edge e)
3717 return else_cost + e->probability.apply ((signed) (then_cost - else_cost));
3720 /* Given a simple IF-THEN-JOIN or IF-THEN-ELSE-JOIN block, attempt to convert
3721 it without using conditional execution. Return TRUE if we were successful
3722 at converting the block. */
3724 static int
3725 noce_process_if_block (struct noce_if_info *if_info)
3727 basic_block test_bb = if_info->test_bb; /* test block */
3728 basic_block then_bb = if_info->then_bb; /* THEN */
3729 basic_block else_bb = if_info->else_bb; /* ELSE or NULL */
3730 basic_block join_bb = if_info->join_bb; /* JOIN */
3731 rtx_insn *jump = if_info->jump;
3732 rtx cond = if_info->cond;
3733 rtx_insn *insn_a, *insn_b;
3734 rtx set_a, set_b;
3735 rtx orig_x, x, a, b;
3737 /* We're looking for patterns of the form
3739 (1) if (...) x = a; else x = b;
3740 (2) x = b; if (...) x = a;
3741 (3) if (...) x = a; // as if with an initial x = x.
3742 (4) if (...) { x = a; y = b; z = c; } // Like 3, for multiple SETS.
3743 The later patterns require jumps to be more expensive.
3744 For the if (...) x = a; else x = b; case we allow multiple insns
3745 inside the then and else blocks as long as their only effect is
3746 to calculate a value for x.
3747 ??? For future expansion, further expand the "multiple X" rules. */
3749 /* First look for multiple SETS. The original costs already include
3750 a base cost of COSTS_N_INSNS (2): one instruction for the compare
3751 (which we will be needing either way) and one instruction for the
3752 branch. When comparing costs we want to use the branch instruction
3753 cost and the sets vs. the cmovs generated here. Therefore subtract
3754 the costs of the compare before checking.
3755 ??? Actually, instead of the branch instruction costs we might want
3756 to use COSTS_N_INSNS (BRANCH_COST ()) as in other places. */
3758 unsigned potential_cost = if_info->original_cost - COSTS_N_INSNS (1);
3759 unsigned old_cost = if_info->original_cost;
3760 if (!else_bb
3761 && HAVE_conditional_move
3762 && bb_ok_for_noce_convert_multiple_sets (then_bb, &potential_cost))
3764 /* Temporarily set the original costs to what we estimated so
3765 we can determine if the transformation is worth it. */
3766 if_info->original_cost = potential_cost;
3767 if (noce_convert_multiple_sets (if_info))
3769 if (dump_file && if_info->transform_name)
3770 fprintf (dump_file, "if-conversion succeeded through %s\n",
3771 if_info->transform_name);
3772 return TRUE;
3775 /* Restore the original costs. */
3776 if_info->original_cost = old_cost;
3779 bool speed_p = optimize_bb_for_speed_p (test_bb);
3780 unsigned int then_cost = 0, else_cost = 0;
3781 if (!bb_valid_for_noce_process_p (then_bb, cond, &then_cost,
3782 &if_info->then_simple))
3783 return false;
3785 if (else_bb
3786 && !bb_valid_for_noce_process_p (else_bb, cond, &else_cost,
3787 &if_info->else_simple))
3788 return false;
3790 if (speed_p)
3791 if_info->original_cost += average_cost (then_cost, else_cost,
3792 find_edge (test_bb, then_bb));
3793 else
3794 if_info->original_cost += then_cost + else_cost;
3796 insn_a = last_active_insn (then_bb, FALSE);
3797 set_a = single_set (insn_a);
3798 gcc_assert (set_a);
3800 x = SET_DEST (set_a);
3801 a = SET_SRC (set_a);
3803 /* Look for the other potential set. Make sure we've got equivalent
3804 destinations. */
3805 /* ??? This is overconservative. Storing to two different mems is
3806 as easy as conditionally computing the address. Storing to a
3807 single mem merely requires a scratch memory to use as one of the
3808 destination addresses; often the memory immediately below the
3809 stack pointer is available for this. */
3810 set_b = NULL_RTX;
3811 if (else_bb)
3813 insn_b = last_active_insn (else_bb, FALSE);
3814 set_b = single_set (insn_b);
3815 gcc_assert (set_b);
3817 if (!rtx_interchangeable_p (x, SET_DEST (set_b)))
3818 return FALSE;
3820 else
3822 insn_b = if_info->cond_earliest;
3824 insn_b = prev_nonnote_nondebug_insn (insn_b);
3825 while (insn_b
3826 && (BLOCK_FOR_INSN (insn_b)
3827 == BLOCK_FOR_INSN (if_info->cond_earliest))
3828 && !modified_in_p (x, insn_b));
3830 /* We're going to be moving the evaluation of B down from above
3831 COND_EARLIEST to JUMP. Make sure the relevant data is still
3832 intact. */
3833 if (! insn_b
3834 || BLOCK_FOR_INSN (insn_b) != BLOCK_FOR_INSN (if_info->cond_earliest)
3835 || !NONJUMP_INSN_P (insn_b)
3836 || (set_b = single_set (insn_b)) == NULL_RTX
3837 || ! rtx_interchangeable_p (x, SET_DEST (set_b))
3838 || ! noce_operand_ok (SET_SRC (set_b))
3839 || reg_overlap_mentioned_p (x, SET_SRC (set_b))
3840 || modified_between_p (SET_SRC (set_b), insn_b, jump)
3841 /* Avoid extending the lifetime of hard registers on small
3842 register class machines. */
3843 || (REG_P (SET_SRC (set_b))
3844 && HARD_REGISTER_P (SET_SRC (set_b))
3845 && targetm.small_register_classes_for_mode_p
3846 (GET_MODE (SET_SRC (set_b))))
3847 /* Likewise with X. In particular this can happen when
3848 noce_get_condition looks farther back in the instruction
3849 stream than one might expect. */
3850 || reg_overlap_mentioned_p (x, cond)
3851 || reg_overlap_mentioned_p (x, a)
3852 || modified_between_p (x, insn_b, jump))
3854 insn_b = NULL;
3855 set_b = NULL_RTX;
3859 /* If x has side effects then only the if-then-else form is safe to
3860 convert. But even in that case we would need to restore any notes
3861 (such as REG_INC) at then end. That can be tricky if
3862 noce_emit_move_insn expands to more than one insn, so disable the
3863 optimization entirely for now if there are side effects. */
3864 if (side_effects_p (x))
3865 return FALSE;
3867 b = (set_b ? SET_SRC (set_b) : x);
3869 /* Only operate on register destinations, and even then avoid extending
3870 the lifetime of hard registers on small register class machines. */
3871 orig_x = x;
3872 if_info->orig_x = orig_x;
3873 if (!REG_P (x)
3874 || (HARD_REGISTER_P (x)
3875 && targetm.small_register_classes_for_mode_p (GET_MODE (x))))
3877 if (GET_MODE (x) == BLKmode)
3878 return FALSE;
3880 if (GET_CODE (x) == ZERO_EXTRACT
3881 && (!CONST_INT_P (XEXP (x, 1))
3882 || !CONST_INT_P (XEXP (x, 2))))
3883 return FALSE;
3885 x = gen_reg_rtx (GET_MODE (GET_CODE (x) == STRICT_LOW_PART
3886 ? XEXP (x, 0) : x));
3889 /* Don't operate on sources that may trap or are volatile. */
3890 if (! noce_operand_ok (a) || ! noce_operand_ok (b))
3891 return FALSE;
3893 retry:
3894 /* Set up the info block for our subroutines. */
3895 if_info->insn_a = insn_a;
3896 if_info->insn_b = insn_b;
3897 if_info->x = x;
3898 if_info->a = a;
3899 if_info->b = b;
3901 /* Try optimizations in some approximation of a useful order. */
3902 /* ??? Should first look to see if X is live incoming at all. If it
3903 isn't, we don't need anything but an unconditional set. */
3905 /* Look and see if A and B are really the same. Avoid creating silly
3906 cmove constructs that no one will fix up later. */
3907 if (noce_simple_bbs (if_info)
3908 && rtx_interchangeable_p (a, b))
3910 /* If we have an INSN_B, we don't have to create any new rtl. Just
3911 move the instruction that we already have. If we don't have an
3912 INSN_B, that means that A == X, and we've got a noop move. In
3913 that case don't do anything and let the code below delete INSN_A. */
3914 if (insn_b && else_bb)
3916 rtx note;
3918 if (else_bb && insn_b == BB_END (else_bb))
3919 BB_END (else_bb) = PREV_INSN (insn_b);
3920 reorder_insns (insn_b, insn_b, PREV_INSN (jump));
3922 /* If there was a REG_EQUAL note, delete it since it may have been
3923 true due to this insn being after a jump. */
3924 if ((note = find_reg_note (insn_b, REG_EQUAL, NULL_RTX)) != 0)
3925 remove_note (insn_b, note);
3927 insn_b = NULL;
3929 /* If we have "x = b; if (...) x = a;", and x has side-effects, then
3930 x must be executed twice. */
3931 else if (insn_b && side_effects_p (orig_x))
3932 return FALSE;
3934 x = orig_x;
3935 goto success;
3938 if (!set_b && MEM_P (orig_x))
3939 /* We want to avoid store speculation to avoid cases like
3940 if (pthread_mutex_trylock(mutex))
3941 ++global_variable;
3942 Rather than go to much effort here, we rely on the SSA optimizers,
3943 which do a good enough job these days. */
3944 return FALSE;
3946 if (noce_try_move (if_info))
3947 goto success;
3948 if (noce_try_ifelse_collapse (if_info))
3949 goto success;
3950 if (noce_try_store_flag (if_info))
3951 goto success;
3952 if (noce_try_bitop (if_info))
3953 goto success;
3954 if (noce_try_minmax (if_info))
3955 goto success;
3956 if (noce_try_abs (if_info))
3957 goto success;
3958 if (noce_try_inverse_constants (if_info))
3959 goto success;
3960 if (!targetm.have_conditional_execution ()
3961 && noce_try_store_flag_constants (if_info))
3962 goto success;
3963 if (HAVE_conditional_move
3964 && noce_try_cmove (if_info))
3965 goto success;
3966 if (! targetm.have_conditional_execution ())
3968 if (noce_try_addcc (if_info))
3969 goto success;
3970 if (noce_try_store_flag_mask (if_info))
3971 goto success;
3972 if (HAVE_conditional_move
3973 && noce_try_cmove_arith (if_info))
3974 goto success;
3975 if (noce_try_sign_mask (if_info))
3976 goto success;
3979 if (!else_bb && set_b)
3981 insn_b = NULL;
3982 set_b = NULL_RTX;
3983 b = orig_x;
3984 goto retry;
3987 return FALSE;
3989 success:
3990 if (dump_file && if_info->transform_name)
3991 fprintf (dump_file, "if-conversion succeeded through %s\n",
3992 if_info->transform_name);
3994 /* If we used a temporary, fix it up now. */
3995 if (orig_x != x)
3997 rtx_insn *seq;
3999 start_sequence ();
4000 noce_emit_move_insn (orig_x, x);
4001 seq = get_insns ();
4002 set_used_flags (orig_x);
4003 unshare_all_rtl_in_chain (seq);
4004 end_sequence ();
4006 emit_insn_before_setloc (seq, BB_END (test_bb), INSN_LOCATION (insn_a));
4009 /* The original THEN and ELSE blocks may now be removed. The test block
4010 must now jump to the join block. If the test block and the join block
4011 can be merged, do so. */
4012 if (else_bb)
4014 delete_basic_block (else_bb);
4015 num_true_changes++;
4017 else
4018 remove_edge (find_edge (test_bb, join_bb));
4020 remove_edge (find_edge (then_bb, join_bb));
4021 redirect_edge_and_branch_force (single_succ_edge (test_bb), join_bb);
4022 delete_basic_block (then_bb);
4023 num_true_changes++;
4025 if (can_merge_blocks_p (test_bb, join_bb))
4027 merge_blocks (test_bb, join_bb);
4028 num_true_changes++;
4031 num_updated_if_blocks++;
4032 return TRUE;
4035 /* Check whether a block is suitable for conditional move conversion.
4036 Every insn must be a simple set of a register to a constant or a
4037 register. For each assignment, store the value in the pointer map
4038 VALS, keyed indexed by register pointer, then store the register
4039 pointer in REGS. COND is the condition we will test. */
4041 static int
4042 check_cond_move_block (basic_block bb,
4043 hash_map<rtx, rtx> *vals,
4044 vec<rtx> *regs,
4045 rtx cond)
4047 rtx_insn *insn;
4048 rtx cc = cc_in_cond (cond);
4050 /* We can only handle simple jumps at the end of the basic block.
4051 It is almost impossible to update the CFG otherwise. */
4052 insn = BB_END (bb);
4053 if (JUMP_P (insn) && !onlyjump_p (insn))
4054 return FALSE;
4056 FOR_BB_INSNS (bb, insn)
4058 rtx set, dest, src;
4060 if (!NONDEBUG_INSN_P (insn) || JUMP_P (insn))
4061 continue;
4062 set = single_set (insn);
4063 if (!set)
4064 return FALSE;
4066 dest = SET_DEST (set);
4067 src = SET_SRC (set);
4068 if (!REG_P (dest)
4069 || (HARD_REGISTER_P (dest)
4070 && targetm.small_register_classes_for_mode_p (GET_MODE (dest))))
4071 return FALSE;
4073 if (!CONSTANT_P (src) && !register_operand (src, VOIDmode))
4074 return FALSE;
4076 if (side_effects_p (src) || side_effects_p (dest))
4077 return FALSE;
4079 if (may_trap_p (src) || may_trap_p (dest))
4080 return FALSE;
4082 /* Don't try to handle this if the source register was
4083 modified earlier in the block. */
4084 if ((REG_P (src)
4085 && vals->get (src))
4086 || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
4087 && vals->get (SUBREG_REG (src))))
4088 return FALSE;
4090 /* Don't try to handle this if the destination register was
4091 modified earlier in the block. */
4092 if (vals->get (dest))
4093 return FALSE;
4095 /* Don't try to handle this if the condition uses the
4096 destination register. */
4097 if (reg_overlap_mentioned_p (dest, cond))
4098 return FALSE;
4100 /* Don't try to handle this if the source register is modified
4101 later in the block. */
4102 if (!CONSTANT_P (src)
4103 && modified_between_p (src, insn, NEXT_INSN (BB_END (bb))))
4104 return FALSE;
4106 /* Skip it if the instruction to be moved might clobber CC. */
4107 if (cc && set_of (cc, insn))
4108 return FALSE;
4110 vals->put (dest, src);
4112 regs->safe_push (dest);
4115 return TRUE;
4118 /* Find local swap-style idioms in BB and mark the first insn (1)
4119 that is only a temporary as not needing a conditional move as
4120 it is going to be dead afterwards anyway.
4122 (1) int tmp = a;
4123 a = b;
4124 b = tmp;
4126 ifcvt
4129 tmp = a;
4130 a = cond ? b : a_old;
4131 b = cond ? tmp : b_old;
4133 Additionally, store the index of insns like (2) when a subsequent
4134 SET reads from their destination.
4136 (2) int c = a;
4137 int d = c;
4139 ifcvt
4142 c = cond ? a : c_old;
4143 d = cond ? d : c; // Need to use c rather than c_old here.
4146 static void
4147 need_cmov_or_rewire (basic_block bb,
4148 hash_set<rtx_insn *> *need_no_cmov,
4149 hash_map<rtx_insn *, int> *rewired_src)
4151 rtx_insn *insn;
4152 int count = 0;
4153 auto_vec<rtx_insn *> insns;
4154 auto_vec<rtx> dests;
4156 /* Iterate over all SETs, storing the destinations
4157 in DEST.
4158 - If we hit a SET that reads from a destination
4159 that we have seen before and the corresponding register
4160 is dead afterwards, the register does not need to be
4161 moved conditionally.
4162 - If we encounter a previously changed register,
4163 rewire the read to the original source. */
4164 FOR_BB_INSNS (bb, insn)
4166 rtx set, src, dest;
4168 if (!active_insn_p (insn))
4169 continue;
4171 set = single_set (insn);
4172 if (set == NULL_RTX)
4173 continue;
4175 src = SET_SRC (set);
4176 if (SUBREG_P (src))
4177 src = SUBREG_REG (src);
4178 dest = SET_DEST (set);
4180 /* Check if the current SET's source is the same
4181 as any previously seen destination.
4182 This is quadratic but the number of insns in BB
4183 is bounded by PARAM_MAX_RTL_IF_CONVERSION_INSNS. */
4184 if (REG_P (src))
4185 for (int i = count - 1; i >= 0; --i)
4186 if (reg_overlap_mentioned_p (src, dests[i]))
4188 if (find_reg_note (insn, REG_DEAD, src) != NULL_RTX)
4189 need_no_cmov->add (insns[i]);
4190 else
4191 rewired_src->put (insn, i);
4194 insns.safe_push (insn);
4195 dests.safe_push (dest);
4197 count++;
4201 /* Given a basic block BB suitable for conditional move conversion,
4202 a condition COND, and pointer maps THEN_VALS and ELSE_VALS containing
4203 the register values depending on COND, emit the insns in the block as
4204 conditional moves. If ELSE_BLOCK is true, THEN_BB was already
4205 processed. The caller has started a sequence for the conversion.
4206 Return true if successful, false if something goes wrong. */
4208 static bool
4209 cond_move_convert_if_block (struct noce_if_info *if_infop,
4210 basic_block bb, rtx cond,
4211 hash_map<rtx, rtx> *then_vals,
4212 hash_map<rtx, rtx> *else_vals,
4213 bool else_block_p)
4215 enum rtx_code code;
4216 rtx_insn *insn;
4217 rtx cond_arg0, cond_arg1;
4219 code = GET_CODE (cond);
4220 cond_arg0 = XEXP (cond, 0);
4221 cond_arg1 = XEXP (cond, 1);
4223 FOR_BB_INSNS (bb, insn)
4225 rtx set, target, dest, t, e;
4227 /* ??? Maybe emit conditional debug insn? */
4228 if (!NONDEBUG_INSN_P (insn) || JUMP_P (insn))
4229 continue;
4230 set = single_set (insn);
4231 gcc_assert (set && REG_P (SET_DEST (set)));
4233 dest = SET_DEST (set);
4235 rtx *then_slot = then_vals->get (dest);
4236 rtx *else_slot = else_vals->get (dest);
4237 t = then_slot ? *then_slot : NULL_RTX;
4238 e = else_slot ? *else_slot : NULL_RTX;
4240 if (else_block_p)
4242 /* If this register was set in the then block, we already
4243 handled this case there. */
4244 if (t)
4245 continue;
4246 t = dest;
4247 gcc_assert (e);
4249 else
4251 gcc_assert (t);
4252 if (!e)
4253 e = dest;
4256 target = noce_emit_cmove (if_infop, dest, code, cond_arg0, cond_arg1,
4257 t, e);
4258 if (!target)
4259 return false;
4261 if (target != dest)
4262 noce_emit_move_insn (dest, target);
4265 return true;
4268 /* Given a simple IF-THEN-JOIN or IF-THEN-ELSE-JOIN block, attempt to convert
4269 it using only conditional moves. Return TRUE if we were successful at
4270 converting the block. */
4272 static int
4273 cond_move_process_if_block (struct noce_if_info *if_info)
4275 basic_block test_bb = if_info->test_bb;
4276 basic_block then_bb = if_info->then_bb;
4277 basic_block else_bb = if_info->else_bb;
4278 basic_block join_bb = if_info->join_bb;
4279 rtx_insn *jump = if_info->jump;
4280 rtx cond = if_info->cond;
4281 rtx_insn *seq, *loc_insn;
4282 int c;
4283 vec<rtx> then_regs = vNULL;
4284 vec<rtx> else_regs = vNULL;
4285 int success_p = FALSE;
4286 int limit = param_max_rtl_if_conversion_insns;
4288 /* Build a mapping for each block to the value used for each
4289 register. */
4290 hash_map<rtx, rtx> then_vals;
4291 hash_map<rtx, rtx> else_vals;
4293 /* Make sure the blocks are suitable. */
4294 if (!check_cond_move_block (then_bb, &then_vals, &then_regs, cond)
4295 || (else_bb
4296 && !check_cond_move_block (else_bb, &else_vals, &else_regs, cond)))
4297 goto done;
4299 /* Make sure the blocks can be used together. If the same register
4300 is set in both blocks, and is not set to a constant in both
4301 cases, then both blocks must set it to the same register. We
4302 have already verified that if it is set to a register, that the
4303 source register does not change after the assignment. Also count
4304 the number of registers set in only one of the blocks. */
4305 c = 0;
4306 for (rtx reg : then_regs)
4308 rtx *then_slot = then_vals.get (reg);
4309 rtx *else_slot = else_vals.get (reg);
4311 gcc_checking_assert (then_slot);
4312 if (!else_slot)
4313 ++c;
4314 else
4316 rtx then_val = *then_slot;
4317 rtx else_val = *else_slot;
4318 if (!CONSTANT_P (then_val) && !CONSTANT_P (else_val)
4319 && !rtx_equal_p (then_val, else_val))
4320 goto done;
4324 /* Finish off c for MAX_CONDITIONAL_EXECUTE. */
4325 for (rtx reg : else_regs)
4327 gcc_checking_assert (else_vals.get (reg));
4328 if (!then_vals.get (reg))
4329 ++c;
4332 /* Make sure it is reasonable to convert this block. What matters
4333 is the number of assignments currently made in only one of the
4334 branches, since if we convert we are going to always execute
4335 them. */
4336 if (c > MAX_CONDITIONAL_EXECUTE
4337 || c > limit)
4338 goto done;
4340 /* Try to emit the conditional moves. First do the then block,
4341 then do anything left in the else blocks. */
4342 start_sequence ();
4343 if (!cond_move_convert_if_block (if_info, then_bb, cond,
4344 &then_vals, &else_vals, false)
4345 || (else_bb
4346 && !cond_move_convert_if_block (if_info, else_bb, cond,
4347 &then_vals, &else_vals, true)))
4349 end_sequence ();
4350 goto done;
4352 seq = end_ifcvt_sequence (if_info);
4353 if (!seq)
4354 goto done;
4356 loc_insn = first_active_insn (then_bb);
4357 if (!loc_insn)
4359 loc_insn = first_active_insn (else_bb);
4360 gcc_assert (loc_insn);
4362 emit_insn_before_setloc (seq, jump, INSN_LOCATION (loc_insn));
4364 if (else_bb)
4366 delete_basic_block (else_bb);
4367 num_true_changes++;
4369 else
4370 remove_edge (find_edge (test_bb, join_bb));
4372 remove_edge (find_edge (then_bb, join_bb));
4373 redirect_edge_and_branch_force (single_succ_edge (test_bb), join_bb);
4374 delete_basic_block (then_bb);
4375 num_true_changes++;
4377 if (can_merge_blocks_p (test_bb, join_bb))
4379 merge_blocks (test_bb, join_bb);
4380 num_true_changes++;
4383 num_updated_if_blocks++;
4384 success_p = TRUE;
4386 done:
4387 then_regs.release ();
4388 else_regs.release ();
4389 return success_p;
4393 /* Determine if a given basic block heads a simple IF-THEN-JOIN or an
4394 IF-THEN-ELSE-JOIN block.
4396 If so, we'll try to convert the insns to not require the branch,
4397 using only transformations that do not require conditional execution.
4399 Return TRUE if we were successful at converting the block. */
4401 static int
4402 noce_find_if_block (basic_block test_bb, edge then_edge, edge else_edge,
4403 int pass)
4405 basic_block then_bb, else_bb, join_bb;
4406 bool then_else_reversed = false;
4407 rtx_insn *jump;
4408 rtx cond;
4409 rtx_insn *cond_earliest;
4410 struct noce_if_info if_info;
4411 bool speed_p = optimize_bb_for_speed_p (test_bb);
4413 /* We only ever should get here before reload. */
4414 gcc_assert (!reload_completed);
4416 /* Recognize an IF-THEN-ELSE-JOIN block. */
4417 if (single_pred_p (then_edge->dest)
4418 && single_succ_p (then_edge->dest)
4419 && single_pred_p (else_edge->dest)
4420 && single_succ_p (else_edge->dest)
4421 && single_succ (then_edge->dest) == single_succ (else_edge->dest))
4423 then_bb = then_edge->dest;
4424 else_bb = else_edge->dest;
4425 join_bb = single_succ (then_bb);
4427 /* Recognize an IF-THEN-JOIN block. */
4428 else if (single_pred_p (then_edge->dest)
4429 && single_succ_p (then_edge->dest)
4430 && single_succ (then_edge->dest) == else_edge->dest)
4432 then_bb = then_edge->dest;
4433 else_bb = NULL_BLOCK;
4434 join_bb = else_edge->dest;
4436 /* Recognize an IF-ELSE-JOIN block. We can have those because the order
4437 of basic blocks in cfglayout mode does not matter, so the fallthrough
4438 edge can go to any basic block (and not just to bb->next_bb, like in
4439 cfgrtl mode). */
4440 else if (single_pred_p (else_edge->dest)
4441 && single_succ_p (else_edge->dest)
4442 && single_succ (else_edge->dest) == then_edge->dest)
4444 /* The noce transformations do not apply to IF-ELSE-JOIN blocks.
4445 To make this work, we have to invert the THEN and ELSE blocks
4446 and reverse the jump condition. */
4447 then_bb = else_edge->dest;
4448 else_bb = NULL_BLOCK;
4449 join_bb = single_succ (then_bb);
4450 then_else_reversed = true;
4452 else
4453 /* Not a form we can handle. */
4454 return FALSE;
4456 /* The edges of the THEN and ELSE blocks cannot have complex edges. */
4457 if (single_succ_edge (then_bb)->flags & EDGE_COMPLEX)
4458 return FALSE;
4459 if (else_bb
4460 && single_succ_edge (else_bb)->flags & EDGE_COMPLEX)
4461 return FALSE;
4463 num_possible_if_blocks++;
4465 if (dump_file)
4467 fprintf (dump_file,
4468 "\nIF-THEN%s-JOIN block found, pass %d, test %d, then %d",
4469 (else_bb) ? "-ELSE" : "",
4470 pass, test_bb->index, then_bb->index);
4472 if (else_bb)
4473 fprintf (dump_file, ", else %d", else_bb->index);
4475 fprintf (dump_file, ", join %d\n", join_bb->index);
4478 /* If the conditional jump is more than just a conditional
4479 jump, then we cannot do if-conversion on this block. */
4480 jump = BB_END (test_bb);
4481 if (! onlyjump_p (jump))
4482 return FALSE;
4484 /* If this is not a standard conditional jump, we can't parse it. */
4485 cond = noce_get_condition (jump, &cond_earliest, then_else_reversed);
4486 if (!cond)
4487 return FALSE;
4489 /* We must be comparing objects whose modes imply the size. */
4490 if (GET_MODE (XEXP (cond, 0)) == BLKmode)
4491 return FALSE;
4493 /* Initialize an IF_INFO struct to pass around. */
4494 memset (&if_info, 0, sizeof if_info);
4495 if_info.test_bb = test_bb;
4496 if_info.then_bb = then_bb;
4497 if_info.else_bb = else_bb;
4498 if_info.join_bb = join_bb;
4499 if_info.cond = cond;
4500 rtx_insn *rev_cond_earliest;
4501 if_info.rev_cond = noce_get_condition (jump, &rev_cond_earliest,
4502 !then_else_reversed);
4503 gcc_assert (if_info.rev_cond == NULL_RTX
4504 || rev_cond_earliest == cond_earliest);
4505 if_info.cond_earliest = cond_earliest;
4506 if_info.jump = jump;
4507 if_info.then_else_reversed = then_else_reversed;
4508 if_info.speed_p = speed_p;
4509 if_info.max_seq_cost
4510 = targetm.max_noce_ifcvt_seq_cost (then_edge);
4511 /* We'll add in the cost of THEN_BB and ELSE_BB later, when we check
4512 that they are valid to transform. We can't easily get back to the insn
4513 for COND (and it may not exist if we had to canonicalize to get COND),
4514 and jump_insns are always given a cost of 1 by seq_cost, so treat
4515 both instructions as having cost COSTS_N_INSNS (1). */
4516 if_info.original_cost = COSTS_N_INSNS (2);
4519 /* Do the real work. */
4521 if (noce_process_if_block (&if_info))
4522 return TRUE;
4524 if (HAVE_conditional_move
4525 && cond_move_process_if_block (&if_info))
4526 return TRUE;
4528 return FALSE;
4532 /* Merge the blocks and mark for local life update. */
4534 static void
4535 merge_if_block (struct ce_if_block * ce_info)
4537 basic_block test_bb = ce_info->test_bb; /* last test block */
4538 basic_block then_bb = ce_info->then_bb; /* THEN */
4539 basic_block else_bb = ce_info->else_bb; /* ELSE or NULL */
4540 basic_block join_bb = ce_info->join_bb; /* join block */
4541 basic_block combo_bb;
4543 /* All block merging is done into the lower block numbers. */
4545 combo_bb = test_bb;
4546 df_set_bb_dirty (test_bb);
4548 /* Merge any basic blocks to handle && and || subtests. Each of
4549 the blocks are on the fallthru path from the predecessor block. */
4550 if (ce_info->num_multiple_test_blocks > 0)
4552 basic_block bb = test_bb;
4553 basic_block last_test_bb = ce_info->last_test_bb;
4554 basic_block fallthru = block_fallthru (bb);
4558 bb = fallthru;
4559 fallthru = block_fallthru (bb);
4560 merge_blocks (combo_bb, bb);
4561 num_true_changes++;
4563 while (bb != last_test_bb);
4566 /* Merge TEST block into THEN block. Normally the THEN block won't have a
4567 label, but it might if there were || tests. That label's count should be
4568 zero, and it normally should be removed. */
4570 if (then_bb)
4572 /* If THEN_BB has no successors, then there's a BARRIER after it.
4573 If COMBO_BB has more than one successor (THEN_BB), then that BARRIER
4574 is no longer needed, and in fact it is incorrect to leave it in
4575 the insn stream. */
4576 if (EDGE_COUNT (then_bb->succs) == 0
4577 && EDGE_COUNT (combo_bb->succs) > 1)
4579 rtx_insn *end = NEXT_INSN (BB_END (then_bb));
4580 while (end && NOTE_P (end) && !NOTE_INSN_BASIC_BLOCK_P (end))
4581 end = NEXT_INSN (end);
4583 if (end && BARRIER_P (end))
4584 delete_insn (end);
4586 merge_blocks (combo_bb, then_bb);
4587 num_true_changes++;
4590 /* The ELSE block, if it existed, had a label. That label count
4591 will almost always be zero, but odd things can happen when labels
4592 get their addresses taken. */
4593 if (else_bb)
4595 /* If ELSE_BB has no successors, then there's a BARRIER after it.
4596 If COMBO_BB has more than one successor (ELSE_BB), then that BARRIER
4597 is no longer needed, and in fact it is incorrect to leave it in
4598 the insn stream. */
4599 if (EDGE_COUNT (else_bb->succs) == 0
4600 && EDGE_COUNT (combo_bb->succs) > 1)
4602 rtx_insn *end = NEXT_INSN (BB_END (else_bb));
4603 while (end && NOTE_P (end) && !NOTE_INSN_BASIC_BLOCK_P (end))
4604 end = NEXT_INSN (end);
4606 if (end && BARRIER_P (end))
4607 delete_insn (end);
4609 merge_blocks (combo_bb, else_bb);
4610 num_true_changes++;
4613 /* If there was no join block reported, that means it was not adjacent
4614 to the others, and so we cannot merge them. */
4616 if (! join_bb)
4618 rtx_insn *last = BB_END (combo_bb);
4620 /* The outgoing edge for the current COMBO block should already
4621 be correct. Verify this. */
4622 if (EDGE_COUNT (combo_bb->succs) == 0)
4623 gcc_assert (find_reg_note (last, REG_NORETURN, NULL)
4624 || (NONJUMP_INSN_P (last)
4625 && GET_CODE (PATTERN (last)) == TRAP_IF
4626 && (TRAP_CONDITION (PATTERN (last))
4627 == const_true_rtx)));
4629 else
4630 /* There should still be something at the end of the THEN or ELSE
4631 blocks taking us to our final destination. */
4632 gcc_assert (JUMP_P (last)
4633 || (EDGE_SUCC (combo_bb, 0)->dest
4634 == EXIT_BLOCK_PTR_FOR_FN (cfun)
4635 && CALL_P (last)
4636 && SIBLING_CALL_P (last))
4637 || ((EDGE_SUCC (combo_bb, 0)->flags & EDGE_EH)
4638 && can_throw_internal (last)));
4641 /* The JOIN block may have had quite a number of other predecessors too.
4642 Since we've already merged the TEST, THEN and ELSE blocks, we should
4643 have only one remaining edge from our if-then-else diamond. If there
4644 is more than one remaining edge, it must come from elsewhere. There
4645 may be zero incoming edges if the THEN block didn't actually join
4646 back up (as with a call to a non-return function). */
4647 else if (EDGE_COUNT (join_bb->preds) < 2
4648 && join_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
4650 /* We can merge the JOIN cleanly and update the dataflow try
4651 again on this pass.*/
4652 merge_blocks (combo_bb, join_bb);
4653 num_true_changes++;
4655 else
4657 /* We cannot merge the JOIN. */
4659 /* The outgoing edge for the current COMBO block should already
4660 be correct. Verify this. */
4661 gcc_assert (single_succ_p (combo_bb)
4662 && single_succ (combo_bb) == join_bb);
4664 /* Remove the jump and cruft from the end of the COMBO block. */
4665 if (join_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
4666 tidy_fallthru_edge (single_succ_edge (combo_bb));
4669 num_updated_if_blocks++;
4672 /* Find a block ending in a simple IF condition and try to transform it
4673 in some way. When converting a multi-block condition, put the new code
4674 in the first such block and delete the rest. Return a pointer to this
4675 first block if some transformation was done. Return NULL otherwise. */
4677 static basic_block
4678 find_if_header (basic_block test_bb, int pass)
4680 ce_if_block ce_info;
4681 edge then_edge;
4682 edge else_edge;
4684 /* The kind of block we're looking for has exactly two successors. */
4685 if (EDGE_COUNT (test_bb->succs) != 2)
4686 return NULL;
4688 then_edge = EDGE_SUCC (test_bb, 0);
4689 else_edge = EDGE_SUCC (test_bb, 1);
4691 if (df_get_bb_dirty (then_edge->dest))
4692 return NULL;
4693 if (df_get_bb_dirty (else_edge->dest))
4694 return NULL;
4696 /* Neither edge should be abnormal. */
4697 if ((then_edge->flags & EDGE_COMPLEX)
4698 || (else_edge->flags & EDGE_COMPLEX))
4699 return NULL;
4701 /* Nor exit the loop. */
4702 if ((then_edge->flags & EDGE_LOOP_EXIT)
4703 || (else_edge->flags & EDGE_LOOP_EXIT))
4704 return NULL;
4706 /* The THEN edge is canonically the one that falls through. */
4707 if (then_edge->flags & EDGE_FALLTHRU)
4709 else if (else_edge->flags & EDGE_FALLTHRU)
4710 std::swap (then_edge, else_edge);
4711 else
4712 /* Otherwise this must be a multiway branch of some sort. */
4713 return NULL;
4715 memset (&ce_info, 0, sizeof (ce_info));
4716 ce_info.test_bb = test_bb;
4717 ce_info.then_bb = then_edge->dest;
4718 ce_info.else_bb = else_edge->dest;
4719 ce_info.pass = pass;
4721 #ifdef IFCVT_MACHDEP_INIT
4722 IFCVT_MACHDEP_INIT (&ce_info);
4723 #endif
4725 if (!reload_completed
4726 && noce_find_if_block (test_bb, then_edge, else_edge, pass))
4727 goto success;
4729 if (reload_completed
4730 && targetm.have_conditional_execution ()
4731 && cond_exec_find_if_block (&ce_info))
4732 goto success;
4734 if (targetm.have_trap ()
4735 && optab_handler (ctrap_optab, word_mode) != CODE_FOR_nothing
4736 && find_cond_trap (test_bb, then_edge, else_edge))
4737 goto success;
4739 if (dom_info_state (CDI_POST_DOMINATORS) >= DOM_NO_FAST_QUERY
4740 && (reload_completed || !targetm.have_conditional_execution ()))
4742 if (find_if_case_1 (test_bb, then_edge, else_edge))
4743 goto success;
4744 if (find_if_case_2 (test_bb, then_edge, else_edge))
4745 goto success;
4748 return NULL;
4750 success:
4751 if (dump_file)
4752 fprintf (dump_file, "Conversion succeeded on pass %d.\n", pass);
4753 /* Set this so we continue looking. */
4754 cond_exec_changed_p = TRUE;
4755 return ce_info.test_bb;
4758 /* Return true if a block has two edges, one of which falls through to the next
4759 block, and the other jumps to a specific block, so that we can tell if the
4760 block is part of an && test or an || test. Returns either -1 or the number
4761 of non-note, non-jump, non-USE/CLOBBER insns in the block. */
4763 static int
4764 block_jumps_and_fallthru_p (basic_block cur_bb, basic_block target_bb)
4766 edge cur_edge;
4767 int fallthru_p = FALSE;
4768 int jump_p = FALSE;
4769 rtx_insn *insn;
4770 rtx_insn *end;
4771 int n_insns = 0;
4772 edge_iterator ei;
4774 if (!cur_bb || !target_bb)
4775 return -1;
4777 /* If no edges, obviously it doesn't jump or fallthru. */
4778 if (EDGE_COUNT (cur_bb->succs) == 0)
4779 return FALSE;
4781 FOR_EACH_EDGE (cur_edge, ei, cur_bb->succs)
4783 if (cur_edge->flags & EDGE_COMPLEX)
4784 /* Anything complex isn't what we want. */
4785 return -1;
4787 else if (cur_edge->flags & EDGE_FALLTHRU)
4788 fallthru_p = TRUE;
4790 else if (cur_edge->dest == target_bb)
4791 jump_p = TRUE;
4793 else
4794 return -1;
4797 if ((jump_p & fallthru_p) == 0)
4798 return -1;
4800 /* Don't allow calls in the block, since this is used to group && and ||
4801 together for conditional execution support. ??? we should support
4802 conditional execution support across calls for IA-64 some day, but
4803 for now it makes the code simpler. */
4804 end = BB_END (cur_bb);
4805 insn = BB_HEAD (cur_bb);
4807 while (insn != NULL_RTX)
4809 if (CALL_P (insn))
4810 return -1;
4812 if (INSN_P (insn)
4813 && !JUMP_P (insn)
4814 && !DEBUG_INSN_P (insn)
4815 && GET_CODE (PATTERN (insn)) != USE
4816 && GET_CODE (PATTERN (insn)) != CLOBBER)
4817 n_insns++;
4819 if (insn == end)
4820 break;
4822 insn = NEXT_INSN (insn);
4825 return n_insns;
4828 /* Determine if a given basic block heads a simple IF-THEN or IF-THEN-ELSE
4829 block. If so, we'll try to convert the insns to not require the branch.
4830 Return TRUE if we were successful at converting the block. */
4832 static int
4833 cond_exec_find_if_block (struct ce_if_block * ce_info)
4835 basic_block test_bb = ce_info->test_bb;
4836 basic_block then_bb = ce_info->then_bb;
4837 basic_block else_bb = ce_info->else_bb;
4838 basic_block join_bb = NULL_BLOCK;
4839 edge cur_edge;
4840 basic_block next;
4841 edge_iterator ei;
4843 ce_info->last_test_bb = test_bb;
4845 /* We only ever should get here after reload,
4846 and if we have conditional execution. */
4847 gcc_assert (reload_completed && targetm.have_conditional_execution ());
4849 /* Discover if any fall through predecessors of the current test basic block
4850 were && tests (which jump to the else block) or || tests (which jump to
4851 the then block). */
4852 if (single_pred_p (test_bb)
4853 && single_pred_edge (test_bb)->flags == EDGE_FALLTHRU)
4855 basic_block bb = single_pred (test_bb);
4856 basic_block target_bb;
4857 int max_insns = MAX_CONDITIONAL_EXECUTE;
4858 int n_insns;
4860 /* Determine if the preceding block is an && or || block. */
4861 if ((n_insns = block_jumps_and_fallthru_p (bb, else_bb)) >= 0)
4863 ce_info->and_and_p = TRUE;
4864 target_bb = else_bb;
4866 else if ((n_insns = block_jumps_and_fallthru_p (bb, then_bb)) >= 0)
4868 ce_info->and_and_p = FALSE;
4869 target_bb = then_bb;
4871 else
4872 target_bb = NULL_BLOCK;
4874 if (target_bb && n_insns <= max_insns)
4876 int total_insns = 0;
4877 int blocks = 0;
4879 ce_info->last_test_bb = test_bb;
4881 /* Found at least one && or || block, look for more. */
4884 ce_info->test_bb = test_bb = bb;
4885 total_insns += n_insns;
4886 blocks++;
4888 if (!single_pred_p (bb))
4889 break;
4891 bb = single_pred (bb);
4892 n_insns = block_jumps_and_fallthru_p (bb, target_bb);
4894 while (n_insns >= 0 && (total_insns + n_insns) <= max_insns);
4896 ce_info->num_multiple_test_blocks = blocks;
4897 ce_info->num_multiple_test_insns = total_insns;
4899 if (ce_info->and_and_p)
4900 ce_info->num_and_and_blocks = blocks;
4901 else
4902 ce_info->num_or_or_blocks = blocks;
4906 /* The THEN block of an IF-THEN combo must have exactly one predecessor,
4907 other than any || blocks which jump to the THEN block. */
4908 if ((EDGE_COUNT (then_bb->preds) - ce_info->num_or_or_blocks) != 1)
4909 return FALSE;
4911 /* The edges of the THEN and ELSE blocks cannot have complex edges. */
4912 FOR_EACH_EDGE (cur_edge, ei, then_bb->preds)
4914 if (cur_edge->flags & EDGE_COMPLEX)
4915 return FALSE;
4918 FOR_EACH_EDGE (cur_edge, ei, else_bb->preds)
4920 if (cur_edge->flags & EDGE_COMPLEX)
4921 return FALSE;
4924 /* The THEN block of an IF-THEN combo must have zero or one successors. */
4925 if (EDGE_COUNT (then_bb->succs) > 0
4926 && (!single_succ_p (then_bb)
4927 || (single_succ_edge (then_bb)->flags & EDGE_COMPLEX)
4928 || (epilogue_completed
4929 && tablejump_p (BB_END (then_bb), NULL, NULL))))
4930 return FALSE;
4932 /* If the THEN block has no successors, conditional execution can still
4933 make a conditional call. Don't do this unless the ELSE block has
4934 only one incoming edge -- the CFG manipulation is too ugly otherwise.
4935 Check for the last insn of the THEN block being an indirect jump, which
4936 is listed as not having any successors, but confuses the rest of the CE
4937 code processing. ??? we should fix this in the future. */
4938 if (EDGE_COUNT (then_bb->succs) == 0)
4940 if (single_pred_p (else_bb) && else_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
4942 rtx_insn *last_insn = BB_END (then_bb);
4944 while (last_insn
4945 && NOTE_P (last_insn)
4946 && last_insn != BB_HEAD (then_bb))
4947 last_insn = PREV_INSN (last_insn);
4949 if (last_insn
4950 && JUMP_P (last_insn)
4951 && ! simplejump_p (last_insn))
4952 return FALSE;
4954 join_bb = else_bb;
4955 else_bb = NULL_BLOCK;
4957 else
4958 return FALSE;
4961 /* If the THEN block's successor is the other edge out of the TEST block,
4962 then we have an IF-THEN combo without an ELSE. */
4963 else if (single_succ (then_bb) == else_bb)
4965 join_bb = else_bb;
4966 else_bb = NULL_BLOCK;
4969 /* If the THEN and ELSE block meet in a subsequent block, and the ELSE
4970 has exactly one predecessor and one successor, and the outgoing edge
4971 is not complex, then we have an IF-THEN-ELSE combo. */
4972 else if (single_succ_p (else_bb)
4973 && single_succ (then_bb) == single_succ (else_bb)
4974 && single_pred_p (else_bb)
4975 && !(single_succ_edge (else_bb)->flags & EDGE_COMPLEX)
4976 && !(epilogue_completed
4977 && tablejump_p (BB_END (else_bb), NULL, NULL)))
4978 join_bb = single_succ (else_bb);
4980 /* Otherwise it is not an IF-THEN or IF-THEN-ELSE combination. */
4981 else
4982 return FALSE;
4984 num_possible_if_blocks++;
4986 if (dump_file)
4988 fprintf (dump_file,
4989 "\nIF-THEN%s block found, pass %d, start block %d "
4990 "[insn %d], then %d [%d]",
4991 (else_bb) ? "-ELSE" : "",
4992 ce_info->pass,
4993 test_bb->index,
4994 BB_HEAD (test_bb) ? (int)INSN_UID (BB_HEAD (test_bb)) : -1,
4995 then_bb->index,
4996 BB_HEAD (then_bb) ? (int)INSN_UID (BB_HEAD (then_bb)) : -1);
4998 if (else_bb)
4999 fprintf (dump_file, ", else %d [%d]",
5000 else_bb->index,
5001 BB_HEAD (else_bb) ? (int)INSN_UID (BB_HEAD (else_bb)) : -1);
5003 fprintf (dump_file, ", join %d [%d]",
5004 join_bb->index,
5005 BB_HEAD (join_bb) ? (int)INSN_UID (BB_HEAD (join_bb)) : -1);
5007 if (ce_info->num_multiple_test_blocks > 0)
5008 fprintf (dump_file, ", %d %s block%s last test %d [%d]",
5009 ce_info->num_multiple_test_blocks,
5010 (ce_info->and_and_p) ? "&&" : "||",
5011 (ce_info->num_multiple_test_blocks == 1) ? "" : "s",
5012 ce_info->last_test_bb->index,
5013 ((BB_HEAD (ce_info->last_test_bb))
5014 ? (int)INSN_UID (BB_HEAD (ce_info->last_test_bb))
5015 : -1));
5017 fputc ('\n', dump_file);
5020 /* Make sure IF, THEN, and ELSE, blocks are adjacent. Actually, we get the
5021 first condition for free, since we've already asserted that there's a
5022 fallthru edge from IF to THEN. Likewise for the && and || blocks, since
5023 we checked the FALLTHRU flag, those are already adjacent to the last IF
5024 block. */
5025 /* ??? As an enhancement, move the ELSE block. Have to deal with
5026 BLOCK notes, if by no other means than backing out the merge if they
5027 exist. Sticky enough I don't want to think about it now. */
5028 next = then_bb;
5029 if (else_bb && (next = next->next_bb) != else_bb)
5030 return FALSE;
5031 if ((next = next->next_bb) != join_bb
5032 && join_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
5034 if (else_bb)
5035 join_bb = NULL;
5036 else
5037 return FALSE;
5040 /* Do the real work. */
5042 ce_info->else_bb = else_bb;
5043 ce_info->join_bb = join_bb;
5045 /* If we have && and || tests, try to first handle combining the && and ||
5046 tests into the conditional code, and if that fails, go back and handle
5047 it without the && and ||, which at present handles the && case if there
5048 was no ELSE block. */
5049 if (cond_exec_process_if_block (ce_info, TRUE))
5050 return TRUE;
5052 if (ce_info->num_multiple_test_blocks)
5054 cancel_changes (0);
5056 if (cond_exec_process_if_block (ce_info, FALSE))
5057 return TRUE;
5060 return FALSE;
5063 /* Convert a branch over a trap, or a branch
5064 to a trap, into a conditional trap. */
5066 static int
5067 find_cond_trap (basic_block test_bb, edge then_edge, edge else_edge)
5069 basic_block then_bb = then_edge->dest;
5070 basic_block else_bb = else_edge->dest;
5071 basic_block other_bb, trap_bb;
5072 rtx_insn *trap, *jump;
5073 rtx cond;
5074 rtx_insn *cond_earliest;
5076 /* Locate the block with the trap instruction. */
5077 /* ??? While we look for no successors, we really ought to allow
5078 EH successors. Need to fix merge_if_block for that to work. */
5079 if ((trap = block_has_only_trap (then_bb)) != NULL)
5080 trap_bb = then_bb, other_bb = else_bb;
5081 else if ((trap = block_has_only_trap (else_bb)) != NULL)
5082 trap_bb = else_bb, other_bb = then_bb;
5083 else
5084 return FALSE;
5086 if (dump_file)
5088 fprintf (dump_file, "\nTRAP-IF block found, start %d, trap %d\n",
5089 test_bb->index, trap_bb->index);
5092 /* If this is not a standard conditional jump, we can't parse it. */
5093 jump = BB_END (test_bb);
5094 cond = noce_get_condition (jump, &cond_earliest, then_bb == trap_bb);
5095 if (! cond)
5096 return FALSE;
5098 /* If the conditional jump is more than just a conditional jump, then
5099 we cannot do if-conversion on this block. Give up for returnjump_p,
5100 changing a conditional return followed by unconditional trap for
5101 conditional trap followed by unconditional return is likely not
5102 beneficial and harder to handle. */
5103 if (! onlyjump_p (jump) || returnjump_p (jump))
5104 return FALSE;
5106 /* We must be comparing objects whose modes imply the size. */
5107 if (GET_MODE (XEXP (cond, 0)) == BLKmode)
5108 return FALSE;
5110 /* Attempt to generate the conditional trap. */
5111 rtx_insn *seq = gen_cond_trap (GET_CODE (cond), copy_rtx (XEXP (cond, 0)),
5112 copy_rtx (XEXP (cond, 1)),
5113 TRAP_CODE (PATTERN (trap)));
5114 if (seq == NULL)
5115 return FALSE;
5117 /* If that results in an invalid insn, back out. */
5118 for (rtx_insn *x = seq; x; x = NEXT_INSN (x))
5119 if (reload_completed
5120 ? !valid_insn_p (x)
5121 : recog_memoized (x) < 0)
5122 return FALSE;
5124 /* Emit the new insns before cond_earliest. */
5125 emit_insn_before_setloc (seq, cond_earliest, INSN_LOCATION (trap));
5127 /* Delete the trap block if possible. */
5128 remove_edge (trap_bb == then_bb ? then_edge : else_edge);
5129 df_set_bb_dirty (test_bb);
5130 df_set_bb_dirty (then_bb);
5131 df_set_bb_dirty (else_bb);
5133 if (EDGE_COUNT (trap_bb->preds) == 0)
5135 delete_basic_block (trap_bb);
5136 num_true_changes++;
5139 /* Wire together the blocks again. */
5140 if (current_ir_type () == IR_RTL_CFGLAYOUT)
5141 single_succ_edge (test_bb)->flags |= EDGE_FALLTHRU;
5142 else if (trap_bb == then_bb)
5144 rtx lab = JUMP_LABEL (jump);
5145 rtx_insn *seq = targetm.gen_jump (lab);
5146 rtx_jump_insn *newjump = emit_jump_insn_after (seq, jump);
5147 LABEL_NUSES (lab) += 1;
5148 JUMP_LABEL (newjump) = lab;
5149 emit_barrier_after (newjump);
5151 delete_insn (jump);
5153 if (can_merge_blocks_p (test_bb, other_bb))
5155 merge_blocks (test_bb, other_bb);
5156 num_true_changes++;
5159 num_updated_if_blocks++;
5160 return TRUE;
5163 /* Subroutine of find_cond_trap: if BB contains only a trap insn,
5164 return it. */
5166 static rtx_insn *
5167 block_has_only_trap (basic_block bb)
5169 rtx_insn *trap;
5171 /* We're not the exit block. */
5172 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
5173 return NULL;
5175 /* The block must have no successors. */
5176 if (EDGE_COUNT (bb->succs) > 0)
5177 return NULL;
5179 /* The only instruction in the THEN block must be the trap. */
5180 trap = first_active_insn (bb);
5181 if (! (trap == BB_END (bb)
5182 && GET_CODE (PATTERN (trap)) == TRAP_IF
5183 && TRAP_CONDITION (PATTERN (trap)) == const_true_rtx))
5184 return NULL;
5186 return trap;
5189 /* Look for IF-THEN-ELSE cases in which one of THEN or ELSE is
5190 transformable, but not necessarily the other. There need be no
5191 JOIN block.
5193 Return TRUE if we were successful at converting the block.
5195 Cases we'd like to look at:
5198 if (test) goto over; // x not live
5199 x = a;
5200 goto label;
5201 over:
5203 becomes
5205 x = a;
5206 if (! test) goto label;
5209 if (test) goto E; // x not live
5210 x = big();
5211 goto L;
5213 x = b;
5214 goto M;
5216 becomes
5218 x = b;
5219 if (test) goto M;
5220 x = big();
5221 goto L;
5223 (3) // This one's really only interesting for targets that can do
5224 // multiway branching, e.g. IA-64 BBB bundles. For other targets
5225 // it results in multiple branches on a cache line, which often
5226 // does not sit well with predictors.
5228 if (test1) goto E; // predicted not taken
5229 x = a;
5230 if (test2) goto F;
5233 x = b;
5236 becomes
5238 x = a;
5239 if (test1) goto E;
5240 if (test2) goto F;
5242 Notes:
5244 (A) Don't do (2) if the branch is predicted against the block we're
5245 eliminating. Do it anyway if we can eliminate a branch; this requires
5246 that the sole successor of the eliminated block postdominate the other
5247 side of the if.
5249 (B) With CE, on (3) we can steal from both sides of the if, creating
5251 if (test1) x = a;
5252 if (!test1) x = b;
5253 if (test1) goto J;
5254 if (test2) goto F;
5258 Again, this is most useful if J postdominates.
5260 (C) CE substitutes for helpful life information.
5262 (D) These heuristics need a lot of work. */
5264 /* Tests for case 1 above. */
5266 static int
5267 find_if_case_1 (basic_block test_bb, edge then_edge, edge else_edge)
5269 basic_block then_bb = then_edge->dest;
5270 basic_block else_bb = else_edge->dest;
5271 basic_block new_bb;
5272 int then_bb_index;
5273 profile_probability then_prob;
5274 rtx else_target = NULL_RTX;
5276 /* If we are partitioning hot/cold basic blocks, we don't want to
5277 mess up unconditional or indirect jumps that cross between hot
5278 and cold sections.
5280 Basic block partitioning may result in some jumps that appear to
5281 be optimizable (or blocks that appear to be mergeable), but which really
5282 must be left untouched (they are required to make it safely across
5283 partition boundaries). See the comments at the top of
5284 bb-reorder.cc:partition_hot_cold_basic_blocks for complete details. */
5286 if ((BB_END (then_bb)
5287 && JUMP_P (BB_END (then_bb))
5288 && CROSSING_JUMP_P (BB_END (then_bb)))
5289 || (JUMP_P (BB_END (test_bb))
5290 && CROSSING_JUMP_P (BB_END (test_bb)))
5291 || (BB_END (else_bb)
5292 && JUMP_P (BB_END (else_bb))
5293 && CROSSING_JUMP_P (BB_END (else_bb))))
5294 return FALSE;
5296 /* Verify test_bb ends in a conditional jump with no other side-effects. */
5297 if (!onlyjump_p (BB_END (test_bb)))
5298 return FALSE;
5300 /* THEN has one successor. */
5301 if (!single_succ_p (then_bb))
5302 return FALSE;
5304 /* THEN does not fall through, but is not strange either. */
5305 if (single_succ_edge (then_bb)->flags & (EDGE_COMPLEX | EDGE_FALLTHRU))
5306 return FALSE;
5308 /* THEN has one predecessor. */
5309 if (!single_pred_p (then_bb))
5310 return FALSE;
5312 /* THEN must do something. */
5313 if (forwarder_block_p (then_bb))
5314 return FALSE;
5316 num_possible_if_blocks++;
5317 if (dump_file)
5318 fprintf (dump_file,
5319 "\nIF-CASE-1 found, start %d, then %d\n",
5320 test_bb->index, then_bb->index);
5322 then_prob = then_edge->probability.invert ();
5324 /* We're speculating from the THEN path, we want to make sure the cost
5325 of speculation is within reason. */
5326 if (! cheap_bb_rtx_cost_p (then_bb, then_prob,
5327 COSTS_N_INSNS (BRANCH_COST (optimize_bb_for_speed_p (then_edge->src),
5328 predictable_edge_p (then_edge)))))
5329 return FALSE;
5331 if (else_bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
5333 rtx_insn *jump = BB_END (else_edge->src);
5334 gcc_assert (JUMP_P (jump));
5335 else_target = JUMP_LABEL (jump);
5338 /* Registers set are dead, or are predicable. */
5339 if (! dead_or_predicable (test_bb, then_bb, else_bb,
5340 single_succ_edge (then_bb), 1))
5341 return FALSE;
5343 /* Conversion went ok, including moving the insns and fixing up the
5344 jump. Adjust the CFG to match. */
5346 /* We can avoid creating a new basic block if then_bb is immediately
5347 followed by else_bb, i.e. deleting then_bb allows test_bb to fall
5348 through to else_bb. */
5350 if (then_bb->next_bb == else_bb
5351 && then_bb->prev_bb == test_bb
5352 && else_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
5354 redirect_edge_succ (FALLTHRU_EDGE (test_bb), else_bb);
5355 new_bb = 0;
5357 else if (else_bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
5358 new_bb = force_nonfallthru_and_redirect (FALLTHRU_EDGE (test_bb),
5359 else_bb, else_target);
5360 else
5361 new_bb = redirect_edge_and_branch_force (FALLTHRU_EDGE (test_bb),
5362 else_bb);
5364 df_set_bb_dirty (test_bb);
5365 df_set_bb_dirty (else_bb);
5367 then_bb_index = then_bb->index;
5368 delete_basic_block (then_bb);
5370 /* Make rest of code believe that the newly created block is the THEN_BB
5371 block we removed. */
5372 if (new_bb)
5374 df_bb_replace (then_bb_index, new_bb);
5375 /* This should have been done above via force_nonfallthru_and_redirect
5376 (possibly called from redirect_edge_and_branch_force). */
5377 gcc_checking_assert (BB_PARTITION (new_bb) == BB_PARTITION (test_bb));
5380 num_true_changes++;
5381 num_updated_if_blocks++;
5382 return TRUE;
5385 /* Test for case 2 above. */
5387 static int
5388 find_if_case_2 (basic_block test_bb, edge then_edge, edge else_edge)
5390 basic_block then_bb = then_edge->dest;
5391 basic_block else_bb = else_edge->dest;
5392 edge else_succ;
5393 profile_probability then_prob, else_prob;
5395 /* We do not want to speculate (empty) loop latches. */
5396 if (current_loops
5397 && else_bb->loop_father->latch == else_bb)
5398 return FALSE;
5400 /* If we are partitioning hot/cold basic blocks, we don't want to
5401 mess up unconditional or indirect jumps that cross between hot
5402 and cold sections.
5404 Basic block partitioning may result in some jumps that appear to
5405 be optimizable (or blocks that appear to be mergeable), but which really
5406 must be left untouched (they are required to make it safely across
5407 partition boundaries). See the comments at the top of
5408 bb-reorder.cc:partition_hot_cold_basic_blocks for complete details. */
5410 if ((BB_END (then_bb)
5411 && JUMP_P (BB_END (then_bb))
5412 && CROSSING_JUMP_P (BB_END (then_bb)))
5413 || (JUMP_P (BB_END (test_bb))
5414 && CROSSING_JUMP_P (BB_END (test_bb)))
5415 || (BB_END (else_bb)
5416 && JUMP_P (BB_END (else_bb))
5417 && CROSSING_JUMP_P (BB_END (else_bb))))
5418 return FALSE;
5420 /* Verify test_bb ends in a conditional jump with no other side-effects. */
5421 if (!onlyjump_p (BB_END (test_bb)))
5422 return FALSE;
5424 /* ELSE has one successor. */
5425 if (!single_succ_p (else_bb))
5426 return FALSE;
5427 else
5428 else_succ = single_succ_edge (else_bb);
5430 /* ELSE outgoing edge is not complex. */
5431 if (else_succ->flags & EDGE_COMPLEX)
5432 return FALSE;
5434 /* ELSE has one predecessor. */
5435 if (!single_pred_p (else_bb))
5436 return FALSE;
5438 /* THEN is not EXIT. */
5439 if (then_bb->index < NUM_FIXED_BLOCKS)
5440 return FALSE;
5442 else_prob = else_edge->probability;
5443 then_prob = else_prob.invert ();
5445 /* ELSE is predicted or SUCC(ELSE) postdominates THEN. */
5446 if (else_prob > then_prob)
5448 else if (else_succ->dest->index < NUM_FIXED_BLOCKS
5449 || dominated_by_p (CDI_POST_DOMINATORS, then_bb,
5450 else_succ->dest))
5452 else
5453 return FALSE;
5455 num_possible_if_blocks++;
5456 if (dump_file)
5457 fprintf (dump_file,
5458 "\nIF-CASE-2 found, start %d, else %d\n",
5459 test_bb->index, else_bb->index);
5461 /* We're speculating from the ELSE path, we want to make sure the cost
5462 of speculation is within reason. */
5463 if (! cheap_bb_rtx_cost_p (else_bb, else_prob,
5464 COSTS_N_INSNS (BRANCH_COST (optimize_bb_for_speed_p (else_edge->src),
5465 predictable_edge_p (else_edge)))))
5466 return FALSE;
5468 /* Registers set are dead, or are predicable. */
5469 if (! dead_or_predicable (test_bb, else_bb, then_bb, else_succ, 0))
5470 return FALSE;
5472 /* Conversion went ok, including moving the insns and fixing up the
5473 jump. Adjust the CFG to match. */
5475 df_set_bb_dirty (test_bb);
5476 df_set_bb_dirty (then_bb);
5477 delete_basic_block (else_bb);
5479 num_true_changes++;
5480 num_updated_if_blocks++;
5482 /* ??? We may now fallthru from one of THEN's successors into a join
5483 block. Rerun cleanup_cfg? Examine things manually? Wait? */
5485 return TRUE;
5488 /* Used by the code above to perform the actual rtl transformations.
5489 Return TRUE if successful.
5491 TEST_BB is the block containing the conditional branch. MERGE_BB
5492 is the block containing the code to manipulate. DEST_EDGE is an
5493 edge representing a jump to the join block; after the conversion,
5494 TEST_BB should be branching to its destination.
5495 REVERSEP is true if the sense of the branch should be reversed. */
5497 static int
5498 dead_or_predicable (basic_block test_bb, basic_block merge_bb,
5499 basic_block other_bb, edge dest_edge, int reversep)
5501 basic_block new_dest = dest_edge->dest;
5502 rtx_insn *head, *end, *jump;
5503 rtx_insn *earliest = NULL;
5504 rtx old_dest;
5505 bitmap merge_set = NULL;
5506 /* Number of pending changes. */
5507 int n_validated_changes = 0;
5508 rtx new_dest_label = NULL_RTX;
5510 jump = BB_END (test_bb);
5512 /* Find the extent of the real code in the merge block. */
5513 head = BB_HEAD (merge_bb);
5514 end = BB_END (merge_bb);
5516 while (DEBUG_INSN_P (end) && end != head)
5517 end = PREV_INSN (end);
5519 /* If merge_bb ends with a tablejump, predicating/moving insn's
5520 into test_bb and then deleting merge_bb will result in the jumptable
5521 that follows merge_bb being removed along with merge_bb and then we
5522 get an unresolved reference to the jumptable. */
5523 if (tablejump_p (end, NULL, NULL))
5524 return FALSE;
5526 if (LABEL_P (head))
5527 head = NEXT_INSN (head);
5528 while (DEBUG_INSN_P (head) && head != end)
5529 head = NEXT_INSN (head);
5530 if (NOTE_P (head))
5532 if (head == end)
5534 head = end = NULL;
5535 goto no_body;
5537 head = NEXT_INSN (head);
5538 while (DEBUG_INSN_P (head) && head != end)
5539 head = NEXT_INSN (head);
5542 if (JUMP_P (end))
5544 if (!onlyjump_p (end))
5545 return FALSE;
5546 if (head == end)
5548 head = end = NULL;
5549 goto no_body;
5551 end = PREV_INSN (end);
5552 while (DEBUG_INSN_P (end) && end != head)
5553 end = PREV_INSN (end);
5556 /* Don't move frame-related insn across the conditional branch. This
5557 can lead to one of the paths of the branch having wrong unwind info. */
5558 if (epilogue_completed)
5560 rtx_insn *insn = head;
5561 while (1)
5563 if (INSN_P (insn) && RTX_FRAME_RELATED_P (insn))
5564 return FALSE;
5565 if (insn == end)
5566 break;
5567 insn = NEXT_INSN (insn);
5571 /* Disable handling dead code by conditional execution if the machine needs
5572 to do anything funny with the tests, etc. */
5573 #ifndef IFCVT_MODIFY_TESTS
5574 if (targetm.have_conditional_execution ())
5576 /* In the conditional execution case, we have things easy. We know
5577 the condition is reversible. We don't have to check life info
5578 because we're going to conditionally execute the code anyway.
5579 All that's left is making sure the insns involved can actually
5580 be predicated. */
5582 rtx cond;
5584 /* If the conditional jump is more than just a conditional jump,
5585 then we cannot do conditional execution conversion on this block. */
5586 if (!onlyjump_p (jump))
5587 goto nce;
5589 cond = cond_exec_get_condition (jump);
5590 if (! cond)
5591 goto nce;
5593 rtx note = find_reg_note (jump, REG_BR_PROB, NULL_RTX);
5594 profile_probability prob_val
5595 = (note ? profile_probability::from_reg_br_prob_note (XINT (note, 0))
5596 : profile_probability::uninitialized ());
5598 if (reversep)
5600 enum rtx_code rev = reversed_comparison_code (cond, jump);
5601 if (rev == UNKNOWN)
5602 return FALSE;
5603 cond = gen_rtx_fmt_ee (rev, GET_MODE (cond), XEXP (cond, 0),
5604 XEXP (cond, 1));
5605 prob_val = prob_val.invert ();
5608 if (cond_exec_process_insns (NULL, head, end, cond, prob_val, 0)
5609 && verify_changes (0))
5610 n_validated_changes = num_validated_changes ();
5611 else
5612 cancel_changes (0);
5614 earliest = jump;
5616 nce:
5617 #endif
5619 /* If we allocated new pseudos (e.g. in the conditional move
5620 expander called from noce_emit_cmove), we must resize the
5621 array first. */
5622 if (max_regno < max_reg_num ())
5623 max_regno = max_reg_num ();
5625 /* Try the NCE path if the CE path did not result in any changes. */
5626 if (n_validated_changes == 0)
5628 rtx cond;
5629 rtx_insn *insn;
5630 regset live;
5631 bool success;
5633 /* In the non-conditional execution case, we have to verify that there
5634 are no trapping operations, no calls, no references to memory, and
5635 that any registers modified are dead at the branch site. */
5637 if (!any_condjump_p (jump))
5638 return FALSE;
5640 /* Find the extent of the conditional. */
5641 cond = noce_get_condition (jump, &earliest, false);
5642 if (!cond)
5643 return FALSE;
5645 live = BITMAP_ALLOC (&reg_obstack);
5646 simulate_backwards_to_point (merge_bb, live, end);
5647 success = can_move_insns_across (head, end, earliest, jump,
5648 merge_bb, live,
5649 df_get_live_in (other_bb), NULL);
5650 BITMAP_FREE (live);
5651 if (!success)
5652 return FALSE;
5654 /* Collect the set of registers set in MERGE_BB. */
5655 merge_set = BITMAP_ALLOC (&reg_obstack);
5657 FOR_BB_INSNS (merge_bb, insn)
5658 if (NONDEBUG_INSN_P (insn))
5659 df_simulate_find_defs (insn, merge_set);
5661 /* If shrink-wrapping, disable this optimization when test_bb is
5662 the first basic block and merge_bb exits. The idea is to not
5663 move code setting up a return register as that may clobber a
5664 register used to pass function parameters, which then must be
5665 saved in caller-saved regs. A caller-saved reg requires the
5666 prologue, killing a shrink-wrap opportunity. */
5667 if ((SHRINK_WRAPPING_ENABLED && !epilogue_completed)
5668 && ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb == test_bb
5669 && single_succ_p (new_dest)
5670 && single_succ (new_dest) == EXIT_BLOCK_PTR_FOR_FN (cfun)
5671 && bitmap_intersect_p (df_get_live_in (new_dest), merge_set))
5673 regset return_regs;
5674 unsigned int i;
5676 return_regs = BITMAP_ALLOC (&reg_obstack);
5678 /* Start off with the intersection of regs used to pass
5679 params and regs used to return values. */
5680 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
5681 if (FUNCTION_ARG_REGNO_P (i)
5682 && targetm.calls.function_value_regno_p (i))
5683 bitmap_set_bit (return_regs, INCOMING_REGNO (i));
5685 bitmap_and_into (return_regs,
5686 df_get_live_out (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
5687 bitmap_and_into (return_regs,
5688 df_get_live_in (EXIT_BLOCK_PTR_FOR_FN (cfun)));
5689 if (!bitmap_empty_p (return_regs))
5691 FOR_BB_INSNS_REVERSE (new_dest, insn)
5692 if (NONDEBUG_INSN_P (insn))
5694 df_ref def;
5696 /* If this insn sets any reg in return_regs, add all
5697 reg uses to the set of regs we're interested in. */
5698 FOR_EACH_INSN_DEF (def, insn)
5699 if (bitmap_bit_p (return_regs, DF_REF_REGNO (def)))
5701 df_simulate_uses (insn, return_regs);
5702 break;
5705 if (bitmap_intersect_p (merge_set, return_regs))
5707 BITMAP_FREE (return_regs);
5708 BITMAP_FREE (merge_set);
5709 return FALSE;
5712 BITMAP_FREE (return_regs);
5716 no_body:
5717 /* We don't want to use normal invert_jump or redirect_jump because
5718 we don't want to delete_insn called. Also, we want to do our own
5719 change group management. */
5721 old_dest = JUMP_LABEL (jump);
5722 if (other_bb != new_dest)
5724 if (!any_condjump_p (jump))
5725 goto cancel;
5727 if (JUMP_P (BB_END (dest_edge->src)))
5728 new_dest_label = JUMP_LABEL (BB_END (dest_edge->src));
5729 else if (new_dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
5730 new_dest_label = ret_rtx;
5731 else
5732 new_dest_label = block_label (new_dest);
5734 rtx_jump_insn *jump_insn = as_a <rtx_jump_insn *> (jump);
5735 if (reversep
5736 ? ! invert_jump_1 (jump_insn, new_dest_label)
5737 : ! redirect_jump_1 (jump_insn, new_dest_label))
5738 goto cancel;
5741 if (verify_changes (n_validated_changes))
5742 confirm_change_group ();
5743 else
5744 goto cancel;
5746 if (other_bb != new_dest)
5748 redirect_jump_2 (as_a <rtx_jump_insn *> (jump), old_dest, new_dest_label,
5749 0, reversep);
5751 redirect_edge_succ (BRANCH_EDGE (test_bb), new_dest);
5752 if (reversep)
5754 std::swap (BRANCH_EDGE (test_bb)->probability,
5755 FALLTHRU_EDGE (test_bb)->probability);
5756 update_br_prob_note (test_bb);
5760 /* Move the insns out of MERGE_BB to before the branch. */
5761 if (head != NULL)
5763 rtx_insn *insn;
5765 if (end == BB_END (merge_bb))
5766 BB_END (merge_bb) = PREV_INSN (head);
5768 /* PR 21767: when moving insns above a conditional branch, the REG_EQUAL
5769 notes being moved might become invalid. */
5770 insn = head;
5773 rtx note;
5775 if (! INSN_P (insn))
5776 continue;
5777 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5778 if (! note)
5779 continue;
5780 remove_note (insn, note);
5781 } while (insn != end && (insn = NEXT_INSN (insn)));
5783 /* PR46315: when moving insns above a conditional branch, the REG_EQUAL
5784 notes referring to the registers being set might become invalid. */
5785 if (merge_set)
5787 unsigned i;
5788 bitmap_iterator bi;
5790 EXECUTE_IF_SET_IN_BITMAP (merge_set, 0, i, bi)
5791 remove_reg_equal_equiv_notes_for_regno (i);
5793 BITMAP_FREE (merge_set);
5796 reorder_insns (head, end, PREV_INSN (earliest));
5799 /* Remove the jump and edge if we can. */
5800 if (other_bb == new_dest)
5802 delete_insn (jump);
5803 remove_edge (BRANCH_EDGE (test_bb));
5804 /* ??? Can't merge blocks here, as then_bb is still in use.
5805 At minimum, the merge will get done just before bb-reorder. */
5808 return TRUE;
5810 cancel:
5811 cancel_changes (0);
5813 if (merge_set)
5814 BITMAP_FREE (merge_set);
5816 return FALSE;
5819 /* Main entry point for all if-conversion. AFTER_COMBINE is true if
5820 we are after combine pass. */
5822 static void
5823 if_convert (bool after_combine)
5825 basic_block bb;
5826 int pass;
5828 if (optimize == 1)
5830 df_live_add_problem ();
5831 df_live_set_all_dirty ();
5834 /* Record whether we are after combine pass. */
5835 ifcvt_after_combine = after_combine;
5836 have_cbranchcc4 = (direct_optab_handler (cbranch_optab, CCmode)
5837 != CODE_FOR_nothing);
5838 num_possible_if_blocks = 0;
5839 num_updated_if_blocks = 0;
5840 num_true_changes = 0;
5842 loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
5843 mark_loop_exit_edges ();
5844 loop_optimizer_finalize ();
5845 free_dominance_info (CDI_DOMINATORS);
5847 /* Compute postdominators. */
5848 calculate_dominance_info (CDI_POST_DOMINATORS);
5850 df_set_flags (DF_LR_RUN_DCE);
5852 /* Go through each of the basic blocks looking for things to convert. If we
5853 have conditional execution, we make multiple passes to allow us to handle
5854 IF-THEN{-ELSE} blocks within other IF-THEN{-ELSE} blocks. */
5855 pass = 0;
5858 df_analyze ();
5859 /* Only need to do dce on the first pass. */
5860 df_clear_flags (DF_LR_RUN_DCE);
5861 cond_exec_changed_p = FALSE;
5862 pass++;
5864 #ifdef IFCVT_MULTIPLE_DUMPS
5865 if (dump_file && pass > 1)
5866 fprintf (dump_file, "\n\n========== Pass %d ==========\n", pass);
5867 #endif
5869 FOR_EACH_BB_FN (bb, cfun)
5871 basic_block new_bb;
5872 while (!df_get_bb_dirty (bb)
5873 && (new_bb = find_if_header (bb, pass)) != NULL)
5874 bb = new_bb;
5877 #ifdef IFCVT_MULTIPLE_DUMPS
5878 if (dump_file && cond_exec_changed_p)
5879 print_rtl_with_bb (dump_file, get_insns (), dump_flags);
5880 #endif
5882 while (cond_exec_changed_p);
5884 #ifdef IFCVT_MULTIPLE_DUMPS
5885 if (dump_file)
5886 fprintf (dump_file, "\n\n========== no more changes\n");
5887 #endif
5889 free_dominance_info (CDI_POST_DOMINATORS);
5891 if (dump_file)
5892 fflush (dump_file);
5894 clear_aux_for_blocks ();
5896 /* If we allocated new pseudos, we must resize the array for sched1. */
5897 if (max_regno < max_reg_num ())
5898 max_regno = max_reg_num ();
5900 /* Write the final stats. */
5901 if (dump_file && num_possible_if_blocks > 0)
5903 fprintf (dump_file,
5904 "\n%d possible IF blocks searched.\n",
5905 num_possible_if_blocks);
5906 fprintf (dump_file,
5907 "%d IF blocks converted.\n",
5908 num_updated_if_blocks);
5909 fprintf (dump_file,
5910 "%d true changes made.\n\n\n",
5911 num_true_changes);
5914 if (optimize == 1)
5915 df_remove_problem (df_live);
5917 /* Some non-cold blocks may now be only reachable from cold blocks.
5918 Fix that up. */
5919 fixup_partitions ();
5921 checking_verify_flow_info ();
5924 /* If-conversion and CFG cleanup. */
5925 static unsigned int
5926 rest_of_handle_if_conversion (void)
5928 int flags = 0;
5930 if (flag_if_conversion)
5932 if (dump_file)
5934 dump_reg_info (dump_file);
5935 dump_flow_info (dump_file, dump_flags);
5937 cleanup_cfg (CLEANUP_EXPENSIVE);
5938 if_convert (false);
5939 if (num_updated_if_blocks)
5940 /* Get rid of any dead CC-related instructions. */
5941 flags |= CLEANUP_FORCE_FAST_DCE;
5944 cleanup_cfg (flags);
5945 return 0;
5948 namespace {
5950 const pass_data pass_data_rtl_ifcvt =
5952 RTL_PASS, /* type */
5953 "ce1", /* name */
5954 OPTGROUP_NONE, /* optinfo_flags */
5955 TV_IFCVT, /* tv_id */
5956 0, /* properties_required */
5957 0, /* properties_provided */
5958 0, /* properties_destroyed */
5959 0, /* todo_flags_start */
5960 TODO_df_finish, /* todo_flags_finish */
5963 class pass_rtl_ifcvt : public rtl_opt_pass
5965 public:
5966 pass_rtl_ifcvt (gcc::context *ctxt)
5967 : rtl_opt_pass (pass_data_rtl_ifcvt, ctxt)
5970 /* opt_pass methods: */
5971 bool gate (function *) final override
5973 return (optimize > 0) && dbg_cnt (if_conversion);
5976 unsigned int execute (function *) final override
5978 return rest_of_handle_if_conversion ();
5981 }; // class pass_rtl_ifcvt
5983 } // anon namespace
5985 rtl_opt_pass *
5986 make_pass_rtl_ifcvt (gcc::context *ctxt)
5988 return new pass_rtl_ifcvt (ctxt);
5992 /* Rerun if-conversion, as combine may have simplified things enough
5993 to now meet sequence length restrictions. */
5995 namespace {
5997 const pass_data pass_data_if_after_combine =
5999 RTL_PASS, /* type */
6000 "ce2", /* name */
6001 OPTGROUP_NONE, /* optinfo_flags */
6002 TV_IFCVT, /* tv_id */
6003 0, /* properties_required */
6004 0, /* properties_provided */
6005 0, /* properties_destroyed */
6006 0, /* todo_flags_start */
6007 TODO_df_finish, /* todo_flags_finish */
6010 class pass_if_after_combine : public rtl_opt_pass
6012 public:
6013 pass_if_after_combine (gcc::context *ctxt)
6014 : rtl_opt_pass (pass_data_if_after_combine, ctxt)
6017 /* opt_pass methods: */
6018 bool gate (function *) final override
6020 return optimize > 0 && flag_if_conversion
6021 && dbg_cnt (if_after_combine);
6024 unsigned int execute (function *) final override
6026 if_convert (true);
6027 return 0;
6030 }; // class pass_if_after_combine
6032 } // anon namespace
6034 rtl_opt_pass *
6035 make_pass_if_after_combine (gcc::context *ctxt)
6037 return new pass_if_after_combine (ctxt);
6041 namespace {
6043 const pass_data pass_data_if_after_reload =
6045 RTL_PASS, /* type */
6046 "ce3", /* name */
6047 OPTGROUP_NONE, /* optinfo_flags */
6048 TV_IFCVT2, /* tv_id */
6049 0, /* properties_required */
6050 0, /* properties_provided */
6051 0, /* properties_destroyed */
6052 0, /* todo_flags_start */
6053 TODO_df_finish, /* todo_flags_finish */
6056 class pass_if_after_reload : public rtl_opt_pass
6058 public:
6059 pass_if_after_reload (gcc::context *ctxt)
6060 : rtl_opt_pass (pass_data_if_after_reload, ctxt)
6063 /* opt_pass methods: */
6064 bool gate (function *) final override
6066 return optimize > 0 && flag_if_conversion2
6067 && dbg_cnt (if_after_reload);
6070 unsigned int execute (function *) final override
6072 if_convert (true);
6073 return 0;
6076 }; // class pass_if_after_reload
6078 } // anon namespace
6080 rtl_opt_pass *
6081 make_pass_if_after_reload (gcc::context *ctxt)
6083 return new pass_if_after_reload (ctxt);